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 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
151 if (!mem_cgroup_disabled())
152 return mem_cgroup_get_lru_size(lruvec
, lru
);
154 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
158 * Add a shrinker callback to be called from the vm
160 void register_shrinker(struct shrinker
*shrinker
)
162 atomic_long_set(&shrinker
->nr_in_batch
, 0);
163 down_write(&shrinker_rwsem
);
164 list_add_tail(&shrinker
->list
, &shrinker_list
);
165 up_write(&shrinker_rwsem
);
167 EXPORT_SYMBOL(register_shrinker
);
172 void unregister_shrinker(struct shrinker
*shrinker
)
174 down_write(&shrinker_rwsem
);
175 list_del(&shrinker
->list
);
176 up_write(&shrinker_rwsem
);
178 EXPORT_SYMBOL(unregister_shrinker
);
180 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
181 struct shrink_control
*sc
,
182 unsigned long nr_to_scan
)
184 sc
->nr_to_scan
= nr_to_scan
;
185 return (*shrinker
->shrink
)(shrinker
, sc
);
188 #define SHRINK_BATCH 128
190 * Call the shrink functions to age shrinkable caches
192 * Here we assume it costs one seek to replace a lru page and that it also
193 * takes a seek to recreate a cache object. With this in mind we age equal
194 * percentages of the lru and ageable caches. This should balance the seeks
195 * generated by these structures.
197 * If the vm encountered mapped pages on the LRU it increase the pressure on
198 * slab to avoid swapping.
200 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
202 * `lru_pages' represents the number of on-LRU pages in all the zones which
203 * are eligible for the caller's allocation attempt. It is used for balancing
204 * slab reclaim versus page reclaim.
206 * Returns the number of slab objects which we shrunk.
208 unsigned long shrink_slab(struct shrink_control
*shrink
,
209 unsigned long nr_pages_scanned
,
210 unsigned long lru_pages
)
212 struct shrinker
*shrinker
;
213 unsigned long ret
= 0;
215 if (nr_pages_scanned
== 0)
216 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
218 if (!down_read_trylock(&shrinker_rwsem
)) {
219 /* Assume we'll be able to shrink next time */
224 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
225 unsigned long long delta
;
231 long batch_size
= shrinker
->batch
? shrinker
->batch
234 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
243 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
246 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
248 do_div(delta
, lru_pages
+ 1);
250 if (total_scan
< 0) {
251 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
253 shrinker
->shrink
, total_scan
);
254 total_scan
= max_pass
;
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * max_pass. This is bad for sustaining a working set in
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
269 if (delta
< max_pass
/ 4)
270 total_scan
= min(total_scan
, max_pass
/ 2);
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
277 if (total_scan
> max_pass
* 2)
278 total_scan
= max_pass
* 2;
280 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
281 nr_pages_scanned
, lru_pages
,
282 max_pass
, delta
, total_scan
);
284 while (total_scan
>= batch_size
) {
287 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
288 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
290 if (shrink_ret
== -1)
292 if (shrink_ret
< nr_before
)
293 ret
+= nr_before
- shrink_ret
;
294 count_vm_events(SLABS_SCANNED
, batch_size
);
295 total_scan
-= batch_size
;
301 * move the unused scan count back into the shrinker in a
302 * manner that handles concurrent updates. If we exhausted the
303 * scan, there is no need to do an update.
306 new_nr
= atomic_long_add_return(total_scan
,
307 &shrinker
->nr_in_batch
);
309 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
311 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
313 up_read(&shrinker_rwsem
);
319 static inline int is_page_cache_freeable(struct page
*page
)
322 * A freeable page cache page is referenced only by the caller
323 * that isolated the page, the page cache radix tree and
324 * optional buffer heads at page->private.
326 return page_count(page
) - page_has_private(page
) == 2;
329 static int may_write_to_queue(struct backing_dev_info
*bdi
,
330 struct scan_control
*sc
)
332 if (current
->flags
& PF_SWAPWRITE
)
334 if (!bdi_write_congested(bdi
))
336 if (bdi
== current
->backing_dev_info
)
342 * We detected a synchronous write error writing a page out. Probably
343 * -ENOSPC. We need to propagate that into the address_space for a subsequent
344 * fsync(), msync() or close().
346 * The tricky part is that after writepage we cannot touch the mapping: nothing
347 * prevents it from being freed up. But we have a ref on the page and once
348 * that page is locked, the mapping is pinned.
350 * We're allowed to run sleeping lock_page() here because we know the caller has
353 static void handle_write_error(struct address_space
*mapping
,
354 struct page
*page
, int error
)
357 if (page_mapping(page
) == mapping
)
358 mapping_set_error(mapping
, error
);
362 /* possible outcome of pageout() */
364 /* failed to write page out, page is locked */
366 /* move page to the active list, page is locked */
368 /* page has been sent to the disk successfully, page is unlocked */
370 /* page is clean and locked */
375 * pageout is called by shrink_page_list() for each dirty page.
376 * Calls ->writepage().
378 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
379 struct scan_control
*sc
)
382 * If the page is dirty, only perform writeback if that write
383 * will be non-blocking. To prevent this allocation from being
384 * stalled by pagecache activity. But note that there may be
385 * stalls if we need to run get_block(). We could test
386 * PagePrivate for that.
388 * If this process is currently in __generic_file_aio_write() against
389 * this page's queue, we can perform writeback even if that
392 * If the page is swapcache, write it back even if that would
393 * block, for some throttling. This happens by accident, because
394 * swap_backing_dev_info is bust: it doesn't reflect the
395 * congestion state of the swapdevs. Easy to fix, if needed.
397 if (!is_page_cache_freeable(page
))
401 * Some data journaling orphaned pages can have
402 * page->mapping == NULL while being dirty with clean buffers.
404 if (page_has_private(page
)) {
405 if (try_to_free_buffers(page
)) {
406 ClearPageDirty(page
);
407 printk("%s: orphaned page\n", __func__
);
413 if (mapping
->a_ops
->writepage
== NULL
)
414 return PAGE_ACTIVATE
;
415 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
418 if (clear_page_dirty_for_io(page
)) {
420 struct writeback_control wbc
= {
421 .sync_mode
= WB_SYNC_NONE
,
422 .nr_to_write
= SWAP_CLUSTER_MAX
,
424 .range_end
= LLONG_MAX
,
428 SetPageReclaim(page
);
429 res
= mapping
->a_ops
->writepage(page
, &wbc
);
431 handle_write_error(mapping
, page
, res
);
432 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
433 ClearPageReclaim(page
);
434 return PAGE_ACTIVATE
;
437 if (!PageWriteback(page
)) {
438 /* synchronous write or broken a_ops? */
439 ClearPageReclaim(page
);
441 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
442 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
450 * Same as remove_mapping, but if the page is removed from the mapping, it
451 * gets returned with a refcount of 0.
453 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
455 BUG_ON(!PageLocked(page
));
456 BUG_ON(mapping
!= page_mapping(page
));
458 spin_lock_irq(&mapping
->tree_lock
);
460 * The non racy check for a busy page.
462 * Must be careful with the order of the tests. When someone has
463 * a ref to the page, it may be possible that they dirty it then
464 * drop the reference. So if PageDirty is tested before page_count
465 * here, then the following race may occur:
467 * get_user_pages(&page);
468 * [user mapping goes away]
470 * !PageDirty(page) [good]
471 * SetPageDirty(page);
473 * !page_count(page) [good, discard it]
475 * [oops, our write_to data is lost]
477 * Reversing the order of the tests ensures such a situation cannot
478 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
479 * load is not satisfied before that of page->_count.
481 * Note that if SetPageDirty is always performed via set_page_dirty,
482 * and thus under tree_lock, then this ordering is not required.
484 if (!page_freeze_refs(page
, 2))
486 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
487 if (unlikely(PageDirty(page
))) {
488 page_unfreeze_refs(page
, 2);
492 if (PageSwapCache(page
)) {
493 swp_entry_t swap
= { .val
= page_private(page
) };
494 __delete_from_swap_cache(page
);
495 spin_unlock_irq(&mapping
->tree_lock
);
496 swapcache_free(swap
, page
);
498 void (*freepage
)(struct page
*);
500 freepage
= mapping
->a_ops
->freepage
;
502 __delete_from_page_cache(page
);
503 spin_unlock_irq(&mapping
->tree_lock
);
504 mem_cgroup_uncharge_cache_page(page
);
506 if (freepage
!= NULL
)
513 spin_unlock_irq(&mapping
->tree_lock
);
518 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
519 * someone else has a ref on the page, abort and return 0. If it was
520 * successfully detached, return 1. Assumes the caller has a single ref on
523 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
525 if (__remove_mapping(mapping
, page
)) {
527 * Unfreezing the refcount with 1 rather than 2 effectively
528 * drops the pagecache ref for us without requiring another
531 page_unfreeze_refs(page
, 1);
538 * putback_lru_page - put previously isolated page onto appropriate LRU list
539 * @page: page to be put back to appropriate lru list
541 * Add previously isolated @page to appropriate LRU list.
542 * Page may still be unevictable for other reasons.
544 * lru_lock must not be held, interrupts must be enabled.
546 void putback_lru_page(struct page
*page
)
549 int active
= !!TestClearPageActive(page
);
550 int was_unevictable
= PageUnevictable(page
);
552 VM_BUG_ON(PageLRU(page
));
555 ClearPageUnevictable(page
);
557 if (page_evictable(page
)) {
559 * For evictable pages, we can use the cache.
560 * In event of a race, worst case is we end up with an
561 * unevictable page on [in]active list.
562 * We know how to handle that.
564 lru
= active
+ page_lru_base_type(page
);
565 lru_cache_add_lru(page
, lru
);
568 * Put unevictable pages directly on zone's unevictable
571 lru
= LRU_UNEVICTABLE
;
572 add_page_to_unevictable_list(page
);
574 * When racing with an mlock or AS_UNEVICTABLE clearing
575 * (page is unlocked) make sure that if the other thread
576 * does not observe our setting of PG_lru and fails
577 * isolation/check_move_unevictable_pages,
578 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
579 * the page back to the evictable list.
581 * The other side is TestClearPageMlocked() or shmem_lock().
587 * page's status can change while we move it among lru. If an evictable
588 * page is on unevictable list, it never be freed. To avoid that,
589 * check after we added it to the list, again.
591 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
)) {
592 if (!isolate_lru_page(page
)) {
596 /* This means someone else dropped this page from LRU
597 * So, it will be freed or putback to LRU again. There is
598 * nothing to do here.
602 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
603 count_vm_event(UNEVICTABLE_PGRESCUED
);
604 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
605 count_vm_event(UNEVICTABLE_PGCULLED
);
607 put_page(page
); /* drop ref from isolate */
610 enum page_references
{
612 PAGEREF_RECLAIM_CLEAN
,
617 static enum page_references
page_check_references(struct page
*page
,
618 struct scan_control
*sc
)
620 int referenced_ptes
, referenced_page
;
621 unsigned long vm_flags
;
623 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
625 referenced_page
= TestClearPageReferenced(page
);
628 * Mlock lost the isolation race with us. Let try_to_unmap()
629 * move the page to the unevictable list.
631 if (vm_flags
& VM_LOCKED
)
632 return PAGEREF_RECLAIM
;
634 if (referenced_ptes
) {
635 if (PageSwapBacked(page
))
636 return PAGEREF_ACTIVATE
;
638 * All mapped pages start out with page table
639 * references from the instantiating fault, so we need
640 * to look twice if a mapped file page is used more
643 * Mark it and spare it for another trip around the
644 * inactive list. Another page table reference will
645 * lead to its activation.
647 * Note: the mark is set for activated pages as well
648 * so that recently deactivated but used pages are
651 SetPageReferenced(page
);
653 if (referenced_page
|| referenced_ptes
> 1)
654 return PAGEREF_ACTIVATE
;
657 * Activate file-backed executable pages after first usage.
659 if (vm_flags
& VM_EXEC
)
660 return PAGEREF_ACTIVATE
;
665 /* Reclaim if clean, defer dirty pages to writeback */
666 if (referenced_page
&& !PageSwapBacked(page
))
667 return PAGEREF_RECLAIM_CLEAN
;
669 return PAGEREF_RECLAIM
;
672 /* Check if a page is dirty or under writeback */
673 static void page_check_dirty_writeback(struct page
*page
,
674 bool *dirty
, bool *writeback
)
677 * Anonymous pages are not handled by flushers and must be written
678 * from reclaim context. Do not stall reclaim based on them
680 if (!page_is_file_cache(page
)) {
686 /* By default assume that the page flags are accurate */
687 *dirty
= PageDirty(page
);
688 *writeback
= PageWriteback(page
);
692 * shrink_page_list() returns the number of reclaimed pages
694 static unsigned long shrink_page_list(struct list_head
*page_list
,
696 struct scan_control
*sc
,
697 enum ttu_flags ttu_flags
,
698 unsigned long *ret_nr_dirty
,
699 unsigned long *ret_nr_unqueued_dirty
,
700 unsigned long *ret_nr_congested
,
701 unsigned long *ret_nr_writeback
,
702 unsigned long *ret_nr_immediate
,
705 LIST_HEAD(ret_pages
);
706 LIST_HEAD(free_pages
);
708 unsigned long nr_unqueued_dirty
= 0;
709 unsigned long nr_dirty
= 0;
710 unsigned long nr_congested
= 0;
711 unsigned long nr_reclaimed
= 0;
712 unsigned long nr_writeback
= 0;
713 unsigned long nr_immediate
= 0;
717 mem_cgroup_uncharge_start();
718 while (!list_empty(page_list
)) {
719 struct address_space
*mapping
;
722 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
723 bool dirty
, writeback
;
727 page
= lru_to_page(page_list
);
728 list_del(&page
->lru
);
730 if (!trylock_page(page
))
733 VM_BUG_ON(PageActive(page
));
734 VM_BUG_ON(page_zone(page
) != zone
);
738 if (unlikely(!page_evictable(page
)))
741 if (!sc
->may_unmap
&& page_mapped(page
))
744 /* Double the slab pressure for mapped and swapcache pages */
745 if (page_mapped(page
) || PageSwapCache(page
))
748 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
749 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
752 * The number of dirty pages determines if a zone is marked
753 * reclaim_congested which affects wait_iff_congested. kswapd
754 * will stall and start writing pages if the tail of the LRU
755 * is all dirty unqueued pages.
757 page_check_dirty_writeback(page
, &dirty
, &writeback
);
758 if (dirty
|| writeback
)
761 if (dirty
&& !writeback
)
764 /* Treat this page as congested if underlying BDI is */
765 mapping
= page_mapping(page
);
766 if (mapping
&& bdi_write_congested(mapping
->backing_dev_info
))
770 * If a page at the tail of the LRU is under writeback, there
771 * are three cases to consider.
773 * 1) If reclaim is encountering an excessive number of pages
774 * under writeback and this page is both under writeback and
775 * PageReclaim then it indicates that pages are being queued
776 * for IO but are being recycled through the LRU before the
777 * IO can complete. Waiting on the page itself risks an
778 * indefinite stall if it is impossible to writeback the
779 * page due to IO error or disconnected storage so instead
780 * note that the LRU is being scanned too quickly and the
781 * caller can stall after page list has been processed.
783 * 2) Global reclaim encounters a page, memcg encounters a
784 * page that is not marked for immediate reclaim or
785 * the caller does not have __GFP_IO. In this case mark
786 * the page for immediate reclaim and continue scanning.
788 * __GFP_IO is checked because a loop driver thread might
789 * enter reclaim, and deadlock if it waits on a page for
790 * which it is needed to do the write (loop masks off
791 * __GFP_IO|__GFP_FS for this reason); but more thought
792 * would probably show more reasons.
794 * Don't require __GFP_FS, since we're not going into the
795 * FS, just waiting on its writeback completion. Worryingly,
796 * ext4 gfs2 and xfs allocate pages with
797 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
798 * may_enter_fs here is liable to OOM on them.
800 * 3) memcg encounters a page that is not already marked
801 * PageReclaim. memcg does not have any dirty pages
802 * throttling so we could easily OOM just because too many
803 * pages are in writeback and there is nothing else to
804 * reclaim. Wait for the writeback to complete.
806 if (PageWriteback(page
)) {
808 if (current_is_kswapd() &&
810 zone_is_reclaim_writeback(zone
)) {
815 } else if (global_reclaim(sc
) ||
816 !PageReclaim(page
) || !(sc
->gfp_mask
& __GFP_IO
)) {
818 * This is slightly racy - end_page_writeback()
819 * might have just cleared PageReclaim, then
820 * setting PageReclaim here end up interpreted
821 * as PageReadahead - but that does not matter
822 * enough to care. What we do want is for this
823 * page to have PageReclaim set next time memcg
824 * reclaim reaches the tests above, so it will
825 * then wait_on_page_writeback() to avoid OOM;
826 * and it's also appropriate in global reclaim.
828 SetPageReclaim(page
);
835 wait_on_page_writeback(page
);
840 references
= page_check_references(page
, sc
);
842 switch (references
) {
843 case PAGEREF_ACTIVATE
:
844 goto activate_locked
;
847 case PAGEREF_RECLAIM
:
848 case PAGEREF_RECLAIM_CLEAN
:
849 ; /* try to reclaim the page below */
853 * Anonymous process memory has backing store?
854 * Try to allocate it some swap space here.
856 if (PageAnon(page
) && !PageSwapCache(page
)) {
857 if (!(sc
->gfp_mask
& __GFP_IO
))
859 if (!add_to_swap(page
, page_list
))
860 goto activate_locked
;
863 /* Adding to swap updated mapping */
864 mapping
= page_mapping(page
);
868 * The page is mapped into the page tables of one or more
869 * processes. Try to unmap it here.
871 if (page_mapped(page
) && mapping
) {
872 switch (try_to_unmap(page
, ttu_flags
)) {
874 goto activate_locked
;
880 ; /* try to free the page below */
884 if (PageDirty(page
)) {
886 * Only kswapd can writeback filesystem pages to
887 * avoid risk of stack overflow but only writeback
888 * if many dirty pages have been encountered.
890 if (page_is_file_cache(page
) &&
891 (!current_is_kswapd() ||
892 !zone_is_reclaim_dirty(zone
))) {
894 * Immediately reclaim when written back.
895 * Similar in principal to deactivate_page()
896 * except we already have the page isolated
897 * and know it's dirty
899 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
900 SetPageReclaim(page
);
905 if (references
== PAGEREF_RECLAIM_CLEAN
)
909 if (!sc
->may_writepage
)
912 /* Page is dirty, try to write it out here */
913 switch (pageout(page
, mapping
, sc
)) {
917 goto activate_locked
;
919 if (PageWriteback(page
))
925 * A synchronous write - probably a ramdisk. Go
926 * ahead and try to reclaim the page.
928 if (!trylock_page(page
))
930 if (PageDirty(page
) || PageWriteback(page
))
932 mapping
= page_mapping(page
);
934 ; /* try to free the page below */
939 * If the page has buffers, try to free the buffer mappings
940 * associated with this page. If we succeed we try to free
943 * We do this even if the page is PageDirty().
944 * try_to_release_page() does not perform I/O, but it is
945 * possible for a page to have PageDirty set, but it is actually
946 * clean (all its buffers are clean). This happens if the
947 * buffers were written out directly, with submit_bh(). ext3
948 * will do this, as well as the blockdev mapping.
949 * try_to_release_page() will discover that cleanness and will
950 * drop the buffers and mark the page clean - it can be freed.
952 * Rarely, pages can have buffers and no ->mapping. These are
953 * the pages which were not successfully invalidated in
954 * truncate_complete_page(). We try to drop those buffers here
955 * and if that worked, and the page is no longer mapped into
956 * process address space (page_count == 1) it can be freed.
957 * Otherwise, leave the page on the LRU so it is swappable.
959 if (page_has_private(page
)) {
960 if (!try_to_release_page(page
, sc
->gfp_mask
))
961 goto activate_locked
;
962 if (!mapping
&& page_count(page
) == 1) {
964 if (put_page_testzero(page
))
968 * rare race with speculative reference.
969 * the speculative reference will free
970 * this page shortly, so we may
971 * increment nr_reclaimed here (and
972 * leave it off the LRU).
980 if (!mapping
|| !__remove_mapping(mapping
, page
))
984 * At this point, we have no other references and there is
985 * no way to pick any more up (removed from LRU, removed
986 * from pagecache). Can use non-atomic bitops now (and
987 * we obviously don't have to worry about waking up a process
988 * waiting on the page lock, because there are no references.
990 __clear_page_locked(page
);
995 * Is there need to periodically free_page_list? It would
996 * appear not as the counts should be low
998 list_add(&page
->lru
, &free_pages
);
1002 if (PageSwapCache(page
))
1003 try_to_free_swap(page
);
1005 putback_lru_page(page
);
1009 /* Not a candidate for swapping, so reclaim swap space. */
1010 if (PageSwapCache(page
) && vm_swap_full())
1011 try_to_free_swap(page
);
1012 VM_BUG_ON(PageActive(page
));
1013 SetPageActive(page
);
1018 list_add(&page
->lru
, &ret_pages
);
1019 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
1022 free_hot_cold_page_list(&free_pages
, 1);
1024 list_splice(&ret_pages
, page_list
);
1025 count_vm_events(PGACTIVATE
, pgactivate
);
1026 mem_cgroup_uncharge_end();
1027 *ret_nr_dirty
+= nr_dirty
;
1028 *ret_nr_congested
+= nr_congested
;
1029 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1030 *ret_nr_writeback
+= nr_writeback
;
1031 *ret_nr_immediate
+= nr_immediate
;
1032 return nr_reclaimed
;
1035 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1036 struct list_head
*page_list
)
1038 struct scan_control sc
= {
1039 .gfp_mask
= GFP_KERNEL
,
1040 .priority
= DEF_PRIORITY
,
1043 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1044 struct page
*page
, *next
;
1045 LIST_HEAD(clean_pages
);
1047 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1048 if (page_is_file_cache(page
) && !PageDirty(page
)) {
1049 ClearPageActive(page
);
1050 list_move(&page
->lru
, &clean_pages
);
1054 ret
= shrink_page_list(&clean_pages
, zone
, &sc
,
1055 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1056 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1057 list_splice(&clean_pages
, page_list
);
1058 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -ret
);
1063 * Attempt to remove the specified page from its LRU. Only take this page
1064 * if it is of the appropriate PageActive status. Pages which are being
1065 * freed elsewhere are also ignored.
1067 * page: page to consider
1068 * mode: one of the LRU isolation modes defined above
1070 * returns 0 on success, -ve errno on failure.
1072 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1076 /* Only take pages on the LRU. */
1080 /* Compaction should not handle unevictable pages but CMA can do so */
1081 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1087 * To minimise LRU disruption, the caller can indicate that it only
1088 * wants to isolate pages it will be able to operate on without
1089 * blocking - clean pages for the most part.
1091 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1092 * is used by reclaim when it is cannot write to backing storage
1094 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1095 * that it is possible to migrate without blocking
1097 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1098 /* All the caller can do on PageWriteback is block */
1099 if (PageWriteback(page
))
1102 if (PageDirty(page
)) {
1103 struct address_space
*mapping
;
1105 /* ISOLATE_CLEAN means only clean pages */
1106 if (mode
& ISOLATE_CLEAN
)
1110 * Only pages without mappings or that have a
1111 * ->migratepage callback are possible to migrate
1114 mapping
= page_mapping(page
);
1115 if (mapping
&& !mapping
->a_ops
->migratepage
)
1120 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1123 if (likely(get_page_unless_zero(page
))) {
1125 * Be careful not to clear PageLRU until after we're
1126 * sure the page is not being freed elsewhere -- the
1127 * page release code relies on it.
1137 * zone->lru_lock is heavily contended. Some of the functions that
1138 * shrink the lists perform better by taking out a batch of pages
1139 * and working on them outside the LRU lock.
1141 * For pagecache intensive workloads, this function is the hottest
1142 * spot in the kernel (apart from copy_*_user functions).
1144 * Appropriate locks must be held before calling this function.
1146 * @nr_to_scan: The number of pages to look through on the list.
1147 * @lruvec: The LRU vector to pull pages from.
1148 * @dst: The temp list to put pages on to.
1149 * @nr_scanned: The number of pages that were scanned.
1150 * @sc: The scan_control struct for this reclaim session
1151 * @mode: One of the LRU isolation modes
1152 * @lru: LRU list id for isolating
1154 * returns how many pages were moved onto *@dst.
1156 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1157 struct lruvec
*lruvec
, struct list_head
*dst
,
1158 unsigned long *nr_scanned
, struct scan_control
*sc
,
1159 isolate_mode_t mode
, enum lru_list lru
)
1161 struct list_head
*src
= &lruvec
->lists
[lru
];
1162 unsigned long nr_taken
= 0;
1165 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1169 page
= lru_to_page(src
);
1170 prefetchw_prev_lru_page(page
, src
, flags
);
1172 VM_BUG_ON(!PageLRU(page
));
1174 switch (__isolate_lru_page(page
, mode
)) {
1176 nr_pages
= hpage_nr_pages(page
);
1177 mem_cgroup_update_lru_size(lruvec
, lru
, -nr_pages
);
1178 list_move(&page
->lru
, dst
);
1179 nr_taken
+= nr_pages
;
1183 /* else it is being freed elsewhere */
1184 list_move(&page
->lru
, src
);
1193 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1194 nr_taken
, mode
, is_file_lru(lru
));
1199 * isolate_lru_page - tries to isolate a page from its LRU list
1200 * @page: page to isolate from its LRU list
1202 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1203 * vmstat statistic corresponding to whatever LRU list the page was on.
1205 * Returns 0 if the page was removed from an LRU list.
1206 * Returns -EBUSY if the page was not on an LRU list.
1208 * The returned page will have PageLRU() cleared. If it was found on
1209 * the active list, it will have PageActive set. If it was found on
1210 * the unevictable list, it will have the PageUnevictable bit set. That flag
1211 * may need to be cleared by the caller before letting the page go.
1213 * The vmstat statistic corresponding to the list on which the page was
1214 * found will be decremented.
1217 * (1) Must be called with an elevated refcount on the page. This is a
1218 * fundamentnal difference from isolate_lru_pages (which is called
1219 * without a stable reference).
1220 * (2) the lru_lock must not be held.
1221 * (3) interrupts must be enabled.
1223 int isolate_lru_page(struct page
*page
)
1227 VM_BUG_ON(!page_count(page
));
1229 if (PageLRU(page
)) {
1230 struct zone
*zone
= page_zone(page
);
1231 struct lruvec
*lruvec
;
1233 spin_lock_irq(&zone
->lru_lock
);
1234 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1235 if (PageLRU(page
)) {
1236 int lru
= page_lru(page
);
1239 del_page_from_lru_list(page
, lruvec
, lru
);
1242 spin_unlock_irq(&zone
->lru_lock
);
1248 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1249 * then get resheduled. When there are massive number of tasks doing page
1250 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1251 * the LRU list will go small and be scanned faster than necessary, leading to
1252 * unnecessary swapping, thrashing and OOM.
1254 static int too_many_isolated(struct zone
*zone
, int file
,
1255 struct scan_control
*sc
)
1257 unsigned long inactive
, isolated
;
1259 if (current_is_kswapd())
1262 if (!global_reclaim(sc
))
1266 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1267 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1269 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1270 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1274 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1275 * won't get blocked by normal direct-reclaimers, forming a circular
1278 if ((sc
->gfp_mask
& GFP_IOFS
) == GFP_IOFS
)
1281 return isolated
> inactive
;
1284 static noinline_for_stack
void
1285 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1287 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1288 struct zone
*zone
= lruvec_zone(lruvec
);
1289 LIST_HEAD(pages_to_free
);
1292 * Put back any unfreeable pages.
1294 while (!list_empty(page_list
)) {
1295 struct page
*page
= lru_to_page(page_list
);
1298 VM_BUG_ON(PageLRU(page
));
1299 list_del(&page
->lru
);
1300 if (unlikely(!page_evictable(page
))) {
1301 spin_unlock_irq(&zone
->lru_lock
);
1302 putback_lru_page(page
);
1303 spin_lock_irq(&zone
->lru_lock
);
1307 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1310 lru
= page_lru(page
);
1311 add_page_to_lru_list(page
, lruvec
, lru
);
1313 if (is_active_lru(lru
)) {
1314 int file
= is_file_lru(lru
);
1315 int numpages
= hpage_nr_pages(page
);
1316 reclaim_stat
->recent_rotated
[file
] += numpages
;
1318 if (put_page_testzero(page
)) {
1319 __ClearPageLRU(page
);
1320 __ClearPageActive(page
);
1321 del_page_from_lru_list(page
, lruvec
, lru
);
1323 if (unlikely(PageCompound(page
))) {
1324 spin_unlock_irq(&zone
->lru_lock
);
1325 (*get_compound_page_dtor(page
))(page
);
1326 spin_lock_irq(&zone
->lru_lock
);
1328 list_add(&page
->lru
, &pages_to_free
);
1333 * To save our caller's stack, now use input list for pages to free.
1335 list_splice(&pages_to_free
, page_list
);
1339 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1340 * of reclaimed pages
1342 static noinline_for_stack
unsigned long
1343 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1344 struct scan_control
*sc
, enum lru_list lru
)
1346 LIST_HEAD(page_list
);
1347 unsigned long nr_scanned
;
1348 unsigned long nr_reclaimed
= 0;
1349 unsigned long nr_taken
;
1350 unsigned long nr_dirty
= 0;
1351 unsigned long nr_congested
= 0;
1352 unsigned long nr_unqueued_dirty
= 0;
1353 unsigned long nr_writeback
= 0;
1354 unsigned long nr_immediate
= 0;
1355 isolate_mode_t isolate_mode
= 0;
1356 int file
= is_file_lru(lru
);
1357 struct zone
*zone
= lruvec_zone(lruvec
);
1358 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1360 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1361 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1363 /* We are about to die and free our memory. Return now. */
1364 if (fatal_signal_pending(current
))
1365 return SWAP_CLUSTER_MAX
;
1371 isolate_mode
|= ISOLATE_UNMAPPED
;
1372 if (!sc
->may_writepage
)
1373 isolate_mode
|= ISOLATE_CLEAN
;
1375 spin_lock_irq(&zone
->lru_lock
);
1377 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1378 &nr_scanned
, sc
, isolate_mode
, lru
);
1380 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1381 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1383 if (global_reclaim(sc
)) {
1384 zone
->pages_scanned
+= nr_scanned
;
1385 if (current_is_kswapd())
1386 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1388 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1390 spin_unlock_irq(&zone
->lru_lock
);
1395 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
, TTU_UNMAP
,
1396 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1397 &nr_writeback
, &nr_immediate
,
1400 spin_lock_irq(&zone
->lru_lock
);
1402 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1404 if (global_reclaim(sc
)) {
1405 if (current_is_kswapd())
1406 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1409 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1413 putback_inactive_pages(lruvec
, &page_list
);
1415 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1417 spin_unlock_irq(&zone
->lru_lock
);
1419 free_hot_cold_page_list(&page_list
, 1);
1422 * If reclaim is isolating dirty pages under writeback, it implies
1423 * that the long-lived page allocation rate is exceeding the page
1424 * laundering rate. Either the global limits are not being effective
1425 * at throttling processes due to the page distribution throughout
1426 * zones or there is heavy usage of a slow backing device. The
1427 * only option is to throttle from reclaim context which is not ideal
1428 * as there is no guarantee the dirtying process is throttled in the
1429 * same way balance_dirty_pages() manages.
1431 * This scales the number of dirty pages that must be under writeback
1432 * before a zone gets flagged ZONE_WRITEBACK. It is a simple backoff
1433 * function that has the most effect in the range DEF_PRIORITY to
1434 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1435 * in trouble and reclaim is considered to be in trouble.
1437 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1438 * DEF_PRIORITY-1 50% must be PageWriteback
1439 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1441 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1442 * isolated page is PageWriteback
1444 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1445 * of pages under pages flagged for immediate reclaim and stall if any
1446 * are encountered in the nr_immediate check below.
1448 if (nr_writeback
&& nr_writeback
>=
1449 (nr_taken
>> (DEF_PRIORITY
- sc
->priority
)))
1450 zone_set_flag(zone
, ZONE_WRITEBACK
);
1453 * memcg will stall in page writeback so only consider forcibly
1454 * stalling for global reclaim
1456 if (global_reclaim(sc
)) {
1458 * Tag a zone as congested if all the dirty pages scanned were
1459 * backed by a congested BDI and wait_iff_congested will stall.
1461 if (nr_dirty
&& nr_dirty
== nr_congested
)
1462 zone_set_flag(zone
, ZONE_CONGESTED
);
1465 * If dirty pages are scanned that are not queued for IO, it
1466 * implies that flushers are not keeping up. In this case, flag
1467 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1468 * pages from reclaim context. It will forcibly stall in the
1471 if (nr_unqueued_dirty
== nr_taken
)
1472 zone_set_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
1475 * In addition, if kswapd scans pages marked marked for
1476 * immediate reclaim and under writeback (nr_immediate), it
1477 * implies that pages are cycling through the LRU faster than
1478 * they are written so also forcibly stall.
1480 if (nr_unqueued_dirty
== nr_taken
|| nr_immediate
)
1481 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1485 * Stall direct reclaim for IO completions if underlying BDIs or zone
1486 * is congested. Allow kswapd to continue until it starts encountering
1487 * unqueued dirty pages or cycling through the LRU too quickly.
1489 if (!sc
->hibernation_mode
&& !current_is_kswapd())
1490 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1492 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1494 nr_scanned
, nr_reclaimed
,
1496 trace_shrink_flags(file
));
1497 return nr_reclaimed
;
1501 * This moves pages from the active list to the inactive list.
1503 * We move them the other way if the page is referenced by one or more
1504 * processes, from rmap.
1506 * If the pages are mostly unmapped, the processing is fast and it is
1507 * appropriate to hold zone->lru_lock across the whole operation. But if
1508 * the pages are mapped, the processing is slow (page_referenced()) so we
1509 * should drop zone->lru_lock around each page. It's impossible to balance
1510 * this, so instead we remove the pages from the LRU while processing them.
1511 * It is safe to rely on PG_active against the non-LRU pages in here because
1512 * nobody will play with that bit on a non-LRU page.
1514 * The downside is that we have to touch page->_count against each page.
1515 * But we had to alter page->flags anyway.
1518 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1519 struct list_head
*list
,
1520 struct list_head
*pages_to_free
,
1523 struct zone
*zone
= lruvec_zone(lruvec
);
1524 unsigned long pgmoved
= 0;
1528 while (!list_empty(list
)) {
1529 page
= lru_to_page(list
);
1530 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
1532 VM_BUG_ON(PageLRU(page
));
1535 nr_pages
= hpage_nr_pages(page
);
1536 mem_cgroup_update_lru_size(lruvec
, lru
, nr_pages
);
1537 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1538 pgmoved
+= nr_pages
;
1540 if (put_page_testzero(page
)) {
1541 __ClearPageLRU(page
);
1542 __ClearPageActive(page
);
1543 del_page_from_lru_list(page
, lruvec
, lru
);
1545 if (unlikely(PageCompound(page
))) {
1546 spin_unlock_irq(&zone
->lru_lock
);
1547 (*get_compound_page_dtor(page
))(page
);
1548 spin_lock_irq(&zone
->lru_lock
);
1550 list_add(&page
->lru
, pages_to_free
);
1553 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1554 if (!is_active_lru(lru
))
1555 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1558 static void shrink_active_list(unsigned long nr_to_scan
,
1559 struct lruvec
*lruvec
,
1560 struct scan_control
*sc
,
1563 unsigned long nr_taken
;
1564 unsigned long nr_scanned
;
1565 unsigned long vm_flags
;
1566 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1567 LIST_HEAD(l_active
);
1568 LIST_HEAD(l_inactive
);
1570 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1571 unsigned long nr_rotated
= 0;
1572 isolate_mode_t isolate_mode
= 0;
1573 int file
= is_file_lru(lru
);
1574 struct zone
*zone
= lruvec_zone(lruvec
);
1579 isolate_mode
|= ISOLATE_UNMAPPED
;
1580 if (!sc
->may_writepage
)
1581 isolate_mode
|= ISOLATE_CLEAN
;
1583 spin_lock_irq(&zone
->lru_lock
);
1585 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1586 &nr_scanned
, sc
, isolate_mode
, lru
);
1587 if (global_reclaim(sc
))
1588 zone
->pages_scanned
+= nr_scanned
;
1590 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1592 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1593 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1594 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1595 spin_unlock_irq(&zone
->lru_lock
);
1597 while (!list_empty(&l_hold
)) {
1599 page
= lru_to_page(&l_hold
);
1600 list_del(&page
->lru
);
1602 if (unlikely(!page_evictable(page
))) {
1603 putback_lru_page(page
);
1607 if (unlikely(buffer_heads_over_limit
)) {
1608 if (page_has_private(page
) && trylock_page(page
)) {
1609 if (page_has_private(page
))
1610 try_to_release_page(page
, 0);
1615 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1617 nr_rotated
+= hpage_nr_pages(page
);
1619 * Identify referenced, file-backed active pages and
1620 * give them one more trip around the active list. So
1621 * that executable code get better chances to stay in
1622 * memory under moderate memory pressure. Anon pages
1623 * are not likely to be evicted by use-once streaming
1624 * IO, plus JVM can create lots of anon VM_EXEC pages,
1625 * so we ignore them here.
1627 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1628 list_add(&page
->lru
, &l_active
);
1633 ClearPageActive(page
); /* we are de-activating */
1634 list_add(&page
->lru
, &l_inactive
);
1638 * Move pages back to the lru list.
1640 spin_lock_irq(&zone
->lru_lock
);
1642 * Count referenced pages from currently used mappings as rotated,
1643 * even though only some of them are actually re-activated. This
1644 * helps balance scan pressure between file and anonymous pages in
1647 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1649 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1650 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1651 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1652 spin_unlock_irq(&zone
->lru_lock
);
1654 free_hot_cold_page_list(&l_hold
, 1);
1658 static int inactive_anon_is_low_global(struct zone
*zone
)
1660 unsigned long active
, inactive
;
1662 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1663 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1665 if (inactive
* zone
->inactive_ratio
< active
)
1672 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1673 * @lruvec: LRU vector to check
1675 * Returns true if the zone does not have enough inactive anon pages,
1676 * meaning some active anon pages need to be deactivated.
1678 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1681 * If we don't have swap space, anonymous page deactivation
1684 if (!total_swap_pages
)
1687 if (!mem_cgroup_disabled())
1688 return mem_cgroup_inactive_anon_is_low(lruvec
);
1690 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1693 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1700 * inactive_file_is_low - check if file pages need to be deactivated
1701 * @lruvec: LRU vector to check
1703 * When the system is doing streaming IO, memory pressure here
1704 * ensures that active file pages get deactivated, until more
1705 * than half of the file pages are on the inactive list.
1707 * Once we get to that situation, protect the system's working
1708 * set from being evicted by disabling active file page aging.
1710 * This uses a different ratio than the anonymous pages, because
1711 * the page cache uses a use-once replacement algorithm.
1713 static int inactive_file_is_low(struct lruvec
*lruvec
)
1715 unsigned long inactive
;
1716 unsigned long active
;
1718 inactive
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1719 active
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1721 return active
> inactive
;
1724 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1726 if (is_file_lru(lru
))
1727 return inactive_file_is_low(lruvec
);
1729 return inactive_anon_is_low(lruvec
);
1732 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1733 struct lruvec
*lruvec
, struct scan_control
*sc
)
1735 if (is_active_lru(lru
)) {
1736 if (inactive_list_is_low(lruvec
, lru
))
1737 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1741 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1744 static int vmscan_swappiness(struct scan_control
*sc
)
1746 if (global_reclaim(sc
))
1747 return vm_swappiness
;
1748 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1759 * Determine how aggressively the anon and file LRU lists should be
1760 * scanned. The relative value of each set of LRU lists is determined
1761 * by looking at the fraction of the pages scanned we did rotate back
1762 * onto the active list instead of evict.
1764 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1765 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1767 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
1770 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1772 u64 denominator
= 0; /* gcc */
1773 struct zone
*zone
= lruvec_zone(lruvec
);
1774 unsigned long anon_prio
, file_prio
;
1775 enum scan_balance scan_balance
;
1776 unsigned long anon
, file
, free
;
1777 bool force_scan
= false;
1778 unsigned long ap
, fp
;
1782 * If the zone or memcg is small, nr[l] can be 0. This
1783 * results in no scanning on this priority and a potential
1784 * priority drop. Global direct reclaim can go to the next
1785 * zone and tends to have no problems. Global kswapd is for
1786 * zone balancing and it needs to scan a minimum amount. When
1787 * reclaiming for a memcg, a priority drop can cause high
1788 * latencies, so it's better to scan a minimum amount there as
1791 if (current_is_kswapd() && zone
->all_unreclaimable
)
1793 if (!global_reclaim(sc
))
1796 /* If we have no swap space, do not bother scanning anon pages. */
1797 if (!sc
->may_swap
|| (get_nr_swap_pages() <= 0)) {
1798 scan_balance
= SCAN_FILE
;
1803 * Global reclaim will swap to prevent OOM even with no
1804 * swappiness, but memcg users want to use this knob to
1805 * disable swapping for individual groups completely when
1806 * using the memory controller's swap limit feature would be
1809 if (!global_reclaim(sc
) && !vmscan_swappiness(sc
)) {
1810 scan_balance
= SCAN_FILE
;
1815 * Do not apply any pressure balancing cleverness when the
1816 * system is close to OOM, scan both anon and file equally
1817 * (unless the swappiness setting disagrees with swapping).
1819 if (!sc
->priority
&& vmscan_swappiness(sc
)) {
1820 scan_balance
= SCAN_EQUAL
;
1824 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1825 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1826 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1827 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1830 * If it's foreseeable that reclaiming the file cache won't be
1831 * enough to get the zone back into a desirable shape, we have
1832 * to swap. Better start now and leave the - probably heavily
1833 * thrashing - remaining file pages alone.
1835 if (global_reclaim(sc
)) {
1836 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1837 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1838 scan_balance
= SCAN_ANON
;
1844 * There is enough inactive page cache, do not reclaim
1845 * anything from the anonymous working set right now.
1847 if (!inactive_file_is_low(lruvec
)) {
1848 scan_balance
= SCAN_FILE
;
1852 scan_balance
= SCAN_FRACT
;
1855 * With swappiness at 100, anonymous and file have the same priority.
1856 * This scanning priority is essentially the inverse of IO cost.
1858 anon_prio
= vmscan_swappiness(sc
);
1859 file_prio
= 200 - anon_prio
;
1862 * OK, so we have swap space and a fair amount of page cache
1863 * pages. We use the recently rotated / recently scanned
1864 * ratios to determine how valuable each cache is.
1866 * Because workloads change over time (and to avoid overflow)
1867 * we keep these statistics as a floating average, which ends
1868 * up weighing recent references more than old ones.
1870 * anon in [0], file in [1]
1872 spin_lock_irq(&zone
->lru_lock
);
1873 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1874 reclaim_stat
->recent_scanned
[0] /= 2;
1875 reclaim_stat
->recent_rotated
[0] /= 2;
1878 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1879 reclaim_stat
->recent_scanned
[1] /= 2;
1880 reclaim_stat
->recent_rotated
[1] /= 2;
1884 * The amount of pressure on anon vs file pages is inversely
1885 * proportional to the fraction of recently scanned pages on
1886 * each list that were recently referenced and in active use.
1888 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1889 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1891 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1892 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1893 spin_unlock_irq(&zone
->lru_lock
);
1897 denominator
= ap
+ fp
+ 1;
1899 for_each_evictable_lru(lru
) {
1900 int file
= is_file_lru(lru
);
1904 size
= get_lru_size(lruvec
, lru
);
1905 scan
= size
>> sc
->priority
;
1907 if (!scan
&& force_scan
)
1908 scan
= min(size
, SWAP_CLUSTER_MAX
);
1910 switch (scan_balance
) {
1912 /* Scan lists relative to size */
1916 * Scan types proportional to swappiness and
1917 * their relative recent reclaim efficiency.
1919 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1923 /* Scan one type exclusively */
1924 if ((scan_balance
== SCAN_FILE
) != file
)
1928 /* Look ma, no brain */
1936 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1938 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
1940 unsigned long nr
[NR_LRU_LISTS
];
1941 unsigned long targets
[NR_LRU_LISTS
];
1942 unsigned long nr_to_scan
;
1944 unsigned long nr_reclaimed
= 0;
1945 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1946 struct blk_plug plug
;
1947 bool scan_adjusted
= false;
1949 get_scan_count(lruvec
, sc
, nr
);
1951 /* Record the original scan target for proportional adjustments later */
1952 memcpy(targets
, nr
, sizeof(nr
));
1954 blk_start_plug(&plug
);
1955 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1956 nr
[LRU_INACTIVE_FILE
]) {
1957 unsigned long nr_anon
, nr_file
, percentage
;
1958 unsigned long nr_scanned
;
1960 for_each_evictable_lru(lru
) {
1962 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
1963 nr
[lru
] -= nr_to_scan
;
1965 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1970 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
1974 * For global direct reclaim, reclaim only the number of pages
1975 * requested. Less care is taken to scan proportionally as it
1976 * is more important to minimise direct reclaim stall latency
1977 * than it is to properly age the LRU lists.
1979 if (global_reclaim(sc
) && !current_is_kswapd())
1983 * For kswapd and memcg, reclaim at least the number of pages
1984 * requested. Ensure that the anon and file LRUs shrink
1985 * proportionally what was requested by get_scan_count(). We
1986 * stop reclaiming one LRU and reduce the amount scanning
1987 * proportional to the original scan target.
1989 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
1990 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
1992 if (nr_file
> nr_anon
) {
1993 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
1994 targets
[LRU_ACTIVE_ANON
] + 1;
1996 percentage
= nr_anon
* 100 / scan_target
;
1998 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
1999 targets
[LRU_ACTIVE_FILE
] + 1;
2001 percentage
= nr_file
* 100 / scan_target
;
2004 /* Stop scanning the smaller of the LRU */
2006 nr
[lru
+ LRU_ACTIVE
] = 0;
2009 * Recalculate the other LRU scan count based on its original
2010 * scan target and the percentage scanning already complete
2012 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2013 nr_scanned
= targets
[lru
] - nr
[lru
];
2014 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2015 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2018 nr_scanned
= targets
[lru
] - nr
[lru
];
2019 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2020 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2022 scan_adjusted
= true;
2024 blk_finish_plug(&plug
);
2025 sc
->nr_reclaimed
+= nr_reclaimed
;
2028 * Even if we did not try to evict anon pages at all, we want to
2029 * rebalance the anon lru active/inactive ratio.
2031 if (inactive_anon_is_low(lruvec
))
2032 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2033 sc
, LRU_ACTIVE_ANON
);
2035 throttle_vm_writeout(sc
->gfp_mask
);
2038 /* Use reclaim/compaction for costly allocs or under memory pressure */
2039 static bool in_reclaim_compaction(struct scan_control
*sc
)
2041 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2042 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2043 sc
->priority
< DEF_PRIORITY
- 2))
2050 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2051 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2052 * true if more pages should be reclaimed such that when the page allocator
2053 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2054 * It will give up earlier than that if there is difficulty reclaiming pages.
2056 static inline bool should_continue_reclaim(struct zone
*zone
,
2057 unsigned long nr_reclaimed
,
2058 unsigned long nr_scanned
,
2059 struct scan_control
*sc
)
2061 unsigned long pages_for_compaction
;
2062 unsigned long inactive_lru_pages
;
2064 /* If not in reclaim/compaction mode, stop */
2065 if (!in_reclaim_compaction(sc
))
2068 /* Consider stopping depending on scan and reclaim activity */
2069 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2071 * For __GFP_REPEAT allocations, stop reclaiming if the
2072 * full LRU list has been scanned and we are still failing
2073 * to reclaim pages. This full LRU scan is potentially
2074 * expensive but a __GFP_REPEAT caller really wants to succeed
2076 if (!nr_reclaimed
&& !nr_scanned
)
2080 * For non-__GFP_REPEAT allocations which can presumably
2081 * fail without consequence, stop if we failed to reclaim
2082 * any pages from the last SWAP_CLUSTER_MAX number of
2083 * pages that were scanned. This will return to the
2084 * caller faster at the risk reclaim/compaction and
2085 * the resulting allocation attempt fails
2092 * If we have not reclaimed enough pages for compaction and the
2093 * inactive lists are large enough, continue reclaiming
2095 pages_for_compaction
= (2UL << sc
->order
);
2096 inactive_lru_pages
= zone_page_state(zone
, NR_INACTIVE_FILE
);
2097 if (get_nr_swap_pages() > 0)
2098 inactive_lru_pages
+= zone_page_state(zone
, NR_INACTIVE_ANON
);
2099 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2100 inactive_lru_pages
> pages_for_compaction
)
2103 /* If compaction would go ahead or the allocation would succeed, stop */
2104 switch (compaction_suitable(zone
, sc
->order
)) {
2105 case COMPACT_PARTIAL
:
2106 case COMPACT_CONTINUE
:
2113 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
2115 unsigned long nr_reclaimed
, nr_scanned
;
2118 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2119 struct mem_cgroup_reclaim_cookie reclaim
= {
2121 .priority
= sc
->priority
,
2123 struct mem_cgroup
*memcg
;
2125 nr_reclaimed
= sc
->nr_reclaimed
;
2126 nr_scanned
= sc
->nr_scanned
;
2128 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2130 struct lruvec
*lruvec
;
2132 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2134 shrink_lruvec(lruvec
, sc
);
2137 * Direct reclaim and kswapd have to scan all memory
2138 * cgroups to fulfill the overall scan target for the
2141 * Limit reclaim, on the other hand, only cares about
2142 * nr_to_reclaim pages to be reclaimed and it will
2143 * retry with decreasing priority if one round over the
2144 * whole hierarchy is not sufficient.
2146 if (!global_reclaim(sc
) &&
2147 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2148 mem_cgroup_iter_break(root
, memcg
);
2151 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
2154 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2155 sc
->nr_scanned
- nr_scanned
,
2156 sc
->nr_reclaimed
- nr_reclaimed
);
2158 } while (should_continue_reclaim(zone
, sc
->nr_reclaimed
- nr_reclaimed
,
2159 sc
->nr_scanned
- nr_scanned
, sc
));
2162 /* Returns true if compaction should go ahead for a high-order request */
2163 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2165 unsigned long balance_gap
, watermark
;
2168 /* Do not consider compaction for orders reclaim is meant to satisfy */
2169 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
2173 * Compaction takes time to run and there are potentially other
2174 * callers using the pages just freed. Continue reclaiming until
2175 * there is a buffer of free pages available to give compaction
2176 * a reasonable chance of completing and allocating the page
2178 balance_gap
= min(low_wmark_pages(zone
),
2179 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2180 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2181 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
2182 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2185 * If compaction is deferred, reclaim up to a point where
2186 * compaction will have a chance of success when re-enabled
2188 if (compaction_deferred(zone
, sc
->order
))
2189 return watermark_ok
;
2191 /* If compaction is not ready to start, keep reclaiming */
2192 if (!compaction_suitable(zone
, sc
->order
))
2195 return watermark_ok
;
2199 * This is the direct reclaim path, for page-allocating processes. We only
2200 * try to reclaim pages from zones which will satisfy the caller's allocation
2203 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2205 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2207 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2208 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2209 * zone defense algorithm.
2211 * If a zone is deemed to be full of pinned pages then just give it a light
2212 * scan then give up on it.
2214 * This function returns true if a zone is being reclaimed for a costly
2215 * high-order allocation and compaction is ready to begin. This indicates to
2216 * the caller that it should consider retrying the allocation instead of
2219 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2223 unsigned long nr_soft_reclaimed
;
2224 unsigned long nr_soft_scanned
;
2225 bool aborted_reclaim
= false;
2228 * If the number of buffer_heads in the machine exceeds the maximum
2229 * allowed level, force direct reclaim to scan the highmem zone as
2230 * highmem pages could be pinning lowmem pages storing buffer_heads
2232 if (buffer_heads_over_limit
)
2233 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2235 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2236 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2237 if (!populated_zone(zone
))
2240 * Take care memory controller reclaiming has small influence
2243 if (global_reclaim(sc
)) {
2244 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2246 if (zone
->all_unreclaimable
&&
2247 sc
->priority
!= DEF_PRIORITY
)
2248 continue; /* Let kswapd poll it */
2249 if (IS_ENABLED(CONFIG_COMPACTION
)) {
2251 * If we already have plenty of memory free for
2252 * compaction in this zone, don't free any more.
2253 * Even though compaction is invoked for any
2254 * non-zero order, only frequent costly order
2255 * reclamation is disruptive enough to become a
2256 * noticeable problem, like transparent huge
2259 if (compaction_ready(zone
, sc
)) {
2260 aborted_reclaim
= true;
2265 * This steals pages from memory cgroups over softlimit
2266 * and returns the number of reclaimed pages and
2267 * scanned pages. This works for global memory pressure
2268 * and balancing, not for a memcg's limit.
2270 nr_soft_scanned
= 0;
2271 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2272 sc
->order
, sc
->gfp_mask
,
2274 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2275 sc
->nr_scanned
+= nr_soft_scanned
;
2276 /* need some check for avoid more shrink_zone() */
2279 shrink_zone(zone
, sc
);
2282 return aborted_reclaim
;
2285 static bool zone_reclaimable(struct zone
*zone
)
2287 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2290 /* All zones in zonelist are unreclaimable? */
2291 static bool all_unreclaimable(struct zonelist
*zonelist
,
2292 struct scan_control
*sc
)
2297 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2298 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2299 if (!populated_zone(zone
))
2301 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2303 if (!zone
->all_unreclaimable
)
2311 * This is the main entry point to direct page reclaim.
2313 * If a full scan of the inactive list fails to free enough memory then we
2314 * are "out of memory" and something needs to be killed.
2316 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2317 * high - the zone may be full of dirty or under-writeback pages, which this
2318 * caller can't do much about. We kick the writeback threads and take explicit
2319 * naps in the hope that some of these pages can be written. But if the
2320 * allocating task holds filesystem locks which prevent writeout this might not
2321 * work, and the allocation attempt will fail.
2323 * returns: 0, if no pages reclaimed
2324 * else, the number of pages reclaimed
2326 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2327 struct scan_control
*sc
,
2328 struct shrink_control
*shrink
)
2330 unsigned long total_scanned
= 0;
2331 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2334 unsigned long writeback_threshold
;
2335 bool aborted_reclaim
;
2337 delayacct_freepages_start();
2339 if (global_reclaim(sc
))
2340 count_vm_event(ALLOCSTALL
);
2343 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2346 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2349 * Don't shrink slabs when reclaiming memory from
2350 * over limit cgroups
2352 if (global_reclaim(sc
)) {
2353 unsigned long lru_pages
= 0;
2354 for_each_zone_zonelist(zone
, z
, zonelist
,
2355 gfp_zone(sc
->gfp_mask
)) {
2356 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2359 lru_pages
+= zone_reclaimable_pages(zone
);
2362 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2363 if (reclaim_state
) {
2364 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2365 reclaim_state
->reclaimed_slab
= 0;
2368 total_scanned
+= sc
->nr_scanned
;
2369 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2373 * If we're getting trouble reclaiming, start doing
2374 * writepage even in laptop mode.
2376 if (sc
->priority
< DEF_PRIORITY
- 2)
2377 sc
->may_writepage
= 1;
2380 * Try to write back as many pages as we just scanned. This
2381 * tends to cause slow streaming writers to write data to the
2382 * disk smoothly, at the dirtying rate, which is nice. But
2383 * that's undesirable in laptop mode, where we *want* lumpy
2384 * writeout. So in laptop mode, write out the whole world.
2386 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2387 if (total_scanned
> writeback_threshold
) {
2388 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2389 WB_REASON_TRY_TO_FREE_PAGES
);
2390 sc
->may_writepage
= 1;
2392 } while (--sc
->priority
>= 0);
2395 delayacct_freepages_end();
2397 if (sc
->nr_reclaimed
)
2398 return sc
->nr_reclaimed
;
2401 * As hibernation is going on, kswapd is freezed so that it can't mark
2402 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2405 if (oom_killer_disabled
)
2408 /* Aborted reclaim to try compaction? don't OOM, then */
2409 if (aborted_reclaim
)
2412 /* top priority shrink_zones still had more to do? don't OOM, then */
2413 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2419 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2422 unsigned long pfmemalloc_reserve
= 0;
2423 unsigned long free_pages
= 0;
2427 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2428 zone
= &pgdat
->node_zones
[i
];
2429 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2430 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2433 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2435 /* kswapd must be awake if processes are being throttled */
2436 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2437 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
,
2438 (enum zone_type
)ZONE_NORMAL
);
2439 wake_up_interruptible(&pgdat
->kswapd_wait
);
2446 * Throttle direct reclaimers if backing storage is backed by the network
2447 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2448 * depleted. kswapd will continue to make progress and wake the processes
2449 * when the low watermark is reached.
2451 * Returns true if a fatal signal was delivered during throttling. If this
2452 * happens, the page allocator should not consider triggering the OOM killer.
2454 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2455 nodemask_t
*nodemask
)
2458 int high_zoneidx
= gfp_zone(gfp_mask
);
2462 * Kernel threads should not be throttled as they may be indirectly
2463 * responsible for cleaning pages necessary for reclaim to make forward
2464 * progress. kjournald for example may enter direct reclaim while
2465 * committing a transaction where throttling it could forcing other
2466 * processes to block on log_wait_commit().
2468 if (current
->flags
& PF_KTHREAD
)
2472 * If a fatal signal is pending, this process should not throttle.
2473 * It should return quickly so it can exit and free its memory
2475 if (fatal_signal_pending(current
))
2478 /* Check if the pfmemalloc reserves are ok */
2479 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
, &zone
);
2480 pgdat
= zone
->zone_pgdat
;
2481 if (pfmemalloc_watermark_ok(pgdat
))
2484 /* Account for the throttling */
2485 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2488 * If the caller cannot enter the filesystem, it's possible that it
2489 * is due to the caller holding an FS lock or performing a journal
2490 * transaction in the case of a filesystem like ext[3|4]. In this case,
2491 * it is not safe to block on pfmemalloc_wait as kswapd could be
2492 * blocked waiting on the same lock. Instead, throttle for up to a
2493 * second before continuing.
2495 if (!(gfp_mask
& __GFP_FS
)) {
2496 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2497 pfmemalloc_watermark_ok(pgdat
), HZ
);
2502 /* Throttle until kswapd wakes the process */
2503 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2504 pfmemalloc_watermark_ok(pgdat
));
2507 if (fatal_signal_pending(current
))
2514 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2515 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2517 unsigned long nr_reclaimed
;
2518 struct scan_control sc
= {
2519 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2520 .may_writepage
= !laptop_mode
,
2521 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2525 .priority
= DEF_PRIORITY
,
2526 .target_mem_cgroup
= NULL
,
2527 .nodemask
= nodemask
,
2529 struct shrink_control shrink
= {
2530 .gfp_mask
= sc
.gfp_mask
,
2534 * Do not enter reclaim if fatal signal was delivered while throttled.
2535 * 1 is returned so that the page allocator does not OOM kill at this
2538 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2541 trace_mm_vmscan_direct_reclaim_begin(order
,
2545 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2547 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2549 return nr_reclaimed
;
2554 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2555 gfp_t gfp_mask
, bool noswap
,
2557 unsigned long *nr_scanned
)
2559 struct scan_control sc
= {
2561 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2562 .may_writepage
= !laptop_mode
,
2564 .may_swap
= !noswap
,
2567 .target_mem_cgroup
= memcg
,
2569 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2571 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2572 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2574 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2579 * NOTE: Although we can get the priority field, using it
2580 * here is not a good idea, since it limits the pages we can scan.
2581 * if we don't reclaim here, the shrink_zone from balance_pgdat
2582 * will pick up pages from other mem cgroup's as well. We hack
2583 * the priority and make it zero.
2585 shrink_lruvec(lruvec
, &sc
);
2587 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2589 *nr_scanned
= sc
.nr_scanned
;
2590 return sc
.nr_reclaimed
;
2593 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2597 struct zonelist
*zonelist
;
2598 unsigned long nr_reclaimed
;
2600 struct scan_control sc
= {
2601 .may_writepage
= !laptop_mode
,
2603 .may_swap
= !noswap
,
2604 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2606 .priority
= DEF_PRIORITY
,
2607 .target_mem_cgroup
= memcg
,
2608 .nodemask
= NULL
, /* we don't care the placement */
2609 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2610 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2612 struct shrink_control shrink
= {
2613 .gfp_mask
= sc
.gfp_mask
,
2617 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2618 * take care of from where we get pages. So the node where we start the
2619 * scan does not need to be the current node.
2621 nid
= mem_cgroup_select_victim_node(memcg
);
2623 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2625 trace_mm_vmscan_memcg_reclaim_begin(0,
2629 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2631 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2633 return nr_reclaimed
;
2637 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2639 struct mem_cgroup
*memcg
;
2641 if (!total_swap_pages
)
2644 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2646 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2648 if (inactive_anon_is_low(lruvec
))
2649 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2650 sc
, LRU_ACTIVE_ANON
);
2652 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2656 static bool zone_balanced(struct zone
*zone
, int order
,
2657 unsigned long balance_gap
, int classzone_idx
)
2659 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
) +
2660 balance_gap
, classzone_idx
, 0))
2663 if (IS_ENABLED(CONFIG_COMPACTION
) && order
&&
2664 !compaction_suitable(zone
, order
))
2671 * pgdat_balanced() is used when checking if a node is balanced.
2673 * For order-0, all zones must be balanced!
2675 * For high-order allocations only zones that meet watermarks and are in a
2676 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2677 * total of balanced pages must be at least 25% of the zones allowed by
2678 * classzone_idx for the node to be considered balanced. Forcing all zones to
2679 * be balanced for high orders can cause excessive reclaim when there are
2681 * The choice of 25% is due to
2682 * o a 16M DMA zone that is balanced will not balance a zone on any
2683 * reasonable sized machine
2684 * o On all other machines, the top zone must be at least a reasonable
2685 * percentage of the middle zones. For example, on 32-bit x86, highmem
2686 * would need to be at least 256M for it to be balance a whole node.
2687 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2688 * to balance a node on its own. These seemed like reasonable ratios.
2690 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2692 unsigned long managed_pages
= 0;
2693 unsigned long balanced_pages
= 0;
2696 /* Check the watermark levels */
2697 for (i
= 0; i
<= classzone_idx
; i
++) {
2698 struct zone
*zone
= pgdat
->node_zones
+ i
;
2700 if (!populated_zone(zone
))
2703 managed_pages
+= zone
->managed_pages
;
2706 * A special case here:
2708 * balance_pgdat() skips over all_unreclaimable after
2709 * DEF_PRIORITY. Effectively, it considers them balanced so
2710 * they must be considered balanced here as well!
2712 if (zone
->all_unreclaimable
) {
2713 balanced_pages
+= zone
->managed_pages
;
2717 if (zone_balanced(zone
, order
, 0, i
))
2718 balanced_pages
+= zone
->managed_pages
;
2724 return balanced_pages
>= (managed_pages
>> 2);
2730 * Prepare kswapd for sleeping. This verifies that there are no processes
2731 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2733 * Returns true if kswapd is ready to sleep
2735 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, long remaining
,
2738 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2743 * There is a potential race between when kswapd checks its watermarks
2744 * and a process gets throttled. There is also a potential race if
2745 * processes get throttled, kswapd wakes, a large process exits therby
2746 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2747 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2748 * so wake them now if necessary. If necessary, processes will wake
2749 * kswapd and get throttled again
2751 if (waitqueue_active(&pgdat
->pfmemalloc_wait
)) {
2752 wake_up(&pgdat
->pfmemalloc_wait
);
2756 return pgdat_balanced(pgdat
, order
, classzone_idx
);
2760 * kswapd shrinks the zone by the number of pages required to reach
2761 * the high watermark.
2763 * Returns true if kswapd scanned at least the requested number of pages to
2764 * reclaim or if the lack of progress was due to pages under writeback.
2765 * This is used to determine if the scanning priority needs to be raised.
2767 static bool kswapd_shrink_zone(struct zone
*zone
,
2769 struct scan_control
*sc
,
2770 unsigned long lru_pages
,
2771 unsigned long *nr_attempted
)
2773 unsigned long nr_slab
;
2774 int testorder
= sc
->order
;
2775 unsigned long balance_gap
;
2776 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2777 struct shrink_control shrink
= {
2778 .gfp_mask
= sc
->gfp_mask
,
2780 bool lowmem_pressure
;
2782 /* Reclaim above the high watermark. */
2783 sc
->nr_to_reclaim
= max(SWAP_CLUSTER_MAX
, high_wmark_pages(zone
));
2786 * Kswapd reclaims only single pages with compaction enabled. Trying
2787 * too hard to reclaim until contiguous free pages have become
2788 * available can hurt performance by evicting too much useful data
2789 * from memory. Do not reclaim more than needed for compaction.
2791 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2792 compaction_suitable(zone
, sc
->order
) !=
2797 * We put equal pressure on every zone, unless one zone has way too
2798 * many pages free already. The "too many pages" is defined as the
2799 * high wmark plus a "gap" where the gap is either the low
2800 * watermark or 1% of the zone, whichever is smaller.
2802 balance_gap
= min(low_wmark_pages(zone
),
2803 (zone
->managed_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2804 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2807 * If there is no low memory pressure or the zone is balanced then no
2808 * reclaim is necessary
2810 lowmem_pressure
= (buffer_heads_over_limit
&& is_highmem(zone
));
2811 if (!lowmem_pressure
&& zone_balanced(zone
, testorder
,
2812 balance_gap
, classzone_idx
))
2815 shrink_zone(zone
, sc
);
2817 reclaim_state
->reclaimed_slab
= 0;
2818 nr_slab
= shrink_slab(&shrink
, sc
->nr_scanned
, lru_pages
);
2819 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2821 /* Account for the number of pages attempted to reclaim */
2822 *nr_attempted
+= sc
->nr_to_reclaim
;
2824 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2825 zone
->all_unreclaimable
= 1;
2827 zone_clear_flag(zone
, ZONE_WRITEBACK
);
2830 * If a zone reaches its high watermark, consider it to be no longer
2831 * congested. It's possible there are dirty pages backed by congested
2832 * BDIs but as pressure is relieved, speculatively avoid congestion
2835 if (!zone
->all_unreclaimable
&&
2836 zone_balanced(zone
, testorder
, 0, classzone_idx
)) {
2837 zone_clear_flag(zone
, ZONE_CONGESTED
);
2838 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2841 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
2845 * For kswapd, balance_pgdat() will work across all this node's zones until
2846 * they are all at high_wmark_pages(zone).
2848 * Returns the final order kswapd was reclaiming at
2850 * There is special handling here for zones which are full of pinned pages.
2851 * This can happen if the pages are all mlocked, or if they are all used by
2852 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2853 * What we do is to detect the case where all pages in the zone have been
2854 * scanned twice and there has been zero successful reclaim. Mark the zone as
2855 * dead and from now on, only perform a short scan. Basically we're polling
2856 * the zone for when the problem goes away.
2858 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2859 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2860 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2861 * lower zones regardless of the number of free pages in the lower zones. This
2862 * interoperates with the page allocator fallback scheme to ensure that aging
2863 * of pages is balanced across the zones.
2865 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2869 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2870 unsigned long nr_soft_reclaimed
;
2871 unsigned long nr_soft_scanned
;
2872 struct scan_control sc
= {
2873 .gfp_mask
= GFP_KERNEL
,
2874 .priority
= DEF_PRIORITY
,
2877 .may_writepage
= !laptop_mode
,
2879 .target_mem_cgroup
= NULL
,
2881 count_vm_event(PAGEOUTRUN
);
2884 unsigned long lru_pages
= 0;
2885 unsigned long nr_attempted
= 0;
2886 bool raise_priority
= true;
2887 bool pgdat_needs_compaction
= (order
> 0);
2889 sc
.nr_reclaimed
= 0;
2892 * Scan in the highmem->dma direction for the highest
2893 * zone which needs scanning
2895 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2896 struct zone
*zone
= pgdat
->node_zones
+ i
;
2898 if (!populated_zone(zone
))
2901 if (zone
->all_unreclaimable
&&
2902 sc
.priority
!= DEF_PRIORITY
)
2906 * Do some background aging of the anon list, to give
2907 * pages a chance to be referenced before reclaiming.
2909 age_active_anon(zone
, &sc
);
2912 * If the number of buffer_heads in the machine
2913 * exceeds the maximum allowed level and this node
2914 * has a highmem zone, force kswapd to reclaim from
2915 * it to relieve lowmem pressure.
2917 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2922 if (!zone_balanced(zone
, order
, 0, 0)) {
2927 * If balanced, clear the dirty and congested
2930 zone_clear_flag(zone
, ZONE_CONGESTED
);
2931 zone_clear_flag(zone
, ZONE_TAIL_LRU_DIRTY
);
2938 for (i
= 0; i
<= end_zone
; i
++) {
2939 struct zone
*zone
= pgdat
->node_zones
+ i
;
2941 if (!populated_zone(zone
))
2944 lru_pages
+= zone_reclaimable_pages(zone
);
2947 * If any zone is currently balanced then kswapd will
2948 * not call compaction as it is expected that the
2949 * necessary pages are already available.
2951 if (pgdat_needs_compaction
&&
2952 zone_watermark_ok(zone
, order
,
2953 low_wmark_pages(zone
),
2955 pgdat_needs_compaction
= false;
2959 * If we're getting trouble reclaiming, start doing writepage
2960 * even in laptop mode.
2962 if (sc
.priority
< DEF_PRIORITY
- 2)
2963 sc
.may_writepage
= 1;
2966 * Now scan the zone in the dma->highmem direction, stopping
2967 * at the last zone which needs scanning.
2969 * We do this because the page allocator works in the opposite
2970 * direction. This prevents the page allocator from allocating
2971 * pages behind kswapd's direction of progress, which would
2972 * cause too much scanning of the lower zones.
2974 for (i
= 0; i
<= end_zone
; i
++) {
2975 struct zone
*zone
= pgdat
->node_zones
+ i
;
2977 if (!populated_zone(zone
))
2980 if (zone
->all_unreclaimable
&&
2981 sc
.priority
!= DEF_PRIORITY
)
2986 nr_soft_scanned
= 0;
2988 * Call soft limit reclaim before calling shrink_zone.
2990 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2993 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2996 * There should be no need to raise the scanning
2997 * priority if enough pages are already being scanned
2998 * that that high watermark would be met at 100%
3001 if (kswapd_shrink_zone(zone
, end_zone
, &sc
,
3002 lru_pages
, &nr_attempted
))
3003 raise_priority
= false;
3007 * If the low watermark is met there is no need for processes
3008 * to be throttled on pfmemalloc_wait as they should not be
3009 * able to safely make forward progress. Wake them
3011 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3012 pfmemalloc_watermark_ok(pgdat
))
3013 wake_up(&pgdat
->pfmemalloc_wait
);
3016 * Fragmentation may mean that the system cannot be rebalanced
3017 * for high-order allocations in all zones. If twice the
3018 * allocation size has been reclaimed and the zones are still
3019 * not balanced then recheck the watermarks at order-0 to
3020 * prevent kswapd reclaiming excessively. Assume that a
3021 * process requested a high-order can direct reclaim/compact.
3023 if (order
&& sc
.nr_reclaimed
>= 2UL << order
)
3024 order
= sc
.order
= 0;
3026 /* Check if kswapd should be suspending */
3027 if (try_to_freeze() || kthread_should_stop())
3031 * Compact if necessary and kswapd is reclaiming at least the
3032 * high watermark number of pages as requsted
3034 if (pgdat_needs_compaction
&& sc
.nr_reclaimed
> nr_attempted
)
3035 compact_pgdat(pgdat
, order
);
3038 * Raise priority if scanning rate is too low or there was no
3039 * progress in reclaiming pages
3041 if (raise_priority
|| !sc
.nr_reclaimed
)
3043 } while (sc
.priority
>= 1 &&
3044 !pgdat_balanced(pgdat
, order
, *classzone_idx
));
3048 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3049 * makes a decision on the order we were last reclaiming at. However,
3050 * if another caller entered the allocator slow path while kswapd
3051 * was awake, order will remain at the higher level
3053 *classzone_idx
= end_zone
;
3057 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3062 if (freezing(current
) || kthread_should_stop())
3065 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3067 /* Try to sleep for a short interval */
3068 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3069 remaining
= schedule_timeout(HZ
/10);
3070 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3071 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3075 * After a short sleep, check if it was a premature sleep. If not, then
3076 * go fully to sleep until explicitly woken up.
3078 if (prepare_kswapd_sleep(pgdat
, order
, remaining
, classzone_idx
)) {
3079 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3082 * vmstat counters are not perfectly accurate and the estimated
3083 * value for counters such as NR_FREE_PAGES can deviate from the
3084 * true value by nr_online_cpus * threshold. To avoid the zone
3085 * watermarks being breached while under pressure, we reduce the
3086 * per-cpu vmstat threshold while kswapd is awake and restore
3087 * them before going back to sleep.
3089 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3092 * Compaction records what page blocks it recently failed to
3093 * isolate pages from and skips them in the future scanning.
3094 * When kswapd is going to sleep, it is reasonable to assume
3095 * that pages and compaction may succeed so reset the cache.
3097 reset_isolation_suitable(pgdat
);
3099 if (!kthread_should_stop())
3102 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3105 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3107 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3109 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3113 * The background pageout daemon, started as a kernel thread
3114 * from the init process.
3116 * This basically trickles out pages so that we have _some_
3117 * free memory available even if there is no other activity
3118 * that frees anything up. This is needed for things like routing
3119 * etc, where we otherwise might have all activity going on in
3120 * asynchronous contexts that cannot page things out.
3122 * If there are applications that are active memory-allocators
3123 * (most normal use), this basically shouldn't matter.
3125 static int kswapd(void *p
)
3127 unsigned long order
, new_order
;
3128 unsigned balanced_order
;
3129 int classzone_idx
, new_classzone_idx
;
3130 int balanced_classzone_idx
;
3131 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3132 struct task_struct
*tsk
= current
;
3134 struct reclaim_state reclaim_state
= {
3135 .reclaimed_slab
= 0,
3137 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3139 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3141 if (!cpumask_empty(cpumask
))
3142 set_cpus_allowed_ptr(tsk
, cpumask
);
3143 current
->reclaim_state
= &reclaim_state
;
3146 * Tell the memory management that we're a "memory allocator",
3147 * and that if we need more memory we should get access to it
3148 * regardless (see "__alloc_pages()"). "kswapd" should
3149 * never get caught in the normal page freeing logic.
3151 * (Kswapd normally doesn't need memory anyway, but sometimes
3152 * you need a small amount of memory in order to be able to
3153 * page out something else, and this flag essentially protects
3154 * us from recursively trying to free more memory as we're
3155 * trying to free the first piece of memory in the first place).
3157 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3160 order
= new_order
= 0;
3162 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
3163 balanced_classzone_idx
= classzone_idx
;
3168 * If the last balance_pgdat was unsuccessful it's unlikely a
3169 * new request of a similar or harder type will succeed soon
3170 * so consider going to sleep on the basis we reclaimed at
3172 if (balanced_classzone_idx
>= new_classzone_idx
&&
3173 balanced_order
== new_order
) {
3174 new_order
= pgdat
->kswapd_max_order
;
3175 new_classzone_idx
= pgdat
->classzone_idx
;
3176 pgdat
->kswapd_max_order
= 0;
3177 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3180 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
3182 * Don't sleep if someone wants a larger 'order'
3183 * allocation or has tigher zone constraints
3186 classzone_idx
= new_classzone_idx
;
3188 kswapd_try_to_sleep(pgdat
, balanced_order
,
3189 balanced_classzone_idx
);
3190 order
= pgdat
->kswapd_max_order
;
3191 classzone_idx
= pgdat
->classzone_idx
;
3193 new_classzone_idx
= classzone_idx
;
3194 pgdat
->kswapd_max_order
= 0;
3195 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
3198 ret
= try_to_freeze();
3199 if (kthread_should_stop())
3203 * We can speed up thawing tasks if we don't call balance_pgdat
3204 * after returning from the refrigerator
3207 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
3208 balanced_classzone_idx
= classzone_idx
;
3209 balanced_order
= balance_pgdat(pgdat
, order
,
3210 &balanced_classzone_idx
);
3214 current
->reclaim_state
= NULL
;
3219 * A zone is low on free memory, so wake its kswapd task to service it.
3221 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3225 if (!populated_zone(zone
))
3228 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
3230 pgdat
= zone
->zone_pgdat
;
3231 if (pgdat
->kswapd_max_order
< order
) {
3232 pgdat
->kswapd_max_order
= order
;
3233 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
3235 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3237 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
3240 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3241 wake_up_interruptible(&pgdat
->kswapd_wait
);
3245 * The reclaimable count would be mostly accurate.
3246 * The less reclaimable pages may be
3247 * - mlocked pages, which will be moved to unevictable list when encountered
3248 * - mapped pages, which may require several travels to be reclaimed
3249 * - dirty pages, which is not "instantly" reclaimable
3251 unsigned long global_reclaimable_pages(void)
3255 nr
= global_page_state(NR_ACTIVE_FILE
) +
3256 global_page_state(NR_INACTIVE_FILE
);
3258 if (get_nr_swap_pages() > 0)
3259 nr
+= global_page_state(NR_ACTIVE_ANON
) +
3260 global_page_state(NR_INACTIVE_ANON
);
3265 unsigned long zone_reclaimable_pages(struct zone
*zone
)
3269 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
3270 zone_page_state(zone
, NR_INACTIVE_FILE
);
3272 if (get_nr_swap_pages() > 0)
3273 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
3274 zone_page_state(zone
, NR_INACTIVE_ANON
);
3279 #ifdef CONFIG_HIBERNATION
3281 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3284 * Rather than trying to age LRUs the aim is to preserve the overall
3285 * LRU order by reclaiming preferentially
3286 * inactive > active > active referenced > active mapped
3288 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3290 struct reclaim_state reclaim_state
;
3291 struct scan_control sc
= {
3292 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3296 .nr_to_reclaim
= nr_to_reclaim
,
3297 .hibernation_mode
= 1,
3299 .priority
= DEF_PRIORITY
,
3301 struct shrink_control shrink
= {
3302 .gfp_mask
= sc
.gfp_mask
,
3304 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3305 struct task_struct
*p
= current
;
3306 unsigned long nr_reclaimed
;
3308 p
->flags
|= PF_MEMALLOC
;
3309 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3310 reclaim_state
.reclaimed_slab
= 0;
3311 p
->reclaim_state
= &reclaim_state
;
3313 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3315 p
->reclaim_state
= NULL
;
3316 lockdep_clear_current_reclaim_state();
3317 p
->flags
&= ~PF_MEMALLOC
;
3319 return nr_reclaimed
;
3321 #endif /* CONFIG_HIBERNATION */
3323 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3324 not required for correctness. So if the last cpu in a node goes
3325 away, we get changed to run anywhere: as the first one comes back,
3326 restore their cpu bindings. */
3327 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3332 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3333 for_each_node_state(nid
, N_MEMORY
) {
3334 pg_data_t
*pgdat
= NODE_DATA(nid
);
3335 const struct cpumask
*mask
;
3337 mask
= cpumask_of_node(pgdat
->node_id
);
3339 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3340 /* One of our CPUs online: restore mask */
3341 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3348 * This kswapd start function will be called by init and node-hot-add.
3349 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3351 int kswapd_run(int nid
)
3353 pg_data_t
*pgdat
= NODE_DATA(nid
);
3359 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3360 if (IS_ERR(pgdat
->kswapd
)) {
3361 /* failure at boot is fatal */
3362 BUG_ON(system_state
== SYSTEM_BOOTING
);
3363 pr_err("Failed to start kswapd on node %d\n", nid
);
3364 ret
= PTR_ERR(pgdat
->kswapd
);
3365 pgdat
->kswapd
= NULL
;
3371 * Called by memory hotplug when all memory in a node is offlined. Caller must
3372 * hold lock_memory_hotplug().
3374 void kswapd_stop(int nid
)
3376 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3379 kthread_stop(kswapd
);
3380 NODE_DATA(nid
)->kswapd
= NULL
;
3384 static int __init
kswapd_init(void)
3389 for_each_node_state(nid
, N_MEMORY
)
3391 hotcpu_notifier(cpu_callback
, 0);
3395 module_init(kswapd_init
)
3401 * If non-zero call zone_reclaim when the number of free pages falls below
3404 int zone_reclaim_mode __read_mostly
;
3406 #define RECLAIM_OFF 0
3407 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3408 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3409 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3412 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3413 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3416 #define ZONE_RECLAIM_PRIORITY 4
3419 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3422 int sysctl_min_unmapped_ratio
= 1;
3425 * If the number of slab pages in a zone grows beyond this percentage then
3426 * slab reclaim needs to occur.
3428 int sysctl_min_slab_ratio
= 5;
3430 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3432 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3433 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3434 zone_page_state(zone
, NR_ACTIVE_FILE
);
3437 * It's possible for there to be more file mapped pages than
3438 * accounted for by the pages on the file LRU lists because
3439 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3441 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3444 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3445 static long zone_pagecache_reclaimable(struct zone
*zone
)
3447 long nr_pagecache_reclaimable
;
3451 * If RECLAIM_SWAP is set, then all file pages are considered
3452 * potentially reclaimable. Otherwise, we have to worry about
3453 * pages like swapcache and zone_unmapped_file_pages() provides
3456 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3457 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3459 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3461 /* If we can't clean pages, remove dirty pages from consideration */
3462 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3463 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3465 /* Watch for any possible underflows due to delta */
3466 if (unlikely(delta
> nr_pagecache_reclaimable
))
3467 delta
= nr_pagecache_reclaimable
;
3469 return nr_pagecache_reclaimable
- delta
;
3473 * Try to free up some pages from this zone through reclaim.
3475 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3477 /* Minimum pages needed in order to stay on node */
3478 const unsigned long nr_pages
= 1 << order
;
3479 struct task_struct
*p
= current
;
3480 struct reclaim_state reclaim_state
;
3481 struct scan_control sc
= {
3482 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3483 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3485 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3486 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3488 .priority
= ZONE_RECLAIM_PRIORITY
,
3490 struct shrink_control shrink
= {
3491 .gfp_mask
= sc
.gfp_mask
,
3493 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3497 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3498 * and we also need to be able to write out pages for RECLAIM_WRITE
3501 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3502 lockdep_set_current_reclaim_state(gfp_mask
);
3503 reclaim_state
.reclaimed_slab
= 0;
3504 p
->reclaim_state
= &reclaim_state
;
3506 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3508 * Free memory by calling shrink zone with increasing
3509 * priorities until we have enough memory freed.
3512 shrink_zone(zone
, &sc
);
3513 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3516 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3517 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3519 * shrink_slab() does not currently allow us to determine how
3520 * many pages were freed in this zone. So we take the current
3521 * number of slab pages and shake the slab until it is reduced
3522 * by the same nr_pages that we used for reclaiming unmapped
3525 * Note that shrink_slab will free memory on all zones and may
3529 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3531 /* No reclaimable slab or very low memory pressure */
3532 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3535 /* Freed enough memory */
3536 nr_slab_pages1
= zone_page_state(zone
,
3537 NR_SLAB_RECLAIMABLE
);
3538 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3543 * Update nr_reclaimed by the number of slab pages we
3544 * reclaimed from this zone.
3546 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3547 if (nr_slab_pages1
< nr_slab_pages0
)
3548 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3551 p
->reclaim_state
= NULL
;
3552 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3553 lockdep_clear_current_reclaim_state();
3554 return sc
.nr_reclaimed
>= nr_pages
;
3557 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3563 * Zone reclaim reclaims unmapped file backed pages and
3564 * slab pages if we are over the defined limits.
3566 * A small portion of unmapped file backed pages is needed for
3567 * file I/O otherwise pages read by file I/O will be immediately
3568 * thrown out if the zone is overallocated. So we do not reclaim
3569 * if less than a specified percentage of the zone is used by
3570 * unmapped file backed pages.
3572 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3573 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3574 return ZONE_RECLAIM_FULL
;
3576 if (zone
->all_unreclaimable
)
3577 return ZONE_RECLAIM_FULL
;
3580 * Do not scan if the allocation should not be delayed.
3582 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3583 return ZONE_RECLAIM_NOSCAN
;
3586 * Only run zone reclaim on the local zone or on zones that do not
3587 * have associated processors. This will favor the local processor
3588 * over remote processors and spread off node memory allocations
3589 * as wide as possible.
3591 node_id
= zone_to_nid(zone
);
3592 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3593 return ZONE_RECLAIM_NOSCAN
;
3595 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3596 return ZONE_RECLAIM_NOSCAN
;
3598 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3599 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3602 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3609 * page_evictable - test whether a page is evictable
3610 * @page: the page to test
3612 * Test whether page is evictable--i.e., should be placed on active/inactive
3613 * lists vs unevictable list.
3615 * Reasons page might not be evictable:
3616 * (1) page's mapping marked unevictable
3617 * (2) page is part of an mlocked VMA
3620 int page_evictable(struct page
*page
)
3622 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3627 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3628 * @pages: array of pages to check
3629 * @nr_pages: number of pages to check
3631 * Checks pages for evictability and moves them to the appropriate lru list.
3633 * This function is only used for SysV IPC SHM_UNLOCK.
3635 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3637 struct lruvec
*lruvec
;
3638 struct zone
*zone
= NULL
;
3643 for (i
= 0; i
< nr_pages
; i
++) {
3644 struct page
*page
= pages
[i
];
3645 struct zone
*pagezone
;
3648 pagezone
= page_zone(page
);
3649 if (pagezone
!= zone
) {
3651 spin_unlock_irq(&zone
->lru_lock
);
3653 spin_lock_irq(&zone
->lru_lock
);
3655 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
3657 if (!PageLRU(page
) || !PageUnevictable(page
))
3660 if (page_evictable(page
)) {
3661 enum lru_list lru
= page_lru_base_type(page
);
3663 VM_BUG_ON(PageActive(page
));
3664 ClearPageUnevictable(page
);
3665 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3666 add_page_to_lru_list(page
, lruvec
, lru
);
3672 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3673 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3674 spin_unlock_irq(&zone
->lru_lock
);
3677 #endif /* CONFIG_SHMEM */
3679 static void warn_scan_unevictable_pages(void)
3681 printk_once(KERN_WARNING
3682 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3683 "disabled for lack of a legitimate use case. If you have "
3684 "one, please send an email to linux-mm@kvack.org.\n",
3689 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3690 * all nodes' unevictable lists for evictable pages
3692 unsigned long scan_unevictable_pages
;
3694 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3695 void __user
*buffer
,
3696 size_t *length
, loff_t
*ppos
)
3698 warn_scan_unevictable_pages();
3699 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3700 scan_unevictable_pages
= 0;
3706 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3707 * a specified node's per zone unevictable lists for evictable pages.
3710 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3711 struct device_attribute
*attr
,
3714 warn_scan_unevictable_pages();
3715 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3718 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3719 struct device_attribute
*attr
,
3720 const char *buf
, size_t count
)
3722 warn_scan_unevictable_pages();
3727 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3728 read_scan_unevictable_node
,
3729 write_scan_unevictable_node
);
3731 int scan_unevictable_register_node(struct node
*node
)
3733 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3736 void scan_unevictable_unregister_node(struct node
*node
)
3738 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);