4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/dax.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup
*target_mem_cgroup
;
84 /* Scan (total_size >> priority) pages at once */
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx
;
90 unsigned int may_writepage
:1;
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap
:1;
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap
:1;
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash
:1;
101 unsigned int hibernation_mode
:1;
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready
:1;
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned
;
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed
;
113 #ifdef ARCH_HAS_PREFETCH
114 #define prefetch_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
124 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
127 #ifdef ARCH_HAS_PREFETCHW
128 #define prefetchw_prev_lru_page(_page, _base, _field) \
130 if ((_page)->lru.prev != _base) { \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
138 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
142 * From 0 .. 100. Higher means more swappy.
144 int vm_swappiness
= 60;
146 * The total number of pages which are beyond the high watermark within all
149 unsigned long vm_total_pages
;
151 static LIST_HEAD(shrinker_list
);
152 static DECLARE_RWSEM(shrinker_rwsem
);
155 static bool global_reclaim(struct scan_control
*sc
)
157 return !sc
->target_mem_cgroup
;
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
173 static bool sane_reclaim(struct scan_control
*sc
)
175 struct mem_cgroup
*memcg
= sc
->target_mem_cgroup
;
179 #ifdef CONFIG_CGROUP_WRITEBACK
180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
186 static bool global_reclaim(struct scan_control
*sc
)
191 static bool sane_reclaim(struct scan_control
*sc
)
197 unsigned long pgdat_reclaimable_pages(struct pglist_data
*pgdat
)
201 nr
= node_page_state_snapshot(pgdat
, NR_ACTIVE_FILE
) +
202 node_page_state_snapshot(pgdat
, NR_INACTIVE_FILE
) +
203 node_page_state_snapshot(pgdat
, NR_ISOLATED_FILE
);
205 if (get_nr_swap_pages() > 0)
206 nr
+= node_page_state_snapshot(pgdat
, NR_ACTIVE_ANON
) +
207 node_page_state_snapshot(pgdat
, NR_INACTIVE_ANON
) +
208 node_page_state_snapshot(pgdat
, NR_ISOLATED_ANON
);
213 bool pgdat_reclaimable(struct pglist_data
*pgdat
)
215 return node_page_state_snapshot(pgdat
, NR_PAGES_SCANNED
) <
216 pgdat_reclaimable_pages(pgdat
) * 6;
219 unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
221 if (!mem_cgroup_disabled())
222 return mem_cgroup_get_lru_size(lruvec
, lru
);
224 return node_page_state(lruvec_pgdat(lruvec
), NR_LRU_BASE
+ lru
);
228 * Add a shrinker callback to be called from the vm.
230 int register_shrinker(struct shrinker
*shrinker
)
232 size_t size
= sizeof(*shrinker
->nr_deferred
);
234 if (shrinker
->flags
& SHRINKER_NUMA_AWARE
)
237 shrinker
->nr_deferred
= kzalloc(size
, GFP_KERNEL
);
238 if (!shrinker
->nr_deferred
)
241 down_write(&shrinker_rwsem
);
242 list_add_tail(&shrinker
->list
, &shrinker_list
);
243 up_write(&shrinker_rwsem
);
246 EXPORT_SYMBOL(register_shrinker
);
251 void unregister_shrinker(struct shrinker
*shrinker
)
253 down_write(&shrinker_rwsem
);
254 list_del(&shrinker
->list
);
255 up_write(&shrinker_rwsem
);
256 kfree(shrinker
->nr_deferred
);
258 EXPORT_SYMBOL(unregister_shrinker
);
260 #define SHRINK_BATCH 128
262 static unsigned long do_shrink_slab(struct shrink_control
*shrinkctl
,
263 struct shrinker
*shrinker
,
264 unsigned long nr_scanned
,
265 unsigned long nr_eligible
)
267 unsigned long freed
= 0;
268 unsigned long long delta
;
273 int nid
= shrinkctl
->nid
;
274 long batch_size
= shrinker
->batch
? shrinker
->batch
277 freeable
= shrinker
->count_objects(shrinker
, shrinkctl
);
282 * copy the current shrinker scan count into a local variable
283 * and zero it so that other concurrent shrinker invocations
284 * don't also do this scanning work.
286 nr
= atomic_long_xchg(&shrinker
->nr_deferred
[nid
], 0);
289 delta
= (4 * nr_scanned
) / shrinker
->seeks
;
291 do_div(delta
, nr_eligible
+ 1);
293 if (total_scan
< 0) {
294 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
295 shrinker
->scan_objects
, total_scan
);
296 total_scan
= freeable
;
300 * We need to avoid excessive windup on filesystem shrinkers
301 * due to large numbers of GFP_NOFS allocations causing the
302 * shrinkers to return -1 all the time. This results in a large
303 * nr being built up so when a shrink that can do some work
304 * comes along it empties the entire cache due to nr >>>
305 * freeable. This is bad for sustaining a working set in
308 * Hence only allow the shrinker to scan the entire cache when
309 * a large delta change is calculated directly.
311 if (delta
< freeable
/ 4)
312 total_scan
= min(total_scan
, freeable
/ 2);
315 * Avoid risking looping forever due to too large nr value:
316 * never try to free more than twice the estimate number of
319 if (total_scan
> freeable
* 2)
320 total_scan
= freeable
* 2;
322 trace_mm_shrink_slab_start(shrinker
, shrinkctl
, nr
,
323 nr_scanned
, nr_eligible
,
324 freeable
, delta
, total_scan
);
327 * Normally, we should not scan less than batch_size objects in one
328 * pass to avoid too frequent shrinker calls, but if the slab has less
329 * than batch_size objects in total and we are really tight on memory,
330 * we will try to reclaim all available objects, otherwise we can end
331 * up failing allocations although there are plenty of reclaimable
332 * objects spread over several slabs with usage less than the
335 * We detect the "tight on memory" situations by looking at the total
336 * number of objects we want to scan (total_scan). If it is greater
337 * than the total number of objects on slab (freeable), we must be
338 * scanning at high prio and therefore should try to reclaim as much as
341 while (total_scan
>= batch_size
||
342 total_scan
>= freeable
) {
344 unsigned long nr_to_scan
= min(batch_size
, total_scan
);
346 shrinkctl
->nr_to_scan
= nr_to_scan
;
347 ret
= shrinker
->scan_objects(shrinker
, shrinkctl
);
348 if (ret
== SHRINK_STOP
)
352 count_vm_events(SLABS_SCANNED
, nr_to_scan
);
353 total_scan
-= nr_to_scan
;
359 * move the unused scan count back into the shrinker in a
360 * manner that handles concurrent updates. If we exhausted the
361 * scan, there is no need to do an update.
364 new_nr
= atomic_long_add_return(total_scan
,
365 &shrinker
->nr_deferred
[nid
]);
367 new_nr
= atomic_long_read(&shrinker
->nr_deferred
[nid
]);
369 trace_mm_shrink_slab_end(shrinker
, nid
, freed
, nr
, new_nr
, total_scan
);
374 * shrink_slab - shrink slab caches
375 * @gfp_mask: allocation context
376 * @nid: node whose slab caches to target
377 * @memcg: memory cgroup whose slab caches to target
378 * @nr_scanned: pressure numerator
379 * @nr_eligible: pressure denominator
381 * Call the shrink functions to age shrinkable caches.
383 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
384 * unaware shrinkers will receive a node id of 0 instead.
386 * @memcg specifies the memory cgroup to target. If it is not NULL,
387 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
388 * objects from the memory cgroup specified. Otherwise, only unaware
389 * shrinkers are called.
391 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
392 * the available objects should be scanned. Page reclaim for example
393 * passes the number of pages scanned and the number of pages on the
394 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
395 * when it encountered mapped pages. The ratio is further biased by
396 * the ->seeks setting of the shrink function, which indicates the
397 * cost to recreate an object relative to that of an LRU page.
399 * Returns the number of reclaimed slab objects.
401 static unsigned long shrink_slab(gfp_t gfp_mask
, int nid
,
402 struct mem_cgroup
*memcg
,
403 unsigned long nr_scanned
,
404 unsigned long nr_eligible
)
406 struct shrinker
*shrinker
;
407 unsigned long freed
= 0;
409 if (memcg
&& (!memcg_kmem_enabled() || !mem_cgroup_online(memcg
)))
413 nr_scanned
= SWAP_CLUSTER_MAX
;
415 if (!down_read_trylock(&shrinker_rwsem
)) {
417 * If we would return 0, our callers would understand that we
418 * have nothing else to shrink and give up trying. By returning
419 * 1 we keep it going and assume we'll be able to shrink next
426 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
427 struct shrink_control sc
= {
428 .gfp_mask
= gfp_mask
,
434 * If kernel memory accounting is disabled, we ignore
435 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
436 * passing NULL for memcg.
438 if (memcg_kmem_enabled() &&
439 !!memcg
!= !!(shrinker
->flags
& SHRINKER_MEMCG_AWARE
))
442 if (!(shrinker
->flags
& SHRINKER_NUMA_AWARE
))
445 freed
+= do_shrink_slab(&sc
, shrinker
, nr_scanned
, nr_eligible
);
448 up_read(&shrinker_rwsem
);
454 void drop_slab_node(int nid
)
459 struct mem_cgroup
*memcg
= NULL
;
463 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
,
465 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
466 } while (freed
> 10);
473 for_each_online_node(nid
)
477 static inline int is_page_cache_freeable(struct page
*page
)
480 * A freeable page cache page is referenced only by the caller
481 * that isolated the page, the page cache radix tree and
482 * optional buffer heads at page->private.
484 return page_count(page
) - page_has_private(page
) == 2;
487 static int may_write_to_inode(struct inode
*inode
, struct scan_control
*sc
)
489 if (current
->flags
& PF_SWAPWRITE
)
491 if (!inode_write_congested(inode
))
493 if (inode_to_bdi(inode
) == current
->backing_dev_info
)
499 * We detected a synchronous write error writing a page out. Probably
500 * -ENOSPC. We need to propagate that into the address_space for a subsequent
501 * fsync(), msync() or close().
503 * The tricky part is that after writepage we cannot touch the mapping: nothing
504 * prevents it from being freed up. But we have a ref on the page and once
505 * that page is locked, the mapping is pinned.
507 * We're allowed to run sleeping lock_page() here because we know the caller has
510 static void handle_write_error(struct address_space
*mapping
,
511 struct page
*page
, int error
)
514 if (page_mapping(page
) == mapping
)
515 mapping_set_error(mapping
, error
);
519 /* possible outcome of pageout() */
521 /* failed to write page out, page is locked */
523 /* move page to the active list, page is locked */
525 /* page has been sent to the disk successfully, page is unlocked */
527 /* page is clean and locked */
532 * pageout is called by shrink_page_list() for each dirty page.
533 * Calls ->writepage().
535 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
536 struct scan_control
*sc
)
539 * If the page is dirty, only perform writeback if that write
540 * will be non-blocking. To prevent this allocation from being
541 * stalled by pagecache activity. But note that there may be
542 * stalls if we need to run get_block(). We could test
543 * PagePrivate for that.
545 * If this process is currently in __generic_file_write_iter() against
546 * this page's queue, we can perform writeback even if that
549 * If the page is swapcache, write it back even if that would
550 * block, for some throttling. This happens by accident, because
551 * swap_backing_dev_info is bust: it doesn't reflect the
552 * congestion state of the swapdevs. Easy to fix, if needed.
554 if (!is_page_cache_freeable(page
))
558 * Some data journaling orphaned pages can have
559 * page->mapping == NULL while being dirty with clean buffers.
561 if (page_has_private(page
)) {
562 if (try_to_free_buffers(page
)) {
563 ClearPageDirty(page
);
564 pr_info("%s: orphaned page\n", __func__
);
570 if (mapping
->a_ops
->writepage
== NULL
)
571 return PAGE_ACTIVATE
;
572 if (!may_write_to_inode(mapping
->host
, sc
))
575 if (clear_page_dirty_for_io(page
)) {
577 struct writeback_control wbc
= {
578 .sync_mode
= WB_SYNC_NONE
,
579 .nr_to_write
= SWAP_CLUSTER_MAX
,
581 .range_end
= LLONG_MAX
,
585 SetPageReclaim(page
);
586 res
= mapping
->a_ops
->writepage(page
, &wbc
);
588 handle_write_error(mapping
, page
, res
);
589 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
590 ClearPageReclaim(page
);
591 return PAGE_ACTIVATE
;
594 if (!PageWriteback(page
)) {
595 /* synchronous write or broken a_ops? */
596 ClearPageReclaim(page
);
598 trace_mm_vmscan_writepage(page
);
599 inc_node_page_state(page
, NR_VMSCAN_WRITE
);
607 * Same as remove_mapping, but if the page is removed from the mapping, it
608 * gets returned with a refcount of 0.
610 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
,
615 BUG_ON(!PageLocked(page
));
616 BUG_ON(mapping
!= page_mapping(page
));
618 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
620 * The non racy check for a busy page.
622 * Must be careful with the order of the tests. When someone has
623 * a ref to the page, it may be possible that they dirty it then
624 * drop the reference. So if PageDirty is tested before page_count
625 * here, then the following race may occur:
627 * get_user_pages(&page);
628 * [user mapping goes away]
630 * !PageDirty(page) [good]
631 * SetPageDirty(page);
633 * !page_count(page) [good, discard it]
635 * [oops, our write_to data is lost]
637 * Reversing the order of the tests ensures such a situation cannot
638 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
639 * load is not satisfied before that of page->_refcount.
641 * Note that if SetPageDirty is always performed via set_page_dirty,
642 * and thus under tree_lock, then this ordering is not required.
644 if (!page_ref_freeze(page
, 2))
646 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
647 if (unlikely(PageDirty(page
))) {
648 page_ref_unfreeze(page
, 2);
652 if (PageSwapCache(page
)) {
653 swp_entry_t swap
= { .val
= page_private(page
) };
654 mem_cgroup_swapout(page
, swap
);
655 __delete_from_swap_cache(page
);
656 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
657 swapcache_free(swap
);
659 void (*freepage
)(struct page
*);
662 freepage
= mapping
->a_ops
->freepage
;
664 * Remember a shadow entry for reclaimed file cache in
665 * order to detect refaults, thus thrashing, later on.
667 * But don't store shadows in an address space that is
668 * already exiting. This is not just an optizimation,
669 * inode reclaim needs to empty out the radix tree or
670 * the nodes are lost. Don't plant shadows behind its
673 * We also don't store shadows for DAX mappings because the
674 * only page cache pages found in these are zero pages
675 * covering holes, and because we don't want to mix DAX
676 * exceptional entries and shadow exceptional entries in the
679 if (reclaimed
&& page_is_file_cache(page
) &&
680 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
681 shadow
= workingset_eviction(mapping
, page
);
682 __delete_from_page_cache(page
, shadow
);
683 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
685 if (freepage
!= NULL
)
692 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
697 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
698 * someone else has a ref on the page, abort and return 0. If it was
699 * successfully detached, return 1. Assumes the caller has a single ref on
702 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
704 if (__remove_mapping(mapping
, page
, false)) {
706 * Unfreezing the refcount with 1 rather than 2 effectively
707 * drops the pagecache ref for us without requiring another
710 page_ref_unfreeze(page
, 1);
717 * putback_lru_page - put previously isolated page onto appropriate LRU list
718 * @page: page to be put back to appropriate lru list
720 * Add previously isolated @page to appropriate LRU list.
721 * Page may still be unevictable for other reasons.
723 * lru_lock must not be held, interrupts must be enabled.
725 void putback_lru_page(struct page
*page
)
728 int was_unevictable
= PageUnevictable(page
);
730 VM_BUG_ON_PAGE(PageLRU(page
), page
);
733 ClearPageUnevictable(page
);
735 if (page_evictable(page
)) {
737 * For evictable pages, we can use the cache.
738 * In event of a race, worst case is we end up with an
739 * unevictable page on [in]active list.
740 * We know how to handle that.
742 is_unevictable
= false;
746 * Put unevictable pages directly on zone's unevictable
749 is_unevictable
= true;
750 add_page_to_unevictable_list(page
);
752 * When racing with an mlock or AS_UNEVICTABLE clearing
753 * (page is unlocked) make sure that if the other thread
754 * does not observe our setting of PG_lru and fails
755 * isolation/check_move_unevictable_pages,
756 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
757 * the page back to the evictable list.
759 * The other side is TestClearPageMlocked() or shmem_lock().
765 * page's status can change while we move it among lru. If an evictable
766 * page is on unevictable list, it never be freed. To avoid that,
767 * check after we added it to the list, again.
769 if (is_unevictable
&& page_evictable(page
)) {
770 if (!isolate_lru_page(page
)) {
774 /* This means someone else dropped this page from LRU
775 * So, it will be freed or putback to LRU again. There is
776 * nothing to do here.
780 if (was_unevictable
&& !is_unevictable
)
781 count_vm_event(UNEVICTABLE_PGRESCUED
);
782 else if (!was_unevictable
&& is_unevictable
)
783 count_vm_event(UNEVICTABLE_PGCULLED
);
785 put_page(page
); /* drop ref from isolate */
788 enum page_references
{
790 PAGEREF_RECLAIM_CLEAN
,
795 static enum page_references
page_check_references(struct page
*page
,
796 struct scan_control
*sc
)
798 int referenced_ptes
, referenced_page
;
799 unsigned long vm_flags
;
801 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
803 referenced_page
= TestClearPageReferenced(page
);
806 * Mlock lost the isolation race with us. Let try_to_unmap()
807 * move the page to the unevictable list.
809 if (vm_flags
& VM_LOCKED
)
810 return PAGEREF_RECLAIM
;
812 if (referenced_ptes
) {
813 if (PageSwapBacked(page
))
814 return PAGEREF_ACTIVATE
;
816 * All mapped pages start out with page table
817 * references from the instantiating fault, so we need
818 * to look twice if a mapped file page is used more
821 * Mark it and spare it for another trip around the
822 * inactive list. Another page table reference will
823 * lead to its activation.
825 * Note: the mark is set for activated pages as well
826 * so that recently deactivated but used pages are
829 SetPageReferenced(page
);
831 if (referenced_page
|| referenced_ptes
> 1)
832 return PAGEREF_ACTIVATE
;
835 * Activate file-backed executable pages after first usage.
837 if (vm_flags
& VM_EXEC
)
838 return PAGEREF_ACTIVATE
;
843 /* Reclaim if clean, defer dirty pages to writeback */
844 if (referenced_page
&& !PageSwapBacked(page
))
845 return PAGEREF_RECLAIM_CLEAN
;
847 return PAGEREF_RECLAIM
;
850 /* Check if a page is dirty or under writeback */
851 static void page_check_dirty_writeback(struct page
*page
,
852 bool *dirty
, bool *writeback
)
854 struct address_space
*mapping
;
857 * Anonymous pages are not handled by flushers and must be written
858 * from reclaim context. Do not stall reclaim based on them
860 if (!page_is_file_cache(page
)) {
866 /* By default assume that the page flags are accurate */
867 *dirty
= PageDirty(page
);
868 *writeback
= PageWriteback(page
);
870 /* Verify dirty/writeback state if the filesystem supports it */
871 if (!page_has_private(page
))
874 mapping
= page_mapping(page
);
875 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
876 mapping
->a_ops
->is_dirty_writeback(page
, dirty
, writeback
);
880 * shrink_page_list() returns the number of reclaimed pages
882 static unsigned long shrink_page_list(struct list_head
*page_list
,
883 struct pglist_data
*pgdat
,
884 struct scan_control
*sc
,
885 enum ttu_flags ttu_flags
,
886 unsigned long *ret_nr_dirty
,
887 unsigned long *ret_nr_unqueued_dirty
,
888 unsigned long *ret_nr_congested
,
889 unsigned long *ret_nr_writeback
,
890 unsigned long *ret_nr_immediate
,
893 LIST_HEAD(ret_pages
);
894 LIST_HEAD(free_pages
);
896 unsigned long nr_unqueued_dirty
= 0;
897 unsigned long nr_dirty
= 0;
898 unsigned long nr_congested
= 0;
899 unsigned long nr_reclaimed
= 0;
900 unsigned long nr_writeback
= 0;
901 unsigned long nr_immediate
= 0;
905 while (!list_empty(page_list
)) {
906 struct address_space
*mapping
;
909 enum page_references references
= PAGEREF_RECLAIM_CLEAN
;
910 bool dirty
, writeback
;
911 bool lazyfree
= false;
912 int ret
= SWAP_SUCCESS
;
916 page
= lru_to_page(page_list
);
917 list_del(&page
->lru
);
919 if (!trylock_page(page
))
922 VM_BUG_ON_PAGE(PageActive(page
), page
);
926 if (unlikely(!page_evictable(page
)))
929 if (!sc
->may_unmap
&& page_mapped(page
))
932 /* Double the slab pressure for mapped and swapcache pages */
933 if (page_mapped(page
) || PageSwapCache(page
))
936 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
937 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
940 * The number of dirty pages determines if a zone is marked
941 * reclaim_congested which affects wait_iff_congested. kswapd
942 * will stall and start writing pages if the tail of the LRU
943 * is all dirty unqueued pages.
945 page_check_dirty_writeback(page
, &dirty
, &writeback
);
946 if (dirty
|| writeback
)
949 if (dirty
&& !writeback
)
953 * Treat this page as congested if the underlying BDI is or if
954 * pages are cycling through the LRU so quickly that the
955 * pages marked for immediate reclaim are making it to the
956 * end of the LRU a second time.
958 mapping
= page_mapping(page
);
959 if (((dirty
|| writeback
) && mapping
&&
960 inode_write_congested(mapping
->host
)) ||
961 (writeback
&& PageReclaim(page
)))
965 * If a page at the tail of the LRU is under writeback, there
966 * are three cases to consider.
968 * 1) If reclaim is encountering an excessive number of pages
969 * under writeback and this page is both under writeback and
970 * PageReclaim then it indicates that pages are being queued
971 * for IO but are being recycled through the LRU before the
972 * IO can complete. Waiting on the page itself risks an
973 * indefinite stall if it is impossible to writeback the
974 * page due to IO error or disconnected storage so instead
975 * note that the LRU is being scanned too quickly and the
976 * caller can stall after page list has been processed.
978 * 2) Global or new memcg reclaim encounters a page that is
979 * not marked for immediate reclaim, or the caller does not
980 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
981 * not to fs). In this case mark the page for immediate
982 * reclaim and continue scanning.
984 * Require may_enter_fs because we would wait on fs, which
985 * may not have submitted IO yet. And the loop driver might
986 * enter reclaim, and deadlock if it waits on a page for
987 * which it is needed to do the write (loop masks off
988 * __GFP_IO|__GFP_FS for this reason); but more thought
989 * would probably show more reasons.
991 * 3) Legacy memcg encounters a page that is already marked
992 * PageReclaim. memcg does not have any dirty pages
993 * throttling so we could easily OOM just because too many
994 * pages are in writeback and there is nothing else to
995 * reclaim. Wait for the writeback to complete.
997 if (PageWriteback(page
)) {
999 if (current_is_kswapd() &&
1000 PageReclaim(page
) &&
1001 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1006 } else if (sane_reclaim(sc
) ||
1007 !PageReclaim(page
) || !may_enter_fs
) {
1009 * This is slightly racy - end_page_writeback()
1010 * might have just cleared PageReclaim, then
1011 * setting PageReclaim here end up interpreted
1012 * as PageReadahead - but that does not matter
1013 * enough to care. What we do want is for this
1014 * page to have PageReclaim set next time memcg
1015 * reclaim reaches the tests above, so it will
1016 * then wait_on_page_writeback() to avoid OOM;
1017 * and it's also appropriate in global reclaim.
1019 SetPageReclaim(page
);
1026 wait_on_page_writeback(page
);
1027 /* then go back and try same page again */
1028 list_add_tail(&page
->lru
, page_list
);
1034 references
= page_check_references(page
, sc
);
1036 switch (references
) {
1037 case PAGEREF_ACTIVATE
:
1038 goto activate_locked
;
1041 case PAGEREF_RECLAIM
:
1042 case PAGEREF_RECLAIM_CLEAN
:
1043 ; /* try to reclaim the page below */
1047 * Anonymous process memory has backing store?
1048 * Try to allocate it some swap space here.
1050 if (PageAnon(page
) && !PageSwapCache(page
)) {
1051 if (!(sc
->gfp_mask
& __GFP_IO
))
1053 if (!add_to_swap(page
, page_list
))
1054 goto activate_locked
;
1058 /* Adding to swap updated mapping */
1059 mapping
= page_mapping(page
);
1060 } else if (unlikely(PageTransHuge(page
))) {
1061 /* Split file THP */
1062 if (split_huge_page_to_list(page
, page_list
))
1066 VM_BUG_ON_PAGE(PageTransHuge(page
), page
);
1069 * The page is mapped into the page tables of one or more
1070 * processes. Try to unmap it here.
1072 if (page_mapped(page
) && mapping
) {
1073 switch (ret
= try_to_unmap(page
, lazyfree
?
1074 (ttu_flags
| TTU_BATCH_FLUSH
| TTU_LZFREE
) :
1075 (ttu_flags
| TTU_BATCH_FLUSH
))) {
1077 goto activate_locked
;
1085 ; /* try to free the page below */
1089 if (PageDirty(page
)) {
1091 * Only kswapd can writeback filesystem pages to
1092 * avoid risk of stack overflow but only writeback
1093 * if many dirty pages have been encountered.
1095 if (page_is_file_cache(page
) &&
1096 (!current_is_kswapd() ||
1097 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1099 * Immediately reclaim when written back.
1100 * Similar in principal to deactivate_page()
1101 * except we already have the page isolated
1102 * and know it's dirty
1104 inc_node_page_state(page
, NR_VMSCAN_IMMEDIATE
);
1105 SetPageReclaim(page
);
1110 if (references
== PAGEREF_RECLAIM_CLEAN
)
1114 if (!sc
->may_writepage
)
1118 * Page is dirty. Flush the TLB if a writable entry
1119 * potentially exists to avoid CPU writes after IO
1120 * starts and then write it out here.
1122 try_to_unmap_flush_dirty();
1123 switch (pageout(page
, mapping
, sc
)) {
1127 goto activate_locked
;
1129 if (PageWriteback(page
))
1131 if (PageDirty(page
))
1135 * A synchronous write - probably a ramdisk. Go
1136 * ahead and try to reclaim the page.
1138 if (!trylock_page(page
))
1140 if (PageDirty(page
) || PageWriteback(page
))
1142 mapping
= page_mapping(page
);
1144 ; /* try to free the page below */
1149 * If the page has buffers, try to free the buffer mappings
1150 * associated with this page. If we succeed we try to free
1153 * We do this even if the page is PageDirty().
1154 * try_to_release_page() does not perform I/O, but it is
1155 * possible for a page to have PageDirty set, but it is actually
1156 * clean (all its buffers are clean). This happens if the
1157 * buffers were written out directly, with submit_bh(). ext3
1158 * will do this, as well as the blockdev mapping.
1159 * try_to_release_page() will discover that cleanness and will
1160 * drop the buffers and mark the page clean - it can be freed.
1162 * Rarely, pages can have buffers and no ->mapping. These are
1163 * the pages which were not successfully invalidated in
1164 * truncate_complete_page(). We try to drop those buffers here
1165 * and if that worked, and the page is no longer mapped into
1166 * process address space (page_count == 1) it can be freed.
1167 * Otherwise, leave the page on the LRU so it is swappable.
1169 if (page_has_private(page
)) {
1170 if (!try_to_release_page(page
, sc
->gfp_mask
))
1171 goto activate_locked
;
1172 if (!mapping
&& page_count(page
) == 1) {
1174 if (put_page_testzero(page
))
1178 * rare race with speculative reference.
1179 * the speculative reference will free
1180 * this page shortly, so we may
1181 * increment nr_reclaimed here (and
1182 * leave it off the LRU).
1191 if (!mapping
|| !__remove_mapping(mapping
, page
, true))
1195 * At this point, we have no other references and there is
1196 * no way to pick any more up (removed from LRU, removed
1197 * from pagecache). Can use non-atomic bitops now (and
1198 * we obviously don't have to worry about waking up a process
1199 * waiting on the page lock, because there are no references.
1201 __ClearPageLocked(page
);
1203 if (ret
== SWAP_LZFREE
)
1204 count_vm_event(PGLAZYFREED
);
1209 * Is there need to periodically free_page_list? It would
1210 * appear not as the counts should be low
1212 list_add(&page
->lru
, &free_pages
);
1216 if (PageSwapCache(page
))
1217 try_to_free_swap(page
);
1219 list_add(&page
->lru
, &ret_pages
);
1223 /* Not a candidate for swapping, so reclaim swap space. */
1224 if (PageSwapCache(page
) && mem_cgroup_swap_full(page
))
1225 try_to_free_swap(page
);
1226 VM_BUG_ON_PAGE(PageActive(page
), page
);
1227 SetPageActive(page
);
1232 list_add(&page
->lru
, &ret_pages
);
1233 VM_BUG_ON_PAGE(PageLRU(page
) || PageUnevictable(page
), page
);
1236 mem_cgroup_uncharge_list(&free_pages
);
1237 try_to_unmap_flush();
1238 free_hot_cold_page_list(&free_pages
, true);
1240 list_splice(&ret_pages
, page_list
);
1241 count_vm_events(PGACTIVATE
, pgactivate
);
1243 *ret_nr_dirty
+= nr_dirty
;
1244 *ret_nr_congested
+= nr_congested
;
1245 *ret_nr_unqueued_dirty
+= nr_unqueued_dirty
;
1246 *ret_nr_writeback
+= nr_writeback
;
1247 *ret_nr_immediate
+= nr_immediate
;
1248 return nr_reclaimed
;
1251 unsigned long reclaim_clean_pages_from_list(struct zone
*zone
,
1252 struct list_head
*page_list
)
1254 struct scan_control sc
= {
1255 .gfp_mask
= GFP_KERNEL
,
1256 .priority
= DEF_PRIORITY
,
1259 unsigned long ret
, dummy1
, dummy2
, dummy3
, dummy4
, dummy5
;
1260 struct page
*page
, *next
;
1261 LIST_HEAD(clean_pages
);
1263 list_for_each_entry_safe(page
, next
, page_list
, lru
) {
1264 if (page_is_file_cache(page
) && !PageDirty(page
) &&
1265 !__PageMovable(page
)) {
1266 ClearPageActive(page
);
1267 list_move(&page
->lru
, &clean_pages
);
1271 ret
= shrink_page_list(&clean_pages
, zone
->zone_pgdat
, &sc
,
1272 TTU_UNMAP
|TTU_IGNORE_ACCESS
,
1273 &dummy1
, &dummy2
, &dummy3
, &dummy4
, &dummy5
, true);
1274 list_splice(&clean_pages
, page_list
);
1275 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
, -ret
);
1280 * Attempt to remove the specified page from its LRU. Only take this page
1281 * if it is of the appropriate PageActive status. Pages which are being
1282 * freed elsewhere are also ignored.
1284 * page: page to consider
1285 * mode: one of the LRU isolation modes defined above
1287 * returns 0 on success, -ve errno on failure.
1289 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
1293 /* Only take pages on the LRU. */
1297 /* Compaction should not handle unevictable pages but CMA can do so */
1298 if (PageUnevictable(page
) && !(mode
& ISOLATE_UNEVICTABLE
))
1304 * To minimise LRU disruption, the caller can indicate that it only
1305 * wants to isolate pages it will be able to operate on without
1306 * blocking - clean pages for the most part.
1308 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1309 * is used by reclaim when it is cannot write to backing storage
1311 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1312 * that it is possible to migrate without blocking
1314 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1315 /* All the caller can do on PageWriteback is block */
1316 if (PageWriteback(page
))
1319 if (PageDirty(page
)) {
1320 struct address_space
*mapping
;
1322 /* ISOLATE_CLEAN means only clean pages */
1323 if (mode
& ISOLATE_CLEAN
)
1327 * Only pages without mappings or that have a
1328 * ->migratepage callback are possible to migrate
1331 mapping
= page_mapping(page
);
1332 if (mapping
&& !mapping
->a_ops
->migratepage
)
1337 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1340 if (likely(get_page_unless_zero(page
))) {
1342 * Be careful not to clear PageLRU until after we're
1343 * sure the page is not being freed elsewhere -- the
1344 * page release code relies on it.
1354 * zone_lru_lock is heavily contended. Some of the functions that
1355 * shrink the lists perform better by taking out a batch of pages
1356 * and working on them outside the LRU lock.
1358 * For pagecache intensive workloads, this function is the hottest
1359 * spot in the kernel (apart from copy_*_user functions).
1361 * Appropriate locks must be held before calling this function.
1363 * @nr_to_scan: The number of pages to look through on the list.
1364 * @lruvec: The LRU vector to pull pages from.
1365 * @dst: The temp list to put pages on to.
1366 * @nr_scanned: The number of pages that were scanned.
1367 * @sc: The scan_control struct for this reclaim session
1368 * @mode: One of the LRU isolation modes
1369 * @lru: LRU list id for isolating
1371 * returns how many pages were moved onto *@dst.
1373 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1374 struct lruvec
*lruvec
, struct list_head
*dst
,
1375 unsigned long *nr_scanned
, struct scan_control
*sc
,
1376 isolate_mode_t mode
, enum lru_list lru
)
1378 struct list_head
*src
= &lruvec
->lists
[lru
];
1379 unsigned long nr_taken
= 0;
1380 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1381 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1382 unsigned long scan
, nr_pages
;
1383 LIST_HEAD(pages_skipped
);
1385 for (scan
= 0; scan
< nr_to_scan
&& nr_taken
< nr_to_scan
&&
1386 !list_empty(src
); scan
++) {
1389 page
= lru_to_page(src
);
1390 prefetchw_prev_lru_page(page
, src
, flags
);
1392 VM_BUG_ON_PAGE(!PageLRU(page
), page
);
1394 if (page_zonenum(page
) > sc
->reclaim_idx
) {
1395 list_move(&page
->lru
, &pages_skipped
);
1396 nr_skipped
[page_zonenum(page
)]++;
1400 switch (__isolate_lru_page(page
, mode
)) {
1402 nr_pages
= hpage_nr_pages(page
);
1403 nr_taken
+= nr_pages
;
1404 nr_zone_taken
[page_zonenum(page
)] += nr_pages
;
1405 list_move(&page
->lru
, dst
);
1409 /* else it is being freed elsewhere */
1410 list_move(&page
->lru
, src
);
1419 * Splice any skipped pages to the start of the LRU list. Note that
1420 * this disrupts the LRU order when reclaiming for lower zones but
1421 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1422 * scanning would soon rescan the same pages to skip and put the
1423 * system at risk of premature OOM.
1425 if (!list_empty(&pages_skipped
)) {
1428 list_splice(&pages_skipped
, src
);
1429 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1430 if (!nr_skipped
[zid
])
1433 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1437 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
, scan
,
1438 nr_taken
, mode
, is_file_lru(lru
));
1439 for (scan
= 0; scan
< MAX_NR_ZONES
; scan
++) {
1440 nr_pages
= nr_zone_taken
[scan
];
1444 update_lru_size(lruvec
, lru
, scan
, -nr_pages
);
1450 * isolate_lru_page - tries to isolate a page from its LRU list
1451 * @page: page to isolate from its LRU list
1453 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1454 * vmstat statistic corresponding to whatever LRU list the page was on.
1456 * Returns 0 if the page was removed from an LRU list.
1457 * Returns -EBUSY if the page was not on an LRU list.
1459 * The returned page will have PageLRU() cleared. If it was found on
1460 * the active list, it will have PageActive set. If it was found on
1461 * the unevictable list, it will have the PageUnevictable bit set. That flag
1462 * may need to be cleared by the caller before letting the page go.
1464 * The vmstat statistic corresponding to the list on which the page was
1465 * found will be decremented.
1468 * (1) Must be called with an elevated refcount on the page. This is a
1469 * fundamentnal difference from isolate_lru_pages (which is called
1470 * without a stable reference).
1471 * (2) the lru_lock must not be held.
1472 * (3) interrupts must be enabled.
1474 int isolate_lru_page(struct page
*page
)
1478 VM_BUG_ON_PAGE(!page_count(page
), page
);
1479 WARN_RATELIMIT(PageTail(page
), "trying to isolate tail page");
1481 if (PageLRU(page
)) {
1482 struct zone
*zone
= page_zone(page
);
1483 struct lruvec
*lruvec
;
1485 spin_lock_irq(zone_lru_lock(zone
));
1486 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
1487 if (PageLRU(page
)) {
1488 int lru
= page_lru(page
);
1491 del_page_from_lru_list(page
, lruvec
, lru
);
1494 spin_unlock_irq(zone_lru_lock(zone
));
1500 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1501 * then get resheduled. When there are massive number of tasks doing page
1502 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1503 * the LRU list will go small and be scanned faster than necessary, leading to
1504 * unnecessary swapping, thrashing and OOM.
1506 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1507 struct scan_control
*sc
)
1509 unsigned long inactive
, isolated
;
1511 if (current_is_kswapd())
1514 if (!sane_reclaim(sc
))
1518 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1519 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1521 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1522 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1526 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1527 * won't get blocked by normal direct-reclaimers, forming a circular
1530 if ((sc
->gfp_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
1533 return isolated
> inactive
;
1536 static noinline_for_stack
void
1537 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1539 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1540 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1541 LIST_HEAD(pages_to_free
);
1544 * Put back any unfreeable pages.
1546 while (!list_empty(page_list
)) {
1547 struct page
*page
= lru_to_page(page_list
);
1550 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1551 list_del(&page
->lru
);
1552 if (unlikely(!page_evictable(page
))) {
1553 spin_unlock_irq(&pgdat
->lru_lock
);
1554 putback_lru_page(page
);
1555 spin_lock_irq(&pgdat
->lru_lock
);
1559 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1562 lru
= page_lru(page
);
1563 add_page_to_lru_list(page
, lruvec
, lru
);
1565 if (is_active_lru(lru
)) {
1566 int file
= is_file_lru(lru
);
1567 int numpages
= hpage_nr_pages(page
);
1568 reclaim_stat
->recent_rotated
[file
] += numpages
;
1570 if (put_page_testzero(page
)) {
1571 __ClearPageLRU(page
);
1572 __ClearPageActive(page
);
1573 del_page_from_lru_list(page
, lruvec
, lru
);
1575 if (unlikely(PageCompound(page
))) {
1576 spin_unlock_irq(&pgdat
->lru_lock
);
1577 mem_cgroup_uncharge(page
);
1578 (*get_compound_page_dtor(page
))(page
);
1579 spin_lock_irq(&pgdat
->lru_lock
);
1581 list_add(&page
->lru
, &pages_to_free
);
1586 * To save our caller's stack, now use input list for pages to free.
1588 list_splice(&pages_to_free
, page_list
);
1592 * If a kernel thread (such as nfsd for loop-back mounts) services
1593 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1594 * In that case we should only throttle if the backing device it is
1595 * writing to is congested. In other cases it is safe to throttle.
1597 static int current_may_throttle(void)
1599 return !(current
->flags
& PF_LESS_THROTTLE
) ||
1600 current
->backing_dev_info
== NULL
||
1601 bdi_write_congested(current
->backing_dev_info
);
1605 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1606 * of reclaimed pages
1608 static noinline_for_stack
unsigned long
1609 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1610 struct scan_control
*sc
, enum lru_list lru
)
1612 LIST_HEAD(page_list
);
1613 unsigned long nr_scanned
;
1614 unsigned long nr_reclaimed
= 0;
1615 unsigned long nr_taken
;
1616 unsigned long nr_dirty
= 0;
1617 unsigned long nr_congested
= 0;
1618 unsigned long nr_unqueued_dirty
= 0;
1619 unsigned long nr_writeback
= 0;
1620 unsigned long nr_immediate
= 0;
1621 isolate_mode_t isolate_mode
= 0;
1622 int file
= is_file_lru(lru
);
1623 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1624 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1626 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1627 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1629 /* We are about to die and free our memory. Return now. */
1630 if (fatal_signal_pending(current
))
1631 return SWAP_CLUSTER_MAX
;
1637 isolate_mode
|= ISOLATE_UNMAPPED
;
1638 if (!sc
->may_writepage
)
1639 isolate_mode
|= ISOLATE_CLEAN
;
1641 spin_lock_irq(&pgdat
->lru_lock
);
1643 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1644 &nr_scanned
, sc
, isolate_mode
, lru
);
1646 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1647 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1649 if (global_reclaim(sc
)) {
1650 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1651 if (current_is_kswapd())
1652 __count_vm_events(PGSCAN_KSWAPD
, nr_scanned
);
1654 __count_vm_events(PGSCAN_DIRECT
, nr_scanned
);
1656 spin_unlock_irq(&pgdat
->lru_lock
);
1661 nr_reclaimed
= shrink_page_list(&page_list
, pgdat
, sc
, TTU_UNMAP
,
1662 &nr_dirty
, &nr_unqueued_dirty
, &nr_congested
,
1663 &nr_writeback
, &nr_immediate
,
1666 spin_lock_irq(&pgdat
->lru_lock
);
1668 if (global_reclaim(sc
)) {
1669 if (current_is_kswapd())
1670 __count_vm_events(PGSTEAL_KSWAPD
, nr_reclaimed
);
1672 __count_vm_events(PGSTEAL_DIRECT
, nr_reclaimed
);
1675 putback_inactive_pages(lruvec
, &page_list
);
1677 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1679 spin_unlock_irq(&pgdat
->lru_lock
);
1681 mem_cgroup_uncharge_list(&page_list
);
1682 free_hot_cold_page_list(&page_list
, true);
1685 * If reclaim is isolating dirty pages under writeback, it implies
1686 * that the long-lived page allocation rate is exceeding the page
1687 * laundering rate. Either the global limits are not being effective
1688 * at throttling processes due to the page distribution throughout
1689 * zones or there is heavy usage of a slow backing device. The
1690 * only option is to throttle from reclaim context which is not ideal
1691 * as there is no guarantee the dirtying process is throttled in the
1692 * same way balance_dirty_pages() manages.
1694 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1695 * of pages under pages flagged for immediate reclaim and stall if any
1696 * are encountered in the nr_immediate check below.
1698 if (nr_writeback
&& nr_writeback
== nr_taken
)
1699 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
1702 * Legacy memcg will stall in page writeback so avoid forcibly
1705 if (sane_reclaim(sc
)) {
1707 * Tag a zone as congested if all the dirty pages scanned were
1708 * backed by a congested BDI and wait_iff_congested will stall.
1710 if (nr_dirty
&& nr_dirty
== nr_congested
)
1711 set_bit(PGDAT_CONGESTED
, &pgdat
->flags
);
1714 * If dirty pages are scanned that are not queued for IO, it
1715 * implies that flushers are not keeping up. In this case, flag
1716 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
1719 if (nr_unqueued_dirty
== nr_taken
)
1720 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
1723 * If kswapd scans pages marked marked for immediate
1724 * reclaim and under writeback (nr_immediate), it implies
1725 * that pages are cycling through the LRU faster than
1726 * they are written so also forcibly stall.
1728 if (nr_immediate
&& current_may_throttle())
1729 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1733 * Stall direct reclaim for IO completions if underlying BDIs or zone
1734 * is congested. Allow kswapd to continue until it starts encountering
1735 * unqueued dirty pages or cycling through the LRU too quickly.
1737 if (!sc
->hibernation_mode
&& !current_is_kswapd() &&
1738 current_may_throttle())
1739 wait_iff_congested(pgdat
, BLK_RW_ASYNC
, HZ
/10);
1741 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1742 nr_scanned
, nr_reclaimed
,
1743 sc
->priority
, file
);
1744 return nr_reclaimed
;
1748 * This moves pages from the active list to the inactive list.
1750 * We move them the other way if the page is referenced by one or more
1751 * processes, from rmap.
1753 * If the pages are mostly unmapped, the processing is fast and it is
1754 * appropriate to hold zone_lru_lock across the whole operation. But if
1755 * the pages are mapped, the processing is slow (page_referenced()) so we
1756 * should drop zone_lru_lock around each page. It's impossible to balance
1757 * this, so instead we remove the pages from the LRU while processing them.
1758 * It is safe to rely on PG_active against the non-LRU pages in here because
1759 * nobody will play with that bit on a non-LRU page.
1761 * The downside is that we have to touch page->_refcount against each page.
1762 * But we had to alter page->flags anyway.
1765 static void move_active_pages_to_lru(struct lruvec
*lruvec
,
1766 struct list_head
*list
,
1767 struct list_head
*pages_to_free
,
1770 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1771 unsigned long pgmoved
= 0;
1775 while (!list_empty(list
)) {
1776 page
= lru_to_page(list
);
1777 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1779 VM_BUG_ON_PAGE(PageLRU(page
), page
);
1782 nr_pages
= hpage_nr_pages(page
);
1783 update_lru_size(lruvec
, lru
, page_zonenum(page
), nr_pages
);
1784 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1785 pgmoved
+= nr_pages
;
1787 if (put_page_testzero(page
)) {
1788 __ClearPageLRU(page
);
1789 __ClearPageActive(page
);
1790 del_page_from_lru_list(page
, lruvec
, lru
);
1792 if (unlikely(PageCompound(page
))) {
1793 spin_unlock_irq(&pgdat
->lru_lock
);
1794 mem_cgroup_uncharge(page
);
1795 (*get_compound_page_dtor(page
))(page
);
1796 spin_lock_irq(&pgdat
->lru_lock
);
1798 list_add(&page
->lru
, pages_to_free
);
1802 if (!is_active_lru(lru
))
1803 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1806 static void shrink_active_list(unsigned long nr_to_scan
,
1807 struct lruvec
*lruvec
,
1808 struct scan_control
*sc
,
1811 unsigned long nr_taken
;
1812 unsigned long nr_scanned
;
1813 unsigned long vm_flags
;
1814 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1815 LIST_HEAD(l_active
);
1816 LIST_HEAD(l_inactive
);
1818 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1819 unsigned long nr_rotated
= 0;
1820 isolate_mode_t isolate_mode
= 0;
1821 int file
= is_file_lru(lru
);
1822 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1827 isolate_mode
|= ISOLATE_UNMAPPED
;
1828 if (!sc
->may_writepage
)
1829 isolate_mode
|= ISOLATE_CLEAN
;
1831 spin_lock_irq(&pgdat
->lru_lock
);
1833 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1834 &nr_scanned
, sc
, isolate_mode
, lru
);
1836 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1837 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1839 if (global_reclaim(sc
))
1840 __mod_node_page_state(pgdat
, NR_PAGES_SCANNED
, nr_scanned
);
1841 __count_vm_events(PGREFILL
, nr_scanned
);
1843 spin_unlock_irq(&pgdat
->lru_lock
);
1845 while (!list_empty(&l_hold
)) {
1847 page
= lru_to_page(&l_hold
);
1848 list_del(&page
->lru
);
1850 if (unlikely(!page_evictable(page
))) {
1851 putback_lru_page(page
);
1855 if (unlikely(buffer_heads_over_limit
)) {
1856 if (page_has_private(page
) && trylock_page(page
)) {
1857 if (page_has_private(page
))
1858 try_to_release_page(page
, 0);
1863 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1865 nr_rotated
+= hpage_nr_pages(page
);
1867 * Identify referenced, file-backed active pages and
1868 * give them one more trip around the active list. So
1869 * that executable code get better chances to stay in
1870 * memory under moderate memory pressure. Anon pages
1871 * are not likely to be evicted by use-once streaming
1872 * IO, plus JVM can create lots of anon VM_EXEC pages,
1873 * so we ignore them here.
1875 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1876 list_add(&page
->lru
, &l_active
);
1881 ClearPageActive(page
); /* we are de-activating */
1882 list_add(&page
->lru
, &l_inactive
);
1886 * Move pages back to the lru list.
1888 spin_lock_irq(&pgdat
->lru_lock
);
1890 * Count referenced pages from currently used mappings as rotated,
1891 * even though only some of them are actually re-activated. This
1892 * helps balance scan pressure between file and anonymous pages in
1895 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1897 move_active_pages_to_lru(lruvec
, &l_active
, &l_hold
, lru
);
1898 move_active_pages_to_lru(lruvec
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1899 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1900 spin_unlock_irq(&pgdat
->lru_lock
);
1902 mem_cgroup_uncharge_list(&l_hold
);
1903 free_hot_cold_page_list(&l_hold
, true);
1907 * The inactive anon list should be small enough that the VM never has
1908 * to do too much work.
1910 * The inactive file list should be small enough to leave most memory
1911 * to the established workingset on the scan-resistant active list,
1912 * but large enough to avoid thrashing the aggregate readahead window.
1914 * Both inactive lists should also be large enough that each inactive
1915 * page has a chance to be referenced again before it is reclaimed.
1917 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1918 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1919 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
1922 * memory ratio inactive
1923 * -------------------------------------
1932 static bool inactive_list_is_low(struct lruvec
*lruvec
, bool file
)
1934 unsigned long inactive_ratio
;
1935 unsigned long inactive
;
1936 unsigned long active
;
1940 * If we don't have swap space, anonymous page deactivation
1943 if (!file
&& !total_swap_pages
)
1946 inactive
= lruvec_lru_size(lruvec
, file
* LRU_FILE
);
1947 active
= lruvec_lru_size(lruvec
, file
* LRU_FILE
+ LRU_ACTIVE
);
1949 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1951 inactive_ratio
= int_sqrt(10 * gb
);
1955 return inactive
* inactive_ratio
< active
;
1958 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1959 struct lruvec
*lruvec
, struct scan_control
*sc
)
1961 if (is_active_lru(lru
)) {
1962 if (inactive_list_is_low(lruvec
, is_file_lru(lru
)))
1963 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1967 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1978 * Determine how aggressively the anon and file LRU lists should be
1979 * scanned. The relative value of each set of LRU lists is determined
1980 * by looking at the fraction of the pages scanned we did rotate back
1981 * onto the active list instead of evict.
1983 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1984 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1986 static void get_scan_count(struct lruvec
*lruvec
, struct mem_cgroup
*memcg
,
1987 struct scan_control
*sc
, unsigned long *nr
,
1988 unsigned long *lru_pages
)
1990 int swappiness
= mem_cgroup_swappiness(memcg
);
1991 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1993 u64 denominator
= 0; /* gcc */
1994 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1995 unsigned long anon_prio
, file_prio
;
1996 enum scan_balance scan_balance
;
1997 unsigned long anon
, file
;
1998 bool force_scan
= false;
1999 unsigned long ap
, fp
;
2005 * If the zone or memcg is small, nr[l] can be 0. This
2006 * results in no scanning on this priority and a potential
2007 * priority drop. Global direct reclaim can go to the next
2008 * zone and tends to have no problems. Global kswapd is for
2009 * zone balancing and it needs to scan a minimum amount. When
2010 * reclaiming for a memcg, a priority drop can cause high
2011 * latencies, so it's better to scan a minimum amount there as
2014 if (current_is_kswapd()) {
2015 if (!pgdat_reclaimable(pgdat
))
2017 if (!mem_cgroup_online(memcg
))
2020 if (!global_reclaim(sc
))
2023 /* If we have no swap space, do not bother scanning anon pages. */
2024 if (!sc
->may_swap
|| mem_cgroup_get_nr_swap_pages(memcg
) <= 0) {
2025 scan_balance
= SCAN_FILE
;
2030 * Global reclaim will swap to prevent OOM even with no
2031 * swappiness, but memcg users want to use this knob to
2032 * disable swapping for individual groups completely when
2033 * using the memory controller's swap limit feature would be
2036 if (!global_reclaim(sc
) && !swappiness
) {
2037 scan_balance
= SCAN_FILE
;
2042 * Do not apply any pressure balancing cleverness when the
2043 * system is close to OOM, scan both anon and file equally
2044 * (unless the swappiness setting disagrees with swapping).
2046 if (!sc
->priority
&& swappiness
) {
2047 scan_balance
= SCAN_EQUAL
;
2052 * Prevent the reclaimer from falling into the cache trap: as
2053 * cache pages start out inactive, every cache fault will tip
2054 * the scan balance towards the file LRU. And as the file LRU
2055 * shrinks, so does the window for rotation from references.
2056 * This means we have a runaway feedback loop where a tiny
2057 * thrashing file LRU becomes infinitely more attractive than
2058 * anon pages. Try to detect this based on file LRU size.
2060 if (global_reclaim(sc
)) {
2061 unsigned long pgdatfile
;
2062 unsigned long pgdatfree
;
2064 unsigned long total_high_wmark
= 0;
2066 pgdatfree
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2067 pgdatfile
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2068 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2070 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2071 struct zone
*zone
= &pgdat
->node_zones
[z
];
2072 if (!populated_zone(zone
))
2075 total_high_wmark
+= high_wmark_pages(zone
);
2078 if (unlikely(pgdatfile
+ pgdatfree
<= total_high_wmark
)) {
2079 scan_balance
= SCAN_ANON
;
2085 * If there is enough inactive page cache, i.e. if the size of the
2086 * inactive list is greater than that of the active list *and* the
2087 * inactive list actually has some pages to scan on this priority, we
2088 * do not reclaim anything from the anonymous working set right now.
2089 * Without the second condition we could end up never scanning an
2090 * lruvec even if it has plenty of old anonymous pages unless the
2091 * system is under heavy pressure.
2093 if (!inactive_list_is_low(lruvec
, true) &&
2094 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
) >> sc
->priority
) {
2095 scan_balance
= SCAN_FILE
;
2099 scan_balance
= SCAN_FRACT
;
2102 * With swappiness at 100, anonymous and file have the same priority.
2103 * This scanning priority is essentially the inverse of IO cost.
2105 anon_prio
= swappiness
;
2106 file_prio
= 200 - anon_prio
;
2109 * OK, so we have swap space and a fair amount of page cache
2110 * pages. We use the recently rotated / recently scanned
2111 * ratios to determine how valuable each cache is.
2113 * Because workloads change over time (and to avoid overflow)
2114 * we keep these statistics as a floating average, which ends
2115 * up weighing recent references more than old ones.
2117 * anon in [0], file in [1]
2120 anon
= lruvec_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
2121 lruvec_lru_size(lruvec
, LRU_INACTIVE_ANON
);
2122 file
= lruvec_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
2123 lruvec_lru_size(lruvec
, LRU_INACTIVE_FILE
);
2125 spin_lock_irq(&pgdat
->lru_lock
);
2126 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
2127 reclaim_stat
->recent_scanned
[0] /= 2;
2128 reclaim_stat
->recent_rotated
[0] /= 2;
2131 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
2132 reclaim_stat
->recent_scanned
[1] /= 2;
2133 reclaim_stat
->recent_rotated
[1] /= 2;
2137 * The amount of pressure on anon vs file pages is inversely
2138 * proportional to the fraction of recently scanned pages on
2139 * each list that were recently referenced and in active use.
2141 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
2142 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
2144 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
2145 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
2146 spin_unlock_irq(&pgdat
->lru_lock
);
2150 denominator
= ap
+ fp
+ 1;
2152 some_scanned
= false;
2153 /* Only use force_scan on second pass. */
2154 for (pass
= 0; !some_scanned
&& pass
< 2; pass
++) {
2156 for_each_evictable_lru(lru
) {
2157 int file
= is_file_lru(lru
);
2161 size
= lruvec_lru_size(lruvec
, lru
);
2162 scan
= size
>> sc
->priority
;
2164 if (!scan
&& pass
&& force_scan
)
2165 scan
= min(size
, SWAP_CLUSTER_MAX
);
2167 switch (scan_balance
) {
2169 /* Scan lists relative to size */
2173 * Scan types proportional to swappiness and
2174 * their relative recent reclaim efficiency.
2176 scan
= div64_u64(scan
* fraction
[file
],
2181 /* Scan one type exclusively */
2182 if ((scan_balance
== SCAN_FILE
) != file
) {
2188 /* Look ma, no brain */
2196 * Skip the second pass and don't force_scan,
2197 * if we found something to scan.
2199 some_scanned
|= !!scan
;
2204 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2205 static void init_tlb_ubc(void)
2208 * This deliberately does not clear the cpumask as it's expensive
2209 * and unnecessary. If there happens to be data in there then the
2210 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2211 * then will be cleared.
2213 current
->tlb_ubc
.flush_required
= false;
2216 static inline void init_tlb_ubc(void)
2219 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2222 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2224 static void shrink_node_memcg(struct pglist_data
*pgdat
, struct mem_cgroup
*memcg
,
2225 struct scan_control
*sc
, unsigned long *lru_pages
)
2227 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2228 unsigned long nr
[NR_LRU_LISTS
];
2229 unsigned long targets
[NR_LRU_LISTS
];
2230 unsigned long nr_to_scan
;
2232 unsigned long nr_reclaimed
= 0;
2233 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
2234 struct blk_plug plug
;
2237 get_scan_count(lruvec
, memcg
, sc
, nr
, lru_pages
);
2239 /* Record the original scan target for proportional adjustments later */
2240 memcpy(targets
, nr
, sizeof(nr
));
2243 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2244 * event that can occur when there is little memory pressure e.g.
2245 * multiple streaming readers/writers. Hence, we do not abort scanning
2246 * when the requested number of pages are reclaimed when scanning at
2247 * DEF_PRIORITY on the assumption that the fact we are direct
2248 * reclaiming implies that kswapd is not keeping up and it is best to
2249 * do a batch of work at once. For memcg reclaim one check is made to
2250 * abort proportional reclaim if either the file or anon lru has already
2251 * dropped to zero at the first pass.
2253 scan_adjusted
= (global_reclaim(sc
) && !current_is_kswapd() &&
2254 sc
->priority
== DEF_PRIORITY
);
2258 blk_start_plug(&plug
);
2259 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
2260 nr
[LRU_INACTIVE_FILE
]) {
2261 unsigned long nr_anon
, nr_file
, percentage
;
2262 unsigned long nr_scanned
;
2264 for_each_evictable_lru(lru
) {
2266 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
2267 nr
[lru
] -= nr_to_scan
;
2269 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
2274 if (nr_reclaimed
< nr_to_reclaim
|| scan_adjusted
)
2278 * For kswapd and memcg, reclaim at least the number of pages
2279 * requested. Ensure that the anon and file LRUs are scanned
2280 * proportionally what was requested by get_scan_count(). We
2281 * stop reclaiming one LRU and reduce the amount scanning
2282 * proportional to the original scan target.
2284 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
2285 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
2288 * It's just vindictive to attack the larger once the smaller
2289 * has gone to zero. And given the way we stop scanning the
2290 * smaller below, this makes sure that we only make one nudge
2291 * towards proportionality once we've got nr_to_reclaim.
2293 if (!nr_file
|| !nr_anon
)
2296 if (nr_file
> nr_anon
) {
2297 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
2298 targets
[LRU_ACTIVE_ANON
] + 1;
2300 percentage
= nr_anon
* 100 / scan_target
;
2302 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
2303 targets
[LRU_ACTIVE_FILE
] + 1;
2305 percentage
= nr_file
* 100 / scan_target
;
2308 /* Stop scanning the smaller of the LRU */
2310 nr
[lru
+ LRU_ACTIVE
] = 0;
2313 * Recalculate the other LRU scan count based on its original
2314 * scan target and the percentage scanning already complete
2316 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
2317 nr_scanned
= targets
[lru
] - nr
[lru
];
2318 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2319 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2322 nr_scanned
= targets
[lru
] - nr
[lru
];
2323 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
2324 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
2326 scan_adjusted
= true;
2328 blk_finish_plug(&plug
);
2329 sc
->nr_reclaimed
+= nr_reclaimed
;
2332 * Even if we did not try to evict anon pages at all, we want to
2333 * rebalance the anon lru active/inactive ratio.
2335 if (inactive_list_is_low(lruvec
, false))
2336 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2337 sc
, LRU_ACTIVE_ANON
);
2339 throttle_vm_writeout(sc
->gfp_mask
);
2342 /* Use reclaim/compaction for costly allocs or under memory pressure */
2343 static bool in_reclaim_compaction(struct scan_control
*sc
)
2345 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
2346 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
2347 sc
->priority
< DEF_PRIORITY
- 2))
2354 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2355 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2356 * true if more pages should be reclaimed such that when the page allocator
2357 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2358 * It will give up earlier than that if there is difficulty reclaiming pages.
2360 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
2361 unsigned long nr_reclaimed
,
2362 unsigned long nr_scanned
,
2363 struct scan_control
*sc
)
2365 unsigned long pages_for_compaction
;
2366 unsigned long inactive_lru_pages
;
2369 /* If not in reclaim/compaction mode, stop */
2370 if (!in_reclaim_compaction(sc
))
2373 /* Consider stopping depending on scan and reclaim activity */
2374 if (sc
->gfp_mask
& __GFP_REPEAT
) {
2376 * For __GFP_REPEAT allocations, stop reclaiming if the
2377 * full LRU list has been scanned and we are still failing
2378 * to reclaim pages. This full LRU scan is potentially
2379 * expensive but a __GFP_REPEAT caller really wants to succeed
2381 if (!nr_reclaimed
&& !nr_scanned
)
2385 * For non-__GFP_REPEAT allocations which can presumably
2386 * fail without consequence, stop if we failed to reclaim
2387 * any pages from the last SWAP_CLUSTER_MAX number of
2388 * pages that were scanned. This will return to the
2389 * caller faster at the risk reclaim/compaction and
2390 * the resulting allocation attempt fails
2397 * If we have not reclaimed enough pages for compaction and the
2398 * inactive lists are large enough, continue reclaiming
2400 pages_for_compaction
= (2UL << sc
->order
);
2401 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
2402 if (get_nr_swap_pages() > 0)
2403 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2404 if (sc
->nr_reclaimed
< pages_for_compaction
&&
2405 inactive_lru_pages
> pages_for_compaction
)
2408 /* If compaction would go ahead or the allocation would succeed, stop */
2409 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
2410 struct zone
*zone
= &pgdat
->node_zones
[z
];
2411 if (!populated_zone(zone
))
2414 switch (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
)) {
2415 case COMPACT_PARTIAL
:
2416 case COMPACT_CONTINUE
:
2419 /* check next zone */
2426 static bool shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
2428 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2429 unsigned long nr_reclaimed
, nr_scanned
;
2430 bool reclaimable
= false;
2433 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
2434 struct mem_cgroup_reclaim_cookie reclaim
= {
2436 .priority
= sc
->priority
,
2438 unsigned long node_lru_pages
= 0;
2439 struct mem_cgroup
*memcg
;
2441 nr_reclaimed
= sc
->nr_reclaimed
;
2442 nr_scanned
= sc
->nr_scanned
;
2444 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
2446 unsigned long lru_pages
;
2447 unsigned long reclaimed
;
2448 unsigned long scanned
;
2450 if (mem_cgroup_low(root
, memcg
)) {
2451 if (!sc
->may_thrash
)
2453 mem_cgroup_events(memcg
, MEMCG_LOW
, 1);
2456 reclaimed
= sc
->nr_reclaimed
;
2457 scanned
= sc
->nr_scanned
;
2459 shrink_node_memcg(pgdat
, memcg
, sc
, &lru_pages
);
2460 node_lru_pages
+= lru_pages
;
2462 if (!global_reclaim(sc
))
2463 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
,
2464 memcg
, sc
->nr_scanned
- scanned
,
2467 /* Record the group's reclaim efficiency */
2468 vmpressure(sc
->gfp_mask
, memcg
, false,
2469 sc
->nr_scanned
- scanned
,
2470 sc
->nr_reclaimed
- reclaimed
);
2473 * Direct reclaim and kswapd have to scan all memory
2474 * cgroups to fulfill the overall scan target for the
2477 * Limit reclaim, on the other hand, only cares about
2478 * nr_to_reclaim pages to be reclaimed and it will
2479 * retry with decreasing priority if one round over the
2480 * whole hierarchy is not sufficient.
2482 if (!global_reclaim(sc
) &&
2483 sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
2484 mem_cgroup_iter_break(root
, memcg
);
2487 } while ((memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
)));
2490 * Shrink the slab caches in the same proportion that
2491 * the eligible LRU pages were scanned.
2493 if (global_reclaim(sc
))
2494 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, NULL
,
2495 sc
->nr_scanned
- nr_scanned
,
2498 if (reclaim_state
) {
2499 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2500 reclaim_state
->reclaimed_slab
= 0;
2503 /* Record the subtree's reclaim efficiency */
2504 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
2505 sc
->nr_scanned
- nr_scanned
,
2506 sc
->nr_reclaimed
- nr_reclaimed
);
2508 if (sc
->nr_reclaimed
- nr_reclaimed
)
2511 } while (should_continue_reclaim(pgdat
, sc
->nr_reclaimed
- nr_reclaimed
,
2512 sc
->nr_scanned
- nr_scanned
, sc
));
2518 * Returns true if compaction should go ahead for a high-order request, or
2519 * the high-order allocation would succeed without compaction.
2521 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
2523 unsigned long watermark
;
2527 * Compaction takes time to run and there are potentially other
2528 * callers using the pages just freed. Continue reclaiming until
2529 * there is a buffer of free pages available to give compaction
2530 * a reasonable chance of completing and allocating the page
2532 watermark
= high_wmark_pages(zone
) + (2UL << sc
->order
);
2533 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
2536 * If compaction is deferred, reclaim up to a point where
2537 * compaction will have a chance of success when re-enabled
2539 if (compaction_deferred(zone
, sc
->order
))
2540 return watermark_ok
;
2543 * If compaction is not ready to start and allocation is not likely
2544 * to succeed without it, then keep reclaiming.
2546 if (compaction_suitable(zone
, sc
->order
, 0, sc
->reclaim_idx
) == COMPACT_SKIPPED
)
2549 return watermark_ok
;
2553 * This is the direct reclaim path, for page-allocating processes. We only
2554 * try to reclaim pages from zones which will satisfy the caller's allocation
2557 * If a zone is deemed to be full of pinned pages then just give it a light
2558 * scan then give up on it.
2560 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
2564 unsigned long nr_soft_reclaimed
;
2565 unsigned long nr_soft_scanned
;
2567 pg_data_t
*last_pgdat
= NULL
;
2570 * If the number of buffer_heads in the machine exceeds the maximum
2571 * allowed level, force direct reclaim to scan the highmem zone as
2572 * highmem pages could be pinning lowmem pages storing buffer_heads
2574 orig_mask
= sc
->gfp_mask
;
2575 if (buffer_heads_over_limit
) {
2576 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2577 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
2580 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2581 sc
->reclaim_idx
, sc
->nodemask
) {
2582 if (!populated_zone(zone
))
2586 * Take care memory controller reclaiming has small influence
2589 if (global_reclaim(sc
)) {
2590 if (!cpuset_zone_allowed(zone
,
2591 GFP_KERNEL
| __GFP_HARDWALL
))
2594 if (sc
->priority
!= DEF_PRIORITY
&&
2595 !pgdat_reclaimable(zone
->zone_pgdat
))
2596 continue; /* Let kswapd poll it */
2599 * If we already have plenty of memory free for
2600 * compaction in this zone, don't free any more.
2601 * Even though compaction is invoked for any
2602 * non-zero order, only frequent costly order
2603 * reclamation is disruptive enough to become a
2604 * noticeable problem, like transparent huge
2607 if (IS_ENABLED(CONFIG_COMPACTION
) &&
2608 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
2609 compaction_ready(zone
, sc
)) {
2610 sc
->compaction_ready
= true;
2615 * Shrink each node in the zonelist once. If the
2616 * zonelist is ordered by zone (not the default) then a
2617 * node may be shrunk multiple times but in that case
2618 * the user prefers lower zones being preserved.
2620 if (zone
->zone_pgdat
== last_pgdat
)
2624 * This steals pages from memory cgroups over softlimit
2625 * and returns the number of reclaimed pages and
2626 * scanned pages. This works for global memory pressure
2627 * and balancing, not for a memcg's limit.
2629 nr_soft_scanned
= 0;
2630 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
2631 sc
->order
, sc
->gfp_mask
,
2633 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2634 sc
->nr_scanned
+= nr_soft_scanned
;
2635 /* need some check for avoid more shrink_zone() */
2638 /* See comment about same check for global reclaim above */
2639 if (zone
->zone_pgdat
== last_pgdat
)
2641 last_pgdat
= zone
->zone_pgdat
;
2642 shrink_node(zone
->zone_pgdat
, sc
);
2646 * Restore to original mask to avoid the impact on the caller if we
2647 * promoted it to __GFP_HIGHMEM.
2649 sc
->gfp_mask
= orig_mask
;
2653 * This is the main entry point to direct page reclaim.
2655 * If a full scan of the inactive list fails to free enough memory then we
2656 * are "out of memory" and something needs to be killed.
2658 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2659 * high - the zone may be full of dirty or under-writeback pages, which this
2660 * caller can't do much about. We kick the writeback threads and take explicit
2661 * naps in the hope that some of these pages can be written. But if the
2662 * allocating task holds filesystem locks which prevent writeout this might not
2663 * work, and the allocation attempt will fail.
2665 * returns: 0, if no pages reclaimed
2666 * else, the number of pages reclaimed
2668 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2669 struct scan_control
*sc
)
2671 int initial_priority
= sc
->priority
;
2672 unsigned long total_scanned
= 0;
2673 unsigned long writeback_threshold
;
2675 delayacct_freepages_start();
2677 if (global_reclaim(sc
))
2678 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
2681 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
2684 shrink_zones(zonelist
, sc
);
2686 total_scanned
+= sc
->nr_scanned
;
2687 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2690 if (sc
->compaction_ready
)
2694 * If we're getting trouble reclaiming, start doing
2695 * writepage even in laptop mode.
2697 if (sc
->priority
< DEF_PRIORITY
- 2)
2698 sc
->may_writepage
= 1;
2701 * Try to write back as many pages as we just scanned. This
2702 * tends to cause slow streaming writers to write data to the
2703 * disk smoothly, at the dirtying rate, which is nice. But
2704 * that's undesirable in laptop mode, where we *want* lumpy
2705 * writeout. So in laptop mode, write out the whole world.
2707 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2708 if (total_scanned
> writeback_threshold
) {
2709 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2710 WB_REASON_TRY_TO_FREE_PAGES
);
2711 sc
->may_writepage
= 1;
2713 } while (--sc
->priority
>= 0);
2715 delayacct_freepages_end();
2717 if (sc
->nr_reclaimed
)
2718 return sc
->nr_reclaimed
;
2720 /* Aborted reclaim to try compaction? don't OOM, then */
2721 if (sc
->compaction_ready
)
2724 /* Untapped cgroup reserves? Don't OOM, retry. */
2725 if (!sc
->may_thrash
) {
2726 sc
->priority
= initial_priority
;
2734 static bool pfmemalloc_watermark_ok(pg_data_t
*pgdat
)
2737 unsigned long pfmemalloc_reserve
= 0;
2738 unsigned long free_pages
= 0;
2742 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
2743 zone
= &pgdat
->node_zones
[i
];
2744 if (!populated_zone(zone
) ||
2745 pgdat_reclaimable_pages(pgdat
) == 0)
2748 pfmemalloc_reserve
+= min_wmark_pages(zone
);
2749 free_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
2752 /* If there are no reserves (unexpected config) then do not throttle */
2753 if (!pfmemalloc_reserve
)
2756 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
2758 /* kswapd must be awake if processes are being throttled */
2759 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
2760 pgdat
->kswapd_classzone_idx
= min(pgdat
->kswapd_classzone_idx
,
2761 (enum zone_type
)ZONE_NORMAL
);
2762 wake_up_interruptible(&pgdat
->kswapd_wait
);
2769 * Throttle direct reclaimers if backing storage is backed by the network
2770 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2771 * depleted. kswapd will continue to make progress and wake the processes
2772 * when the low watermark is reached.
2774 * Returns true if a fatal signal was delivered during throttling. If this
2775 * happens, the page allocator should not consider triggering the OOM killer.
2777 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
2778 nodemask_t
*nodemask
)
2782 pg_data_t
*pgdat
= NULL
;
2785 * Kernel threads should not be throttled as they may be indirectly
2786 * responsible for cleaning pages necessary for reclaim to make forward
2787 * progress. kjournald for example may enter direct reclaim while
2788 * committing a transaction where throttling it could forcing other
2789 * processes to block on log_wait_commit().
2791 if (current
->flags
& PF_KTHREAD
)
2795 * If a fatal signal is pending, this process should not throttle.
2796 * It should return quickly so it can exit and free its memory
2798 if (fatal_signal_pending(current
))
2802 * Check if the pfmemalloc reserves are ok by finding the first node
2803 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2804 * GFP_KERNEL will be required for allocating network buffers when
2805 * swapping over the network so ZONE_HIGHMEM is unusable.
2807 * Throttling is based on the first usable node and throttled processes
2808 * wait on a queue until kswapd makes progress and wakes them. There
2809 * is an affinity then between processes waking up and where reclaim
2810 * progress has been made assuming the process wakes on the same node.
2811 * More importantly, processes running on remote nodes will not compete
2812 * for remote pfmemalloc reserves and processes on different nodes
2813 * should make reasonable progress.
2815 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2816 gfp_zone(gfp_mask
), nodemask
) {
2817 if (zone_idx(zone
) > ZONE_NORMAL
)
2820 /* Throttle based on the first usable node */
2821 pgdat
= zone
->zone_pgdat
;
2822 if (pfmemalloc_watermark_ok(pgdat
))
2827 /* If no zone was usable by the allocation flags then do not throttle */
2831 /* Account for the throttling */
2832 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
2835 * If the caller cannot enter the filesystem, it's possible that it
2836 * is due to the caller holding an FS lock or performing a journal
2837 * transaction in the case of a filesystem like ext[3|4]. In this case,
2838 * it is not safe to block on pfmemalloc_wait as kswapd could be
2839 * blocked waiting on the same lock. Instead, throttle for up to a
2840 * second before continuing.
2842 if (!(gfp_mask
& __GFP_FS
)) {
2843 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
2844 pfmemalloc_watermark_ok(pgdat
), HZ
);
2849 /* Throttle until kswapd wakes the process */
2850 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
2851 pfmemalloc_watermark_ok(pgdat
));
2854 if (fatal_signal_pending(current
))
2861 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2862 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2864 unsigned long nr_reclaimed
;
2865 struct scan_control sc
= {
2866 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2867 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
2868 .reclaim_idx
= gfp_zone(gfp_mask
),
2870 .nodemask
= nodemask
,
2871 .priority
= DEF_PRIORITY
,
2872 .may_writepage
= !laptop_mode
,
2878 * Do not enter reclaim if fatal signal was delivered while throttled.
2879 * 1 is returned so that the page allocator does not OOM kill at this
2882 if (throttle_direct_reclaim(gfp_mask
, zonelist
, nodemask
))
2885 trace_mm_vmscan_direct_reclaim_begin(order
,
2890 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2892 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2894 return nr_reclaimed
;
2899 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
2900 gfp_t gfp_mask
, bool noswap
,
2902 unsigned long *nr_scanned
)
2904 struct scan_control sc
= {
2905 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2906 .target_mem_cgroup
= memcg
,
2907 .may_writepage
= !laptop_mode
,
2909 .reclaim_idx
= MAX_NR_ZONES
- 1,
2910 .may_swap
= !noswap
,
2912 unsigned long lru_pages
;
2914 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2915 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2917 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2923 * NOTE: Although we can get the priority field, using it
2924 * here is not a good idea, since it limits the pages we can scan.
2925 * if we don't reclaim here, the shrink_node from balance_pgdat
2926 * will pick up pages from other mem cgroup's as well. We hack
2927 * the priority and make it zero.
2929 shrink_node_memcg(pgdat
, memcg
, &sc
, &lru_pages
);
2931 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2933 *nr_scanned
= sc
.nr_scanned
;
2934 return sc
.nr_reclaimed
;
2937 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2938 unsigned long nr_pages
,
2942 struct zonelist
*zonelist
;
2943 unsigned long nr_reclaimed
;
2945 struct scan_control sc
= {
2946 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
2947 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2948 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2949 .reclaim_idx
= MAX_NR_ZONES
- 1,
2950 .target_mem_cgroup
= memcg
,
2951 .priority
= DEF_PRIORITY
,
2952 .may_writepage
= !laptop_mode
,
2954 .may_swap
= may_swap
,
2958 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2959 * take care of from where we get pages. So the node where we start the
2960 * scan does not need to be the current node.
2962 nid
= mem_cgroup_select_victim_node(memcg
);
2964 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2966 trace_mm_vmscan_memcg_reclaim_begin(0,
2971 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2973 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2975 return nr_reclaimed
;
2979 static void age_active_anon(struct pglist_data
*pgdat
,
2980 struct scan_control
*sc
)
2982 struct mem_cgroup
*memcg
;
2984 if (!total_swap_pages
)
2987 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2989 struct lruvec
*lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
2991 if (inactive_list_is_low(lruvec
, false))
2992 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2993 sc
, LRU_ACTIVE_ANON
);
2995 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2999 static bool zone_balanced(struct zone
*zone
, int order
, int classzone_idx
)
3001 unsigned long mark
= high_wmark_pages(zone
);
3003 if (!zone_watermark_ok_safe(zone
, order
, mark
, classzone_idx
))
3007 * If any eligible zone is balanced then the node is not considered
3008 * to be congested or dirty
3010 clear_bit(PGDAT_CONGESTED
, &zone
->zone_pgdat
->flags
);
3011 clear_bit(PGDAT_DIRTY
, &zone
->zone_pgdat
->flags
);
3017 * Prepare kswapd for sleeping. This verifies that there are no processes
3018 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3020 * Returns true if kswapd is ready to sleep
3022 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3027 * The throttled processes are normally woken up in balance_pgdat() as
3028 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3029 * race between when kswapd checks the watermarks and a process gets
3030 * throttled. There is also a potential race if processes get
3031 * throttled, kswapd wakes, a large process exits thereby balancing the
3032 * zones, which causes kswapd to exit balance_pgdat() before reaching
3033 * the wake up checks. If kswapd is going to sleep, no process should
3034 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3035 * the wake up is premature, processes will wake kswapd and get
3036 * throttled again. The difference from wake ups in balance_pgdat() is
3037 * that here we are under prepare_to_wait().
3039 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
3040 wake_up_all(&pgdat
->pfmemalloc_wait
);
3042 for (i
= 0; i
<= classzone_idx
; i
++) {
3043 struct zone
*zone
= pgdat
->node_zones
+ i
;
3045 if (!populated_zone(zone
))
3048 if (!zone_balanced(zone
, order
, classzone_idx
))
3056 * kswapd shrinks a node of pages that are at or below the highest usable
3057 * zone that is currently unbalanced.
3059 * Returns true if kswapd scanned at least the requested number of pages to
3060 * reclaim or if the lack of progress was due to pages under writeback.
3061 * This is used to determine if the scanning priority needs to be raised.
3063 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
3064 struct scan_control
*sc
)
3069 /* Reclaim a number of pages proportional to the number of zones */
3070 sc
->nr_to_reclaim
= 0;
3071 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
3072 zone
= pgdat
->node_zones
+ z
;
3073 if (!populated_zone(zone
))
3076 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
3080 * Historically care was taken to put equal pressure on all zones but
3081 * now pressure is applied based on node LRU order.
3083 shrink_node(pgdat
, sc
);
3086 * Fragmentation may mean that the system cannot be rebalanced for
3087 * high-order allocations. If twice the allocation size has been
3088 * reclaimed then recheck watermarks only at order-0 to prevent
3089 * excessive reclaim. Assume that a process requested a high-order
3090 * can direct reclaim/compact.
3092 if (sc
->order
&& sc
->nr_reclaimed
>= 2UL << sc
->order
)
3095 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
3099 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3100 * that are eligible for use by the caller until at least one zone is
3103 * Returns the order kswapd finished reclaiming at.
3105 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3106 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3107 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3108 * or lower is eligible for reclaim until at least one usable zone is
3111 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int classzone_idx
)
3114 unsigned long nr_soft_reclaimed
;
3115 unsigned long nr_soft_scanned
;
3117 struct scan_control sc
= {
3118 .gfp_mask
= GFP_KERNEL
,
3120 .priority
= DEF_PRIORITY
,
3121 .may_writepage
= !laptop_mode
,
3125 count_vm_event(PAGEOUTRUN
);
3128 bool raise_priority
= true;
3130 sc
.nr_reclaimed
= 0;
3131 sc
.reclaim_idx
= classzone_idx
;
3134 * If the number of buffer_heads exceeds the maximum allowed
3135 * then consider reclaiming from all zones. This has a dual
3136 * purpose -- on 64-bit systems it is expected that
3137 * buffer_heads are stripped during active rotation. On 32-bit
3138 * systems, highmem pages can pin lowmem memory and shrinking
3139 * buffers can relieve lowmem pressure. Reclaim may still not
3140 * go ahead if all eligible zones for the original allocation
3141 * request are balanced to avoid excessive reclaim from kswapd.
3143 if (buffer_heads_over_limit
) {
3144 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
3145 zone
= pgdat
->node_zones
+ i
;
3146 if (!populated_zone(zone
))
3155 * Only reclaim if there are no eligible zones. Check from
3156 * high to low zone as allocations prefer higher zones.
3157 * Scanning from low to high zone would allow congestion to be
3158 * cleared during a very small window when a small low
3159 * zone was balanced even under extreme pressure when the
3160 * overall node may be congested. Note that sc.reclaim_idx
3161 * is not used as buffer_heads_over_limit may have adjusted
3164 for (i
= classzone_idx
; i
>= 0; i
--) {
3165 zone
= pgdat
->node_zones
+ i
;
3166 if (!populated_zone(zone
))
3169 if (zone_balanced(zone
, sc
.order
, classzone_idx
))
3174 * Do some background aging of the anon list, to give
3175 * pages a chance to be referenced before reclaiming. All
3176 * pages are rotated regardless of classzone as this is
3177 * about consistent aging.
3179 age_active_anon(pgdat
, &sc
);
3182 * If we're getting trouble reclaiming, start doing writepage
3183 * even in laptop mode.
3185 if (sc
.priority
< DEF_PRIORITY
- 2 || !pgdat_reclaimable(pgdat
))
3186 sc
.may_writepage
= 1;
3188 /* Call soft limit reclaim before calling shrink_node. */
3190 nr_soft_scanned
= 0;
3191 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
3192 sc
.gfp_mask
, &nr_soft_scanned
);
3193 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
3196 * There should be no need to raise the scanning priority if
3197 * enough pages are already being scanned that that high
3198 * watermark would be met at 100% efficiency.
3200 if (kswapd_shrink_node(pgdat
, &sc
))
3201 raise_priority
= false;
3204 * If the low watermark is met there is no need for processes
3205 * to be throttled on pfmemalloc_wait as they should not be
3206 * able to safely make forward progress. Wake them
3208 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
3209 pfmemalloc_watermark_ok(pgdat
))
3210 wake_up_all(&pgdat
->pfmemalloc_wait
);
3212 /* Check if kswapd should be suspending */
3213 if (try_to_freeze() || kthread_should_stop())
3217 * Raise priority if scanning rate is too low or there was no
3218 * progress in reclaiming pages
3220 if (raise_priority
|| !sc
.nr_reclaimed
)
3222 } while (sc
.priority
>= 1);
3226 * Return the order kswapd stopped reclaiming at as
3227 * prepare_kswapd_sleep() takes it into account. If another caller
3228 * entered the allocator slow path while kswapd was awake, order will
3229 * remain at the higher level.
3234 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
3235 unsigned int classzone_idx
)
3240 if (freezing(current
) || kthread_should_stop())
3243 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3245 /* Try to sleep for a short interval */
3246 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3248 * Compaction records what page blocks it recently failed to
3249 * isolate pages from and skips them in the future scanning.
3250 * When kswapd is going to sleep, it is reasonable to assume
3251 * that pages and compaction may succeed so reset the cache.
3253 reset_isolation_suitable(pgdat
);
3256 * We have freed the memory, now we should compact it to make
3257 * allocation of the requested order possible.
3259 wakeup_kcompactd(pgdat
, alloc_order
, classzone_idx
);
3261 remaining
= schedule_timeout(HZ
/10);
3264 * If woken prematurely then reset kswapd_classzone_idx and
3265 * order. The values will either be from a wakeup request or
3266 * the previous request that slept prematurely.
3269 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3270 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, reclaim_order
);
3273 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3274 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
3278 * After a short sleep, check if it was a premature sleep. If not, then
3279 * go fully to sleep until explicitly woken up.
3282 prepare_kswapd_sleep(pgdat
, reclaim_order
, classzone_idx
)) {
3283 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
3286 * vmstat counters are not perfectly accurate and the estimated
3287 * value for counters such as NR_FREE_PAGES can deviate from the
3288 * true value by nr_online_cpus * threshold. To avoid the zone
3289 * watermarks being breached while under pressure, we reduce the
3290 * per-cpu vmstat threshold while kswapd is awake and restore
3291 * them before going back to sleep.
3293 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
3295 if (!kthread_should_stop())
3298 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
3301 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
3303 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
3305 finish_wait(&pgdat
->kswapd_wait
, &wait
);
3309 * The background pageout daemon, started as a kernel thread
3310 * from the init process.
3312 * This basically trickles out pages so that we have _some_
3313 * free memory available even if there is no other activity
3314 * that frees anything up. This is needed for things like routing
3315 * etc, where we otherwise might have all activity going on in
3316 * asynchronous contexts that cannot page things out.
3318 * If there are applications that are active memory-allocators
3319 * (most normal use), this basically shouldn't matter.
3321 static int kswapd(void *p
)
3323 unsigned int alloc_order
, reclaim_order
, classzone_idx
;
3324 pg_data_t
*pgdat
= (pg_data_t
*)p
;
3325 struct task_struct
*tsk
= current
;
3327 struct reclaim_state reclaim_state
= {
3328 .reclaimed_slab
= 0,
3330 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
3332 lockdep_set_current_reclaim_state(GFP_KERNEL
);
3334 if (!cpumask_empty(cpumask
))
3335 set_cpus_allowed_ptr(tsk
, cpumask
);
3336 current
->reclaim_state
= &reclaim_state
;
3339 * Tell the memory management that we're a "memory allocator",
3340 * and that if we need more memory we should get access to it
3341 * regardless (see "__alloc_pages()"). "kswapd" should
3342 * never get caught in the normal page freeing logic.
3344 * (Kswapd normally doesn't need memory anyway, but sometimes
3345 * you need a small amount of memory in order to be able to
3346 * page out something else, and this flag essentially protects
3347 * us from recursively trying to free more memory as we're
3348 * trying to free the first piece of memory in the first place).
3350 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
3353 pgdat
->kswapd_order
= alloc_order
= reclaim_order
= 0;
3354 pgdat
->kswapd_classzone_idx
= classzone_idx
= 0;
3359 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
3362 /* Read the new order and classzone_idx */
3363 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3364 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3365 pgdat
->kswapd_order
= 0;
3366 pgdat
->kswapd_classzone_idx
= 0;
3368 ret
= try_to_freeze();
3369 if (kthread_should_stop())
3373 * We can speed up thawing tasks if we don't call balance_pgdat
3374 * after returning from the refrigerator
3380 * Reclaim begins at the requested order but if a high-order
3381 * reclaim fails then kswapd falls back to reclaiming for
3382 * order-0. If that happens, kswapd will consider sleeping
3383 * for the order it finished reclaiming at (reclaim_order)
3384 * but kcompactd is woken to compact for the original
3385 * request (alloc_order).
3387 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, classzone_idx
,
3389 reclaim_order
= balance_pgdat(pgdat
, alloc_order
, classzone_idx
);
3390 if (reclaim_order
< alloc_order
)
3391 goto kswapd_try_sleep
;
3393 alloc_order
= reclaim_order
= pgdat
->kswapd_order
;
3394 classzone_idx
= pgdat
->kswapd_classzone_idx
;
3397 tsk
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
);
3398 current
->reclaim_state
= NULL
;
3399 lockdep_clear_current_reclaim_state();
3405 * A zone is low on free memory, so wake its kswapd task to service it.
3407 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
3412 if (!populated_zone(zone
))
3415 if (!cpuset_zone_allowed(zone
, GFP_KERNEL
| __GFP_HARDWALL
))
3417 pgdat
= zone
->zone_pgdat
;
3418 pgdat
->kswapd_classzone_idx
= max(pgdat
->kswapd_classzone_idx
, classzone_idx
);
3419 pgdat
->kswapd_order
= max(pgdat
->kswapd_order
, order
);
3420 if (!waitqueue_active(&pgdat
->kswapd_wait
))
3423 /* Only wake kswapd if all zones are unbalanced */
3424 for (z
= 0; z
<= classzone_idx
; z
++) {
3425 zone
= pgdat
->node_zones
+ z
;
3426 if (!populated_zone(zone
))
3429 if (zone_balanced(zone
, order
, classzone_idx
))
3433 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
3434 wake_up_interruptible(&pgdat
->kswapd_wait
);
3437 #ifdef CONFIG_HIBERNATION
3439 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3442 * Rather than trying to age LRUs the aim is to preserve the overall
3443 * LRU order by reclaiming preferentially
3444 * inactive > active > active referenced > active mapped
3446 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3448 struct reclaim_state reclaim_state
;
3449 struct scan_control sc
= {
3450 .nr_to_reclaim
= nr_to_reclaim
,
3451 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3452 .reclaim_idx
= MAX_NR_ZONES
- 1,
3453 .priority
= DEF_PRIORITY
,
3457 .hibernation_mode
= 1,
3459 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3460 struct task_struct
*p
= current
;
3461 unsigned long nr_reclaimed
;
3463 p
->flags
|= PF_MEMALLOC
;
3464 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3465 reclaim_state
.reclaimed_slab
= 0;
3466 p
->reclaim_state
= &reclaim_state
;
3468 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
3470 p
->reclaim_state
= NULL
;
3471 lockdep_clear_current_reclaim_state();
3472 p
->flags
&= ~PF_MEMALLOC
;
3474 return nr_reclaimed
;
3476 #endif /* CONFIG_HIBERNATION */
3478 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3479 not required for correctness. So if the last cpu in a node goes
3480 away, we get changed to run anywhere: as the first one comes back,
3481 restore their cpu bindings. */
3482 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
3487 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3488 for_each_node_state(nid
, N_MEMORY
) {
3489 pg_data_t
*pgdat
= NODE_DATA(nid
);
3490 const struct cpumask
*mask
;
3492 mask
= cpumask_of_node(pgdat
->node_id
);
3494 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3495 /* One of our CPUs online: restore mask */
3496 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3503 * This kswapd start function will be called by init and node-hot-add.
3504 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3506 int kswapd_run(int nid
)
3508 pg_data_t
*pgdat
= NODE_DATA(nid
);
3514 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3515 if (IS_ERR(pgdat
->kswapd
)) {
3516 /* failure at boot is fatal */
3517 BUG_ON(system_state
== SYSTEM_BOOTING
);
3518 pr_err("Failed to start kswapd on node %d\n", nid
);
3519 ret
= PTR_ERR(pgdat
->kswapd
);
3520 pgdat
->kswapd
= NULL
;
3526 * Called by memory hotplug when all memory in a node is offlined. Caller must
3527 * hold mem_hotplug_begin/end().
3529 void kswapd_stop(int nid
)
3531 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3534 kthread_stop(kswapd
);
3535 NODE_DATA(nid
)->kswapd
= NULL
;
3539 static int __init
kswapd_init(void)
3544 for_each_node_state(nid
, N_MEMORY
)
3546 hotcpu_notifier(cpu_callback
, 0);
3550 module_init(kswapd_init
)
3556 * If non-zero call node_reclaim when the number of free pages falls below
3559 int node_reclaim_mode __read_mostly
;
3561 #define RECLAIM_OFF 0
3562 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3563 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3564 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3567 * Priority for NODE_RECLAIM. This determines the fraction of pages
3568 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3571 #define NODE_RECLAIM_PRIORITY 4
3574 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3577 int sysctl_min_unmapped_ratio
= 1;
3580 * If the number of slab pages in a zone grows beyond this percentage then
3581 * slab reclaim needs to occur.
3583 int sysctl_min_slab_ratio
= 5;
3585 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
3587 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
3588 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
3589 node_page_state(pgdat
, NR_ACTIVE_FILE
);
3592 * It's possible for there to be more file mapped pages than
3593 * accounted for by the pages on the file LRU lists because
3594 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3596 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3599 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3600 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
3602 unsigned long nr_pagecache_reclaimable
;
3603 unsigned long delta
= 0;
3606 * If RECLAIM_UNMAP is set, then all file pages are considered
3607 * potentially reclaimable. Otherwise, we have to worry about
3608 * pages like swapcache and node_unmapped_file_pages() provides
3611 if (node_reclaim_mode
& RECLAIM_UNMAP
)
3612 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
3614 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
3616 /* If we can't clean pages, remove dirty pages from consideration */
3617 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
3618 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
3620 /* Watch for any possible underflows due to delta */
3621 if (unlikely(delta
> nr_pagecache_reclaimable
))
3622 delta
= nr_pagecache_reclaimable
;
3624 return nr_pagecache_reclaimable
- delta
;
3628 * Try to free up some pages from this node through reclaim.
3630 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3632 /* Minimum pages needed in order to stay on node */
3633 const unsigned long nr_pages
= 1 << order
;
3634 struct task_struct
*p
= current
;
3635 struct reclaim_state reclaim_state
;
3636 int classzone_idx
= gfp_zone(gfp_mask
);
3637 struct scan_control sc
= {
3638 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
3639 .gfp_mask
= (gfp_mask
= memalloc_noio_flags(gfp_mask
)),
3641 .priority
= NODE_RECLAIM_PRIORITY
,
3642 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
3643 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
3645 .reclaim_idx
= classzone_idx
,
3650 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3651 * and we also need to be able to write out pages for RECLAIM_WRITE
3652 * and RECLAIM_UNMAP.
3654 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3655 lockdep_set_current_reclaim_state(gfp_mask
);
3656 reclaim_state
.reclaimed_slab
= 0;
3657 p
->reclaim_state
= &reclaim_state
;
3659 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
) {
3661 * Free memory by calling shrink zone with increasing
3662 * priorities until we have enough memory freed.
3665 shrink_node(pgdat
, &sc
);
3666 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3669 p
->reclaim_state
= NULL
;
3670 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3671 lockdep_clear_current_reclaim_state();
3672 return sc
.nr_reclaimed
>= nr_pages
;
3675 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
3680 * Node reclaim reclaims unmapped file backed pages and
3681 * slab pages if we are over the defined limits.
3683 * A small portion of unmapped file backed pages is needed for
3684 * file I/O otherwise pages read by file I/O will be immediately
3685 * thrown out if the node is overallocated. So we do not reclaim
3686 * if less than a specified percentage of the node is used by
3687 * unmapped file backed pages.
3689 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
3690 sum_zone_node_page_state(pgdat
->node_id
, NR_SLAB_RECLAIMABLE
) <= pgdat
->min_slab_pages
)
3691 return NODE_RECLAIM_FULL
;
3693 if (!pgdat_reclaimable(pgdat
))
3694 return NODE_RECLAIM_FULL
;
3697 * Do not scan if the allocation should not be delayed.
3699 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
3700 return NODE_RECLAIM_NOSCAN
;
3703 * Only run node reclaim on the local node or on nodes that do not
3704 * have associated processors. This will favor the local processor
3705 * over remote processors and spread off node memory allocations
3706 * as wide as possible.
3708 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
3709 return NODE_RECLAIM_NOSCAN
;
3711 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
3712 return NODE_RECLAIM_NOSCAN
;
3714 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
3715 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
3718 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3725 * page_evictable - test whether a page is evictable
3726 * @page: the page to test
3728 * Test whether page is evictable--i.e., should be placed on active/inactive
3729 * lists vs unevictable list.
3731 * Reasons page might not be evictable:
3732 * (1) page's mapping marked unevictable
3733 * (2) page is part of an mlocked VMA
3736 int page_evictable(struct page
*page
)
3738 return !mapping_unevictable(page_mapping(page
)) && !PageMlocked(page
);
3743 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3744 * @pages: array of pages to check
3745 * @nr_pages: number of pages to check
3747 * Checks pages for evictability and moves them to the appropriate lru list.
3749 * This function is only used for SysV IPC SHM_UNLOCK.
3751 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3753 struct lruvec
*lruvec
;
3754 struct zone
*zone
= NULL
;
3759 for (i
= 0; i
< nr_pages
; i
++) {
3760 struct page
*page
= pages
[i
];
3761 struct zone
*pagezone
;
3764 pagezone
= page_zone(page
);
3765 if (pagezone
!= zone
) {
3767 spin_unlock_irq(zone_lru_lock(zone
));
3769 spin_lock_irq(zone_lru_lock(zone
));
3771 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
3773 if (!PageLRU(page
) || !PageUnevictable(page
))
3776 if (page_evictable(page
)) {
3777 enum lru_list lru
= page_lru_base_type(page
);
3779 VM_BUG_ON_PAGE(PageActive(page
), page
);
3780 ClearPageUnevictable(page
);
3781 del_page_from_lru_list(page
, lruvec
, LRU_UNEVICTABLE
);
3782 add_page_to_lru_list(page
, lruvec
, lru
);
3788 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3789 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3790 spin_unlock_irq(zone_lru_lock(zone
));
3793 #endif /* CONFIG_SHMEM */