4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/slab.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/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
43 #include <asm/tlbflush.h>
44 #include <asm/div64.h>
46 #include <linux/swapops.h>
51 /* Incremented by the number of inactive pages that were scanned */
52 unsigned long nr_scanned;
54 /* This context's GFP mask */
59 /* Can pages be swapped as part of reclaim? */
62 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
63 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
64 * In this context, it doesn't matter that we scan the
65 * whole list at once. */
70 int all_unreclaimable;
74 /* Which cgroup do we reclaim from */
75 struct mem_cgroup *mem_cgroup;
77 /* Pluggable isolate pages callback */
78 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
79 unsigned long *scanned, int order, int mode,
80 struct zone *z, struct mem_cgroup *mem_cont,
84 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
86 #ifdef ARCH_HAS_PREFETCH
87 #define prefetch_prev_lru_page(_page, _base, _field) \
89 if ((_page)->lru.prev != _base) { \
92 prev = lru_to_page(&(_page->lru)); \
93 prefetch(&prev->_field); \
97 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
100 #ifdef ARCH_HAS_PREFETCHW
101 #define prefetchw_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetchw(&prev->_field); \
111 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
115 * From 0 .. 100. Higher means more swappy.
117 int vm_swappiness = 60;
118 long vm_total_pages; /* The total number of pages which the VM controls */
120 static LIST_HEAD(shrinker_list);
121 static DECLARE_RWSEM(shrinker_rwsem);
123 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
124 #define scan_global_lru(sc) (!(sc)->mem_cgroup)
126 #define scan_global_lru(sc) (1)
130 * Add a shrinker callback to be called from the vm
132 void register_shrinker(struct shrinker *shrinker)
135 down_write(&shrinker_rwsem);
136 list_add_tail(&shrinker->list, &shrinker_list);
137 up_write(&shrinker_rwsem);
139 EXPORT_SYMBOL(register_shrinker);
144 void unregister_shrinker(struct shrinker *shrinker)
146 down_write(&shrinker_rwsem);
147 list_del(&shrinker->list);
148 up_write(&shrinker_rwsem);
150 EXPORT_SYMBOL(unregister_shrinker);
152 #define SHRINK_BATCH 128
154 * Call the shrink functions to age shrinkable caches
156 * Here we assume it costs one seek to replace a lru page and that it also
157 * takes a seek to recreate a cache object. With this in mind we age equal
158 * percentages of the lru and ageable caches. This should balance the seeks
159 * generated by these structures.
161 * If the vm encountered mapped pages on the LRU it increase the pressure on
162 * slab to avoid swapping.
164 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
166 * `lru_pages' represents the number of on-LRU pages in all the zones which
167 * are eligible for the caller's allocation attempt. It is used for balancing
168 * slab reclaim versus page reclaim.
170 * Returns the number of slab objects which we shrunk.
172 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
173 unsigned long lru_pages)
175 struct shrinker *shrinker;
176 unsigned long ret = 0;
179 scanned = SWAP_CLUSTER_MAX;
181 if (!down_read_trylock(&shrinker_rwsem))
182 return 1; /* Assume we'll be able to shrink next time */
184 list_for_each_entry(shrinker, &shrinker_list, list) {
185 unsigned long long delta;
186 unsigned long total_scan;
187 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
189 delta = (4 * scanned) / shrinker->seeks;
191 do_div(delta, lru_pages + 1);
192 shrinker->nr += delta;
193 if (shrinker->nr < 0) {
194 printk(KERN_ERR "%s: nr=%ld\n",
195 __func__, shrinker->nr);
196 shrinker->nr = max_pass;
200 * Avoid risking looping forever due to too large nr value:
201 * never try to free more than twice the estimate number of
204 if (shrinker->nr > max_pass * 2)
205 shrinker->nr = max_pass * 2;
207 total_scan = shrinker->nr;
210 while (total_scan >= SHRINK_BATCH) {
211 long this_scan = SHRINK_BATCH;
215 nr_before = (*shrinker->shrink)(0, gfp_mask);
216 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
217 if (shrink_ret == -1)
219 if (shrink_ret < nr_before)
220 ret += nr_before - shrink_ret;
221 count_vm_events(SLABS_SCANNED, this_scan);
222 total_scan -= this_scan;
227 shrinker->nr += total_scan;
229 up_read(&shrinker_rwsem);
233 /* Called without lock on whether page is mapped, so answer is unstable */
234 static inline int page_mapping_inuse(struct page *page)
236 struct address_space *mapping;
238 /* Page is in somebody's page tables. */
239 if (page_mapped(page))
242 /* Be more reluctant to reclaim swapcache than pagecache */
243 if (PageSwapCache(page))
246 mapping = page_mapping(page);
250 /* File is mmap'd by somebody? */
251 return mapping_mapped(mapping);
254 static inline int is_page_cache_freeable(struct page *page)
256 return page_count(page) - !!PagePrivate(page) == 2;
259 static int may_write_to_queue(struct backing_dev_info *bdi)
261 if (current->flags & PF_SWAPWRITE)
263 if (!bdi_write_congested(bdi))
265 if (bdi == current->backing_dev_info)
271 * We detected a synchronous write error writing a page out. Probably
272 * -ENOSPC. We need to propagate that into the address_space for a subsequent
273 * fsync(), msync() or close().
275 * The tricky part is that after writepage we cannot touch the mapping: nothing
276 * prevents it from being freed up. But we have a ref on the page and once
277 * that page is locked, the mapping is pinned.
279 * We're allowed to run sleeping lock_page() here because we know the caller has
282 static void handle_write_error(struct address_space *mapping,
283 struct page *page, int error)
286 if (page_mapping(page) == mapping)
287 mapping_set_error(mapping, error);
291 /* Request for sync pageout. */
297 /* possible outcome of pageout() */
299 /* failed to write page out, page is locked */
301 /* move page to the active list, page is locked */
303 /* page has been sent to the disk successfully, page is unlocked */
305 /* page is clean and locked */
310 * pageout is called by shrink_page_list() for each dirty page.
311 * Calls ->writepage().
313 static pageout_t pageout(struct page *page, struct address_space *mapping,
314 enum pageout_io sync_writeback)
317 * If the page is dirty, only perform writeback if that write
318 * will be non-blocking. To prevent this allocation from being
319 * stalled by pagecache activity. But note that there may be
320 * stalls if we need to run get_block(). We could test
321 * PagePrivate for that.
323 * If this process is currently in generic_file_write() against
324 * this page's queue, we can perform writeback even if that
327 * If the page is swapcache, write it back even if that would
328 * block, for some throttling. This happens by accident, because
329 * swap_backing_dev_info is bust: it doesn't reflect the
330 * congestion state of the swapdevs. Easy to fix, if needed.
331 * See swapfile.c:page_queue_congested().
333 if (!is_page_cache_freeable(page))
337 * Some data journaling orphaned pages can have
338 * page->mapping == NULL while being dirty with clean buffers.
340 if (PagePrivate(page)) {
341 if (try_to_free_buffers(page)) {
342 ClearPageDirty(page);
343 printk("%s: orphaned page\n", __func__);
349 if (mapping->a_ops->writepage == NULL)
350 return PAGE_ACTIVATE;
351 if (!may_write_to_queue(mapping->backing_dev_info))
354 if (clear_page_dirty_for_io(page)) {
356 struct writeback_control wbc = {
357 .sync_mode = WB_SYNC_NONE,
358 .nr_to_write = SWAP_CLUSTER_MAX,
360 .range_end = LLONG_MAX,
365 SetPageReclaim(page);
366 res = mapping->a_ops->writepage(page, &wbc);
368 handle_write_error(mapping, page, res);
369 if (res == AOP_WRITEPAGE_ACTIVATE) {
370 ClearPageReclaim(page);
371 return PAGE_ACTIVATE;
375 * Wait on writeback if requested to. This happens when
376 * direct reclaiming a large contiguous area and the
377 * first attempt to free a range of pages fails.
379 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
380 wait_on_page_writeback(page);
382 if (!PageWriteback(page)) {
383 /* synchronous write or broken a_ops? */
384 ClearPageReclaim(page);
386 inc_zone_page_state(page, NR_VMSCAN_WRITE);
394 * Same as remove_mapping, but if the page is removed from the mapping, it
395 * gets returned with a refcount of 0.
397 static int __remove_mapping(struct address_space *mapping, struct page *page)
399 BUG_ON(!PageLocked(page));
400 BUG_ON(mapping != page_mapping(page));
402 spin_lock_irq(&mapping->tree_lock);
404 * The non racy check for a busy page.
406 * Must be careful with the order of the tests. When someone has
407 * a ref to the page, it may be possible that they dirty it then
408 * drop the reference. So if PageDirty is tested before page_count
409 * here, then the following race may occur:
411 * get_user_pages(&page);
412 * [user mapping goes away]
414 * !PageDirty(page) [good]
415 * SetPageDirty(page);
417 * !page_count(page) [good, discard it]
419 * [oops, our write_to data is lost]
421 * Reversing the order of the tests ensures such a situation cannot
422 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
423 * load is not satisfied before that of page->_count.
425 * Note that if SetPageDirty is always performed via set_page_dirty,
426 * and thus under tree_lock, then this ordering is not required.
428 if (!page_freeze_refs(page, 2))
430 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
431 if (unlikely(PageDirty(page))) {
432 page_unfreeze_refs(page, 2);
436 if (PageSwapCache(page)) {
437 swp_entry_t swap = { .val = page_private(page) };
438 __delete_from_swap_cache(page);
439 spin_unlock_irq(&mapping->tree_lock);
442 __remove_from_page_cache(page);
443 spin_unlock_irq(&mapping->tree_lock);
449 spin_unlock_irq(&mapping->tree_lock);
454 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
455 * someone else has a ref on the page, abort and return 0. If it was
456 * successfully detached, return 1. Assumes the caller has a single ref on
459 int remove_mapping(struct address_space *mapping, struct page *page)
461 if (__remove_mapping(mapping, page)) {
463 * Unfreezing the refcount with 1 rather than 2 effectively
464 * drops the pagecache ref for us without requiring another
467 page_unfreeze_refs(page, 1);
474 * shrink_page_list() returns the number of reclaimed pages
476 static unsigned long shrink_page_list(struct list_head *page_list,
477 struct scan_control *sc,
478 enum pageout_io sync_writeback)
480 LIST_HEAD(ret_pages);
481 struct pagevec freed_pvec;
483 unsigned long nr_reclaimed = 0;
487 pagevec_init(&freed_pvec, 1);
488 while (!list_empty(page_list)) {
489 struct address_space *mapping;
496 page = lru_to_page(page_list);
497 list_del(&page->lru);
499 if (!trylock_page(page))
502 VM_BUG_ON(PageActive(page));
506 if (!sc->may_swap && page_mapped(page))
509 /* Double the slab pressure for mapped and swapcache pages */
510 if (page_mapped(page) || PageSwapCache(page))
513 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
514 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
516 if (PageWriteback(page)) {
518 * Synchronous reclaim is performed in two passes,
519 * first an asynchronous pass over the list to
520 * start parallel writeback, and a second synchronous
521 * pass to wait for the IO to complete. Wait here
522 * for any page for which writeback has already
525 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
526 wait_on_page_writeback(page);
531 referenced = page_referenced(page, 1, sc->mem_cgroup);
532 /* In active use or really unfreeable? Activate it. */
533 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
534 referenced && page_mapping_inuse(page))
535 goto activate_locked;
539 * Anonymous process memory has backing store?
540 * Try to allocate it some swap space here.
542 if (PageAnon(page) && !PageSwapCache(page))
543 if (!add_to_swap(page, GFP_ATOMIC))
544 goto activate_locked;
545 #endif /* CONFIG_SWAP */
547 mapping = page_mapping(page);
550 * The page is mapped into the page tables of one or more
551 * processes. Try to unmap it here.
553 if (page_mapped(page) && mapping) {
554 switch (try_to_unmap(page, 0)) {
556 goto activate_locked;
560 ; /* try to free the page below */
564 if (PageDirty(page)) {
565 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
569 if (!sc->may_writepage)
572 /* Page is dirty, try to write it out here */
573 switch (pageout(page, mapping, sync_writeback)) {
577 goto activate_locked;
579 if (PageWriteback(page) || PageDirty(page))
582 * A synchronous write - probably a ramdisk. Go
583 * ahead and try to reclaim the page.
585 if (!trylock_page(page))
587 if (PageDirty(page) || PageWriteback(page))
589 mapping = page_mapping(page);
591 ; /* try to free the page below */
596 * If the page has buffers, try to free the buffer mappings
597 * associated with this page. If we succeed we try to free
600 * We do this even if the page is PageDirty().
601 * try_to_release_page() does not perform I/O, but it is
602 * possible for a page to have PageDirty set, but it is actually
603 * clean (all its buffers are clean). This happens if the
604 * buffers were written out directly, with submit_bh(). ext3
605 * will do this, as well as the blockdev mapping.
606 * try_to_release_page() will discover that cleanness and will
607 * drop the buffers and mark the page clean - it can be freed.
609 * Rarely, pages can have buffers and no ->mapping. These are
610 * the pages which were not successfully invalidated in
611 * truncate_complete_page(). We try to drop those buffers here
612 * and if that worked, and the page is no longer mapped into
613 * process address space (page_count == 1) it can be freed.
614 * Otherwise, leave the page on the LRU so it is swappable.
616 if (PagePrivate(page)) {
617 if (!try_to_release_page(page, sc->gfp_mask))
618 goto activate_locked;
619 if (!mapping && page_count(page) == 1) {
621 if (put_page_testzero(page))
625 * rare race with speculative reference.
626 * the speculative reference will free
627 * this page shortly, so we may
628 * increment nr_reclaimed here (and
629 * leave it off the LRU).
637 if (!mapping || !__remove_mapping(mapping, page))
643 if (!pagevec_add(&freed_pvec, page)) {
644 __pagevec_free(&freed_pvec);
645 pagevec_reinit(&freed_pvec);
655 list_add(&page->lru, &ret_pages);
656 VM_BUG_ON(PageLRU(page));
658 list_splice(&ret_pages, page_list);
659 if (pagevec_count(&freed_pvec))
660 __pagevec_free(&freed_pvec);
661 count_vm_events(PGACTIVATE, pgactivate);
665 /* LRU Isolation modes. */
666 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
667 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
668 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
671 * Attempt to remove the specified page from its LRU. Only take this page
672 * if it is of the appropriate PageActive status. Pages which are being
673 * freed elsewhere are also ignored.
675 * page: page to consider
676 * mode: one of the LRU isolation modes defined above
678 * returns 0 on success, -ve errno on failure.
680 int __isolate_lru_page(struct page *page, int mode)
684 /* Only take pages on the LRU. */
689 * When checking the active state, we need to be sure we are
690 * dealing with comparible boolean values. Take the logical not
693 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
697 if (likely(get_page_unless_zero(page))) {
699 * Be careful not to clear PageLRU until after we're
700 * sure the page is not being freed elsewhere -- the
701 * page release code relies on it.
711 * zone->lru_lock is heavily contended. Some of the functions that
712 * shrink the lists perform better by taking out a batch of pages
713 * and working on them outside the LRU lock.
715 * For pagecache intensive workloads, this function is the hottest
716 * spot in the kernel (apart from copy_*_user functions).
718 * Appropriate locks must be held before calling this function.
720 * @nr_to_scan: The number of pages to look through on the list.
721 * @src: The LRU list to pull pages off.
722 * @dst: The temp list to put pages on to.
723 * @scanned: The number of pages that were scanned.
724 * @order: The caller's attempted allocation order
725 * @mode: One of the LRU isolation modes
727 * returns how many pages were moved onto *@dst.
729 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
730 struct list_head *src, struct list_head *dst,
731 unsigned long *scanned, int order, int mode)
733 unsigned long nr_taken = 0;
736 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
739 unsigned long end_pfn;
740 unsigned long page_pfn;
743 page = lru_to_page(src);
744 prefetchw_prev_lru_page(page, src, flags);
746 VM_BUG_ON(!PageLRU(page));
748 switch (__isolate_lru_page(page, mode)) {
750 list_move(&page->lru, dst);
755 /* else it is being freed elsewhere */
756 list_move(&page->lru, src);
767 * Attempt to take all pages in the order aligned region
768 * surrounding the tag page. Only take those pages of
769 * the same active state as that tag page. We may safely
770 * round the target page pfn down to the requested order
771 * as the mem_map is guarenteed valid out to MAX_ORDER,
772 * where that page is in a different zone we will detect
773 * it from its zone id and abort this block scan.
775 zone_id = page_zone_id(page);
776 page_pfn = page_to_pfn(page);
777 pfn = page_pfn & ~((1 << order) - 1);
778 end_pfn = pfn + (1 << order);
779 for (; pfn < end_pfn; pfn++) {
780 struct page *cursor_page;
782 /* The target page is in the block, ignore it. */
783 if (unlikely(pfn == page_pfn))
786 /* Avoid holes within the zone. */
787 if (unlikely(!pfn_valid_within(pfn)))
790 cursor_page = pfn_to_page(pfn);
791 /* Check that we have not crossed a zone boundary. */
792 if (unlikely(page_zone_id(cursor_page) != zone_id))
794 switch (__isolate_lru_page(cursor_page, mode)) {
796 list_move(&cursor_page->lru, dst);
802 /* else it is being freed elsewhere */
803 list_move(&cursor_page->lru, src);
814 static unsigned long isolate_pages_global(unsigned long nr,
815 struct list_head *dst,
816 unsigned long *scanned, int order,
817 int mode, struct zone *z,
818 struct mem_cgroup *mem_cont,
822 return isolate_lru_pages(nr, &z->lru[LRU_ACTIVE].list, dst,
823 scanned, order, mode);
825 return isolate_lru_pages(nr, &z->lru[LRU_INACTIVE].list, dst,
826 scanned, order, mode);
830 * clear_active_flags() is a helper for shrink_active_list(), clearing
831 * any active bits from the pages in the list.
833 static unsigned long clear_active_flags(struct list_head *page_list)
838 list_for_each_entry(page, page_list, lru)
839 if (PageActive(page)) {
840 ClearPageActive(page);
848 * isolate_lru_page - tries to isolate a page from its LRU list
849 * @page: page to isolate from its LRU list
851 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
852 * vmstat statistic corresponding to whatever LRU list the page was on.
854 * Returns 0 if the page was removed from an LRU list.
855 * Returns -EBUSY if the page was not on an LRU list.
857 * The returned page will have PageLRU() cleared. If it was found on
858 * the active list, it will have PageActive set. That flag may need
859 * to be cleared by the caller before letting the page go.
861 * The vmstat statistic corresponding to the list on which the page was
862 * found will be decremented.
865 * (1) Must be called with an elevated refcount on the page. This is a
866 * fundamentnal difference from isolate_lru_pages (which is called
867 * without a stable reference).
868 * (2) the lru_lock must not be held.
869 * (3) interrupts must be enabled.
871 int isolate_lru_page(struct page *page)
876 struct zone *zone = page_zone(page);
878 spin_lock_irq(&zone->lru_lock);
879 if (PageLRU(page) && get_page_unless_zero(page)) {
882 if (PageActive(page))
883 del_page_from_active_list(zone, page);
885 del_page_from_inactive_list(zone, page);
887 spin_unlock_irq(&zone->lru_lock);
893 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
896 static unsigned long shrink_inactive_list(unsigned long max_scan,
897 struct zone *zone, struct scan_control *sc)
899 LIST_HEAD(page_list);
901 unsigned long nr_scanned = 0;
902 unsigned long nr_reclaimed = 0;
904 pagevec_init(&pvec, 1);
907 spin_lock_irq(&zone->lru_lock);
910 unsigned long nr_taken;
911 unsigned long nr_scan;
912 unsigned long nr_freed;
913 unsigned long nr_active;
915 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
916 &page_list, &nr_scan, sc->order,
917 (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
918 ISOLATE_BOTH : ISOLATE_INACTIVE,
919 zone, sc->mem_cgroup, 0);
920 nr_active = clear_active_flags(&page_list);
921 __count_vm_events(PGDEACTIVATE, nr_active);
923 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
924 __mod_zone_page_state(zone, NR_INACTIVE,
925 -(nr_taken - nr_active));
926 if (scan_global_lru(sc))
927 zone->pages_scanned += nr_scan;
928 spin_unlock_irq(&zone->lru_lock);
930 nr_scanned += nr_scan;
931 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
934 * If we are direct reclaiming for contiguous pages and we do
935 * not reclaim everything in the list, try again and wait
936 * for IO to complete. This will stall high-order allocations
937 * but that should be acceptable to the caller
939 if (nr_freed < nr_taken && !current_is_kswapd() &&
940 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
941 congestion_wait(WRITE, HZ/10);
944 * The attempt at page out may have made some
945 * of the pages active, mark them inactive again.
947 nr_active = clear_active_flags(&page_list);
948 count_vm_events(PGDEACTIVATE, nr_active);
950 nr_freed += shrink_page_list(&page_list, sc,
954 nr_reclaimed += nr_freed;
956 if (current_is_kswapd()) {
957 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
958 __count_vm_events(KSWAPD_STEAL, nr_freed);
959 } else if (scan_global_lru(sc))
960 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
962 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
967 spin_lock(&zone->lru_lock);
969 * Put back any unfreeable pages.
971 while (!list_empty(&page_list)) {
972 page = lru_to_page(&page_list);
973 VM_BUG_ON(PageLRU(page));
975 list_del(&page->lru);
976 add_page_to_lru_list(zone, page, page_lru(page));
977 if (!pagevec_add(&pvec, page)) {
978 spin_unlock_irq(&zone->lru_lock);
979 __pagevec_release(&pvec);
980 spin_lock_irq(&zone->lru_lock);
983 } while (nr_scanned < max_scan);
984 spin_unlock(&zone->lru_lock);
987 pagevec_release(&pvec);
992 * We are about to scan this zone at a certain priority level. If that priority
993 * level is smaller (ie: more urgent) than the previous priority, then note
994 * that priority level within the zone. This is done so that when the next
995 * process comes in to scan this zone, it will immediately start out at this
996 * priority level rather than having to build up its own scanning priority.
997 * Here, this priority affects only the reclaim-mapped threshold.
999 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1001 if (priority < zone->prev_priority)
1002 zone->prev_priority = priority;
1005 static inline int zone_is_near_oom(struct zone *zone)
1007 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
1008 + zone_page_state(zone, NR_INACTIVE))*3;
1012 * Determine we should try to reclaim mapped pages.
1013 * This is called only when sc->mem_cgroup is NULL.
1015 static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
1022 int reclaim_mapped = 0;
1025 if (scan_global_lru(sc) && zone_is_near_oom(zone))
1028 * `distress' is a measure of how much trouble we're having
1029 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1031 if (scan_global_lru(sc))
1032 prev_priority = zone->prev_priority;
1034 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
1036 distress = 100 >> min(prev_priority, priority);
1039 * The point of this algorithm is to decide when to start
1040 * reclaiming mapped memory instead of just pagecache. Work out
1044 if (scan_global_lru(sc))
1045 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
1046 global_page_state(NR_ANON_PAGES)) * 100) /
1049 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
1052 * Now decide how much we really want to unmap some pages. The
1053 * mapped ratio is downgraded - just because there's a lot of
1054 * mapped memory doesn't necessarily mean that page reclaim
1057 * The distress ratio is important - we don't want to start
1060 * A 100% value of vm_swappiness overrides this algorithm
1063 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1066 * If there's huge imbalance between active and inactive
1067 * (think active 100 times larger than inactive) we should
1068 * become more permissive, or the system will take too much
1069 * cpu before it start swapping during memory pressure.
1070 * Distress is about avoiding early-oom, this is about
1071 * making swappiness graceful despite setting it to low
1074 * Avoid div by zero with nr_inactive+1, and max resulting
1075 * value is vm_total_pages.
1077 if (scan_global_lru(sc)) {
1078 imbalance = zone_page_state(zone, NR_ACTIVE);
1079 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1081 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1084 * Reduce the effect of imbalance if swappiness is low,
1085 * this means for a swappiness very low, the imbalance
1086 * must be much higher than 100 for this logic to make
1089 * Max temporary value is vm_total_pages*100.
1091 imbalance *= (vm_swappiness + 1);
1095 * If not much of the ram is mapped, makes the imbalance
1096 * less relevant, it's high priority we refill the inactive
1097 * list with mapped pages only in presence of high ratio of
1100 * Max temporary value is vm_total_pages*100.
1102 imbalance *= mapped_ratio;
1105 /* apply imbalance feedback to swap_tendency */
1106 swap_tendency += imbalance;
1109 * Now use this metric to decide whether to start moving mapped
1110 * memory onto the inactive list.
1112 if (swap_tendency >= 100)
1115 return reclaim_mapped;
1119 * This moves pages from the active list to the inactive list.
1121 * We move them the other way if the page is referenced by one or more
1122 * processes, from rmap.
1124 * If the pages are mostly unmapped, the processing is fast and it is
1125 * appropriate to hold zone->lru_lock across the whole operation. But if
1126 * the pages are mapped, the processing is slow (page_referenced()) so we
1127 * should drop zone->lru_lock around each page. It's impossible to balance
1128 * this, so instead we remove the pages from the LRU while processing them.
1129 * It is safe to rely on PG_active against the non-LRU pages in here because
1130 * nobody will play with that bit on a non-LRU page.
1132 * The downside is that we have to touch page->_count against each page.
1133 * But we had to alter page->flags anyway.
1137 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1138 struct scan_control *sc, int priority)
1140 unsigned long pgmoved;
1141 int pgdeactivate = 0;
1142 unsigned long pgscanned;
1143 LIST_HEAD(l_hold); /* The pages which were snipped off */
1144 LIST_HEAD(l_active);
1145 LIST_HEAD(l_inactive);
1147 struct pagevec pvec;
1148 int reclaim_mapped = 0;
1151 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
1154 spin_lock_irq(&zone->lru_lock);
1155 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1156 ISOLATE_ACTIVE, zone,
1159 * zone->pages_scanned is used for detect zone's oom
1160 * mem_cgroup remembers nr_scan by itself.
1162 if (scan_global_lru(sc))
1163 zone->pages_scanned += pgscanned;
1165 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1166 spin_unlock_irq(&zone->lru_lock);
1168 while (!list_empty(&l_hold)) {
1170 page = lru_to_page(&l_hold);
1171 list_del(&page->lru);
1172 if (page_mapped(page)) {
1173 if (!reclaim_mapped ||
1174 (total_swap_pages == 0 && PageAnon(page)) ||
1175 page_referenced(page, 0, sc->mem_cgroup)) {
1176 list_add(&page->lru, &l_active);
1180 list_add(&page->lru, &l_inactive);
1183 pagevec_init(&pvec, 1);
1185 spin_lock_irq(&zone->lru_lock);
1186 while (!list_empty(&l_inactive)) {
1187 page = lru_to_page(&l_inactive);
1188 prefetchw_prev_lru_page(page, &l_inactive, flags);
1189 VM_BUG_ON(PageLRU(page));
1191 VM_BUG_ON(!PageActive(page));
1192 ClearPageActive(page);
1194 list_move(&page->lru, &zone->lru[LRU_INACTIVE].list);
1195 mem_cgroup_move_lists(page, false);
1197 if (!pagevec_add(&pvec, page)) {
1198 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1199 spin_unlock_irq(&zone->lru_lock);
1200 pgdeactivate += pgmoved;
1202 if (buffer_heads_over_limit)
1203 pagevec_strip(&pvec);
1204 __pagevec_release(&pvec);
1205 spin_lock_irq(&zone->lru_lock);
1208 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1209 pgdeactivate += pgmoved;
1210 if (buffer_heads_over_limit) {
1211 spin_unlock_irq(&zone->lru_lock);
1212 pagevec_strip(&pvec);
1213 spin_lock_irq(&zone->lru_lock);
1217 while (!list_empty(&l_active)) {
1218 page = lru_to_page(&l_active);
1219 prefetchw_prev_lru_page(page, &l_active, flags);
1220 VM_BUG_ON(PageLRU(page));
1222 VM_BUG_ON(!PageActive(page));
1224 list_move(&page->lru, &zone->lru[LRU_ACTIVE].list);
1225 mem_cgroup_move_lists(page, true);
1227 if (!pagevec_add(&pvec, page)) {
1228 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1230 spin_unlock_irq(&zone->lru_lock);
1231 __pagevec_release(&pvec);
1232 spin_lock_irq(&zone->lru_lock);
1235 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1237 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1238 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1239 spin_unlock_irq(&zone->lru_lock);
1241 pagevec_release(&pvec);
1244 static unsigned long shrink_list(enum lru_list l, unsigned long nr_to_scan,
1245 struct zone *zone, struct scan_control *sc, int priority)
1247 if (l == LRU_ACTIVE) {
1248 shrink_active_list(nr_to_scan, zone, sc, priority);
1251 return shrink_inactive_list(nr_to_scan, zone, sc);
1255 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1257 static unsigned long shrink_zone(int priority, struct zone *zone,
1258 struct scan_control *sc)
1260 unsigned long nr[NR_LRU_LISTS];
1261 unsigned long nr_to_scan;
1262 unsigned long nr_reclaimed = 0;
1265 if (scan_global_lru(sc)) {
1267 * Add one to nr_to_scan just to make sure that the kernel
1268 * will slowly sift through the active list.
1271 zone->lru[l].nr_scan += (zone_page_state(zone,
1272 NR_LRU_BASE + l) >> priority) + 1;
1273 nr[l] = zone->lru[l].nr_scan;
1274 if (nr[l] >= sc->swap_cluster_max)
1275 zone->lru[l].nr_scan = 0;
1281 * This reclaim occurs not because zone memory shortage but
1282 * because memory controller hits its limit.
1283 * Then, don't modify zone reclaim related data.
1285 nr[LRU_ACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1286 zone, priority, LRU_ACTIVE);
1288 nr[LRU_INACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1289 zone, priority, LRU_INACTIVE);
1292 while (nr[LRU_ACTIVE] || nr[LRU_INACTIVE]) {
1295 nr_to_scan = min(nr[l],
1296 (unsigned long)sc->swap_cluster_max);
1297 nr[l] -= nr_to_scan;
1299 nr_reclaimed += shrink_list(l, nr_to_scan,
1300 zone, sc, priority);
1305 throttle_vm_writeout(sc->gfp_mask);
1306 return nr_reclaimed;
1310 * This is the direct reclaim path, for page-allocating processes. We only
1311 * try to reclaim pages from zones which will satisfy the caller's allocation
1314 * We reclaim from a zone even if that zone is over pages_high. Because:
1315 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1317 * b) The zones may be over pages_high but they must go *over* pages_high to
1318 * satisfy the `incremental min' zone defense algorithm.
1320 * Returns the number of reclaimed pages.
1322 * If a zone is deemed to be full of pinned pages then just give it a light
1323 * scan then give up on it.
1325 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1326 struct scan_control *sc)
1328 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1329 unsigned long nr_reclaimed = 0;
1333 sc->all_unreclaimable = 1;
1334 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1335 if (!populated_zone(zone))
1338 * Take care memory controller reclaiming has small influence
1341 if (scan_global_lru(sc)) {
1342 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1344 note_zone_scanning_priority(zone, priority);
1346 if (zone_is_all_unreclaimable(zone) &&
1347 priority != DEF_PRIORITY)
1348 continue; /* Let kswapd poll it */
1349 sc->all_unreclaimable = 0;
1352 * Ignore cpuset limitation here. We just want to reduce
1353 * # of used pages by us regardless of memory shortage.
1355 sc->all_unreclaimable = 0;
1356 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1360 nr_reclaimed += shrink_zone(priority, zone, sc);
1363 return nr_reclaimed;
1367 * This is the main entry point to direct page reclaim.
1369 * If a full scan of the inactive list fails to free enough memory then we
1370 * are "out of memory" and something needs to be killed.
1372 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1373 * high - the zone may be full of dirty or under-writeback pages, which this
1374 * caller can't do much about. We kick pdflush and take explicit naps in the
1375 * hope that some of these pages can be written. But if the allocating task
1376 * holds filesystem locks which prevent writeout this might not work, and the
1377 * allocation attempt will fail.
1379 * returns: 0, if no pages reclaimed
1380 * else, the number of pages reclaimed
1382 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1383 struct scan_control *sc)
1386 unsigned long ret = 0;
1387 unsigned long total_scanned = 0;
1388 unsigned long nr_reclaimed = 0;
1389 struct reclaim_state *reclaim_state = current->reclaim_state;
1390 unsigned long lru_pages = 0;
1393 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1395 delayacct_freepages_start();
1397 if (scan_global_lru(sc))
1398 count_vm_event(ALLOCSTALL);
1400 * mem_cgroup will not do shrink_slab.
1402 if (scan_global_lru(sc)) {
1403 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1405 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1408 lru_pages += zone_page_state(zone, NR_ACTIVE)
1409 + zone_page_state(zone, NR_INACTIVE);
1413 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1416 disable_swap_token();
1417 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1419 * Don't shrink slabs when reclaiming memory from
1420 * over limit cgroups
1422 if (scan_global_lru(sc)) {
1423 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1424 if (reclaim_state) {
1425 nr_reclaimed += reclaim_state->reclaimed_slab;
1426 reclaim_state->reclaimed_slab = 0;
1429 total_scanned += sc->nr_scanned;
1430 if (nr_reclaimed >= sc->swap_cluster_max) {
1436 * Try to write back as many pages as we just scanned. This
1437 * tends to cause slow streaming writers to write data to the
1438 * disk smoothly, at the dirtying rate, which is nice. But
1439 * that's undesirable in laptop mode, where we *want* lumpy
1440 * writeout. So in laptop mode, write out the whole world.
1442 if (total_scanned > sc->swap_cluster_max +
1443 sc->swap_cluster_max / 2) {
1444 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1445 sc->may_writepage = 1;
1448 /* Take a nap, wait for some writeback to complete */
1449 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1450 congestion_wait(WRITE, HZ/10);
1452 /* top priority shrink_zones still had more to do? don't OOM, then */
1453 if (!sc->all_unreclaimable && scan_global_lru(sc))
1457 * Now that we've scanned all the zones at this priority level, note
1458 * that level within the zone so that the next thread which performs
1459 * scanning of this zone will immediately start out at this priority
1460 * level. This affects only the decision whether or not to bring
1461 * mapped pages onto the inactive list.
1466 if (scan_global_lru(sc)) {
1467 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1469 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1472 zone->prev_priority = priority;
1475 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1477 delayacct_freepages_end();
1482 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1485 struct scan_control sc = {
1486 .gfp_mask = gfp_mask,
1487 .may_writepage = !laptop_mode,
1488 .swap_cluster_max = SWAP_CLUSTER_MAX,
1490 .swappiness = vm_swappiness,
1493 .isolate_pages = isolate_pages_global,
1496 return do_try_to_free_pages(zonelist, &sc);
1499 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1501 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1504 struct scan_control sc = {
1505 .may_writepage = !laptop_mode,
1507 .swap_cluster_max = SWAP_CLUSTER_MAX,
1508 .swappiness = vm_swappiness,
1510 .mem_cgroup = mem_cont,
1511 .isolate_pages = mem_cgroup_isolate_pages,
1513 struct zonelist *zonelist;
1515 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1516 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1517 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1518 return do_try_to_free_pages(zonelist, &sc);
1523 * For kswapd, balance_pgdat() will work across all this node's zones until
1524 * they are all at pages_high.
1526 * Returns the number of pages which were actually freed.
1528 * There is special handling here for zones which are full of pinned pages.
1529 * This can happen if the pages are all mlocked, or if they are all used by
1530 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1531 * What we do is to detect the case where all pages in the zone have been
1532 * scanned twice and there has been zero successful reclaim. Mark the zone as
1533 * dead and from now on, only perform a short scan. Basically we're polling
1534 * the zone for when the problem goes away.
1536 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1537 * zones which have free_pages > pages_high, but once a zone is found to have
1538 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1539 * of the number of free pages in the lower zones. This interoperates with
1540 * the page allocator fallback scheme to ensure that aging of pages is balanced
1543 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1548 unsigned long total_scanned;
1549 unsigned long nr_reclaimed;
1550 struct reclaim_state *reclaim_state = current->reclaim_state;
1551 struct scan_control sc = {
1552 .gfp_mask = GFP_KERNEL,
1554 .swap_cluster_max = SWAP_CLUSTER_MAX,
1555 .swappiness = vm_swappiness,
1558 .isolate_pages = isolate_pages_global,
1561 * temp_priority is used to remember the scanning priority at which
1562 * this zone was successfully refilled to free_pages == pages_high.
1564 int temp_priority[MAX_NR_ZONES];
1569 sc.may_writepage = !laptop_mode;
1570 count_vm_event(PAGEOUTRUN);
1572 for (i = 0; i < pgdat->nr_zones; i++)
1573 temp_priority[i] = DEF_PRIORITY;
1575 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1576 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1577 unsigned long lru_pages = 0;
1579 /* The swap token gets in the way of swapout... */
1581 disable_swap_token();
1586 * Scan in the highmem->dma direction for the highest
1587 * zone which needs scanning
1589 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1590 struct zone *zone = pgdat->node_zones + i;
1592 if (!populated_zone(zone))
1595 if (zone_is_all_unreclaimable(zone) &&
1596 priority != DEF_PRIORITY)
1599 if (!zone_watermark_ok(zone, order, zone->pages_high,
1608 for (i = 0; i <= end_zone; i++) {
1609 struct zone *zone = pgdat->node_zones + i;
1611 lru_pages += zone_page_state(zone, NR_ACTIVE)
1612 + zone_page_state(zone, NR_INACTIVE);
1616 * Now scan the zone in the dma->highmem direction, stopping
1617 * at the last zone which needs scanning.
1619 * We do this because the page allocator works in the opposite
1620 * direction. This prevents the page allocator from allocating
1621 * pages behind kswapd's direction of progress, which would
1622 * cause too much scanning of the lower zones.
1624 for (i = 0; i <= end_zone; i++) {
1625 struct zone *zone = pgdat->node_zones + i;
1628 if (!populated_zone(zone))
1631 if (zone_is_all_unreclaimable(zone) &&
1632 priority != DEF_PRIORITY)
1635 if (!zone_watermark_ok(zone, order, zone->pages_high,
1638 temp_priority[i] = priority;
1640 note_zone_scanning_priority(zone, priority);
1642 * We put equal pressure on every zone, unless one
1643 * zone has way too many pages free already.
1645 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1647 nr_reclaimed += shrink_zone(priority, zone, &sc);
1648 reclaim_state->reclaimed_slab = 0;
1649 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1651 nr_reclaimed += reclaim_state->reclaimed_slab;
1652 total_scanned += sc.nr_scanned;
1653 if (zone_is_all_unreclaimable(zone))
1655 if (nr_slab == 0 && zone->pages_scanned >=
1656 (zone_page_state(zone, NR_ACTIVE)
1657 + zone_page_state(zone, NR_INACTIVE)) * 6)
1659 ZONE_ALL_UNRECLAIMABLE);
1661 * If we've done a decent amount of scanning and
1662 * the reclaim ratio is low, start doing writepage
1663 * even in laptop mode
1665 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1666 total_scanned > nr_reclaimed + nr_reclaimed / 2)
1667 sc.may_writepage = 1;
1670 break; /* kswapd: all done */
1672 * OK, kswapd is getting into trouble. Take a nap, then take
1673 * another pass across the zones.
1675 if (total_scanned && priority < DEF_PRIORITY - 2)
1676 congestion_wait(WRITE, HZ/10);
1679 * We do this so kswapd doesn't build up large priorities for
1680 * example when it is freeing in parallel with allocators. It
1681 * matches the direct reclaim path behaviour in terms of impact
1682 * on zone->*_priority.
1684 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1689 * Note within each zone the priority level at which this zone was
1690 * brought into a happy state. So that the next thread which scans this
1691 * zone will start out at that priority level.
1693 for (i = 0; i < pgdat->nr_zones; i++) {
1694 struct zone *zone = pgdat->node_zones + i;
1696 zone->prev_priority = temp_priority[i];
1698 if (!all_zones_ok) {
1706 return nr_reclaimed;
1710 * The background pageout daemon, started as a kernel thread
1711 * from the init process.
1713 * This basically trickles out pages so that we have _some_
1714 * free memory available even if there is no other activity
1715 * that frees anything up. This is needed for things like routing
1716 * etc, where we otherwise might have all activity going on in
1717 * asynchronous contexts that cannot page things out.
1719 * If there are applications that are active memory-allocators
1720 * (most normal use), this basically shouldn't matter.
1722 static int kswapd(void *p)
1724 unsigned long order;
1725 pg_data_t *pgdat = (pg_data_t*)p;
1726 struct task_struct *tsk = current;
1728 struct reclaim_state reclaim_state = {
1729 .reclaimed_slab = 0,
1731 node_to_cpumask_ptr(cpumask, pgdat->node_id);
1733 if (!cpus_empty(*cpumask))
1734 set_cpus_allowed_ptr(tsk, cpumask);
1735 current->reclaim_state = &reclaim_state;
1738 * Tell the memory management that we're a "memory allocator",
1739 * and that if we need more memory we should get access to it
1740 * regardless (see "__alloc_pages()"). "kswapd" should
1741 * never get caught in the normal page freeing logic.
1743 * (Kswapd normally doesn't need memory anyway, but sometimes
1744 * you need a small amount of memory in order to be able to
1745 * page out something else, and this flag essentially protects
1746 * us from recursively trying to free more memory as we're
1747 * trying to free the first piece of memory in the first place).
1749 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1754 unsigned long new_order;
1756 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1757 new_order = pgdat->kswapd_max_order;
1758 pgdat->kswapd_max_order = 0;
1759 if (order < new_order) {
1761 * Don't sleep if someone wants a larger 'order'
1766 if (!freezing(current))
1769 order = pgdat->kswapd_max_order;
1771 finish_wait(&pgdat->kswapd_wait, &wait);
1773 if (!try_to_freeze()) {
1774 /* We can speed up thawing tasks if we don't call
1775 * balance_pgdat after returning from the refrigerator
1777 balance_pgdat(pgdat, order);
1784 * A zone is low on free memory, so wake its kswapd task to service it.
1786 void wakeup_kswapd(struct zone *zone, int order)
1790 if (!populated_zone(zone))
1793 pgdat = zone->zone_pgdat;
1794 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1796 if (pgdat->kswapd_max_order < order)
1797 pgdat->kswapd_max_order = order;
1798 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1800 if (!waitqueue_active(&pgdat->kswapd_wait))
1802 wake_up_interruptible(&pgdat->kswapd_wait);
1807 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1808 * from LRU lists system-wide, for given pass and priority, and returns the
1809 * number of reclaimed pages
1811 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1813 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1814 int pass, struct scan_control *sc)
1817 unsigned long nr_to_scan, ret = 0;
1820 for_each_zone(zone) {
1822 if (!populated_zone(zone))
1825 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1829 /* For pass = 0 we don't shrink the active list */
1830 if (pass == 0 && l == LRU_ACTIVE)
1833 zone->lru[l].nr_scan +=
1834 (zone_page_state(zone, NR_LRU_BASE + l)
1836 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
1837 zone->lru[l].nr_scan = 0;
1838 nr_to_scan = min(nr_pages,
1839 zone_page_state(zone,
1841 ret += shrink_list(l, nr_to_scan, zone,
1843 if (ret >= nr_pages)
1852 static unsigned long count_lru_pages(void)
1854 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1858 * Try to free `nr_pages' of memory, system-wide, and return the number of
1861 * Rather than trying to age LRUs the aim is to preserve the overall
1862 * LRU order by reclaiming preferentially
1863 * inactive > active > active referenced > active mapped
1865 unsigned long shrink_all_memory(unsigned long nr_pages)
1867 unsigned long lru_pages, nr_slab;
1868 unsigned long ret = 0;
1870 struct reclaim_state reclaim_state;
1871 struct scan_control sc = {
1872 .gfp_mask = GFP_KERNEL,
1874 .swap_cluster_max = nr_pages,
1876 .swappiness = vm_swappiness,
1877 .isolate_pages = isolate_pages_global,
1880 current->reclaim_state = &reclaim_state;
1882 lru_pages = count_lru_pages();
1883 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1884 /* If slab caches are huge, it's better to hit them first */
1885 while (nr_slab >= lru_pages) {
1886 reclaim_state.reclaimed_slab = 0;
1887 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1888 if (!reclaim_state.reclaimed_slab)
1891 ret += reclaim_state.reclaimed_slab;
1892 if (ret >= nr_pages)
1895 nr_slab -= reclaim_state.reclaimed_slab;
1899 * We try to shrink LRUs in 5 passes:
1900 * 0 = Reclaim from inactive_list only
1901 * 1 = Reclaim from active list but don't reclaim mapped
1902 * 2 = 2nd pass of type 1
1903 * 3 = Reclaim mapped (normal reclaim)
1904 * 4 = 2nd pass of type 3
1906 for (pass = 0; pass < 5; pass++) {
1909 /* Force reclaiming mapped pages in the passes #3 and #4 */
1912 sc.swappiness = 100;
1915 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1916 unsigned long nr_to_scan = nr_pages - ret;
1919 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1920 if (ret >= nr_pages)
1923 reclaim_state.reclaimed_slab = 0;
1924 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1926 ret += reclaim_state.reclaimed_slab;
1927 if (ret >= nr_pages)
1930 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1931 congestion_wait(WRITE, HZ / 10);
1936 * If ret = 0, we could not shrink LRUs, but there may be something
1941 reclaim_state.reclaimed_slab = 0;
1942 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1943 ret += reclaim_state.reclaimed_slab;
1944 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1948 current->reclaim_state = NULL;
1954 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1955 not required for correctness. So if the last cpu in a node goes
1956 away, we get changed to run anywhere: as the first one comes back,
1957 restore their cpu bindings. */
1958 static int __devinit cpu_callback(struct notifier_block *nfb,
1959 unsigned long action, void *hcpu)
1963 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1964 for_each_node_state(nid, N_HIGH_MEMORY) {
1965 pg_data_t *pgdat = NODE_DATA(nid);
1966 node_to_cpumask_ptr(mask, pgdat->node_id);
1968 if (any_online_cpu(*mask) < nr_cpu_ids)
1969 /* One of our CPUs online: restore mask */
1970 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1977 * This kswapd start function will be called by init and node-hot-add.
1978 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1980 int kswapd_run(int nid)
1982 pg_data_t *pgdat = NODE_DATA(nid);
1988 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1989 if (IS_ERR(pgdat->kswapd)) {
1990 /* failure at boot is fatal */
1991 BUG_ON(system_state == SYSTEM_BOOTING);
1992 printk("Failed to start kswapd on node %d\n",nid);
1998 static int __init kswapd_init(void)
2003 for_each_node_state(nid, N_HIGH_MEMORY)
2005 hotcpu_notifier(cpu_callback, 0);
2009 module_init(kswapd_init)
2015 * If non-zero call zone_reclaim when the number of free pages falls below
2018 int zone_reclaim_mode __read_mostly;
2020 #define RECLAIM_OFF 0
2021 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2022 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2023 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2026 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2027 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2030 #define ZONE_RECLAIM_PRIORITY 4
2033 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2036 int sysctl_min_unmapped_ratio = 1;
2039 * If the number of slab pages in a zone grows beyond this percentage then
2040 * slab reclaim needs to occur.
2042 int sysctl_min_slab_ratio = 5;
2045 * Try to free up some pages from this zone through reclaim.
2047 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2049 /* Minimum pages needed in order to stay on node */
2050 const unsigned long nr_pages = 1 << order;
2051 struct task_struct *p = current;
2052 struct reclaim_state reclaim_state;
2054 unsigned long nr_reclaimed = 0;
2055 struct scan_control sc = {
2056 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2057 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2058 .swap_cluster_max = max_t(unsigned long, nr_pages,
2060 .gfp_mask = gfp_mask,
2061 .swappiness = vm_swappiness,
2062 .isolate_pages = isolate_pages_global,
2064 unsigned long slab_reclaimable;
2066 disable_swap_token();
2069 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2070 * and we also need to be able to write out pages for RECLAIM_WRITE
2073 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2074 reclaim_state.reclaimed_slab = 0;
2075 p->reclaim_state = &reclaim_state;
2077 if (zone_page_state(zone, NR_FILE_PAGES) -
2078 zone_page_state(zone, NR_FILE_MAPPED) >
2079 zone->min_unmapped_pages) {
2081 * Free memory by calling shrink zone with increasing
2082 * priorities until we have enough memory freed.
2084 priority = ZONE_RECLAIM_PRIORITY;
2086 note_zone_scanning_priority(zone, priority);
2087 nr_reclaimed += shrink_zone(priority, zone, &sc);
2089 } while (priority >= 0 && nr_reclaimed < nr_pages);
2092 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2093 if (slab_reclaimable > zone->min_slab_pages) {
2095 * shrink_slab() does not currently allow us to determine how
2096 * many pages were freed in this zone. So we take the current
2097 * number of slab pages and shake the slab until it is reduced
2098 * by the same nr_pages that we used for reclaiming unmapped
2101 * Note that shrink_slab will free memory on all zones and may
2104 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2105 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2106 slab_reclaimable - nr_pages)
2110 * Update nr_reclaimed by the number of slab pages we
2111 * reclaimed from this zone.
2113 nr_reclaimed += slab_reclaimable -
2114 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2117 p->reclaim_state = NULL;
2118 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2119 return nr_reclaimed >= nr_pages;
2122 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2128 * Zone reclaim reclaims unmapped file backed pages and
2129 * slab pages if we are over the defined limits.
2131 * A small portion of unmapped file backed pages is needed for
2132 * file I/O otherwise pages read by file I/O will be immediately
2133 * thrown out if the zone is overallocated. So we do not reclaim
2134 * if less than a specified percentage of the zone is used by
2135 * unmapped file backed pages.
2137 if (zone_page_state(zone, NR_FILE_PAGES) -
2138 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2139 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2140 <= zone->min_slab_pages)
2143 if (zone_is_all_unreclaimable(zone))
2147 * Do not scan if the allocation should not be delayed.
2149 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2153 * Only run zone reclaim on the local zone or on zones that do not
2154 * have associated processors. This will favor the local processor
2155 * over remote processors and spread off node memory allocations
2156 * as wide as possible.
2158 node_id = zone_to_nid(zone);
2159 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2162 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2164 ret = __zone_reclaim(zone, gfp_mask, order);
2165 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);