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/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
42 /* possible outcome of pageout() */
44 /* failed to write page out, page is locked */
46 /* move page to the active list, page is locked */
48 /* page has been sent to the disk successfully, page is unlocked */
50 /* page is clean and locked */
55 /* Incremented by the number of inactive pages that were scanned */
56 unsigned long nr_scanned;
58 unsigned long nr_mapped; /* From page_state */
60 /* This context's GFP mask */
65 /* Can pages be swapped as part of reclaim? */
68 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
69 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
70 * In this context, it doesn't matter that we scan the
71 * whole list at once. */
76 * The list of shrinker callbacks used by to apply pressure to
81 struct list_head list;
82 int seeks; /* seeks to recreate an obj */
83 long nr; /* objs pending delete */
86 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
88 #ifdef ARCH_HAS_PREFETCH
89 #define prefetch_prev_lru_page(_page, _base, _field) \
91 if ((_page)->lru.prev != _base) { \
94 prev = lru_to_page(&(_page->lru)); \
95 prefetch(&prev->_field); \
99 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
102 #ifdef ARCH_HAS_PREFETCHW
103 #define prefetchw_prev_lru_page(_page, _base, _field) \
105 if ((_page)->lru.prev != _base) { \
108 prev = lru_to_page(&(_page->lru)); \
109 prefetchw(&prev->_field); \
113 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
117 * From 0 .. 100. Higher means more swappy.
119 int vm_swappiness = 60;
120 static long total_memory;
122 static LIST_HEAD(shrinker_list);
123 static DECLARE_RWSEM(shrinker_rwsem);
126 * Add a shrinker callback to be called from the vm
128 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
130 struct shrinker *shrinker;
132 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
134 shrinker->shrinker = theshrinker;
135 shrinker->seeks = seeks;
137 down_write(&shrinker_rwsem);
138 list_add_tail(&shrinker->list, &shrinker_list);
139 up_write(&shrinker_rwsem);
143 EXPORT_SYMBOL(set_shrinker);
148 void remove_shrinker(struct shrinker *shrinker)
150 down_write(&shrinker_rwsem);
151 list_del(&shrinker->list);
152 up_write(&shrinker_rwsem);
155 EXPORT_SYMBOL(remove_shrinker);
157 #define SHRINK_BATCH 128
159 * Call the shrink functions to age shrinkable caches
161 * Here we assume it costs one seek to replace a lru page and that it also
162 * takes a seek to recreate a cache object. With this in mind we age equal
163 * percentages of the lru and ageable caches. This should balance the seeks
164 * generated by these structures.
166 * If the vm encounted mapped pages on the LRU it increase the pressure on
167 * slab to avoid swapping.
169 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
171 * `lru_pages' represents the number of on-LRU pages in all the zones which
172 * are eligible for the caller's allocation attempt. It is used for balancing
173 * slab reclaim versus page reclaim.
175 * Returns the number of slab objects which we shrunk.
177 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
178 unsigned long lru_pages)
180 struct shrinker *shrinker;
181 unsigned long ret = 0;
184 scanned = SWAP_CLUSTER_MAX;
186 if (!down_read_trylock(&shrinker_rwsem))
187 return 1; /* Assume we'll be able to shrink next time */
189 list_for_each_entry(shrinker, &shrinker_list, list) {
190 unsigned long long delta;
191 unsigned long total_scan;
192 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
194 delta = (4 * scanned) / shrinker->seeks;
196 do_div(delta, lru_pages + 1);
197 shrinker->nr += delta;
198 if (shrinker->nr < 0) {
199 printk(KERN_ERR "%s: nr=%ld\n",
200 __FUNCTION__, shrinker->nr);
201 shrinker->nr = max_pass;
205 * Avoid risking looping forever due to too large nr value:
206 * never try to free more than twice the estimate number of
209 if (shrinker->nr > max_pass * 2)
210 shrinker->nr = max_pass * 2;
212 total_scan = shrinker->nr;
215 while (total_scan >= SHRINK_BATCH) {
216 long this_scan = SHRINK_BATCH;
220 nr_before = (*shrinker->shrinker)(0, gfp_mask);
221 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
222 if (shrink_ret == -1)
224 if (shrink_ret < nr_before)
225 ret += nr_before - shrink_ret;
226 mod_page_state(slabs_scanned, this_scan);
227 total_scan -= this_scan;
232 shrinker->nr += total_scan;
234 up_read(&shrinker_rwsem);
238 /* Called without lock on whether page is mapped, so answer is unstable */
239 static inline int page_mapping_inuse(struct page *page)
241 struct address_space *mapping;
243 /* Page is in somebody's page tables. */
244 if (page_mapped(page))
247 /* Be more reluctant to reclaim swapcache than pagecache */
248 if (PageSwapCache(page))
251 mapping = page_mapping(page);
255 /* File is mmap'd by somebody? */
256 return mapping_mapped(mapping);
259 static inline int is_page_cache_freeable(struct page *page)
261 return page_count(page) - !!PagePrivate(page) == 2;
264 static int may_write_to_queue(struct backing_dev_info *bdi)
266 if (current->flags & PF_SWAPWRITE)
268 if (!bdi_write_congested(bdi))
270 if (bdi == current->backing_dev_info)
276 * We detected a synchronous write error writing a page out. Probably
277 * -ENOSPC. We need to propagate that into the address_space for a subsequent
278 * fsync(), msync() or close().
280 * The tricky part is that after writepage we cannot touch the mapping: nothing
281 * prevents it from being freed up. But we have a ref on the page and once
282 * that page is locked, the mapping is pinned.
284 * We're allowed to run sleeping lock_page() here because we know the caller has
287 static void handle_write_error(struct address_space *mapping,
288 struct page *page, int error)
291 if (page_mapping(page) == mapping) {
292 if (error == -ENOSPC)
293 set_bit(AS_ENOSPC, &mapping->flags);
295 set_bit(AS_EIO, &mapping->flags);
301 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
303 static pageout_t pageout(struct page *page, struct address_space *mapping)
306 * If the page is dirty, only perform writeback if that write
307 * will be non-blocking. To prevent this allocation from being
308 * stalled by pagecache activity. But note that there may be
309 * stalls if we need to run get_block(). We could test
310 * PagePrivate for that.
312 * If this process is currently in generic_file_write() against
313 * this page's queue, we can perform writeback even if that
316 * If the page is swapcache, write it back even if that would
317 * block, for some throttling. This happens by accident, because
318 * swap_backing_dev_info is bust: it doesn't reflect the
319 * congestion state of the swapdevs. Easy to fix, if needed.
320 * See swapfile.c:page_queue_congested().
322 if (!is_page_cache_freeable(page))
326 * Some data journaling orphaned pages can have
327 * page->mapping == NULL while being dirty with clean buffers.
329 if (PagePrivate(page)) {
330 if (try_to_free_buffers(page)) {
331 ClearPageDirty(page);
332 printk("%s: orphaned page\n", __FUNCTION__);
338 if (mapping->a_ops->writepage == NULL)
339 return PAGE_ACTIVATE;
340 if (!may_write_to_queue(mapping->backing_dev_info))
343 if (clear_page_dirty_for_io(page)) {
345 struct writeback_control wbc = {
346 .sync_mode = WB_SYNC_NONE,
347 .nr_to_write = SWAP_CLUSTER_MAX,
352 SetPageReclaim(page);
353 res = mapping->a_ops->writepage(page, &wbc);
355 handle_write_error(mapping, page, res);
356 if (res == AOP_WRITEPAGE_ACTIVATE) {
357 ClearPageReclaim(page);
358 return PAGE_ACTIVATE;
360 if (!PageWriteback(page)) {
361 /* synchronous write or broken a_ops? */
362 ClearPageReclaim(page);
371 static int remove_mapping(struct address_space *mapping, struct page *page)
374 return 0; /* truncate got there first */
376 write_lock_irq(&mapping->tree_lock);
379 * The non-racy check for busy page. It is critical to check
380 * PageDirty _after_ making sure that the page is freeable and
381 * not in use by anybody. (pagecache + us == 2)
383 if (unlikely(page_count(page) != 2))
386 if (unlikely(PageDirty(page)))
389 if (PageSwapCache(page)) {
390 swp_entry_t swap = { .val = page_private(page) };
391 __delete_from_swap_cache(page);
392 write_unlock_irq(&mapping->tree_lock);
394 __put_page(page); /* The pagecache ref */
398 __remove_from_page_cache(page);
399 write_unlock_irq(&mapping->tree_lock);
404 write_unlock_irq(&mapping->tree_lock);
409 * shrink_list return the number of reclaimed pages
411 static unsigned long shrink_list(struct list_head *page_list,
412 struct scan_control *sc)
414 LIST_HEAD(ret_pages);
415 struct pagevec freed_pvec;
417 unsigned long nr_reclaimed = 0;
421 pagevec_init(&freed_pvec, 1);
422 while (!list_empty(page_list)) {
423 struct address_space *mapping;
430 page = lru_to_page(page_list);
431 list_del(&page->lru);
433 if (TestSetPageLocked(page))
436 BUG_ON(PageActive(page));
440 if (!sc->may_swap && page_mapped(page))
443 /* Double the slab pressure for mapped and swapcache pages */
444 if (page_mapped(page) || PageSwapCache(page))
447 if (PageWriteback(page))
450 referenced = page_referenced(page, 1);
451 /* In active use or really unfreeable? Activate it. */
452 if (referenced && page_mapping_inuse(page))
453 goto activate_locked;
457 * Anonymous process memory has backing store?
458 * Try to allocate it some swap space here.
460 if (PageAnon(page) && !PageSwapCache(page)) {
463 if (!add_to_swap(page, GFP_ATOMIC))
464 goto activate_locked;
466 #endif /* CONFIG_SWAP */
468 mapping = page_mapping(page);
469 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
470 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
473 * The page is mapped into the page tables of one or more
474 * processes. Try to unmap it here.
476 if (page_mapped(page) && mapping) {
478 * No unmapping if we do not swap
483 switch (try_to_unmap(page, 0)) {
485 goto activate_locked;
489 ; /* try to free the page below */
493 if (PageDirty(page)) {
498 if (!sc->may_writepage)
501 /* Page is dirty, try to write it out here */
502 switch(pageout(page, mapping)) {
506 goto activate_locked;
508 if (PageWriteback(page) || PageDirty(page))
511 * A synchronous write - probably a ramdisk. Go
512 * ahead and try to reclaim the page.
514 if (TestSetPageLocked(page))
516 if (PageDirty(page) || PageWriteback(page))
518 mapping = page_mapping(page);
520 ; /* try to free the page below */
525 * If the page has buffers, try to free the buffer mappings
526 * associated with this page. If we succeed we try to free
529 * We do this even if the page is PageDirty().
530 * try_to_release_page() does not perform I/O, but it is
531 * possible for a page to have PageDirty set, but it is actually
532 * clean (all its buffers are clean). This happens if the
533 * buffers were written out directly, with submit_bh(). ext3
534 * will do this, as well as the blockdev mapping.
535 * try_to_release_page() will discover that cleanness and will
536 * drop the buffers and mark the page clean - it can be freed.
538 * Rarely, pages can have buffers and no ->mapping. These are
539 * the pages which were not successfully invalidated in
540 * truncate_complete_page(). We try to drop those buffers here
541 * and if that worked, and the page is no longer mapped into
542 * process address space (page_count == 1) it can be freed.
543 * Otherwise, leave the page on the LRU so it is swappable.
545 if (PagePrivate(page)) {
546 if (!try_to_release_page(page, sc->gfp_mask))
547 goto activate_locked;
548 if (!mapping && page_count(page) == 1)
552 if (!remove_mapping(mapping, page))
558 if (!pagevec_add(&freed_pvec, page))
559 __pagevec_release_nonlru(&freed_pvec);
568 list_add(&page->lru, &ret_pages);
569 BUG_ON(PageLRU(page));
571 list_splice(&ret_pages, page_list);
572 if (pagevec_count(&freed_pvec))
573 __pagevec_release_nonlru(&freed_pvec);
574 mod_page_state(pgactivate, pgactivate);
578 #ifdef CONFIG_MIGRATION
579 static inline void move_to_lru(struct page *page)
581 list_del(&page->lru);
582 if (PageActive(page)) {
584 * lru_cache_add_active checks that
585 * the PG_active bit is off.
587 ClearPageActive(page);
588 lru_cache_add_active(page);
596 * Add isolated pages on the list back to the LRU.
598 * returns the number of pages put back.
600 unsigned long putback_lru_pages(struct list_head *l)
604 unsigned long count = 0;
606 list_for_each_entry_safe(page, page2, l, lru) {
614 * Non migratable page
616 int fail_migrate_page(struct page *newpage, struct page *page)
620 EXPORT_SYMBOL(fail_migrate_page);
623 * swapout a single page
624 * page is locked upon entry, unlocked on exit
626 static int swap_page(struct page *page)
628 struct address_space *mapping = page_mapping(page);
630 if (page_mapped(page) && mapping)
631 if (try_to_unmap(page, 1) != SWAP_SUCCESS)
634 if (PageDirty(page)) {
635 /* Page is dirty, try to write it out here */
636 switch(pageout(page, mapping)) {
645 ; /* try to free the page below */
649 if (PagePrivate(page)) {
650 if (!try_to_release_page(page, GFP_KERNEL) ||
651 (!mapping && page_count(page) == 1))
655 if (remove_mapping(mapping, page)) {
667 EXPORT_SYMBOL(swap_page);
670 * Page migration was first developed in the context of the memory hotplug
671 * project. The main authors of the migration code are:
673 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
674 * Hirokazu Takahashi <taka@valinux.co.jp>
675 * Dave Hansen <haveblue@us.ibm.com>
676 * Christoph Lameter <clameter@sgi.com>
680 * Remove references for a page and establish the new page with the correct
681 * basic settings to be able to stop accesses to the page.
683 int migrate_page_remove_references(struct page *newpage,
684 struct page *page, int nr_refs)
686 struct address_space *mapping = page_mapping(page);
687 struct page **radix_pointer;
690 * Avoid doing any of the following work if the page count
691 * indicates that the page is in use or truncate has removed
694 if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
698 * Establish swap ptes for anonymous pages or destroy pte
701 * In order to reestablish file backed mappings the fault handlers
702 * will take the radix tree_lock which may then be used to stop
703 * processses from accessing this page until the new page is ready.
705 * A process accessing via a swap pte (an anonymous page) will take a
706 * page_lock on the old page which will block the process until the
707 * migration attempt is complete. At that time the PageSwapCache bit
708 * will be examined. If the page was migrated then the PageSwapCache
709 * bit will be clear and the operation to retrieve the page will be
710 * retried which will find the new page in the radix tree. Then a new
711 * direct mapping may be generated based on the radix tree contents.
713 * If the page was not migrated then the PageSwapCache bit
714 * is still set and the operation may continue.
716 if (try_to_unmap(page, 1) == SWAP_FAIL)
717 /* A vma has VM_LOCKED set -> Permanent failure */
721 * Give up if we were unable to remove all mappings.
723 if (page_mapcount(page))
726 write_lock_irq(&mapping->tree_lock);
728 radix_pointer = (struct page **)radix_tree_lookup_slot(
732 if (!page_mapping(page) || page_count(page) != nr_refs ||
733 *radix_pointer != page) {
734 write_unlock_irq(&mapping->tree_lock);
739 * Now we know that no one else is looking at the page.
741 * Certain minimal information about a page must be available
742 * in order for other subsystems to properly handle the page if they
743 * find it through the radix tree update before we are finished
747 newpage->index = page->index;
748 newpage->mapping = page->mapping;
749 if (PageSwapCache(page)) {
750 SetPageSwapCache(newpage);
751 set_page_private(newpage, page_private(page));
754 *radix_pointer = newpage;
756 write_unlock_irq(&mapping->tree_lock);
760 EXPORT_SYMBOL(migrate_page_remove_references);
763 * Copy the page to its new location
765 void migrate_page_copy(struct page *newpage, struct page *page)
767 copy_highpage(newpage, page);
770 SetPageError(newpage);
771 if (PageReferenced(page))
772 SetPageReferenced(newpage);
773 if (PageUptodate(page))
774 SetPageUptodate(newpage);
775 if (PageActive(page))
776 SetPageActive(newpage);
777 if (PageChecked(page))
778 SetPageChecked(newpage);
779 if (PageMappedToDisk(page))
780 SetPageMappedToDisk(newpage);
782 if (PageDirty(page)) {
783 clear_page_dirty_for_io(page);
784 set_page_dirty(newpage);
787 ClearPageSwapCache(page);
788 ClearPageActive(page);
789 ClearPagePrivate(page);
790 set_page_private(page, 0);
791 page->mapping = NULL;
794 * If any waiters have accumulated on the new page then
797 if (PageWriteback(newpage))
798 end_page_writeback(newpage);
800 EXPORT_SYMBOL(migrate_page_copy);
803 * Common logic to directly migrate a single page suitable for
804 * pages that do not use PagePrivate.
806 * Pages are locked upon entry and exit.
808 int migrate_page(struct page *newpage, struct page *page)
812 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
814 rc = migrate_page_remove_references(newpage, page, 2);
819 migrate_page_copy(newpage, page);
822 * Remove auxiliary swap entries and replace
823 * them with real ptes.
825 * Note that a real pte entry will allow processes that are not
826 * waiting on the page lock to use the new page via the page tables
827 * before the new page is unlocked.
829 remove_from_swap(newpage);
832 EXPORT_SYMBOL(migrate_page);
837 * Two lists are passed to this function. The first list
838 * contains the pages isolated from the LRU to be migrated.
839 * The second list contains new pages that the pages isolated
840 * can be moved to. If the second list is NULL then all
841 * pages are swapped out.
843 * The function returns after 10 attempts or if no pages
844 * are movable anymore because to has become empty
845 * or no retryable pages exist anymore.
847 * Return: Number of pages not migrated when "to" ran empty.
849 unsigned long migrate_pages(struct list_head *from, struct list_head *to,
850 struct list_head *moved, struct list_head *failed)
853 unsigned long nr_failed = 0;
857 int swapwrite = current->flags & PF_SWAPWRITE;
861 current->flags |= PF_SWAPWRITE;
866 list_for_each_entry_safe(page, page2, from, lru) {
867 struct page *newpage = NULL;
868 struct address_space *mapping;
873 if (page_count(page) == 1)
874 /* page was freed from under us. So we are done. */
877 if (to && list_empty(to))
881 * Skip locked pages during the first two passes to give the
882 * functions holding the lock time to release the page. Later we
883 * use lock_page() to have a higher chance of acquiring the
890 if (TestSetPageLocked(page))
894 * Only wait on writeback if we have already done a pass where
895 * we we may have triggered writeouts for lots of pages.
898 wait_on_page_writeback(page);
900 if (PageWriteback(page))
905 * Anonymous pages must have swap cache references otherwise
906 * the information contained in the page maps cannot be
909 if (PageAnon(page) && !PageSwapCache(page)) {
910 if (!add_to_swap(page, GFP_KERNEL)) {
917 rc = swap_page(page);
921 newpage = lru_to_page(to);
925 * Pages are properly locked and writeback is complete.
926 * Try to migrate the page.
928 mapping = page_mapping(page);
932 if (mapping->a_ops->migratepage) {
934 * Most pages have a mapping and most filesystems
935 * should provide a migration function. Anonymous
936 * pages are part of swap space which also has its
937 * own migration function. This is the most common
938 * path for page migration.
940 rc = mapping->a_ops->migratepage(newpage, page);
945 * Default handling if a filesystem does not provide
946 * a migration function. We can only migrate clean
947 * pages so try to write out any dirty pages first.
949 if (PageDirty(page)) {
950 switch (pageout(page, mapping)) {
956 unlock_page(newpage);
960 ; /* try to migrate the page below */
965 * Buffers are managed in a filesystem specific way.
966 * We must have no buffers or drop them.
968 if (!page_has_buffers(page) ||
969 try_to_release_page(page, GFP_KERNEL)) {
970 rc = migrate_page(newpage, page);
975 * On early passes with mapped pages simply
976 * retry. There may be a lock held for some
977 * buffers that may go away. Later
982 * Persistently unable to drop buffers..... As a
983 * measure of last resort we fall back to
986 unlock_page(newpage);
988 rc = swap_page(page);
993 unlock_page(newpage);
1002 /* Permanent failure */
1003 list_move(&page->lru, failed);
1007 /* Successful migration. Return page to LRU */
1008 move_to_lru(newpage);
1010 list_move(&page->lru, moved);
1013 if (retry && pass++ < 10)
1017 current->flags &= ~PF_SWAPWRITE;
1019 return nr_failed + retry;
1023 * Isolate one page from the LRU lists and put it on the
1024 * indicated list with elevated refcount.
1027 * 0 = page not on LRU list
1028 * 1 = page removed from LRU list and added to the specified list.
1030 int isolate_lru_page(struct page *page)
1034 if (PageLRU(page)) {
1035 struct zone *zone = page_zone(page);
1036 spin_lock_irq(&zone->lru_lock);
1037 if (PageLRU(page)) {
1041 if (PageActive(page))
1042 del_page_from_active_list(zone, page);
1044 del_page_from_inactive_list(zone, page);
1046 spin_unlock_irq(&zone->lru_lock);
1054 * zone->lru_lock is heavily contended. Some of the functions that
1055 * shrink the lists perform better by taking out a batch of pages
1056 * and working on them outside the LRU lock.
1058 * For pagecache intensive workloads, this function is the hottest
1059 * spot in the kernel (apart from copy_*_user functions).
1061 * Appropriate locks must be held before calling this function.
1063 * @nr_to_scan: The number of pages to look through on the list.
1064 * @src: The LRU list to pull pages off.
1065 * @dst: The temp list to put pages on to.
1066 * @scanned: The number of pages that were scanned.
1068 * returns how many pages were moved onto *@dst.
1070 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1071 struct list_head *src, struct list_head *dst,
1072 unsigned long *scanned)
1074 unsigned long nr_taken = 0;
1076 unsigned long scan = 0;
1078 while (scan++ < nr_to_scan && !list_empty(src)) {
1079 struct list_head *target;
1080 page = lru_to_page(src);
1081 prefetchw_prev_lru_page(page, src, flags);
1083 BUG_ON(!PageLRU(page));
1085 list_del(&page->lru);
1087 if (likely(get_page_unless_zero(page))) {
1089 * Be careful not to clear PageLRU until after we're
1090 * sure the page is not being freed elsewhere -- the
1091 * page release code relies on it.
1096 } /* else it is being freed elsewhere */
1098 list_add(&page->lru, target);
1106 * shrink_cache() return the number of reclaimed pages
1108 static unsigned long shrink_cache(unsigned long max_scan, struct zone *zone,
1109 struct scan_control *sc)
1111 LIST_HEAD(page_list);
1112 struct pagevec pvec;
1113 unsigned long nr_scanned = 0;
1114 unsigned long nr_reclaimed = 0;
1116 pagevec_init(&pvec, 1);
1119 spin_lock_irq(&zone->lru_lock);
1122 unsigned long nr_taken;
1123 unsigned long nr_scan;
1124 unsigned long nr_freed;
1126 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
1127 &zone->inactive_list,
1128 &page_list, &nr_scan);
1129 zone->nr_inactive -= nr_taken;
1130 zone->pages_scanned += nr_scan;
1131 spin_unlock_irq(&zone->lru_lock);
1136 nr_scanned += nr_scan;
1137 nr_freed = shrink_list(&page_list, sc);
1138 nr_reclaimed += nr_freed;
1139 local_irq_disable();
1140 if (current_is_kswapd()) {
1141 __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
1142 __mod_page_state(kswapd_steal, nr_freed);
1144 __mod_page_state_zone(zone, pgscan_direct, nr_scan);
1145 __mod_page_state_zone(zone, pgsteal, nr_freed);
1147 spin_lock(&zone->lru_lock);
1149 * Put back any unfreeable pages.
1151 while (!list_empty(&page_list)) {
1152 page = lru_to_page(&page_list);
1153 BUG_ON(PageLRU(page));
1155 list_del(&page->lru);
1156 if (PageActive(page))
1157 add_page_to_active_list(zone, page);
1159 add_page_to_inactive_list(zone, page);
1160 if (!pagevec_add(&pvec, page)) {
1161 spin_unlock_irq(&zone->lru_lock);
1162 __pagevec_release(&pvec);
1163 spin_lock_irq(&zone->lru_lock);
1166 } while (nr_scanned < max_scan);
1167 spin_unlock_irq(&zone->lru_lock);
1169 pagevec_release(&pvec);
1170 return nr_reclaimed;
1174 * This moves pages from the active list to the inactive list.
1176 * We move them the other way if the page is referenced by one or more
1177 * processes, from rmap.
1179 * If the pages are mostly unmapped, the processing is fast and it is
1180 * appropriate to hold zone->lru_lock across the whole operation. But if
1181 * the pages are mapped, the processing is slow (page_referenced()) so we
1182 * should drop zone->lru_lock around each page. It's impossible to balance
1183 * this, so instead we remove the pages from the LRU while processing them.
1184 * It is safe to rely on PG_active against the non-LRU pages in here because
1185 * nobody will play with that bit on a non-LRU page.
1187 * The downside is that we have to touch page->_count against each page.
1188 * But we had to alter page->flags anyway.
1191 refill_inactive_zone(unsigned long nr_pages, struct zone *zone,
1192 struct scan_control *sc)
1194 unsigned long pgmoved;
1195 int pgdeactivate = 0;
1196 unsigned long pgscanned;
1197 LIST_HEAD(l_hold); /* The pages which were snipped off */
1198 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
1199 LIST_HEAD(l_active); /* Pages to go onto the active_list */
1201 struct pagevec pvec;
1202 int reclaim_mapped = 0;
1204 if (unlikely(sc->may_swap)) {
1210 * `distress' is a measure of how much trouble we're having
1211 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1213 distress = 100 >> zone->prev_priority;
1216 * The point of this algorithm is to decide when to start
1217 * reclaiming mapped memory instead of just pagecache. Work out
1221 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
1224 * Now decide how much we really want to unmap some pages. The
1225 * mapped ratio is downgraded - just because there's a lot of
1226 * mapped memory doesn't necessarily mean that page reclaim
1229 * The distress ratio is important - we don't want to start
1232 * A 100% value of vm_swappiness overrides this algorithm
1235 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
1238 * Now use this metric to decide whether to start moving mapped
1239 * memory onto the inactive list.
1241 if (swap_tendency >= 100)
1246 spin_lock_irq(&zone->lru_lock);
1247 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1248 &l_hold, &pgscanned);
1249 zone->pages_scanned += pgscanned;
1250 zone->nr_active -= pgmoved;
1251 spin_unlock_irq(&zone->lru_lock);
1253 while (!list_empty(&l_hold)) {
1255 page = lru_to_page(&l_hold);
1256 list_del(&page->lru);
1257 if (page_mapped(page)) {
1258 if (!reclaim_mapped ||
1259 (total_swap_pages == 0 && PageAnon(page)) ||
1260 page_referenced(page, 0)) {
1261 list_add(&page->lru, &l_active);
1265 list_add(&page->lru, &l_inactive);
1268 pagevec_init(&pvec, 1);
1270 spin_lock_irq(&zone->lru_lock);
1271 while (!list_empty(&l_inactive)) {
1272 page = lru_to_page(&l_inactive);
1273 prefetchw_prev_lru_page(page, &l_inactive, flags);
1274 BUG_ON(PageLRU(page));
1276 BUG_ON(!PageActive(page));
1277 ClearPageActive(page);
1279 list_move(&page->lru, &zone->inactive_list);
1281 if (!pagevec_add(&pvec, page)) {
1282 zone->nr_inactive += pgmoved;
1283 spin_unlock_irq(&zone->lru_lock);
1284 pgdeactivate += pgmoved;
1286 if (buffer_heads_over_limit)
1287 pagevec_strip(&pvec);
1288 __pagevec_release(&pvec);
1289 spin_lock_irq(&zone->lru_lock);
1292 zone->nr_inactive += pgmoved;
1293 pgdeactivate += pgmoved;
1294 if (buffer_heads_over_limit) {
1295 spin_unlock_irq(&zone->lru_lock);
1296 pagevec_strip(&pvec);
1297 spin_lock_irq(&zone->lru_lock);
1301 while (!list_empty(&l_active)) {
1302 page = lru_to_page(&l_active);
1303 prefetchw_prev_lru_page(page, &l_active, flags);
1304 BUG_ON(PageLRU(page));
1306 BUG_ON(!PageActive(page));
1307 list_move(&page->lru, &zone->active_list);
1309 if (!pagevec_add(&pvec, page)) {
1310 zone->nr_active += pgmoved;
1312 spin_unlock_irq(&zone->lru_lock);
1313 __pagevec_release(&pvec);
1314 spin_lock_irq(&zone->lru_lock);
1317 zone->nr_active += pgmoved;
1318 spin_unlock(&zone->lru_lock);
1320 __mod_page_state_zone(zone, pgrefill, pgscanned);
1321 __mod_page_state(pgdeactivate, pgdeactivate);
1324 pagevec_release(&pvec);
1328 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1330 static unsigned long shrink_zone(int priority, struct zone *zone,
1331 struct scan_control *sc)
1333 unsigned long nr_active;
1334 unsigned long nr_inactive;
1335 unsigned long nr_to_scan;
1336 unsigned long nr_reclaimed = 0;
1338 atomic_inc(&zone->reclaim_in_progress);
1341 * Add one to `nr_to_scan' just to make sure that the kernel will
1342 * slowly sift through the active list.
1344 zone->nr_scan_active += (zone->nr_active >> priority) + 1;
1345 nr_active = zone->nr_scan_active;
1346 if (nr_active >= sc->swap_cluster_max)
1347 zone->nr_scan_active = 0;
1351 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
1352 nr_inactive = zone->nr_scan_inactive;
1353 if (nr_inactive >= sc->swap_cluster_max)
1354 zone->nr_scan_inactive = 0;
1358 while (nr_active || nr_inactive) {
1360 nr_to_scan = min(nr_active,
1361 (unsigned long)sc->swap_cluster_max);
1362 nr_active -= nr_to_scan;
1363 refill_inactive_zone(nr_to_scan, zone, sc);
1367 nr_to_scan = min(nr_inactive,
1368 (unsigned long)sc->swap_cluster_max);
1369 nr_inactive -= nr_to_scan;
1370 nr_reclaimed += shrink_cache(nr_to_scan, zone, sc);
1374 throttle_vm_writeout();
1376 atomic_dec(&zone->reclaim_in_progress);
1377 return nr_reclaimed;
1381 * This is the direct reclaim path, for page-allocating processes. We only
1382 * try to reclaim pages from zones which will satisfy the caller's allocation
1385 * We reclaim from a zone even if that zone is over pages_high. Because:
1386 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1388 * b) The zones may be over pages_high but they must go *over* pages_high to
1389 * satisfy the `incremental min' zone defense algorithm.
1391 * Returns the number of reclaimed pages.
1393 * If a zone is deemed to be full of pinned pages then just give it a light
1394 * scan then give up on it.
1396 static unsigned long shrink_caches(int priority, struct zone **zones,
1397 struct scan_control *sc)
1399 unsigned long nr_reclaimed = 0;
1402 for (i = 0; zones[i] != NULL; i++) {
1403 struct zone *zone = zones[i];
1405 if (!populated_zone(zone))
1408 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1411 zone->temp_priority = priority;
1412 if (zone->prev_priority > priority)
1413 zone->prev_priority = priority;
1415 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1416 continue; /* Let kswapd poll it */
1418 nr_reclaimed += shrink_zone(priority, zone, sc);
1420 return nr_reclaimed;
1424 * This is the main entry point to direct page reclaim.
1426 * If a full scan of the inactive list fails to free enough memory then we
1427 * are "out of memory" and something needs to be killed.
1429 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1430 * high - the zone may be full of dirty or under-writeback pages, which this
1431 * caller can't do much about. We kick pdflush and take explicit naps in the
1432 * hope that some of these pages can be written. But if the allocating task
1433 * holds filesystem locks which prevent writeout this might not work, and the
1434 * allocation attempt will fail.
1436 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1440 unsigned long total_scanned = 0;
1441 unsigned long nr_reclaimed = 0;
1442 struct reclaim_state *reclaim_state = current->reclaim_state;
1443 unsigned long lru_pages = 0;
1445 struct scan_control sc = {
1446 .gfp_mask = gfp_mask,
1447 .may_writepage = !laptop_mode,
1448 .swap_cluster_max = SWAP_CLUSTER_MAX,
1452 inc_page_state(allocstall);
1454 for (i = 0; zones[i] != NULL; i++) {
1455 struct zone *zone = zones[i];
1457 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1460 zone->temp_priority = DEF_PRIORITY;
1461 lru_pages += zone->nr_active + zone->nr_inactive;
1464 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1465 sc.nr_mapped = read_page_state(nr_mapped);
1468 disable_swap_token();
1469 nr_reclaimed += shrink_caches(priority, zones, &sc);
1470 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1471 if (reclaim_state) {
1472 nr_reclaimed += reclaim_state->reclaimed_slab;
1473 reclaim_state->reclaimed_slab = 0;
1475 total_scanned += sc.nr_scanned;
1476 if (nr_reclaimed >= sc.swap_cluster_max) {
1482 * Try to write back as many pages as we just scanned. This
1483 * tends to cause slow streaming writers to write data to the
1484 * disk smoothly, at the dirtying rate, which is nice. But
1485 * that's undesirable in laptop mode, where we *want* lumpy
1486 * writeout. So in laptop mode, write out the whole world.
1488 if (total_scanned > sc.swap_cluster_max +
1489 sc.swap_cluster_max / 2) {
1490 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1491 sc.may_writepage = 1;
1494 /* Take a nap, wait for some writeback to complete */
1495 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1496 blk_congestion_wait(WRITE, HZ/10);
1499 for (i = 0; zones[i] != 0; i++) {
1500 struct zone *zone = zones[i];
1502 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1505 zone->prev_priority = zone->temp_priority;
1511 * For kswapd, balance_pgdat() will work across all this node's zones until
1512 * they are all at pages_high.
1514 * If `nr_pages' is non-zero then it is the number of pages which are to be
1515 * reclaimed, regardless of the zone occupancies. This is a software suspend
1518 * Returns the number of pages which were actually freed.
1520 * There is special handling here for zones which are full of pinned pages.
1521 * This can happen if the pages are all mlocked, or if they are all used by
1522 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1523 * What we do is to detect the case where all pages in the zone have been
1524 * scanned twice and there has been zero successful reclaim. Mark the zone as
1525 * dead and from now on, only perform a short scan. Basically we're polling
1526 * the zone for when the problem goes away.
1528 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1529 * zones which have free_pages > pages_high, but once a zone is found to have
1530 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1531 * of the number of free pages in the lower zones. This interoperates with
1532 * the page allocator fallback scheme to ensure that aging of pages is balanced
1535 static unsigned long balance_pgdat(pg_data_t *pgdat, unsigned long nr_pages,
1538 unsigned long to_free = nr_pages;
1542 unsigned long total_scanned;
1543 unsigned long nr_reclaimed;
1544 struct reclaim_state *reclaim_state = current->reclaim_state;
1545 struct scan_control sc = {
1546 .gfp_mask = GFP_KERNEL,
1548 .swap_cluster_max = nr_pages ? nr_pages : SWAP_CLUSTER_MAX,
1554 sc.may_writepage = !laptop_mode,
1555 sc.nr_mapped = read_page_state(nr_mapped);
1557 inc_page_state(pageoutrun);
1559 for (i = 0; i < pgdat->nr_zones; i++) {
1560 struct zone *zone = pgdat->node_zones + i;
1562 zone->temp_priority = DEF_PRIORITY;
1565 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1566 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1567 unsigned long lru_pages = 0;
1569 /* The swap token gets in the way of swapout... */
1571 disable_swap_token();
1575 if (nr_pages == 0) {
1577 * Scan in the highmem->dma direction for the highest
1578 * zone which needs scanning
1580 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1581 struct zone *zone = pgdat->node_zones + i;
1583 if (!populated_zone(zone))
1586 if (zone->all_unreclaimable &&
1587 priority != DEF_PRIORITY)
1590 if (!zone_watermark_ok(zone, order,
1591 zone->pages_high, 0, 0)) {
1598 end_zone = pgdat->nr_zones - 1;
1601 for (i = 0; i <= end_zone; i++) {
1602 struct zone *zone = pgdat->node_zones + i;
1604 lru_pages += zone->nr_active + zone->nr_inactive;
1608 * Now scan the zone in the dma->highmem direction, stopping
1609 * at the last zone which needs scanning.
1611 * We do this because the page allocator works in the opposite
1612 * direction. This prevents the page allocator from allocating
1613 * pages behind kswapd's direction of progress, which would
1614 * cause too much scanning of the lower zones.
1616 for (i = 0; i <= end_zone; i++) {
1617 struct zone *zone = pgdat->node_zones + i;
1620 if (!populated_zone(zone))
1623 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1626 if (nr_pages == 0) { /* Not software suspend */
1627 if (!zone_watermark_ok(zone, order,
1628 zone->pages_high, end_zone, 0))
1631 zone->temp_priority = priority;
1632 if (zone->prev_priority > priority)
1633 zone->prev_priority = priority;
1635 nr_reclaimed += shrink_zone(priority, zone, &sc);
1636 reclaim_state->reclaimed_slab = 0;
1637 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1639 nr_reclaimed += reclaim_state->reclaimed_slab;
1640 total_scanned += sc.nr_scanned;
1641 if (zone->all_unreclaimable)
1643 if (nr_slab == 0 && zone->pages_scanned >=
1644 (zone->nr_active + zone->nr_inactive) * 4)
1645 zone->all_unreclaimable = 1;
1647 * If we've done a decent amount of scanning and
1648 * the reclaim ratio is low, start doing writepage
1649 * even in laptop mode
1651 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1652 total_scanned > nr_reclaimed + nr_reclaimed / 2)
1653 sc.may_writepage = 1;
1655 if (nr_pages && to_free > nr_reclaimed)
1656 continue; /* swsusp: need to do more work */
1658 break; /* kswapd: all done */
1660 * OK, kswapd is getting into trouble. Take a nap, then take
1661 * another pass across the zones.
1663 if (total_scanned && priority < DEF_PRIORITY - 2)
1664 blk_congestion_wait(WRITE, HZ/10);
1667 * We do this so kswapd doesn't build up large priorities for
1668 * example when it is freeing in parallel with allocators. It
1669 * matches the direct reclaim path behaviour in terms of impact
1670 * on zone->*_priority.
1672 if ((nr_reclaimed >= SWAP_CLUSTER_MAX) && !nr_pages)
1676 for (i = 0; i < pgdat->nr_zones; i++) {
1677 struct zone *zone = pgdat->node_zones + i;
1679 zone->prev_priority = zone->temp_priority;
1681 if (!all_zones_ok) {
1686 return nr_reclaimed;
1690 * The background pageout daemon, started as a kernel thread
1691 * from the init process.
1693 * This basically trickles out pages so that we have _some_
1694 * free memory available even if there is no other activity
1695 * that frees anything up. This is needed for things like routing
1696 * etc, where we otherwise might have all activity going on in
1697 * asynchronous contexts that cannot page things out.
1699 * If there are applications that are active memory-allocators
1700 * (most normal use), this basically shouldn't matter.
1702 static int kswapd(void *p)
1704 unsigned long order;
1705 pg_data_t *pgdat = (pg_data_t*)p;
1706 struct task_struct *tsk = current;
1708 struct reclaim_state reclaim_state = {
1709 .reclaimed_slab = 0,
1713 daemonize("kswapd%d", pgdat->node_id);
1714 cpumask = node_to_cpumask(pgdat->node_id);
1715 if (!cpus_empty(cpumask))
1716 set_cpus_allowed(tsk, cpumask);
1717 current->reclaim_state = &reclaim_state;
1720 * Tell the memory management that we're a "memory allocator",
1721 * and that if we need more memory we should get access to it
1722 * regardless (see "__alloc_pages()"). "kswapd" should
1723 * never get caught in the normal page freeing logic.
1725 * (Kswapd normally doesn't need memory anyway, but sometimes
1726 * you need a small amount of memory in order to be able to
1727 * page out something else, and this flag essentially protects
1728 * us from recursively trying to free more memory as we're
1729 * trying to free the first piece of memory in the first place).
1731 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1735 unsigned long new_order;
1739 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1740 new_order = pgdat->kswapd_max_order;
1741 pgdat->kswapd_max_order = 0;
1742 if (order < new_order) {
1744 * Don't sleep if someone wants a larger 'order'
1750 order = pgdat->kswapd_max_order;
1752 finish_wait(&pgdat->kswapd_wait, &wait);
1754 balance_pgdat(pgdat, 0, order);
1760 * A zone is low on free memory, so wake its kswapd task to service it.
1762 void wakeup_kswapd(struct zone *zone, int order)
1766 if (!populated_zone(zone))
1769 pgdat = zone->zone_pgdat;
1770 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1772 if (pgdat->kswapd_max_order < order)
1773 pgdat->kswapd_max_order = order;
1774 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1776 if (!waitqueue_active(&pgdat->kswapd_wait))
1778 wake_up_interruptible(&pgdat->kswapd_wait);
1783 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1786 unsigned long shrink_all_memory(unsigned long nr_pages)
1789 unsigned long nr_to_free = nr_pages;
1790 unsigned long ret = 0;
1791 struct reclaim_state reclaim_state = {
1792 .reclaimed_slab = 0,
1795 current->reclaim_state = &reclaim_state;
1796 for_each_pgdat(pgdat) {
1797 unsigned long freed;
1799 freed = balance_pgdat(pgdat, nr_to_free, 0);
1801 nr_to_free -= freed;
1802 if ((long)nr_to_free <= 0)
1805 current->reclaim_state = NULL;
1810 #ifdef CONFIG_HOTPLUG_CPU
1811 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1812 not required for correctness. So if the last cpu in a node goes
1813 away, we get changed to run anywhere: as the first one comes back,
1814 restore their cpu bindings. */
1815 static int __devinit cpu_callback(struct notifier_block *nfb,
1816 unsigned long action, void *hcpu)
1821 if (action == CPU_ONLINE) {
1822 for_each_pgdat(pgdat) {
1823 mask = node_to_cpumask(pgdat->node_id);
1824 if (any_online_cpu(mask) != NR_CPUS)
1825 /* One of our CPUs online: restore mask */
1826 set_cpus_allowed(pgdat->kswapd, mask);
1831 #endif /* CONFIG_HOTPLUG_CPU */
1833 static int __init kswapd_init(void)
1838 for_each_pgdat(pgdat) {
1841 pid = kernel_thread(kswapd, pgdat, CLONE_KERNEL);
1843 pgdat->kswapd = find_task_by_pid(pid);
1845 total_memory = nr_free_pagecache_pages();
1846 hotcpu_notifier(cpu_callback, 0);
1850 module_init(kswapd_init)
1856 * If non-zero call zone_reclaim when the number of free pages falls below
1859 * In the future we may add flags to the mode. However, the page allocator
1860 * should only have to check that zone_reclaim_mode != 0 before calling
1863 int zone_reclaim_mode __read_mostly;
1865 #define RECLAIM_OFF 0
1866 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1867 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1868 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1869 #define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
1872 * Mininum time between zone reclaim scans
1874 int zone_reclaim_interval __read_mostly = 30*HZ;
1877 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1878 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1881 #define ZONE_RECLAIM_PRIORITY 4
1884 * Try to free up some pages from this zone through reclaim.
1886 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1888 const unsigned long nr_pages = 1 << order;
1889 struct task_struct *p = current;
1890 struct reclaim_state reclaim_state;
1892 unsigned long nr_reclaimed = 0;
1893 struct scan_control sc = {
1894 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1895 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1896 .nr_mapped = read_page_state(nr_mapped),
1897 .swap_cluster_max = max_t(unsigned long, nr_pages,
1899 .gfp_mask = gfp_mask,
1902 disable_swap_token();
1905 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1906 * and we also need to be able to write out pages for RECLAIM_WRITE
1909 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1910 reclaim_state.reclaimed_slab = 0;
1911 p->reclaim_state = &reclaim_state;
1914 * Free memory by calling shrink zone with increasing priorities
1915 * until we have enough memory freed.
1917 priority = ZONE_RECLAIM_PRIORITY;
1919 nr_reclaimed += shrink_zone(priority, zone, &sc);
1921 } while (priority >= 0 && nr_reclaimed < nr_pages);
1923 if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1925 * shrink_slab does not currently allow us to determine
1926 * how many pages were freed in the zone. So we just
1927 * shake the slab and then go offnode for a single allocation.
1929 * shrink_slab will free memory on all zones and may take
1932 shrink_slab(sc.nr_scanned, gfp_mask, order);
1935 p->reclaim_state = NULL;
1936 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1938 if (nr_reclaimed == 0)
1939 zone->last_unsuccessful_zone_reclaim = jiffies;
1941 return nr_reclaimed >= nr_pages;
1944 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1950 * Do not reclaim if there was a recent unsuccessful attempt at zone
1951 * reclaim. In that case we let allocations go off node for the
1952 * zone_reclaim_interval. Otherwise we would scan for each off-node
1955 if (time_before(jiffies,
1956 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1960 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1961 * not have reclaimable pages and if we should not delay the allocation
1964 if (!(gfp_mask & __GFP_WAIT) ||
1965 zone->all_unreclaimable ||
1966 atomic_read(&zone->reclaim_in_progress) > 0 ||
1967 (current->flags & PF_MEMALLOC))
1971 * Only run zone reclaim on the local zone or on zones that do not
1972 * have associated processors. This will favor the local processor
1973 * over remote processors and spread off node memory allocations
1974 * as wide as possible.
1976 node_id = zone->zone_pgdat->node_id;
1977 mask = node_to_cpumask(node_id);
1978 if (!cpus_empty(mask) && node_id != numa_node_id())
1980 return __zone_reclaim(zone, gfp_mask, order);