hwmon: (pmbus/adm1275) Add support for second current limit
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / vmscan.c
blobb55699cd9067c31721866bb3ef4c1660a1177f06
1 /*
2 * linux/mm/vmscan.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #include <linux/mm.h>
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/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
52 #include "internal.h"
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66 * order-0 pages and then compact the zone
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
75 struct scan_control {
76 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
79 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
82 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
85 unsigned long hibernation_mode;
87 /* This context's GFP mask */
88 gfp_t gfp_mask;
90 int may_writepage;
92 /* Can mapped pages be reclaimed? */
93 int may_unmap;
95 /* Can pages be swapped as part of reclaim? */
96 int may_swap;
98 int order;
101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
104 reclaim_mode_t reclaim_mode;
106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
111 * are scanned.
113 nodemask_t *nodemask;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
120 do { \
121 if ((_page)->lru.prev != _base) { \
122 struct page *prev; \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
127 } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
134 do { \
135 if ((_page)->lru.prev != _base) { \
136 struct page *prev; \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
141 } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
150 long vm_total_pages; /* The total number of pages which the VM controls */
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
164 if (!scanning_global_lru(sc))
165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167 return &zone->reclaim_stat;
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
173 if (!scanning_global_lru(sc))
174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175 zone_to_nid(zone), zone_idx(zone), BIT(lru));
177 return zone_page_state(zone, NR_LRU_BASE + lru);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker *shrinker)
186 shrinker->nr = 0;
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
191 EXPORT_SYMBOL(register_shrinker);
194 * Remove one
196 void unregister_shrinker(struct shrinker *shrinker)
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(unregister_shrinker);
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control *shrink,
233 unsigned long nr_pages_scanned,
234 unsigned long lru_pages)
236 struct shrinker *shrinker;
237 unsigned long ret = 0;
239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
244 ret = 1;
245 goto out;
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
250 unsigned long total_scan;
251 unsigned long max_pass;
252 int shrink_ret = 0;
253 long nr;
254 long new_nr;
255 long batch_size = shrinker->batch ? shrinker->batch
256 : SHRINK_BATCH;
259 * copy the current shrinker scan count into a local variable
260 * and zero it so that other concurrent shrinker invocations
261 * don't also do this scanning work.
263 do {
264 nr = shrinker->nr;
265 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
267 total_scan = nr;
268 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269 delta = (4 * nr_pages_scanned) / shrinker->seeks;
270 delta *= max_pass;
271 do_div(delta, lru_pages + 1);
272 total_scan += delta;
273 if (total_scan < 0) {
274 printk(KERN_ERR "shrink_slab: %pF negative objects to "
275 "delete nr=%ld\n",
276 shrinker->shrink, total_scan);
277 total_scan = max_pass;
281 * We need to avoid excessive windup on filesystem shrinkers
282 * due to large numbers of GFP_NOFS allocations causing the
283 * shrinkers to return -1 all the time. This results in a large
284 * nr being built up so when a shrink that can do some work
285 * comes along it empties the entire cache due to nr >>>
286 * max_pass. This is bad for sustaining a working set in
287 * memory.
289 * Hence only allow the shrinker to scan the entire cache when
290 * a large delta change is calculated directly.
292 if (delta < max_pass / 4)
293 total_scan = min(total_scan, max_pass / 2);
296 * Avoid risking looping forever due to too large nr value:
297 * never try to free more than twice the estimate number of
298 * freeable entries.
300 if (total_scan > max_pass * 2)
301 total_scan = max_pass * 2;
303 trace_mm_shrink_slab_start(shrinker, shrink, nr,
304 nr_pages_scanned, lru_pages,
305 max_pass, delta, total_scan);
307 while (total_scan >= batch_size) {
308 int nr_before;
310 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311 shrink_ret = do_shrinker_shrink(shrinker, shrink,
312 batch_size);
313 if (shrink_ret == -1)
314 break;
315 if (shrink_ret < nr_before)
316 ret += nr_before - shrink_ret;
317 count_vm_events(SLABS_SCANNED, batch_size);
318 total_scan -= batch_size;
320 cond_resched();
324 * move the unused scan count back into the shrinker in a
325 * manner that handles concurrent updates. If we exhausted the
326 * scan, there is no need to do an update.
328 do {
329 nr = shrinker->nr;
330 new_nr = total_scan + nr;
331 if (total_scan <= 0)
332 break;
333 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337 up_read(&shrinker_rwsem);
338 out:
339 cond_resched();
340 return ret;
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
344 bool sync)
346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
349 * Initially assume we are entering either lumpy reclaim or
350 * reclaim/compaction.Depending on the order, we will either set the
351 * sync mode or just reclaim order-0 pages later.
353 if (COMPACTION_BUILD)
354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
355 else
356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
359 * Avoid using lumpy reclaim or reclaim/compaction if possible by
360 * restricting when its set to either costly allocations or when
361 * under memory pressure
363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364 sc->reclaim_mode |= syncmode;
365 else if (sc->order && priority < DEF_PRIORITY - 2)
366 sc->reclaim_mode |= syncmode;
367 else
368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
371 static void reset_reclaim_mode(struct scan_control *sc)
373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
376 static inline int is_page_cache_freeable(struct page *page)
379 * A freeable page cache page is referenced only by the caller
380 * that isolated the page, the page cache radix tree and
381 * optional buffer heads at page->private.
383 return page_count(page) - page_has_private(page) == 2;
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387 struct scan_control *sc)
389 if (current->flags & PF_SWAPWRITE)
390 return 1;
391 if (!bdi_write_congested(bdi))
392 return 1;
393 if (bdi == current->backing_dev_info)
394 return 1;
396 /* lumpy reclaim for hugepage often need a lot of write */
397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398 return 1;
399 return 0;
403 * We detected a synchronous write error writing a page out. Probably
404 * -ENOSPC. We need to propagate that into the address_space for a subsequent
405 * fsync(), msync() or close().
407 * The tricky part is that after writepage we cannot touch the mapping: nothing
408 * prevents it from being freed up. But we have a ref on the page and once
409 * that page is locked, the mapping is pinned.
411 * We're allowed to run sleeping lock_page() here because we know the caller has
412 * __GFP_FS.
414 static void handle_write_error(struct address_space *mapping,
415 struct page *page, int error)
417 lock_page(page);
418 if (page_mapping(page) == mapping)
419 mapping_set_error(mapping, error);
420 unlock_page(page);
423 /* possible outcome of pageout() */
424 typedef enum {
425 /* failed to write page out, page is locked */
426 PAGE_KEEP,
427 /* move page to the active list, page is locked */
428 PAGE_ACTIVATE,
429 /* page has been sent to the disk successfully, page is unlocked */
430 PAGE_SUCCESS,
431 /* page is clean and locked */
432 PAGE_CLEAN,
433 } pageout_t;
436 * pageout is called by shrink_page_list() for each dirty page.
437 * Calls ->writepage().
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440 struct scan_control *sc)
443 * If the page is dirty, only perform writeback if that write
444 * will be non-blocking. To prevent this allocation from being
445 * stalled by pagecache activity. But note that there may be
446 * stalls if we need to run get_block(). We could test
447 * PagePrivate for that.
449 * If this process is currently in __generic_file_aio_write() against
450 * this page's queue, we can perform writeback even if that
451 * will block.
453 * If the page is swapcache, write it back even if that would
454 * block, for some throttling. This happens by accident, because
455 * swap_backing_dev_info is bust: it doesn't reflect the
456 * congestion state of the swapdevs. Easy to fix, if needed.
458 if (!is_page_cache_freeable(page))
459 return PAGE_KEEP;
460 if (!mapping) {
462 * Some data journaling orphaned pages can have
463 * page->mapping == NULL while being dirty with clean buffers.
465 if (page_has_private(page)) {
466 if (try_to_free_buffers(page)) {
467 ClearPageDirty(page);
468 printk("%s: orphaned page\n", __func__);
469 return PAGE_CLEAN;
472 return PAGE_KEEP;
474 if (mapping->a_ops->writepage == NULL)
475 return PAGE_ACTIVATE;
476 if (!may_write_to_queue(mapping->backing_dev_info, sc))
477 return PAGE_KEEP;
479 if (clear_page_dirty_for_io(page)) {
480 int res;
481 struct writeback_control wbc = {
482 .sync_mode = WB_SYNC_NONE,
483 .nr_to_write = SWAP_CLUSTER_MAX,
484 .range_start = 0,
485 .range_end = LLONG_MAX,
486 .for_reclaim = 1,
489 SetPageReclaim(page);
490 res = mapping->a_ops->writepage(page, &wbc);
491 if (res < 0)
492 handle_write_error(mapping, page, res);
493 if (res == AOP_WRITEPAGE_ACTIVATE) {
494 ClearPageReclaim(page);
495 return PAGE_ACTIVATE;
499 * Wait on writeback if requested to. This happens when
500 * direct reclaiming a large contiguous area and the
501 * first attempt to free a range of pages fails.
503 if (PageWriteback(page) &&
504 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
505 wait_on_page_writeback(page);
507 if (!PageWriteback(page)) {
508 /* synchronous write or broken a_ops? */
509 ClearPageReclaim(page);
511 trace_mm_vmscan_writepage(page,
512 trace_reclaim_flags(page, sc->reclaim_mode));
513 inc_zone_page_state(page, NR_VMSCAN_WRITE);
514 return PAGE_SUCCESS;
517 return PAGE_CLEAN;
521 * Same as remove_mapping, but if the page is removed from the mapping, it
522 * gets returned with a refcount of 0.
524 static int __remove_mapping(struct address_space *mapping, struct page *page)
526 BUG_ON(!PageLocked(page));
527 BUG_ON(mapping != page_mapping(page));
529 spin_lock_irq(&mapping->tree_lock);
531 * The non racy check for a busy page.
533 * Must be careful with the order of the tests. When someone has
534 * a ref to the page, it may be possible that they dirty it then
535 * drop the reference. So if PageDirty is tested before page_count
536 * here, then the following race may occur:
538 * get_user_pages(&page);
539 * [user mapping goes away]
540 * write_to(page);
541 * !PageDirty(page) [good]
542 * SetPageDirty(page);
543 * put_page(page);
544 * !page_count(page) [good, discard it]
546 * [oops, our write_to data is lost]
548 * Reversing the order of the tests ensures such a situation cannot
549 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
550 * load is not satisfied before that of page->_count.
552 * Note that if SetPageDirty is always performed via set_page_dirty,
553 * and thus under tree_lock, then this ordering is not required.
555 if (!page_freeze_refs(page, 2))
556 goto cannot_free;
557 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
558 if (unlikely(PageDirty(page))) {
559 page_unfreeze_refs(page, 2);
560 goto cannot_free;
563 if (PageSwapCache(page)) {
564 swp_entry_t swap = { .val = page_private(page) };
565 __delete_from_swap_cache(page);
566 spin_unlock_irq(&mapping->tree_lock);
567 swapcache_free(swap, page);
568 } else {
569 void (*freepage)(struct page *);
571 freepage = mapping->a_ops->freepage;
573 __delete_from_page_cache(page);
574 spin_unlock_irq(&mapping->tree_lock);
575 mem_cgroup_uncharge_cache_page(page);
577 if (freepage != NULL)
578 freepage(page);
581 return 1;
583 cannot_free:
584 spin_unlock_irq(&mapping->tree_lock);
585 return 0;
589 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
590 * someone else has a ref on the page, abort and return 0. If it was
591 * successfully detached, return 1. Assumes the caller has a single ref on
592 * this page.
594 int remove_mapping(struct address_space *mapping, struct page *page)
596 if (__remove_mapping(mapping, page)) {
598 * Unfreezing the refcount with 1 rather than 2 effectively
599 * drops the pagecache ref for us without requiring another
600 * atomic operation.
602 page_unfreeze_refs(page, 1);
603 return 1;
605 return 0;
609 * putback_lru_page - put previously isolated page onto appropriate LRU list
610 * @page: page to be put back to appropriate lru list
612 * Add previously isolated @page to appropriate LRU list.
613 * Page may still be unevictable for other reasons.
615 * lru_lock must not be held, interrupts must be enabled.
617 void putback_lru_page(struct page *page)
619 int lru;
620 int active = !!TestClearPageActive(page);
621 int was_unevictable = PageUnevictable(page);
623 VM_BUG_ON(PageLRU(page));
625 redo:
626 ClearPageUnevictable(page);
628 if (page_evictable(page, NULL)) {
630 * For evictable pages, we can use the cache.
631 * In event of a race, worst case is we end up with an
632 * unevictable page on [in]active list.
633 * We know how to handle that.
635 lru = active + page_lru_base_type(page);
636 lru_cache_add_lru(page, lru);
637 } else {
639 * Put unevictable pages directly on zone's unevictable
640 * list.
642 lru = LRU_UNEVICTABLE;
643 add_page_to_unevictable_list(page);
645 * When racing with an mlock clearing (page is
646 * unlocked), make sure that if the other thread does
647 * not observe our setting of PG_lru and fails
648 * isolation, we see PG_mlocked cleared below and move
649 * the page back to the evictable list.
651 * The other side is TestClearPageMlocked().
653 smp_mb();
657 * page's status can change while we move it among lru. If an evictable
658 * page is on unevictable list, it never be freed. To avoid that,
659 * check after we added it to the list, again.
661 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
662 if (!isolate_lru_page(page)) {
663 put_page(page);
664 goto redo;
666 /* This means someone else dropped this page from LRU
667 * So, it will be freed or putback to LRU again. There is
668 * nothing to do here.
672 if (was_unevictable && lru != LRU_UNEVICTABLE)
673 count_vm_event(UNEVICTABLE_PGRESCUED);
674 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
675 count_vm_event(UNEVICTABLE_PGCULLED);
677 put_page(page); /* drop ref from isolate */
680 enum page_references {
681 PAGEREF_RECLAIM,
682 PAGEREF_RECLAIM_CLEAN,
683 PAGEREF_KEEP,
684 PAGEREF_ACTIVATE,
687 static enum page_references page_check_references(struct page *page,
688 struct scan_control *sc)
690 int referenced_ptes, referenced_page;
691 unsigned long vm_flags;
693 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
694 referenced_page = TestClearPageReferenced(page);
696 /* Lumpy reclaim - ignore references */
697 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
698 return PAGEREF_RECLAIM;
701 * Mlock lost the isolation race with us. Let try_to_unmap()
702 * move the page to the unevictable list.
704 if (vm_flags & VM_LOCKED)
705 return PAGEREF_RECLAIM;
707 if (referenced_ptes) {
708 if (PageAnon(page))
709 return PAGEREF_ACTIVATE;
711 * All mapped pages start out with page table
712 * references from the instantiating fault, so we need
713 * to look twice if a mapped file page is used more
714 * than once.
716 * Mark it and spare it for another trip around the
717 * inactive list. Another page table reference will
718 * lead to its activation.
720 * Note: the mark is set for activated pages as well
721 * so that recently deactivated but used pages are
722 * quickly recovered.
724 SetPageReferenced(page);
726 if (referenced_page)
727 return PAGEREF_ACTIVATE;
729 return PAGEREF_KEEP;
732 /* Reclaim if clean, defer dirty pages to writeback */
733 if (referenced_page && !PageSwapBacked(page))
734 return PAGEREF_RECLAIM_CLEAN;
736 return PAGEREF_RECLAIM;
739 static noinline_for_stack void free_page_list(struct list_head *free_pages)
741 struct pagevec freed_pvec;
742 struct page *page, *tmp;
744 pagevec_init(&freed_pvec, 1);
746 list_for_each_entry_safe(page, tmp, free_pages, lru) {
747 list_del(&page->lru);
748 if (!pagevec_add(&freed_pvec, page)) {
749 __pagevec_free(&freed_pvec);
750 pagevec_reinit(&freed_pvec);
754 pagevec_free(&freed_pvec);
758 * shrink_page_list() returns the number of reclaimed pages
760 static unsigned long shrink_page_list(struct list_head *page_list,
761 struct zone *zone,
762 struct scan_control *sc)
764 LIST_HEAD(ret_pages);
765 LIST_HEAD(free_pages);
766 int pgactivate = 0;
767 unsigned long nr_dirty = 0;
768 unsigned long nr_congested = 0;
769 unsigned long nr_reclaimed = 0;
771 cond_resched();
773 while (!list_empty(page_list)) {
774 enum page_references references;
775 struct address_space *mapping;
776 struct page *page;
777 int may_enter_fs;
779 cond_resched();
781 page = lru_to_page(page_list);
782 list_del(&page->lru);
784 if (!trylock_page(page))
785 goto keep;
787 VM_BUG_ON(PageActive(page));
788 VM_BUG_ON(page_zone(page) != zone);
790 sc->nr_scanned++;
792 if (unlikely(!page_evictable(page, NULL)))
793 goto cull_mlocked;
795 if (!sc->may_unmap && page_mapped(page))
796 goto keep_locked;
798 /* Double the slab pressure for mapped and swapcache pages */
799 if (page_mapped(page) || PageSwapCache(page))
800 sc->nr_scanned++;
802 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805 if (PageWriteback(page)) {
807 * Synchronous reclaim is performed in two passes,
808 * first an asynchronous pass over the list to
809 * start parallel writeback, and a second synchronous
810 * pass to wait for the IO to complete. Wait here
811 * for any page for which writeback has already
812 * started.
814 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
815 may_enter_fs)
816 wait_on_page_writeback(page);
817 else {
818 unlock_page(page);
819 goto keep_lumpy;
823 references = page_check_references(page, sc);
824 switch (references) {
825 case PAGEREF_ACTIVATE:
826 goto activate_locked;
827 case PAGEREF_KEEP:
828 goto keep_locked;
829 case PAGEREF_RECLAIM:
830 case PAGEREF_RECLAIM_CLEAN:
831 ; /* try to reclaim the page below */
835 * Anonymous process memory has backing store?
836 * Try to allocate it some swap space here.
838 if (PageAnon(page) && !PageSwapCache(page)) {
839 if (!(sc->gfp_mask & __GFP_IO))
840 goto keep_locked;
841 if (!add_to_swap(page))
842 goto activate_locked;
843 may_enter_fs = 1;
846 mapping = page_mapping(page);
849 * The page is mapped into the page tables of one or more
850 * processes. Try to unmap it here.
852 if (page_mapped(page) && mapping) {
853 switch (try_to_unmap(page, TTU_UNMAP)) {
854 case SWAP_FAIL:
855 goto activate_locked;
856 case SWAP_AGAIN:
857 goto keep_locked;
858 case SWAP_MLOCK:
859 goto cull_mlocked;
860 case SWAP_SUCCESS:
861 ; /* try to free the page below */
865 if (PageDirty(page)) {
866 nr_dirty++;
868 if (references == PAGEREF_RECLAIM_CLEAN)
869 goto keep_locked;
870 if (!may_enter_fs)
871 goto keep_locked;
872 if (!sc->may_writepage)
873 goto keep_locked;
875 /* Page is dirty, try to write it out here */
876 switch (pageout(page, mapping, sc)) {
877 case PAGE_KEEP:
878 nr_congested++;
879 goto keep_locked;
880 case PAGE_ACTIVATE:
881 goto activate_locked;
882 case PAGE_SUCCESS:
883 if (PageWriteback(page))
884 goto keep_lumpy;
885 if (PageDirty(page))
886 goto keep;
889 * A synchronous write - probably a ramdisk. Go
890 * ahead and try to reclaim the page.
892 if (!trylock_page(page))
893 goto keep;
894 if (PageDirty(page) || PageWriteback(page))
895 goto keep_locked;
896 mapping = page_mapping(page);
897 case PAGE_CLEAN:
898 ; /* try to free the page below */
903 * If the page has buffers, try to free the buffer mappings
904 * associated with this page. If we succeed we try to free
905 * the page as well.
907 * We do this even if the page is PageDirty().
908 * try_to_release_page() does not perform I/O, but it is
909 * possible for a page to have PageDirty set, but it is actually
910 * clean (all its buffers are clean). This happens if the
911 * buffers were written out directly, with submit_bh(). ext3
912 * will do this, as well as the blockdev mapping.
913 * try_to_release_page() will discover that cleanness and will
914 * drop the buffers and mark the page clean - it can be freed.
916 * Rarely, pages can have buffers and no ->mapping. These are
917 * the pages which were not successfully invalidated in
918 * truncate_complete_page(). We try to drop those buffers here
919 * and if that worked, and the page is no longer mapped into
920 * process address space (page_count == 1) it can be freed.
921 * Otherwise, leave the page on the LRU so it is swappable.
923 if (page_has_private(page)) {
924 if (!try_to_release_page(page, sc->gfp_mask))
925 goto activate_locked;
926 if (!mapping && page_count(page) == 1) {
927 unlock_page(page);
928 if (put_page_testzero(page))
929 goto free_it;
930 else {
932 * rare race with speculative reference.
933 * the speculative reference will free
934 * this page shortly, so we may
935 * increment nr_reclaimed here (and
936 * leave it off the LRU).
938 nr_reclaimed++;
939 continue;
944 if (!mapping || !__remove_mapping(mapping, page))
945 goto keep_locked;
948 * At this point, we have no other references and there is
949 * no way to pick any more up (removed from LRU, removed
950 * from pagecache). Can use non-atomic bitops now (and
951 * we obviously don't have to worry about waking up a process
952 * waiting on the page lock, because there are no references.
954 __clear_page_locked(page);
955 free_it:
956 nr_reclaimed++;
959 * Is there need to periodically free_page_list? It would
960 * appear not as the counts should be low
962 list_add(&page->lru, &free_pages);
963 continue;
965 cull_mlocked:
966 if (PageSwapCache(page))
967 try_to_free_swap(page);
968 unlock_page(page);
969 putback_lru_page(page);
970 reset_reclaim_mode(sc);
971 continue;
973 activate_locked:
974 /* Not a candidate for swapping, so reclaim swap space. */
975 if (PageSwapCache(page) && vm_swap_full())
976 try_to_free_swap(page);
977 VM_BUG_ON(PageActive(page));
978 SetPageActive(page);
979 pgactivate++;
980 keep_locked:
981 unlock_page(page);
982 keep:
983 reset_reclaim_mode(sc);
984 keep_lumpy:
985 list_add(&page->lru, &ret_pages);
986 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
990 * Tag a zone as congested if all the dirty pages encountered were
991 * backed by a congested BDI. In this case, reclaimers should just
992 * back off and wait for congestion to clear because further reclaim
993 * will encounter the same problem
995 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
996 zone_set_flag(zone, ZONE_CONGESTED);
998 free_page_list(&free_pages);
1000 list_splice(&ret_pages, page_list);
1001 count_vm_events(PGACTIVATE, pgactivate);
1002 return nr_reclaimed;
1006 * Attempt to remove the specified page from its LRU. Only take this page
1007 * if it is of the appropriate PageActive status. Pages which are being
1008 * freed elsewhere are also ignored.
1010 * page: page to consider
1011 * mode: one of the LRU isolation modes defined above
1013 * returns 0 on success, -ve errno on failure.
1015 int __isolate_lru_page(struct page *page, int mode, int file)
1017 int ret = -EINVAL;
1019 /* Only take pages on the LRU. */
1020 if (!PageLRU(page))
1021 return ret;
1024 * When checking the active state, we need to be sure we are
1025 * dealing with comparible boolean values. Take the logical not
1026 * of each.
1028 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
1029 return ret;
1031 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
1032 return ret;
1035 * When this function is being called for lumpy reclaim, we
1036 * initially look into all LRU pages, active, inactive and
1037 * unevictable; only give shrink_page_list evictable pages.
1039 if (PageUnevictable(page))
1040 return ret;
1042 ret = -EBUSY;
1044 if (likely(get_page_unless_zero(page))) {
1046 * Be careful not to clear PageLRU until after we're
1047 * sure the page is not being freed elsewhere -- the
1048 * page release code relies on it.
1050 ClearPageLRU(page);
1051 ret = 0;
1054 return ret;
1058 * zone->lru_lock is heavily contended. Some of the functions that
1059 * shrink the lists perform better by taking out a batch of pages
1060 * and working on them outside the LRU lock.
1062 * For pagecache intensive workloads, this function is the hottest
1063 * spot in the kernel (apart from copy_*_user functions).
1065 * Appropriate locks must be held before calling this function.
1067 * @nr_to_scan: The number of pages to look through on the list.
1068 * @src: The LRU list to pull pages off.
1069 * @dst: The temp list to put pages on to.
1070 * @scanned: The number of pages that were scanned.
1071 * @order: The caller's attempted allocation order
1072 * @mode: One of the LRU isolation modes
1073 * @file: True [1] if isolating file [!anon] pages
1075 * returns how many pages were moved onto *@dst.
1077 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1078 struct list_head *src, struct list_head *dst,
1079 unsigned long *scanned, int order, int mode, int file)
1081 unsigned long nr_taken = 0;
1082 unsigned long nr_lumpy_taken = 0;
1083 unsigned long nr_lumpy_dirty = 0;
1084 unsigned long nr_lumpy_failed = 0;
1085 unsigned long scan;
1087 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1088 struct page *page;
1089 unsigned long pfn;
1090 unsigned long end_pfn;
1091 unsigned long page_pfn;
1092 int zone_id;
1094 page = lru_to_page(src);
1095 prefetchw_prev_lru_page(page, src, flags);
1097 VM_BUG_ON(!PageLRU(page));
1099 switch (__isolate_lru_page(page, mode, file)) {
1100 case 0:
1101 list_move(&page->lru, dst);
1102 mem_cgroup_del_lru(page);
1103 nr_taken += hpage_nr_pages(page);
1104 break;
1106 case -EBUSY:
1107 /* else it is being freed elsewhere */
1108 list_move(&page->lru, src);
1109 mem_cgroup_rotate_lru_list(page, page_lru(page));
1110 continue;
1112 default:
1113 BUG();
1116 if (!order)
1117 continue;
1120 * Attempt to take all pages in the order aligned region
1121 * surrounding the tag page. Only take those pages of
1122 * the same active state as that tag page. We may safely
1123 * round the target page pfn down to the requested order
1124 * as the mem_map is guaranteed valid out to MAX_ORDER,
1125 * where that page is in a different zone we will detect
1126 * it from its zone id and abort this block scan.
1128 zone_id = page_zone_id(page);
1129 page_pfn = page_to_pfn(page);
1130 pfn = page_pfn & ~((1 << order) - 1);
1131 end_pfn = pfn + (1 << order);
1132 for (; pfn < end_pfn; pfn++) {
1133 struct page *cursor_page;
1135 /* The target page is in the block, ignore it. */
1136 if (unlikely(pfn == page_pfn))
1137 continue;
1139 /* Avoid holes within the zone. */
1140 if (unlikely(!pfn_valid_within(pfn)))
1141 break;
1143 cursor_page = pfn_to_page(pfn);
1145 /* Check that we have not crossed a zone boundary. */
1146 if (unlikely(page_zone_id(cursor_page) != zone_id))
1147 break;
1150 * If we don't have enough swap space, reclaiming of
1151 * anon page which don't already have a swap slot is
1152 * pointless.
1154 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1155 !PageSwapCache(cursor_page))
1156 break;
1158 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1159 list_move(&cursor_page->lru, dst);
1160 mem_cgroup_del_lru(cursor_page);
1161 nr_taken += hpage_nr_pages(page);
1162 nr_lumpy_taken++;
1163 if (PageDirty(cursor_page))
1164 nr_lumpy_dirty++;
1165 scan++;
1166 } else {
1168 * Check if the page is freed already.
1170 * We can't use page_count() as that
1171 * requires compound_head and we don't
1172 * have a pin on the page here. If a
1173 * page is tail, we may or may not
1174 * have isolated the head, so assume
1175 * it's not free, it'd be tricky to
1176 * track the head status without a
1177 * page pin.
1179 if (!PageTail(cursor_page) &&
1180 !atomic_read(&cursor_page->_count))
1181 continue;
1182 break;
1186 /* If we break out of the loop above, lumpy reclaim failed */
1187 if (pfn < end_pfn)
1188 nr_lumpy_failed++;
1191 *scanned = scan;
1193 trace_mm_vmscan_lru_isolate(order,
1194 nr_to_scan, scan,
1195 nr_taken,
1196 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1197 mode);
1198 return nr_taken;
1201 static unsigned long isolate_pages_global(unsigned long nr,
1202 struct list_head *dst,
1203 unsigned long *scanned, int order,
1204 int mode, struct zone *z,
1205 int active, int file)
1207 int lru = LRU_BASE;
1208 if (active)
1209 lru += LRU_ACTIVE;
1210 if (file)
1211 lru += LRU_FILE;
1212 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1213 mode, file);
1217 * clear_active_flags() is a helper for shrink_active_list(), clearing
1218 * any active bits from the pages in the list.
1220 static unsigned long clear_active_flags(struct list_head *page_list,
1221 unsigned int *count)
1223 int nr_active = 0;
1224 int lru;
1225 struct page *page;
1227 list_for_each_entry(page, page_list, lru) {
1228 int numpages = hpage_nr_pages(page);
1229 lru = page_lru_base_type(page);
1230 if (PageActive(page)) {
1231 lru += LRU_ACTIVE;
1232 ClearPageActive(page);
1233 nr_active += numpages;
1235 if (count)
1236 count[lru] += numpages;
1239 return nr_active;
1243 * isolate_lru_page - tries to isolate a page from its LRU list
1244 * @page: page to isolate from its LRU list
1246 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1247 * vmstat statistic corresponding to whatever LRU list the page was on.
1249 * Returns 0 if the page was removed from an LRU list.
1250 * Returns -EBUSY if the page was not on an LRU list.
1252 * The returned page will have PageLRU() cleared. If it was found on
1253 * the active list, it will have PageActive set. If it was found on
1254 * the unevictable list, it will have the PageUnevictable bit set. That flag
1255 * may need to be cleared by the caller before letting the page go.
1257 * The vmstat statistic corresponding to the list on which the page was
1258 * found will be decremented.
1260 * Restrictions:
1261 * (1) Must be called with an elevated refcount on the page. This is a
1262 * fundamentnal difference from isolate_lru_pages (which is called
1263 * without a stable reference).
1264 * (2) the lru_lock must not be held.
1265 * (3) interrupts must be enabled.
1267 int isolate_lru_page(struct page *page)
1269 int ret = -EBUSY;
1271 VM_BUG_ON(!page_count(page));
1273 if (PageLRU(page)) {
1274 struct zone *zone = page_zone(page);
1276 spin_lock_irq(&zone->lru_lock);
1277 if (PageLRU(page)) {
1278 int lru = page_lru(page);
1279 ret = 0;
1280 get_page(page);
1281 ClearPageLRU(page);
1283 del_page_from_lru_list(zone, page, lru);
1285 spin_unlock_irq(&zone->lru_lock);
1287 return ret;
1291 * Are there way too many processes in the direct reclaim path already?
1293 static int too_many_isolated(struct zone *zone, int file,
1294 struct scan_control *sc)
1296 unsigned long inactive, isolated;
1298 if (current_is_kswapd())
1299 return 0;
1301 if (!scanning_global_lru(sc))
1302 return 0;
1304 if (file) {
1305 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1306 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1307 } else {
1308 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1309 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1312 return isolated > inactive;
1316 * TODO: Try merging with migrations version of putback_lru_pages
1318 static noinline_for_stack void
1319 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1320 unsigned long nr_anon, unsigned long nr_file,
1321 struct list_head *page_list)
1323 struct page *page;
1324 struct pagevec pvec;
1325 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1327 pagevec_init(&pvec, 1);
1330 * Put back any unfreeable pages.
1332 spin_lock(&zone->lru_lock);
1333 while (!list_empty(page_list)) {
1334 int lru;
1335 page = lru_to_page(page_list);
1336 VM_BUG_ON(PageLRU(page));
1337 list_del(&page->lru);
1338 if (unlikely(!page_evictable(page, NULL))) {
1339 spin_unlock_irq(&zone->lru_lock);
1340 putback_lru_page(page);
1341 spin_lock_irq(&zone->lru_lock);
1342 continue;
1344 SetPageLRU(page);
1345 lru = page_lru(page);
1346 add_page_to_lru_list(zone, page, lru);
1347 if (is_active_lru(lru)) {
1348 int file = is_file_lru(lru);
1349 int numpages = hpage_nr_pages(page);
1350 reclaim_stat->recent_rotated[file] += numpages;
1352 if (!pagevec_add(&pvec, page)) {
1353 spin_unlock_irq(&zone->lru_lock);
1354 __pagevec_release(&pvec);
1355 spin_lock_irq(&zone->lru_lock);
1358 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1359 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1361 spin_unlock_irq(&zone->lru_lock);
1362 pagevec_release(&pvec);
1365 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1366 struct scan_control *sc,
1367 unsigned long *nr_anon,
1368 unsigned long *nr_file,
1369 struct list_head *isolated_list)
1371 unsigned long nr_active;
1372 unsigned int count[NR_LRU_LISTS] = { 0, };
1373 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1375 nr_active = clear_active_flags(isolated_list, count);
1376 __count_vm_events(PGDEACTIVATE, nr_active);
1378 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1379 -count[LRU_ACTIVE_FILE]);
1380 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1381 -count[LRU_INACTIVE_FILE]);
1382 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1383 -count[LRU_ACTIVE_ANON]);
1384 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1385 -count[LRU_INACTIVE_ANON]);
1387 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1388 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1389 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1390 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1392 reclaim_stat->recent_scanned[0] += *nr_anon;
1393 reclaim_stat->recent_scanned[1] += *nr_file;
1397 * Returns true if the caller should wait to clean dirty/writeback pages.
1399 * If we are direct reclaiming for contiguous pages and we do not reclaim
1400 * everything in the list, try again and wait for writeback IO to complete.
1401 * This will stall high-order allocations noticeably. Only do that when really
1402 * need to free the pages under high memory pressure.
1404 static inline bool should_reclaim_stall(unsigned long nr_taken,
1405 unsigned long nr_freed,
1406 int priority,
1407 struct scan_control *sc)
1409 int lumpy_stall_priority;
1411 /* kswapd should not stall on sync IO */
1412 if (current_is_kswapd())
1413 return false;
1415 /* Only stall on lumpy reclaim */
1416 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1417 return false;
1419 /* If we have relaimed everything on the isolated list, no stall */
1420 if (nr_freed == nr_taken)
1421 return false;
1424 * For high-order allocations, there are two stall thresholds.
1425 * High-cost allocations stall immediately where as lower
1426 * order allocations such as stacks require the scanning
1427 * priority to be much higher before stalling.
1429 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1430 lumpy_stall_priority = DEF_PRIORITY;
1431 else
1432 lumpy_stall_priority = DEF_PRIORITY / 3;
1434 return priority <= lumpy_stall_priority;
1438 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1439 * of reclaimed pages
1441 static noinline_for_stack unsigned long
1442 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1443 struct scan_control *sc, int priority, int file)
1445 LIST_HEAD(page_list);
1446 unsigned long nr_scanned;
1447 unsigned long nr_reclaimed = 0;
1448 unsigned long nr_taken;
1449 unsigned long nr_anon;
1450 unsigned long nr_file;
1452 while (unlikely(too_many_isolated(zone, file, sc))) {
1453 congestion_wait(BLK_RW_ASYNC, HZ/10);
1455 /* We are about to die and free our memory. Return now. */
1456 if (fatal_signal_pending(current))
1457 return SWAP_CLUSTER_MAX;
1460 set_reclaim_mode(priority, sc, false);
1461 lru_add_drain();
1462 spin_lock_irq(&zone->lru_lock);
1464 if (scanning_global_lru(sc)) {
1465 nr_taken = isolate_pages_global(nr_to_scan,
1466 &page_list, &nr_scanned, sc->order,
1467 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1468 ISOLATE_BOTH : ISOLATE_INACTIVE,
1469 zone, 0, file);
1470 zone->pages_scanned += nr_scanned;
1471 if (current_is_kswapd())
1472 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1473 nr_scanned);
1474 else
1475 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1476 nr_scanned);
1477 } else {
1478 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1479 &page_list, &nr_scanned, sc->order,
1480 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1481 ISOLATE_BOTH : ISOLATE_INACTIVE,
1482 zone, sc->mem_cgroup,
1483 0, file);
1485 * mem_cgroup_isolate_pages() keeps track of
1486 * scanned pages on its own.
1490 if (nr_taken == 0) {
1491 spin_unlock_irq(&zone->lru_lock);
1492 return 0;
1495 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1497 spin_unlock_irq(&zone->lru_lock);
1499 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1501 /* Check if we should syncronously wait for writeback */
1502 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1503 set_reclaim_mode(priority, sc, true);
1504 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1507 local_irq_disable();
1508 if (current_is_kswapd())
1509 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1510 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1512 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1514 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1515 zone_idx(zone),
1516 nr_scanned, nr_reclaimed,
1517 priority,
1518 trace_shrink_flags(file, sc->reclaim_mode));
1519 return nr_reclaimed;
1523 * This moves pages from the active list to the inactive list.
1525 * We move them the other way if the page is referenced by one or more
1526 * processes, from rmap.
1528 * If the pages are mostly unmapped, the processing is fast and it is
1529 * appropriate to hold zone->lru_lock across the whole operation. But if
1530 * the pages are mapped, the processing is slow (page_referenced()) so we
1531 * should drop zone->lru_lock around each page. It's impossible to balance
1532 * this, so instead we remove the pages from the LRU while processing them.
1533 * It is safe to rely on PG_active against the non-LRU pages in here because
1534 * nobody will play with that bit on a non-LRU page.
1536 * The downside is that we have to touch page->_count against each page.
1537 * But we had to alter page->flags anyway.
1540 static void move_active_pages_to_lru(struct zone *zone,
1541 struct list_head *list,
1542 enum lru_list lru)
1544 unsigned long pgmoved = 0;
1545 struct pagevec pvec;
1546 struct page *page;
1548 pagevec_init(&pvec, 1);
1550 while (!list_empty(list)) {
1551 page = lru_to_page(list);
1553 VM_BUG_ON(PageLRU(page));
1554 SetPageLRU(page);
1556 list_move(&page->lru, &zone->lru[lru].list);
1557 mem_cgroup_add_lru_list(page, lru);
1558 pgmoved += hpage_nr_pages(page);
1560 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1561 spin_unlock_irq(&zone->lru_lock);
1562 if (buffer_heads_over_limit)
1563 pagevec_strip(&pvec);
1564 __pagevec_release(&pvec);
1565 spin_lock_irq(&zone->lru_lock);
1568 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1569 if (!is_active_lru(lru))
1570 __count_vm_events(PGDEACTIVATE, pgmoved);
1573 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1574 struct scan_control *sc, int priority, int file)
1576 unsigned long nr_taken;
1577 unsigned long pgscanned;
1578 unsigned long vm_flags;
1579 LIST_HEAD(l_hold); /* The pages which were snipped off */
1580 LIST_HEAD(l_active);
1581 LIST_HEAD(l_inactive);
1582 struct page *page;
1583 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1584 unsigned long nr_rotated = 0;
1586 lru_add_drain();
1587 spin_lock_irq(&zone->lru_lock);
1588 if (scanning_global_lru(sc)) {
1589 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1590 &pgscanned, sc->order,
1591 ISOLATE_ACTIVE, zone,
1592 1, file);
1593 zone->pages_scanned += pgscanned;
1594 } else {
1595 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1596 &pgscanned, sc->order,
1597 ISOLATE_ACTIVE, zone,
1598 sc->mem_cgroup, 1, file);
1600 * mem_cgroup_isolate_pages() keeps track of
1601 * scanned pages on its own.
1605 reclaim_stat->recent_scanned[file] += nr_taken;
1607 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1608 if (file)
1609 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1610 else
1611 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1612 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1613 spin_unlock_irq(&zone->lru_lock);
1615 while (!list_empty(&l_hold)) {
1616 cond_resched();
1617 page = lru_to_page(&l_hold);
1618 list_del(&page->lru);
1620 if (unlikely(!page_evictable(page, NULL))) {
1621 putback_lru_page(page);
1622 continue;
1625 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1626 nr_rotated += hpage_nr_pages(page);
1628 * Identify referenced, file-backed active pages and
1629 * give them one more trip around the active list. So
1630 * that executable code get better chances to stay in
1631 * memory under moderate memory pressure. Anon pages
1632 * are not likely to be evicted by use-once streaming
1633 * IO, plus JVM can create lots of anon VM_EXEC pages,
1634 * so we ignore them here.
1636 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1637 list_add(&page->lru, &l_active);
1638 continue;
1642 ClearPageActive(page); /* we are de-activating */
1643 list_add(&page->lru, &l_inactive);
1647 * Move pages back to the lru list.
1649 spin_lock_irq(&zone->lru_lock);
1651 * Count referenced pages from currently used mappings as rotated,
1652 * even though only some of them are actually re-activated. This
1653 * helps balance scan pressure between file and anonymous pages in
1654 * get_scan_ratio.
1656 reclaim_stat->recent_rotated[file] += nr_rotated;
1658 move_active_pages_to_lru(zone, &l_active,
1659 LRU_ACTIVE + file * LRU_FILE);
1660 move_active_pages_to_lru(zone, &l_inactive,
1661 LRU_BASE + file * LRU_FILE);
1662 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1663 spin_unlock_irq(&zone->lru_lock);
1666 #ifdef CONFIG_SWAP
1667 static int inactive_anon_is_low_global(struct zone *zone)
1669 unsigned long active, inactive;
1671 active = zone_page_state(zone, NR_ACTIVE_ANON);
1672 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1674 if (inactive * zone->inactive_ratio < active)
1675 return 1;
1677 return 0;
1681 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1682 * @zone: zone to check
1683 * @sc: scan control of this context
1685 * Returns true if the zone does not have enough inactive anon pages,
1686 * meaning some active anon pages need to be deactivated.
1688 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1690 int low;
1693 * If we don't have swap space, anonymous page deactivation
1694 * is pointless.
1696 if (!total_swap_pages)
1697 return 0;
1699 if (scanning_global_lru(sc))
1700 low = inactive_anon_is_low_global(zone);
1701 else
1702 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1703 return low;
1705 #else
1706 static inline int inactive_anon_is_low(struct zone *zone,
1707 struct scan_control *sc)
1709 return 0;
1711 #endif
1713 static int inactive_file_is_low_global(struct zone *zone)
1715 unsigned long active, inactive;
1717 active = zone_page_state(zone, NR_ACTIVE_FILE);
1718 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1720 return (active > inactive);
1724 * inactive_file_is_low - check if file pages need to be deactivated
1725 * @zone: zone to check
1726 * @sc: scan control of this context
1728 * When the system is doing streaming IO, memory pressure here
1729 * ensures that active file pages get deactivated, until more
1730 * than half of the file pages are on the inactive list.
1732 * Once we get to that situation, protect the system's working
1733 * set from being evicted by disabling active file page aging.
1735 * This uses a different ratio than the anonymous pages, because
1736 * the page cache uses a use-once replacement algorithm.
1738 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1740 int low;
1742 if (scanning_global_lru(sc))
1743 low = inactive_file_is_low_global(zone);
1744 else
1745 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1746 return low;
1749 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1750 int file)
1752 if (file)
1753 return inactive_file_is_low(zone, sc);
1754 else
1755 return inactive_anon_is_low(zone, sc);
1758 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1759 struct zone *zone, struct scan_control *sc, int priority)
1761 int file = is_file_lru(lru);
1763 if (is_active_lru(lru)) {
1764 if (inactive_list_is_low(zone, sc, file))
1765 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1766 return 0;
1769 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1772 static int vmscan_swappiness(struct scan_control *sc)
1774 if (scanning_global_lru(sc))
1775 return vm_swappiness;
1776 return mem_cgroup_swappiness(sc->mem_cgroup);
1780 * Determine how aggressively the anon and file LRU lists should be
1781 * scanned. The relative value of each set of LRU lists is determined
1782 * by looking at the fraction of the pages scanned we did rotate back
1783 * onto the active list instead of evict.
1785 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1787 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1788 unsigned long *nr, int priority)
1790 unsigned long anon, file, free;
1791 unsigned long anon_prio, file_prio;
1792 unsigned long ap, fp;
1793 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1794 u64 fraction[2], denominator;
1795 enum lru_list l;
1796 int noswap = 0;
1797 bool force_scan = false;
1798 unsigned long nr_force_scan[2];
1800 /* kswapd does zone balancing and needs to scan this zone */
1801 if (scanning_global_lru(sc) && current_is_kswapd())
1802 force_scan = true;
1803 /* memcg may have small limit and need to avoid priority drop */
1804 if (!scanning_global_lru(sc))
1805 force_scan = true;
1807 /* If we have no swap space, do not bother scanning anon pages. */
1808 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1809 noswap = 1;
1810 fraction[0] = 0;
1811 fraction[1] = 1;
1812 denominator = 1;
1813 nr_force_scan[0] = 0;
1814 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1815 goto out;
1818 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1819 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1820 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1821 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1823 if (scanning_global_lru(sc)) {
1824 free = zone_page_state(zone, NR_FREE_PAGES);
1825 /* If we have very few page cache pages,
1826 force-scan anon pages. */
1827 if (unlikely(file + free <= high_wmark_pages(zone))) {
1828 fraction[0] = 1;
1829 fraction[1] = 0;
1830 denominator = 1;
1831 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1832 nr_force_scan[1] = 0;
1833 goto out;
1838 * With swappiness at 100, anonymous and file have the same priority.
1839 * This scanning priority is essentially the inverse of IO cost.
1841 anon_prio = vmscan_swappiness(sc);
1842 file_prio = 200 - vmscan_swappiness(sc);
1845 * OK, so we have swap space and a fair amount of page cache
1846 * pages. We use the recently rotated / recently scanned
1847 * ratios to determine how valuable each cache is.
1849 * Because workloads change over time (and to avoid overflow)
1850 * we keep these statistics as a floating average, which ends
1851 * up weighing recent references more than old ones.
1853 * anon in [0], file in [1]
1855 spin_lock_irq(&zone->lru_lock);
1856 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1857 reclaim_stat->recent_scanned[0] /= 2;
1858 reclaim_stat->recent_rotated[0] /= 2;
1861 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1862 reclaim_stat->recent_scanned[1] /= 2;
1863 reclaim_stat->recent_rotated[1] /= 2;
1867 * The amount of pressure on anon vs file pages is inversely
1868 * proportional to the fraction of recently scanned pages on
1869 * each list that were recently referenced and in active use.
1871 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1872 ap /= reclaim_stat->recent_rotated[0] + 1;
1874 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1875 fp /= reclaim_stat->recent_rotated[1] + 1;
1876 spin_unlock_irq(&zone->lru_lock);
1878 fraction[0] = ap;
1879 fraction[1] = fp;
1880 denominator = ap + fp + 1;
1881 if (force_scan) {
1882 unsigned long scan = SWAP_CLUSTER_MAX;
1883 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1884 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1886 out:
1887 for_each_evictable_lru(l) {
1888 int file = is_file_lru(l);
1889 unsigned long scan;
1891 scan = zone_nr_lru_pages(zone, sc, l);
1892 if (priority || noswap) {
1893 scan >>= priority;
1894 scan = div64_u64(scan * fraction[file], denominator);
1898 * If zone is small or memcg is small, nr[l] can be 0.
1899 * This results no-scan on this priority and priority drop down.
1900 * For global direct reclaim, it can visit next zone and tend
1901 * not to have problems. For global kswapd, it's for zone
1902 * balancing and it need to scan a small amounts. When using
1903 * memcg, priority drop can cause big latency. So, it's better
1904 * to scan small amount. See may_noscan above.
1906 if (!scan && force_scan)
1907 scan = nr_force_scan[file];
1908 nr[l] = scan;
1913 * Reclaim/compaction depends on a number of pages being freed. To avoid
1914 * disruption to the system, a small number of order-0 pages continue to be
1915 * rotated and reclaimed in the normal fashion. However, by the time we get
1916 * back to the allocator and call try_to_compact_zone(), we ensure that
1917 * there are enough free pages for it to be likely successful
1919 static inline bool should_continue_reclaim(struct zone *zone,
1920 unsigned long nr_reclaimed,
1921 unsigned long nr_scanned,
1922 struct scan_control *sc)
1924 unsigned long pages_for_compaction;
1925 unsigned long inactive_lru_pages;
1927 /* If not in reclaim/compaction mode, stop */
1928 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1929 return false;
1931 /* Consider stopping depending on scan and reclaim activity */
1932 if (sc->gfp_mask & __GFP_REPEAT) {
1934 * For __GFP_REPEAT allocations, stop reclaiming if the
1935 * full LRU list has been scanned and we are still failing
1936 * to reclaim pages. This full LRU scan is potentially
1937 * expensive but a __GFP_REPEAT caller really wants to succeed
1939 if (!nr_reclaimed && !nr_scanned)
1940 return false;
1941 } else {
1943 * For non-__GFP_REPEAT allocations which can presumably
1944 * fail without consequence, stop if we failed to reclaim
1945 * any pages from the last SWAP_CLUSTER_MAX number of
1946 * pages that were scanned. This will return to the
1947 * caller faster at the risk reclaim/compaction and
1948 * the resulting allocation attempt fails
1950 if (!nr_reclaimed)
1951 return false;
1955 * If we have not reclaimed enough pages for compaction and the
1956 * inactive lists are large enough, continue reclaiming
1958 pages_for_compaction = (2UL << sc->order);
1959 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1960 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1961 if (sc->nr_reclaimed < pages_for_compaction &&
1962 inactive_lru_pages > pages_for_compaction)
1963 return true;
1965 /* If compaction would go ahead or the allocation would succeed, stop */
1966 switch (compaction_suitable(zone, sc->order)) {
1967 case COMPACT_PARTIAL:
1968 case COMPACT_CONTINUE:
1969 return false;
1970 default:
1971 return true;
1976 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1978 static void shrink_zone(int priority, struct zone *zone,
1979 struct scan_control *sc)
1981 unsigned long nr[NR_LRU_LISTS];
1982 unsigned long nr_to_scan;
1983 enum lru_list l;
1984 unsigned long nr_reclaimed, nr_scanned;
1985 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1987 restart:
1988 nr_reclaimed = 0;
1989 nr_scanned = sc->nr_scanned;
1990 get_scan_count(zone, sc, nr, priority);
1992 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1993 nr[LRU_INACTIVE_FILE]) {
1994 for_each_evictable_lru(l) {
1995 if (nr[l]) {
1996 nr_to_scan = min_t(unsigned long,
1997 nr[l], SWAP_CLUSTER_MAX);
1998 nr[l] -= nr_to_scan;
2000 nr_reclaimed += shrink_list(l, nr_to_scan,
2001 zone, sc, priority);
2005 * On large memory systems, scan >> priority can become
2006 * really large. This is fine for the starting priority;
2007 * we want to put equal scanning pressure on each zone.
2008 * However, if the VM has a harder time of freeing pages,
2009 * with multiple processes reclaiming pages, the total
2010 * freeing target can get unreasonably large.
2012 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2013 break;
2015 sc->nr_reclaimed += nr_reclaimed;
2018 * Even if we did not try to evict anon pages at all, we want to
2019 * rebalance the anon lru active/inactive ratio.
2021 if (inactive_anon_is_low(zone, sc))
2022 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2024 /* reclaim/compaction might need reclaim to continue */
2025 if (should_continue_reclaim(zone, nr_reclaimed,
2026 sc->nr_scanned - nr_scanned, sc))
2027 goto restart;
2029 throttle_vm_writeout(sc->gfp_mask);
2033 * This is the direct reclaim path, for page-allocating processes. We only
2034 * try to reclaim pages from zones which will satisfy the caller's allocation
2035 * request.
2037 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2038 * Because:
2039 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2040 * allocation or
2041 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2042 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2043 * zone defense algorithm.
2045 * If a zone is deemed to be full of pinned pages then just give it a light
2046 * scan then give up on it.
2048 static void shrink_zones(int priority, struct zonelist *zonelist,
2049 struct scan_control *sc)
2051 struct zoneref *z;
2052 struct zone *zone;
2053 unsigned long nr_soft_reclaimed;
2054 unsigned long nr_soft_scanned;
2056 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2057 gfp_zone(sc->gfp_mask), sc->nodemask) {
2058 if (!populated_zone(zone))
2059 continue;
2061 * Take care memory controller reclaiming has small influence
2062 * to global LRU.
2064 if (scanning_global_lru(sc)) {
2065 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2066 continue;
2067 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2068 continue; /* Let kswapd poll it */
2070 * This steals pages from memory cgroups over softlimit
2071 * and returns the number of reclaimed pages and
2072 * scanned pages. This works for global memory pressure
2073 * and balancing, not for a memcg's limit.
2075 nr_soft_scanned = 0;
2076 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2077 sc->order, sc->gfp_mask,
2078 &nr_soft_scanned);
2079 sc->nr_reclaimed += nr_soft_reclaimed;
2080 sc->nr_scanned += nr_soft_scanned;
2081 /* need some check for avoid more shrink_zone() */
2084 shrink_zone(priority, zone, sc);
2088 static bool zone_reclaimable(struct zone *zone)
2090 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2093 /* All zones in zonelist are unreclaimable? */
2094 static bool all_unreclaimable(struct zonelist *zonelist,
2095 struct scan_control *sc)
2097 struct zoneref *z;
2098 struct zone *zone;
2100 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2101 gfp_zone(sc->gfp_mask), sc->nodemask) {
2102 if (!populated_zone(zone))
2103 continue;
2104 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2105 continue;
2106 if (!zone->all_unreclaimable)
2107 return false;
2110 return true;
2114 * This is the main entry point to direct page reclaim.
2116 * If a full scan of the inactive list fails to free enough memory then we
2117 * are "out of memory" and something needs to be killed.
2119 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2120 * high - the zone may be full of dirty or under-writeback pages, which this
2121 * caller can't do much about. We kick the writeback threads and take explicit
2122 * naps in the hope that some of these pages can be written. But if the
2123 * allocating task holds filesystem locks which prevent writeout this might not
2124 * work, and the allocation attempt will fail.
2126 * returns: 0, if no pages reclaimed
2127 * else, the number of pages reclaimed
2129 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2130 struct scan_control *sc,
2131 struct shrink_control *shrink)
2133 int priority;
2134 unsigned long total_scanned = 0;
2135 struct reclaim_state *reclaim_state = current->reclaim_state;
2136 struct zoneref *z;
2137 struct zone *zone;
2138 unsigned long writeback_threshold;
2140 get_mems_allowed();
2141 delayacct_freepages_start();
2143 if (scanning_global_lru(sc))
2144 count_vm_event(ALLOCSTALL);
2146 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2147 sc->nr_scanned = 0;
2148 if (!priority)
2149 disable_swap_token(sc->mem_cgroup);
2150 shrink_zones(priority, zonelist, sc);
2152 * Don't shrink slabs when reclaiming memory from
2153 * over limit cgroups
2155 if (scanning_global_lru(sc)) {
2156 unsigned long lru_pages = 0;
2157 for_each_zone_zonelist(zone, z, zonelist,
2158 gfp_zone(sc->gfp_mask)) {
2159 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2160 continue;
2162 lru_pages += zone_reclaimable_pages(zone);
2165 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2166 if (reclaim_state) {
2167 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2168 reclaim_state->reclaimed_slab = 0;
2171 total_scanned += sc->nr_scanned;
2172 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2173 goto out;
2176 * Try to write back as many pages as we just scanned. This
2177 * tends to cause slow streaming writers to write data to the
2178 * disk smoothly, at the dirtying rate, which is nice. But
2179 * that's undesirable in laptop mode, where we *want* lumpy
2180 * writeout. So in laptop mode, write out the whole world.
2182 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2183 if (total_scanned > writeback_threshold) {
2184 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2185 sc->may_writepage = 1;
2188 /* Take a nap, wait for some writeback to complete */
2189 if (!sc->hibernation_mode && sc->nr_scanned &&
2190 priority < DEF_PRIORITY - 2) {
2191 struct zone *preferred_zone;
2193 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2194 &cpuset_current_mems_allowed,
2195 &preferred_zone);
2196 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2200 out:
2201 delayacct_freepages_end();
2202 put_mems_allowed();
2204 if (sc->nr_reclaimed)
2205 return sc->nr_reclaimed;
2208 * As hibernation is going on, kswapd is freezed so that it can't mark
2209 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2210 * check.
2212 if (oom_killer_disabled)
2213 return 0;
2215 /* top priority shrink_zones still had more to do? don't OOM, then */
2216 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2217 return 1;
2219 return 0;
2222 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2223 gfp_t gfp_mask, nodemask_t *nodemask)
2225 unsigned long nr_reclaimed;
2226 struct scan_control sc = {
2227 .gfp_mask = gfp_mask,
2228 .may_writepage = !laptop_mode,
2229 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2230 .may_unmap = 1,
2231 .may_swap = 1,
2232 .order = order,
2233 .mem_cgroup = NULL,
2234 .nodemask = nodemask,
2236 struct shrink_control shrink = {
2237 .gfp_mask = sc.gfp_mask,
2240 trace_mm_vmscan_direct_reclaim_begin(order,
2241 sc.may_writepage,
2242 gfp_mask);
2244 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2246 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2248 return nr_reclaimed;
2251 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2253 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2254 gfp_t gfp_mask, bool noswap,
2255 struct zone *zone,
2256 unsigned long *nr_scanned)
2258 struct scan_control sc = {
2259 .nr_scanned = 0,
2260 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2261 .may_writepage = !laptop_mode,
2262 .may_unmap = 1,
2263 .may_swap = !noswap,
2264 .order = 0,
2265 .mem_cgroup = mem,
2268 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2269 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2271 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2272 sc.may_writepage,
2273 sc.gfp_mask);
2276 * NOTE: Although we can get the priority field, using it
2277 * here is not a good idea, since it limits the pages we can scan.
2278 * if we don't reclaim here, the shrink_zone from balance_pgdat
2279 * will pick up pages from other mem cgroup's as well. We hack
2280 * the priority and make it zero.
2282 shrink_zone(0, zone, &sc);
2284 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2286 *nr_scanned = sc.nr_scanned;
2287 return sc.nr_reclaimed;
2290 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2291 gfp_t gfp_mask,
2292 bool noswap)
2294 struct zonelist *zonelist;
2295 unsigned long nr_reclaimed;
2296 int nid;
2297 struct scan_control sc = {
2298 .may_writepage = !laptop_mode,
2299 .may_unmap = 1,
2300 .may_swap = !noswap,
2301 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2302 .order = 0,
2303 .mem_cgroup = mem_cont,
2304 .nodemask = NULL, /* we don't care the placement */
2305 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2306 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2308 struct shrink_control shrink = {
2309 .gfp_mask = sc.gfp_mask,
2313 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2314 * take care of from where we get pages. So the node where we start the
2315 * scan does not need to be the current node.
2317 nid = mem_cgroup_select_victim_node(mem_cont);
2319 zonelist = NODE_DATA(nid)->node_zonelists;
2321 trace_mm_vmscan_memcg_reclaim_begin(0,
2322 sc.may_writepage,
2323 sc.gfp_mask);
2325 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2327 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2329 return nr_reclaimed;
2331 #endif
2334 * pgdat_balanced is used when checking if a node is balanced for high-order
2335 * allocations. Only zones that meet watermarks and are in a zone allowed
2336 * by the callers classzone_idx are added to balanced_pages. The total of
2337 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2338 * for the node to be considered balanced. Forcing all zones to be balanced
2339 * for high orders can cause excessive reclaim when there are imbalanced zones.
2340 * The choice of 25% is due to
2341 * o a 16M DMA zone that is balanced will not balance a zone on any
2342 * reasonable sized machine
2343 * o On all other machines, the top zone must be at least a reasonable
2344 * percentage of the middle zones. For example, on 32-bit x86, highmem
2345 * would need to be at least 256M for it to be balance a whole node.
2346 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2347 * to balance a node on its own. These seemed like reasonable ratios.
2349 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2350 int classzone_idx)
2352 unsigned long present_pages = 0;
2353 int i;
2355 for (i = 0; i <= classzone_idx; i++)
2356 present_pages += pgdat->node_zones[i].present_pages;
2358 /* A special case here: if zone has no page, we think it's balanced */
2359 return balanced_pages >= (present_pages >> 2);
2362 /* is kswapd sleeping prematurely? */
2363 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2364 int classzone_idx)
2366 int i;
2367 unsigned long balanced = 0;
2368 bool all_zones_ok = true;
2370 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2371 if (remaining)
2372 return true;
2374 /* Check the watermark levels */
2375 for (i = 0; i <= classzone_idx; i++) {
2376 struct zone *zone = pgdat->node_zones + i;
2378 if (!populated_zone(zone))
2379 continue;
2382 * balance_pgdat() skips over all_unreclaimable after
2383 * DEF_PRIORITY. Effectively, it considers them balanced so
2384 * they must be considered balanced here as well if kswapd
2385 * is to sleep
2387 if (zone->all_unreclaimable) {
2388 balanced += zone->present_pages;
2389 continue;
2392 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2393 i, 0))
2394 all_zones_ok = false;
2395 else
2396 balanced += zone->present_pages;
2400 * For high-order requests, the balanced zones must contain at least
2401 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2402 * must be balanced
2404 if (order)
2405 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2406 else
2407 return !all_zones_ok;
2411 * For kswapd, balance_pgdat() will work across all this node's zones until
2412 * they are all at high_wmark_pages(zone).
2414 * Returns the final order kswapd was reclaiming at
2416 * There is special handling here for zones which are full of pinned pages.
2417 * This can happen if the pages are all mlocked, or if they are all used by
2418 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2419 * What we do is to detect the case where all pages in the zone have been
2420 * scanned twice and there has been zero successful reclaim. Mark the zone as
2421 * dead and from now on, only perform a short scan. Basically we're polling
2422 * the zone for when the problem goes away.
2424 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2425 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2426 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2427 * lower zones regardless of the number of free pages in the lower zones. This
2428 * interoperates with the page allocator fallback scheme to ensure that aging
2429 * of pages is balanced across the zones.
2431 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2432 int *classzone_idx)
2434 int all_zones_ok;
2435 unsigned long balanced;
2436 int priority;
2437 int i;
2438 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2439 unsigned long total_scanned;
2440 struct reclaim_state *reclaim_state = current->reclaim_state;
2441 unsigned long nr_soft_reclaimed;
2442 unsigned long nr_soft_scanned;
2443 struct scan_control sc = {
2444 .gfp_mask = GFP_KERNEL,
2445 .may_unmap = 1,
2446 .may_swap = 1,
2448 * kswapd doesn't want to be bailed out while reclaim. because
2449 * we want to put equal scanning pressure on each zone.
2451 .nr_to_reclaim = ULONG_MAX,
2452 .order = order,
2453 .mem_cgroup = NULL,
2455 struct shrink_control shrink = {
2456 .gfp_mask = sc.gfp_mask,
2458 loop_again:
2459 total_scanned = 0;
2460 sc.nr_reclaimed = 0;
2461 sc.may_writepage = !laptop_mode;
2462 count_vm_event(PAGEOUTRUN);
2464 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2465 unsigned long lru_pages = 0;
2466 int has_under_min_watermark_zone = 0;
2468 /* The swap token gets in the way of swapout... */
2469 if (!priority)
2470 disable_swap_token(NULL);
2472 all_zones_ok = 1;
2473 balanced = 0;
2476 * Scan in the highmem->dma direction for the highest
2477 * zone which needs scanning
2479 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2480 struct zone *zone = pgdat->node_zones + i;
2482 if (!populated_zone(zone))
2483 continue;
2485 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2486 continue;
2489 * Do some background aging of the anon list, to give
2490 * pages a chance to be referenced before reclaiming.
2492 if (inactive_anon_is_low(zone, &sc))
2493 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2494 &sc, priority, 0);
2496 if (!zone_watermark_ok_safe(zone, order,
2497 high_wmark_pages(zone), 0, 0)) {
2498 end_zone = i;
2499 break;
2500 } else {
2501 /* If balanced, clear the congested flag */
2502 zone_clear_flag(zone, ZONE_CONGESTED);
2505 if (i < 0)
2506 goto out;
2508 for (i = 0; i <= end_zone; i++) {
2509 struct zone *zone = pgdat->node_zones + i;
2511 lru_pages += zone_reclaimable_pages(zone);
2515 * Now scan the zone in the dma->highmem direction, stopping
2516 * at the last zone which needs scanning.
2518 * We do this because the page allocator works in the opposite
2519 * direction. This prevents the page allocator from allocating
2520 * pages behind kswapd's direction of progress, which would
2521 * cause too much scanning of the lower zones.
2523 for (i = 0; i <= end_zone; i++) {
2524 struct zone *zone = pgdat->node_zones + i;
2525 int nr_slab;
2526 unsigned long balance_gap;
2528 if (!populated_zone(zone))
2529 continue;
2531 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2532 continue;
2534 sc.nr_scanned = 0;
2536 nr_soft_scanned = 0;
2538 * Call soft limit reclaim before calling shrink_zone.
2540 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2541 order, sc.gfp_mask,
2542 &nr_soft_scanned);
2543 sc.nr_reclaimed += nr_soft_reclaimed;
2544 total_scanned += nr_soft_scanned;
2547 * We put equal pressure on every zone, unless
2548 * one zone has way too many pages free
2549 * already. The "too many pages" is defined
2550 * as the high wmark plus a "gap" where the
2551 * gap is either the low watermark or 1%
2552 * of the zone, whichever is smaller.
2554 balance_gap = min(low_wmark_pages(zone),
2555 (zone->present_pages +
2556 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2557 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2558 if (!zone_watermark_ok_safe(zone, order,
2559 high_wmark_pages(zone) + balance_gap,
2560 end_zone, 0)) {
2561 shrink_zone(priority, zone, &sc);
2563 reclaim_state->reclaimed_slab = 0;
2564 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2565 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2566 total_scanned += sc.nr_scanned;
2568 if (nr_slab == 0 && !zone_reclaimable(zone))
2569 zone->all_unreclaimable = 1;
2573 * If we've done a decent amount of scanning and
2574 * the reclaim ratio is low, start doing writepage
2575 * even in laptop mode
2577 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2578 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2579 sc.may_writepage = 1;
2581 if (zone->all_unreclaimable) {
2582 if (end_zone && end_zone == i)
2583 end_zone--;
2584 continue;
2587 if (!zone_watermark_ok_safe(zone, order,
2588 high_wmark_pages(zone), end_zone, 0)) {
2589 all_zones_ok = 0;
2591 * We are still under min water mark. This
2592 * means that we have a GFP_ATOMIC allocation
2593 * failure risk. Hurry up!
2595 if (!zone_watermark_ok_safe(zone, order,
2596 min_wmark_pages(zone), end_zone, 0))
2597 has_under_min_watermark_zone = 1;
2598 } else {
2600 * If a zone reaches its high watermark,
2601 * consider it to be no longer congested. It's
2602 * possible there are dirty pages backed by
2603 * congested BDIs but as pressure is relieved,
2604 * spectulatively avoid congestion waits
2606 zone_clear_flag(zone, ZONE_CONGESTED);
2607 if (i <= *classzone_idx)
2608 balanced += zone->present_pages;
2612 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2613 break; /* kswapd: all done */
2615 * OK, kswapd is getting into trouble. Take a nap, then take
2616 * another pass across the zones.
2618 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2619 if (has_under_min_watermark_zone)
2620 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2621 else
2622 congestion_wait(BLK_RW_ASYNC, HZ/10);
2626 * We do this so kswapd doesn't build up large priorities for
2627 * example when it is freeing in parallel with allocators. It
2628 * matches the direct reclaim path behaviour in terms of impact
2629 * on zone->*_priority.
2631 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2632 break;
2634 out:
2637 * order-0: All zones must meet high watermark for a balanced node
2638 * high-order: Balanced zones must make up at least 25% of the node
2639 * for the node to be balanced
2641 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2642 cond_resched();
2644 try_to_freeze();
2647 * Fragmentation may mean that the system cannot be
2648 * rebalanced for high-order allocations in all zones.
2649 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2650 * it means the zones have been fully scanned and are still
2651 * not balanced. For high-order allocations, there is
2652 * little point trying all over again as kswapd may
2653 * infinite loop.
2655 * Instead, recheck all watermarks at order-0 as they
2656 * are the most important. If watermarks are ok, kswapd will go
2657 * back to sleep. High-order users can still perform direct
2658 * reclaim if they wish.
2660 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2661 order = sc.order = 0;
2663 goto loop_again;
2667 * If kswapd was reclaiming at a higher order, it has the option of
2668 * sleeping without all zones being balanced. Before it does, it must
2669 * ensure that the watermarks for order-0 on *all* zones are met and
2670 * that the congestion flags are cleared. The congestion flag must
2671 * be cleared as kswapd is the only mechanism that clears the flag
2672 * and it is potentially going to sleep here.
2674 if (order) {
2675 for (i = 0; i <= end_zone; i++) {
2676 struct zone *zone = pgdat->node_zones + i;
2678 if (!populated_zone(zone))
2679 continue;
2681 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2682 continue;
2684 /* Confirm the zone is balanced for order-0 */
2685 if (!zone_watermark_ok(zone, 0,
2686 high_wmark_pages(zone), 0, 0)) {
2687 order = sc.order = 0;
2688 goto loop_again;
2691 /* If balanced, clear the congested flag */
2692 zone_clear_flag(zone, ZONE_CONGESTED);
2697 * Return the order we were reclaiming at so sleeping_prematurely()
2698 * makes a decision on the order we were last reclaiming at. However,
2699 * if another caller entered the allocator slow path while kswapd
2700 * was awake, order will remain at the higher level
2702 *classzone_idx = end_zone;
2703 return order;
2706 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2708 long remaining = 0;
2709 DEFINE_WAIT(wait);
2711 if (freezing(current) || kthread_should_stop())
2712 return;
2714 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2716 /* Try to sleep for a short interval */
2717 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2718 remaining = schedule_timeout(HZ/10);
2719 finish_wait(&pgdat->kswapd_wait, &wait);
2720 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2724 * After a short sleep, check if it was a premature sleep. If not, then
2725 * go fully to sleep until explicitly woken up.
2727 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2728 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2731 * vmstat counters are not perfectly accurate and the estimated
2732 * value for counters such as NR_FREE_PAGES can deviate from the
2733 * true value by nr_online_cpus * threshold. To avoid the zone
2734 * watermarks being breached while under pressure, we reduce the
2735 * per-cpu vmstat threshold while kswapd is awake and restore
2736 * them before going back to sleep.
2738 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2739 schedule();
2740 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2741 } else {
2742 if (remaining)
2743 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2744 else
2745 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2747 finish_wait(&pgdat->kswapd_wait, &wait);
2751 * The background pageout daemon, started as a kernel thread
2752 * from the init process.
2754 * This basically trickles out pages so that we have _some_
2755 * free memory available even if there is no other activity
2756 * that frees anything up. This is needed for things like routing
2757 * etc, where we otherwise might have all activity going on in
2758 * asynchronous contexts that cannot page things out.
2760 * If there are applications that are active memory-allocators
2761 * (most normal use), this basically shouldn't matter.
2763 static int kswapd(void *p)
2765 unsigned long order, new_order;
2766 int classzone_idx, new_classzone_idx;
2767 pg_data_t *pgdat = (pg_data_t*)p;
2768 struct task_struct *tsk = current;
2770 struct reclaim_state reclaim_state = {
2771 .reclaimed_slab = 0,
2773 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2775 lockdep_set_current_reclaim_state(GFP_KERNEL);
2777 if (!cpumask_empty(cpumask))
2778 set_cpus_allowed_ptr(tsk, cpumask);
2779 current->reclaim_state = &reclaim_state;
2782 * Tell the memory management that we're a "memory allocator",
2783 * and that if we need more memory we should get access to it
2784 * regardless (see "__alloc_pages()"). "kswapd" should
2785 * never get caught in the normal page freeing logic.
2787 * (Kswapd normally doesn't need memory anyway, but sometimes
2788 * you need a small amount of memory in order to be able to
2789 * page out something else, and this flag essentially protects
2790 * us from recursively trying to free more memory as we're
2791 * trying to free the first piece of memory in the first place).
2793 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2794 set_freezable();
2796 order = new_order = 0;
2797 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2798 for ( ; ; ) {
2799 int ret;
2802 * If the last balance_pgdat was unsuccessful it's unlikely a
2803 * new request of a similar or harder type will succeed soon
2804 * so consider going to sleep on the basis we reclaimed at
2806 if (classzone_idx >= new_classzone_idx && order == new_order) {
2807 new_order = pgdat->kswapd_max_order;
2808 new_classzone_idx = pgdat->classzone_idx;
2809 pgdat->kswapd_max_order = 0;
2810 pgdat->classzone_idx = pgdat->nr_zones - 1;
2813 if (order < new_order || classzone_idx > new_classzone_idx) {
2815 * Don't sleep if someone wants a larger 'order'
2816 * allocation or has tigher zone constraints
2818 order = new_order;
2819 classzone_idx = new_classzone_idx;
2820 } else {
2821 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2822 order = pgdat->kswapd_max_order;
2823 classzone_idx = pgdat->classzone_idx;
2824 pgdat->kswapd_max_order = 0;
2825 pgdat->classzone_idx = pgdat->nr_zones - 1;
2828 ret = try_to_freeze();
2829 if (kthread_should_stop())
2830 break;
2833 * We can speed up thawing tasks if we don't call balance_pgdat
2834 * after returning from the refrigerator
2836 if (!ret) {
2837 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2838 order = balance_pgdat(pgdat, order, &classzone_idx);
2841 return 0;
2845 * A zone is low on free memory, so wake its kswapd task to service it.
2847 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2849 pg_data_t *pgdat;
2851 if (!populated_zone(zone))
2852 return;
2854 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2855 return;
2856 pgdat = zone->zone_pgdat;
2857 if (pgdat->kswapd_max_order < order) {
2858 pgdat->kswapd_max_order = order;
2859 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2861 if (!waitqueue_active(&pgdat->kswapd_wait))
2862 return;
2863 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2864 return;
2866 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2867 wake_up_interruptible(&pgdat->kswapd_wait);
2871 * The reclaimable count would be mostly accurate.
2872 * The less reclaimable pages may be
2873 * - mlocked pages, which will be moved to unevictable list when encountered
2874 * - mapped pages, which may require several travels to be reclaimed
2875 * - dirty pages, which is not "instantly" reclaimable
2877 unsigned long global_reclaimable_pages(void)
2879 int nr;
2881 nr = global_page_state(NR_ACTIVE_FILE) +
2882 global_page_state(NR_INACTIVE_FILE);
2884 if (nr_swap_pages > 0)
2885 nr += global_page_state(NR_ACTIVE_ANON) +
2886 global_page_state(NR_INACTIVE_ANON);
2888 return nr;
2891 unsigned long zone_reclaimable_pages(struct zone *zone)
2893 int nr;
2895 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2896 zone_page_state(zone, NR_INACTIVE_FILE);
2898 if (nr_swap_pages > 0)
2899 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2900 zone_page_state(zone, NR_INACTIVE_ANON);
2902 return nr;
2905 #ifdef CONFIG_HIBERNATION
2907 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2908 * freed pages.
2910 * Rather than trying to age LRUs the aim is to preserve the overall
2911 * LRU order by reclaiming preferentially
2912 * inactive > active > active referenced > active mapped
2914 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2916 struct reclaim_state reclaim_state;
2917 struct scan_control sc = {
2918 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2919 .may_swap = 1,
2920 .may_unmap = 1,
2921 .may_writepage = 1,
2922 .nr_to_reclaim = nr_to_reclaim,
2923 .hibernation_mode = 1,
2924 .order = 0,
2926 struct shrink_control shrink = {
2927 .gfp_mask = sc.gfp_mask,
2929 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2930 struct task_struct *p = current;
2931 unsigned long nr_reclaimed;
2933 p->flags |= PF_MEMALLOC;
2934 lockdep_set_current_reclaim_state(sc.gfp_mask);
2935 reclaim_state.reclaimed_slab = 0;
2936 p->reclaim_state = &reclaim_state;
2938 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2940 p->reclaim_state = NULL;
2941 lockdep_clear_current_reclaim_state();
2942 p->flags &= ~PF_MEMALLOC;
2944 return nr_reclaimed;
2946 #endif /* CONFIG_HIBERNATION */
2948 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2949 not required for correctness. So if the last cpu in a node goes
2950 away, we get changed to run anywhere: as the first one comes back,
2951 restore their cpu bindings. */
2952 static int __devinit cpu_callback(struct notifier_block *nfb,
2953 unsigned long action, void *hcpu)
2955 int nid;
2957 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2958 for_each_node_state(nid, N_HIGH_MEMORY) {
2959 pg_data_t *pgdat = NODE_DATA(nid);
2960 const struct cpumask *mask;
2962 mask = cpumask_of_node(pgdat->node_id);
2964 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2965 /* One of our CPUs online: restore mask */
2966 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2969 return NOTIFY_OK;
2973 * This kswapd start function will be called by init and node-hot-add.
2974 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2976 int kswapd_run(int nid)
2978 pg_data_t *pgdat = NODE_DATA(nid);
2979 int ret = 0;
2981 if (pgdat->kswapd)
2982 return 0;
2984 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2985 if (IS_ERR(pgdat->kswapd)) {
2986 /* failure at boot is fatal */
2987 BUG_ON(system_state == SYSTEM_BOOTING);
2988 printk("Failed to start kswapd on node %d\n",nid);
2989 ret = -1;
2991 return ret;
2995 * Called by memory hotplug when all memory in a node is offlined.
2997 void kswapd_stop(int nid)
2999 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3001 if (kswapd)
3002 kthread_stop(kswapd);
3005 static int __init kswapd_init(void)
3007 int nid;
3009 swap_setup();
3010 for_each_node_state(nid, N_HIGH_MEMORY)
3011 kswapd_run(nid);
3012 hotcpu_notifier(cpu_callback, 0);
3013 return 0;
3016 module_init(kswapd_init)
3018 #ifdef CONFIG_NUMA
3020 * Zone reclaim mode
3022 * If non-zero call zone_reclaim when the number of free pages falls below
3023 * the watermarks.
3025 int zone_reclaim_mode __read_mostly;
3027 #define RECLAIM_OFF 0
3028 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3029 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3030 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3033 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3034 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3035 * a zone.
3037 #define ZONE_RECLAIM_PRIORITY 4
3040 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3041 * occur.
3043 int sysctl_min_unmapped_ratio = 1;
3046 * If the number of slab pages in a zone grows beyond this percentage then
3047 * slab reclaim needs to occur.
3049 int sysctl_min_slab_ratio = 5;
3051 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3053 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3054 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3055 zone_page_state(zone, NR_ACTIVE_FILE);
3058 * It's possible for there to be more file mapped pages than
3059 * accounted for by the pages on the file LRU lists because
3060 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3062 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3065 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3066 static long zone_pagecache_reclaimable(struct zone *zone)
3068 long nr_pagecache_reclaimable;
3069 long delta = 0;
3072 * If RECLAIM_SWAP is set, then all file pages are considered
3073 * potentially reclaimable. Otherwise, we have to worry about
3074 * pages like swapcache and zone_unmapped_file_pages() provides
3075 * a better estimate
3077 if (zone_reclaim_mode & RECLAIM_SWAP)
3078 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3079 else
3080 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3082 /* If we can't clean pages, remove dirty pages from consideration */
3083 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3084 delta += zone_page_state(zone, NR_FILE_DIRTY);
3086 /* Watch for any possible underflows due to delta */
3087 if (unlikely(delta > nr_pagecache_reclaimable))
3088 delta = nr_pagecache_reclaimable;
3090 return nr_pagecache_reclaimable - delta;
3094 * Try to free up some pages from this zone through reclaim.
3096 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3098 /* Minimum pages needed in order to stay on node */
3099 const unsigned long nr_pages = 1 << order;
3100 struct task_struct *p = current;
3101 struct reclaim_state reclaim_state;
3102 int priority;
3103 struct scan_control sc = {
3104 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3105 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3106 .may_swap = 1,
3107 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3108 SWAP_CLUSTER_MAX),
3109 .gfp_mask = gfp_mask,
3110 .order = order,
3112 struct shrink_control shrink = {
3113 .gfp_mask = sc.gfp_mask,
3115 unsigned long nr_slab_pages0, nr_slab_pages1;
3117 cond_resched();
3119 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3120 * and we also need to be able to write out pages for RECLAIM_WRITE
3121 * and RECLAIM_SWAP.
3123 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3124 lockdep_set_current_reclaim_state(gfp_mask);
3125 reclaim_state.reclaimed_slab = 0;
3126 p->reclaim_state = &reclaim_state;
3128 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3130 * Free memory by calling shrink zone with increasing
3131 * priorities until we have enough memory freed.
3133 priority = ZONE_RECLAIM_PRIORITY;
3134 do {
3135 shrink_zone(priority, zone, &sc);
3136 priority--;
3137 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3140 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3141 if (nr_slab_pages0 > zone->min_slab_pages) {
3143 * shrink_slab() does not currently allow us to determine how
3144 * many pages were freed in this zone. So we take the current
3145 * number of slab pages and shake the slab until it is reduced
3146 * by the same nr_pages that we used for reclaiming unmapped
3147 * pages.
3149 * Note that shrink_slab will free memory on all zones and may
3150 * take a long time.
3152 for (;;) {
3153 unsigned long lru_pages = zone_reclaimable_pages(zone);
3155 /* No reclaimable slab or very low memory pressure */
3156 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3157 break;
3159 /* Freed enough memory */
3160 nr_slab_pages1 = zone_page_state(zone,
3161 NR_SLAB_RECLAIMABLE);
3162 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3163 break;
3167 * Update nr_reclaimed by the number of slab pages we
3168 * reclaimed from this zone.
3170 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3171 if (nr_slab_pages1 < nr_slab_pages0)
3172 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3175 p->reclaim_state = NULL;
3176 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3177 lockdep_clear_current_reclaim_state();
3178 return sc.nr_reclaimed >= nr_pages;
3181 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3183 int node_id;
3184 int ret;
3187 * Zone reclaim reclaims unmapped file backed pages and
3188 * slab pages if we are over the defined limits.
3190 * A small portion of unmapped file backed pages is needed for
3191 * file I/O otherwise pages read by file I/O will be immediately
3192 * thrown out if the zone is overallocated. So we do not reclaim
3193 * if less than a specified percentage of the zone is used by
3194 * unmapped file backed pages.
3196 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3197 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3198 return ZONE_RECLAIM_FULL;
3200 if (zone->all_unreclaimable)
3201 return ZONE_RECLAIM_FULL;
3204 * Do not scan if the allocation should not be delayed.
3206 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3207 return ZONE_RECLAIM_NOSCAN;
3210 * Only run zone reclaim on the local zone or on zones that do not
3211 * have associated processors. This will favor the local processor
3212 * over remote processors and spread off node memory allocations
3213 * as wide as possible.
3215 node_id = zone_to_nid(zone);
3216 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3217 return ZONE_RECLAIM_NOSCAN;
3219 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3220 return ZONE_RECLAIM_NOSCAN;
3222 ret = __zone_reclaim(zone, gfp_mask, order);
3223 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3225 if (!ret)
3226 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3228 return ret;
3230 #endif
3233 * page_evictable - test whether a page is evictable
3234 * @page: the page to test
3235 * @vma: the VMA in which the page is or will be mapped, may be NULL
3237 * Test whether page is evictable--i.e., should be placed on active/inactive
3238 * lists vs unevictable list. The vma argument is !NULL when called from the
3239 * fault path to determine how to instantate a new page.
3241 * Reasons page might not be evictable:
3242 * (1) page's mapping marked unevictable
3243 * (2) page is part of an mlocked VMA
3246 int page_evictable(struct page *page, struct vm_area_struct *vma)
3249 if (mapping_unevictable(page_mapping(page)))
3250 return 0;
3252 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3253 return 0;
3255 return 1;
3259 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3260 * @page: page to check evictability and move to appropriate lru list
3261 * @zone: zone page is in
3263 * Checks a page for evictability and moves the page to the appropriate
3264 * zone lru list.
3266 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3267 * have PageUnevictable set.
3269 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3271 VM_BUG_ON(PageActive(page));
3273 retry:
3274 ClearPageUnevictable(page);
3275 if (page_evictable(page, NULL)) {
3276 enum lru_list l = page_lru_base_type(page);
3278 __dec_zone_state(zone, NR_UNEVICTABLE);
3279 list_move(&page->lru, &zone->lru[l].list);
3280 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3281 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3282 __count_vm_event(UNEVICTABLE_PGRESCUED);
3283 } else {
3285 * rotate unevictable list
3287 SetPageUnevictable(page);
3288 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3289 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3290 if (page_evictable(page, NULL))
3291 goto retry;
3296 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3297 * @mapping: struct address_space to scan for evictable pages
3299 * Scan all pages in mapping. Check unevictable pages for
3300 * evictability and move them to the appropriate zone lru list.
3302 void scan_mapping_unevictable_pages(struct address_space *mapping)
3304 pgoff_t next = 0;
3305 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3306 PAGE_CACHE_SHIFT;
3307 struct zone *zone;
3308 struct pagevec pvec;
3310 if (mapping->nrpages == 0)
3311 return;
3313 pagevec_init(&pvec, 0);
3314 while (next < end &&
3315 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3316 int i;
3317 int pg_scanned = 0;
3319 zone = NULL;
3321 for (i = 0; i < pagevec_count(&pvec); i++) {
3322 struct page *page = pvec.pages[i];
3323 pgoff_t page_index = page->index;
3324 struct zone *pagezone = page_zone(page);
3326 pg_scanned++;
3327 if (page_index > next)
3328 next = page_index;
3329 next++;
3331 if (pagezone != zone) {
3332 if (zone)
3333 spin_unlock_irq(&zone->lru_lock);
3334 zone = pagezone;
3335 spin_lock_irq(&zone->lru_lock);
3338 if (PageLRU(page) && PageUnevictable(page))
3339 check_move_unevictable_page(page, zone);
3341 if (zone)
3342 spin_unlock_irq(&zone->lru_lock);
3343 pagevec_release(&pvec);
3345 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3351 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3352 * @zone - zone of which to scan the unevictable list
3354 * Scan @zone's unevictable LRU lists to check for pages that have become
3355 * evictable. Move those that have to @zone's inactive list where they
3356 * become candidates for reclaim, unless shrink_inactive_zone() decides
3357 * to reactivate them. Pages that are still unevictable are rotated
3358 * back onto @zone's unevictable list.
3360 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3361 static void scan_zone_unevictable_pages(struct zone *zone)
3363 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3364 unsigned long scan;
3365 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3367 while (nr_to_scan > 0) {
3368 unsigned long batch_size = min(nr_to_scan,
3369 SCAN_UNEVICTABLE_BATCH_SIZE);
3371 spin_lock_irq(&zone->lru_lock);
3372 for (scan = 0; scan < batch_size; scan++) {
3373 struct page *page = lru_to_page(l_unevictable);
3375 if (!trylock_page(page))
3376 continue;
3378 prefetchw_prev_lru_page(page, l_unevictable, flags);
3380 if (likely(PageLRU(page) && PageUnevictable(page)))
3381 check_move_unevictable_page(page, zone);
3383 unlock_page(page);
3385 spin_unlock_irq(&zone->lru_lock);
3387 nr_to_scan -= batch_size;
3393 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3395 * A really big hammer: scan all zones' unevictable LRU lists to check for
3396 * pages that have become evictable. Move those back to the zones'
3397 * inactive list where they become candidates for reclaim.
3398 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3399 * and we add swap to the system. As such, it runs in the context of a task
3400 * that has possibly/probably made some previously unevictable pages
3401 * evictable.
3403 static void scan_all_zones_unevictable_pages(void)
3405 struct zone *zone;
3407 for_each_zone(zone) {
3408 scan_zone_unevictable_pages(zone);
3413 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3414 * all nodes' unevictable lists for evictable pages
3416 unsigned long scan_unevictable_pages;
3418 int scan_unevictable_handler(struct ctl_table *table, int write,
3419 void __user *buffer,
3420 size_t *length, loff_t *ppos)
3422 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3424 if (write && *(unsigned long *)table->data)
3425 scan_all_zones_unevictable_pages();
3427 scan_unevictable_pages = 0;
3428 return 0;
3431 #ifdef CONFIG_NUMA
3433 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3434 * a specified node's per zone unevictable lists for evictable pages.
3437 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3438 struct sysdev_attribute *attr,
3439 char *buf)
3441 return sprintf(buf, "0\n"); /* always zero; should fit... */
3444 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3445 struct sysdev_attribute *attr,
3446 const char *buf, size_t count)
3448 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3449 struct zone *zone;
3450 unsigned long res;
3451 unsigned long req = strict_strtoul(buf, 10, &res);
3453 if (!req)
3454 return 1; /* zero is no-op */
3456 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3457 if (!populated_zone(zone))
3458 continue;
3459 scan_zone_unevictable_pages(zone);
3461 return 1;
3465 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3466 read_scan_unevictable_node,
3467 write_scan_unevictable_node);
3469 int scan_unevictable_register_node(struct node *node)
3471 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3474 void scan_unevictable_unregister_node(struct node *node)
3476 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3478 #endif