crypto: skcipher - remove redundant NULL check
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / vmscan.c
blobf5d90dedebbafab2235c6d408a371d40e52d4e54
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>
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
48 #include <linux/swapops.h>
50 #include "internal.h"
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/vmscan.h>
56 * reclaim_mode determines how the inactive list is shrunk
57 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
58 * RECLAIM_MODE_ASYNC: Do not block
59 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
60 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
61 * page from the LRU and reclaim all pages within a
62 * naturally aligned range
63 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
64 * order-0 pages and then compact the zone
66 typedef unsigned __bitwise__ reclaim_mode_t;
67 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
68 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
69 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
70 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
71 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
73 struct scan_control {
74 /* Incremented by the number of inactive pages that were scanned */
75 unsigned long nr_scanned;
77 /* Number of pages freed so far during a call to shrink_zones() */
78 unsigned long nr_reclaimed;
80 /* How many pages shrink_list() should reclaim */
81 unsigned long nr_to_reclaim;
83 unsigned long hibernation_mode;
85 /* This context's GFP mask */
86 gfp_t gfp_mask;
88 int may_writepage;
90 /* Can mapped pages be reclaimed? */
91 int may_unmap;
93 /* Can pages be swapped as part of reclaim? */
94 int may_swap;
96 int swappiness;
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_pages(sc->mem_cgroup, zone, lru);
176 return zone_page_state(zone, NR_LRU_BASE + lru);
181 * Add a shrinker callback to be called from the vm
183 void register_shrinker(struct shrinker *shrinker)
185 shrinker->nr = 0;
186 down_write(&shrinker_rwsem);
187 list_add_tail(&shrinker->list, &shrinker_list);
188 up_write(&shrinker_rwsem);
190 EXPORT_SYMBOL(register_shrinker);
193 * Remove one
195 void unregister_shrinker(struct shrinker *shrinker)
197 down_write(&shrinker_rwsem);
198 list_del(&shrinker->list);
199 up_write(&shrinker_rwsem);
201 EXPORT_SYMBOL(unregister_shrinker);
203 #define SHRINK_BATCH 128
205 * Call the shrink functions to age shrinkable caches
207 * Here we assume it costs one seek to replace a lru page and that it also
208 * takes a seek to recreate a cache object. With this in mind we age equal
209 * percentages of the lru and ageable caches. This should balance the seeks
210 * generated by these structures.
212 * If the vm encountered mapped pages on the LRU it increase the pressure on
213 * slab to avoid swapping.
215 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
217 * `lru_pages' represents the number of on-LRU pages in all the zones which
218 * are eligible for the caller's allocation attempt. It is used for balancing
219 * slab reclaim versus page reclaim.
221 * Returns the number of slab objects which we shrunk.
223 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
224 unsigned long lru_pages)
226 struct shrinker *shrinker;
227 unsigned long ret = 0;
229 if (scanned == 0)
230 scanned = SWAP_CLUSTER_MAX;
232 if (!down_read_trylock(&shrinker_rwsem))
233 return 1; /* Assume we'll be able to shrink next time */
235 list_for_each_entry(shrinker, &shrinker_list, list) {
236 unsigned long long delta;
237 unsigned long total_scan;
238 unsigned long max_pass;
240 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
241 delta = (4 * scanned) / shrinker->seeks;
242 delta *= max_pass;
243 do_div(delta, lru_pages + 1);
244 shrinker->nr += delta;
245 if (shrinker->nr < 0) {
246 printk(KERN_ERR "shrink_slab: %pF negative objects to "
247 "delete nr=%ld\n",
248 shrinker->shrink, shrinker->nr);
249 shrinker->nr = max_pass;
253 * Avoid risking looping forever due to too large nr value:
254 * never try to free more than twice the estimate number of
255 * freeable entries.
257 if (shrinker->nr > max_pass * 2)
258 shrinker->nr = max_pass * 2;
260 total_scan = shrinker->nr;
261 shrinker->nr = 0;
263 while (total_scan >= SHRINK_BATCH) {
264 long this_scan = SHRINK_BATCH;
265 int shrink_ret;
266 int nr_before;
268 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
269 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
270 gfp_mask);
271 if (shrink_ret == -1)
272 break;
273 if (shrink_ret < nr_before)
274 ret += nr_before - shrink_ret;
275 count_vm_events(SLABS_SCANNED, this_scan);
276 total_scan -= this_scan;
278 cond_resched();
281 shrinker->nr += total_scan;
283 up_read(&shrinker_rwsem);
284 return ret;
287 static void set_reclaim_mode(int priority, struct scan_control *sc,
288 bool sync)
290 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
293 * Initially assume we are entering either lumpy reclaim or
294 * reclaim/compaction.Depending on the order, we will either set the
295 * sync mode or just reclaim order-0 pages later.
297 if (COMPACTION_BUILD)
298 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
299 else
300 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
303 * Avoid using lumpy reclaim or reclaim/compaction if possible by
304 * restricting when its set to either costly allocations or when
305 * under memory pressure
307 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
308 sc->reclaim_mode |= syncmode;
309 else if (sc->order && priority < DEF_PRIORITY - 2)
310 sc->reclaim_mode |= syncmode;
311 else
312 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
315 static void reset_reclaim_mode(struct scan_control *sc)
317 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
320 static inline int is_page_cache_freeable(struct page *page)
323 * A freeable page cache page is referenced only by the caller
324 * that isolated the page, the page cache radix tree and
325 * optional buffer heads at page->private.
327 return page_count(page) - page_has_private(page) == 2;
330 static int may_write_to_queue(struct backing_dev_info *bdi,
331 struct scan_control *sc)
333 if (current->flags & PF_SWAPWRITE)
334 return 1;
335 if (!bdi_write_congested(bdi))
336 return 1;
337 if (bdi == current->backing_dev_info)
338 return 1;
340 /* lumpy reclaim for hugepage often need a lot of write */
341 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
342 return 1;
343 return 0;
347 * We detected a synchronous write error writing a page out. Probably
348 * -ENOSPC. We need to propagate that into the address_space for a subsequent
349 * fsync(), msync() or close().
351 * The tricky part is that after writepage we cannot touch the mapping: nothing
352 * prevents it from being freed up. But we have a ref on the page and once
353 * that page is locked, the mapping is pinned.
355 * We're allowed to run sleeping lock_page() here because we know the caller has
356 * __GFP_FS.
358 static void handle_write_error(struct address_space *mapping,
359 struct page *page, int error)
361 lock_page_nosync(page);
362 if (page_mapping(page) == mapping)
363 mapping_set_error(mapping, error);
364 unlock_page(page);
367 /* possible outcome of pageout() */
368 typedef enum {
369 /* failed to write page out, page is locked */
370 PAGE_KEEP,
371 /* move page to the active list, page is locked */
372 PAGE_ACTIVATE,
373 /* page has been sent to the disk successfully, page is unlocked */
374 PAGE_SUCCESS,
375 /* page is clean and locked */
376 PAGE_CLEAN,
377 } pageout_t;
380 * pageout is called by shrink_page_list() for each dirty page.
381 * Calls ->writepage().
383 static pageout_t pageout(struct page *page, struct address_space *mapping,
384 struct scan_control *sc)
387 * If the page is dirty, only perform writeback if that write
388 * will be non-blocking. To prevent this allocation from being
389 * stalled by pagecache activity. But note that there may be
390 * stalls if we need to run get_block(). We could test
391 * PagePrivate for that.
393 * If this process is currently in __generic_file_aio_write() against
394 * this page's queue, we can perform writeback even if that
395 * will block.
397 * If the page is swapcache, write it back even if that would
398 * block, for some throttling. This happens by accident, because
399 * swap_backing_dev_info is bust: it doesn't reflect the
400 * congestion state of the swapdevs. Easy to fix, if needed.
402 if (!is_page_cache_freeable(page))
403 return PAGE_KEEP;
404 if (!mapping) {
406 * Some data journaling orphaned pages can have
407 * page->mapping == NULL while being dirty with clean buffers.
409 if (page_has_private(page)) {
410 if (try_to_free_buffers(page)) {
411 ClearPageDirty(page);
412 printk("%s: orphaned page\n", __func__);
413 return PAGE_CLEAN;
416 return PAGE_KEEP;
418 if (mapping->a_ops->writepage == NULL)
419 return PAGE_ACTIVATE;
420 if (!may_write_to_queue(mapping->backing_dev_info, sc))
421 return PAGE_KEEP;
423 if (clear_page_dirty_for_io(page)) {
424 int res;
425 struct writeback_control wbc = {
426 .sync_mode = WB_SYNC_NONE,
427 .nr_to_write = SWAP_CLUSTER_MAX,
428 .range_start = 0,
429 .range_end = LLONG_MAX,
430 .for_reclaim = 1,
433 SetPageReclaim(page);
434 res = mapping->a_ops->writepage(page, &wbc);
435 if (res < 0)
436 handle_write_error(mapping, page, res);
437 if (res == AOP_WRITEPAGE_ACTIVATE) {
438 ClearPageReclaim(page);
439 return PAGE_ACTIVATE;
443 * Wait on writeback if requested to. This happens when
444 * direct reclaiming a large contiguous area and the
445 * first attempt to free a range of pages fails.
447 if (PageWriteback(page) &&
448 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
449 wait_on_page_writeback(page);
451 if (!PageWriteback(page)) {
452 /* synchronous write or broken a_ops? */
453 ClearPageReclaim(page);
455 trace_mm_vmscan_writepage(page,
456 trace_reclaim_flags(page, sc->reclaim_mode));
457 inc_zone_page_state(page, NR_VMSCAN_WRITE);
458 return PAGE_SUCCESS;
461 return PAGE_CLEAN;
465 * Same as remove_mapping, but if the page is removed from the mapping, it
466 * gets returned with a refcount of 0.
468 static int __remove_mapping(struct address_space *mapping, struct page *page)
470 BUG_ON(!PageLocked(page));
471 BUG_ON(mapping != page_mapping(page));
473 spin_lock_irq(&mapping->tree_lock);
475 * The non racy check for a busy page.
477 * Must be careful with the order of the tests. When someone has
478 * a ref to the page, it may be possible that they dirty it then
479 * drop the reference. So if PageDirty is tested before page_count
480 * here, then the following race may occur:
482 * get_user_pages(&page);
483 * [user mapping goes away]
484 * write_to(page);
485 * !PageDirty(page) [good]
486 * SetPageDirty(page);
487 * put_page(page);
488 * !page_count(page) [good, discard it]
490 * [oops, our write_to data is lost]
492 * Reversing the order of the tests ensures such a situation cannot
493 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
494 * load is not satisfied before that of page->_count.
496 * Note that if SetPageDirty is always performed via set_page_dirty,
497 * and thus under tree_lock, then this ordering is not required.
499 if (!page_freeze_refs(page, 2))
500 goto cannot_free;
501 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
502 if (unlikely(PageDirty(page))) {
503 page_unfreeze_refs(page, 2);
504 goto cannot_free;
507 if (PageSwapCache(page)) {
508 swp_entry_t swap = { .val = page_private(page) };
509 __delete_from_swap_cache(page);
510 spin_unlock_irq(&mapping->tree_lock);
511 swapcache_free(swap, page);
512 } else {
513 void (*freepage)(struct page *);
515 freepage = mapping->a_ops->freepage;
517 __remove_from_page_cache(page);
518 spin_unlock_irq(&mapping->tree_lock);
519 mem_cgroup_uncharge_cache_page(page);
521 if (freepage != NULL)
522 freepage(page);
525 return 1;
527 cannot_free:
528 spin_unlock_irq(&mapping->tree_lock);
529 return 0;
533 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
534 * someone else has a ref on the page, abort and return 0. If it was
535 * successfully detached, return 1. Assumes the caller has a single ref on
536 * this page.
538 int remove_mapping(struct address_space *mapping, struct page *page)
540 if (__remove_mapping(mapping, page)) {
542 * Unfreezing the refcount with 1 rather than 2 effectively
543 * drops the pagecache ref for us without requiring another
544 * atomic operation.
546 page_unfreeze_refs(page, 1);
547 return 1;
549 return 0;
553 * putback_lru_page - put previously isolated page onto appropriate LRU list
554 * @page: page to be put back to appropriate lru list
556 * Add previously isolated @page to appropriate LRU list.
557 * Page may still be unevictable for other reasons.
559 * lru_lock must not be held, interrupts must be enabled.
561 void putback_lru_page(struct page *page)
563 int lru;
564 int active = !!TestClearPageActive(page);
565 int was_unevictable = PageUnevictable(page);
567 VM_BUG_ON(PageLRU(page));
569 redo:
570 ClearPageUnevictable(page);
572 if (page_evictable(page, NULL)) {
574 * For evictable pages, we can use the cache.
575 * In event of a race, worst case is we end up with an
576 * unevictable page on [in]active list.
577 * We know how to handle that.
579 lru = active + page_lru_base_type(page);
580 lru_cache_add_lru(page, lru);
581 } else {
583 * Put unevictable pages directly on zone's unevictable
584 * list.
586 lru = LRU_UNEVICTABLE;
587 add_page_to_unevictable_list(page);
589 * When racing with an mlock clearing (page is
590 * unlocked), make sure that if the other thread does
591 * not observe our setting of PG_lru and fails
592 * isolation, we see PG_mlocked cleared below and move
593 * the page back to the evictable list.
595 * The other side is TestClearPageMlocked().
597 smp_mb();
601 * page's status can change while we move it among lru. If an evictable
602 * page is on unevictable list, it never be freed. To avoid that,
603 * check after we added it to the list, again.
605 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606 if (!isolate_lru_page(page)) {
607 put_page(page);
608 goto redo;
610 /* This means someone else dropped this page from LRU
611 * So, it will be freed or putback to LRU again. There is
612 * nothing to do here.
616 if (was_unevictable && lru != LRU_UNEVICTABLE)
617 count_vm_event(UNEVICTABLE_PGRESCUED);
618 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619 count_vm_event(UNEVICTABLE_PGCULLED);
621 put_page(page); /* drop ref from isolate */
624 enum page_references {
625 PAGEREF_RECLAIM,
626 PAGEREF_RECLAIM_CLEAN,
627 PAGEREF_KEEP,
628 PAGEREF_ACTIVATE,
631 static enum page_references page_check_references(struct page *page,
632 struct scan_control *sc)
634 int referenced_ptes, referenced_page;
635 unsigned long vm_flags;
637 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
638 referenced_page = TestClearPageReferenced(page);
640 /* Lumpy reclaim - ignore references */
641 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
642 return PAGEREF_RECLAIM;
645 * Mlock lost the isolation race with us. Let try_to_unmap()
646 * move the page to the unevictable list.
648 if (vm_flags & VM_LOCKED)
649 return PAGEREF_RECLAIM;
651 if (referenced_ptes) {
652 if (PageAnon(page))
653 return PAGEREF_ACTIVATE;
655 * All mapped pages start out with page table
656 * references from the instantiating fault, so we need
657 * to look twice if a mapped file page is used more
658 * than once.
660 * Mark it and spare it for another trip around the
661 * inactive list. Another page table reference will
662 * lead to its activation.
664 * Note: the mark is set for activated pages as well
665 * so that recently deactivated but used pages are
666 * quickly recovered.
668 SetPageReferenced(page);
670 if (referenced_page)
671 return PAGEREF_ACTIVATE;
673 return PAGEREF_KEEP;
676 /* Reclaim if clean, defer dirty pages to writeback */
677 if (referenced_page && !PageSwapBacked(page))
678 return PAGEREF_RECLAIM_CLEAN;
680 return PAGEREF_RECLAIM;
683 static noinline_for_stack void free_page_list(struct list_head *free_pages)
685 struct pagevec freed_pvec;
686 struct page *page, *tmp;
688 pagevec_init(&freed_pvec, 1);
690 list_for_each_entry_safe(page, tmp, free_pages, lru) {
691 list_del(&page->lru);
692 if (!pagevec_add(&freed_pvec, page)) {
693 __pagevec_free(&freed_pvec);
694 pagevec_reinit(&freed_pvec);
698 pagevec_free(&freed_pvec);
702 * shrink_page_list() returns the number of reclaimed pages
704 static unsigned long shrink_page_list(struct list_head *page_list,
705 struct zone *zone,
706 struct scan_control *sc)
708 LIST_HEAD(ret_pages);
709 LIST_HEAD(free_pages);
710 int pgactivate = 0;
711 unsigned long nr_dirty = 0;
712 unsigned long nr_congested = 0;
713 unsigned long nr_reclaimed = 0;
715 cond_resched();
717 while (!list_empty(page_list)) {
718 enum page_references references;
719 struct address_space *mapping;
720 struct page *page;
721 int may_enter_fs;
723 cond_resched();
725 page = lru_to_page(page_list);
726 list_del(&page->lru);
728 if (!trylock_page(page))
729 goto keep;
731 VM_BUG_ON(PageActive(page));
732 VM_BUG_ON(page_zone(page) != zone);
734 sc->nr_scanned++;
736 if (unlikely(!page_evictable(page, NULL)))
737 goto cull_mlocked;
739 if (!sc->may_unmap && page_mapped(page))
740 goto keep_locked;
742 /* Double the slab pressure for mapped and swapcache pages */
743 if (page_mapped(page) || PageSwapCache(page))
744 sc->nr_scanned++;
746 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
749 if (PageWriteback(page)) {
751 * Synchronous reclaim is performed in two passes,
752 * first an asynchronous pass over the list to
753 * start parallel writeback, and a second synchronous
754 * pass to wait for the IO to complete. Wait here
755 * for any page for which writeback has already
756 * started.
758 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
759 may_enter_fs)
760 wait_on_page_writeback(page);
761 else {
762 unlock_page(page);
763 goto keep_lumpy;
767 references = page_check_references(page, sc);
768 switch (references) {
769 case PAGEREF_ACTIVATE:
770 goto activate_locked;
771 case PAGEREF_KEEP:
772 goto keep_locked;
773 case PAGEREF_RECLAIM:
774 case PAGEREF_RECLAIM_CLEAN:
775 ; /* try to reclaim the page below */
779 * Anonymous process memory has backing store?
780 * Try to allocate it some swap space here.
782 if (PageAnon(page) && !PageSwapCache(page)) {
783 if (!(sc->gfp_mask & __GFP_IO))
784 goto keep_locked;
785 if (!add_to_swap(page))
786 goto activate_locked;
787 may_enter_fs = 1;
790 mapping = page_mapping(page);
793 * The page is mapped into the page tables of one or more
794 * processes. Try to unmap it here.
796 if (page_mapped(page) && mapping) {
797 switch (try_to_unmap(page, TTU_UNMAP)) {
798 case SWAP_FAIL:
799 goto activate_locked;
800 case SWAP_AGAIN:
801 goto keep_locked;
802 case SWAP_MLOCK:
803 goto cull_mlocked;
804 case SWAP_SUCCESS:
805 ; /* try to free the page below */
809 if (PageDirty(page)) {
810 nr_dirty++;
812 if (references == PAGEREF_RECLAIM_CLEAN)
813 goto keep_locked;
814 if (!may_enter_fs)
815 goto keep_locked;
816 if (!sc->may_writepage)
817 goto keep_locked;
819 /* Page is dirty, try to write it out here */
820 switch (pageout(page, mapping, sc)) {
821 case PAGE_KEEP:
822 nr_congested++;
823 goto keep_locked;
824 case PAGE_ACTIVATE:
825 goto activate_locked;
826 case PAGE_SUCCESS:
827 if (PageWriteback(page))
828 goto keep_lumpy;
829 if (PageDirty(page))
830 goto keep;
833 * A synchronous write - probably a ramdisk. Go
834 * ahead and try to reclaim the page.
836 if (!trylock_page(page))
837 goto keep;
838 if (PageDirty(page) || PageWriteback(page))
839 goto keep_locked;
840 mapping = page_mapping(page);
841 case PAGE_CLEAN:
842 ; /* try to free the page below */
847 * If the page has buffers, try to free the buffer mappings
848 * associated with this page. If we succeed we try to free
849 * the page as well.
851 * We do this even if the page is PageDirty().
852 * try_to_release_page() does not perform I/O, but it is
853 * possible for a page to have PageDirty set, but it is actually
854 * clean (all its buffers are clean). This happens if the
855 * buffers were written out directly, with submit_bh(). ext3
856 * will do this, as well as the blockdev mapping.
857 * try_to_release_page() will discover that cleanness and will
858 * drop the buffers and mark the page clean - it can be freed.
860 * Rarely, pages can have buffers and no ->mapping. These are
861 * the pages which were not successfully invalidated in
862 * truncate_complete_page(). We try to drop those buffers here
863 * and if that worked, and the page is no longer mapped into
864 * process address space (page_count == 1) it can be freed.
865 * Otherwise, leave the page on the LRU so it is swappable.
867 if (page_has_private(page)) {
868 if (!try_to_release_page(page, sc->gfp_mask))
869 goto activate_locked;
870 if (!mapping && page_count(page) == 1) {
871 unlock_page(page);
872 if (put_page_testzero(page))
873 goto free_it;
874 else {
876 * rare race with speculative reference.
877 * the speculative reference will free
878 * this page shortly, so we may
879 * increment nr_reclaimed here (and
880 * leave it off the LRU).
882 nr_reclaimed++;
883 continue;
888 if (!mapping || !__remove_mapping(mapping, page))
889 goto keep_locked;
892 * At this point, we have no other references and there is
893 * no way to pick any more up (removed from LRU, removed
894 * from pagecache). Can use non-atomic bitops now (and
895 * we obviously don't have to worry about waking up a process
896 * waiting on the page lock, because there are no references.
898 __clear_page_locked(page);
899 free_it:
900 nr_reclaimed++;
903 * Is there need to periodically free_page_list? It would
904 * appear not as the counts should be low
906 list_add(&page->lru, &free_pages);
907 continue;
909 cull_mlocked:
910 if (PageSwapCache(page))
911 try_to_free_swap(page);
912 unlock_page(page);
913 putback_lru_page(page);
914 reset_reclaim_mode(sc);
915 continue;
917 activate_locked:
918 /* Not a candidate for swapping, so reclaim swap space. */
919 if (PageSwapCache(page) && vm_swap_full())
920 try_to_free_swap(page);
921 VM_BUG_ON(PageActive(page));
922 SetPageActive(page);
923 pgactivate++;
924 keep_locked:
925 unlock_page(page);
926 keep:
927 reset_reclaim_mode(sc);
928 keep_lumpy:
929 list_add(&page->lru, &ret_pages);
930 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
934 * Tag a zone as congested if all the dirty pages encountered were
935 * backed by a congested BDI. In this case, reclaimers should just
936 * back off and wait for congestion to clear because further reclaim
937 * will encounter the same problem
939 if (nr_dirty == nr_congested && nr_dirty != 0)
940 zone_set_flag(zone, ZONE_CONGESTED);
942 free_page_list(&free_pages);
944 list_splice(&ret_pages, page_list);
945 count_vm_events(PGACTIVATE, pgactivate);
946 return nr_reclaimed;
950 * Attempt to remove the specified page from its LRU. Only take this page
951 * if it is of the appropriate PageActive status. Pages which are being
952 * freed elsewhere are also ignored.
954 * page: page to consider
955 * mode: one of the LRU isolation modes defined above
957 * returns 0 on success, -ve errno on failure.
959 int __isolate_lru_page(struct page *page, int mode, int file)
961 int ret = -EINVAL;
963 /* Only take pages on the LRU. */
964 if (!PageLRU(page))
965 return ret;
968 * When checking the active state, we need to be sure we are
969 * dealing with comparible boolean values. Take the logical not
970 * of each.
972 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
973 return ret;
975 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
976 return ret;
979 * When this function is being called for lumpy reclaim, we
980 * initially look into all LRU pages, active, inactive and
981 * unevictable; only give shrink_page_list evictable pages.
983 if (PageUnevictable(page))
984 return ret;
986 ret = -EBUSY;
988 if (likely(get_page_unless_zero(page))) {
990 * Be careful not to clear PageLRU until after we're
991 * sure the page is not being freed elsewhere -- the
992 * page release code relies on it.
994 ClearPageLRU(page);
995 ret = 0;
998 return ret;
1002 * zone->lru_lock is heavily contended. Some of the functions that
1003 * shrink the lists perform better by taking out a batch of pages
1004 * and working on them outside the LRU lock.
1006 * For pagecache intensive workloads, this function is the hottest
1007 * spot in the kernel (apart from copy_*_user functions).
1009 * Appropriate locks must be held before calling this function.
1011 * @nr_to_scan: The number of pages to look through on the list.
1012 * @src: The LRU list to pull pages off.
1013 * @dst: The temp list to put pages on to.
1014 * @scanned: The number of pages that were scanned.
1015 * @order: The caller's attempted allocation order
1016 * @mode: One of the LRU isolation modes
1017 * @file: True [1] if isolating file [!anon] pages
1019 * returns how many pages were moved onto *@dst.
1021 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1022 struct list_head *src, struct list_head *dst,
1023 unsigned long *scanned, int order, int mode, int file)
1025 unsigned long nr_taken = 0;
1026 unsigned long nr_lumpy_taken = 0;
1027 unsigned long nr_lumpy_dirty = 0;
1028 unsigned long nr_lumpy_failed = 0;
1029 unsigned long scan;
1031 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1032 struct page *page;
1033 unsigned long pfn;
1034 unsigned long end_pfn;
1035 unsigned long page_pfn;
1036 int zone_id;
1038 page = lru_to_page(src);
1039 prefetchw_prev_lru_page(page, src, flags);
1041 VM_BUG_ON(!PageLRU(page));
1043 switch (__isolate_lru_page(page, mode, file)) {
1044 case 0:
1045 list_move(&page->lru, dst);
1046 mem_cgroup_del_lru(page);
1047 nr_taken += hpage_nr_pages(page);
1048 break;
1050 case -EBUSY:
1051 /* else it is being freed elsewhere */
1052 list_move(&page->lru, src);
1053 mem_cgroup_rotate_lru_list(page, page_lru(page));
1054 continue;
1056 default:
1057 BUG();
1060 if (!order)
1061 continue;
1064 * Attempt to take all pages in the order aligned region
1065 * surrounding the tag page. Only take those pages of
1066 * the same active state as that tag page. We may safely
1067 * round the target page pfn down to the requested order
1068 * as the mem_map is guarenteed valid out to MAX_ORDER,
1069 * where that page is in a different zone we will detect
1070 * it from its zone id and abort this block scan.
1072 zone_id = page_zone_id(page);
1073 page_pfn = page_to_pfn(page);
1074 pfn = page_pfn & ~((1 << order) - 1);
1075 end_pfn = pfn + (1 << order);
1076 for (; pfn < end_pfn; pfn++) {
1077 struct page *cursor_page;
1079 /* The target page is in the block, ignore it. */
1080 if (unlikely(pfn == page_pfn))
1081 continue;
1083 /* Avoid holes within the zone. */
1084 if (unlikely(!pfn_valid_within(pfn)))
1085 break;
1087 cursor_page = pfn_to_page(pfn);
1089 /* Check that we have not crossed a zone boundary. */
1090 if (unlikely(page_zone_id(cursor_page) != zone_id))
1091 break;
1094 * If we don't have enough swap space, reclaiming of
1095 * anon page which don't already have a swap slot is
1096 * pointless.
1098 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1099 !PageSwapCache(cursor_page))
1100 break;
1102 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1103 list_move(&cursor_page->lru, dst);
1104 mem_cgroup_del_lru(cursor_page);
1105 nr_taken += hpage_nr_pages(page);
1106 nr_lumpy_taken++;
1107 if (PageDirty(cursor_page))
1108 nr_lumpy_dirty++;
1109 scan++;
1110 } else {
1111 /* the page is freed already. */
1112 if (!page_count(cursor_page))
1113 continue;
1114 break;
1118 /* If we break out of the loop above, lumpy reclaim failed */
1119 if (pfn < end_pfn)
1120 nr_lumpy_failed++;
1123 *scanned = scan;
1125 trace_mm_vmscan_lru_isolate(order,
1126 nr_to_scan, scan,
1127 nr_taken,
1128 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1129 mode);
1130 return nr_taken;
1133 static unsigned long isolate_pages_global(unsigned long nr,
1134 struct list_head *dst,
1135 unsigned long *scanned, int order,
1136 int mode, struct zone *z,
1137 int active, int file)
1139 int lru = LRU_BASE;
1140 if (active)
1141 lru += LRU_ACTIVE;
1142 if (file)
1143 lru += LRU_FILE;
1144 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1145 mode, file);
1149 * clear_active_flags() is a helper for shrink_active_list(), clearing
1150 * any active bits from the pages in the list.
1152 static unsigned long clear_active_flags(struct list_head *page_list,
1153 unsigned int *count)
1155 int nr_active = 0;
1156 int lru;
1157 struct page *page;
1159 list_for_each_entry(page, page_list, lru) {
1160 int numpages = hpage_nr_pages(page);
1161 lru = page_lru_base_type(page);
1162 if (PageActive(page)) {
1163 lru += LRU_ACTIVE;
1164 ClearPageActive(page);
1165 nr_active += numpages;
1167 if (count)
1168 count[lru] += numpages;
1171 return nr_active;
1175 * isolate_lru_page - tries to isolate a page from its LRU list
1176 * @page: page to isolate from its LRU list
1178 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1179 * vmstat statistic corresponding to whatever LRU list the page was on.
1181 * Returns 0 if the page was removed from an LRU list.
1182 * Returns -EBUSY if the page was not on an LRU list.
1184 * The returned page will have PageLRU() cleared. If it was found on
1185 * the active list, it will have PageActive set. If it was found on
1186 * the unevictable list, it will have the PageUnevictable bit set. That flag
1187 * may need to be cleared by the caller before letting the page go.
1189 * The vmstat statistic corresponding to the list on which the page was
1190 * found will be decremented.
1192 * Restrictions:
1193 * (1) Must be called with an elevated refcount on the page. This is a
1194 * fundamentnal difference from isolate_lru_pages (which is called
1195 * without a stable reference).
1196 * (2) the lru_lock must not be held.
1197 * (3) interrupts must be enabled.
1199 int isolate_lru_page(struct page *page)
1201 int ret = -EBUSY;
1203 if (PageLRU(page)) {
1204 struct zone *zone = page_zone(page);
1206 spin_lock_irq(&zone->lru_lock);
1207 if (PageLRU(page) && get_page_unless_zero(page)) {
1208 int lru = page_lru(page);
1209 ret = 0;
1210 ClearPageLRU(page);
1212 del_page_from_lru_list(zone, page, lru);
1214 spin_unlock_irq(&zone->lru_lock);
1216 return ret;
1220 * Are there way too many processes in the direct reclaim path already?
1222 static int too_many_isolated(struct zone *zone, int file,
1223 struct scan_control *sc)
1225 unsigned long inactive, isolated;
1227 if (current_is_kswapd())
1228 return 0;
1230 if (!scanning_global_lru(sc))
1231 return 0;
1233 if (file) {
1234 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1235 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1236 } else {
1237 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1238 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1241 return isolated > inactive;
1245 * TODO: Try merging with migrations version of putback_lru_pages
1247 static noinline_for_stack void
1248 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1249 unsigned long nr_anon, unsigned long nr_file,
1250 struct list_head *page_list)
1252 struct page *page;
1253 struct pagevec pvec;
1254 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1256 pagevec_init(&pvec, 1);
1259 * Put back any unfreeable pages.
1261 spin_lock(&zone->lru_lock);
1262 while (!list_empty(page_list)) {
1263 int lru;
1264 page = lru_to_page(page_list);
1265 VM_BUG_ON(PageLRU(page));
1266 list_del(&page->lru);
1267 if (unlikely(!page_evictable(page, NULL))) {
1268 spin_unlock_irq(&zone->lru_lock);
1269 putback_lru_page(page);
1270 spin_lock_irq(&zone->lru_lock);
1271 continue;
1273 SetPageLRU(page);
1274 lru = page_lru(page);
1275 add_page_to_lru_list(zone, page, lru);
1276 if (is_active_lru(lru)) {
1277 int file = is_file_lru(lru);
1278 int numpages = hpage_nr_pages(page);
1279 reclaim_stat->recent_rotated[file] += numpages;
1281 if (!pagevec_add(&pvec, page)) {
1282 spin_unlock_irq(&zone->lru_lock);
1283 __pagevec_release(&pvec);
1284 spin_lock_irq(&zone->lru_lock);
1287 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1288 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1290 spin_unlock_irq(&zone->lru_lock);
1291 pagevec_release(&pvec);
1294 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1295 struct scan_control *sc,
1296 unsigned long *nr_anon,
1297 unsigned long *nr_file,
1298 struct list_head *isolated_list)
1300 unsigned long nr_active;
1301 unsigned int count[NR_LRU_LISTS] = { 0, };
1302 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1304 nr_active = clear_active_flags(isolated_list, count);
1305 __count_vm_events(PGDEACTIVATE, nr_active);
1307 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1308 -count[LRU_ACTIVE_FILE]);
1309 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1310 -count[LRU_INACTIVE_FILE]);
1311 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1312 -count[LRU_ACTIVE_ANON]);
1313 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1314 -count[LRU_INACTIVE_ANON]);
1316 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1317 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1318 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1319 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1321 reclaim_stat->recent_scanned[0] += *nr_anon;
1322 reclaim_stat->recent_scanned[1] += *nr_file;
1326 * Returns true if the caller should wait to clean dirty/writeback pages.
1328 * If we are direct reclaiming for contiguous pages and we do not reclaim
1329 * everything in the list, try again and wait for writeback IO to complete.
1330 * This will stall high-order allocations noticeably. Only do that when really
1331 * need to free the pages under high memory pressure.
1333 static inline bool should_reclaim_stall(unsigned long nr_taken,
1334 unsigned long nr_freed,
1335 int priority,
1336 struct scan_control *sc)
1338 int lumpy_stall_priority;
1340 /* kswapd should not stall on sync IO */
1341 if (current_is_kswapd())
1342 return false;
1344 /* Only stall on lumpy reclaim */
1345 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1346 return false;
1348 /* If we have relaimed everything on the isolated list, no stall */
1349 if (nr_freed == nr_taken)
1350 return false;
1353 * For high-order allocations, there are two stall thresholds.
1354 * High-cost allocations stall immediately where as lower
1355 * order allocations such as stacks require the scanning
1356 * priority to be much higher before stalling.
1358 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1359 lumpy_stall_priority = DEF_PRIORITY;
1360 else
1361 lumpy_stall_priority = DEF_PRIORITY / 3;
1363 return priority <= lumpy_stall_priority;
1367 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1368 * of reclaimed pages
1370 static noinline_for_stack unsigned long
1371 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1372 struct scan_control *sc, int priority, int file)
1374 LIST_HEAD(page_list);
1375 unsigned long nr_scanned;
1376 unsigned long nr_reclaimed = 0;
1377 unsigned long nr_taken;
1378 unsigned long nr_anon;
1379 unsigned long nr_file;
1381 while (unlikely(too_many_isolated(zone, file, sc))) {
1382 congestion_wait(BLK_RW_ASYNC, HZ/10);
1384 /* We are about to die and free our memory. Return now. */
1385 if (fatal_signal_pending(current))
1386 return SWAP_CLUSTER_MAX;
1389 set_reclaim_mode(priority, sc, false);
1390 lru_add_drain();
1391 spin_lock_irq(&zone->lru_lock);
1393 if (scanning_global_lru(sc)) {
1394 nr_taken = isolate_pages_global(nr_to_scan,
1395 &page_list, &nr_scanned, sc->order,
1396 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1397 ISOLATE_BOTH : ISOLATE_INACTIVE,
1398 zone, 0, file);
1399 zone->pages_scanned += nr_scanned;
1400 if (current_is_kswapd())
1401 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1402 nr_scanned);
1403 else
1404 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1405 nr_scanned);
1406 } else {
1407 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1408 &page_list, &nr_scanned, sc->order,
1409 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1410 ISOLATE_BOTH : ISOLATE_INACTIVE,
1411 zone, sc->mem_cgroup,
1412 0, file);
1414 * mem_cgroup_isolate_pages() keeps track of
1415 * scanned pages on its own.
1419 if (nr_taken == 0) {
1420 spin_unlock_irq(&zone->lru_lock);
1421 return 0;
1424 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1426 spin_unlock_irq(&zone->lru_lock);
1428 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1430 /* Check if we should syncronously wait for writeback */
1431 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1432 set_reclaim_mode(priority, sc, true);
1433 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1436 local_irq_disable();
1437 if (current_is_kswapd())
1438 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1439 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1441 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1443 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1444 zone_idx(zone),
1445 nr_scanned, nr_reclaimed,
1446 priority,
1447 trace_shrink_flags(file, sc->reclaim_mode));
1448 return nr_reclaimed;
1452 * This moves pages from the active list to the inactive list.
1454 * We move them the other way if the page is referenced by one or more
1455 * processes, from rmap.
1457 * If the pages are mostly unmapped, the processing is fast and it is
1458 * appropriate to hold zone->lru_lock across the whole operation. But if
1459 * the pages are mapped, the processing is slow (page_referenced()) so we
1460 * should drop zone->lru_lock around each page. It's impossible to balance
1461 * this, so instead we remove the pages from the LRU while processing them.
1462 * It is safe to rely on PG_active against the non-LRU pages in here because
1463 * nobody will play with that bit on a non-LRU page.
1465 * The downside is that we have to touch page->_count against each page.
1466 * But we had to alter page->flags anyway.
1469 static void move_active_pages_to_lru(struct zone *zone,
1470 struct list_head *list,
1471 enum lru_list lru)
1473 unsigned long pgmoved = 0;
1474 struct pagevec pvec;
1475 struct page *page;
1477 pagevec_init(&pvec, 1);
1479 while (!list_empty(list)) {
1480 page = lru_to_page(list);
1482 VM_BUG_ON(PageLRU(page));
1483 SetPageLRU(page);
1485 list_move(&page->lru, &zone->lru[lru].list);
1486 mem_cgroup_add_lru_list(page, lru);
1487 pgmoved += hpage_nr_pages(page);
1489 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1490 spin_unlock_irq(&zone->lru_lock);
1491 if (buffer_heads_over_limit)
1492 pagevec_strip(&pvec);
1493 __pagevec_release(&pvec);
1494 spin_lock_irq(&zone->lru_lock);
1497 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1498 if (!is_active_lru(lru))
1499 __count_vm_events(PGDEACTIVATE, pgmoved);
1502 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1503 struct scan_control *sc, int priority, int file)
1505 unsigned long nr_taken;
1506 unsigned long pgscanned;
1507 unsigned long vm_flags;
1508 LIST_HEAD(l_hold); /* The pages which were snipped off */
1509 LIST_HEAD(l_active);
1510 LIST_HEAD(l_inactive);
1511 struct page *page;
1512 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1513 unsigned long nr_rotated = 0;
1515 lru_add_drain();
1516 spin_lock_irq(&zone->lru_lock);
1517 if (scanning_global_lru(sc)) {
1518 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1519 &pgscanned, sc->order,
1520 ISOLATE_ACTIVE, zone,
1521 1, file);
1522 zone->pages_scanned += pgscanned;
1523 } else {
1524 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1525 &pgscanned, sc->order,
1526 ISOLATE_ACTIVE, zone,
1527 sc->mem_cgroup, 1, file);
1529 * mem_cgroup_isolate_pages() keeps track of
1530 * scanned pages on its own.
1534 reclaim_stat->recent_scanned[file] += nr_taken;
1536 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1537 if (file)
1538 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1539 else
1540 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1541 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1542 spin_unlock_irq(&zone->lru_lock);
1544 while (!list_empty(&l_hold)) {
1545 cond_resched();
1546 page = lru_to_page(&l_hold);
1547 list_del(&page->lru);
1549 if (unlikely(!page_evictable(page, NULL))) {
1550 putback_lru_page(page);
1551 continue;
1554 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1555 nr_rotated += hpage_nr_pages(page);
1557 * Identify referenced, file-backed active pages and
1558 * give them one more trip around the active list. So
1559 * that executable code get better chances to stay in
1560 * memory under moderate memory pressure. Anon pages
1561 * are not likely to be evicted by use-once streaming
1562 * IO, plus JVM can create lots of anon VM_EXEC pages,
1563 * so we ignore them here.
1565 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1566 list_add(&page->lru, &l_active);
1567 continue;
1571 ClearPageActive(page); /* we are de-activating */
1572 list_add(&page->lru, &l_inactive);
1576 * Move pages back to the lru list.
1578 spin_lock_irq(&zone->lru_lock);
1580 * Count referenced pages from currently used mappings as rotated,
1581 * even though only some of them are actually re-activated. This
1582 * helps balance scan pressure between file and anonymous pages in
1583 * get_scan_ratio.
1585 reclaim_stat->recent_rotated[file] += nr_rotated;
1587 move_active_pages_to_lru(zone, &l_active,
1588 LRU_ACTIVE + file * LRU_FILE);
1589 move_active_pages_to_lru(zone, &l_inactive,
1590 LRU_BASE + file * LRU_FILE);
1591 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1592 spin_unlock_irq(&zone->lru_lock);
1595 #ifdef CONFIG_SWAP
1596 static int inactive_anon_is_low_global(struct zone *zone)
1598 unsigned long active, inactive;
1600 active = zone_page_state(zone, NR_ACTIVE_ANON);
1601 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1603 if (inactive * zone->inactive_ratio < active)
1604 return 1;
1606 return 0;
1610 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1611 * @zone: zone to check
1612 * @sc: scan control of this context
1614 * Returns true if the zone does not have enough inactive anon pages,
1615 * meaning some active anon pages need to be deactivated.
1617 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1619 int low;
1622 * If we don't have swap space, anonymous page deactivation
1623 * is pointless.
1625 if (!total_swap_pages)
1626 return 0;
1628 if (scanning_global_lru(sc))
1629 low = inactive_anon_is_low_global(zone);
1630 else
1631 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1632 return low;
1634 #else
1635 static inline int inactive_anon_is_low(struct zone *zone,
1636 struct scan_control *sc)
1638 return 0;
1640 #endif
1642 static int inactive_file_is_low_global(struct zone *zone)
1644 unsigned long active, inactive;
1646 active = zone_page_state(zone, NR_ACTIVE_FILE);
1647 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1649 return (active > inactive);
1653 * inactive_file_is_low - check if file pages need to be deactivated
1654 * @zone: zone to check
1655 * @sc: scan control of this context
1657 * When the system is doing streaming IO, memory pressure here
1658 * ensures that active file pages get deactivated, until more
1659 * than half of the file pages are on the inactive list.
1661 * Once we get to that situation, protect the system's working
1662 * set from being evicted by disabling active file page aging.
1664 * This uses a different ratio than the anonymous pages, because
1665 * the page cache uses a use-once replacement algorithm.
1667 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1669 int low;
1671 if (scanning_global_lru(sc))
1672 low = inactive_file_is_low_global(zone);
1673 else
1674 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1675 return low;
1678 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1679 int file)
1681 if (file)
1682 return inactive_file_is_low(zone, sc);
1683 else
1684 return inactive_anon_is_low(zone, sc);
1687 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1688 struct zone *zone, struct scan_control *sc, int priority)
1690 int file = is_file_lru(lru);
1692 if (is_active_lru(lru)) {
1693 if (inactive_list_is_low(zone, sc, file))
1694 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1695 return 0;
1698 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1702 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1703 * until we collected @swap_cluster_max pages to scan.
1705 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1706 unsigned long *nr_saved_scan)
1708 unsigned long nr;
1710 *nr_saved_scan += nr_to_scan;
1711 nr = *nr_saved_scan;
1713 if (nr >= SWAP_CLUSTER_MAX)
1714 *nr_saved_scan = 0;
1715 else
1716 nr = 0;
1718 return nr;
1722 * Determine how aggressively the anon and file LRU lists should be
1723 * scanned. The relative value of each set of LRU lists is determined
1724 * by looking at the fraction of the pages scanned we did rotate back
1725 * onto the active list instead of evict.
1727 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1729 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1730 unsigned long *nr, int priority)
1732 unsigned long anon, file, free;
1733 unsigned long anon_prio, file_prio;
1734 unsigned long ap, fp;
1735 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1736 u64 fraction[2], denominator;
1737 enum lru_list l;
1738 int noswap = 0;
1740 /* If we have no swap space, do not bother scanning anon pages. */
1741 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1742 noswap = 1;
1743 fraction[0] = 0;
1744 fraction[1] = 1;
1745 denominator = 1;
1746 goto out;
1749 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1750 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1751 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1752 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1754 if (scanning_global_lru(sc)) {
1755 free = zone_page_state(zone, NR_FREE_PAGES);
1756 /* If we have very few page cache pages,
1757 force-scan anon pages. */
1758 if (unlikely(file + free <= high_wmark_pages(zone))) {
1759 fraction[0] = 1;
1760 fraction[1] = 0;
1761 denominator = 1;
1762 goto out;
1767 * With swappiness at 100, anonymous and file have the same priority.
1768 * This scanning priority is essentially the inverse of IO cost.
1770 anon_prio = sc->swappiness;
1771 file_prio = 200 - sc->swappiness;
1774 * OK, so we have swap space and a fair amount of page cache
1775 * pages. We use the recently rotated / recently scanned
1776 * ratios to determine how valuable each cache is.
1778 * Because workloads change over time (and to avoid overflow)
1779 * we keep these statistics as a floating average, which ends
1780 * up weighing recent references more than old ones.
1782 * anon in [0], file in [1]
1784 spin_lock_irq(&zone->lru_lock);
1785 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1786 reclaim_stat->recent_scanned[0] /= 2;
1787 reclaim_stat->recent_rotated[0] /= 2;
1790 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1791 reclaim_stat->recent_scanned[1] /= 2;
1792 reclaim_stat->recent_rotated[1] /= 2;
1796 * The amount of pressure on anon vs file pages is inversely
1797 * proportional to the fraction of recently scanned pages on
1798 * each list that were recently referenced and in active use.
1800 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1801 ap /= reclaim_stat->recent_rotated[0] + 1;
1803 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1804 fp /= reclaim_stat->recent_rotated[1] + 1;
1805 spin_unlock_irq(&zone->lru_lock);
1807 fraction[0] = ap;
1808 fraction[1] = fp;
1809 denominator = ap + fp + 1;
1810 out:
1811 for_each_evictable_lru(l) {
1812 int file = is_file_lru(l);
1813 unsigned long scan;
1815 scan = zone_nr_lru_pages(zone, sc, l);
1816 if (priority || noswap) {
1817 scan >>= priority;
1818 scan = div64_u64(scan * fraction[file], denominator);
1820 nr[l] = nr_scan_try_batch(scan,
1821 &reclaim_stat->nr_saved_scan[l]);
1826 * Reclaim/compaction depends on a number of pages being freed. To avoid
1827 * disruption to the system, a small number of order-0 pages continue to be
1828 * rotated and reclaimed in the normal fashion. However, by the time we get
1829 * back to the allocator and call try_to_compact_zone(), we ensure that
1830 * there are enough free pages for it to be likely successful
1832 static inline bool should_continue_reclaim(struct zone *zone,
1833 unsigned long nr_reclaimed,
1834 unsigned long nr_scanned,
1835 struct scan_control *sc)
1837 unsigned long pages_for_compaction;
1838 unsigned long inactive_lru_pages;
1840 /* If not in reclaim/compaction mode, stop */
1841 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1842 return false;
1845 * If we failed to reclaim and have scanned the full list, stop.
1846 * NOTE: Checking just nr_reclaimed would exit reclaim/compaction far
1847 * faster but obviously would be less likely to succeed
1848 * allocation. If this is desirable, use GFP_REPEAT to decide
1849 * if both reclaimed and scanned should be checked or just
1850 * reclaimed
1852 if (!nr_reclaimed && !nr_scanned)
1853 return false;
1856 * If we have not reclaimed enough pages for compaction and the
1857 * inactive lists are large enough, continue reclaiming
1859 pages_for_compaction = (2UL << sc->order);
1860 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1861 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1862 if (sc->nr_reclaimed < pages_for_compaction &&
1863 inactive_lru_pages > pages_for_compaction)
1864 return true;
1866 /* If compaction would go ahead or the allocation would succeed, stop */
1867 switch (compaction_suitable(zone, sc->order)) {
1868 case COMPACT_PARTIAL:
1869 case COMPACT_CONTINUE:
1870 return false;
1871 default:
1872 return true;
1877 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1879 static void shrink_zone(int priority, struct zone *zone,
1880 struct scan_control *sc)
1882 unsigned long nr[NR_LRU_LISTS];
1883 unsigned long nr_to_scan;
1884 enum lru_list l;
1885 unsigned long nr_reclaimed;
1886 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1887 unsigned long nr_scanned = sc->nr_scanned;
1889 restart:
1890 nr_reclaimed = 0;
1891 get_scan_count(zone, sc, nr, priority);
1893 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1894 nr[LRU_INACTIVE_FILE]) {
1895 for_each_evictable_lru(l) {
1896 if (nr[l]) {
1897 nr_to_scan = min_t(unsigned long,
1898 nr[l], SWAP_CLUSTER_MAX);
1899 nr[l] -= nr_to_scan;
1901 nr_reclaimed += shrink_list(l, nr_to_scan,
1902 zone, sc, priority);
1906 * On large memory systems, scan >> priority can become
1907 * really large. This is fine for the starting priority;
1908 * we want to put equal scanning pressure on each zone.
1909 * However, if the VM has a harder time of freeing pages,
1910 * with multiple processes reclaiming pages, the total
1911 * freeing target can get unreasonably large.
1913 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1914 break;
1916 sc->nr_reclaimed += nr_reclaimed;
1919 * Even if we did not try to evict anon pages at all, we want to
1920 * rebalance the anon lru active/inactive ratio.
1922 if (inactive_anon_is_low(zone, sc))
1923 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1925 /* reclaim/compaction might need reclaim to continue */
1926 if (should_continue_reclaim(zone, nr_reclaimed,
1927 sc->nr_scanned - nr_scanned, sc))
1928 goto restart;
1930 throttle_vm_writeout(sc->gfp_mask);
1934 * This is the direct reclaim path, for page-allocating processes. We only
1935 * try to reclaim pages from zones which will satisfy the caller's allocation
1936 * request.
1938 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1939 * Because:
1940 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1941 * allocation or
1942 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1943 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1944 * zone defense algorithm.
1946 * If a zone is deemed to be full of pinned pages then just give it a light
1947 * scan then give up on it.
1949 static void shrink_zones(int priority, struct zonelist *zonelist,
1950 struct scan_control *sc)
1952 struct zoneref *z;
1953 struct zone *zone;
1955 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1956 gfp_zone(sc->gfp_mask), sc->nodemask) {
1957 if (!populated_zone(zone))
1958 continue;
1960 * Take care memory controller reclaiming has small influence
1961 * to global LRU.
1963 if (scanning_global_lru(sc)) {
1964 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1965 continue;
1966 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1967 continue; /* Let kswapd poll it */
1970 shrink_zone(priority, zone, sc);
1974 static bool zone_reclaimable(struct zone *zone)
1976 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1980 * As hibernation is going on, kswapd is freezed so that it can't mark
1981 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1982 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1984 static bool all_unreclaimable(struct zonelist *zonelist,
1985 struct scan_control *sc)
1987 struct zoneref *z;
1988 struct zone *zone;
1989 bool all_unreclaimable = true;
1991 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1992 gfp_zone(sc->gfp_mask), sc->nodemask) {
1993 if (!populated_zone(zone))
1994 continue;
1995 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1996 continue;
1997 if (zone_reclaimable(zone)) {
1998 all_unreclaimable = false;
1999 break;
2003 return all_unreclaimable;
2007 * This is the main entry point to direct page reclaim.
2009 * If a full scan of the inactive list fails to free enough memory then we
2010 * are "out of memory" and something needs to be killed.
2012 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2013 * high - the zone may be full of dirty or under-writeback pages, which this
2014 * caller can't do much about. We kick the writeback threads and take explicit
2015 * naps in the hope that some of these pages can be written. But if the
2016 * allocating task holds filesystem locks which prevent writeout this might not
2017 * work, and the allocation attempt will fail.
2019 * returns: 0, if no pages reclaimed
2020 * else, the number of pages reclaimed
2022 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2023 struct scan_control *sc)
2025 int priority;
2026 unsigned long total_scanned = 0;
2027 struct reclaim_state *reclaim_state = current->reclaim_state;
2028 struct zoneref *z;
2029 struct zone *zone;
2030 unsigned long writeback_threshold;
2032 get_mems_allowed();
2033 delayacct_freepages_start();
2035 if (scanning_global_lru(sc))
2036 count_vm_event(ALLOCSTALL);
2038 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2039 sc->nr_scanned = 0;
2040 if (!priority)
2041 disable_swap_token();
2042 shrink_zones(priority, zonelist, sc);
2044 * Don't shrink slabs when reclaiming memory from
2045 * over limit cgroups
2047 if (scanning_global_lru(sc)) {
2048 unsigned long lru_pages = 0;
2049 for_each_zone_zonelist(zone, z, zonelist,
2050 gfp_zone(sc->gfp_mask)) {
2051 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2052 continue;
2054 lru_pages += zone_reclaimable_pages(zone);
2057 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2058 if (reclaim_state) {
2059 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2060 reclaim_state->reclaimed_slab = 0;
2063 total_scanned += sc->nr_scanned;
2064 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2065 goto out;
2068 * Try to write back as many pages as we just scanned. This
2069 * tends to cause slow streaming writers to write data to the
2070 * disk smoothly, at the dirtying rate, which is nice. But
2071 * that's undesirable in laptop mode, where we *want* lumpy
2072 * writeout. So in laptop mode, write out the whole world.
2074 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2075 if (total_scanned > writeback_threshold) {
2076 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2077 sc->may_writepage = 1;
2080 /* Take a nap, wait for some writeback to complete */
2081 if (!sc->hibernation_mode && sc->nr_scanned &&
2082 priority < DEF_PRIORITY - 2) {
2083 struct zone *preferred_zone;
2085 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2086 NULL, &preferred_zone);
2087 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2091 out:
2092 delayacct_freepages_end();
2093 put_mems_allowed();
2095 if (sc->nr_reclaimed)
2096 return sc->nr_reclaimed;
2098 /* top priority shrink_zones still had more to do? don't OOM, then */
2099 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2100 return 1;
2102 return 0;
2105 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2106 gfp_t gfp_mask, nodemask_t *nodemask)
2108 unsigned long nr_reclaimed;
2109 struct scan_control sc = {
2110 .gfp_mask = gfp_mask,
2111 .may_writepage = !laptop_mode,
2112 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2113 .may_unmap = 1,
2114 .may_swap = 1,
2115 .swappiness = vm_swappiness,
2116 .order = order,
2117 .mem_cgroup = NULL,
2118 .nodemask = nodemask,
2121 trace_mm_vmscan_direct_reclaim_begin(order,
2122 sc.may_writepage,
2123 gfp_mask);
2125 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2127 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2129 return nr_reclaimed;
2132 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2134 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2135 gfp_t gfp_mask, bool noswap,
2136 unsigned int swappiness,
2137 struct zone *zone)
2139 struct scan_control sc = {
2140 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2141 .may_writepage = !laptop_mode,
2142 .may_unmap = 1,
2143 .may_swap = !noswap,
2144 .swappiness = swappiness,
2145 .order = 0,
2146 .mem_cgroup = mem,
2148 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2149 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2151 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2152 sc.may_writepage,
2153 sc.gfp_mask);
2156 * NOTE: Although we can get the priority field, using it
2157 * here is not a good idea, since it limits the pages we can scan.
2158 * if we don't reclaim here, the shrink_zone from balance_pgdat
2159 * will pick up pages from other mem cgroup's as well. We hack
2160 * the priority and make it zero.
2162 shrink_zone(0, zone, &sc);
2164 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2166 return sc.nr_reclaimed;
2169 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2170 gfp_t gfp_mask,
2171 bool noswap,
2172 unsigned int swappiness)
2174 struct zonelist *zonelist;
2175 unsigned long nr_reclaimed;
2176 struct scan_control sc = {
2177 .may_writepage = !laptop_mode,
2178 .may_unmap = 1,
2179 .may_swap = !noswap,
2180 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2181 .swappiness = swappiness,
2182 .order = 0,
2183 .mem_cgroup = mem_cont,
2184 .nodemask = NULL, /* we don't care the placement */
2187 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2188 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2189 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2191 trace_mm_vmscan_memcg_reclaim_begin(0,
2192 sc.may_writepage,
2193 sc.gfp_mask);
2195 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2197 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2199 return nr_reclaimed;
2201 #endif
2204 * pgdat_balanced is used when checking if a node is balanced for high-order
2205 * allocations. Only zones that meet watermarks and are in a zone allowed
2206 * by the callers classzone_idx are added to balanced_pages. The total of
2207 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2208 * for the node to be considered balanced. Forcing all zones to be balanced
2209 * for high orders can cause excessive reclaim when there are imbalanced zones.
2210 * The choice of 25% is due to
2211 * o a 16M DMA zone that is balanced will not balance a zone on any
2212 * reasonable sized machine
2213 * o On all other machines, the top zone must be at least a reasonable
2214 * precentage of the middle zones. For example, on 32-bit x86, highmem
2215 * would need to be at least 256M for it to be balance a whole node.
2216 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2217 * to balance a node on its own. These seemed like reasonable ratios.
2219 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2220 int classzone_idx)
2222 unsigned long present_pages = 0;
2223 int i;
2225 for (i = 0; i <= classzone_idx; i++)
2226 present_pages += pgdat->node_zones[i].present_pages;
2228 return balanced_pages > (present_pages >> 2);
2231 /* is kswapd sleeping prematurely? */
2232 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2233 int classzone_idx)
2235 int i;
2236 unsigned long balanced = 0;
2237 bool all_zones_ok = true;
2239 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2240 if (remaining)
2241 return true;
2243 /* Check the watermark levels */
2244 for (i = 0; i < pgdat->nr_zones; i++) {
2245 struct zone *zone = pgdat->node_zones + i;
2247 if (!populated_zone(zone))
2248 continue;
2251 * balance_pgdat() skips over all_unreclaimable after
2252 * DEF_PRIORITY. Effectively, it considers them balanced so
2253 * they must be considered balanced here as well if kswapd
2254 * is to sleep
2256 if (zone->all_unreclaimable) {
2257 balanced += zone->present_pages;
2258 continue;
2261 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2262 classzone_idx, 0))
2263 all_zones_ok = false;
2264 else
2265 balanced += zone->present_pages;
2269 * For high-order requests, the balanced zones must contain at least
2270 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2271 * must be balanced
2273 if (order)
2274 return pgdat_balanced(pgdat, balanced, classzone_idx);
2275 else
2276 return !all_zones_ok;
2280 * For kswapd, balance_pgdat() will work across all this node's zones until
2281 * they are all at high_wmark_pages(zone).
2283 * Returns the final order kswapd was reclaiming at
2285 * There is special handling here for zones which are full of pinned pages.
2286 * This can happen if the pages are all mlocked, or if they are all used by
2287 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2288 * What we do is to detect the case where all pages in the zone have been
2289 * scanned twice and there has been zero successful reclaim. Mark the zone as
2290 * dead and from now on, only perform a short scan. Basically we're polling
2291 * the zone for when the problem goes away.
2293 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2294 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2295 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2296 * lower zones regardless of the number of free pages in the lower zones. This
2297 * interoperates with the page allocator fallback scheme to ensure that aging
2298 * of pages is balanced across the zones.
2300 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2301 int *classzone_idx)
2303 int all_zones_ok;
2304 unsigned long balanced;
2305 int priority;
2306 int i;
2307 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2308 unsigned long total_scanned;
2309 struct reclaim_state *reclaim_state = current->reclaim_state;
2310 struct scan_control sc = {
2311 .gfp_mask = GFP_KERNEL,
2312 .may_unmap = 1,
2313 .may_swap = 1,
2315 * kswapd doesn't want to be bailed out while reclaim. because
2316 * we want to put equal scanning pressure on each zone.
2318 .nr_to_reclaim = ULONG_MAX,
2319 .swappiness = vm_swappiness,
2320 .order = order,
2321 .mem_cgroup = NULL,
2323 loop_again:
2324 total_scanned = 0;
2325 sc.nr_reclaimed = 0;
2326 sc.may_writepage = !laptop_mode;
2327 count_vm_event(PAGEOUTRUN);
2329 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2330 unsigned long lru_pages = 0;
2331 int has_under_min_watermark_zone = 0;
2333 /* The swap token gets in the way of swapout... */
2334 if (!priority)
2335 disable_swap_token();
2337 all_zones_ok = 1;
2338 balanced = 0;
2341 * Scan in the highmem->dma direction for the highest
2342 * zone which needs scanning
2344 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2345 struct zone *zone = pgdat->node_zones + i;
2347 if (!populated_zone(zone))
2348 continue;
2350 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2351 continue;
2354 * Do some background aging of the anon list, to give
2355 * pages a chance to be referenced before reclaiming.
2357 if (inactive_anon_is_low(zone, &sc))
2358 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2359 &sc, priority, 0);
2361 if (!zone_watermark_ok_safe(zone, order,
2362 high_wmark_pages(zone), 0, 0)) {
2363 end_zone = i;
2364 *classzone_idx = i;
2365 break;
2368 if (i < 0)
2369 goto out;
2371 for (i = 0; i <= end_zone; i++) {
2372 struct zone *zone = pgdat->node_zones + i;
2374 lru_pages += zone_reclaimable_pages(zone);
2378 * Now scan the zone in the dma->highmem direction, stopping
2379 * at the last zone which needs scanning.
2381 * We do this because the page allocator works in the opposite
2382 * direction. This prevents the page allocator from allocating
2383 * pages behind kswapd's direction of progress, which would
2384 * cause too much scanning of the lower zones.
2386 for (i = 0; i <= end_zone; i++) {
2387 int compaction;
2388 struct zone *zone = pgdat->node_zones + i;
2389 int nr_slab;
2391 if (!populated_zone(zone))
2392 continue;
2394 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2395 continue;
2397 sc.nr_scanned = 0;
2400 * Call soft limit reclaim before calling shrink_zone.
2401 * For now we ignore the return value
2403 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2406 * We put equal pressure on every zone, unless one
2407 * zone has way too many pages free already.
2409 if (!zone_watermark_ok_safe(zone, order,
2410 8*high_wmark_pages(zone), end_zone, 0))
2411 shrink_zone(priority, zone, &sc);
2412 reclaim_state->reclaimed_slab = 0;
2413 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2414 lru_pages);
2415 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2416 total_scanned += sc.nr_scanned;
2418 compaction = 0;
2419 if (order &&
2420 zone_watermark_ok(zone, 0,
2421 high_wmark_pages(zone),
2422 end_zone, 0) &&
2423 !zone_watermark_ok(zone, order,
2424 high_wmark_pages(zone),
2425 end_zone, 0)) {
2426 compact_zone_order(zone,
2427 order,
2428 sc.gfp_mask, false,
2429 COMPACT_MODE_KSWAPD);
2430 compaction = 1;
2433 if (zone->all_unreclaimable)
2434 continue;
2435 if (!compaction && nr_slab == 0 &&
2436 !zone_reclaimable(zone))
2437 zone->all_unreclaimable = 1;
2439 * If we've done a decent amount of scanning and
2440 * the reclaim ratio is low, start doing writepage
2441 * even in laptop mode
2443 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2444 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2445 sc.may_writepage = 1;
2447 if (!zone_watermark_ok_safe(zone, order,
2448 high_wmark_pages(zone), end_zone, 0)) {
2449 all_zones_ok = 0;
2451 * We are still under min water mark. This
2452 * means that we have a GFP_ATOMIC allocation
2453 * failure risk. Hurry up!
2455 if (!zone_watermark_ok_safe(zone, order,
2456 min_wmark_pages(zone), end_zone, 0))
2457 has_under_min_watermark_zone = 1;
2458 } else {
2460 * If a zone reaches its high watermark,
2461 * consider it to be no longer congested. It's
2462 * possible there are dirty pages backed by
2463 * congested BDIs but as pressure is relieved,
2464 * spectulatively avoid congestion waits
2466 zone_clear_flag(zone, ZONE_CONGESTED);
2467 if (i <= *classzone_idx)
2468 balanced += zone->present_pages;
2472 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2473 break; /* kswapd: all done */
2475 * OK, kswapd is getting into trouble. Take a nap, then take
2476 * another pass across the zones.
2478 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2479 if (has_under_min_watermark_zone)
2480 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2481 else
2482 congestion_wait(BLK_RW_ASYNC, HZ/10);
2486 * We do this so kswapd doesn't build up large priorities for
2487 * example when it is freeing in parallel with allocators. It
2488 * matches the direct reclaim path behaviour in terms of impact
2489 * on zone->*_priority.
2491 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2492 break;
2494 out:
2497 * order-0: All zones must meet high watermark for a balanced node
2498 * high-order: Balanced zones must make up at least 25% of the node
2499 * for the node to be balanced
2501 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2502 cond_resched();
2504 try_to_freeze();
2507 * Fragmentation may mean that the system cannot be
2508 * rebalanced for high-order allocations in all zones.
2509 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2510 * it means the zones have been fully scanned and are still
2511 * not balanced. For high-order allocations, there is
2512 * little point trying all over again as kswapd may
2513 * infinite loop.
2515 * Instead, recheck all watermarks at order-0 as they
2516 * are the most important. If watermarks are ok, kswapd will go
2517 * back to sleep. High-order users can still perform direct
2518 * reclaim if they wish.
2520 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2521 order = sc.order = 0;
2523 goto loop_again;
2527 * If kswapd was reclaiming at a higher order, it has the option of
2528 * sleeping without all zones being balanced. Before it does, it must
2529 * ensure that the watermarks for order-0 on *all* zones are met and
2530 * that the congestion flags are cleared. The congestion flag must
2531 * be cleared as kswapd is the only mechanism that clears the flag
2532 * and it is potentially going to sleep here.
2534 if (order) {
2535 for (i = 0; i <= end_zone; i++) {
2536 struct zone *zone = pgdat->node_zones + i;
2538 if (!populated_zone(zone))
2539 continue;
2541 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2542 continue;
2544 /* Confirm the zone is balanced for order-0 */
2545 if (!zone_watermark_ok(zone, 0,
2546 high_wmark_pages(zone), 0, 0)) {
2547 order = sc.order = 0;
2548 goto loop_again;
2551 /* If balanced, clear the congested flag */
2552 zone_clear_flag(zone, ZONE_CONGESTED);
2557 * Return the order we were reclaiming at so sleeping_prematurely()
2558 * makes a decision on the order we were last reclaiming at. However,
2559 * if another caller entered the allocator slow path while kswapd
2560 * was awake, order will remain at the higher level
2562 *classzone_idx = end_zone;
2563 return order;
2566 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2568 long remaining = 0;
2569 DEFINE_WAIT(wait);
2571 if (freezing(current) || kthread_should_stop())
2572 return;
2574 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2576 /* Try to sleep for a short interval */
2577 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2578 remaining = schedule_timeout(HZ/10);
2579 finish_wait(&pgdat->kswapd_wait, &wait);
2580 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2584 * After a short sleep, check if it was a premature sleep. If not, then
2585 * go fully to sleep until explicitly woken up.
2587 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2588 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2591 * vmstat counters are not perfectly accurate and the estimated
2592 * value for counters such as NR_FREE_PAGES can deviate from the
2593 * true value by nr_online_cpus * threshold. To avoid the zone
2594 * watermarks being breached while under pressure, we reduce the
2595 * per-cpu vmstat threshold while kswapd is awake and restore
2596 * them before going back to sleep.
2598 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2599 schedule();
2600 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2601 } else {
2602 if (remaining)
2603 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2604 else
2605 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2607 finish_wait(&pgdat->kswapd_wait, &wait);
2611 * The background pageout daemon, started as a kernel thread
2612 * from the init process.
2614 * This basically trickles out pages so that we have _some_
2615 * free memory available even if there is no other activity
2616 * that frees anything up. This is needed for things like routing
2617 * etc, where we otherwise might have all activity going on in
2618 * asynchronous contexts that cannot page things out.
2620 * If there are applications that are active memory-allocators
2621 * (most normal use), this basically shouldn't matter.
2623 static int kswapd(void *p)
2625 unsigned long order;
2626 int classzone_idx;
2627 pg_data_t *pgdat = (pg_data_t*)p;
2628 struct task_struct *tsk = current;
2630 struct reclaim_state reclaim_state = {
2631 .reclaimed_slab = 0,
2633 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2635 lockdep_set_current_reclaim_state(GFP_KERNEL);
2637 if (!cpumask_empty(cpumask))
2638 set_cpus_allowed_ptr(tsk, cpumask);
2639 current->reclaim_state = &reclaim_state;
2642 * Tell the memory management that we're a "memory allocator",
2643 * and that if we need more memory we should get access to it
2644 * regardless (see "__alloc_pages()"). "kswapd" should
2645 * never get caught in the normal page freeing logic.
2647 * (Kswapd normally doesn't need memory anyway, but sometimes
2648 * you need a small amount of memory in order to be able to
2649 * page out something else, and this flag essentially protects
2650 * us from recursively trying to free more memory as we're
2651 * trying to free the first piece of memory in the first place).
2653 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2654 set_freezable();
2656 order = 0;
2657 classzone_idx = MAX_NR_ZONES - 1;
2658 for ( ; ; ) {
2659 unsigned long new_order;
2660 int new_classzone_idx;
2661 int ret;
2663 new_order = pgdat->kswapd_max_order;
2664 new_classzone_idx = pgdat->classzone_idx;
2665 pgdat->kswapd_max_order = 0;
2666 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2667 if (order < new_order || classzone_idx > new_classzone_idx) {
2669 * Don't sleep if someone wants a larger 'order'
2670 * allocation or has tigher zone constraints
2672 order = new_order;
2673 classzone_idx = new_classzone_idx;
2674 } else {
2675 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2676 order = pgdat->kswapd_max_order;
2677 classzone_idx = pgdat->classzone_idx;
2678 pgdat->kswapd_max_order = 0;
2679 pgdat->classzone_idx = MAX_NR_ZONES - 1;
2682 ret = try_to_freeze();
2683 if (kthread_should_stop())
2684 break;
2687 * We can speed up thawing tasks if we don't call balance_pgdat
2688 * after returning from the refrigerator
2690 if (!ret) {
2691 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2692 order = balance_pgdat(pgdat, order, &classzone_idx);
2695 return 0;
2699 * A zone is low on free memory, so wake its kswapd task to service it.
2701 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2703 pg_data_t *pgdat;
2705 if (!populated_zone(zone))
2706 return;
2708 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2709 return;
2710 pgdat = zone->zone_pgdat;
2711 if (pgdat->kswapd_max_order < order) {
2712 pgdat->kswapd_max_order = order;
2713 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2715 if (!waitqueue_active(&pgdat->kswapd_wait))
2716 return;
2717 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2718 return;
2720 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2721 wake_up_interruptible(&pgdat->kswapd_wait);
2725 * The reclaimable count would be mostly accurate.
2726 * The less reclaimable pages may be
2727 * - mlocked pages, which will be moved to unevictable list when encountered
2728 * - mapped pages, which may require several travels to be reclaimed
2729 * - dirty pages, which is not "instantly" reclaimable
2731 unsigned long global_reclaimable_pages(void)
2733 int nr;
2735 nr = global_page_state(NR_ACTIVE_FILE) +
2736 global_page_state(NR_INACTIVE_FILE);
2738 if (nr_swap_pages > 0)
2739 nr += global_page_state(NR_ACTIVE_ANON) +
2740 global_page_state(NR_INACTIVE_ANON);
2742 return nr;
2745 unsigned long zone_reclaimable_pages(struct zone *zone)
2747 int nr;
2749 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2750 zone_page_state(zone, NR_INACTIVE_FILE);
2752 if (nr_swap_pages > 0)
2753 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2754 zone_page_state(zone, NR_INACTIVE_ANON);
2756 return nr;
2759 #ifdef CONFIG_HIBERNATION
2761 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2762 * freed pages.
2764 * Rather than trying to age LRUs the aim is to preserve the overall
2765 * LRU order by reclaiming preferentially
2766 * inactive > active > active referenced > active mapped
2768 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2770 struct reclaim_state reclaim_state;
2771 struct scan_control sc = {
2772 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2773 .may_swap = 1,
2774 .may_unmap = 1,
2775 .may_writepage = 1,
2776 .nr_to_reclaim = nr_to_reclaim,
2777 .hibernation_mode = 1,
2778 .swappiness = vm_swappiness,
2779 .order = 0,
2781 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2782 struct task_struct *p = current;
2783 unsigned long nr_reclaimed;
2785 p->flags |= PF_MEMALLOC;
2786 lockdep_set_current_reclaim_state(sc.gfp_mask);
2787 reclaim_state.reclaimed_slab = 0;
2788 p->reclaim_state = &reclaim_state;
2790 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2792 p->reclaim_state = NULL;
2793 lockdep_clear_current_reclaim_state();
2794 p->flags &= ~PF_MEMALLOC;
2796 return nr_reclaimed;
2798 #endif /* CONFIG_HIBERNATION */
2800 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2801 not required for correctness. So if the last cpu in a node goes
2802 away, we get changed to run anywhere: as the first one comes back,
2803 restore their cpu bindings. */
2804 static int __devinit cpu_callback(struct notifier_block *nfb,
2805 unsigned long action, void *hcpu)
2807 int nid;
2809 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2810 for_each_node_state(nid, N_HIGH_MEMORY) {
2811 pg_data_t *pgdat = NODE_DATA(nid);
2812 const struct cpumask *mask;
2814 mask = cpumask_of_node(pgdat->node_id);
2816 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2817 /* One of our CPUs online: restore mask */
2818 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2821 return NOTIFY_OK;
2825 * This kswapd start function will be called by init and node-hot-add.
2826 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2828 int kswapd_run(int nid)
2830 pg_data_t *pgdat = NODE_DATA(nid);
2831 int ret = 0;
2833 if (pgdat->kswapd)
2834 return 0;
2836 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2837 if (IS_ERR(pgdat->kswapd)) {
2838 /* failure at boot is fatal */
2839 BUG_ON(system_state == SYSTEM_BOOTING);
2840 printk("Failed to start kswapd on node %d\n",nid);
2841 ret = -1;
2843 return ret;
2847 * Called by memory hotplug when all memory in a node is offlined.
2849 void kswapd_stop(int nid)
2851 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2853 if (kswapd)
2854 kthread_stop(kswapd);
2857 static int __init kswapd_init(void)
2859 int nid;
2861 swap_setup();
2862 for_each_node_state(nid, N_HIGH_MEMORY)
2863 kswapd_run(nid);
2864 hotcpu_notifier(cpu_callback, 0);
2865 return 0;
2868 module_init(kswapd_init)
2870 #ifdef CONFIG_NUMA
2872 * Zone reclaim mode
2874 * If non-zero call zone_reclaim when the number of free pages falls below
2875 * the watermarks.
2877 int zone_reclaim_mode __read_mostly;
2879 #define RECLAIM_OFF 0
2880 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2881 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2882 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2885 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2886 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2887 * a zone.
2889 #define ZONE_RECLAIM_PRIORITY 4
2892 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2893 * occur.
2895 int sysctl_min_unmapped_ratio = 1;
2898 * If the number of slab pages in a zone grows beyond this percentage then
2899 * slab reclaim needs to occur.
2901 int sysctl_min_slab_ratio = 5;
2903 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2905 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2906 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2907 zone_page_state(zone, NR_ACTIVE_FILE);
2910 * It's possible for there to be more file mapped pages than
2911 * accounted for by the pages on the file LRU lists because
2912 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2914 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2917 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2918 static long zone_pagecache_reclaimable(struct zone *zone)
2920 long nr_pagecache_reclaimable;
2921 long delta = 0;
2924 * If RECLAIM_SWAP is set, then all file pages are considered
2925 * potentially reclaimable. Otherwise, we have to worry about
2926 * pages like swapcache and zone_unmapped_file_pages() provides
2927 * a better estimate
2929 if (zone_reclaim_mode & RECLAIM_SWAP)
2930 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2931 else
2932 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2934 /* If we can't clean pages, remove dirty pages from consideration */
2935 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2936 delta += zone_page_state(zone, NR_FILE_DIRTY);
2938 /* Watch for any possible underflows due to delta */
2939 if (unlikely(delta > nr_pagecache_reclaimable))
2940 delta = nr_pagecache_reclaimable;
2942 return nr_pagecache_reclaimable - delta;
2946 * Try to free up some pages from this zone through reclaim.
2948 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2950 /* Minimum pages needed in order to stay on node */
2951 const unsigned long nr_pages = 1 << order;
2952 struct task_struct *p = current;
2953 struct reclaim_state reclaim_state;
2954 int priority;
2955 struct scan_control sc = {
2956 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2957 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2958 .may_swap = 1,
2959 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2960 SWAP_CLUSTER_MAX),
2961 .gfp_mask = gfp_mask,
2962 .swappiness = vm_swappiness,
2963 .order = order,
2965 unsigned long nr_slab_pages0, nr_slab_pages1;
2967 cond_resched();
2969 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2970 * and we also need to be able to write out pages for RECLAIM_WRITE
2971 * and RECLAIM_SWAP.
2973 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2974 lockdep_set_current_reclaim_state(gfp_mask);
2975 reclaim_state.reclaimed_slab = 0;
2976 p->reclaim_state = &reclaim_state;
2978 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2980 * Free memory by calling shrink zone with increasing
2981 * priorities until we have enough memory freed.
2983 priority = ZONE_RECLAIM_PRIORITY;
2984 do {
2985 shrink_zone(priority, zone, &sc);
2986 priority--;
2987 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2990 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2991 if (nr_slab_pages0 > zone->min_slab_pages) {
2993 * shrink_slab() does not currently allow us to determine how
2994 * many pages were freed in this zone. So we take the current
2995 * number of slab pages and shake the slab until it is reduced
2996 * by the same nr_pages that we used for reclaiming unmapped
2997 * pages.
2999 * Note that shrink_slab will free memory on all zones and may
3000 * take a long time.
3002 for (;;) {
3003 unsigned long lru_pages = zone_reclaimable_pages(zone);
3005 /* No reclaimable slab or very low memory pressure */
3006 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
3007 break;
3009 /* Freed enough memory */
3010 nr_slab_pages1 = zone_page_state(zone,
3011 NR_SLAB_RECLAIMABLE);
3012 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3013 break;
3017 * Update nr_reclaimed by the number of slab pages we
3018 * reclaimed from this zone.
3020 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3021 if (nr_slab_pages1 < nr_slab_pages0)
3022 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3025 p->reclaim_state = NULL;
3026 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3027 lockdep_clear_current_reclaim_state();
3028 return sc.nr_reclaimed >= nr_pages;
3031 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3033 int node_id;
3034 int ret;
3037 * Zone reclaim reclaims unmapped file backed pages and
3038 * slab pages if we are over the defined limits.
3040 * A small portion of unmapped file backed pages is needed for
3041 * file I/O otherwise pages read by file I/O will be immediately
3042 * thrown out if the zone is overallocated. So we do not reclaim
3043 * if less than a specified percentage of the zone is used by
3044 * unmapped file backed pages.
3046 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3047 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3048 return ZONE_RECLAIM_FULL;
3050 if (zone->all_unreclaimable)
3051 return ZONE_RECLAIM_FULL;
3054 * Do not scan if the allocation should not be delayed.
3056 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3057 return ZONE_RECLAIM_NOSCAN;
3060 * Only run zone reclaim on the local zone or on zones that do not
3061 * have associated processors. This will favor the local processor
3062 * over remote processors and spread off node memory allocations
3063 * as wide as possible.
3065 node_id = zone_to_nid(zone);
3066 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3067 return ZONE_RECLAIM_NOSCAN;
3069 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3070 return ZONE_RECLAIM_NOSCAN;
3072 ret = __zone_reclaim(zone, gfp_mask, order);
3073 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3075 if (!ret)
3076 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3078 return ret;
3080 #endif
3083 * page_evictable - test whether a page is evictable
3084 * @page: the page to test
3085 * @vma: the VMA in which the page is or will be mapped, may be NULL
3087 * Test whether page is evictable--i.e., should be placed on active/inactive
3088 * lists vs unevictable list. The vma argument is !NULL when called from the
3089 * fault path to determine how to instantate a new page.
3091 * Reasons page might not be evictable:
3092 * (1) page's mapping marked unevictable
3093 * (2) page is part of an mlocked VMA
3096 int page_evictable(struct page *page, struct vm_area_struct *vma)
3099 if (mapping_unevictable(page_mapping(page)))
3100 return 0;
3102 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3103 return 0;
3105 return 1;
3109 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3110 * @page: page to check evictability and move to appropriate lru list
3111 * @zone: zone page is in
3113 * Checks a page for evictability and moves the page to the appropriate
3114 * zone lru list.
3116 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3117 * have PageUnevictable set.
3119 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3121 VM_BUG_ON(PageActive(page));
3123 retry:
3124 ClearPageUnevictable(page);
3125 if (page_evictable(page, NULL)) {
3126 enum lru_list l = page_lru_base_type(page);
3128 __dec_zone_state(zone, NR_UNEVICTABLE);
3129 list_move(&page->lru, &zone->lru[l].list);
3130 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3131 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3132 __count_vm_event(UNEVICTABLE_PGRESCUED);
3133 } else {
3135 * rotate unevictable list
3137 SetPageUnevictable(page);
3138 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3139 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3140 if (page_evictable(page, NULL))
3141 goto retry;
3146 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3147 * @mapping: struct address_space to scan for evictable pages
3149 * Scan all pages in mapping. Check unevictable pages for
3150 * evictability and move them to the appropriate zone lru list.
3152 void scan_mapping_unevictable_pages(struct address_space *mapping)
3154 pgoff_t next = 0;
3155 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3156 PAGE_CACHE_SHIFT;
3157 struct zone *zone;
3158 struct pagevec pvec;
3160 if (mapping->nrpages == 0)
3161 return;
3163 pagevec_init(&pvec, 0);
3164 while (next < end &&
3165 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3166 int i;
3167 int pg_scanned = 0;
3169 zone = NULL;
3171 for (i = 0; i < pagevec_count(&pvec); i++) {
3172 struct page *page = pvec.pages[i];
3173 pgoff_t page_index = page->index;
3174 struct zone *pagezone = page_zone(page);
3176 pg_scanned++;
3177 if (page_index > next)
3178 next = page_index;
3179 next++;
3181 if (pagezone != zone) {
3182 if (zone)
3183 spin_unlock_irq(&zone->lru_lock);
3184 zone = pagezone;
3185 spin_lock_irq(&zone->lru_lock);
3188 if (PageLRU(page) && PageUnevictable(page))
3189 check_move_unevictable_page(page, zone);
3191 if (zone)
3192 spin_unlock_irq(&zone->lru_lock);
3193 pagevec_release(&pvec);
3195 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3201 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3202 * @zone - zone of which to scan the unevictable list
3204 * Scan @zone's unevictable LRU lists to check for pages that have become
3205 * evictable. Move those that have to @zone's inactive list where they
3206 * become candidates for reclaim, unless shrink_inactive_zone() decides
3207 * to reactivate them. Pages that are still unevictable are rotated
3208 * back onto @zone's unevictable list.
3210 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3211 static void scan_zone_unevictable_pages(struct zone *zone)
3213 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3214 unsigned long scan;
3215 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3217 while (nr_to_scan > 0) {
3218 unsigned long batch_size = min(nr_to_scan,
3219 SCAN_UNEVICTABLE_BATCH_SIZE);
3221 spin_lock_irq(&zone->lru_lock);
3222 for (scan = 0; scan < batch_size; scan++) {
3223 struct page *page = lru_to_page(l_unevictable);
3225 if (!trylock_page(page))
3226 continue;
3228 prefetchw_prev_lru_page(page, l_unevictable, flags);
3230 if (likely(PageLRU(page) && PageUnevictable(page)))
3231 check_move_unevictable_page(page, zone);
3233 unlock_page(page);
3235 spin_unlock_irq(&zone->lru_lock);
3237 nr_to_scan -= batch_size;
3243 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3245 * A really big hammer: scan all zones' unevictable LRU lists to check for
3246 * pages that have become evictable. Move those back to the zones'
3247 * inactive list where they become candidates for reclaim.
3248 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3249 * and we add swap to the system. As such, it runs in the context of a task
3250 * that has possibly/probably made some previously unevictable pages
3251 * evictable.
3253 static void scan_all_zones_unevictable_pages(void)
3255 struct zone *zone;
3257 for_each_zone(zone) {
3258 scan_zone_unevictable_pages(zone);
3263 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3264 * all nodes' unevictable lists for evictable pages
3266 unsigned long scan_unevictable_pages;
3268 int scan_unevictable_handler(struct ctl_table *table, int write,
3269 void __user *buffer,
3270 size_t *length, loff_t *ppos)
3272 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3274 if (write && *(unsigned long *)table->data)
3275 scan_all_zones_unevictable_pages();
3277 scan_unevictable_pages = 0;
3278 return 0;
3281 #ifdef CONFIG_NUMA
3283 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3284 * a specified node's per zone unevictable lists for evictable pages.
3287 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3288 struct sysdev_attribute *attr,
3289 char *buf)
3291 return sprintf(buf, "0\n"); /* always zero; should fit... */
3294 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3295 struct sysdev_attribute *attr,
3296 const char *buf, size_t count)
3298 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3299 struct zone *zone;
3300 unsigned long res;
3301 unsigned long req = strict_strtoul(buf, 10, &res);
3303 if (!req)
3304 return 1; /* zero is no-op */
3306 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3307 if (!populated_zone(zone))
3308 continue;
3309 scan_zone_unevictable_pages(zone);
3311 return 1;
3315 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3316 read_scan_unevictable_node,
3317 write_scan_unevictable_node);
3319 int scan_unevictable_register_node(struct node *node)
3321 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3324 void scan_unevictable_unregister_node(struct node *node)
3326 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3328 #endif