4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned;
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed;
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 unsigned long hibernation_mode;
68 /* This context's GFP mask */
73 /* Can mapped pages be reclaimed? */
76 /* Can pages be swapped as part of reclaim? */
81 /* Scan (total_size >> priority) pages at once */
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
88 struct mem_cgroup *target_mem_cgroup;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field) \
102 if ((_page)->lru.prev != _base) { \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
128 * From 0 .. 100. Higher means more swappy.
130 int vm_swappiness = 60;
131 long vm_total_pages; /* The total number of pages which the VM controls */
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
137 static bool global_reclaim(struct scan_control *sc)
139 return !sc->target_mem_cgroup;
142 static bool global_reclaim(struct scan_control *sc)
148 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
150 if (!mem_cgroup_disabled())
151 return mem_cgroup_get_lru_size(lruvec, lru);
153 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
157 * Add a shrinker callback to be called from the vm
159 void register_shrinker(struct shrinker *shrinker)
161 atomic_long_set(&shrinker->nr_in_batch, 0);
162 down_write(&shrinker_rwsem);
163 list_add_tail(&shrinker->list, &shrinker_list);
164 up_write(&shrinker_rwsem);
166 EXPORT_SYMBOL(register_shrinker);
171 void unregister_shrinker(struct shrinker *shrinker)
173 down_write(&shrinker_rwsem);
174 list_del(&shrinker->list);
175 up_write(&shrinker_rwsem);
177 EXPORT_SYMBOL(unregister_shrinker);
179 static inline int do_shrinker_shrink(struct shrinker *shrinker,
180 struct shrink_control *sc,
181 unsigned long nr_to_scan)
183 sc->nr_to_scan = nr_to_scan;
184 return (*shrinker->shrink)(shrinker, sc);
187 #define SHRINK_BATCH 128
189 * Call the shrink functions to age shrinkable caches
191 * Here we assume it costs one seek to replace a lru page and that it also
192 * takes a seek to recreate a cache object. With this in mind we age equal
193 * percentages of the lru and ageable caches. This should balance the seeks
194 * generated by these structures.
196 * If the vm encountered mapped pages on the LRU it increase the pressure on
197 * slab to avoid swapping.
199 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
201 * `lru_pages' represents the number of on-LRU pages in all the zones which
202 * are eligible for the caller's allocation attempt. It is used for balancing
203 * slab reclaim versus page reclaim.
205 * Returns the number of slab objects which we shrunk.
207 unsigned long shrink_slab(struct shrink_control *shrink,
208 unsigned long nr_pages_scanned,
209 unsigned long lru_pages)
211 struct shrinker *shrinker;
212 unsigned long ret = 0;
214 if (nr_pages_scanned == 0)
215 nr_pages_scanned = SWAP_CLUSTER_MAX;
217 if (!down_read_trylock(&shrinker_rwsem)) {
218 /* Assume we'll be able to shrink next time */
223 list_for_each_entry(shrinker, &shrinker_list, list) {
224 unsigned long long delta;
230 long batch_size = shrinker->batch ? shrinker->batch
233 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
238 * copy the current shrinker scan count into a local variable
239 * and zero it so that other concurrent shrinker invocations
240 * don't also do this scanning work.
242 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
245 delta = (4 * nr_pages_scanned) / shrinker->seeks;
247 do_div(delta, lru_pages + 1);
249 if (total_scan < 0) {
250 printk(KERN_ERR "shrink_slab: %pF negative objects to "
252 shrinker->shrink, total_scan);
253 total_scan = max_pass;
257 * We need to avoid excessive windup on filesystem shrinkers
258 * due to large numbers of GFP_NOFS allocations causing the
259 * shrinkers to return -1 all the time. This results in a large
260 * nr being built up so when a shrink that can do some work
261 * comes along it empties the entire cache due to nr >>>
262 * max_pass. This is bad for sustaining a working set in
265 * Hence only allow the shrinker to scan the entire cache when
266 * a large delta change is calculated directly.
268 if (delta < max_pass / 4)
269 total_scan = min(total_scan, max_pass / 2);
272 * Avoid risking looping forever due to too large nr value:
273 * never try to free more than twice the estimate number of
276 if (total_scan > max_pass * 2)
277 total_scan = max_pass * 2;
279 trace_mm_shrink_slab_start(shrinker, shrink, nr,
280 nr_pages_scanned, lru_pages,
281 max_pass, delta, total_scan);
283 while (total_scan >= batch_size) {
286 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287 shrink_ret = do_shrinker_shrink(shrinker, shrink,
289 if (shrink_ret == -1)
291 if (shrink_ret < nr_before)
292 ret += nr_before - shrink_ret;
293 count_vm_events(SLABS_SCANNED, batch_size);
294 total_scan -= batch_size;
300 * move the unused scan count back into the shrinker in a
301 * manner that handles concurrent updates. If we exhausted the
302 * scan, there is no need to do an update.
305 new_nr = atomic_long_add_return(total_scan,
306 &shrinker->nr_in_batch);
308 new_nr = atomic_long_read(&shrinker->nr_in_batch);
310 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
312 up_read(&shrinker_rwsem);
318 static inline int is_page_cache_freeable(struct page *page)
321 * A freeable page cache page is referenced only by the caller
322 * that isolated the page, the page cache radix tree and
323 * optional buffer heads at page->private.
325 return page_count(page) - page_has_private(page) == 2;
328 static int may_write_to_queue(struct backing_dev_info *bdi,
329 struct scan_control *sc)
331 if (current->flags & PF_SWAPWRITE)
333 if (!bdi_write_congested(bdi))
335 if (bdi == current->backing_dev_info)
341 * We detected a synchronous write error writing a page out. Probably
342 * -ENOSPC. We need to propagate that into the address_space for a subsequent
343 * fsync(), msync() or close().
345 * The tricky part is that after writepage we cannot touch the mapping: nothing
346 * prevents it from being freed up. But we have a ref on the page and once
347 * that page is locked, the mapping is pinned.
349 * We're allowed to run sleeping lock_page() here because we know the caller has
352 static void handle_write_error(struct address_space *mapping,
353 struct page *page, int error)
356 if (page_mapping(page) == mapping)
357 mapping_set_error(mapping, error);
361 /* possible outcome of pageout() */
363 /* failed to write page out, page is locked */
365 /* move page to the active list, page is locked */
367 /* page has been sent to the disk successfully, page is unlocked */
369 /* page is clean and locked */
374 * pageout is called by shrink_page_list() for each dirty page.
375 * Calls ->writepage().
377 static pageout_t pageout(struct page *page, struct address_space *mapping,
378 struct scan_control *sc)
381 * If the page is dirty, only perform writeback if that write
382 * will be non-blocking. To prevent this allocation from being
383 * stalled by pagecache activity. But note that there may be
384 * stalls if we need to run get_block(). We could test
385 * PagePrivate for that.
387 * If this process is currently in __generic_file_aio_write() against
388 * this page's queue, we can perform writeback even if that
391 * If the page is swapcache, write it back even if that would
392 * block, for some throttling. This happens by accident, because
393 * swap_backing_dev_info is bust: it doesn't reflect the
394 * congestion state of the swapdevs. Easy to fix, if needed.
396 if (!is_page_cache_freeable(page))
400 * Some data journaling orphaned pages can have
401 * page->mapping == NULL while being dirty with clean buffers.
403 if (page_has_private(page)) {
404 if (try_to_free_buffers(page)) {
405 ClearPageDirty(page);
406 printk("%s: orphaned page\n", __func__);
412 if (mapping->a_ops->writepage == NULL)
413 return PAGE_ACTIVATE;
414 if (!may_write_to_queue(mapping->backing_dev_info, sc))
417 if (clear_page_dirty_for_io(page)) {
419 struct writeback_control wbc = {
420 .sync_mode = WB_SYNC_NONE,
421 .nr_to_write = SWAP_CLUSTER_MAX,
423 .range_end = LLONG_MAX,
427 SetPageReclaim(page);
428 res = mapping->a_ops->writepage(page, &wbc);
430 handle_write_error(mapping, page, res);
431 if (res == AOP_WRITEPAGE_ACTIVATE) {
432 ClearPageReclaim(page);
433 return PAGE_ACTIVATE;
436 if (!PageWriteback(page)) {
437 /* synchronous write or broken a_ops? */
438 ClearPageReclaim(page);
440 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
441 inc_zone_page_state(page, NR_VMSCAN_WRITE);
449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
454 BUG_ON(!PageLocked(page));
455 BUG_ON(mapping != page_mapping(page));
457 spin_lock_irq(&mapping->tree_lock);
459 * The non racy check for a busy page.
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
466 * get_user_pages(&page);
467 * [user mapping goes away]
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
472 * !page_count(page) [good, discard it]
474 * [oops, our write_to data is lost]
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
483 if (!page_freeze_refs(page, 2))
485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page))) {
487 page_unfreeze_refs(page, 2);
491 if (PageSwapCache(page)) {
492 swp_entry_t swap = { .val = page_private(page) };
493 __delete_from_swap_cache(page);
494 spin_unlock_irq(&mapping->tree_lock);
495 swapcache_free(swap, page);
497 void (*freepage)(struct page *);
499 freepage = mapping->a_ops->freepage;
501 __delete_from_page_cache(page);
502 spin_unlock_irq(&mapping->tree_lock);
503 mem_cgroup_uncharge_cache_page(page);
505 if (freepage != NULL)
512 spin_unlock_irq(&mapping->tree_lock);
517 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
518 * someone else has a ref on the page, abort and return 0. If it was
519 * successfully detached, return 1. Assumes the caller has a single ref on
522 int remove_mapping(struct address_space *mapping, struct page *page)
524 if (__remove_mapping(mapping, page)) {
526 * Unfreezing the refcount with 1 rather than 2 effectively
527 * drops the pagecache ref for us without requiring another
530 page_unfreeze_refs(page, 1);
537 * putback_lru_page - put previously isolated page onto appropriate LRU list
538 * @page: page to be put back to appropriate lru list
540 * Add previously isolated @page to appropriate LRU list.
541 * Page may still be unevictable for other reasons.
543 * lru_lock must not be held, interrupts must be enabled.
545 void putback_lru_page(struct page *page)
548 int active = !!TestClearPageActive(page);
549 int was_unevictable = PageUnevictable(page);
551 VM_BUG_ON(PageLRU(page));
554 ClearPageUnevictable(page);
556 if (page_evictable(page)) {
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
563 lru = active + page_lru_base_type(page);
564 lru_cache_add_lru(page, lru);
567 * Put unevictable pages directly on zone's unevictable
570 lru = LRU_UNEVICTABLE;
571 add_page_to_unevictable_list(page);
573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
576 * isolation/check_move_unevictable_pages,
577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578 * the page back to the evictable list.
580 * The other side is TestClearPageMlocked() or shmem_lock().
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
590 if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
591 if (!isolate_lru_page(page)) {
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
601 if (was_unevictable && lru != LRU_UNEVICTABLE)
602 count_vm_event(UNEVICTABLE_PGRESCUED);
603 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604 count_vm_event(UNEVICTABLE_PGCULLED);
606 put_page(page); /* drop ref from isolate */
609 enum page_references {
611 PAGEREF_RECLAIM_CLEAN,
616 static enum page_references page_check_references(struct page *page,
617 struct scan_control *sc)
619 int referenced_ptes, referenced_page;
620 unsigned long vm_flags;
622 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
624 referenced_page = TestClearPageReferenced(page);
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
630 if (vm_flags & VM_LOCKED)
631 return PAGEREF_RECLAIM;
633 if (referenced_ptes) {
634 if (PageSwapBacked(page))
635 return PAGEREF_ACTIVATE;
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
650 SetPageReferenced(page);
652 if (referenced_page || referenced_ptes > 1)
653 return PAGEREF_ACTIVATE;
656 * Activate file-backed executable pages after first usage.
658 if (vm_flags & VM_EXEC)
659 return PAGEREF_ACTIVATE;
664 /* Reclaim if clean, defer dirty pages to writeback */
665 if (referenced_page && !PageSwapBacked(page))
666 return PAGEREF_RECLAIM_CLEAN;
668 return PAGEREF_RECLAIM;
672 * shrink_page_list() returns the number of reclaimed pages
674 static unsigned long shrink_page_list(struct list_head *page_list,
676 struct scan_control *sc,
677 enum ttu_flags ttu_flags,
678 unsigned long *ret_nr_dirty,
679 unsigned long *ret_nr_writeback,
682 LIST_HEAD(ret_pages);
683 LIST_HEAD(free_pages);
685 unsigned long nr_dirty = 0;
686 unsigned long nr_congested = 0;
687 unsigned long nr_reclaimed = 0;
688 unsigned long nr_writeback = 0;
692 mem_cgroup_uncharge_start();
693 while (!list_empty(page_list)) {
694 struct address_space *mapping;
697 enum page_references references = PAGEREF_RECLAIM_CLEAN;
701 page = lru_to_page(page_list);
702 list_del(&page->lru);
704 if (!trylock_page(page))
707 VM_BUG_ON(PageActive(page));
708 VM_BUG_ON(page_zone(page) != zone);
712 if (unlikely(!page_evictable(page)))
715 if (!sc->may_unmap && page_mapped(page))
718 /* Double the slab pressure for mapped and swapcache pages */
719 if (page_mapped(page) || PageSwapCache(page))
722 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
723 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
725 if (PageWriteback(page)) {
727 * memcg doesn't have any dirty pages throttling so we
728 * could easily OOM just because too many pages are in
729 * writeback and there is nothing else to reclaim.
731 * Check __GFP_IO, certainly because a loop driver
732 * thread might enter reclaim, and deadlock if it waits
733 * on a page for which it is needed to do the write
734 * (loop masks off __GFP_IO|__GFP_FS for this reason);
735 * but more thought would probably show more reasons.
737 * Don't require __GFP_FS, since we're not going into
738 * the FS, just waiting on its writeback completion.
739 * Worryingly, ext4 gfs2 and xfs allocate pages with
740 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
741 * testing may_enter_fs here is liable to OOM on them.
743 if (global_reclaim(sc) ||
744 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
746 * This is slightly racy - end_page_writeback()
747 * might have just cleared PageReclaim, then
748 * setting PageReclaim here end up interpreted
749 * as PageReadahead - but that does not matter
750 * enough to care. What we do want is for this
751 * page to have PageReclaim set next time memcg
752 * reclaim reaches the tests above, so it will
753 * then wait_on_page_writeback() to avoid OOM;
754 * and it's also appropriate in global reclaim.
756 SetPageReclaim(page);
760 wait_on_page_writeback(page);
764 references = page_check_references(page, sc);
766 switch (references) {
767 case PAGEREF_ACTIVATE:
768 goto activate_locked;
771 case PAGEREF_RECLAIM:
772 case PAGEREF_RECLAIM_CLEAN:
773 ; /* try to reclaim the page below */
777 * Anonymous process memory has backing store?
778 * Try to allocate it some swap space here.
780 if (PageAnon(page) && !PageSwapCache(page)) {
781 if (!(sc->gfp_mask & __GFP_IO))
783 if (!add_to_swap(page))
784 goto activate_locked;
788 mapping = page_mapping(page);
791 * The page is mapped into the page tables of one or more
792 * processes. Try to unmap it here.
794 if (page_mapped(page) && mapping) {
795 switch (try_to_unmap(page, ttu_flags)) {
797 goto activate_locked;
803 ; /* try to free the page below */
807 if (PageDirty(page)) {
811 * Only kswapd can writeback filesystem pages to
812 * avoid risk of stack overflow but do not writeback
813 * unless under significant pressure.
815 if (page_is_file_cache(page) &&
816 (!current_is_kswapd() ||
817 sc->priority >= DEF_PRIORITY - 2)) {
819 * Immediately reclaim when written back.
820 * Similar in principal to deactivate_page()
821 * except we already have the page isolated
822 * and know it's dirty
824 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
825 SetPageReclaim(page);
830 if (references == PAGEREF_RECLAIM_CLEAN)
834 if (!sc->may_writepage)
837 /* Page is dirty, try to write it out here */
838 switch (pageout(page, mapping, sc)) {
843 goto activate_locked;
845 if (PageWriteback(page))
851 * A synchronous write - probably a ramdisk. Go
852 * ahead and try to reclaim the page.
854 if (!trylock_page(page))
856 if (PageDirty(page) || PageWriteback(page))
858 mapping = page_mapping(page);
860 ; /* try to free the page below */
865 * If the page has buffers, try to free the buffer mappings
866 * associated with this page. If we succeed we try to free
869 * We do this even if the page is PageDirty().
870 * try_to_release_page() does not perform I/O, but it is
871 * possible for a page to have PageDirty set, but it is actually
872 * clean (all its buffers are clean). This happens if the
873 * buffers were written out directly, with submit_bh(). ext3
874 * will do this, as well as the blockdev mapping.
875 * try_to_release_page() will discover that cleanness and will
876 * drop the buffers and mark the page clean - it can be freed.
878 * Rarely, pages can have buffers and no ->mapping. These are
879 * the pages which were not successfully invalidated in
880 * truncate_complete_page(). We try to drop those buffers here
881 * and if that worked, and the page is no longer mapped into
882 * process address space (page_count == 1) it can be freed.
883 * Otherwise, leave the page on the LRU so it is swappable.
885 if (page_has_private(page)) {
886 if (!try_to_release_page(page, sc->gfp_mask))
887 goto activate_locked;
888 if (!mapping && page_count(page) == 1) {
890 if (put_page_testzero(page))
894 * rare race with speculative reference.
895 * the speculative reference will free
896 * this page shortly, so we may
897 * increment nr_reclaimed here (and
898 * leave it off the LRU).
906 if (!mapping || !__remove_mapping(mapping, page))
910 * At this point, we have no other references and there is
911 * no way to pick any more up (removed from LRU, removed
912 * from pagecache). Can use non-atomic bitops now (and
913 * we obviously don't have to worry about waking up a process
914 * waiting on the page lock, because there are no references.
916 __clear_page_locked(page);
921 * Is there need to periodically free_page_list? It would
922 * appear not as the counts should be low
924 list_add(&page->lru, &free_pages);
928 if (PageSwapCache(page))
929 try_to_free_swap(page);
931 putback_lru_page(page);
935 /* Not a candidate for swapping, so reclaim swap space. */
936 if (PageSwapCache(page) && vm_swap_full())
937 try_to_free_swap(page);
938 VM_BUG_ON(PageActive(page));
944 list_add(&page->lru, &ret_pages);
945 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
949 * Tag a zone as congested if all the dirty pages encountered were
950 * backed by a congested BDI. In this case, reclaimers should just
951 * back off and wait for congestion to clear because further reclaim
952 * will encounter the same problem
954 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
955 zone_set_flag(zone, ZONE_CONGESTED);
957 free_hot_cold_page_list(&free_pages, 1);
959 list_splice(&ret_pages, page_list);
960 count_vm_events(PGACTIVATE, pgactivate);
961 mem_cgroup_uncharge_end();
962 *ret_nr_dirty += nr_dirty;
963 *ret_nr_writeback += nr_writeback;
967 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
968 struct list_head *page_list)
970 struct scan_control sc = {
971 .gfp_mask = GFP_KERNEL,
972 .priority = DEF_PRIORITY,
975 unsigned long ret, dummy1, dummy2;
976 struct page *page, *next;
977 LIST_HEAD(clean_pages);
979 list_for_each_entry_safe(page, next, page_list, lru) {
980 if (page_is_file_cache(page) && !PageDirty(page)) {
981 ClearPageActive(page);
982 list_move(&page->lru, &clean_pages);
986 ret = shrink_page_list(&clean_pages, zone, &sc,
987 TTU_UNMAP|TTU_IGNORE_ACCESS,
988 &dummy1, &dummy2, true);
989 list_splice(&clean_pages, page_list);
990 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
995 * Attempt to remove the specified page from its LRU. Only take this page
996 * if it is of the appropriate PageActive status. Pages which are being
997 * freed elsewhere are also ignored.
999 * page: page to consider
1000 * mode: one of the LRU isolation modes defined above
1002 * returns 0 on success, -ve errno on failure.
1004 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1008 /* Only take pages on the LRU. */
1012 /* Compaction should not handle unevictable pages but CMA can do so */
1013 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1019 * To minimise LRU disruption, the caller can indicate that it only
1020 * wants to isolate pages it will be able to operate on without
1021 * blocking - clean pages for the most part.
1023 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1024 * is used by reclaim when it is cannot write to backing storage
1026 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1027 * that it is possible to migrate without blocking
1029 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1030 /* All the caller can do on PageWriteback is block */
1031 if (PageWriteback(page))
1034 if (PageDirty(page)) {
1035 struct address_space *mapping;
1037 /* ISOLATE_CLEAN means only clean pages */
1038 if (mode & ISOLATE_CLEAN)
1042 * Only pages without mappings or that have a
1043 * ->migratepage callback are possible to migrate
1046 mapping = page_mapping(page);
1047 if (mapping && !mapping->a_ops->migratepage)
1052 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1055 if (likely(get_page_unless_zero(page))) {
1057 * Be careful not to clear PageLRU until after we're
1058 * sure the page is not being freed elsewhere -- the
1059 * page release code relies on it.
1069 * zone->lru_lock is heavily contended. Some of the functions that
1070 * shrink the lists perform better by taking out a batch of pages
1071 * and working on them outside the LRU lock.
1073 * For pagecache intensive workloads, this function is the hottest
1074 * spot in the kernel (apart from copy_*_user functions).
1076 * Appropriate locks must be held before calling this function.
1078 * @nr_to_scan: The number of pages to look through on the list.
1079 * @lruvec: The LRU vector to pull pages from.
1080 * @dst: The temp list to put pages on to.
1081 * @nr_scanned: The number of pages that were scanned.
1082 * @sc: The scan_control struct for this reclaim session
1083 * @mode: One of the LRU isolation modes
1084 * @lru: LRU list id for isolating
1086 * returns how many pages were moved onto *@dst.
1088 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1089 struct lruvec *lruvec, struct list_head *dst,
1090 unsigned long *nr_scanned, struct scan_control *sc,
1091 isolate_mode_t mode, enum lru_list lru)
1093 struct list_head *src = &lruvec->lists[lru];
1094 unsigned long nr_taken = 0;
1097 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1101 page = lru_to_page(src);
1102 prefetchw_prev_lru_page(page, src, flags);
1104 VM_BUG_ON(!PageLRU(page));
1106 switch (__isolate_lru_page(page, mode)) {
1108 nr_pages = hpage_nr_pages(page);
1109 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1110 list_move(&page->lru, dst);
1111 nr_taken += nr_pages;
1115 /* else it is being freed elsewhere */
1116 list_move(&page->lru, src);
1125 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1126 nr_taken, mode, is_file_lru(lru));
1131 * isolate_lru_page - tries to isolate a page from its LRU list
1132 * @page: page to isolate from its LRU list
1134 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1135 * vmstat statistic corresponding to whatever LRU list the page was on.
1137 * Returns 0 if the page was removed from an LRU list.
1138 * Returns -EBUSY if the page was not on an LRU list.
1140 * The returned page will have PageLRU() cleared. If it was found on
1141 * the active list, it will have PageActive set. If it was found on
1142 * the unevictable list, it will have the PageUnevictable bit set. That flag
1143 * may need to be cleared by the caller before letting the page go.
1145 * The vmstat statistic corresponding to the list on which the page was
1146 * found will be decremented.
1149 * (1) Must be called with an elevated refcount on the page. This is a
1150 * fundamentnal difference from isolate_lru_pages (which is called
1151 * without a stable reference).
1152 * (2) the lru_lock must not be held.
1153 * (3) interrupts must be enabled.
1155 int isolate_lru_page(struct page *page)
1159 VM_BUG_ON(!page_count(page));
1161 if (PageLRU(page)) {
1162 struct zone *zone = page_zone(page);
1163 struct lruvec *lruvec;
1165 spin_lock_irq(&zone->lru_lock);
1166 lruvec = mem_cgroup_page_lruvec(page, zone);
1167 if (PageLRU(page)) {
1168 int lru = page_lru(page);
1171 del_page_from_lru_list(page, lruvec, lru);
1174 spin_unlock_irq(&zone->lru_lock);
1180 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1181 * then get resheduled. When there are massive number of tasks doing page
1182 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1183 * the LRU list will go small and be scanned faster than necessary, leading to
1184 * unnecessary swapping, thrashing and OOM.
1186 static int too_many_isolated(struct zone *zone, int file,
1187 struct scan_control *sc)
1189 unsigned long inactive, isolated;
1191 if (current_is_kswapd())
1194 if (!global_reclaim(sc))
1198 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1199 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1201 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1202 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1206 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1207 * won't get blocked by normal direct-reclaimers, forming a circular
1210 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1213 return isolated > inactive;
1216 static noinline_for_stack void
1217 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1219 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1220 struct zone *zone = lruvec_zone(lruvec);
1221 LIST_HEAD(pages_to_free);
1224 * Put back any unfreeable pages.
1226 while (!list_empty(page_list)) {
1227 struct page *page = lru_to_page(page_list);
1230 VM_BUG_ON(PageLRU(page));
1231 list_del(&page->lru);
1232 if (unlikely(!page_evictable(page))) {
1233 spin_unlock_irq(&zone->lru_lock);
1234 putback_lru_page(page);
1235 spin_lock_irq(&zone->lru_lock);
1239 lruvec = mem_cgroup_page_lruvec(page, zone);
1242 lru = page_lru(page);
1243 add_page_to_lru_list(page, lruvec, lru);
1245 if (is_active_lru(lru)) {
1246 int file = is_file_lru(lru);
1247 int numpages = hpage_nr_pages(page);
1248 reclaim_stat->recent_rotated[file] += numpages;
1250 if (put_page_testzero(page)) {
1251 __ClearPageLRU(page);
1252 __ClearPageActive(page);
1253 del_page_from_lru_list(page, lruvec, lru);
1255 if (unlikely(PageCompound(page))) {
1256 spin_unlock_irq(&zone->lru_lock);
1257 (*get_compound_page_dtor(page))(page);
1258 spin_lock_irq(&zone->lru_lock);
1260 list_add(&page->lru, &pages_to_free);
1265 * To save our caller's stack, now use input list for pages to free.
1267 list_splice(&pages_to_free, page_list);
1271 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1272 * of reclaimed pages
1274 static noinline_for_stack unsigned long
1275 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1276 struct scan_control *sc, enum lru_list lru)
1278 LIST_HEAD(page_list);
1279 unsigned long nr_scanned;
1280 unsigned long nr_reclaimed = 0;
1281 unsigned long nr_taken;
1282 unsigned long nr_dirty = 0;
1283 unsigned long nr_writeback = 0;
1284 isolate_mode_t isolate_mode = 0;
1285 int file = is_file_lru(lru);
1286 struct zone *zone = lruvec_zone(lruvec);
1287 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1289 while (unlikely(too_many_isolated(zone, file, sc))) {
1290 congestion_wait(BLK_RW_ASYNC, HZ/10);
1292 /* We are about to die and free our memory. Return now. */
1293 if (fatal_signal_pending(current))
1294 return SWAP_CLUSTER_MAX;
1300 isolate_mode |= ISOLATE_UNMAPPED;
1301 if (!sc->may_writepage)
1302 isolate_mode |= ISOLATE_CLEAN;
1304 spin_lock_irq(&zone->lru_lock);
1306 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1307 &nr_scanned, sc, isolate_mode, lru);
1309 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1310 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1312 if (global_reclaim(sc)) {
1313 zone->pages_scanned += nr_scanned;
1314 if (current_is_kswapd())
1315 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1317 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1319 spin_unlock_irq(&zone->lru_lock);
1324 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1325 &nr_dirty, &nr_writeback, false);
1327 spin_lock_irq(&zone->lru_lock);
1329 reclaim_stat->recent_scanned[file] += nr_taken;
1331 if (global_reclaim(sc)) {
1332 if (current_is_kswapd())
1333 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1336 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1340 putback_inactive_pages(lruvec, &page_list);
1342 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1344 spin_unlock_irq(&zone->lru_lock);
1346 free_hot_cold_page_list(&page_list, 1);
1349 * If reclaim is isolating dirty pages under writeback, it implies
1350 * that the long-lived page allocation rate is exceeding the page
1351 * laundering rate. Either the global limits are not being effective
1352 * at throttling processes due to the page distribution throughout
1353 * zones or there is heavy usage of a slow backing device. The
1354 * only option is to throttle from reclaim context which is not ideal
1355 * as there is no guarantee the dirtying process is throttled in the
1356 * same way balance_dirty_pages() manages.
1358 * This scales the number of dirty pages that must be under writeback
1359 * before throttling depending on priority. It is a simple backoff
1360 * function that has the most effect in the range DEF_PRIORITY to
1361 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1362 * in trouble and reclaim is considered to be in trouble.
1364 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1365 * DEF_PRIORITY-1 50% must be PageWriteback
1366 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1368 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1369 * isolated page is PageWriteback
1371 if (nr_writeback && nr_writeback >=
1372 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1373 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1375 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1377 nr_scanned, nr_reclaimed,
1379 trace_shrink_flags(file));
1380 return nr_reclaimed;
1384 * This moves pages from the active list to the inactive list.
1386 * We move them the other way if the page is referenced by one or more
1387 * processes, from rmap.
1389 * If the pages are mostly unmapped, the processing is fast and it is
1390 * appropriate to hold zone->lru_lock across the whole operation. But if
1391 * the pages are mapped, the processing is slow (page_referenced()) so we
1392 * should drop zone->lru_lock around each page. It's impossible to balance
1393 * this, so instead we remove the pages from the LRU while processing them.
1394 * It is safe to rely on PG_active against the non-LRU pages in here because
1395 * nobody will play with that bit on a non-LRU page.
1397 * The downside is that we have to touch page->_count against each page.
1398 * But we had to alter page->flags anyway.
1401 static void move_active_pages_to_lru(struct lruvec *lruvec,
1402 struct list_head *list,
1403 struct list_head *pages_to_free,
1406 struct zone *zone = lruvec_zone(lruvec);
1407 unsigned long pgmoved = 0;
1411 while (!list_empty(list)) {
1412 page = lru_to_page(list);
1413 lruvec = mem_cgroup_page_lruvec(page, zone);
1415 VM_BUG_ON(PageLRU(page));
1418 nr_pages = hpage_nr_pages(page);
1419 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1420 list_move(&page->lru, &lruvec->lists[lru]);
1421 pgmoved += nr_pages;
1423 if (put_page_testzero(page)) {
1424 __ClearPageLRU(page);
1425 __ClearPageActive(page);
1426 del_page_from_lru_list(page, lruvec, lru);
1428 if (unlikely(PageCompound(page))) {
1429 spin_unlock_irq(&zone->lru_lock);
1430 (*get_compound_page_dtor(page))(page);
1431 spin_lock_irq(&zone->lru_lock);
1433 list_add(&page->lru, pages_to_free);
1436 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1437 if (!is_active_lru(lru))
1438 __count_vm_events(PGDEACTIVATE, pgmoved);
1441 static void shrink_active_list(unsigned long nr_to_scan,
1442 struct lruvec *lruvec,
1443 struct scan_control *sc,
1446 unsigned long nr_taken;
1447 unsigned long nr_scanned;
1448 unsigned long vm_flags;
1449 LIST_HEAD(l_hold); /* The pages which were snipped off */
1450 LIST_HEAD(l_active);
1451 LIST_HEAD(l_inactive);
1453 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1454 unsigned long nr_rotated = 0;
1455 isolate_mode_t isolate_mode = 0;
1456 int file = is_file_lru(lru);
1457 struct zone *zone = lruvec_zone(lruvec);
1462 isolate_mode |= ISOLATE_UNMAPPED;
1463 if (!sc->may_writepage)
1464 isolate_mode |= ISOLATE_CLEAN;
1466 spin_lock_irq(&zone->lru_lock);
1468 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1469 &nr_scanned, sc, isolate_mode, lru);
1470 if (global_reclaim(sc))
1471 zone->pages_scanned += nr_scanned;
1473 reclaim_stat->recent_scanned[file] += nr_taken;
1475 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1476 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1477 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1478 spin_unlock_irq(&zone->lru_lock);
1480 while (!list_empty(&l_hold)) {
1482 page = lru_to_page(&l_hold);
1483 list_del(&page->lru);
1485 if (unlikely(!page_evictable(page))) {
1486 putback_lru_page(page);
1490 if (unlikely(buffer_heads_over_limit)) {
1491 if (page_has_private(page) && trylock_page(page)) {
1492 if (page_has_private(page))
1493 try_to_release_page(page, 0);
1498 if (page_referenced(page, 0, sc->target_mem_cgroup,
1500 nr_rotated += hpage_nr_pages(page);
1502 * Identify referenced, file-backed active pages and
1503 * give them one more trip around the active list. So
1504 * that executable code get better chances to stay in
1505 * memory under moderate memory pressure. Anon pages
1506 * are not likely to be evicted by use-once streaming
1507 * IO, plus JVM can create lots of anon VM_EXEC pages,
1508 * so we ignore them here.
1510 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1511 list_add(&page->lru, &l_active);
1516 ClearPageActive(page); /* we are de-activating */
1517 list_add(&page->lru, &l_inactive);
1521 * Move pages back to the lru list.
1523 spin_lock_irq(&zone->lru_lock);
1525 * Count referenced pages from currently used mappings as rotated,
1526 * even though only some of them are actually re-activated. This
1527 * helps balance scan pressure between file and anonymous pages in
1530 reclaim_stat->recent_rotated[file] += nr_rotated;
1532 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1533 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1534 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1535 spin_unlock_irq(&zone->lru_lock);
1537 free_hot_cold_page_list(&l_hold, 1);
1541 static int inactive_anon_is_low_global(struct zone *zone)
1543 unsigned long active, inactive;
1545 active = zone_page_state(zone, NR_ACTIVE_ANON);
1546 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1548 if (inactive * zone->inactive_ratio < active)
1555 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1556 * @lruvec: LRU vector to check
1558 * Returns true if the zone does not have enough inactive anon pages,
1559 * meaning some active anon pages need to be deactivated.
1561 static int inactive_anon_is_low(struct lruvec *lruvec)
1564 * If we don't have swap space, anonymous page deactivation
1567 if (!total_swap_pages)
1570 if (!mem_cgroup_disabled())
1571 return mem_cgroup_inactive_anon_is_low(lruvec);
1573 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1576 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1582 static int inactive_file_is_low_global(struct zone *zone)
1584 unsigned long active, inactive;
1586 active = zone_page_state(zone, NR_ACTIVE_FILE);
1587 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1589 return (active > inactive);
1593 * inactive_file_is_low - check if file pages need to be deactivated
1594 * @lruvec: LRU vector to check
1596 * When the system is doing streaming IO, memory pressure here
1597 * ensures that active file pages get deactivated, until more
1598 * than half of the file pages are on the inactive list.
1600 * Once we get to that situation, protect the system's working
1601 * set from being evicted by disabling active file page aging.
1603 * This uses a different ratio than the anonymous pages, because
1604 * the page cache uses a use-once replacement algorithm.
1606 static int inactive_file_is_low(struct lruvec *lruvec)
1608 if (!mem_cgroup_disabled())
1609 return mem_cgroup_inactive_file_is_low(lruvec);
1611 return inactive_file_is_low_global(lruvec_zone(lruvec));
1614 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1616 if (is_file_lru(lru))
1617 return inactive_file_is_low(lruvec);
1619 return inactive_anon_is_low(lruvec);
1622 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1623 struct lruvec *lruvec, struct scan_control *sc)
1625 if (is_active_lru(lru)) {
1626 if (inactive_list_is_low(lruvec, lru))
1627 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1631 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1634 static int vmscan_swappiness(struct scan_control *sc)
1636 if (global_reclaim(sc))
1637 return vm_swappiness;
1638 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1649 * Determine how aggressively the anon and file LRU lists should be
1650 * scanned. The relative value of each set of LRU lists is determined
1651 * by looking at the fraction of the pages scanned we did rotate back
1652 * onto the active list instead of evict.
1654 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1655 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1657 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1660 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1662 u64 denominator = 0; /* gcc */
1663 struct zone *zone = lruvec_zone(lruvec);
1664 unsigned long anon_prio, file_prio;
1665 enum scan_balance scan_balance;
1666 unsigned long anon, file, free;
1667 bool force_scan = false;
1668 unsigned long ap, fp;
1672 * If the zone or memcg is small, nr[l] can be 0. This
1673 * results in no scanning on this priority and a potential
1674 * priority drop. Global direct reclaim can go to the next
1675 * zone and tends to have no problems. Global kswapd is for
1676 * zone balancing and it needs to scan a minimum amount. When
1677 * reclaiming for a memcg, a priority drop can cause high
1678 * latencies, so it's better to scan a minimum amount there as
1681 if (current_is_kswapd() && zone->all_unreclaimable)
1683 if (!global_reclaim(sc))
1686 /* If we have no swap space, do not bother scanning anon pages. */
1687 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1688 scan_balance = SCAN_FILE;
1693 * Global reclaim will swap to prevent OOM even with no
1694 * swappiness, but memcg users want to use this knob to
1695 * disable swapping for individual groups completely when
1696 * using the memory controller's swap limit feature would be
1699 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1700 scan_balance = SCAN_FILE;
1705 * Do not apply any pressure balancing cleverness when the
1706 * system is close to OOM, scan both anon and file equally
1707 * (unless the swappiness setting disagrees with swapping).
1709 if (!sc->priority && vmscan_swappiness(sc)) {
1710 scan_balance = SCAN_EQUAL;
1714 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1715 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1716 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1717 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1720 * If it's foreseeable that reclaiming the file cache won't be
1721 * enough to get the zone back into a desirable shape, we have
1722 * to swap. Better start now and leave the - probably heavily
1723 * thrashing - remaining file pages alone.
1725 if (global_reclaim(sc)) {
1726 free = zone_page_state(zone, NR_FREE_PAGES);
1727 if (unlikely(file + free <= high_wmark_pages(zone))) {
1728 scan_balance = SCAN_ANON;
1734 * There is enough inactive page cache, do not reclaim
1735 * anything from the anonymous working set right now.
1737 if (!inactive_file_is_low(lruvec)) {
1738 scan_balance = SCAN_FILE;
1742 scan_balance = SCAN_FRACT;
1745 * With swappiness at 100, anonymous and file have the same priority.
1746 * This scanning priority is essentially the inverse of IO cost.
1748 anon_prio = vmscan_swappiness(sc);
1749 file_prio = 200 - anon_prio;
1752 * OK, so we have swap space and a fair amount of page cache
1753 * pages. We use the recently rotated / recently scanned
1754 * ratios to determine how valuable each cache is.
1756 * Because workloads change over time (and to avoid overflow)
1757 * we keep these statistics as a floating average, which ends
1758 * up weighing recent references more than old ones.
1760 * anon in [0], file in [1]
1762 spin_lock_irq(&zone->lru_lock);
1763 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1764 reclaim_stat->recent_scanned[0] /= 2;
1765 reclaim_stat->recent_rotated[0] /= 2;
1768 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1769 reclaim_stat->recent_scanned[1] /= 2;
1770 reclaim_stat->recent_rotated[1] /= 2;
1774 * The amount of pressure on anon vs file pages is inversely
1775 * proportional to the fraction of recently scanned pages on
1776 * each list that were recently referenced and in active use.
1778 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1779 ap /= reclaim_stat->recent_rotated[0] + 1;
1781 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1782 fp /= reclaim_stat->recent_rotated[1] + 1;
1783 spin_unlock_irq(&zone->lru_lock);
1787 denominator = ap + fp + 1;
1789 for_each_evictable_lru(lru) {
1790 int file = is_file_lru(lru);
1794 size = get_lru_size(lruvec, lru);
1795 scan = size >> sc->priority;
1797 if (!scan && force_scan)
1798 scan = min(size, SWAP_CLUSTER_MAX);
1800 switch (scan_balance) {
1802 /* Scan lists relative to size */
1806 * Scan types proportional to swappiness and
1807 * their relative recent reclaim efficiency.
1809 scan = div64_u64(scan * fraction[file], denominator);
1813 /* Scan one type exclusively */
1814 if ((scan_balance == SCAN_FILE) != file)
1818 /* Look ma, no brain */
1826 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1828 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1830 unsigned long nr[NR_LRU_LISTS];
1831 unsigned long nr_to_scan;
1833 unsigned long nr_reclaimed = 0;
1834 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1835 struct blk_plug plug;
1837 get_scan_count(lruvec, sc, nr);
1839 blk_start_plug(&plug);
1840 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1841 nr[LRU_INACTIVE_FILE]) {
1842 for_each_evictable_lru(lru) {
1844 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1845 nr[lru] -= nr_to_scan;
1847 nr_reclaimed += shrink_list(lru, nr_to_scan,
1852 * On large memory systems, scan >> priority can become
1853 * really large. This is fine for the starting priority;
1854 * we want to put equal scanning pressure on each zone.
1855 * However, if the VM has a harder time of freeing pages,
1856 * with multiple processes reclaiming pages, the total
1857 * freeing target can get unreasonably large.
1859 if (nr_reclaimed >= nr_to_reclaim &&
1860 sc->priority < DEF_PRIORITY)
1863 blk_finish_plug(&plug);
1864 sc->nr_reclaimed += nr_reclaimed;
1867 * Even if we did not try to evict anon pages at all, we want to
1868 * rebalance the anon lru active/inactive ratio.
1870 if (inactive_anon_is_low(lruvec))
1871 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1872 sc, LRU_ACTIVE_ANON);
1874 throttle_vm_writeout(sc->gfp_mask);
1877 /* Use reclaim/compaction for costly allocs or under memory pressure */
1878 static bool in_reclaim_compaction(struct scan_control *sc)
1880 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1881 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1882 sc->priority < DEF_PRIORITY - 2))
1889 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1890 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1891 * true if more pages should be reclaimed such that when the page allocator
1892 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1893 * It will give up earlier than that if there is difficulty reclaiming pages.
1895 static inline bool should_continue_reclaim(struct zone *zone,
1896 unsigned long nr_reclaimed,
1897 unsigned long nr_scanned,
1898 struct scan_control *sc)
1900 unsigned long pages_for_compaction;
1901 unsigned long inactive_lru_pages;
1903 /* If not in reclaim/compaction mode, stop */
1904 if (!in_reclaim_compaction(sc))
1907 /* Consider stopping depending on scan and reclaim activity */
1908 if (sc->gfp_mask & __GFP_REPEAT) {
1910 * For __GFP_REPEAT allocations, stop reclaiming if the
1911 * full LRU list has been scanned and we are still failing
1912 * to reclaim pages. This full LRU scan is potentially
1913 * expensive but a __GFP_REPEAT caller really wants to succeed
1915 if (!nr_reclaimed && !nr_scanned)
1919 * For non-__GFP_REPEAT allocations which can presumably
1920 * fail without consequence, stop if we failed to reclaim
1921 * any pages from the last SWAP_CLUSTER_MAX number of
1922 * pages that were scanned. This will return to the
1923 * caller faster at the risk reclaim/compaction and
1924 * the resulting allocation attempt fails
1931 * If we have not reclaimed enough pages for compaction and the
1932 * inactive lists are large enough, continue reclaiming
1934 pages_for_compaction = (2UL << sc->order);
1935 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
1936 if (nr_swap_pages > 0)
1937 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
1938 if (sc->nr_reclaimed < pages_for_compaction &&
1939 inactive_lru_pages > pages_for_compaction)
1942 /* If compaction would go ahead or the allocation would succeed, stop */
1943 switch (compaction_suitable(zone, sc->order)) {
1944 case COMPACT_PARTIAL:
1945 case COMPACT_CONTINUE:
1952 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1954 unsigned long nr_reclaimed, nr_scanned;
1957 struct mem_cgroup *root = sc->target_mem_cgroup;
1958 struct mem_cgroup_reclaim_cookie reclaim = {
1960 .priority = sc->priority,
1962 struct mem_cgroup *memcg;
1964 nr_reclaimed = sc->nr_reclaimed;
1965 nr_scanned = sc->nr_scanned;
1967 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1969 struct lruvec *lruvec;
1971 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1973 shrink_lruvec(lruvec, sc);
1976 * Limit reclaim has historically picked one
1977 * memcg and scanned it with decreasing
1978 * priority levels until nr_to_reclaim had
1979 * been reclaimed. This priority cycle is
1980 * thus over after a single memcg.
1982 * Direct reclaim and kswapd, on the other
1983 * hand, have to scan all memory cgroups to
1984 * fulfill the overall scan target for the
1987 if (!global_reclaim(sc)) {
1988 mem_cgroup_iter_break(root, memcg);
1991 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1993 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
1994 sc->nr_scanned - nr_scanned, sc));
1997 /* Returns true if compaction should go ahead for a high-order request */
1998 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2000 unsigned long balance_gap, watermark;
2003 /* Do not consider compaction for orders reclaim is meant to satisfy */
2004 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2008 * Compaction takes time to run and there are potentially other
2009 * callers using the pages just freed. Continue reclaiming until
2010 * there is a buffer of free pages available to give compaction
2011 * a reasonable chance of completing and allocating the page
2013 balance_gap = min(low_wmark_pages(zone),
2014 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2015 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2016 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2017 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2020 * If compaction is deferred, reclaim up to a point where
2021 * compaction will have a chance of success when re-enabled
2023 if (compaction_deferred(zone, sc->order))
2024 return watermark_ok;
2026 /* If compaction is not ready to start, keep reclaiming */
2027 if (!compaction_suitable(zone, sc->order))
2030 return watermark_ok;
2034 * This is the direct reclaim path, for page-allocating processes. We only
2035 * try to reclaim pages from zones which will satisfy the caller's allocation
2038 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2040 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2042 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2043 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2044 * zone defense algorithm.
2046 * If a zone is deemed to be full of pinned pages then just give it a light
2047 * scan then give up on it.
2049 * This function returns true if a zone is being reclaimed for a costly
2050 * high-order allocation and compaction is ready to begin. This indicates to
2051 * the caller that it should consider retrying the allocation instead of
2054 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2058 unsigned long nr_soft_reclaimed;
2059 unsigned long nr_soft_scanned;
2060 bool aborted_reclaim = false;
2063 * If the number of buffer_heads in the machine exceeds the maximum
2064 * allowed level, force direct reclaim to scan the highmem zone as
2065 * highmem pages could be pinning lowmem pages storing buffer_heads
2067 if (buffer_heads_over_limit)
2068 sc->gfp_mask |= __GFP_HIGHMEM;
2070 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2071 gfp_zone(sc->gfp_mask), sc->nodemask) {
2072 if (!populated_zone(zone))
2075 * Take care memory controller reclaiming has small influence
2078 if (global_reclaim(sc)) {
2079 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2081 if (zone->all_unreclaimable &&
2082 sc->priority != DEF_PRIORITY)
2083 continue; /* Let kswapd poll it */
2084 if (IS_ENABLED(CONFIG_COMPACTION)) {
2086 * If we already have plenty of memory free for
2087 * compaction in this zone, don't free any more.
2088 * Even though compaction is invoked for any
2089 * non-zero order, only frequent costly order
2090 * reclamation is disruptive enough to become a
2091 * noticeable problem, like transparent huge
2094 if (compaction_ready(zone, sc)) {
2095 aborted_reclaim = true;
2100 * This steals pages from memory cgroups over softlimit
2101 * and returns the number of reclaimed pages and
2102 * scanned pages. This works for global memory pressure
2103 * and balancing, not for a memcg's limit.
2105 nr_soft_scanned = 0;
2106 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2107 sc->order, sc->gfp_mask,
2109 sc->nr_reclaimed += nr_soft_reclaimed;
2110 sc->nr_scanned += nr_soft_scanned;
2111 /* need some check for avoid more shrink_zone() */
2114 shrink_zone(zone, sc);
2117 return aborted_reclaim;
2120 static bool zone_reclaimable(struct zone *zone)
2122 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2125 /* All zones in zonelist are unreclaimable? */
2126 static bool all_unreclaimable(struct zonelist *zonelist,
2127 struct scan_control *sc)
2132 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2133 gfp_zone(sc->gfp_mask), sc->nodemask) {
2134 if (!populated_zone(zone))
2136 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2138 if (!zone->all_unreclaimable)
2146 * This is the main entry point to direct page reclaim.
2148 * If a full scan of the inactive list fails to free enough memory then we
2149 * are "out of memory" and something needs to be killed.
2151 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2152 * high - the zone may be full of dirty or under-writeback pages, which this
2153 * caller can't do much about. We kick the writeback threads and take explicit
2154 * naps in the hope that some of these pages can be written. But if the
2155 * allocating task holds filesystem locks which prevent writeout this might not
2156 * work, and the allocation attempt will fail.
2158 * returns: 0, if no pages reclaimed
2159 * else, the number of pages reclaimed
2161 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2162 struct scan_control *sc,
2163 struct shrink_control *shrink)
2165 unsigned long total_scanned = 0;
2166 struct reclaim_state *reclaim_state = current->reclaim_state;
2169 unsigned long writeback_threshold;
2170 bool aborted_reclaim;
2172 delayacct_freepages_start();
2174 if (global_reclaim(sc))
2175 count_vm_event(ALLOCSTALL);
2179 aborted_reclaim = shrink_zones(zonelist, sc);
2182 * Don't shrink slabs when reclaiming memory from
2183 * over limit cgroups
2185 if (global_reclaim(sc)) {
2186 unsigned long lru_pages = 0;
2187 for_each_zone_zonelist(zone, z, zonelist,
2188 gfp_zone(sc->gfp_mask)) {
2189 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2192 lru_pages += zone_reclaimable_pages(zone);
2195 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2196 if (reclaim_state) {
2197 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2198 reclaim_state->reclaimed_slab = 0;
2201 total_scanned += sc->nr_scanned;
2202 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2206 * Try to write back as many pages as we just scanned. This
2207 * tends to cause slow streaming writers to write data to the
2208 * disk smoothly, at the dirtying rate, which is nice. But
2209 * that's undesirable in laptop mode, where we *want* lumpy
2210 * writeout. So in laptop mode, write out the whole world.
2212 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2213 if (total_scanned > writeback_threshold) {
2214 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2215 WB_REASON_TRY_TO_FREE_PAGES);
2216 sc->may_writepage = 1;
2219 /* Take a nap, wait for some writeback to complete */
2220 if (!sc->hibernation_mode && sc->nr_scanned &&
2221 sc->priority < DEF_PRIORITY - 2) {
2222 struct zone *preferred_zone;
2224 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2225 &cpuset_current_mems_allowed,
2227 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2229 } while (--sc->priority >= 0);
2232 delayacct_freepages_end();
2234 if (sc->nr_reclaimed)
2235 return sc->nr_reclaimed;
2238 * As hibernation is going on, kswapd is freezed so that it can't mark
2239 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2242 if (oom_killer_disabled)
2245 /* Aborted reclaim to try compaction? don't OOM, then */
2246 if (aborted_reclaim)
2249 /* top priority shrink_zones still had more to do? don't OOM, then */
2250 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2256 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2259 unsigned long pfmemalloc_reserve = 0;
2260 unsigned long free_pages = 0;
2264 for (i = 0; i <= ZONE_NORMAL; i++) {
2265 zone = &pgdat->node_zones[i];
2266 pfmemalloc_reserve += min_wmark_pages(zone);
2267 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2270 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2272 /* kswapd must be awake if processes are being throttled */
2273 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2274 pgdat->classzone_idx = min(pgdat->classzone_idx,
2275 (enum zone_type)ZONE_NORMAL);
2276 wake_up_interruptible(&pgdat->kswapd_wait);
2283 * Throttle direct reclaimers if backing storage is backed by the network
2284 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2285 * depleted. kswapd will continue to make progress and wake the processes
2286 * when the low watermark is reached.
2288 * Returns true if a fatal signal was delivered during throttling. If this
2289 * happens, the page allocator should not consider triggering the OOM killer.
2291 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2292 nodemask_t *nodemask)
2295 int high_zoneidx = gfp_zone(gfp_mask);
2299 * Kernel threads should not be throttled as they may be indirectly
2300 * responsible for cleaning pages necessary for reclaim to make forward
2301 * progress. kjournald for example may enter direct reclaim while
2302 * committing a transaction where throttling it could forcing other
2303 * processes to block on log_wait_commit().
2305 if (current->flags & PF_KTHREAD)
2309 * If a fatal signal is pending, this process should not throttle.
2310 * It should return quickly so it can exit and free its memory
2312 if (fatal_signal_pending(current))
2315 /* Check if the pfmemalloc reserves are ok */
2316 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2317 pgdat = zone->zone_pgdat;
2318 if (pfmemalloc_watermark_ok(pgdat))
2321 /* Account for the throttling */
2322 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2325 * If the caller cannot enter the filesystem, it's possible that it
2326 * is due to the caller holding an FS lock or performing a journal
2327 * transaction in the case of a filesystem like ext[3|4]. In this case,
2328 * it is not safe to block on pfmemalloc_wait as kswapd could be
2329 * blocked waiting on the same lock. Instead, throttle for up to a
2330 * second before continuing.
2332 if (!(gfp_mask & __GFP_FS)) {
2333 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2334 pfmemalloc_watermark_ok(pgdat), HZ);
2339 /* Throttle until kswapd wakes the process */
2340 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2341 pfmemalloc_watermark_ok(pgdat));
2344 if (fatal_signal_pending(current))
2351 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2352 gfp_t gfp_mask, nodemask_t *nodemask)
2354 unsigned long nr_reclaimed;
2355 struct scan_control sc = {
2356 .gfp_mask = gfp_mask,
2357 .may_writepage = !laptop_mode,
2358 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2362 .priority = DEF_PRIORITY,
2363 .target_mem_cgroup = NULL,
2364 .nodemask = nodemask,
2366 struct shrink_control shrink = {
2367 .gfp_mask = sc.gfp_mask,
2371 * Do not enter reclaim if fatal signal was delivered while throttled.
2372 * 1 is returned so that the page allocator does not OOM kill at this
2375 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2378 trace_mm_vmscan_direct_reclaim_begin(order,
2382 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2384 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2386 return nr_reclaimed;
2391 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2392 gfp_t gfp_mask, bool noswap,
2394 unsigned long *nr_scanned)
2396 struct scan_control sc = {
2398 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2399 .may_writepage = !laptop_mode,
2401 .may_swap = !noswap,
2404 .target_mem_cgroup = memcg,
2406 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2408 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2409 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2411 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2416 * NOTE: Although we can get the priority field, using it
2417 * here is not a good idea, since it limits the pages we can scan.
2418 * if we don't reclaim here, the shrink_zone from balance_pgdat
2419 * will pick up pages from other mem cgroup's as well. We hack
2420 * the priority and make it zero.
2422 shrink_lruvec(lruvec, &sc);
2424 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2426 *nr_scanned = sc.nr_scanned;
2427 return sc.nr_reclaimed;
2430 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2434 struct zonelist *zonelist;
2435 unsigned long nr_reclaimed;
2437 struct scan_control sc = {
2438 .may_writepage = !laptop_mode,
2440 .may_swap = !noswap,
2441 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2443 .priority = DEF_PRIORITY,
2444 .target_mem_cgroup = memcg,
2445 .nodemask = NULL, /* we don't care the placement */
2446 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2447 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2449 struct shrink_control shrink = {
2450 .gfp_mask = sc.gfp_mask,
2454 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2455 * take care of from where we get pages. So the node where we start the
2456 * scan does not need to be the current node.
2458 nid = mem_cgroup_select_victim_node(memcg);
2460 zonelist = NODE_DATA(nid)->node_zonelists;
2462 trace_mm_vmscan_memcg_reclaim_begin(0,
2466 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2468 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2470 return nr_reclaimed;
2474 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2476 struct mem_cgroup *memcg;
2478 if (!total_swap_pages)
2481 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2483 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2485 if (inactive_anon_is_low(lruvec))
2486 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2487 sc, LRU_ACTIVE_ANON);
2489 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2493 static bool zone_balanced(struct zone *zone, int order,
2494 unsigned long balance_gap, int classzone_idx)
2496 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2497 balance_gap, classzone_idx, 0))
2500 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2501 !compaction_suitable(zone, order))
2508 * pgdat_balanced() is used when checking if a node is balanced.
2510 * For order-0, all zones must be balanced!
2512 * For high-order allocations only zones that meet watermarks and are in a
2513 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2514 * total of balanced pages must be at least 25% of the zones allowed by
2515 * classzone_idx for the node to be considered balanced. Forcing all zones to
2516 * be balanced for high orders can cause excessive reclaim when there are
2518 * The choice of 25% is due to
2519 * o a 16M DMA zone that is balanced will not balance a zone on any
2520 * reasonable sized machine
2521 * o On all other machines, the top zone must be at least a reasonable
2522 * percentage of the middle zones. For example, on 32-bit x86, highmem
2523 * would need to be at least 256M for it to be balance a whole node.
2524 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2525 * to balance a node on its own. These seemed like reasonable ratios.
2527 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2529 unsigned long present_pages = 0;
2530 unsigned long balanced_pages = 0;
2533 /* Check the watermark levels */
2534 for (i = 0; i <= classzone_idx; i++) {
2535 struct zone *zone = pgdat->node_zones + i;
2537 if (!populated_zone(zone))
2540 present_pages += zone->present_pages;
2543 * A special case here:
2545 * balance_pgdat() skips over all_unreclaimable after
2546 * DEF_PRIORITY. Effectively, it considers them balanced so
2547 * they must be considered balanced here as well!
2549 if (zone->all_unreclaimable) {
2550 balanced_pages += zone->present_pages;
2554 if (zone_balanced(zone, order, 0, i))
2555 balanced_pages += zone->present_pages;
2561 return balanced_pages >= (present_pages >> 2);
2567 * Prepare kswapd for sleeping. This verifies that there are no processes
2568 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2570 * Returns true if kswapd is ready to sleep
2572 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2575 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2580 * There is a potential race between when kswapd checks its watermarks
2581 * and a process gets throttled. There is also a potential race if
2582 * processes get throttled, kswapd wakes, a large process exits therby
2583 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2584 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2585 * so wake them now if necessary. If necessary, processes will wake
2586 * kswapd and get throttled again
2588 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2589 wake_up(&pgdat->pfmemalloc_wait);
2593 return pgdat_balanced(pgdat, order, classzone_idx);
2597 * For kswapd, balance_pgdat() will work across all this node's zones until
2598 * they are all at high_wmark_pages(zone).
2600 * Returns the final order kswapd was reclaiming at
2602 * There is special handling here for zones which are full of pinned pages.
2603 * This can happen if the pages are all mlocked, or if they are all used by
2604 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2605 * What we do is to detect the case where all pages in the zone have been
2606 * scanned twice and there has been zero successful reclaim. Mark the zone as
2607 * dead and from now on, only perform a short scan. Basically we're polling
2608 * the zone for when the problem goes away.
2610 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2611 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2612 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2613 * lower zones regardless of the number of free pages in the lower zones. This
2614 * interoperates with the page allocator fallback scheme to ensure that aging
2615 * of pages is balanced across the zones.
2617 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2620 struct zone *unbalanced_zone;
2622 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2623 unsigned long total_scanned;
2624 struct reclaim_state *reclaim_state = current->reclaim_state;
2625 unsigned long nr_soft_reclaimed;
2626 unsigned long nr_soft_scanned;
2627 struct scan_control sc = {
2628 .gfp_mask = GFP_KERNEL,
2632 * kswapd doesn't want to be bailed out while reclaim. because
2633 * we want to put equal scanning pressure on each zone.
2635 .nr_to_reclaim = ULONG_MAX,
2637 .target_mem_cgroup = NULL,
2639 struct shrink_control shrink = {
2640 .gfp_mask = sc.gfp_mask,
2644 sc.priority = DEF_PRIORITY;
2645 sc.nr_reclaimed = 0;
2646 sc.may_writepage = !laptop_mode;
2647 count_vm_event(PAGEOUTRUN);
2650 unsigned long lru_pages = 0;
2651 int has_under_min_watermark_zone = 0;
2653 unbalanced_zone = NULL;
2656 * Scan in the highmem->dma direction for the highest
2657 * zone which needs scanning
2659 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2660 struct zone *zone = pgdat->node_zones + i;
2662 if (!populated_zone(zone))
2665 if (zone->all_unreclaimable &&
2666 sc.priority != DEF_PRIORITY)
2670 * Do some background aging of the anon list, to give
2671 * pages a chance to be referenced before reclaiming.
2673 age_active_anon(zone, &sc);
2676 * If the number of buffer_heads in the machine
2677 * exceeds the maximum allowed level and this node
2678 * has a highmem zone, force kswapd to reclaim from
2679 * it to relieve lowmem pressure.
2681 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2686 if (!zone_balanced(zone, order, 0, 0)) {
2690 /* If balanced, clear the congested flag */
2691 zone_clear_flag(zone, ZONE_CONGESTED);
2697 for (i = 0; i <= end_zone; i++) {
2698 struct zone *zone = pgdat->node_zones + i;
2700 lru_pages += zone_reclaimable_pages(zone);
2704 * Now scan the zone in the dma->highmem direction, stopping
2705 * at the last zone which needs scanning.
2707 * We do this because the page allocator works in the opposite
2708 * direction. This prevents the page allocator from allocating
2709 * pages behind kswapd's direction of progress, which would
2710 * cause too much scanning of the lower zones.
2712 for (i = 0; i <= end_zone; i++) {
2713 struct zone *zone = pgdat->node_zones + i;
2714 int nr_slab, testorder;
2715 unsigned long balance_gap;
2717 if (!populated_zone(zone))
2720 if (zone->all_unreclaimable &&
2721 sc.priority != DEF_PRIORITY)
2726 nr_soft_scanned = 0;
2728 * Call soft limit reclaim before calling shrink_zone.
2730 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2733 sc.nr_reclaimed += nr_soft_reclaimed;
2734 total_scanned += nr_soft_scanned;
2737 * We put equal pressure on every zone, unless
2738 * one zone has way too many pages free
2739 * already. The "too many pages" is defined
2740 * as the high wmark plus a "gap" where the
2741 * gap is either the low watermark or 1%
2742 * of the zone, whichever is smaller.
2744 balance_gap = min(low_wmark_pages(zone),
2745 (zone->present_pages +
2746 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2747 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2749 * Kswapd reclaims only single pages with compaction
2750 * enabled. Trying too hard to reclaim until contiguous
2751 * free pages have become available can hurt performance
2752 * by evicting too much useful data from memory.
2753 * Do not reclaim more than needed for compaction.
2756 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2757 compaction_suitable(zone, order) !=
2761 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2762 !zone_balanced(zone, testorder,
2763 balance_gap, end_zone)) {
2764 shrink_zone(zone, &sc);
2766 reclaim_state->reclaimed_slab = 0;
2767 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2768 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2769 total_scanned += sc.nr_scanned;
2771 if (nr_slab == 0 && !zone_reclaimable(zone))
2772 zone->all_unreclaimable = 1;
2776 * If we've done a decent amount of scanning and
2777 * the reclaim ratio is low, start doing writepage
2778 * even in laptop mode
2780 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2781 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2782 sc.may_writepage = 1;
2784 if (zone->all_unreclaimable) {
2785 if (end_zone && end_zone == i)
2790 if (!zone_balanced(zone, testorder, 0, end_zone)) {
2791 unbalanced_zone = zone;
2793 * We are still under min water mark. This
2794 * means that we have a GFP_ATOMIC allocation
2795 * failure risk. Hurry up!
2797 if (!zone_watermark_ok_safe(zone, order,
2798 min_wmark_pages(zone), end_zone, 0))
2799 has_under_min_watermark_zone = 1;
2802 * If a zone reaches its high watermark,
2803 * consider it to be no longer congested. It's
2804 * possible there are dirty pages backed by
2805 * congested BDIs but as pressure is relieved,
2806 * speculatively avoid congestion waits
2808 zone_clear_flag(zone, ZONE_CONGESTED);
2814 * If the low watermark is met there is no need for processes
2815 * to be throttled on pfmemalloc_wait as they should not be
2816 * able to safely make forward progress. Wake them
2818 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2819 pfmemalloc_watermark_ok(pgdat))
2820 wake_up(&pgdat->pfmemalloc_wait);
2822 if (pgdat_balanced(pgdat, order, *classzone_idx))
2823 break; /* kswapd: all done */
2825 * OK, kswapd is getting into trouble. Take a nap, then take
2826 * another pass across the zones.
2828 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2829 if (has_under_min_watermark_zone)
2830 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2831 else if (unbalanced_zone)
2832 wait_iff_congested(unbalanced_zone, BLK_RW_ASYNC, HZ/10);
2836 * We do this so kswapd doesn't build up large priorities for
2837 * example when it is freeing in parallel with allocators. It
2838 * matches the direct reclaim path behaviour in terms of impact
2839 * on zone->*_priority.
2841 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2843 } while (--sc.priority >= 0);
2846 if (!pgdat_balanced(pgdat, order, *classzone_idx)) {
2852 * Fragmentation may mean that the system cannot be
2853 * rebalanced for high-order allocations in all zones.
2854 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2855 * it means the zones have been fully scanned and are still
2856 * not balanced. For high-order allocations, there is
2857 * little point trying all over again as kswapd may
2860 * Instead, recheck all watermarks at order-0 as they
2861 * are the most important. If watermarks are ok, kswapd will go
2862 * back to sleep. High-order users can still perform direct
2863 * reclaim if they wish.
2865 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2866 order = sc.order = 0;
2872 * If kswapd was reclaiming at a higher order, it has the option of
2873 * sleeping without all zones being balanced. Before it does, it must
2874 * ensure that the watermarks for order-0 on *all* zones are met and
2875 * that the congestion flags are cleared. The congestion flag must
2876 * be cleared as kswapd is the only mechanism that clears the flag
2877 * and it is potentially going to sleep here.
2880 int zones_need_compaction = 1;
2882 for (i = 0; i <= end_zone; i++) {
2883 struct zone *zone = pgdat->node_zones + i;
2885 if (!populated_zone(zone))
2888 /* Check if the memory needs to be defragmented. */
2889 if (zone_watermark_ok(zone, order,
2890 low_wmark_pages(zone), *classzone_idx, 0))
2891 zones_need_compaction = 0;
2894 if (zones_need_compaction)
2895 compact_pgdat(pgdat, order);
2899 * Return the order we were reclaiming at so prepare_kswapd_sleep()
2900 * makes a decision on the order we were last reclaiming at. However,
2901 * if another caller entered the allocator slow path while kswapd
2902 * was awake, order will remain at the higher level
2904 *classzone_idx = end_zone;
2908 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2913 if (freezing(current) || kthread_should_stop())
2916 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2918 /* Try to sleep for a short interval */
2919 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2920 remaining = schedule_timeout(HZ/10);
2921 finish_wait(&pgdat->kswapd_wait, &wait);
2922 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2926 * After a short sleep, check if it was a premature sleep. If not, then
2927 * go fully to sleep until explicitly woken up.
2929 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2930 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2933 * vmstat counters are not perfectly accurate and the estimated
2934 * value for counters such as NR_FREE_PAGES can deviate from the
2935 * true value by nr_online_cpus * threshold. To avoid the zone
2936 * watermarks being breached while under pressure, we reduce the
2937 * per-cpu vmstat threshold while kswapd is awake and restore
2938 * them before going back to sleep.
2940 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2943 * Compaction records what page blocks it recently failed to
2944 * isolate pages from and skips them in the future scanning.
2945 * When kswapd is going to sleep, it is reasonable to assume
2946 * that pages and compaction may succeed so reset the cache.
2948 reset_isolation_suitable(pgdat);
2950 if (!kthread_should_stop())
2953 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2956 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2958 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2960 finish_wait(&pgdat->kswapd_wait, &wait);
2964 * The background pageout daemon, started as a kernel thread
2965 * from the init process.
2967 * This basically trickles out pages so that we have _some_
2968 * free memory available even if there is no other activity
2969 * that frees anything up. This is needed for things like routing
2970 * etc, where we otherwise might have all activity going on in
2971 * asynchronous contexts that cannot page things out.
2973 * If there are applications that are active memory-allocators
2974 * (most normal use), this basically shouldn't matter.
2976 static int kswapd(void *p)
2978 unsigned long order, new_order;
2979 unsigned balanced_order;
2980 int classzone_idx, new_classzone_idx;
2981 int balanced_classzone_idx;
2982 pg_data_t *pgdat = (pg_data_t*)p;
2983 struct task_struct *tsk = current;
2985 struct reclaim_state reclaim_state = {
2986 .reclaimed_slab = 0,
2988 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2990 lockdep_set_current_reclaim_state(GFP_KERNEL);
2992 if (!cpumask_empty(cpumask))
2993 set_cpus_allowed_ptr(tsk, cpumask);
2994 current->reclaim_state = &reclaim_state;
2997 * Tell the memory management that we're a "memory allocator",
2998 * and that if we need more memory we should get access to it
2999 * regardless (see "__alloc_pages()"). "kswapd" should
3000 * never get caught in the normal page freeing logic.
3002 * (Kswapd normally doesn't need memory anyway, but sometimes
3003 * you need a small amount of memory in order to be able to
3004 * page out something else, and this flag essentially protects
3005 * us from recursively trying to free more memory as we're
3006 * trying to free the first piece of memory in the first place).
3008 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3011 order = new_order = 0;
3013 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3014 balanced_classzone_idx = classzone_idx;
3019 * If the last balance_pgdat was unsuccessful it's unlikely a
3020 * new request of a similar or harder type will succeed soon
3021 * so consider going to sleep on the basis we reclaimed at
3023 if (balanced_classzone_idx >= new_classzone_idx &&
3024 balanced_order == new_order) {
3025 new_order = pgdat->kswapd_max_order;
3026 new_classzone_idx = pgdat->classzone_idx;
3027 pgdat->kswapd_max_order = 0;
3028 pgdat->classzone_idx = pgdat->nr_zones - 1;
3031 if (order < new_order || classzone_idx > new_classzone_idx) {
3033 * Don't sleep if someone wants a larger 'order'
3034 * allocation or has tigher zone constraints
3037 classzone_idx = new_classzone_idx;
3039 kswapd_try_to_sleep(pgdat, balanced_order,
3040 balanced_classzone_idx);
3041 order = pgdat->kswapd_max_order;
3042 classzone_idx = pgdat->classzone_idx;
3044 new_classzone_idx = classzone_idx;
3045 pgdat->kswapd_max_order = 0;
3046 pgdat->classzone_idx = pgdat->nr_zones - 1;
3049 ret = try_to_freeze();
3050 if (kthread_should_stop())
3054 * We can speed up thawing tasks if we don't call balance_pgdat
3055 * after returning from the refrigerator
3058 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3059 balanced_classzone_idx = classzone_idx;
3060 balanced_order = balance_pgdat(pgdat, order,
3061 &balanced_classzone_idx);
3065 current->reclaim_state = NULL;
3070 * A zone is low on free memory, so wake its kswapd task to service it.
3072 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3076 if (!populated_zone(zone))
3079 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3081 pgdat = zone->zone_pgdat;
3082 if (pgdat->kswapd_max_order < order) {
3083 pgdat->kswapd_max_order = order;
3084 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3086 if (!waitqueue_active(&pgdat->kswapd_wait))
3088 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3091 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3092 wake_up_interruptible(&pgdat->kswapd_wait);
3096 * The reclaimable count would be mostly accurate.
3097 * The less reclaimable pages may be
3098 * - mlocked pages, which will be moved to unevictable list when encountered
3099 * - mapped pages, which may require several travels to be reclaimed
3100 * - dirty pages, which is not "instantly" reclaimable
3102 unsigned long global_reclaimable_pages(void)
3106 nr = global_page_state(NR_ACTIVE_FILE) +
3107 global_page_state(NR_INACTIVE_FILE);
3109 if (nr_swap_pages > 0)
3110 nr += global_page_state(NR_ACTIVE_ANON) +
3111 global_page_state(NR_INACTIVE_ANON);
3116 unsigned long zone_reclaimable_pages(struct zone *zone)
3120 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3121 zone_page_state(zone, NR_INACTIVE_FILE);
3123 if (nr_swap_pages > 0)
3124 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3125 zone_page_state(zone, NR_INACTIVE_ANON);
3130 #ifdef CONFIG_HIBERNATION
3132 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3135 * Rather than trying to age LRUs the aim is to preserve the overall
3136 * LRU order by reclaiming preferentially
3137 * inactive > active > active referenced > active mapped
3139 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3141 struct reclaim_state reclaim_state;
3142 struct scan_control sc = {
3143 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3147 .nr_to_reclaim = nr_to_reclaim,
3148 .hibernation_mode = 1,
3150 .priority = DEF_PRIORITY,
3152 struct shrink_control shrink = {
3153 .gfp_mask = sc.gfp_mask,
3155 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3156 struct task_struct *p = current;
3157 unsigned long nr_reclaimed;
3159 p->flags |= PF_MEMALLOC;
3160 lockdep_set_current_reclaim_state(sc.gfp_mask);
3161 reclaim_state.reclaimed_slab = 0;
3162 p->reclaim_state = &reclaim_state;
3164 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3166 p->reclaim_state = NULL;
3167 lockdep_clear_current_reclaim_state();
3168 p->flags &= ~PF_MEMALLOC;
3170 return nr_reclaimed;
3172 #endif /* CONFIG_HIBERNATION */
3174 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3175 not required for correctness. So if the last cpu in a node goes
3176 away, we get changed to run anywhere: as the first one comes back,
3177 restore their cpu bindings. */
3178 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3183 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3184 for_each_node_state(nid, N_MEMORY) {
3185 pg_data_t *pgdat = NODE_DATA(nid);
3186 const struct cpumask *mask;
3188 mask = cpumask_of_node(pgdat->node_id);
3190 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3191 /* One of our CPUs online: restore mask */
3192 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3199 * This kswapd start function will be called by init and node-hot-add.
3200 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3202 int kswapd_run(int nid)
3204 pg_data_t *pgdat = NODE_DATA(nid);
3210 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3211 if (IS_ERR(pgdat->kswapd)) {
3212 /* failure at boot is fatal */
3213 BUG_ON(system_state == SYSTEM_BOOTING);
3214 pgdat->kswapd = NULL;
3215 pr_err("Failed to start kswapd on node %d\n", nid);
3216 ret = PTR_ERR(pgdat->kswapd);
3222 * Called by memory hotplug when all memory in a node is offlined. Caller must
3223 * hold lock_memory_hotplug().
3225 void kswapd_stop(int nid)
3227 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3230 kthread_stop(kswapd);
3231 NODE_DATA(nid)->kswapd = NULL;
3235 static int __init kswapd_init(void)
3240 for_each_node_state(nid, N_MEMORY)
3242 hotcpu_notifier(cpu_callback, 0);
3246 module_init(kswapd_init)
3252 * If non-zero call zone_reclaim when the number of free pages falls below
3255 int zone_reclaim_mode __read_mostly;
3257 #define RECLAIM_OFF 0
3258 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3259 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3260 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3263 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3264 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3267 #define ZONE_RECLAIM_PRIORITY 4
3270 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3273 int sysctl_min_unmapped_ratio = 1;
3276 * If the number of slab pages in a zone grows beyond this percentage then
3277 * slab reclaim needs to occur.
3279 int sysctl_min_slab_ratio = 5;
3281 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3283 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3284 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3285 zone_page_state(zone, NR_ACTIVE_FILE);
3288 * It's possible for there to be more file mapped pages than
3289 * accounted for by the pages on the file LRU lists because
3290 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3292 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3295 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3296 static long zone_pagecache_reclaimable(struct zone *zone)
3298 long nr_pagecache_reclaimable;
3302 * If RECLAIM_SWAP is set, then all file pages are considered
3303 * potentially reclaimable. Otherwise, we have to worry about
3304 * pages like swapcache and zone_unmapped_file_pages() provides
3307 if (zone_reclaim_mode & RECLAIM_SWAP)
3308 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3310 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3312 /* If we can't clean pages, remove dirty pages from consideration */
3313 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3314 delta += zone_page_state(zone, NR_FILE_DIRTY);
3316 /* Watch for any possible underflows due to delta */
3317 if (unlikely(delta > nr_pagecache_reclaimable))
3318 delta = nr_pagecache_reclaimable;
3320 return nr_pagecache_reclaimable - delta;
3324 * Try to free up some pages from this zone through reclaim.
3326 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3328 /* Minimum pages needed in order to stay on node */
3329 const unsigned long nr_pages = 1 << order;
3330 struct task_struct *p = current;
3331 struct reclaim_state reclaim_state;
3332 struct scan_control sc = {
3333 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3334 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3336 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3337 .gfp_mask = gfp_mask,
3339 .priority = ZONE_RECLAIM_PRIORITY,
3341 struct shrink_control shrink = {
3342 .gfp_mask = sc.gfp_mask,
3344 unsigned long nr_slab_pages0, nr_slab_pages1;
3348 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3349 * and we also need to be able to write out pages for RECLAIM_WRITE
3352 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3353 lockdep_set_current_reclaim_state(gfp_mask);
3354 reclaim_state.reclaimed_slab = 0;
3355 p->reclaim_state = &reclaim_state;
3357 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3359 * Free memory by calling shrink zone with increasing
3360 * priorities until we have enough memory freed.
3363 shrink_zone(zone, &sc);
3364 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3367 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3368 if (nr_slab_pages0 > zone->min_slab_pages) {
3370 * shrink_slab() does not currently allow us to determine how
3371 * many pages were freed in this zone. So we take the current
3372 * number of slab pages and shake the slab until it is reduced
3373 * by the same nr_pages that we used for reclaiming unmapped
3376 * Note that shrink_slab will free memory on all zones and may
3380 unsigned long lru_pages = zone_reclaimable_pages(zone);
3382 /* No reclaimable slab or very low memory pressure */
3383 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3386 /* Freed enough memory */
3387 nr_slab_pages1 = zone_page_state(zone,
3388 NR_SLAB_RECLAIMABLE);
3389 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3394 * Update nr_reclaimed by the number of slab pages we
3395 * reclaimed from this zone.
3397 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3398 if (nr_slab_pages1 < nr_slab_pages0)
3399 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3402 p->reclaim_state = NULL;
3403 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3404 lockdep_clear_current_reclaim_state();
3405 return sc.nr_reclaimed >= nr_pages;
3408 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3414 * Zone reclaim reclaims unmapped file backed pages and
3415 * slab pages if we are over the defined limits.
3417 * A small portion of unmapped file backed pages is needed for
3418 * file I/O otherwise pages read by file I/O will be immediately
3419 * thrown out if the zone is overallocated. So we do not reclaim
3420 * if less than a specified percentage of the zone is used by
3421 * unmapped file backed pages.
3423 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3424 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3425 return ZONE_RECLAIM_FULL;
3427 if (zone->all_unreclaimable)
3428 return ZONE_RECLAIM_FULL;
3431 * Do not scan if the allocation should not be delayed.
3433 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3434 return ZONE_RECLAIM_NOSCAN;
3437 * Only run zone reclaim on the local zone or on zones that do not
3438 * have associated processors. This will favor the local processor
3439 * over remote processors and spread off node memory allocations
3440 * as wide as possible.
3442 node_id = zone_to_nid(zone);
3443 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3444 return ZONE_RECLAIM_NOSCAN;
3446 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3447 return ZONE_RECLAIM_NOSCAN;
3449 ret = __zone_reclaim(zone, gfp_mask, order);
3450 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3453 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3460 * page_evictable - test whether a page is evictable
3461 * @page: the page to test
3463 * Test whether page is evictable--i.e., should be placed on active/inactive
3464 * lists vs unevictable list.
3466 * Reasons page might not be evictable:
3467 * (1) page's mapping marked unevictable
3468 * (2) page is part of an mlocked VMA
3471 int page_evictable(struct page *page)
3473 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3478 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3479 * @pages: array of pages to check
3480 * @nr_pages: number of pages to check
3482 * Checks pages for evictability and moves them to the appropriate lru list.
3484 * This function is only used for SysV IPC SHM_UNLOCK.
3486 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3488 struct lruvec *lruvec;
3489 struct zone *zone = NULL;
3494 for (i = 0; i < nr_pages; i++) {
3495 struct page *page = pages[i];
3496 struct zone *pagezone;
3499 pagezone = page_zone(page);
3500 if (pagezone != zone) {
3502 spin_unlock_irq(&zone->lru_lock);
3504 spin_lock_irq(&zone->lru_lock);
3506 lruvec = mem_cgroup_page_lruvec(page, zone);
3508 if (!PageLRU(page) || !PageUnevictable(page))
3511 if (page_evictable(page)) {
3512 enum lru_list lru = page_lru_base_type(page);
3514 VM_BUG_ON(PageActive(page));
3515 ClearPageUnevictable(page);
3516 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3517 add_page_to_lru_list(page, lruvec, lru);
3523 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3524 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3525 spin_unlock_irq(&zone->lru_lock);
3528 #endif /* CONFIG_SHMEM */
3530 static void warn_scan_unevictable_pages(void)
3532 printk_once(KERN_WARNING
3533 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3534 "disabled for lack of a legitimate use case. If you have "
3535 "one, please send an email to linux-mm@kvack.org.\n",
3540 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3541 * all nodes' unevictable lists for evictable pages
3543 unsigned long scan_unevictable_pages;
3545 int scan_unevictable_handler(struct ctl_table *table, int write,
3546 void __user *buffer,
3547 size_t *length, loff_t *ppos)
3549 warn_scan_unevictable_pages();
3550 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3551 scan_unevictable_pages = 0;
3557 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3558 * a specified node's per zone unevictable lists for evictable pages.
3561 static ssize_t read_scan_unevictable_node(struct device *dev,
3562 struct device_attribute *attr,
3565 warn_scan_unevictable_pages();
3566 return sprintf(buf, "0\n"); /* always zero; should fit... */
3569 static ssize_t write_scan_unevictable_node(struct device *dev,
3570 struct device_attribute *attr,
3571 const char *buf, size_t count)
3573 warn_scan_unevictable_pages();
3578 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3579 read_scan_unevictable_node,
3580 write_scan_unevictable_node);
3582 int scan_unevictable_register_node(struct node *node)
3584 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3587 void scan_unevictable_unregister_node(struct node *node)
3589 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);