kbuild: improve scripts/gcc-version.sh output a bit when called without args
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / vmscan.c
bloba6e65d024995ad49a1e9675dea9d36b343ceda60
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;
70 int order;
73 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
75 #ifdef ARCH_HAS_PREFETCH
76 #define prefetch_prev_lru_page(_page, _base, _field) \
77 do { \
78 if ((_page)->lru.prev != _base) { \
79 struct page *prev; \
81 prev = lru_to_page(&(_page->lru)); \
82 prefetch(&prev->_field); \
83 } \
84 } while (0)
85 #else
86 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
87 #endif
89 #ifdef ARCH_HAS_PREFETCHW
90 #define prefetchw_prev_lru_page(_page, _base, _field) \
91 do { \
92 if ((_page)->lru.prev != _base) { \
93 struct page *prev; \
95 prev = lru_to_page(&(_page->lru)); \
96 prefetchw(&prev->_field); \
97 } \
98 } while (0)
99 #else
100 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
101 #endif
104 * From 0 .. 100. Higher means more swappy.
106 int vm_swappiness = 60;
107 long vm_total_pages; /* The total number of pages which the VM controls */
109 static LIST_HEAD(shrinker_list);
110 static DECLARE_RWSEM(shrinker_rwsem);
113 * Add a shrinker callback to be called from the vm
115 void register_shrinker(struct shrinker *shrinker)
117 shrinker->nr = 0;
118 down_write(&shrinker_rwsem);
119 list_add_tail(&shrinker->list, &shrinker_list);
120 up_write(&shrinker_rwsem);
122 EXPORT_SYMBOL(register_shrinker);
125 * Remove one
127 void unregister_shrinker(struct shrinker *shrinker)
129 down_write(&shrinker_rwsem);
130 list_del(&shrinker->list);
131 up_write(&shrinker_rwsem);
133 EXPORT_SYMBOL(unregister_shrinker);
135 #define SHRINK_BATCH 128
137 * Call the shrink functions to age shrinkable caches
139 * Here we assume it costs one seek to replace a lru page and that it also
140 * takes a seek to recreate a cache object. With this in mind we age equal
141 * percentages of the lru and ageable caches. This should balance the seeks
142 * generated by these structures.
144 * If the vm encounted mapped pages on the LRU it increase the pressure on
145 * slab to avoid swapping.
147 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
149 * `lru_pages' represents the number of on-LRU pages in all the zones which
150 * are eligible for the caller's allocation attempt. It is used for balancing
151 * slab reclaim versus page reclaim.
153 * Returns the number of slab objects which we shrunk.
155 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
156 unsigned long lru_pages)
158 struct shrinker *shrinker;
159 unsigned long ret = 0;
161 if (scanned == 0)
162 scanned = SWAP_CLUSTER_MAX;
164 if (!down_read_trylock(&shrinker_rwsem))
165 return 1; /* Assume we'll be able to shrink next time */
167 list_for_each_entry(shrinker, &shrinker_list, list) {
168 unsigned long long delta;
169 unsigned long total_scan;
170 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
172 delta = (4 * scanned) / shrinker->seeks;
173 delta *= max_pass;
174 do_div(delta, lru_pages + 1);
175 shrinker->nr += delta;
176 if (shrinker->nr < 0) {
177 printk(KERN_ERR "%s: nr=%ld\n",
178 __FUNCTION__, shrinker->nr);
179 shrinker->nr = max_pass;
183 * Avoid risking looping forever due to too large nr value:
184 * never try to free more than twice the estimate number of
185 * freeable entries.
187 if (shrinker->nr > max_pass * 2)
188 shrinker->nr = max_pass * 2;
190 total_scan = shrinker->nr;
191 shrinker->nr = 0;
193 while (total_scan >= SHRINK_BATCH) {
194 long this_scan = SHRINK_BATCH;
195 int shrink_ret;
196 int nr_before;
198 nr_before = (*shrinker->shrink)(0, gfp_mask);
199 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
200 if (shrink_ret == -1)
201 break;
202 if (shrink_ret < nr_before)
203 ret += nr_before - shrink_ret;
204 count_vm_events(SLABS_SCANNED, this_scan);
205 total_scan -= this_scan;
207 cond_resched();
210 shrinker->nr += total_scan;
212 up_read(&shrinker_rwsem);
213 return ret;
216 /* Called without lock on whether page is mapped, so answer is unstable */
217 static inline int page_mapping_inuse(struct page *page)
219 struct address_space *mapping;
221 /* Page is in somebody's page tables. */
222 if (page_mapped(page))
223 return 1;
225 /* Be more reluctant to reclaim swapcache than pagecache */
226 if (PageSwapCache(page))
227 return 1;
229 mapping = page_mapping(page);
230 if (!mapping)
231 return 0;
233 /* File is mmap'd by somebody? */
234 return mapping_mapped(mapping);
237 static inline int is_page_cache_freeable(struct page *page)
239 return page_count(page) - !!PagePrivate(page) == 2;
242 static int may_write_to_queue(struct backing_dev_info *bdi)
244 if (current->flags & PF_SWAPWRITE)
245 return 1;
246 if (!bdi_write_congested(bdi))
247 return 1;
248 if (bdi == current->backing_dev_info)
249 return 1;
250 return 0;
254 * We detected a synchronous write error writing a page out. Probably
255 * -ENOSPC. We need to propagate that into the address_space for a subsequent
256 * fsync(), msync() or close().
258 * The tricky part is that after writepage we cannot touch the mapping: nothing
259 * prevents it from being freed up. But we have a ref on the page and once
260 * that page is locked, the mapping is pinned.
262 * We're allowed to run sleeping lock_page() here because we know the caller has
263 * __GFP_FS.
265 static void handle_write_error(struct address_space *mapping,
266 struct page *page, int error)
268 lock_page(page);
269 if (page_mapping(page) == mapping)
270 mapping_set_error(mapping, error);
271 unlock_page(page);
274 /* Request for sync pageout. */
275 enum pageout_io {
276 PAGEOUT_IO_ASYNC,
277 PAGEOUT_IO_SYNC,
280 /* possible outcome of pageout() */
281 typedef enum {
282 /* failed to write page out, page is locked */
283 PAGE_KEEP,
284 /* move page to the active list, page is locked */
285 PAGE_ACTIVATE,
286 /* page has been sent to the disk successfully, page is unlocked */
287 PAGE_SUCCESS,
288 /* page is clean and locked */
289 PAGE_CLEAN,
290 } pageout_t;
293 * pageout is called by shrink_page_list() for each dirty page.
294 * Calls ->writepage().
296 static pageout_t pageout(struct page *page, struct address_space *mapping,
297 enum pageout_io sync_writeback)
300 * If the page is dirty, only perform writeback if that write
301 * will be non-blocking. To prevent this allocation from being
302 * stalled by pagecache activity. But note that there may be
303 * stalls if we need to run get_block(). We could test
304 * PagePrivate for that.
306 * If this process is currently in generic_file_write() against
307 * this page's queue, we can perform writeback even if that
308 * will block.
310 * If the page is swapcache, write it back even if that would
311 * block, for some throttling. This happens by accident, because
312 * swap_backing_dev_info is bust: it doesn't reflect the
313 * congestion state of the swapdevs. Easy to fix, if needed.
314 * See swapfile.c:page_queue_congested().
316 if (!is_page_cache_freeable(page))
317 return PAGE_KEEP;
318 if (!mapping) {
320 * Some data journaling orphaned pages can have
321 * page->mapping == NULL while being dirty with clean buffers.
323 if (PagePrivate(page)) {
324 if (try_to_free_buffers(page)) {
325 ClearPageDirty(page);
326 printk("%s: orphaned page\n", __FUNCTION__);
327 return PAGE_CLEAN;
330 return PAGE_KEEP;
332 if (mapping->a_ops->writepage == NULL)
333 return PAGE_ACTIVATE;
334 if (!may_write_to_queue(mapping->backing_dev_info))
335 return PAGE_KEEP;
337 if (clear_page_dirty_for_io(page)) {
338 int res;
339 struct writeback_control wbc = {
340 .sync_mode = WB_SYNC_NONE,
341 .nr_to_write = SWAP_CLUSTER_MAX,
342 .range_start = 0,
343 .range_end = LLONG_MAX,
344 .nonblocking = 1,
345 .for_reclaim = 1,
348 SetPageReclaim(page);
349 res = mapping->a_ops->writepage(page, &wbc);
350 if (res < 0)
351 handle_write_error(mapping, page, res);
352 if (res == AOP_WRITEPAGE_ACTIVATE) {
353 ClearPageReclaim(page);
354 return PAGE_ACTIVATE;
358 * Wait on writeback if requested to. This happens when
359 * direct reclaiming a large contiguous area and the
360 * first attempt to free a range of pages fails.
362 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
363 wait_on_page_writeback(page);
365 if (!PageWriteback(page)) {
366 /* synchronous write or broken a_ops? */
367 ClearPageReclaim(page);
369 inc_zone_page_state(page, NR_VMSCAN_WRITE);
370 return PAGE_SUCCESS;
373 return PAGE_CLEAN;
377 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
378 * someone else has a ref on the page, abort and return 0. If it was
379 * successfully detached, return 1. Assumes the caller has a single ref on
380 * this page.
382 int remove_mapping(struct address_space *mapping, struct page *page)
384 BUG_ON(!PageLocked(page));
385 BUG_ON(mapping != page_mapping(page));
387 write_lock_irq(&mapping->tree_lock);
389 * The non racy check for a busy page.
391 * Must be careful with the order of the tests. When someone has
392 * a ref to the page, it may be possible that they dirty it then
393 * drop the reference. So if PageDirty is tested before page_count
394 * here, then the following race may occur:
396 * get_user_pages(&page);
397 * [user mapping goes away]
398 * write_to(page);
399 * !PageDirty(page) [good]
400 * SetPageDirty(page);
401 * put_page(page);
402 * !page_count(page) [good, discard it]
404 * [oops, our write_to data is lost]
406 * Reversing the order of the tests ensures such a situation cannot
407 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
408 * load is not satisfied before that of page->_count.
410 * Note that if SetPageDirty is always performed via set_page_dirty,
411 * and thus under tree_lock, then this ordering is not required.
413 if (unlikely(page_count(page) != 2))
414 goto cannot_free;
415 smp_rmb();
416 if (unlikely(PageDirty(page)))
417 goto cannot_free;
419 if (PageSwapCache(page)) {
420 swp_entry_t swap = { .val = page_private(page) };
421 __delete_from_swap_cache(page);
422 write_unlock_irq(&mapping->tree_lock);
423 swap_free(swap);
424 __put_page(page); /* The pagecache ref */
425 return 1;
428 __remove_from_page_cache(page);
429 write_unlock_irq(&mapping->tree_lock);
430 __put_page(page);
431 return 1;
433 cannot_free:
434 write_unlock_irq(&mapping->tree_lock);
435 return 0;
439 * shrink_page_list() returns the number of reclaimed pages
441 static unsigned long shrink_page_list(struct list_head *page_list,
442 struct scan_control *sc,
443 enum pageout_io sync_writeback)
445 LIST_HEAD(ret_pages);
446 struct pagevec freed_pvec;
447 int pgactivate = 0;
448 unsigned long nr_reclaimed = 0;
450 cond_resched();
452 pagevec_init(&freed_pvec, 1);
453 while (!list_empty(page_list)) {
454 struct address_space *mapping;
455 struct page *page;
456 int may_enter_fs;
457 int referenced;
459 cond_resched();
461 page = lru_to_page(page_list);
462 list_del(&page->lru);
464 if (TestSetPageLocked(page))
465 goto keep;
467 VM_BUG_ON(PageActive(page));
469 sc->nr_scanned++;
471 if (!sc->may_swap && page_mapped(page))
472 goto keep_locked;
474 /* Double the slab pressure for mapped and swapcache pages */
475 if (page_mapped(page) || PageSwapCache(page))
476 sc->nr_scanned++;
478 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
479 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
481 if (PageWriteback(page)) {
483 * Synchronous reclaim is performed in two passes,
484 * first an asynchronous pass over the list to
485 * start parallel writeback, and a second synchronous
486 * pass to wait for the IO to complete. Wait here
487 * for any page for which writeback has already
488 * started.
490 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
491 wait_on_page_writeback(page);
492 else
493 goto keep_locked;
496 referenced = page_referenced(page, 1);
497 /* In active use or really unfreeable? Activate it. */
498 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
499 referenced && page_mapping_inuse(page))
500 goto activate_locked;
502 #ifdef CONFIG_SWAP
504 * Anonymous process memory has backing store?
505 * Try to allocate it some swap space here.
507 if (PageAnon(page) && !PageSwapCache(page))
508 if (!add_to_swap(page, GFP_ATOMIC))
509 goto activate_locked;
510 #endif /* CONFIG_SWAP */
512 mapping = page_mapping(page);
515 * The page is mapped into the page tables of one or more
516 * processes. Try to unmap it here.
518 if (page_mapped(page) && mapping) {
519 switch (try_to_unmap(page, 0)) {
520 case SWAP_FAIL:
521 goto activate_locked;
522 case SWAP_AGAIN:
523 goto keep_locked;
524 case SWAP_SUCCESS:
525 ; /* try to free the page below */
529 if (PageDirty(page)) {
530 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
531 goto keep_locked;
532 if (!may_enter_fs)
533 goto keep_locked;
534 if (!sc->may_writepage)
535 goto keep_locked;
537 /* Page is dirty, try to write it out here */
538 switch (pageout(page, mapping, sync_writeback)) {
539 case PAGE_KEEP:
540 goto keep_locked;
541 case PAGE_ACTIVATE:
542 goto activate_locked;
543 case PAGE_SUCCESS:
544 if (PageWriteback(page) || PageDirty(page))
545 goto keep;
547 * A synchronous write - probably a ramdisk. Go
548 * ahead and try to reclaim the page.
550 if (TestSetPageLocked(page))
551 goto keep;
552 if (PageDirty(page) || PageWriteback(page))
553 goto keep_locked;
554 mapping = page_mapping(page);
555 case PAGE_CLEAN:
556 ; /* try to free the page below */
561 * If the page has buffers, try to free the buffer mappings
562 * associated with this page. If we succeed we try to free
563 * the page as well.
565 * We do this even if the page is PageDirty().
566 * try_to_release_page() does not perform I/O, but it is
567 * possible for a page to have PageDirty set, but it is actually
568 * clean (all its buffers are clean). This happens if the
569 * buffers were written out directly, with submit_bh(). ext3
570 * will do this, as well as the blockdev mapping.
571 * try_to_release_page() will discover that cleanness and will
572 * drop the buffers and mark the page clean - it can be freed.
574 * Rarely, pages can have buffers and no ->mapping. These are
575 * the pages which were not successfully invalidated in
576 * truncate_complete_page(). We try to drop those buffers here
577 * and if that worked, and the page is no longer mapped into
578 * process address space (page_count == 1) it can be freed.
579 * Otherwise, leave the page on the LRU so it is swappable.
581 if (PagePrivate(page)) {
582 if (!try_to_release_page(page, sc->gfp_mask))
583 goto activate_locked;
584 if (!mapping && page_count(page) == 1)
585 goto free_it;
588 if (!mapping || !remove_mapping(mapping, page))
589 goto keep_locked;
591 free_it:
592 unlock_page(page);
593 nr_reclaimed++;
594 if (!pagevec_add(&freed_pvec, page))
595 __pagevec_release_nonlru(&freed_pvec);
596 continue;
598 activate_locked:
599 SetPageActive(page);
600 pgactivate++;
601 keep_locked:
602 unlock_page(page);
603 keep:
604 list_add(&page->lru, &ret_pages);
605 VM_BUG_ON(PageLRU(page));
607 list_splice(&ret_pages, page_list);
608 if (pagevec_count(&freed_pvec))
609 __pagevec_release_nonlru(&freed_pvec);
610 count_vm_events(PGACTIVATE, pgactivate);
611 return nr_reclaimed;
614 /* LRU Isolation modes. */
615 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
616 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
617 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
620 * Attempt to remove the specified page from its LRU. Only take this page
621 * if it is of the appropriate PageActive status. Pages which are being
622 * freed elsewhere are also ignored.
624 * page: page to consider
625 * mode: one of the LRU isolation modes defined above
627 * returns 0 on success, -ve errno on failure.
629 static int __isolate_lru_page(struct page *page, int mode)
631 int ret = -EINVAL;
633 /* Only take pages on the LRU. */
634 if (!PageLRU(page))
635 return ret;
638 * When checking the active state, we need to be sure we are
639 * dealing with comparible boolean values. Take the logical not
640 * of each.
642 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
643 return ret;
645 ret = -EBUSY;
646 if (likely(get_page_unless_zero(page))) {
648 * Be careful not to clear PageLRU until after we're
649 * sure the page is not being freed elsewhere -- the
650 * page release code relies on it.
652 ClearPageLRU(page);
653 ret = 0;
656 return ret;
660 * zone->lru_lock is heavily contended. Some of the functions that
661 * shrink the lists perform better by taking out a batch of pages
662 * and working on them outside the LRU lock.
664 * For pagecache intensive workloads, this function is the hottest
665 * spot in the kernel (apart from copy_*_user functions).
667 * Appropriate locks must be held before calling this function.
669 * @nr_to_scan: The number of pages to look through on the list.
670 * @src: The LRU list to pull pages off.
671 * @dst: The temp list to put pages on to.
672 * @scanned: The number of pages that were scanned.
673 * @order: The caller's attempted allocation order
674 * @mode: One of the LRU isolation modes
676 * returns how many pages were moved onto *@dst.
678 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
679 struct list_head *src, struct list_head *dst,
680 unsigned long *scanned, int order, int mode)
682 unsigned long nr_taken = 0;
683 unsigned long scan;
685 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
686 struct page *page;
687 unsigned long pfn;
688 unsigned long end_pfn;
689 unsigned long page_pfn;
690 int zone_id;
692 page = lru_to_page(src);
693 prefetchw_prev_lru_page(page, src, flags);
695 VM_BUG_ON(!PageLRU(page));
697 switch (__isolate_lru_page(page, mode)) {
698 case 0:
699 list_move(&page->lru, dst);
700 nr_taken++;
701 break;
703 case -EBUSY:
704 /* else it is being freed elsewhere */
705 list_move(&page->lru, src);
706 continue;
708 default:
709 BUG();
712 if (!order)
713 continue;
716 * Attempt to take all pages in the order aligned region
717 * surrounding the tag page. Only take those pages of
718 * the same active state as that tag page. We may safely
719 * round the target page pfn down to the requested order
720 * as the mem_map is guarenteed valid out to MAX_ORDER,
721 * where that page is in a different zone we will detect
722 * it from its zone id and abort this block scan.
724 zone_id = page_zone_id(page);
725 page_pfn = page_to_pfn(page);
726 pfn = page_pfn & ~((1 << order) - 1);
727 end_pfn = pfn + (1 << order);
728 for (; pfn < end_pfn; pfn++) {
729 struct page *cursor_page;
731 /* The target page is in the block, ignore it. */
732 if (unlikely(pfn == page_pfn))
733 continue;
735 /* Avoid holes within the zone. */
736 if (unlikely(!pfn_valid_within(pfn)))
737 break;
739 cursor_page = pfn_to_page(pfn);
740 /* Check that we have not crossed a zone boundary. */
741 if (unlikely(page_zone_id(cursor_page) != zone_id))
742 continue;
743 switch (__isolate_lru_page(cursor_page, mode)) {
744 case 0:
745 list_move(&cursor_page->lru, dst);
746 nr_taken++;
747 scan++;
748 break;
750 case -EBUSY:
751 /* else it is being freed elsewhere */
752 list_move(&cursor_page->lru, src);
753 default:
754 break;
759 *scanned = scan;
760 return nr_taken;
764 * clear_active_flags() is a helper for shrink_active_list(), clearing
765 * any active bits from the pages in the list.
767 static unsigned long clear_active_flags(struct list_head *page_list)
769 int nr_active = 0;
770 struct page *page;
772 list_for_each_entry(page, page_list, lru)
773 if (PageActive(page)) {
774 ClearPageActive(page);
775 nr_active++;
778 return nr_active;
782 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
783 * of reclaimed pages
785 static unsigned long shrink_inactive_list(unsigned long max_scan,
786 struct zone *zone, struct scan_control *sc)
788 LIST_HEAD(page_list);
789 struct pagevec pvec;
790 unsigned long nr_scanned = 0;
791 unsigned long nr_reclaimed = 0;
793 pagevec_init(&pvec, 1);
795 lru_add_drain();
796 spin_lock_irq(&zone->lru_lock);
797 do {
798 struct page *page;
799 unsigned long nr_taken;
800 unsigned long nr_scan;
801 unsigned long nr_freed;
802 unsigned long nr_active;
804 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
805 &zone->inactive_list,
806 &page_list, &nr_scan, sc->order,
807 (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
808 ISOLATE_BOTH : ISOLATE_INACTIVE);
809 nr_active = clear_active_flags(&page_list);
810 __count_vm_events(PGDEACTIVATE, nr_active);
812 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
813 __mod_zone_page_state(zone, NR_INACTIVE,
814 -(nr_taken - nr_active));
815 zone->pages_scanned += nr_scan;
816 spin_unlock_irq(&zone->lru_lock);
818 nr_scanned += nr_scan;
819 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
822 * If we are direct reclaiming for contiguous pages and we do
823 * not reclaim everything in the list, try again and wait
824 * for IO to complete. This will stall high-order allocations
825 * but that should be acceptable to the caller
827 if (nr_freed < nr_taken && !current_is_kswapd() &&
828 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
829 congestion_wait(WRITE, HZ/10);
832 * The attempt at page out may have made some
833 * of the pages active, mark them inactive again.
835 nr_active = clear_active_flags(&page_list);
836 count_vm_events(PGDEACTIVATE, nr_active);
838 nr_freed += shrink_page_list(&page_list, sc,
839 PAGEOUT_IO_SYNC);
842 nr_reclaimed += nr_freed;
843 local_irq_disable();
844 if (current_is_kswapd()) {
845 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
846 __count_vm_events(KSWAPD_STEAL, nr_freed);
847 } else
848 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
849 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
851 if (nr_taken == 0)
852 goto done;
854 spin_lock(&zone->lru_lock);
856 * Put back any unfreeable pages.
858 while (!list_empty(&page_list)) {
859 page = lru_to_page(&page_list);
860 VM_BUG_ON(PageLRU(page));
861 SetPageLRU(page);
862 list_del(&page->lru);
863 if (PageActive(page))
864 add_page_to_active_list(zone, page);
865 else
866 add_page_to_inactive_list(zone, page);
867 if (!pagevec_add(&pvec, page)) {
868 spin_unlock_irq(&zone->lru_lock);
869 __pagevec_release(&pvec);
870 spin_lock_irq(&zone->lru_lock);
873 } while (nr_scanned < max_scan);
874 spin_unlock(&zone->lru_lock);
875 done:
876 local_irq_enable();
877 pagevec_release(&pvec);
878 return nr_reclaimed;
882 * We are about to scan this zone at a certain priority level. If that priority
883 * level is smaller (ie: more urgent) than the previous priority, then note
884 * that priority level within the zone. This is done so that when the next
885 * process comes in to scan this zone, it will immediately start out at this
886 * priority level rather than having to build up its own scanning priority.
887 * Here, this priority affects only the reclaim-mapped threshold.
889 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
891 if (priority < zone->prev_priority)
892 zone->prev_priority = priority;
895 static inline int zone_is_near_oom(struct zone *zone)
897 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
898 + zone_page_state(zone, NR_INACTIVE))*3;
902 * This moves pages from the active list to the inactive list.
904 * We move them the other way if the page is referenced by one or more
905 * processes, from rmap.
907 * If the pages are mostly unmapped, the processing is fast and it is
908 * appropriate to hold zone->lru_lock across the whole operation. But if
909 * the pages are mapped, the processing is slow (page_referenced()) so we
910 * should drop zone->lru_lock around each page. It's impossible to balance
911 * this, so instead we remove the pages from the LRU while processing them.
912 * It is safe to rely on PG_active against the non-LRU pages in here because
913 * nobody will play with that bit on a non-LRU page.
915 * The downside is that we have to touch page->_count against each page.
916 * But we had to alter page->flags anyway.
918 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
919 struct scan_control *sc, int priority)
921 unsigned long pgmoved;
922 int pgdeactivate = 0;
923 unsigned long pgscanned;
924 LIST_HEAD(l_hold); /* The pages which were snipped off */
925 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
926 LIST_HEAD(l_active); /* Pages to go onto the active_list */
927 struct page *page;
928 struct pagevec pvec;
929 int reclaim_mapped = 0;
931 if (sc->may_swap) {
932 long mapped_ratio;
933 long distress;
934 long swap_tendency;
936 if (zone_is_near_oom(zone))
937 goto force_reclaim_mapped;
940 * `distress' is a measure of how much trouble we're having
941 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
943 distress = 100 >> min(zone->prev_priority, priority);
946 * The point of this algorithm is to decide when to start
947 * reclaiming mapped memory instead of just pagecache. Work out
948 * how much memory
949 * is mapped.
951 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
952 global_page_state(NR_ANON_PAGES)) * 100) /
953 vm_total_pages;
956 * Now decide how much we really want to unmap some pages. The
957 * mapped ratio is downgraded - just because there's a lot of
958 * mapped memory doesn't necessarily mean that page reclaim
959 * isn't succeeding.
961 * The distress ratio is important - we don't want to start
962 * going oom.
964 * A 100% value of vm_swappiness overrides this algorithm
965 * altogether.
967 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
970 * Now use this metric to decide whether to start moving mapped
971 * memory onto the inactive list.
973 if (swap_tendency >= 100)
974 force_reclaim_mapped:
975 reclaim_mapped = 1;
978 lru_add_drain();
979 spin_lock_irq(&zone->lru_lock);
980 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
981 &l_hold, &pgscanned, sc->order, ISOLATE_ACTIVE);
982 zone->pages_scanned += pgscanned;
983 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
984 spin_unlock_irq(&zone->lru_lock);
986 while (!list_empty(&l_hold)) {
987 cond_resched();
988 page = lru_to_page(&l_hold);
989 list_del(&page->lru);
990 if (page_mapped(page)) {
991 if (!reclaim_mapped ||
992 (total_swap_pages == 0 && PageAnon(page)) ||
993 page_referenced(page, 0)) {
994 list_add(&page->lru, &l_active);
995 continue;
998 list_add(&page->lru, &l_inactive);
1001 pagevec_init(&pvec, 1);
1002 pgmoved = 0;
1003 spin_lock_irq(&zone->lru_lock);
1004 while (!list_empty(&l_inactive)) {
1005 page = lru_to_page(&l_inactive);
1006 prefetchw_prev_lru_page(page, &l_inactive, flags);
1007 VM_BUG_ON(PageLRU(page));
1008 SetPageLRU(page);
1009 VM_BUG_ON(!PageActive(page));
1010 ClearPageActive(page);
1012 list_move(&page->lru, &zone->inactive_list);
1013 pgmoved++;
1014 if (!pagevec_add(&pvec, page)) {
1015 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1016 spin_unlock_irq(&zone->lru_lock);
1017 pgdeactivate += pgmoved;
1018 pgmoved = 0;
1019 if (buffer_heads_over_limit)
1020 pagevec_strip(&pvec);
1021 __pagevec_release(&pvec);
1022 spin_lock_irq(&zone->lru_lock);
1025 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1026 pgdeactivate += pgmoved;
1027 if (buffer_heads_over_limit) {
1028 spin_unlock_irq(&zone->lru_lock);
1029 pagevec_strip(&pvec);
1030 spin_lock_irq(&zone->lru_lock);
1033 pgmoved = 0;
1034 while (!list_empty(&l_active)) {
1035 page = lru_to_page(&l_active);
1036 prefetchw_prev_lru_page(page, &l_active, flags);
1037 VM_BUG_ON(PageLRU(page));
1038 SetPageLRU(page);
1039 VM_BUG_ON(!PageActive(page));
1040 list_move(&page->lru, &zone->active_list);
1041 pgmoved++;
1042 if (!pagevec_add(&pvec, page)) {
1043 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1044 pgmoved = 0;
1045 spin_unlock_irq(&zone->lru_lock);
1046 __pagevec_release(&pvec);
1047 spin_lock_irq(&zone->lru_lock);
1050 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1052 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1053 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1054 spin_unlock_irq(&zone->lru_lock);
1056 pagevec_release(&pvec);
1060 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1062 static unsigned long shrink_zone(int priority, struct zone *zone,
1063 struct scan_control *sc)
1065 unsigned long nr_active;
1066 unsigned long nr_inactive;
1067 unsigned long nr_to_scan;
1068 unsigned long nr_reclaimed = 0;
1070 atomic_inc(&zone->reclaim_in_progress);
1073 * Add one to `nr_to_scan' just to make sure that the kernel will
1074 * slowly sift through the active list.
1076 zone->nr_scan_active +=
1077 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1078 nr_active = zone->nr_scan_active;
1079 if (nr_active >= sc->swap_cluster_max)
1080 zone->nr_scan_active = 0;
1081 else
1082 nr_active = 0;
1084 zone->nr_scan_inactive +=
1085 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1086 nr_inactive = zone->nr_scan_inactive;
1087 if (nr_inactive >= sc->swap_cluster_max)
1088 zone->nr_scan_inactive = 0;
1089 else
1090 nr_inactive = 0;
1092 while (nr_active || nr_inactive) {
1093 if (nr_active) {
1094 nr_to_scan = min(nr_active,
1095 (unsigned long)sc->swap_cluster_max);
1096 nr_active -= nr_to_scan;
1097 shrink_active_list(nr_to_scan, zone, sc, priority);
1100 if (nr_inactive) {
1101 nr_to_scan = min(nr_inactive,
1102 (unsigned long)sc->swap_cluster_max);
1103 nr_inactive -= nr_to_scan;
1104 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1105 sc);
1109 throttle_vm_writeout(sc->gfp_mask);
1111 atomic_dec(&zone->reclaim_in_progress);
1112 return nr_reclaimed;
1116 * This is the direct reclaim path, for page-allocating processes. We only
1117 * try to reclaim pages from zones which will satisfy the caller's allocation
1118 * request.
1120 * We reclaim from a zone even if that zone is over pages_high. Because:
1121 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1122 * allocation or
1123 * b) The zones may be over pages_high but they must go *over* pages_high to
1124 * satisfy the `incremental min' zone defense algorithm.
1126 * Returns the number of reclaimed pages.
1128 * If a zone is deemed to be full of pinned pages then just give it a light
1129 * scan then give up on it.
1131 static unsigned long shrink_zones(int priority, struct zone **zones,
1132 struct scan_control *sc)
1134 unsigned long nr_reclaimed = 0;
1135 int i;
1137 sc->all_unreclaimable = 1;
1138 for (i = 0; zones[i] != NULL; i++) {
1139 struct zone *zone = zones[i];
1141 if (!populated_zone(zone))
1142 continue;
1144 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1145 continue;
1147 note_zone_scanning_priority(zone, priority);
1149 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1150 continue; /* Let kswapd poll it */
1152 sc->all_unreclaimable = 0;
1154 nr_reclaimed += shrink_zone(priority, zone, sc);
1156 return nr_reclaimed;
1160 * This is the main entry point to direct page reclaim.
1162 * If a full scan of the inactive list fails to free enough memory then we
1163 * are "out of memory" and something needs to be killed.
1165 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1166 * high - the zone may be full of dirty or under-writeback pages, which this
1167 * caller can't do much about. We kick pdflush and take explicit naps in the
1168 * hope that some of these pages can be written. But if the allocating task
1169 * holds filesystem locks which prevent writeout this might not work, and the
1170 * allocation attempt will fail.
1172 unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1174 int priority;
1175 int ret = 0;
1176 unsigned long total_scanned = 0;
1177 unsigned long nr_reclaimed = 0;
1178 struct reclaim_state *reclaim_state = current->reclaim_state;
1179 unsigned long lru_pages = 0;
1180 int i;
1181 struct scan_control sc = {
1182 .gfp_mask = gfp_mask,
1183 .may_writepage = !laptop_mode,
1184 .swap_cluster_max = SWAP_CLUSTER_MAX,
1185 .may_swap = 1,
1186 .swappiness = vm_swappiness,
1187 .order = order,
1190 count_vm_event(ALLOCSTALL);
1192 for (i = 0; zones[i] != NULL; i++) {
1193 struct zone *zone = zones[i];
1195 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1196 continue;
1198 lru_pages += zone_page_state(zone, NR_ACTIVE)
1199 + zone_page_state(zone, NR_INACTIVE);
1202 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1203 sc.nr_scanned = 0;
1204 if (!priority)
1205 disable_swap_token();
1206 nr_reclaimed += shrink_zones(priority, zones, &sc);
1207 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1208 if (reclaim_state) {
1209 nr_reclaimed += reclaim_state->reclaimed_slab;
1210 reclaim_state->reclaimed_slab = 0;
1212 total_scanned += sc.nr_scanned;
1213 if (nr_reclaimed >= sc.swap_cluster_max) {
1214 ret = 1;
1215 goto out;
1219 * Try to write back as many pages as we just scanned. This
1220 * tends to cause slow streaming writers to write data to the
1221 * disk smoothly, at the dirtying rate, which is nice. But
1222 * that's undesirable in laptop mode, where we *want* lumpy
1223 * writeout. So in laptop mode, write out the whole world.
1225 if (total_scanned > sc.swap_cluster_max +
1226 sc.swap_cluster_max / 2) {
1227 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1228 sc.may_writepage = 1;
1231 /* Take a nap, wait for some writeback to complete */
1232 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1233 congestion_wait(WRITE, HZ/10);
1235 /* top priority shrink_caches still had more to do? don't OOM, then */
1236 if (!sc.all_unreclaimable)
1237 ret = 1;
1238 out:
1240 * Now that we've scanned all the zones at this priority level, note
1241 * that level within the zone so that the next thread which performs
1242 * scanning of this zone will immediately start out at this priority
1243 * level. This affects only the decision whether or not to bring
1244 * mapped pages onto the inactive list.
1246 if (priority < 0)
1247 priority = 0;
1248 for (i = 0; zones[i] != 0; i++) {
1249 struct zone *zone = zones[i];
1251 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1252 continue;
1254 zone->prev_priority = priority;
1256 return ret;
1260 * For kswapd, balance_pgdat() will work across all this node's zones until
1261 * they are all at pages_high.
1263 * Returns the number of pages which were actually freed.
1265 * There is special handling here for zones which are full of pinned pages.
1266 * This can happen if the pages are all mlocked, or if they are all used by
1267 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1268 * What we do is to detect the case where all pages in the zone have been
1269 * scanned twice and there has been zero successful reclaim. Mark the zone as
1270 * dead and from now on, only perform a short scan. Basically we're polling
1271 * the zone for when the problem goes away.
1273 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1274 * zones which have free_pages > pages_high, but once a zone is found to have
1275 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1276 * of the number of free pages in the lower zones. This interoperates with
1277 * the page allocator fallback scheme to ensure that aging of pages is balanced
1278 * across the zones.
1280 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1282 int all_zones_ok;
1283 int priority;
1284 int i;
1285 unsigned long total_scanned;
1286 unsigned long nr_reclaimed;
1287 struct reclaim_state *reclaim_state = current->reclaim_state;
1288 struct scan_control sc = {
1289 .gfp_mask = GFP_KERNEL,
1290 .may_swap = 1,
1291 .swap_cluster_max = SWAP_CLUSTER_MAX,
1292 .swappiness = vm_swappiness,
1293 .order = order,
1296 * temp_priority is used to remember the scanning priority at which
1297 * this zone was successfully refilled to free_pages == pages_high.
1299 int temp_priority[MAX_NR_ZONES];
1301 loop_again:
1302 total_scanned = 0;
1303 nr_reclaimed = 0;
1304 sc.may_writepage = !laptop_mode;
1305 count_vm_event(PAGEOUTRUN);
1307 for (i = 0; i < pgdat->nr_zones; i++)
1308 temp_priority[i] = DEF_PRIORITY;
1310 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1311 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1312 unsigned long lru_pages = 0;
1314 /* The swap token gets in the way of swapout... */
1315 if (!priority)
1316 disable_swap_token();
1318 all_zones_ok = 1;
1321 * Scan in the highmem->dma direction for the highest
1322 * zone which needs scanning
1324 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1325 struct zone *zone = pgdat->node_zones + i;
1327 if (!populated_zone(zone))
1328 continue;
1330 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1331 continue;
1333 if (!zone_watermark_ok(zone, order, zone->pages_high,
1334 0, 0)) {
1335 end_zone = i;
1336 break;
1339 if (i < 0)
1340 goto out;
1342 for (i = 0; i <= end_zone; i++) {
1343 struct zone *zone = pgdat->node_zones + i;
1345 lru_pages += zone_page_state(zone, NR_ACTIVE)
1346 + zone_page_state(zone, NR_INACTIVE);
1350 * Now scan the zone in the dma->highmem direction, stopping
1351 * at the last zone which needs scanning.
1353 * We do this because the page allocator works in the opposite
1354 * direction. This prevents the page allocator from allocating
1355 * pages behind kswapd's direction of progress, which would
1356 * cause too much scanning of the lower zones.
1358 for (i = 0; i <= end_zone; i++) {
1359 struct zone *zone = pgdat->node_zones + i;
1360 int nr_slab;
1362 if (!populated_zone(zone))
1363 continue;
1365 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1366 continue;
1368 if (!zone_watermark_ok(zone, order, zone->pages_high,
1369 end_zone, 0))
1370 all_zones_ok = 0;
1371 temp_priority[i] = priority;
1372 sc.nr_scanned = 0;
1373 note_zone_scanning_priority(zone, priority);
1374 nr_reclaimed += shrink_zone(priority, zone, &sc);
1375 reclaim_state->reclaimed_slab = 0;
1376 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1377 lru_pages);
1378 nr_reclaimed += reclaim_state->reclaimed_slab;
1379 total_scanned += sc.nr_scanned;
1380 if (zone->all_unreclaimable)
1381 continue;
1382 if (nr_slab == 0 && zone->pages_scanned >=
1383 (zone_page_state(zone, NR_ACTIVE)
1384 + zone_page_state(zone, NR_INACTIVE)) * 6)
1385 zone->all_unreclaimable = 1;
1387 * If we've done a decent amount of scanning and
1388 * the reclaim ratio is low, start doing writepage
1389 * even in laptop mode
1391 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1392 total_scanned > nr_reclaimed + nr_reclaimed / 2)
1393 sc.may_writepage = 1;
1395 if (all_zones_ok)
1396 break; /* kswapd: all done */
1398 * OK, kswapd is getting into trouble. Take a nap, then take
1399 * another pass across the zones.
1401 if (total_scanned && priority < DEF_PRIORITY - 2)
1402 congestion_wait(WRITE, HZ/10);
1405 * We do this so kswapd doesn't build up large priorities for
1406 * example when it is freeing in parallel with allocators. It
1407 * matches the direct reclaim path behaviour in terms of impact
1408 * on zone->*_priority.
1410 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1411 break;
1413 out:
1415 * Note within each zone the priority level at which this zone was
1416 * brought into a happy state. So that the next thread which scans this
1417 * zone will start out at that priority level.
1419 for (i = 0; i < pgdat->nr_zones; i++) {
1420 struct zone *zone = pgdat->node_zones + i;
1422 zone->prev_priority = temp_priority[i];
1424 if (!all_zones_ok) {
1425 cond_resched();
1427 try_to_freeze();
1429 goto loop_again;
1432 return nr_reclaimed;
1436 * The background pageout daemon, started as a kernel thread
1437 * from the init process.
1439 * This basically trickles out pages so that we have _some_
1440 * free memory available even if there is no other activity
1441 * that frees anything up. This is needed for things like routing
1442 * etc, where we otherwise might have all activity going on in
1443 * asynchronous contexts that cannot page things out.
1445 * If there are applications that are active memory-allocators
1446 * (most normal use), this basically shouldn't matter.
1448 static int kswapd(void *p)
1450 unsigned long order;
1451 pg_data_t *pgdat = (pg_data_t*)p;
1452 struct task_struct *tsk = current;
1453 DEFINE_WAIT(wait);
1454 struct reclaim_state reclaim_state = {
1455 .reclaimed_slab = 0,
1457 cpumask_t cpumask;
1459 cpumask = node_to_cpumask(pgdat->node_id);
1460 if (!cpus_empty(cpumask))
1461 set_cpus_allowed(tsk, cpumask);
1462 current->reclaim_state = &reclaim_state;
1465 * Tell the memory management that we're a "memory allocator",
1466 * and that if we need more memory we should get access to it
1467 * regardless (see "__alloc_pages()"). "kswapd" should
1468 * never get caught in the normal page freeing logic.
1470 * (Kswapd normally doesn't need memory anyway, but sometimes
1471 * you need a small amount of memory in order to be able to
1472 * page out something else, and this flag essentially protects
1473 * us from recursively trying to free more memory as we're
1474 * trying to free the first piece of memory in the first place).
1476 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1477 set_freezable();
1479 order = 0;
1480 for ( ; ; ) {
1481 unsigned long new_order;
1483 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1484 new_order = pgdat->kswapd_max_order;
1485 pgdat->kswapd_max_order = 0;
1486 if (order < new_order) {
1488 * Don't sleep if someone wants a larger 'order'
1489 * allocation
1491 order = new_order;
1492 } else {
1493 if (!freezing(current))
1494 schedule();
1496 order = pgdat->kswapd_max_order;
1498 finish_wait(&pgdat->kswapd_wait, &wait);
1500 if (!try_to_freeze()) {
1501 /* We can speed up thawing tasks if we don't call
1502 * balance_pgdat after returning from the refrigerator
1504 balance_pgdat(pgdat, order);
1507 return 0;
1511 * A zone is low on free memory, so wake its kswapd task to service it.
1513 void wakeup_kswapd(struct zone *zone, int order)
1515 pg_data_t *pgdat;
1517 if (!populated_zone(zone))
1518 return;
1520 pgdat = zone->zone_pgdat;
1521 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1522 return;
1523 if (pgdat->kswapd_max_order < order)
1524 pgdat->kswapd_max_order = order;
1525 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1526 return;
1527 if (!waitqueue_active(&pgdat->kswapd_wait))
1528 return;
1529 wake_up_interruptible(&pgdat->kswapd_wait);
1532 #ifdef CONFIG_PM
1534 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1535 * from LRU lists system-wide, for given pass and priority, and returns the
1536 * number of reclaimed pages
1538 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1540 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1541 int pass, struct scan_control *sc)
1543 struct zone *zone;
1544 unsigned long nr_to_scan, ret = 0;
1546 for_each_zone(zone) {
1548 if (!populated_zone(zone))
1549 continue;
1551 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1552 continue;
1554 /* For pass = 0 we don't shrink the active list */
1555 if (pass > 0) {
1556 zone->nr_scan_active +=
1557 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1558 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1559 zone->nr_scan_active = 0;
1560 nr_to_scan = min(nr_pages,
1561 zone_page_state(zone, NR_ACTIVE));
1562 shrink_active_list(nr_to_scan, zone, sc, prio);
1566 zone->nr_scan_inactive +=
1567 (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1568 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1569 zone->nr_scan_inactive = 0;
1570 nr_to_scan = min(nr_pages,
1571 zone_page_state(zone, NR_INACTIVE));
1572 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1573 if (ret >= nr_pages)
1574 return ret;
1578 return ret;
1581 static unsigned long count_lru_pages(void)
1583 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1587 * Try to free `nr_pages' of memory, system-wide, and return the number of
1588 * freed pages.
1590 * Rather than trying to age LRUs the aim is to preserve the overall
1591 * LRU order by reclaiming preferentially
1592 * inactive > active > active referenced > active mapped
1594 unsigned long shrink_all_memory(unsigned long nr_pages)
1596 unsigned long lru_pages, nr_slab;
1597 unsigned long ret = 0;
1598 int pass;
1599 struct reclaim_state reclaim_state;
1600 struct scan_control sc = {
1601 .gfp_mask = GFP_KERNEL,
1602 .may_swap = 0,
1603 .swap_cluster_max = nr_pages,
1604 .may_writepage = 1,
1605 .swappiness = vm_swappiness,
1608 current->reclaim_state = &reclaim_state;
1610 lru_pages = count_lru_pages();
1611 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1612 /* If slab caches are huge, it's better to hit them first */
1613 while (nr_slab >= lru_pages) {
1614 reclaim_state.reclaimed_slab = 0;
1615 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1616 if (!reclaim_state.reclaimed_slab)
1617 break;
1619 ret += reclaim_state.reclaimed_slab;
1620 if (ret >= nr_pages)
1621 goto out;
1623 nr_slab -= reclaim_state.reclaimed_slab;
1627 * We try to shrink LRUs in 5 passes:
1628 * 0 = Reclaim from inactive_list only
1629 * 1 = Reclaim from active list but don't reclaim mapped
1630 * 2 = 2nd pass of type 1
1631 * 3 = Reclaim mapped (normal reclaim)
1632 * 4 = 2nd pass of type 3
1634 for (pass = 0; pass < 5; pass++) {
1635 int prio;
1637 /* Force reclaiming mapped pages in the passes #3 and #4 */
1638 if (pass > 2) {
1639 sc.may_swap = 1;
1640 sc.swappiness = 100;
1643 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1644 unsigned long nr_to_scan = nr_pages - ret;
1646 sc.nr_scanned = 0;
1647 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1648 if (ret >= nr_pages)
1649 goto out;
1651 reclaim_state.reclaimed_slab = 0;
1652 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1653 count_lru_pages());
1654 ret += reclaim_state.reclaimed_slab;
1655 if (ret >= nr_pages)
1656 goto out;
1658 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1659 congestion_wait(WRITE, HZ / 10);
1664 * If ret = 0, we could not shrink LRUs, but there may be something
1665 * in slab caches
1667 if (!ret) {
1668 do {
1669 reclaim_state.reclaimed_slab = 0;
1670 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1671 ret += reclaim_state.reclaimed_slab;
1672 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1675 out:
1676 current->reclaim_state = NULL;
1678 return ret;
1680 #endif
1682 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1683 not required for correctness. So if the last cpu in a node goes
1684 away, we get changed to run anywhere: as the first one comes back,
1685 restore their cpu bindings. */
1686 static int __devinit cpu_callback(struct notifier_block *nfb,
1687 unsigned long action, void *hcpu)
1689 pg_data_t *pgdat;
1690 cpumask_t mask;
1692 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1693 for_each_online_pgdat(pgdat) {
1694 mask = node_to_cpumask(pgdat->node_id);
1695 if (any_online_cpu(mask) != NR_CPUS)
1696 /* One of our CPUs online: restore mask */
1697 set_cpus_allowed(pgdat->kswapd, mask);
1700 return NOTIFY_OK;
1704 * This kswapd start function will be called by init and node-hot-add.
1705 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1707 int kswapd_run(int nid)
1709 pg_data_t *pgdat = NODE_DATA(nid);
1710 int ret = 0;
1712 if (pgdat->kswapd)
1713 return 0;
1715 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1716 if (IS_ERR(pgdat->kswapd)) {
1717 /* failure at boot is fatal */
1718 BUG_ON(system_state == SYSTEM_BOOTING);
1719 printk("Failed to start kswapd on node %d\n",nid);
1720 ret = -1;
1722 return ret;
1725 static int __init kswapd_init(void)
1727 int nid;
1729 swap_setup();
1730 for_each_online_node(nid)
1731 kswapd_run(nid);
1732 hotcpu_notifier(cpu_callback, 0);
1733 return 0;
1736 module_init(kswapd_init)
1738 #ifdef CONFIG_NUMA
1740 * Zone reclaim mode
1742 * If non-zero call zone_reclaim when the number of free pages falls below
1743 * the watermarks.
1745 int zone_reclaim_mode __read_mostly;
1747 #define RECLAIM_OFF 0
1748 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1749 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1750 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1753 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1754 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1755 * a zone.
1757 #define ZONE_RECLAIM_PRIORITY 4
1760 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1761 * occur.
1763 int sysctl_min_unmapped_ratio = 1;
1766 * If the number of slab pages in a zone grows beyond this percentage then
1767 * slab reclaim needs to occur.
1769 int sysctl_min_slab_ratio = 5;
1772 * Try to free up some pages from this zone through reclaim.
1774 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1776 /* Minimum pages needed in order to stay on node */
1777 const unsigned long nr_pages = 1 << order;
1778 struct task_struct *p = current;
1779 struct reclaim_state reclaim_state;
1780 int priority;
1781 unsigned long nr_reclaimed = 0;
1782 struct scan_control sc = {
1783 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1784 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1785 .swap_cluster_max = max_t(unsigned long, nr_pages,
1786 SWAP_CLUSTER_MAX),
1787 .gfp_mask = gfp_mask,
1788 .swappiness = vm_swappiness,
1790 unsigned long slab_reclaimable;
1792 disable_swap_token();
1793 cond_resched();
1795 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1796 * and we also need to be able to write out pages for RECLAIM_WRITE
1797 * and RECLAIM_SWAP.
1799 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1800 reclaim_state.reclaimed_slab = 0;
1801 p->reclaim_state = &reclaim_state;
1803 if (zone_page_state(zone, NR_FILE_PAGES) -
1804 zone_page_state(zone, NR_FILE_MAPPED) >
1805 zone->min_unmapped_pages) {
1807 * Free memory by calling shrink zone with increasing
1808 * priorities until we have enough memory freed.
1810 priority = ZONE_RECLAIM_PRIORITY;
1811 do {
1812 note_zone_scanning_priority(zone, priority);
1813 nr_reclaimed += shrink_zone(priority, zone, &sc);
1814 priority--;
1815 } while (priority >= 0 && nr_reclaimed < nr_pages);
1818 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1819 if (slab_reclaimable > zone->min_slab_pages) {
1821 * shrink_slab() does not currently allow us to determine how
1822 * many pages were freed in this zone. So we take the current
1823 * number of slab pages and shake the slab until it is reduced
1824 * by the same nr_pages that we used for reclaiming unmapped
1825 * pages.
1827 * Note that shrink_slab will free memory on all zones and may
1828 * take a long time.
1830 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
1831 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1832 slab_reclaimable - nr_pages)
1836 * Update nr_reclaimed by the number of slab pages we
1837 * reclaimed from this zone.
1839 nr_reclaimed += slab_reclaimable -
1840 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1843 p->reclaim_state = NULL;
1844 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1845 return nr_reclaimed >= nr_pages;
1848 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1850 cpumask_t mask;
1851 int node_id;
1854 * Zone reclaim reclaims unmapped file backed pages and
1855 * slab pages if we are over the defined limits.
1857 * A small portion of unmapped file backed pages is needed for
1858 * file I/O otherwise pages read by file I/O will be immediately
1859 * thrown out if the zone is overallocated. So we do not reclaim
1860 * if less than a specified percentage of the zone is used by
1861 * unmapped file backed pages.
1863 if (zone_page_state(zone, NR_FILE_PAGES) -
1864 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1865 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1866 <= zone->min_slab_pages)
1867 return 0;
1870 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1871 * not have reclaimable pages and if we should not delay the allocation
1872 * then do not scan.
1874 if (!(gfp_mask & __GFP_WAIT) ||
1875 zone->all_unreclaimable ||
1876 atomic_read(&zone->reclaim_in_progress) > 0 ||
1877 (current->flags & PF_MEMALLOC))
1878 return 0;
1881 * Only run zone reclaim on the local zone or on zones that do not
1882 * have associated processors. This will favor the local processor
1883 * over remote processors and spread off node memory allocations
1884 * as wide as possible.
1886 node_id = zone_to_nid(zone);
1887 mask = node_to_cpumask(node_id);
1888 if (!cpus_empty(mask) && node_id != numa_node_id())
1889 return 0;
1890 return __zone_reclaim(zone, gfp_mask, order);
1892 #endif