V4L/DVB (4863): Adding support for Pinnacle PCTV 400e DVB-S
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / vmscan.c
blob093f5fe6dd7795cc5d0ca15f5918eedc822b87f4
1 /*
2 * linux/mm/vmscan.c
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/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>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
44 #include <linux/swapops.h>
46 #include "internal.h"
48 struct scan_control {
49 /* Incremented by the number of inactive pages that were scanned */
50 unsigned long nr_scanned;
52 /* This context's GFP mask */
53 gfp_t gfp_mask;
55 int may_writepage;
57 /* Can pages be swapped as part of reclaim? */
58 int may_swap;
60 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
61 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
62 * In this context, it doesn't matter that we scan the
63 * whole list at once. */
64 int swap_cluster_max;
66 int swappiness;
68 int all_unreclaimable;
72 * The list of shrinker callbacks used by to apply pressure to
73 * ageable caches.
75 struct shrinker {
76 shrinker_t shrinker;
77 struct list_head list;
78 int seeks; /* seeks to recreate an obj */
79 long nr; /* objs pending delete */
82 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
84 #ifdef ARCH_HAS_PREFETCH
85 #define prefetch_prev_lru_page(_page, _base, _field) \
86 do { \
87 if ((_page)->lru.prev != _base) { \
88 struct page *prev; \
90 prev = lru_to_page(&(_page->lru)); \
91 prefetch(&prev->_field); \
92 } \
93 } while (0)
94 #else
95 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
96 #endif
98 #ifdef ARCH_HAS_PREFETCHW
99 #define prefetchw_prev_lru_page(_page, _base, _field) \
100 do { \
101 if ((_page)->lru.prev != _base) { \
102 struct page *prev; \
104 prev = lru_to_page(&(_page->lru)); \
105 prefetchw(&prev->_field); \
107 } while (0)
108 #else
109 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
110 #endif
113 * From 0 .. 100. Higher means more swappy.
115 int vm_swappiness = 60;
116 long vm_total_pages; /* The total number of pages which the VM controls */
118 static LIST_HEAD(shrinker_list);
119 static DECLARE_RWSEM(shrinker_rwsem);
122 * Add a shrinker callback to be called from the vm
124 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
126 struct shrinker *shrinker;
128 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
129 if (shrinker) {
130 shrinker->shrinker = theshrinker;
131 shrinker->seeks = seeks;
132 shrinker->nr = 0;
133 down_write(&shrinker_rwsem);
134 list_add_tail(&shrinker->list, &shrinker_list);
135 up_write(&shrinker_rwsem);
137 return shrinker;
139 EXPORT_SYMBOL(set_shrinker);
142 * Remove one
144 void remove_shrinker(struct shrinker *shrinker)
146 down_write(&shrinker_rwsem);
147 list_del(&shrinker->list);
148 up_write(&shrinker_rwsem);
149 kfree(shrinker);
151 EXPORT_SYMBOL(remove_shrinker);
153 #define SHRINK_BATCH 128
155 * Call the shrink functions to age shrinkable caches
157 * Here we assume it costs one seek to replace a lru page and that it also
158 * takes a seek to recreate a cache object. With this in mind we age equal
159 * percentages of the lru and ageable caches. This should balance the seeks
160 * generated by these structures.
162 * If the vm encounted mapped pages on the LRU it increase the pressure on
163 * slab to avoid swapping.
165 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
167 * `lru_pages' represents the number of on-LRU pages in all the zones which
168 * are eligible for the caller's allocation attempt. It is used for balancing
169 * slab reclaim versus page reclaim.
171 * Returns the number of slab objects which we shrunk.
173 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
174 unsigned long lru_pages)
176 struct shrinker *shrinker;
177 unsigned long ret = 0;
179 if (scanned == 0)
180 scanned = SWAP_CLUSTER_MAX;
182 if (!down_read_trylock(&shrinker_rwsem))
183 return 1; /* Assume we'll be able to shrink next time */
185 list_for_each_entry(shrinker, &shrinker_list, list) {
186 unsigned long long delta;
187 unsigned long total_scan;
188 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
190 delta = (4 * scanned) / shrinker->seeks;
191 delta *= max_pass;
192 do_div(delta, lru_pages + 1);
193 shrinker->nr += delta;
194 if (shrinker->nr < 0) {
195 printk(KERN_ERR "%s: nr=%ld\n",
196 __FUNCTION__, shrinker->nr);
197 shrinker->nr = max_pass;
201 * Avoid risking looping forever due to too large nr value:
202 * never try to free more than twice the estimate number of
203 * freeable entries.
205 if (shrinker->nr > max_pass * 2)
206 shrinker->nr = max_pass * 2;
208 total_scan = shrinker->nr;
209 shrinker->nr = 0;
211 while (total_scan >= SHRINK_BATCH) {
212 long this_scan = SHRINK_BATCH;
213 int shrink_ret;
214 int nr_before;
216 nr_before = (*shrinker->shrinker)(0, gfp_mask);
217 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
218 if (shrink_ret == -1)
219 break;
220 if (shrink_ret < nr_before)
221 ret += nr_before - shrink_ret;
222 count_vm_events(SLABS_SCANNED, this_scan);
223 total_scan -= this_scan;
225 cond_resched();
228 shrinker->nr += total_scan;
230 up_read(&shrinker_rwsem);
231 return ret;
234 /* Called without lock on whether page is mapped, so answer is unstable */
235 static inline int page_mapping_inuse(struct page *page)
237 struct address_space *mapping;
239 /* Page is in somebody's page tables. */
240 if (page_mapped(page))
241 return 1;
243 /* Be more reluctant to reclaim swapcache than pagecache */
244 if (PageSwapCache(page))
245 return 1;
247 mapping = page_mapping(page);
248 if (!mapping)
249 return 0;
251 /* File is mmap'd by somebody? */
252 return mapping_mapped(mapping);
255 static inline int is_page_cache_freeable(struct page *page)
257 return page_count(page) - !!PagePrivate(page) == 2;
260 static int may_write_to_queue(struct backing_dev_info *bdi)
262 if (current->flags & PF_SWAPWRITE)
263 return 1;
264 if (!bdi_write_congested(bdi))
265 return 1;
266 if (bdi == current->backing_dev_info)
267 return 1;
268 return 0;
272 * We detected a synchronous write error writing a page out. Probably
273 * -ENOSPC. We need to propagate that into the address_space for a subsequent
274 * fsync(), msync() or close().
276 * The tricky part is that after writepage we cannot touch the mapping: nothing
277 * prevents it from being freed up. But we have a ref on the page and once
278 * that page is locked, the mapping is pinned.
280 * We're allowed to run sleeping lock_page() here because we know the caller has
281 * __GFP_FS.
283 static void handle_write_error(struct address_space *mapping,
284 struct page *page, int error)
286 lock_page(page);
287 if (page_mapping(page) == mapping) {
288 if (error == -ENOSPC)
289 set_bit(AS_ENOSPC, &mapping->flags);
290 else
291 set_bit(AS_EIO, &mapping->flags);
293 unlock_page(page);
296 /* possible outcome of pageout() */
297 typedef enum {
298 /* failed to write page out, page is locked */
299 PAGE_KEEP,
300 /* move page to the active list, page is locked */
301 PAGE_ACTIVATE,
302 /* page has been sent to the disk successfully, page is unlocked */
303 PAGE_SUCCESS,
304 /* page is clean and locked */
305 PAGE_CLEAN,
306 } pageout_t;
309 * pageout is called by shrink_page_list() for each dirty page.
310 * Calls ->writepage().
312 static pageout_t pageout(struct page *page, struct address_space *mapping)
315 * If the page is dirty, only perform writeback if that write
316 * will be non-blocking. To prevent this allocation from being
317 * stalled by pagecache activity. But note that there may be
318 * stalls if we need to run get_block(). We could test
319 * PagePrivate for that.
321 * If this process is currently in generic_file_write() against
322 * this page's queue, we can perform writeback even if that
323 * will block.
325 * If the page is swapcache, write it back even if that would
326 * block, for some throttling. This happens by accident, because
327 * swap_backing_dev_info is bust: it doesn't reflect the
328 * congestion state of the swapdevs. Easy to fix, if needed.
329 * See swapfile.c:page_queue_congested().
331 if (!is_page_cache_freeable(page))
332 return PAGE_KEEP;
333 if (!mapping) {
335 * Some data journaling orphaned pages can have
336 * page->mapping == NULL while being dirty with clean buffers.
338 if (PagePrivate(page)) {
339 if (try_to_free_buffers(page)) {
340 ClearPageDirty(page);
341 printk("%s: orphaned page\n", __FUNCTION__);
342 return PAGE_CLEAN;
345 return PAGE_KEEP;
347 if (mapping->a_ops->writepage == NULL)
348 return PAGE_ACTIVATE;
349 if (!may_write_to_queue(mapping->backing_dev_info))
350 return PAGE_KEEP;
352 if (clear_page_dirty_for_io(page)) {
353 int res;
354 struct writeback_control wbc = {
355 .sync_mode = WB_SYNC_NONE,
356 .nr_to_write = SWAP_CLUSTER_MAX,
357 .range_start = 0,
358 .range_end = LLONG_MAX,
359 .nonblocking = 1,
360 .for_reclaim = 1,
363 SetPageReclaim(page);
364 res = mapping->a_ops->writepage(page, &wbc);
365 if (res < 0)
366 handle_write_error(mapping, page, res);
367 if (res == AOP_WRITEPAGE_ACTIVATE) {
368 ClearPageReclaim(page);
369 return PAGE_ACTIVATE;
371 if (!PageWriteback(page)) {
372 /* synchronous write or broken a_ops? */
373 ClearPageReclaim(page);
375 inc_zone_page_state(page, NR_VMSCAN_WRITE);
376 return PAGE_SUCCESS;
379 return PAGE_CLEAN;
383 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
384 * someone else has a ref on the page, abort and return 0. If it was
385 * successfully detached, return 1. Assumes the caller has a single ref on
386 * this page.
388 int remove_mapping(struct address_space *mapping, struct page *page)
390 BUG_ON(!PageLocked(page));
391 BUG_ON(mapping != page_mapping(page));
393 write_lock_irq(&mapping->tree_lock);
395 * The non racy check for a busy page.
397 * Must be careful with the order of the tests. When someone has
398 * a ref to the page, it may be possible that they dirty it then
399 * drop the reference. So if PageDirty is tested before page_count
400 * here, then the following race may occur:
402 * get_user_pages(&page);
403 * [user mapping goes away]
404 * write_to(page);
405 * !PageDirty(page) [good]
406 * SetPageDirty(page);
407 * put_page(page);
408 * !page_count(page) [good, discard it]
410 * [oops, our write_to data is lost]
412 * Reversing the order of the tests ensures such a situation cannot
413 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
414 * load is not satisfied before that of page->_count.
416 * Note that if SetPageDirty is always performed via set_page_dirty,
417 * and thus under tree_lock, then this ordering is not required.
419 if (unlikely(page_count(page) != 2))
420 goto cannot_free;
421 smp_rmb();
422 if (unlikely(PageDirty(page)))
423 goto cannot_free;
425 if (PageSwapCache(page)) {
426 swp_entry_t swap = { .val = page_private(page) };
427 __delete_from_swap_cache(page);
428 write_unlock_irq(&mapping->tree_lock);
429 swap_free(swap);
430 __put_page(page); /* The pagecache ref */
431 return 1;
434 __remove_from_page_cache(page);
435 write_unlock_irq(&mapping->tree_lock);
436 __put_page(page);
437 return 1;
439 cannot_free:
440 write_unlock_irq(&mapping->tree_lock);
441 return 0;
445 * shrink_page_list() returns the number of reclaimed pages
447 static unsigned long shrink_page_list(struct list_head *page_list,
448 struct scan_control *sc)
450 LIST_HEAD(ret_pages);
451 struct pagevec freed_pvec;
452 int pgactivate = 0;
453 unsigned long nr_reclaimed = 0;
455 cond_resched();
457 pagevec_init(&freed_pvec, 1);
458 while (!list_empty(page_list)) {
459 struct address_space *mapping;
460 struct page *page;
461 int may_enter_fs;
462 int referenced;
464 cond_resched();
466 page = lru_to_page(page_list);
467 list_del(&page->lru);
469 if (TestSetPageLocked(page))
470 goto keep;
472 VM_BUG_ON(PageActive(page));
474 sc->nr_scanned++;
476 if (!sc->may_swap && page_mapped(page))
477 goto keep_locked;
479 /* Double the slab pressure for mapped and swapcache pages */
480 if (page_mapped(page) || PageSwapCache(page))
481 sc->nr_scanned++;
483 if (PageWriteback(page))
484 goto keep_locked;
486 referenced = page_referenced(page, 1);
487 /* In active use or really unfreeable? Activate it. */
488 if (referenced && page_mapping_inuse(page))
489 goto activate_locked;
491 #ifdef CONFIG_SWAP
493 * Anonymous process memory has backing store?
494 * Try to allocate it some swap space here.
496 if (PageAnon(page) && !PageSwapCache(page))
497 if (!add_to_swap(page, GFP_ATOMIC))
498 goto activate_locked;
499 #endif /* CONFIG_SWAP */
501 mapping = page_mapping(page);
502 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
503 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
506 * The page is mapped into the page tables of one or more
507 * processes. Try to unmap it here.
509 if (page_mapped(page) && mapping) {
510 switch (try_to_unmap(page, 0)) {
511 case SWAP_FAIL:
512 goto activate_locked;
513 case SWAP_AGAIN:
514 goto keep_locked;
515 case SWAP_SUCCESS:
516 ; /* try to free the page below */
520 if (PageDirty(page)) {
521 if (referenced)
522 goto keep_locked;
523 if (!may_enter_fs)
524 goto keep_locked;
525 if (!sc->may_writepage)
526 goto keep_locked;
528 /* Page is dirty, try to write it out here */
529 switch(pageout(page, mapping)) {
530 case PAGE_KEEP:
531 goto keep_locked;
532 case PAGE_ACTIVATE:
533 goto activate_locked;
534 case PAGE_SUCCESS:
535 if (PageWriteback(page) || PageDirty(page))
536 goto keep;
538 * A synchronous write - probably a ramdisk. Go
539 * ahead and try to reclaim the page.
541 if (TestSetPageLocked(page))
542 goto keep;
543 if (PageDirty(page) || PageWriteback(page))
544 goto keep_locked;
545 mapping = page_mapping(page);
546 case PAGE_CLEAN:
547 ; /* try to free the page below */
552 * If the page has buffers, try to free the buffer mappings
553 * associated with this page. If we succeed we try to free
554 * the page as well.
556 * We do this even if the page is PageDirty().
557 * try_to_release_page() does not perform I/O, but it is
558 * possible for a page to have PageDirty set, but it is actually
559 * clean (all its buffers are clean). This happens if the
560 * buffers were written out directly, with submit_bh(). ext3
561 * will do this, as well as the blockdev mapping.
562 * try_to_release_page() will discover that cleanness and will
563 * drop the buffers and mark the page clean - it can be freed.
565 * Rarely, pages can have buffers and no ->mapping. These are
566 * the pages which were not successfully invalidated in
567 * truncate_complete_page(). We try to drop those buffers here
568 * and if that worked, and the page is no longer mapped into
569 * process address space (page_count == 1) it can be freed.
570 * Otherwise, leave the page on the LRU so it is swappable.
572 if (PagePrivate(page)) {
573 if (!try_to_release_page(page, sc->gfp_mask))
574 goto activate_locked;
575 if (!mapping && page_count(page) == 1)
576 goto free_it;
579 if (!mapping || !remove_mapping(mapping, page))
580 goto keep_locked;
582 free_it:
583 unlock_page(page);
584 nr_reclaimed++;
585 if (!pagevec_add(&freed_pvec, page))
586 __pagevec_release_nonlru(&freed_pvec);
587 continue;
589 activate_locked:
590 SetPageActive(page);
591 pgactivate++;
592 keep_locked:
593 unlock_page(page);
594 keep:
595 list_add(&page->lru, &ret_pages);
596 VM_BUG_ON(PageLRU(page));
598 list_splice(&ret_pages, page_list);
599 if (pagevec_count(&freed_pvec))
600 __pagevec_release_nonlru(&freed_pvec);
601 count_vm_events(PGACTIVATE, pgactivate);
602 return nr_reclaimed;
606 * zone->lru_lock is heavily contended. Some of the functions that
607 * shrink the lists perform better by taking out a batch of pages
608 * and working on them outside the LRU lock.
610 * For pagecache intensive workloads, this function is the hottest
611 * spot in the kernel (apart from copy_*_user functions).
613 * Appropriate locks must be held before calling this function.
615 * @nr_to_scan: The number of pages to look through on the list.
616 * @src: The LRU list to pull pages off.
617 * @dst: The temp list to put pages on to.
618 * @scanned: The number of pages that were scanned.
620 * returns how many pages were moved onto *@dst.
622 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
623 struct list_head *src, struct list_head *dst,
624 unsigned long *scanned)
626 unsigned long nr_taken = 0;
627 struct page *page;
628 unsigned long scan;
630 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
631 struct list_head *target;
632 page = lru_to_page(src);
633 prefetchw_prev_lru_page(page, src, flags);
635 VM_BUG_ON(!PageLRU(page));
637 list_del(&page->lru);
638 target = src;
639 if (likely(get_page_unless_zero(page))) {
641 * Be careful not to clear PageLRU until after we're
642 * sure the page is not being freed elsewhere -- the
643 * page release code relies on it.
645 ClearPageLRU(page);
646 target = dst;
647 nr_taken++;
648 } /* else it is being freed elsewhere */
650 list_add(&page->lru, target);
653 *scanned = scan;
654 return nr_taken;
658 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
659 * of reclaimed pages
661 static unsigned long shrink_inactive_list(unsigned long max_scan,
662 struct zone *zone, struct scan_control *sc)
664 LIST_HEAD(page_list);
665 struct pagevec pvec;
666 unsigned long nr_scanned = 0;
667 unsigned long nr_reclaimed = 0;
669 pagevec_init(&pvec, 1);
671 lru_add_drain();
672 spin_lock_irq(&zone->lru_lock);
673 do {
674 struct page *page;
675 unsigned long nr_taken;
676 unsigned long nr_scan;
677 unsigned long nr_freed;
679 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
680 &zone->inactive_list,
681 &page_list, &nr_scan);
682 zone->nr_inactive -= nr_taken;
683 zone->pages_scanned += nr_scan;
684 spin_unlock_irq(&zone->lru_lock);
686 nr_scanned += nr_scan;
687 nr_freed = shrink_page_list(&page_list, sc);
688 nr_reclaimed += nr_freed;
689 local_irq_disable();
690 if (current_is_kswapd()) {
691 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
692 __count_vm_events(KSWAPD_STEAL, nr_freed);
693 } else
694 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
695 __count_vm_events(PGACTIVATE, nr_freed);
697 if (nr_taken == 0)
698 goto done;
700 spin_lock(&zone->lru_lock);
702 * Put back any unfreeable pages.
704 while (!list_empty(&page_list)) {
705 page = lru_to_page(&page_list);
706 VM_BUG_ON(PageLRU(page));
707 SetPageLRU(page);
708 list_del(&page->lru);
709 if (PageActive(page))
710 add_page_to_active_list(zone, page);
711 else
712 add_page_to_inactive_list(zone, page);
713 if (!pagevec_add(&pvec, page)) {
714 spin_unlock_irq(&zone->lru_lock);
715 __pagevec_release(&pvec);
716 spin_lock_irq(&zone->lru_lock);
719 } while (nr_scanned < max_scan);
720 spin_unlock(&zone->lru_lock);
721 done:
722 local_irq_enable();
723 pagevec_release(&pvec);
724 return nr_reclaimed;
728 * We are about to scan this zone at a certain priority level. If that priority
729 * level is smaller (ie: more urgent) than the previous priority, then note
730 * that priority level within the zone. This is done so that when the next
731 * process comes in to scan this zone, it will immediately start out at this
732 * priority level rather than having to build up its own scanning priority.
733 * Here, this priority affects only the reclaim-mapped threshold.
735 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
737 if (priority < zone->prev_priority)
738 zone->prev_priority = priority;
741 static inline int zone_is_near_oom(struct zone *zone)
743 return zone->pages_scanned >= (zone->nr_active + zone->nr_inactive)*3;
747 * This moves pages from the active list to the inactive list.
749 * We move them the other way if the page is referenced by one or more
750 * processes, from rmap.
752 * If the pages are mostly unmapped, the processing is fast and it is
753 * appropriate to hold zone->lru_lock across the whole operation. But if
754 * the pages are mapped, the processing is slow (page_referenced()) so we
755 * should drop zone->lru_lock around each page. It's impossible to balance
756 * this, so instead we remove the pages from the LRU while processing them.
757 * It is safe to rely on PG_active against the non-LRU pages in here because
758 * nobody will play with that bit on a non-LRU page.
760 * The downside is that we have to touch page->_count against each page.
761 * But we had to alter page->flags anyway.
763 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
764 struct scan_control *sc, int priority)
766 unsigned long pgmoved;
767 int pgdeactivate = 0;
768 unsigned long pgscanned;
769 LIST_HEAD(l_hold); /* The pages which were snipped off */
770 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
771 LIST_HEAD(l_active); /* Pages to go onto the active_list */
772 struct page *page;
773 struct pagevec pvec;
774 int reclaim_mapped = 0;
776 if (sc->may_swap) {
777 long mapped_ratio;
778 long distress;
779 long swap_tendency;
781 if (zone_is_near_oom(zone))
782 goto force_reclaim_mapped;
785 * `distress' is a measure of how much trouble we're having
786 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
788 distress = 100 >> min(zone->prev_priority, priority);
791 * The point of this algorithm is to decide when to start
792 * reclaiming mapped memory instead of just pagecache. Work out
793 * how much memory
794 * is mapped.
796 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
797 global_page_state(NR_ANON_PAGES)) * 100) /
798 vm_total_pages;
801 * Now decide how much we really want to unmap some pages. The
802 * mapped ratio is downgraded - just because there's a lot of
803 * mapped memory doesn't necessarily mean that page reclaim
804 * isn't succeeding.
806 * The distress ratio is important - we don't want to start
807 * going oom.
809 * A 100% value of vm_swappiness overrides this algorithm
810 * altogether.
812 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
815 * Now use this metric to decide whether to start moving mapped
816 * memory onto the inactive list.
818 if (swap_tendency >= 100)
819 force_reclaim_mapped:
820 reclaim_mapped = 1;
823 lru_add_drain();
824 spin_lock_irq(&zone->lru_lock);
825 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
826 &l_hold, &pgscanned);
827 zone->pages_scanned += pgscanned;
828 zone->nr_active -= pgmoved;
829 spin_unlock_irq(&zone->lru_lock);
831 while (!list_empty(&l_hold)) {
832 cond_resched();
833 page = lru_to_page(&l_hold);
834 list_del(&page->lru);
835 if (page_mapped(page)) {
836 if (!reclaim_mapped ||
837 (total_swap_pages == 0 && PageAnon(page)) ||
838 page_referenced(page, 0)) {
839 list_add(&page->lru, &l_active);
840 continue;
843 list_add(&page->lru, &l_inactive);
846 pagevec_init(&pvec, 1);
847 pgmoved = 0;
848 spin_lock_irq(&zone->lru_lock);
849 while (!list_empty(&l_inactive)) {
850 page = lru_to_page(&l_inactive);
851 prefetchw_prev_lru_page(page, &l_inactive, flags);
852 VM_BUG_ON(PageLRU(page));
853 SetPageLRU(page);
854 VM_BUG_ON(!PageActive(page));
855 ClearPageActive(page);
857 list_move(&page->lru, &zone->inactive_list);
858 pgmoved++;
859 if (!pagevec_add(&pvec, page)) {
860 zone->nr_inactive += pgmoved;
861 spin_unlock_irq(&zone->lru_lock);
862 pgdeactivate += pgmoved;
863 pgmoved = 0;
864 if (buffer_heads_over_limit)
865 pagevec_strip(&pvec);
866 __pagevec_release(&pvec);
867 spin_lock_irq(&zone->lru_lock);
870 zone->nr_inactive += pgmoved;
871 pgdeactivate += pgmoved;
872 if (buffer_heads_over_limit) {
873 spin_unlock_irq(&zone->lru_lock);
874 pagevec_strip(&pvec);
875 spin_lock_irq(&zone->lru_lock);
878 pgmoved = 0;
879 while (!list_empty(&l_active)) {
880 page = lru_to_page(&l_active);
881 prefetchw_prev_lru_page(page, &l_active, flags);
882 VM_BUG_ON(PageLRU(page));
883 SetPageLRU(page);
884 VM_BUG_ON(!PageActive(page));
885 list_move(&page->lru, &zone->active_list);
886 pgmoved++;
887 if (!pagevec_add(&pvec, page)) {
888 zone->nr_active += pgmoved;
889 pgmoved = 0;
890 spin_unlock_irq(&zone->lru_lock);
891 __pagevec_release(&pvec);
892 spin_lock_irq(&zone->lru_lock);
895 zone->nr_active += pgmoved;
897 __count_zone_vm_events(PGREFILL, zone, pgscanned);
898 __count_vm_events(PGDEACTIVATE, pgdeactivate);
899 spin_unlock_irq(&zone->lru_lock);
901 pagevec_release(&pvec);
905 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
907 static unsigned long shrink_zone(int priority, struct zone *zone,
908 struct scan_control *sc)
910 unsigned long nr_active;
911 unsigned long nr_inactive;
912 unsigned long nr_to_scan;
913 unsigned long nr_reclaimed = 0;
915 atomic_inc(&zone->reclaim_in_progress);
918 * Add one to `nr_to_scan' just to make sure that the kernel will
919 * slowly sift through the active list.
921 zone->nr_scan_active += (zone->nr_active >> priority) + 1;
922 nr_active = zone->nr_scan_active;
923 if (nr_active >= sc->swap_cluster_max)
924 zone->nr_scan_active = 0;
925 else
926 nr_active = 0;
928 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
929 nr_inactive = zone->nr_scan_inactive;
930 if (nr_inactive >= sc->swap_cluster_max)
931 zone->nr_scan_inactive = 0;
932 else
933 nr_inactive = 0;
935 while (nr_active || nr_inactive) {
936 if (nr_active) {
937 nr_to_scan = min(nr_active,
938 (unsigned long)sc->swap_cluster_max);
939 nr_active -= nr_to_scan;
940 shrink_active_list(nr_to_scan, zone, sc, priority);
943 if (nr_inactive) {
944 nr_to_scan = min(nr_inactive,
945 (unsigned long)sc->swap_cluster_max);
946 nr_inactive -= nr_to_scan;
947 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
948 sc);
952 throttle_vm_writeout();
954 atomic_dec(&zone->reclaim_in_progress);
955 return nr_reclaimed;
959 * This is the direct reclaim path, for page-allocating processes. We only
960 * try to reclaim pages from zones which will satisfy the caller's allocation
961 * request.
963 * We reclaim from a zone even if that zone is over pages_high. Because:
964 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
965 * allocation or
966 * b) The zones may be over pages_high but they must go *over* pages_high to
967 * satisfy the `incremental min' zone defense algorithm.
969 * Returns the number of reclaimed pages.
971 * If a zone is deemed to be full of pinned pages then just give it a light
972 * scan then give up on it.
974 static unsigned long shrink_zones(int priority, struct zone **zones,
975 struct scan_control *sc)
977 unsigned long nr_reclaimed = 0;
978 int i;
980 sc->all_unreclaimable = 1;
981 for (i = 0; zones[i] != NULL; i++) {
982 struct zone *zone = zones[i];
984 if (!populated_zone(zone))
985 continue;
987 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
988 continue;
990 note_zone_scanning_priority(zone, priority);
992 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
993 continue; /* Let kswapd poll it */
995 sc->all_unreclaimable = 0;
997 nr_reclaimed += shrink_zone(priority, zone, sc);
999 return nr_reclaimed;
1003 * This is the main entry point to direct page reclaim.
1005 * If a full scan of the inactive list fails to free enough memory then we
1006 * are "out of memory" and something needs to be killed.
1008 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1009 * high - the zone may be full of dirty or under-writeback pages, which this
1010 * caller can't do much about. We kick pdflush and take explicit naps in the
1011 * hope that some of these pages can be written. But if the allocating task
1012 * holds filesystem locks which prevent writeout this might not work, and the
1013 * allocation attempt will fail.
1015 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1017 int priority;
1018 int ret = 0;
1019 unsigned long total_scanned = 0;
1020 unsigned long nr_reclaimed = 0;
1021 struct reclaim_state *reclaim_state = current->reclaim_state;
1022 unsigned long lru_pages = 0;
1023 int i;
1024 struct scan_control sc = {
1025 .gfp_mask = gfp_mask,
1026 .may_writepage = !laptop_mode,
1027 .swap_cluster_max = SWAP_CLUSTER_MAX,
1028 .may_swap = 1,
1029 .swappiness = vm_swappiness,
1032 count_vm_event(ALLOCSTALL);
1034 for (i = 0; zones[i] != NULL; i++) {
1035 struct zone *zone = zones[i];
1037 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1038 continue;
1040 lru_pages += zone->nr_active + zone->nr_inactive;
1043 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1044 sc.nr_scanned = 0;
1045 if (!priority)
1046 disable_swap_token();
1047 nr_reclaimed += shrink_zones(priority, zones, &sc);
1048 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1049 if (reclaim_state) {
1050 nr_reclaimed += reclaim_state->reclaimed_slab;
1051 reclaim_state->reclaimed_slab = 0;
1053 total_scanned += sc.nr_scanned;
1054 if (nr_reclaimed >= sc.swap_cluster_max) {
1055 ret = 1;
1056 goto out;
1060 * Try to write back as many pages as we just scanned. This
1061 * tends to cause slow streaming writers to write data to the
1062 * disk smoothly, at the dirtying rate, which is nice. But
1063 * that's undesirable in laptop mode, where we *want* lumpy
1064 * writeout. So in laptop mode, write out the whole world.
1066 if (total_scanned > sc.swap_cluster_max +
1067 sc.swap_cluster_max / 2) {
1068 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1069 sc.may_writepage = 1;
1072 /* Take a nap, wait for some writeback to complete */
1073 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1074 congestion_wait(WRITE, HZ/10);
1076 /* top priority shrink_caches still had more to do? don't OOM, then */
1077 if (!sc.all_unreclaimable)
1078 ret = 1;
1079 out:
1081 * Now that we've scanned all the zones at this priority level, note
1082 * that level within the zone so that the next thread which performs
1083 * scanning of this zone will immediately start out at this priority
1084 * level. This affects only the decision whether or not to bring
1085 * mapped pages onto the inactive list.
1087 if (priority < 0)
1088 priority = 0;
1089 for (i = 0; zones[i] != 0; i++) {
1090 struct zone *zone = zones[i];
1092 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1093 continue;
1095 zone->prev_priority = priority;
1097 return ret;
1101 * For kswapd, balance_pgdat() will work across all this node's zones until
1102 * they are all at pages_high.
1104 * Returns the number of pages which were actually freed.
1106 * There is special handling here for zones which are full of pinned pages.
1107 * This can happen if the pages are all mlocked, or if they are all used by
1108 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1109 * What we do is to detect the case where all pages in the zone have been
1110 * scanned twice and there has been zero successful reclaim. Mark the zone as
1111 * dead and from now on, only perform a short scan. Basically we're polling
1112 * the zone for when the problem goes away.
1114 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1115 * zones which have free_pages > pages_high, but once a zone is found to have
1116 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1117 * of the number of free pages in the lower zones. This interoperates with
1118 * the page allocator fallback scheme to ensure that aging of pages is balanced
1119 * across the zones.
1121 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1123 int all_zones_ok;
1124 int priority;
1125 int i;
1126 unsigned long total_scanned;
1127 unsigned long nr_reclaimed;
1128 struct reclaim_state *reclaim_state = current->reclaim_state;
1129 struct scan_control sc = {
1130 .gfp_mask = GFP_KERNEL,
1131 .may_swap = 1,
1132 .swap_cluster_max = SWAP_CLUSTER_MAX,
1133 .swappiness = vm_swappiness,
1136 * temp_priority is used to remember the scanning priority at which
1137 * this zone was successfully refilled to free_pages == pages_high.
1139 int temp_priority[MAX_NR_ZONES];
1141 loop_again:
1142 total_scanned = 0;
1143 nr_reclaimed = 0;
1144 sc.may_writepage = !laptop_mode;
1145 count_vm_event(PAGEOUTRUN);
1147 for (i = 0; i < pgdat->nr_zones; i++)
1148 temp_priority[i] = DEF_PRIORITY;
1150 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1151 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1152 unsigned long lru_pages = 0;
1154 /* The swap token gets in the way of swapout... */
1155 if (!priority)
1156 disable_swap_token();
1158 all_zones_ok = 1;
1161 * Scan in the highmem->dma direction for the highest
1162 * zone which needs scanning
1164 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1165 struct zone *zone = pgdat->node_zones + i;
1167 if (!populated_zone(zone))
1168 continue;
1170 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1171 continue;
1173 if (!zone_watermark_ok(zone, order, zone->pages_high,
1174 0, 0)) {
1175 end_zone = i;
1176 break;
1179 if (i < 0)
1180 goto out;
1182 for (i = 0; i <= end_zone; i++) {
1183 struct zone *zone = pgdat->node_zones + i;
1185 lru_pages += zone->nr_active + zone->nr_inactive;
1189 * Now scan the zone in the dma->highmem direction, stopping
1190 * at the last zone which needs scanning.
1192 * We do this because the page allocator works in the opposite
1193 * direction. This prevents the page allocator from allocating
1194 * pages behind kswapd's direction of progress, which would
1195 * cause too much scanning of the lower zones.
1197 for (i = 0; i <= end_zone; i++) {
1198 struct zone *zone = pgdat->node_zones + i;
1199 int nr_slab;
1201 if (!populated_zone(zone))
1202 continue;
1204 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1205 continue;
1207 if (!zone_watermark_ok(zone, order, zone->pages_high,
1208 end_zone, 0))
1209 all_zones_ok = 0;
1210 temp_priority[i] = priority;
1211 sc.nr_scanned = 0;
1212 note_zone_scanning_priority(zone, priority);
1213 nr_reclaimed += shrink_zone(priority, zone, &sc);
1214 reclaim_state->reclaimed_slab = 0;
1215 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1216 lru_pages);
1217 nr_reclaimed += reclaim_state->reclaimed_slab;
1218 total_scanned += sc.nr_scanned;
1219 if (zone->all_unreclaimable)
1220 continue;
1221 if (nr_slab == 0 && zone->pages_scanned >=
1222 (zone->nr_active + zone->nr_inactive) * 6)
1223 zone->all_unreclaimable = 1;
1225 * If we've done a decent amount of scanning and
1226 * the reclaim ratio is low, start doing writepage
1227 * even in laptop mode
1229 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1230 total_scanned > nr_reclaimed + nr_reclaimed / 2)
1231 sc.may_writepage = 1;
1233 if (all_zones_ok)
1234 break; /* kswapd: all done */
1236 * OK, kswapd is getting into trouble. Take a nap, then take
1237 * another pass across the zones.
1239 if (total_scanned && priority < DEF_PRIORITY - 2)
1240 congestion_wait(WRITE, HZ/10);
1243 * We do this so kswapd doesn't build up large priorities for
1244 * example when it is freeing in parallel with allocators. It
1245 * matches the direct reclaim path behaviour in terms of impact
1246 * on zone->*_priority.
1248 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1249 break;
1251 out:
1253 * Note within each zone the priority level at which this zone was
1254 * brought into a happy state. So that the next thread which scans this
1255 * zone will start out at that priority level.
1257 for (i = 0; i < pgdat->nr_zones; i++) {
1258 struct zone *zone = pgdat->node_zones + i;
1260 zone->prev_priority = temp_priority[i];
1262 if (!all_zones_ok) {
1263 cond_resched();
1265 try_to_freeze();
1267 goto loop_again;
1270 return nr_reclaimed;
1274 * The background pageout daemon, started as a kernel thread
1275 * from the init process.
1277 * This basically trickles out pages so that we have _some_
1278 * free memory available even if there is no other activity
1279 * that frees anything up. This is needed for things like routing
1280 * etc, where we otherwise might have all activity going on in
1281 * asynchronous contexts that cannot page things out.
1283 * If there are applications that are active memory-allocators
1284 * (most normal use), this basically shouldn't matter.
1286 static int kswapd(void *p)
1288 unsigned long order;
1289 pg_data_t *pgdat = (pg_data_t*)p;
1290 struct task_struct *tsk = current;
1291 DEFINE_WAIT(wait);
1292 struct reclaim_state reclaim_state = {
1293 .reclaimed_slab = 0,
1295 cpumask_t cpumask;
1297 cpumask = node_to_cpumask(pgdat->node_id);
1298 if (!cpus_empty(cpumask))
1299 set_cpus_allowed(tsk, cpumask);
1300 current->reclaim_state = &reclaim_state;
1303 * Tell the memory management that we're a "memory allocator",
1304 * and that if we need more memory we should get access to it
1305 * regardless (see "__alloc_pages()"). "kswapd" should
1306 * never get caught in the normal page freeing logic.
1308 * (Kswapd normally doesn't need memory anyway, but sometimes
1309 * you need a small amount of memory in order to be able to
1310 * page out something else, and this flag essentially protects
1311 * us from recursively trying to free more memory as we're
1312 * trying to free the first piece of memory in the first place).
1314 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1316 order = 0;
1317 for ( ; ; ) {
1318 unsigned long new_order;
1320 try_to_freeze();
1322 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1323 new_order = pgdat->kswapd_max_order;
1324 pgdat->kswapd_max_order = 0;
1325 if (order < new_order) {
1327 * Don't sleep if someone wants a larger 'order'
1328 * allocation
1330 order = new_order;
1331 } else {
1332 schedule();
1333 order = pgdat->kswapd_max_order;
1335 finish_wait(&pgdat->kswapd_wait, &wait);
1337 balance_pgdat(pgdat, order);
1339 return 0;
1343 * A zone is low on free memory, so wake its kswapd task to service it.
1345 void wakeup_kswapd(struct zone *zone, int order)
1347 pg_data_t *pgdat;
1349 if (!populated_zone(zone))
1350 return;
1352 pgdat = zone->zone_pgdat;
1353 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1354 return;
1355 if (pgdat->kswapd_max_order < order)
1356 pgdat->kswapd_max_order = order;
1357 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1358 return;
1359 if (!waitqueue_active(&pgdat->kswapd_wait))
1360 return;
1361 wake_up_interruptible(&pgdat->kswapd_wait);
1364 #ifdef CONFIG_PM
1366 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1367 * from LRU lists system-wide, for given pass and priority, and returns the
1368 * number of reclaimed pages
1370 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1372 static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1373 int prio, struct scan_control *sc)
1375 struct zone *zone;
1376 unsigned long nr_to_scan, ret = 0;
1378 for_each_zone(zone) {
1380 if (!populated_zone(zone))
1381 continue;
1383 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1384 continue;
1386 /* For pass = 0 we don't shrink the active list */
1387 if (pass > 0) {
1388 zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1389 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1390 zone->nr_scan_active = 0;
1391 nr_to_scan = min(nr_pages, zone->nr_active);
1392 shrink_active_list(nr_to_scan, zone, sc, prio);
1396 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1397 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1398 zone->nr_scan_inactive = 0;
1399 nr_to_scan = min(nr_pages, zone->nr_inactive);
1400 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1401 if (ret >= nr_pages)
1402 return ret;
1406 return ret;
1410 * Try to free `nr_pages' of memory, system-wide, and return the number of
1411 * freed pages.
1413 * Rather than trying to age LRUs the aim is to preserve the overall
1414 * LRU order by reclaiming preferentially
1415 * inactive > active > active referenced > active mapped
1417 unsigned long shrink_all_memory(unsigned long nr_pages)
1419 unsigned long lru_pages, nr_slab;
1420 unsigned long ret = 0;
1421 int pass;
1422 struct reclaim_state reclaim_state;
1423 struct zone *zone;
1424 struct scan_control sc = {
1425 .gfp_mask = GFP_KERNEL,
1426 .may_swap = 0,
1427 .swap_cluster_max = nr_pages,
1428 .may_writepage = 1,
1429 .swappiness = vm_swappiness,
1432 current->reclaim_state = &reclaim_state;
1434 lru_pages = 0;
1435 for_each_zone(zone)
1436 lru_pages += zone->nr_active + zone->nr_inactive;
1438 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1439 /* If slab caches are huge, it's better to hit them first */
1440 while (nr_slab >= lru_pages) {
1441 reclaim_state.reclaimed_slab = 0;
1442 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1443 if (!reclaim_state.reclaimed_slab)
1444 break;
1446 ret += reclaim_state.reclaimed_slab;
1447 if (ret >= nr_pages)
1448 goto out;
1450 nr_slab -= reclaim_state.reclaimed_slab;
1454 * We try to shrink LRUs in 5 passes:
1455 * 0 = Reclaim from inactive_list only
1456 * 1 = Reclaim from active list but don't reclaim mapped
1457 * 2 = 2nd pass of type 1
1458 * 3 = Reclaim mapped (normal reclaim)
1459 * 4 = 2nd pass of type 3
1461 for (pass = 0; pass < 5; pass++) {
1462 int prio;
1464 /* Needed for shrinking slab caches later on */
1465 if (!lru_pages)
1466 for_each_zone(zone) {
1467 lru_pages += zone->nr_active;
1468 lru_pages += zone->nr_inactive;
1471 /* Force reclaiming mapped pages in the passes #3 and #4 */
1472 if (pass > 2) {
1473 sc.may_swap = 1;
1474 sc.swappiness = 100;
1477 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1478 unsigned long nr_to_scan = nr_pages - ret;
1480 sc.nr_scanned = 0;
1481 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1482 if (ret >= nr_pages)
1483 goto out;
1485 reclaim_state.reclaimed_slab = 0;
1486 shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1487 ret += reclaim_state.reclaimed_slab;
1488 if (ret >= nr_pages)
1489 goto out;
1491 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1492 congestion_wait(WRITE, HZ / 10);
1495 lru_pages = 0;
1499 * If ret = 0, we could not shrink LRUs, but there may be something
1500 * in slab caches
1502 if (!ret)
1503 do {
1504 reclaim_state.reclaimed_slab = 0;
1505 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1506 ret += reclaim_state.reclaimed_slab;
1507 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1509 out:
1510 current->reclaim_state = NULL;
1512 return ret;
1514 #endif
1516 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1517 not required for correctness. So if the last cpu in a node goes
1518 away, we get changed to run anywhere: as the first one comes back,
1519 restore their cpu bindings. */
1520 static int __devinit cpu_callback(struct notifier_block *nfb,
1521 unsigned long action, void *hcpu)
1523 pg_data_t *pgdat;
1524 cpumask_t mask;
1526 if (action == CPU_ONLINE) {
1527 for_each_online_pgdat(pgdat) {
1528 mask = node_to_cpumask(pgdat->node_id);
1529 if (any_online_cpu(mask) != NR_CPUS)
1530 /* One of our CPUs online: restore mask */
1531 set_cpus_allowed(pgdat->kswapd, mask);
1534 return NOTIFY_OK;
1538 * This kswapd start function will be called by init and node-hot-add.
1539 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1541 int kswapd_run(int nid)
1543 pg_data_t *pgdat = NODE_DATA(nid);
1544 int ret = 0;
1546 if (pgdat->kswapd)
1547 return 0;
1549 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1550 if (IS_ERR(pgdat->kswapd)) {
1551 /* failure at boot is fatal */
1552 BUG_ON(system_state == SYSTEM_BOOTING);
1553 printk("Failed to start kswapd on node %d\n",nid);
1554 ret = -1;
1556 return ret;
1559 static int __init kswapd_init(void)
1561 int nid;
1563 swap_setup();
1564 for_each_online_node(nid)
1565 kswapd_run(nid);
1566 hotcpu_notifier(cpu_callback, 0);
1567 return 0;
1570 module_init(kswapd_init)
1572 #ifdef CONFIG_NUMA
1574 * Zone reclaim mode
1576 * If non-zero call zone_reclaim when the number of free pages falls below
1577 * the watermarks.
1579 int zone_reclaim_mode __read_mostly;
1581 #define RECLAIM_OFF 0
1582 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1583 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1584 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1587 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1588 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1589 * a zone.
1591 #define ZONE_RECLAIM_PRIORITY 4
1594 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1595 * occur.
1597 int sysctl_min_unmapped_ratio = 1;
1600 * If the number of slab pages in a zone grows beyond this percentage then
1601 * slab reclaim needs to occur.
1603 int sysctl_min_slab_ratio = 5;
1606 * Try to free up some pages from this zone through reclaim.
1608 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1610 /* Minimum pages needed in order to stay on node */
1611 const unsigned long nr_pages = 1 << order;
1612 struct task_struct *p = current;
1613 struct reclaim_state reclaim_state;
1614 int priority;
1615 unsigned long nr_reclaimed = 0;
1616 struct scan_control sc = {
1617 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1618 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1619 .swap_cluster_max = max_t(unsigned long, nr_pages,
1620 SWAP_CLUSTER_MAX),
1621 .gfp_mask = gfp_mask,
1622 .swappiness = vm_swappiness,
1624 unsigned long slab_reclaimable;
1626 disable_swap_token();
1627 cond_resched();
1629 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1630 * and we also need to be able to write out pages for RECLAIM_WRITE
1631 * and RECLAIM_SWAP.
1633 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1634 reclaim_state.reclaimed_slab = 0;
1635 p->reclaim_state = &reclaim_state;
1637 if (zone_page_state(zone, NR_FILE_PAGES) -
1638 zone_page_state(zone, NR_FILE_MAPPED) >
1639 zone->min_unmapped_pages) {
1641 * Free memory by calling shrink zone with increasing
1642 * priorities until we have enough memory freed.
1644 priority = ZONE_RECLAIM_PRIORITY;
1645 do {
1646 note_zone_scanning_priority(zone, priority);
1647 nr_reclaimed += shrink_zone(priority, zone, &sc);
1648 priority--;
1649 } while (priority >= 0 && nr_reclaimed < nr_pages);
1652 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1653 if (slab_reclaimable > zone->min_slab_pages) {
1655 * shrink_slab() does not currently allow us to determine how
1656 * many pages were freed in this zone. So we take the current
1657 * number of slab pages and shake the slab until it is reduced
1658 * by the same nr_pages that we used for reclaiming unmapped
1659 * pages.
1661 * Note that shrink_slab will free memory on all zones and may
1662 * take a long time.
1664 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
1665 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1666 slab_reclaimable - nr_pages)
1670 * Update nr_reclaimed by the number of slab pages we
1671 * reclaimed from this zone.
1673 nr_reclaimed += slab_reclaimable -
1674 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1677 p->reclaim_state = NULL;
1678 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1679 return nr_reclaimed >= nr_pages;
1682 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1684 cpumask_t mask;
1685 int node_id;
1688 * Zone reclaim reclaims unmapped file backed pages and
1689 * slab pages if we are over the defined limits.
1691 * A small portion of unmapped file backed pages is needed for
1692 * file I/O otherwise pages read by file I/O will be immediately
1693 * thrown out if the zone is overallocated. So we do not reclaim
1694 * if less than a specified percentage of the zone is used by
1695 * unmapped file backed pages.
1697 if (zone_page_state(zone, NR_FILE_PAGES) -
1698 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1699 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1700 <= zone->min_slab_pages)
1701 return 0;
1704 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1705 * not have reclaimable pages and if we should not delay the allocation
1706 * then do not scan.
1708 if (!(gfp_mask & __GFP_WAIT) ||
1709 zone->all_unreclaimable ||
1710 atomic_read(&zone->reclaim_in_progress) > 0 ||
1711 (current->flags & PF_MEMALLOC))
1712 return 0;
1715 * Only run zone reclaim on the local zone or on zones that do not
1716 * have associated processors. This will favor the local processor
1717 * over remote processors and spread off node memory allocations
1718 * as wide as possible.
1720 node_id = zone_to_nid(zone);
1721 mask = node_to_cpumask(node_id);
1722 if (!cpus_empty(mask) && node_id != numa_node_id())
1723 return 0;
1724 return __zone_reclaim(zone, gfp_mask, order);
1726 #endif