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/suspend.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/notifier.h>
36 #include <asm/pgalloc.h>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
43 * From 0 .. 100. Higher means more swappy.
45 int vm_swappiness = 60;
46 static long total_memory;
48 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
50 #ifdef ARCH_HAS_PREFETCH
51 #define prefetch_prev_lru_page(_page, _base, _field) \
53 if ((_page)->lru.prev != _base) { \
56 prev = lru_to_page(&(_page->lru)); \
57 prefetch(&prev->_field); \
61 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
64 #ifdef ARCH_HAS_PREFETCHW
65 #define prefetchw_prev_lru_page(_page, _base, _field) \
67 if ((_page)->lru.prev != _base) { \
70 prev = lru_to_page(&(_page->lru)); \
71 prefetchw(&prev->_field); \
75 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
79 * The list of shrinker callbacks used by to apply pressure to
84 struct list_head list;
85 int seeks; /* seeks to recreate an obj */
86 long nr; /* objs pending delete */
89 static LIST_HEAD(shrinker_list);
90 static DECLARE_MUTEX(shrinker_sem);
93 * Add a shrinker callback to be called from the vm
95 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
97 struct shrinker *shrinker;
99 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
101 shrinker->shrinker = theshrinker;
102 shrinker->seeks = seeks;
105 list_add(&shrinker->list, &shrinker_list);
111 EXPORT_SYMBOL(set_shrinker);
116 void remove_shrinker(struct shrinker *shrinker)
119 list_del(&shrinker->list);
124 EXPORT_SYMBOL(remove_shrinker);
126 #define SHRINK_BATCH 128
128 * Call the shrink functions to age shrinkable caches
130 * Here we assume it costs one seek to replace a lru page and that it also
131 * takes a seek to recreate a cache object. With this in mind we age equal
132 * percentages of the lru and ageable caches. This should balance the seeks
133 * generated by these structures.
135 * If the vm encounted mapped pages on the LRU it increase the pressure on
136 * slab to avoid swapping.
138 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
140 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask)
142 struct shrinker *shrinker;
145 if (down_trylock(&shrinker_sem))
148 pages = nr_used_zone_pages();
149 list_for_each_entry(shrinker, &shrinker_list, list) {
150 unsigned long long delta;
152 delta = (4 * scanned) / shrinker->seeks;
153 delta *= (*shrinker->shrinker)(0, gfp_mask);
154 do_div(delta, pages + 1);
155 shrinker->nr += delta;
156 if (shrinker->nr < 0)
157 shrinker->nr = LONG_MAX; /* It wrapped! */
159 if (shrinker->nr <= SHRINK_BATCH)
161 while (shrinker->nr) {
162 long this_scan = shrinker->nr;
167 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
168 mod_page_state(slabs_scanned, this_scan);
169 shrinker->nr -= this_scan;
170 if (shrink_ret == -1)
179 /* Must be called with page's rmap lock held. */
180 static inline int page_mapping_inuse(struct page *page)
182 struct address_space *mapping;
184 /* Page is in somebody's page tables. */
185 if (page_mapped(page))
188 /* Be more reluctant to reclaim swapcache than pagecache */
189 if (PageSwapCache(page))
192 mapping = page_mapping(page);
196 /* File is mmap'd by somebody? */
197 return mapping_mapped(mapping);
200 static inline int is_page_cache_freeable(struct page *page)
202 return page_count(page) - !!PagePrivate(page) == 2;
205 static int may_write_to_queue(struct backing_dev_info *bdi)
207 if (current_is_kswapd())
209 if (current_is_pdflush()) /* This is unlikely, but why not... */
211 if (!bdi_write_congested(bdi))
213 if (bdi == current->backing_dev_info)
219 * We detected a synchronous write error writing a page out. Probably
220 * -ENOSPC. We need to propagate that into the address_space for a subsequent
221 * fsync(), msync() or close().
223 * The tricky part is that after writepage we cannot touch the mapping: nothing
224 * prevents it from being freed up. But we have a ref on the page and once
225 * that page is locked, the mapping is pinned.
227 * We're allowed to run sleeping lock_page() here because we know the caller has
230 static void handle_write_error(struct address_space *mapping,
231 struct page *page, int error)
234 if (page_mapping(page) == mapping) {
235 if (error == -ENOSPC)
236 set_bit(AS_ENOSPC, &mapping->flags);
238 set_bit(AS_EIO, &mapping->flags);
244 * shrink_list returns the number of reclaimed pages
247 shrink_list(struct list_head *page_list, unsigned int gfp_mask,
248 int *nr_scanned, int do_writepage)
250 LIST_HEAD(ret_pages);
251 struct pagevec freed_pvec;
257 pagevec_init(&freed_pvec, 1);
258 while (!list_empty(page_list)) {
259 struct address_space *mapping;
264 page = lru_to_page(page_list);
265 list_del(&page->lru);
267 if (TestSetPageLocked(page))
270 /* Double the slab pressure for mapped and swapcache pages */
271 if (page_mapped(page) || PageSwapCache(page))
274 BUG_ON(PageActive(page));
276 if (PageWriteback(page))
280 referenced = page_referenced(page);
281 if (referenced && page_mapping_inuse(page)) {
282 /* In active use or really unfreeable. Activate it. */
283 page_map_unlock(page);
284 goto activate_locked;
289 * Anonymous process memory has backing store?
290 * Try to allocate it some swap space here.
292 * XXX: implement swap clustering ?
294 if (PageAnon(page) && !PageSwapCache(page)) {
295 page_map_unlock(page);
296 if (!add_to_swap(page))
297 goto activate_locked;
300 #endif /* CONFIG_SWAP */
302 mapping = page_mapping(page);
303 may_enter_fs = (gfp_mask & __GFP_FS) ||
304 (PageSwapCache(page) && (gfp_mask & __GFP_IO));
307 * The page is mapped into the page tables of one or more
308 * processes. Try to unmap it here.
310 if (page_mapped(page) && mapping) {
311 switch (try_to_unmap(page)) {
313 page_map_unlock(page);
314 goto activate_locked;
316 page_map_unlock(page);
319 ; /* try to free the page below */
322 page_map_unlock(page);
325 * If the page is dirty, only perform writeback if that write
326 * will be non-blocking. To prevent this allocation from being
327 * stalled by pagecache activity. But note that there may be
328 * stalls if we need to run get_block(). We could test
329 * PagePrivate for that.
331 * If this process is currently in generic_file_write() against
332 * this page's queue, we can perform writeback even if that
335 * If the page is swapcache, write it back even if that would
336 * block, for some throttling. This happens by accident, because
337 * swap_backing_dev_info is bust: it doesn't reflect the
338 * congestion state of the swapdevs. Easy to fix, if needed.
339 * See swapfile.c:page_queue_congested().
341 if (PageDirty(page)) {
344 if (!is_page_cache_freeable(page))
348 if (mapping->a_ops->writepage == NULL)
349 goto activate_locked;
352 if (!may_write_to_queue(mapping->backing_dev_info))
354 if (laptop_mode && !do_writepage)
356 if (clear_page_dirty_for_io(page)) {
358 struct writeback_control wbc = {
359 .sync_mode = WB_SYNC_NONE,
360 .nr_to_write = SWAP_CLUSTER_MAX,
365 SetPageReclaim(page);
366 res = mapping->a_ops->writepage(page, &wbc);
368 handle_write_error(mapping, page, res);
369 if (res == WRITEPAGE_ACTIVATE) {
370 ClearPageReclaim(page);
371 goto activate_locked;
373 if (!PageWriteback(page)) {
374 /* synchronous write or broken a_ops? */
375 ClearPageReclaim(page);
382 * If the page has buffers, try to free the buffer mappings
383 * associated with this page. If we succeed we try to free
386 * We do this even if the page is PageDirty().
387 * try_to_release_page() does not perform I/O, but it is
388 * possible for a page to have PageDirty set, but it is actually
389 * clean (all its buffers are clean). This happens if the
390 * buffers were written out directly, with submit_bh(). ext3
391 * will do this, as well as the blockdev mapping.
392 * try_to_release_page() will discover that cleanness and will
393 * drop the buffers and mark the page clean - it can be freed.
395 * Rarely, pages can have buffers and no ->mapping. These are
396 * the pages which were not successfully invalidated in
397 * truncate_complete_page(). We try to drop those buffers here
398 * and if that worked, and the page is no longer mapped into
399 * process address space (page_count == 0) it can be freed.
400 * Otherwise, leave the page on the LRU so it is swappable.
402 if (PagePrivate(page)) {
403 if (!try_to_release_page(page, gfp_mask))
404 goto activate_locked;
405 if (!mapping && page_count(page) == 1)
410 goto keep_locked; /* truncate got there first */
412 spin_lock_irq(&mapping->tree_lock);
415 * The non-racy check for busy page. It is critical to check
416 * PageDirty _after_ making sure that the page is freeable and
417 * not in use by anybody. (pagecache + us == 2)
419 if (page_count(page) != 2 || PageDirty(page)) {
420 spin_unlock_irq(&mapping->tree_lock);
425 if (PageSwapCache(page)) {
426 swp_entry_t swap = { .val = page->private };
427 __delete_from_swap_cache(page);
428 spin_unlock_irq(&mapping->tree_lock);
430 __put_page(page); /* The pagecache ref */
433 #endif /* CONFIG_SWAP */
435 __remove_from_page_cache(page);
436 spin_unlock_irq(&mapping->tree_lock);
442 if (!pagevec_add(&freed_pvec, page))
443 __pagevec_release_nonlru(&freed_pvec);
452 list_add(&page->lru, &ret_pages);
453 BUG_ON(PageLRU(page));
455 list_splice(&ret_pages, page_list);
456 if (pagevec_count(&freed_pvec))
457 __pagevec_release_nonlru(&freed_pvec);
458 mod_page_state(pgactivate, pgactivate);
463 * zone->lru_lock is heavily contented. We relieve it by quickly privatising
464 * a batch of pages and working on them outside the lock. Any pages which were
465 * not freed will be added back to the LRU.
467 * shrink_cache() is passed the number of pages to scan and returns the number
468 * of pages which were reclaimed.
470 * For pagecache intensive workloads, the first loop here is the hottest spot
471 * in the kernel (apart from the copy_*_user functions).
474 shrink_cache(struct zone *zone, unsigned int gfp_mask,
475 int max_scan, int *total_scanned, int do_writepage)
477 LIST_HEAD(page_list);
481 pagevec_init(&pvec, 1);
484 spin_lock_irq(&zone->lru_lock);
485 while (max_scan > 0) {
491 while (nr_scan++ < SWAP_CLUSTER_MAX &&
492 !list_empty(&zone->inactive_list)) {
493 page = lru_to_page(&zone->inactive_list);
495 prefetchw_prev_lru_page(page,
496 &zone->inactive_list, flags);
498 if (!TestClearPageLRU(page))
500 list_del(&page->lru);
501 if (get_page_testone(page)) {
503 * It is being freed elsewhere
507 list_add(&page->lru, &zone->inactive_list);
510 list_add(&page->lru, &page_list);
513 zone->nr_inactive -= nr_taken;
514 zone->pages_scanned += nr_taken;
515 spin_unlock_irq(&zone->lru_lock);
521 if (current_is_kswapd())
522 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
524 mod_page_state_zone(zone, pgscan_direct, nr_scan);
525 nr_freed = shrink_list(&page_list, gfp_mask,
526 total_scanned, do_writepage);
527 *total_scanned += nr_taken;
528 if (current_is_kswapd())
529 mod_page_state(kswapd_steal, nr_freed);
530 mod_page_state_zone(zone, pgsteal, nr_freed);
533 if (nr_freed <= 0 && list_empty(&page_list))
536 spin_lock_irq(&zone->lru_lock);
538 * Put back any unfreeable pages.
540 while (!list_empty(&page_list)) {
541 page = lru_to_page(&page_list);
542 if (TestSetPageLRU(page))
544 list_del(&page->lru);
545 if (PageActive(page))
546 add_page_to_active_list(zone, page);
548 add_page_to_inactive_list(zone, page);
549 if (!pagevec_add(&pvec, page)) {
550 spin_unlock_irq(&zone->lru_lock);
551 __pagevec_release(&pvec);
552 spin_lock_irq(&zone->lru_lock);
556 spin_unlock_irq(&zone->lru_lock);
558 pagevec_release(&pvec);
563 * This moves pages from the active list to the inactive list.
565 * We move them the other way if the page is referenced by one or more
566 * processes, from rmap.
568 * If the pages are mostly unmapped, the processing is fast and it is
569 * appropriate to hold zone->lru_lock across the whole operation. But if
570 * the pages are mapped, the processing is slow (page_referenced()) so we
571 * should drop zone->lru_lock around each page. It's impossible to balance
572 * this, so instead we remove the pages from the LRU while processing them.
573 * It is safe to rely on PG_active against the non-LRU pages in here because
574 * nobody will play with that bit on a non-LRU page.
576 * The downside is that we have to touch page->_count against each page.
577 * But we had to alter page->flags anyway.
580 refill_inactive_zone(struct zone *zone, const int nr_pages_in,
581 struct page_state *ps)
584 int pgdeactivate = 0;
585 int nr_pages = nr_pages_in;
586 LIST_HEAD(l_hold); /* The pages which were snipped off */
587 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
588 LIST_HEAD(l_active); /* Pages to go onto the active_list */
591 int reclaim_mapped = 0;
598 spin_lock_irq(&zone->lru_lock);
599 while (nr_pages && !list_empty(&zone->active_list)) {
600 page = lru_to_page(&zone->active_list);
601 prefetchw_prev_lru_page(page, &zone->active_list, flags);
602 if (!TestClearPageLRU(page))
604 list_del(&page->lru);
605 if (get_page_testone(page)) {
607 * It was already free! release_pages() or put_page()
608 * are about to remove it from the LRU and free it. So
609 * put the refcount back and put the page back on the
614 list_add(&page->lru, &zone->active_list);
616 list_add(&page->lru, &l_hold);
621 zone->nr_active -= pgmoved;
622 spin_unlock_irq(&zone->lru_lock);
625 * `distress' is a measure of how much trouble we're having reclaiming
626 * pages. 0 -> no problems. 100 -> great trouble.
628 distress = 100 >> zone->prev_priority;
631 * The point of this algorithm is to decide when to start reclaiming
632 * mapped memory instead of just pagecache. Work out how much memory
635 mapped_ratio = (ps->nr_mapped * 100) / total_memory;
638 * Now decide how much we really want to unmap some pages. The mapped
639 * ratio is downgraded - just because there's a lot of mapped memory
640 * doesn't necessarily mean that page reclaim isn't succeeding.
642 * The distress ratio is important - we don't want to start going oom.
644 * A 100% value of vm_swappiness overrides this algorithm altogether.
646 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
649 * Now use this metric to decide whether to start moving mapped memory
650 * onto the inactive list.
652 if (swap_tendency >= 100)
655 while (!list_empty(&l_hold)) {
656 page = lru_to_page(&l_hold);
657 list_del(&page->lru);
658 if (page_mapped(page)) {
659 if (!reclaim_mapped) {
660 list_add(&page->lru, &l_active);
664 if (page_referenced(page)) {
665 page_map_unlock(page);
666 list_add(&page->lru, &l_active);
669 page_map_unlock(page);
672 * FIXME: need to consider page_count(page) here if/when we
673 * reap orphaned pages via the LRU (Daniel's locking stuff)
675 if (total_swap_pages == 0 && PageAnon(page)) {
676 list_add(&page->lru, &l_active);
679 list_add(&page->lru, &l_inactive);
682 pagevec_init(&pvec, 1);
684 spin_lock_irq(&zone->lru_lock);
685 while (!list_empty(&l_inactive)) {
686 page = lru_to_page(&l_inactive);
687 prefetchw_prev_lru_page(page, &l_inactive, flags);
688 if (TestSetPageLRU(page))
690 if (!TestClearPageActive(page))
692 list_move(&page->lru, &zone->inactive_list);
694 if (!pagevec_add(&pvec, page)) {
695 zone->nr_inactive += pgmoved;
696 spin_unlock_irq(&zone->lru_lock);
697 pgdeactivate += pgmoved;
699 if (buffer_heads_over_limit)
700 pagevec_strip(&pvec);
701 __pagevec_release(&pvec);
702 spin_lock_irq(&zone->lru_lock);
705 zone->nr_inactive += pgmoved;
706 pgdeactivate += pgmoved;
707 if (buffer_heads_over_limit) {
708 spin_unlock_irq(&zone->lru_lock);
709 pagevec_strip(&pvec);
710 spin_lock_irq(&zone->lru_lock);
714 while (!list_empty(&l_active)) {
715 page = lru_to_page(&l_active);
716 prefetchw_prev_lru_page(page, &l_active, flags);
717 if (TestSetPageLRU(page))
719 BUG_ON(!PageActive(page));
720 list_move(&page->lru, &zone->active_list);
722 if (!pagevec_add(&pvec, page)) {
723 zone->nr_active += pgmoved;
725 spin_unlock_irq(&zone->lru_lock);
726 __pagevec_release(&pvec);
727 spin_lock_irq(&zone->lru_lock);
730 zone->nr_active += pgmoved;
731 spin_unlock_irq(&zone->lru_lock);
732 pagevec_release(&pvec);
734 mod_page_state_zone(zone, pgrefill, nr_pages_in - nr_pages);
735 mod_page_state(pgdeactivate, pgdeactivate);
739 * Scan `nr_pages' from this zone. Returns the number of reclaimed pages.
740 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
743 shrink_zone(struct zone *zone, int max_scan, unsigned int gfp_mask,
744 int *total_scanned, struct page_state *ps, int do_writepage)
746 unsigned long scan_active;
750 * Try to keep the active list 2/3 of the size of the cache. And
751 * make sure that refill_inactive is given a decent number of pages.
753 * The "scan_active + 1" here is important. With pagecache-intensive
754 * workloads the inactive list is huge, and `ratio' evaluates to zero
755 * all the time. Which pins the active list memory. So we add one to
756 * `scan_active' just to make sure that the kernel will slowly sift
757 * through the active list.
759 if (zone->nr_active >= 4*(zone->nr_inactive*2 + 1)) {
760 /* Don't scan more than 4 times the inactive list scan size */
761 scan_active = 4*max_scan;
763 unsigned long long tmp;
765 /* Cast to long long so the multiply doesn't overflow */
767 tmp = (unsigned long long)max_scan * zone->nr_active;
768 do_div(tmp, zone->nr_inactive*2 + 1);
769 scan_active = (unsigned long)tmp;
772 atomic_add(scan_active + 1, &zone->nr_scan_active);
773 count = atomic_read(&zone->nr_scan_active);
774 if (count >= SWAP_CLUSTER_MAX) {
775 atomic_set(&zone->nr_scan_active, 0);
776 refill_inactive_zone(zone, count, ps);
779 atomic_add(max_scan, &zone->nr_scan_inactive);
780 count = atomic_read(&zone->nr_scan_inactive);
781 if (count >= SWAP_CLUSTER_MAX) {
782 atomic_set(&zone->nr_scan_inactive, 0);
783 return shrink_cache(zone, gfp_mask, count,
784 total_scanned, do_writepage);
790 * This is the direct reclaim path, for page-allocating processes. We only
791 * try to reclaim pages from zones which will satisfy the caller's allocation
794 * We reclaim from a zone even if that zone is over pages_high. Because:
795 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
797 * b) The zones may be over pages_high but they must go *over* pages_high to
798 * satisfy the `incremental min' zone defense algorithm.
800 * Returns the number of reclaimed pages.
802 * If a zone is deemed to be full of pinned pages then just give it a light
803 * scan then give up on it.
806 shrink_caches(struct zone **zones, int priority, int *total_scanned,
807 int gfp_mask, struct page_state *ps, int do_writepage)
812 for (i = 0; zones[i] != NULL; i++) {
813 struct zone *zone = zones[i];
816 if (zone->free_pages < zone->pages_high)
817 zone->temp_priority = priority;
819 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
820 continue; /* Let kswapd poll it */
822 max_scan = (zone->nr_active + zone->nr_inactive) >> priority;
823 ret += shrink_zone(zone, max_scan, gfp_mask,
824 total_scanned, ps, do_writepage);
830 * This is the main entry point to direct page reclaim.
832 * If a full scan of the inactive list fails to free enough memory then we
833 * are "out of memory" and something needs to be killed.
835 * If the caller is !__GFP_FS then the probability of a failure is reasonably
836 * high - the zone may be full of dirty or under-writeback pages, which this
837 * caller can't do much about. So for !__GFP_FS callers, we just perform a
838 * small LRU walk and if that didn't work out, fail the allocation back to the
839 * caller. GFP_NOFS allocators need to know how to deal with it. Kicking
840 * bdflush, waiting and retrying will work.
842 * This is a fairly lame algorithm - it can result in excessive CPU burning and
843 * excessive rotation of the inactive list, which is _supposed_ to be an LRU,
846 int try_to_free_pages(struct zone **zones,
847 unsigned int gfp_mask, unsigned int order)
851 int nr_reclaimed = 0;
852 struct reclaim_state *reclaim_state = current->reclaim_state;
854 unsigned long total_scanned = 0;
855 int do_writepage = 0;
857 inc_page_state(allocstall);
859 for (i = 0; zones[i] != 0; i++)
860 zones[i]->temp_priority = DEF_PRIORITY;
862 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
864 struct page_state ps;
867 nr_reclaimed += shrink_caches(zones, priority, &scanned,
868 gfp_mask, &ps, do_writepage);
869 shrink_slab(scanned, gfp_mask);
871 nr_reclaimed += reclaim_state->reclaimed_slab;
872 reclaim_state->reclaimed_slab = 0;
874 if (nr_reclaimed >= SWAP_CLUSTER_MAX) {
878 if (!(gfp_mask & __GFP_FS))
879 break; /* Let the caller handle it */
881 * Try to write back as many pages as we just scanned. This
882 * tends to cause slow streaming writers to write data to the
883 * disk smoothly, at the dirtying rate, which is nice. But
884 * that's undesirable in laptop mode, where we *want* lumpy
885 * writeout. So in laptop mode, write out the whole world.
887 total_scanned += scanned;
888 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
889 wakeup_bdflush(laptop_mode ? 0 : total_scanned);
893 /* Take a nap, wait for some writeback to complete */
894 if (scanned && priority < DEF_PRIORITY - 2)
895 blk_congestion_wait(WRITE, HZ/10);
897 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
900 for (i = 0; zones[i] != 0; i++)
901 zones[i]->prev_priority = zones[i]->temp_priority;
906 * For kswapd, balance_pgdat() will work across all this node's zones until
907 * they are all at pages_high.
909 * If `nr_pages' is non-zero then it is the number of pages which are to be
910 * reclaimed, regardless of the zone occupancies. This is a software suspend
913 * Returns the number of pages which were actually freed.
915 * There is special handling here for zones which are full of pinned pages.
916 * This can happen if the pages are all mlocked, or if they are all used by
917 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
918 * What we do is to detect the case where all pages in the zone have been
919 * scanned twice and there has been zero successful reclaim. Mark the zone as
920 * dead and from now on, only perform a short scan. Basically we're polling
921 * the zone for when the problem goes away.
923 * kswapd scans the zones in the highmem->normal->dma direction. It skips
924 * zones which have free_pages > pages_high, but once a zone is found to have
925 * free_pages <= pages_high, we scan that zone and the lower zones regardless
926 * of the number of free pages in the lower zones. This interoperates with
927 * the page allocator fallback scheme to ensure that aging of pages is balanced
930 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, struct page_state *ps)
932 int to_free = nr_pages;
935 struct reclaim_state *reclaim_state = current->reclaim_state;
936 unsigned long total_scanned = 0;
937 unsigned long total_reclaimed = 0;
938 int do_writepage = 0;
940 inc_page_state(pageoutrun);
942 for (i = 0; i < pgdat->nr_zones; i++) {
943 struct zone *zone = pgdat->node_zones + i;
945 zone->temp_priority = DEF_PRIORITY;
948 for (priority = DEF_PRIORITY; priority; priority--) {
949 int all_zones_ok = 1;
950 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
955 * Scan in the highmem->dma direction for the highest
956 * zone which needs scanning
958 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
959 struct zone *zone = pgdat->node_zones + i;
961 if (zone->all_unreclaimable &&
962 priority != DEF_PRIORITY)
965 if (zone->free_pages <= zone->pages_high) {
972 end_zone = pgdat->nr_zones - 1;
976 * Now scan the zone in the dma->highmem direction, stopping
977 * at the last zone which needs scanning.
979 * We do this because the page allocator works in the opposite
980 * direction. This prevents the page allocator from allocating
981 * pages behind kswapd's direction of progress, which would
982 * cause too much scanning of the lower zones.
984 for (i = 0; i <= end_zone; i++) {
985 struct zone *zone = pgdat->node_zones + i;
990 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
993 if (nr_pages == 0) { /* Not software suspend */
994 if (zone->free_pages <= zone->pages_high)
997 zone->temp_priority = priority;
998 max_scan = (zone->nr_active + zone->nr_inactive)
1000 reclaimed = shrink_zone(zone, max_scan, GFP_KERNEL,
1001 &scanned, ps, do_writepage);
1002 total_scanned += scanned;
1003 reclaim_state->reclaimed_slab = 0;
1004 shrink_slab(scanned, GFP_KERNEL);
1005 reclaimed += reclaim_state->reclaimed_slab;
1006 total_reclaimed += reclaimed;
1007 to_free -= reclaimed;
1008 if (zone->all_unreclaimable)
1010 if (zone->pages_scanned > zone->present_pages * 2)
1011 zone->all_unreclaimable = 1;
1013 * If we've done a decent amount of scanning and
1014 * the reclaim ratio is low, start doing writepage
1015 * even in laptop mode
1017 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1018 total_scanned > total_reclaimed+total_reclaimed/2)
1021 if (nr_pages && to_free > 0)
1022 continue; /* swsusp: need to do more work */
1024 break; /* kswapd: all done */
1026 * OK, kswapd is getting into trouble. Take a nap, then take
1027 * another pass across the zones.
1029 if (total_scanned && priority < DEF_PRIORITY - 2)
1030 blk_congestion_wait(WRITE, HZ/10);
1033 for (i = 0; i < pgdat->nr_zones; i++) {
1034 struct zone *zone = pgdat->node_zones + i;
1036 zone->prev_priority = zone->temp_priority;
1038 return total_reclaimed;
1042 * The background pageout daemon, started as a kernel thread
1043 * from the init process.
1045 * This basically trickles out pages so that we have _some_
1046 * free memory available even if there is no other activity
1047 * that frees anything up. This is needed for things like routing
1048 * etc, where we otherwise might have all activity going on in
1049 * asynchronous contexts that cannot page things out.
1051 * If there are applications that are active memory-allocators
1052 * (most normal use), this basically shouldn't matter.
1056 pg_data_t *pgdat = (pg_data_t*)p;
1057 struct task_struct *tsk = current;
1059 struct reclaim_state reclaim_state = {
1060 .reclaimed_slab = 0,
1064 daemonize("kswapd%d", pgdat->node_id);
1065 cpumask = node_to_cpumask(pgdat->node_id);
1066 if (!cpus_empty(cpumask))
1067 set_cpus_allowed(tsk, cpumask);
1068 current->reclaim_state = &reclaim_state;
1071 * Tell the memory management that we're a "memory allocator",
1072 * and that if we need more memory we should get access to it
1073 * regardless (see "__alloc_pages()"). "kswapd" should
1074 * never get caught in the normal page freeing logic.
1076 * (Kswapd normally doesn't need memory anyway, but sometimes
1077 * you need a small amount of memory in order to be able to
1078 * page out something else, and this flag essentially protects
1079 * us from recursively trying to free more memory as we're
1080 * trying to free the first piece of memory in the first place).
1082 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1085 struct page_state ps;
1087 if (current->flags & PF_FREEZE)
1088 refrigerator(PF_FREEZE);
1089 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1091 finish_wait(&pgdat->kswapd_wait, &wait);
1092 get_page_state(&ps);
1093 balance_pgdat(pgdat, 0, &ps);
1098 * A zone is low on free memory, so wake its kswapd task to service it.
1100 void wakeup_kswapd(struct zone *zone)
1102 if (zone->free_pages > zone->pages_low)
1104 if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1106 wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1111 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1114 int shrink_all_memory(int nr_pages)
1117 int nr_to_free = nr_pages;
1119 struct reclaim_state reclaim_state = {
1120 .reclaimed_slab = 0,
1123 current->reclaim_state = &reclaim_state;
1124 for_each_pgdat(pgdat) {
1126 struct page_state ps;
1128 get_page_state(&ps);
1129 freed = balance_pgdat(pgdat, nr_to_free, &ps);
1131 nr_to_free -= freed;
1132 if (nr_to_free <= 0)
1135 current->reclaim_state = NULL;
1140 #ifdef CONFIG_HOTPLUG_CPU
1141 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1142 not required for correctness. So if the last cpu in a node goes
1143 away, we get changed to run anywhere: as the first one comes back,
1144 restore their cpu bindings. */
1145 static int __devinit cpu_callback(struct notifier_block *nfb,
1146 unsigned long action,
1152 if (action == CPU_ONLINE) {
1153 for_each_pgdat(pgdat) {
1154 mask = node_to_cpumask(pgdat->node_id);
1155 if (any_online_cpu(mask) != NR_CPUS)
1156 /* One of our CPUs online: restore mask */
1157 set_cpus_allowed(pgdat->kswapd, mask);
1162 #endif /* CONFIG_HOTPLUG_CPU */
1164 static int __init kswapd_init(void)
1168 for_each_pgdat(pgdat)
1170 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1171 total_memory = nr_free_pagecache_pages();
1172 hotcpu_notifier(cpu_callback, 0);
1176 module_init(kswapd_init)