0960846d649f460efdd4391880dd2f41bd230e81
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
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.
12  */
13
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/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
26                                         buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
36 #include <linux/delay.h>
37 #include <linux/kthread.h>
38
39 #include <asm/tlbflush.h>
40 #include <asm/div64.h>
41
42 #include <linux/swapops.h>
43
44 #include "internal.h"
45
46 struct scan_control {
47         /* Incremented by the number of inactive pages that were scanned */
48         unsigned long nr_scanned;
49
50         /* This context's GFP mask */
51         gfp_t gfp_mask;
52
53         int may_writepage;
54
55         /* Can pages be swapped as part of reclaim? */
56         int may_swap;
57
58         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
59          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
60          * In this context, it doesn't matter that we scan the
61          * whole list at once. */
62         int swap_cluster_max;
63
64         int swappiness;
65 };
66
67 /*
68  * The list of shrinker callbacks used by to apply pressure to
69  * ageable caches.
70  */
71 struct shrinker {
72         shrinker_t              shrinker;
73         struct list_head        list;
74         int                     seeks;  /* seeks to recreate an obj */
75         long                    nr;     /* objs pending delete */
76 };
77
78 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
79
80 #ifdef ARCH_HAS_PREFETCH
81 #define prefetch_prev_lru_page(_page, _base, _field)                    \
82         do {                                                            \
83                 if ((_page)->lru.prev != _base) {                       \
84                         struct page *prev;                              \
85                                                                         \
86                         prev = lru_to_page(&(_page->lru));              \
87                         prefetch(&prev->_field);                        \
88                 }                                                       \
89         } while (0)
90 #else
91 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
92 #endif
93
94 #ifdef ARCH_HAS_PREFETCHW
95 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
96         do {                                                            \
97                 if ((_page)->lru.prev != _base) {                       \
98                         struct page *prev;                              \
99                                                                         \
100                         prev = lru_to_page(&(_page->lru));              \
101                         prefetchw(&prev->_field);                       \
102                 }                                                       \
103         } while (0)
104 #else
105 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
106 #endif
107
108 /*
109  * From 0 .. 100.  Higher means more swappy.
110  */
111 int vm_swappiness = 60;
112 long vm_total_pages;    /* The total number of pages which the VM controls */
113
114 static LIST_HEAD(shrinker_list);
115 static DECLARE_RWSEM(shrinker_rwsem);
116
117 /*
118  * Add a shrinker callback to be called from the vm
119  */
120 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
121 {
122         struct shrinker *shrinker;
123
124         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
125         if (shrinker) {
126                 shrinker->shrinker = theshrinker;
127                 shrinker->seeks = seeks;
128                 shrinker->nr = 0;
129                 down_write(&shrinker_rwsem);
130                 list_add_tail(&shrinker->list, &shrinker_list);
131                 up_write(&shrinker_rwsem);
132         }
133         return shrinker;
134 }
135 EXPORT_SYMBOL(set_shrinker);
136
137 /*
138  * Remove one
139  */
140 void remove_shrinker(struct shrinker *shrinker)
141 {
142         down_write(&shrinker_rwsem);
143         list_del(&shrinker->list);
144         up_write(&shrinker_rwsem);
145         kfree(shrinker);
146 }
147 EXPORT_SYMBOL(remove_shrinker);
148
149 #define SHRINK_BATCH 128
150 /*
151  * Call the shrink functions to age shrinkable caches
152  *
153  * Here we assume it costs one seek to replace a lru page and that it also
154  * takes a seek to recreate a cache object.  With this in mind we age equal
155  * percentages of the lru and ageable caches.  This should balance the seeks
156  * generated by these structures.
157  *
158  * If the vm encounted mapped pages on the LRU it increase the pressure on
159  * slab to avoid swapping.
160  *
161  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
162  *
163  * `lru_pages' represents the number of on-LRU pages in all the zones which
164  * are eligible for the caller's allocation attempt.  It is used for balancing
165  * slab reclaim versus page reclaim.
166  *
167  * Returns the number of slab objects which we shrunk.
168  */
169 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
170                         unsigned long lru_pages)
171 {
172         struct shrinker *shrinker;
173         unsigned long ret = 0;
174
175         if (scanned == 0)
176                 scanned = SWAP_CLUSTER_MAX;
177
178         if (!down_read_trylock(&shrinker_rwsem))
179                 return 1;       /* Assume we'll be able to shrink next time */
180
181         list_for_each_entry(shrinker, &shrinker_list, list) {
182                 unsigned long long delta;
183                 unsigned long total_scan;
184                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
185
186                 delta = (4 * scanned) / shrinker->seeks;
187                 delta *= max_pass;
188                 do_div(delta, lru_pages + 1);
189                 shrinker->nr += delta;
190                 if (shrinker->nr < 0) {
191                         printk(KERN_ERR "%s: nr=%ld\n",
192                                         __FUNCTION__, shrinker->nr);
193                         shrinker->nr = max_pass;
194                 }
195
196                 /*
197                  * Avoid risking looping forever due to too large nr value:
198                  * never try to free more than twice the estimate number of
199                  * freeable entries.
200                  */
201                 if (shrinker->nr > max_pass * 2)
202                         shrinker->nr = max_pass * 2;
203
204                 total_scan = shrinker->nr;
205                 shrinker->nr = 0;
206
207                 while (total_scan >= SHRINK_BATCH) {
208                         long this_scan = SHRINK_BATCH;
209                         int shrink_ret;
210                         int nr_before;
211
212                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
213                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
214                         if (shrink_ret == -1)
215                                 break;
216                         if (shrink_ret < nr_before)
217                                 ret += nr_before - shrink_ret;
218                         mod_page_state(slabs_scanned, this_scan);
219                         total_scan -= this_scan;
220
221                         cond_resched();
222                 }
223
224                 shrinker->nr += total_scan;
225         }
226         up_read(&shrinker_rwsem);
227         return ret;
228 }
229
230 /* Called without lock on whether page is mapped, so answer is unstable */
231 static inline int page_mapping_inuse(struct page *page)
232 {
233         struct address_space *mapping;
234
235         /* Page is in somebody's page tables. */
236         if (page_mapped(page))
237                 return 1;
238
239         /* Be more reluctant to reclaim swapcache than pagecache */
240         if (PageSwapCache(page))
241                 return 1;
242
243         mapping = page_mapping(page);
244         if (!mapping)
245                 return 0;
246
247         /* File is mmap'd by somebody? */
248         return mapping_mapped(mapping);
249 }
250
251 static inline int is_page_cache_freeable(struct page *page)
252 {
253         return page_count(page) - !!PagePrivate(page) == 2;
254 }
255
256 static int may_write_to_queue(struct backing_dev_info *bdi)
257 {
258         if (current->flags & PF_SWAPWRITE)
259                 return 1;
260         if (!bdi_write_congested(bdi))
261                 return 1;
262         if (bdi == current->backing_dev_info)
263                 return 1;
264         return 0;
265 }
266
267 /*
268  * We detected a synchronous write error writing a page out.  Probably
269  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
270  * fsync(), msync() or close().
271  *
272  * The tricky part is that after writepage we cannot touch the mapping: nothing
273  * prevents it from being freed up.  But we have a ref on the page and once
274  * that page is locked, the mapping is pinned.
275  *
276  * We're allowed to run sleeping lock_page() here because we know the caller has
277  * __GFP_FS.
278  */
279 static void handle_write_error(struct address_space *mapping,
280                                 struct page *page, int error)
281 {
282         lock_page(page);
283         if (page_mapping(page) == mapping) {
284                 if (error == -ENOSPC)
285                         set_bit(AS_ENOSPC, &mapping->flags);
286                 else
287                         set_bit(AS_EIO, &mapping->flags);
288         }
289         unlock_page(page);
290 }
291
292 /* possible outcome of pageout() */
293 typedef enum {
294         /* failed to write page out, page is locked */
295         PAGE_KEEP,
296         /* move page to the active list, page is locked */
297         PAGE_ACTIVATE,
298         /* page has been sent to the disk successfully, page is unlocked */
299         PAGE_SUCCESS,
300         /* page is clean and locked */
301         PAGE_CLEAN,
302 } pageout_t;
303
304 /*
305  * pageout is called by shrink_page_list() for each dirty page.
306  * Calls ->writepage().
307  */
308 static pageout_t pageout(struct page *page, struct address_space *mapping)
309 {
310         /*
311          * If the page is dirty, only perform writeback if that write
312          * will be non-blocking.  To prevent this allocation from being
313          * stalled by pagecache activity.  But note that there may be
314          * stalls if we need to run get_block().  We could test
315          * PagePrivate for that.
316          *
317          * If this process is currently in generic_file_write() against
318          * this page's queue, we can perform writeback even if that
319          * will block.
320          *
321          * If the page is swapcache, write it back even if that would
322          * block, for some throttling. This happens by accident, because
323          * swap_backing_dev_info is bust: it doesn't reflect the
324          * congestion state of the swapdevs.  Easy to fix, if needed.
325          * See swapfile.c:page_queue_congested().
326          */
327         if (!is_page_cache_freeable(page))
328                 return PAGE_KEEP;
329         if (!mapping) {
330                 /*
331                  * Some data journaling orphaned pages can have
332                  * page->mapping == NULL while being dirty with clean buffers.
333                  */
334                 if (PagePrivate(page)) {
335                         if (try_to_free_buffers(page)) {
336                                 ClearPageDirty(page);
337                                 printk("%s: orphaned page\n", __FUNCTION__);
338                                 return PAGE_CLEAN;
339                         }
340                 }
341                 return PAGE_KEEP;
342         }
343         if (mapping->a_ops->writepage == NULL)
344                 return PAGE_ACTIVATE;
345         if (!may_write_to_queue(mapping->backing_dev_info))
346                 return PAGE_KEEP;
347
348         if (clear_page_dirty_for_io(page)) {
349                 int res;
350                 struct writeback_control wbc = {
351                         .sync_mode = WB_SYNC_NONE,
352                         .nr_to_write = SWAP_CLUSTER_MAX,
353                         .range_start = 0,
354                         .range_end = LLONG_MAX,
355                         .nonblocking = 1,
356                         .for_reclaim = 1,
357                 };
358
359                 SetPageReclaim(page);
360                 res = mapping->a_ops->writepage(page, &wbc);
361                 if (res < 0)
362                         handle_write_error(mapping, page, res);
363                 if (res == AOP_WRITEPAGE_ACTIVATE) {
364                         ClearPageReclaim(page);
365                         return PAGE_ACTIVATE;
366                 }
367                 if (!PageWriteback(page)) {
368                         /* synchronous write or broken a_ops? */
369                         ClearPageReclaim(page);
370                 }
371
372                 return PAGE_SUCCESS;
373         }
374
375         return PAGE_CLEAN;
376 }
377
378 int remove_mapping(struct address_space *mapping, struct page *page)
379 {
380         if (!mapping)
381                 return 0;               /* truncate got there first */
382
383         write_lock_irq(&mapping->tree_lock);
384
385         /*
386          * The non-racy check for busy page.  It is critical to check
387          * PageDirty _after_ making sure that the page is freeable and
388          * not in use by anybody.       (pagecache + us == 2)
389          */
390         if (unlikely(page_count(page) != 2))
391                 goto cannot_free;
392         smp_rmb();
393         if (unlikely(PageDirty(page)))
394                 goto cannot_free;
395
396         if (PageSwapCache(page)) {
397                 swp_entry_t swap = { .val = page_private(page) };
398                 __delete_from_swap_cache(page);
399                 write_unlock_irq(&mapping->tree_lock);
400                 swap_free(swap);
401                 __put_page(page);       /* The pagecache ref */
402                 return 1;
403         }
404
405         __remove_from_page_cache(page);
406         write_unlock_irq(&mapping->tree_lock);
407         __put_page(page);
408         return 1;
409
410 cannot_free:
411         write_unlock_irq(&mapping->tree_lock);
412         return 0;
413 }
414
415 /*
416  * shrink_page_list() returns the number of reclaimed pages
417  */
418 static unsigned long shrink_page_list(struct list_head *page_list,
419                                         struct scan_control *sc)
420 {
421         LIST_HEAD(ret_pages);
422         struct pagevec freed_pvec;
423         int pgactivate = 0;
424         unsigned long nr_reclaimed = 0;
425
426         cond_resched();
427
428         pagevec_init(&freed_pvec, 1);
429         while (!list_empty(page_list)) {
430                 struct address_space *mapping;
431                 struct page *page;
432                 int may_enter_fs;
433                 int referenced;
434
435                 cond_resched();
436
437                 page = lru_to_page(page_list);
438                 list_del(&page->lru);
439
440                 if (TestSetPageLocked(page))
441                         goto keep;
442
443                 BUG_ON(PageActive(page));
444
445                 sc->nr_scanned++;
446
447                 if (!sc->may_swap && page_mapped(page))
448                         goto keep_locked;
449
450                 /* Double the slab pressure for mapped and swapcache pages */
451                 if (page_mapped(page) || PageSwapCache(page))
452                         sc->nr_scanned++;
453
454                 if (PageWriteback(page))
455                         goto keep_locked;
456
457                 referenced = page_referenced(page, 1);
458                 /* In active use or really unfreeable?  Activate it. */
459                 if (referenced && page_mapping_inuse(page))
460                         goto activate_locked;
461
462 #ifdef CONFIG_SWAP
463                 /*
464                  * Anonymous process memory has backing store?
465                  * Try to allocate it some swap space here.
466                  */
467                 if (PageAnon(page) && !PageSwapCache(page))
468                         if (!add_to_swap(page, GFP_ATOMIC))
469                                 goto activate_locked;
470 #endif /* CONFIG_SWAP */
471
472                 mapping = page_mapping(page);
473                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
474                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
475
476                 /*
477                  * The page is mapped into the page tables of one or more
478                  * processes. Try to unmap it here.
479                  */
480                 if (page_mapped(page) && mapping) {
481                         switch (try_to_unmap(page, 0)) {
482                         case SWAP_FAIL:
483                                 goto activate_locked;
484                         case SWAP_AGAIN:
485                                 goto keep_locked;
486                         case SWAP_SUCCESS:
487                                 ; /* try to free the page below */
488                         }
489                 }
490
491                 if (PageDirty(page)) {
492                         if (referenced)
493                                 goto keep_locked;
494                         if (!may_enter_fs)
495                                 goto keep_locked;
496                         if (!sc->may_writepage)
497                                 goto keep_locked;
498
499                         /* Page is dirty, try to write it out here */
500                         switch(pageout(page, mapping)) {
501                         case PAGE_KEEP:
502                                 goto keep_locked;
503                         case PAGE_ACTIVATE:
504                                 goto activate_locked;
505                         case PAGE_SUCCESS:
506                                 if (PageWriteback(page) || PageDirty(page))
507                                         goto keep;
508                                 /*
509                                  * A synchronous write - probably a ramdisk.  Go
510                                  * ahead and try to reclaim the page.
511                                  */
512                                 if (TestSetPageLocked(page))
513                                         goto keep;
514                                 if (PageDirty(page) || PageWriteback(page))
515                                         goto keep_locked;
516                                 mapping = page_mapping(page);
517                         case PAGE_CLEAN:
518                                 ; /* try to free the page below */
519                         }
520                 }
521
522                 /*
523                  * If the page has buffers, try to free the buffer mappings
524                  * associated with this page. If we succeed we try to free
525                  * the page as well.
526                  *
527                  * We do this even if the page is PageDirty().
528                  * try_to_release_page() does not perform I/O, but it is
529                  * possible for a page to have PageDirty set, but it is actually
530                  * clean (all its buffers are clean).  This happens if the
531                  * buffers were written out directly, with submit_bh(). ext3
532                  * will do this, as well as the blockdev mapping. 
533                  * try_to_release_page() will discover that cleanness and will
534                  * drop the buffers and mark the page clean - it can be freed.
535                  *
536                  * Rarely, pages can have buffers and no ->mapping.  These are
537                  * the pages which were not successfully invalidated in
538                  * truncate_complete_page().  We try to drop those buffers here
539                  * and if that worked, and the page is no longer mapped into
540                  * process address space (page_count == 1) it can be freed.
541                  * Otherwise, leave the page on the LRU so it is swappable.
542                  */
543                 if (PagePrivate(page)) {
544                         if (!try_to_release_page(page, sc->gfp_mask))
545                                 goto activate_locked;
546                         if (!mapping && page_count(page) == 1)
547                                 goto free_it;
548                 }
549
550                 if (!remove_mapping(mapping, page))
551                         goto keep_locked;
552
553 free_it:
554                 unlock_page(page);
555                 nr_reclaimed++;
556                 if (!pagevec_add(&freed_pvec, page))
557                         __pagevec_release_nonlru(&freed_pvec);
558                 continue;
559
560 activate_locked:
561                 SetPageActive(page);
562                 pgactivate++;
563 keep_locked:
564                 unlock_page(page);
565 keep:
566                 list_add(&page->lru, &ret_pages);
567                 BUG_ON(PageLRU(page));
568         }
569         list_splice(&ret_pages, page_list);
570         if (pagevec_count(&freed_pvec))
571                 __pagevec_release_nonlru(&freed_pvec);
572         mod_page_state(pgactivate, pgactivate);
573         return nr_reclaimed;
574 }
575
576 /*
577  * zone->lru_lock is heavily contended.  Some of the functions that
578  * shrink the lists perform better by taking out a batch of pages
579  * and working on them outside the LRU lock.
580  *
581  * For pagecache intensive workloads, this function is the hottest
582  * spot in the kernel (apart from copy_*_user functions).
583  *
584  * Appropriate locks must be held before calling this function.
585  *
586  * @nr_to_scan: The number of pages to look through on the list.
587  * @src:        The LRU list to pull pages off.
588  * @dst:        The temp list to put pages on to.
589  * @scanned:    The number of pages that were scanned.
590  *
591  * returns how many pages were moved onto *@dst.
592  */
593 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
594                 struct list_head *src, struct list_head *dst,
595                 unsigned long *scanned)
596 {
597         unsigned long nr_taken = 0;
598         struct page *page;
599         unsigned long scan;
600
601         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
602                 struct list_head *target;
603                 page = lru_to_page(src);
604                 prefetchw_prev_lru_page(page, src, flags);
605
606                 BUG_ON(!PageLRU(page));
607
608                 list_del(&page->lru);
609                 target = src;
610                 if (likely(get_page_unless_zero(page))) {
611                         /*
612                          * Be careful not to clear PageLRU until after we're
613                          * sure the page is not being freed elsewhere -- the
614                          * page release code relies on it.
615                          */
616                         ClearPageLRU(page);
617                         target = dst;
618                         nr_taken++;
619                 } /* else it is being freed elsewhere */
620
621                 list_add(&page->lru, target);
622         }
623
624         *scanned = scan;
625         return nr_taken;
626 }
627
628 /*
629  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
630  * of reclaimed pages
631  */
632 static unsigned long shrink_inactive_list(unsigned long max_scan,
633                                 struct zone *zone, struct scan_control *sc)
634 {
635         LIST_HEAD(page_list);
636         struct pagevec pvec;
637         unsigned long nr_scanned = 0;
638         unsigned long nr_reclaimed = 0;
639
640         pagevec_init(&pvec, 1);
641
642         lru_add_drain();
643         spin_lock_irq(&zone->lru_lock);
644         do {
645                 struct page *page;
646                 unsigned long nr_taken;
647                 unsigned long nr_scan;
648                 unsigned long nr_freed;
649
650                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
651                                              &zone->inactive_list,
652                                              &page_list, &nr_scan);
653                 zone->nr_inactive -= nr_taken;
654                 zone->pages_scanned += nr_scan;
655                 spin_unlock_irq(&zone->lru_lock);
656
657                 nr_scanned += nr_scan;
658                 nr_freed = shrink_page_list(&page_list, sc);
659                 nr_reclaimed += nr_freed;
660                 local_irq_disable();
661                 if (current_is_kswapd()) {
662                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
663                         __mod_page_state(kswapd_steal, nr_freed);
664                 } else
665                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
666                 __mod_page_state_zone(zone, pgsteal, nr_freed);
667
668                 if (nr_taken == 0)
669                         goto done;
670
671                 spin_lock(&zone->lru_lock);
672                 /*
673                  * Put back any unfreeable pages.
674                  */
675                 while (!list_empty(&page_list)) {
676                         page = lru_to_page(&page_list);
677                         BUG_ON(PageLRU(page));
678                         SetPageLRU(page);
679                         list_del(&page->lru);
680                         if (PageActive(page))
681                                 add_page_to_active_list(zone, page);
682                         else
683                                 add_page_to_inactive_list(zone, page);
684                         if (!pagevec_add(&pvec, page)) {
685                                 spin_unlock_irq(&zone->lru_lock);
686                                 __pagevec_release(&pvec);
687                                 spin_lock_irq(&zone->lru_lock);
688                         }
689                 }
690         } while (nr_scanned < max_scan);
691         spin_unlock(&zone->lru_lock);
692 done:
693         local_irq_enable();
694         pagevec_release(&pvec);
695         return nr_reclaimed;
696 }
697
698 /*
699  * This moves pages from the active list to the inactive list.
700  *
701  * We move them the other way if the page is referenced by one or more
702  * processes, from rmap.
703  *
704  * If the pages are mostly unmapped, the processing is fast and it is
705  * appropriate to hold zone->lru_lock across the whole operation.  But if
706  * the pages are mapped, the processing is slow (page_referenced()) so we
707  * should drop zone->lru_lock around each page.  It's impossible to balance
708  * this, so instead we remove the pages from the LRU while processing them.
709  * It is safe to rely on PG_active against the non-LRU pages in here because
710  * nobody will play with that bit on a non-LRU page.
711  *
712  * The downside is that we have to touch page->_count against each page.
713  * But we had to alter page->flags anyway.
714  */
715 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
716                                 struct scan_control *sc)
717 {
718         unsigned long pgmoved;
719         int pgdeactivate = 0;
720         unsigned long pgscanned;
721         LIST_HEAD(l_hold);      /* The pages which were snipped off */
722         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
723         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
724         struct page *page;
725         struct pagevec pvec;
726         int reclaim_mapped = 0;
727
728         if (sc->may_swap) {
729                 long mapped_ratio;
730                 long distress;
731                 long swap_tendency;
732
733                 /*
734                  * `distress' is a measure of how much trouble we're having
735                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
736                  */
737                 distress = 100 >> zone->prev_priority;
738
739                 /*
740                  * The point of this algorithm is to decide when to start
741                  * reclaiming mapped memory instead of just pagecache.  Work out
742                  * how much memory
743                  * is mapped.
744                  */
745                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
746                                 global_page_state(NR_ANON_PAGES)) * 100) /
747                                         vm_total_pages;
748
749                 /*
750                  * Now decide how much we really want to unmap some pages.  The
751                  * mapped ratio is downgraded - just because there's a lot of
752                  * mapped memory doesn't necessarily mean that page reclaim
753                  * isn't succeeding.
754                  *
755                  * The distress ratio is important - we don't want to start
756                  * going oom.
757                  *
758                  * A 100% value of vm_swappiness overrides this algorithm
759                  * altogether.
760                  */
761                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
762
763                 /*
764                  * Now use this metric to decide whether to start moving mapped
765                  * memory onto the inactive list.
766                  */
767                 if (swap_tendency >= 100)
768                         reclaim_mapped = 1;
769         }
770
771         lru_add_drain();
772         spin_lock_irq(&zone->lru_lock);
773         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
774                                     &l_hold, &pgscanned);
775         zone->pages_scanned += pgscanned;
776         zone->nr_active -= pgmoved;
777         spin_unlock_irq(&zone->lru_lock);
778
779         while (!list_empty(&l_hold)) {
780                 cond_resched();
781                 page = lru_to_page(&l_hold);
782                 list_del(&page->lru);
783                 if (page_mapped(page)) {
784                         if (!reclaim_mapped ||
785                             (total_swap_pages == 0 && PageAnon(page)) ||
786                             page_referenced(page, 0)) {
787                                 list_add(&page->lru, &l_active);
788                                 continue;
789                         }
790                 }
791                 list_add(&page->lru, &l_inactive);
792         }
793
794         pagevec_init(&pvec, 1);
795         pgmoved = 0;
796         spin_lock_irq(&zone->lru_lock);
797         while (!list_empty(&l_inactive)) {
798                 page = lru_to_page(&l_inactive);
799                 prefetchw_prev_lru_page(page, &l_inactive, flags);
800                 BUG_ON(PageLRU(page));
801                 SetPageLRU(page);
802                 BUG_ON(!PageActive(page));
803                 ClearPageActive(page);
804
805                 list_move(&page->lru, &zone->inactive_list);
806                 pgmoved++;
807                 if (!pagevec_add(&pvec, page)) {
808                         zone->nr_inactive += pgmoved;
809                         spin_unlock_irq(&zone->lru_lock);
810                         pgdeactivate += pgmoved;
811                         pgmoved = 0;
812                         if (buffer_heads_over_limit)
813                                 pagevec_strip(&pvec);
814                         __pagevec_release(&pvec);
815                         spin_lock_irq(&zone->lru_lock);
816                 }
817         }
818         zone->nr_inactive += pgmoved;
819         pgdeactivate += pgmoved;
820         if (buffer_heads_over_limit) {
821                 spin_unlock_irq(&zone->lru_lock);
822                 pagevec_strip(&pvec);
823                 spin_lock_irq(&zone->lru_lock);
824         }
825
826         pgmoved = 0;
827         while (!list_empty(&l_active)) {
828                 page = lru_to_page(&l_active);
829                 prefetchw_prev_lru_page(page, &l_active, flags);
830                 BUG_ON(PageLRU(page));
831                 SetPageLRU(page);
832                 BUG_ON(!PageActive(page));
833                 list_move(&page->lru, &zone->active_list);
834                 pgmoved++;
835                 if (!pagevec_add(&pvec, page)) {
836                         zone->nr_active += pgmoved;
837                         pgmoved = 0;
838                         spin_unlock_irq(&zone->lru_lock);
839                         __pagevec_release(&pvec);
840                         spin_lock_irq(&zone->lru_lock);
841                 }
842         }
843         zone->nr_active += pgmoved;
844         spin_unlock(&zone->lru_lock);
845
846         __mod_page_state_zone(zone, pgrefill, pgscanned);
847         __mod_page_state(pgdeactivate, pgdeactivate);
848         local_irq_enable();
849
850         pagevec_release(&pvec);
851 }
852
853 /*
854  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
855  */
856 static unsigned long shrink_zone(int priority, struct zone *zone,
857                                 struct scan_control *sc)
858 {
859         unsigned long nr_active;
860         unsigned long nr_inactive;
861         unsigned long nr_to_scan;
862         unsigned long nr_reclaimed = 0;
863
864         atomic_inc(&zone->reclaim_in_progress);
865
866         /*
867          * Add one to `nr_to_scan' just to make sure that the kernel will
868          * slowly sift through the active list.
869          */
870         zone->nr_scan_active += (zone->nr_active >> priority) + 1;
871         nr_active = zone->nr_scan_active;
872         if (nr_active >= sc->swap_cluster_max)
873                 zone->nr_scan_active = 0;
874         else
875                 nr_active = 0;
876
877         zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
878         nr_inactive = zone->nr_scan_inactive;
879         if (nr_inactive >= sc->swap_cluster_max)
880                 zone->nr_scan_inactive = 0;
881         else
882                 nr_inactive = 0;
883
884         while (nr_active || nr_inactive) {
885                 if (nr_active) {
886                         nr_to_scan = min(nr_active,
887                                         (unsigned long)sc->swap_cluster_max);
888                         nr_active -= nr_to_scan;
889                         shrink_active_list(nr_to_scan, zone, sc);
890                 }
891
892                 if (nr_inactive) {
893                         nr_to_scan = min(nr_inactive,
894                                         (unsigned long)sc->swap_cluster_max);
895                         nr_inactive -= nr_to_scan;
896                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
897                                                                 sc);
898                 }
899         }
900
901         throttle_vm_writeout();
902
903         atomic_dec(&zone->reclaim_in_progress);
904         return nr_reclaimed;
905 }
906
907 /*
908  * This is the direct reclaim path, for page-allocating processes.  We only
909  * try to reclaim pages from zones which will satisfy the caller's allocation
910  * request.
911  *
912  * We reclaim from a zone even if that zone is over pages_high.  Because:
913  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
914  *    allocation or
915  * b) The zones may be over pages_high but they must go *over* pages_high to
916  *    satisfy the `incremental min' zone defense algorithm.
917  *
918  * Returns the number of reclaimed pages.
919  *
920  * If a zone is deemed to be full of pinned pages then just give it a light
921  * scan then give up on it.
922  */
923 static unsigned long shrink_zones(int priority, struct zone **zones,
924                                         struct scan_control *sc)
925 {
926         unsigned long nr_reclaimed = 0;
927         int i;
928
929         for (i = 0; zones[i] != NULL; i++) {
930                 struct zone *zone = zones[i];
931
932                 if (!populated_zone(zone))
933                         continue;
934
935                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
936                         continue;
937
938                 zone->temp_priority = priority;
939                 if (zone->prev_priority > priority)
940                         zone->prev_priority = priority;
941
942                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
943                         continue;       /* Let kswapd poll it */
944
945                 nr_reclaimed += shrink_zone(priority, zone, sc);
946         }
947         return nr_reclaimed;
948 }
949  
950 /*
951  * This is the main entry point to direct page reclaim.
952  *
953  * If a full scan of the inactive list fails to free enough memory then we
954  * are "out of memory" and something needs to be killed.
955  *
956  * If the caller is !__GFP_FS then the probability of a failure is reasonably
957  * high - the zone may be full of dirty or under-writeback pages, which this
958  * caller can't do much about.  We kick pdflush and take explicit naps in the
959  * hope that some of these pages can be written.  But if the allocating task
960  * holds filesystem locks which prevent writeout this might not work, and the
961  * allocation attempt will fail.
962  */
963 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
964 {
965         int priority;
966         int ret = 0;
967         unsigned long total_scanned = 0;
968         unsigned long nr_reclaimed = 0;
969         struct reclaim_state *reclaim_state = current->reclaim_state;
970         unsigned long lru_pages = 0;
971         int i;
972         struct scan_control sc = {
973                 .gfp_mask = gfp_mask,
974                 .may_writepage = !laptop_mode,
975                 .swap_cluster_max = SWAP_CLUSTER_MAX,
976                 .may_swap = 1,
977                 .swappiness = vm_swappiness,
978         };
979
980         inc_page_state(allocstall);
981
982         for (i = 0; zones[i] != NULL; i++) {
983                 struct zone *zone = zones[i];
984
985                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
986                         continue;
987
988                 zone->temp_priority = DEF_PRIORITY;
989                 lru_pages += zone->nr_active + zone->nr_inactive;
990         }
991
992         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
993                 sc.nr_scanned = 0;
994                 if (!priority)
995                         disable_swap_token();
996                 nr_reclaimed += shrink_zones(priority, zones, &sc);
997                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
998                 if (reclaim_state) {
999                         nr_reclaimed += reclaim_state->reclaimed_slab;
1000                         reclaim_state->reclaimed_slab = 0;
1001                 }
1002                 total_scanned += sc.nr_scanned;
1003                 if (nr_reclaimed >= sc.swap_cluster_max) {
1004                         ret = 1;
1005                         goto out;
1006                 }
1007
1008                 /*
1009                  * Try to write back as many pages as we just scanned.  This
1010                  * tends to cause slow streaming writers to write data to the
1011                  * disk smoothly, at the dirtying rate, which is nice.   But
1012                  * that's undesirable in laptop mode, where we *want* lumpy
1013                  * writeout.  So in laptop mode, write out the whole world.
1014                  */
1015                 if (total_scanned > sc.swap_cluster_max +
1016                                         sc.swap_cluster_max / 2) {
1017                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1018                         sc.may_writepage = 1;
1019                 }
1020
1021                 /* Take a nap, wait for some writeback to complete */
1022                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1023                         blk_congestion_wait(WRITE, HZ/10);
1024         }
1025 out:
1026         for (i = 0; zones[i] != 0; i++) {
1027                 struct zone *zone = zones[i];
1028
1029                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1030                         continue;
1031
1032                 zone->prev_priority = zone->temp_priority;
1033         }
1034         return ret;
1035 }
1036
1037 /*
1038  * For kswapd, balance_pgdat() will work across all this node's zones until
1039  * they are all at pages_high.
1040  *
1041  * Returns the number of pages which were actually freed.
1042  *
1043  * There is special handling here for zones which are full of pinned pages.
1044  * This can happen if the pages are all mlocked, or if they are all used by
1045  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1046  * What we do is to detect the case where all pages in the zone have been
1047  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1048  * dead and from now on, only perform a short scan.  Basically we're polling
1049  * the zone for when the problem goes away.
1050  *
1051  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1052  * zones which have free_pages > pages_high, but once a zone is found to have
1053  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1054  * of the number of free pages in the lower zones.  This interoperates with
1055  * the page allocator fallback scheme to ensure that aging of pages is balanced
1056  * across the zones.
1057  */
1058 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1059 {
1060         int all_zones_ok;
1061         int priority;
1062         int i;
1063         unsigned long total_scanned;
1064         unsigned long nr_reclaimed;
1065         struct reclaim_state *reclaim_state = current->reclaim_state;
1066         struct scan_control sc = {
1067                 .gfp_mask = GFP_KERNEL,
1068                 .may_swap = 1,
1069                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1070                 .swappiness = vm_swappiness,
1071         };
1072
1073 loop_again:
1074         total_scanned = 0;
1075         nr_reclaimed = 0;
1076         sc.may_writepage = !laptop_mode;
1077         inc_page_state(pageoutrun);
1078
1079         for (i = 0; i < pgdat->nr_zones; i++) {
1080                 struct zone *zone = pgdat->node_zones + i;
1081
1082                 zone->temp_priority = DEF_PRIORITY;
1083         }
1084
1085         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1086                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1087                 unsigned long lru_pages = 0;
1088
1089                 /* The swap token gets in the way of swapout... */
1090                 if (!priority)
1091                         disable_swap_token();
1092
1093                 all_zones_ok = 1;
1094
1095                 /*
1096                  * Scan in the highmem->dma direction for the highest
1097                  * zone which needs scanning
1098                  */
1099                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1100                         struct zone *zone = pgdat->node_zones + i;
1101
1102                         if (!populated_zone(zone))
1103                                 continue;
1104
1105                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1106                                 continue;
1107
1108                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1109                                                0, 0)) {
1110                                 end_zone = i;
1111                                 goto scan;
1112                         }
1113                 }
1114                 goto out;
1115 scan:
1116                 for (i = 0; i <= end_zone; i++) {
1117                         struct zone *zone = pgdat->node_zones + i;
1118
1119                         lru_pages += zone->nr_active + zone->nr_inactive;
1120                 }
1121
1122                 /*
1123                  * Now scan the zone in the dma->highmem direction, stopping
1124                  * at the last zone which needs scanning.
1125                  *
1126                  * We do this because the page allocator works in the opposite
1127                  * direction.  This prevents the page allocator from allocating
1128                  * pages behind kswapd's direction of progress, which would
1129                  * cause too much scanning of the lower zones.
1130                  */
1131                 for (i = 0; i <= end_zone; i++) {
1132                         struct zone *zone = pgdat->node_zones + i;
1133                         int nr_slab;
1134
1135                         if (!populated_zone(zone))
1136                                 continue;
1137
1138                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1139                                 continue;
1140
1141                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1142                                                end_zone, 0))
1143                                 all_zones_ok = 0;
1144                         zone->temp_priority = priority;
1145                         if (zone->prev_priority > priority)
1146                                 zone->prev_priority = priority;
1147                         sc.nr_scanned = 0;
1148                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1149                         reclaim_state->reclaimed_slab = 0;
1150                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1151                                                 lru_pages);
1152                         nr_reclaimed += reclaim_state->reclaimed_slab;
1153                         total_scanned += sc.nr_scanned;
1154                         if (zone->all_unreclaimable)
1155                                 continue;
1156                         if (nr_slab == 0 && zone->pages_scanned >=
1157                                     (zone->nr_active + zone->nr_inactive) * 4)
1158                                 zone->all_unreclaimable = 1;
1159                         /*
1160                          * If we've done a decent amount of scanning and
1161                          * the reclaim ratio is low, start doing writepage
1162                          * even in laptop mode
1163                          */
1164                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1165                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1166                                 sc.may_writepage = 1;
1167                 }
1168                 if (all_zones_ok)
1169                         break;          /* kswapd: all done */
1170                 /*
1171                  * OK, kswapd is getting into trouble.  Take a nap, then take
1172                  * another pass across the zones.
1173                  */
1174                 if (total_scanned && priority < DEF_PRIORITY - 2)
1175                         blk_congestion_wait(WRITE, HZ/10);
1176
1177                 /*
1178                  * We do this so kswapd doesn't build up large priorities for
1179                  * example when it is freeing in parallel with allocators. It
1180                  * matches the direct reclaim path behaviour in terms of impact
1181                  * on zone->*_priority.
1182                  */
1183                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1184                         break;
1185         }
1186 out:
1187         for (i = 0; i < pgdat->nr_zones; i++) {
1188                 struct zone *zone = pgdat->node_zones + i;
1189
1190                 zone->prev_priority = zone->temp_priority;
1191         }
1192         if (!all_zones_ok) {
1193                 cond_resched();
1194                 goto loop_again;
1195         }
1196
1197         return nr_reclaimed;
1198 }
1199
1200 /*
1201  * The background pageout daemon, started as a kernel thread
1202  * from the init process. 
1203  *
1204  * This basically trickles out pages so that we have _some_
1205  * free memory available even if there is no other activity
1206  * that frees anything up. This is needed for things like routing
1207  * etc, where we otherwise might have all activity going on in
1208  * asynchronous contexts that cannot page things out.
1209  *
1210  * If there are applications that are active memory-allocators
1211  * (most normal use), this basically shouldn't matter.
1212  */
1213 static int kswapd(void *p)
1214 {
1215         unsigned long order;
1216         pg_data_t *pgdat = (pg_data_t*)p;
1217         struct task_struct *tsk = current;
1218         DEFINE_WAIT(wait);
1219         struct reclaim_state reclaim_state = {
1220                 .reclaimed_slab = 0,
1221         };
1222         cpumask_t cpumask;
1223
1224         cpumask = node_to_cpumask(pgdat->node_id);
1225         if (!cpus_empty(cpumask))
1226                 set_cpus_allowed(tsk, cpumask);
1227         current->reclaim_state = &reclaim_state;
1228
1229         /*
1230          * Tell the memory management that we're a "memory allocator",
1231          * and that if we need more memory we should get access to it
1232          * regardless (see "__alloc_pages()"). "kswapd" should
1233          * never get caught in the normal page freeing logic.
1234          *
1235          * (Kswapd normally doesn't need memory anyway, but sometimes
1236          * you need a small amount of memory in order to be able to
1237          * page out something else, and this flag essentially protects
1238          * us from recursively trying to free more memory as we're
1239          * trying to free the first piece of memory in the first place).
1240          */
1241         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1242
1243         order = 0;
1244         for ( ; ; ) {
1245                 unsigned long new_order;
1246
1247                 try_to_freeze();
1248
1249                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1250                 new_order = pgdat->kswapd_max_order;
1251                 pgdat->kswapd_max_order = 0;
1252                 if (order < new_order) {
1253                         /*
1254                          * Don't sleep if someone wants a larger 'order'
1255                          * allocation
1256                          */
1257                         order = new_order;
1258                 } else {
1259                         schedule();
1260                         order = pgdat->kswapd_max_order;
1261                 }
1262                 finish_wait(&pgdat->kswapd_wait, &wait);
1263
1264                 balance_pgdat(pgdat, order);
1265         }
1266         return 0;
1267 }
1268
1269 /*
1270  * A zone is low on free memory, so wake its kswapd task to service it.
1271  */
1272 void wakeup_kswapd(struct zone *zone, int order)
1273 {
1274         pg_data_t *pgdat;
1275
1276         if (!populated_zone(zone))
1277                 return;
1278
1279         pgdat = zone->zone_pgdat;
1280         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1281                 return;
1282         if (pgdat->kswapd_max_order < order)
1283                 pgdat->kswapd_max_order = order;
1284         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1285                 return;
1286         if (!waitqueue_active(&pgdat->kswapd_wait))
1287                 return;
1288         wake_up_interruptible(&pgdat->kswapd_wait);
1289 }
1290
1291 #ifdef CONFIG_PM
1292 /*
1293  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1294  * from LRU lists system-wide, for given pass and priority, and returns the
1295  * number of reclaimed pages
1296  *
1297  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1298  */
1299 static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1300                                       int prio, struct scan_control *sc)
1301 {
1302         struct zone *zone;
1303         unsigned long nr_to_scan, ret = 0;
1304
1305         for_each_zone(zone) {
1306
1307                 if (!populated_zone(zone))
1308                         continue;
1309
1310                 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1311                         continue;
1312
1313                 /* For pass = 0 we don't shrink the active list */
1314                 if (pass > 0) {
1315                         zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1316                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1317                                 zone->nr_scan_active = 0;
1318                                 nr_to_scan = min(nr_pages, zone->nr_active);
1319                                 shrink_active_list(nr_to_scan, zone, sc);
1320                         }
1321                 }
1322
1323                 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1324                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1325                         zone->nr_scan_inactive = 0;
1326                         nr_to_scan = min(nr_pages, zone->nr_inactive);
1327                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1328                         if (ret >= nr_pages)
1329                                 return ret;
1330                 }
1331         }
1332
1333         return ret;
1334 }
1335
1336 /*
1337  * Try to free `nr_pages' of memory, system-wide, and return the number of
1338  * freed pages.
1339  *
1340  * Rather than trying to age LRUs the aim is to preserve the overall
1341  * LRU order by reclaiming preferentially
1342  * inactive > active > active referenced > active mapped
1343  */
1344 unsigned long shrink_all_memory(unsigned long nr_pages)
1345 {
1346         unsigned long lru_pages, nr_slab;
1347         unsigned long ret = 0;
1348         int pass;
1349         struct reclaim_state reclaim_state;
1350         struct zone *zone;
1351         struct scan_control sc = {
1352                 .gfp_mask = GFP_KERNEL,
1353                 .may_swap = 0,
1354                 .swap_cluster_max = nr_pages,
1355                 .may_writepage = 1,
1356                 .swappiness = vm_swappiness,
1357         };
1358
1359         current->reclaim_state = &reclaim_state;
1360
1361         lru_pages = 0;
1362         for_each_zone(zone)
1363                 lru_pages += zone->nr_active + zone->nr_inactive;
1364
1365         nr_slab = read_page_state(nr_slab);
1366         /* If slab caches are huge, it's better to hit them first */
1367         while (nr_slab >= lru_pages) {
1368                 reclaim_state.reclaimed_slab = 0;
1369                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1370                 if (!reclaim_state.reclaimed_slab)
1371                         break;
1372
1373                 ret += reclaim_state.reclaimed_slab;
1374                 if (ret >= nr_pages)
1375                         goto out;
1376
1377                 nr_slab -= reclaim_state.reclaimed_slab;
1378         }
1379
1380         /*
1381          * We try to shrink LRUs in 5 passes:
1382          * 0 = Reclaim from inactive_list only
1383          * 1 = Reclaim from active list but don't reclaim mapped
1384          * 2 = 2nd pass of type 1
1385          * 3 = Reclaim mapped (normal reclaim)
1386          * 4 = 2nd pass of type 3
1387          */
1388         for (pass = 0; pass < 5; pass++) {
1389                 int prio;
1390
1391                 /* Needed for shrinking slab caches later on */
1392                 if (!lru_pages)
1393                         for_each_zone(zone) {
1394                                 lru_pages += zone->nr_active;
1395                                 lru_pages += zone->nr_inactive;
1396                         }
1397
1398                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1399                 if (pass > 2) {
1400                         sc.may_swap = 1;
1401                         sc.swappiness = 100;
1402                 }
1403
1404                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1405                         unsigned long nr_to_scan = nr_pages - ret;
1406
1407                         sc.nr_scanned = 0;
1408                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1409                         if (ret >= nr_pages)
1410                                 goto out;
1411
1412                         reclaim_state.reclaimed_slab = 0;
1413                         shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1414                         ret += reclaim_state.reclaimed_slab;
1415                         if (ret >= nr_pages)
1416                                 goto out;
1417
1418                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1419                                 blk_congestion_wait(WRITE, HZ / 10);
1420                 }
1421
1422                 lru_pages = 0;
1423         }
1424
1425         /*
1426          * If ret = 0, we could not shrink LRUs, but there may be something
1427          * in slab caches
1428          */
1429         if (!ret)
1430                 do {
1431                         reclaim_state.reclaimed_slab = 0;
1432                         shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1433                         ret += reclaim_state.reclaimed_slab;
1434                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1435
1436 out:
1437         current->reclaim_state = NULL;
1438
1439         return ret;
1440 }
1441 #endif
1442
1443 #ifdef CONFIG_HOTPLUG_CPU
1444 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1445    not required for correctness.  So if the last cpu in a node goes
1446    away, we get changed to run anywhere: as the first one comes back,
1447    restore their cpu bindings. */
1448 static int __devinit cpu_callback(struct notifier_block *nfb,
1449                                   unsigned long action, void *hcpu)
1450 {
1451         pg_data_t *pgdat;
1452         cpumask_t mask;
1453
1454         if (action == CPU_ONLINE) {
1455                 for_each_online_pgdat(pgdat) {
1456                         mask = node_to_cpumask(pgdat->node_id);
1457                         if (any_online_cpu(mask) != NR_CPUS)
1458                                 /* One of our CPUs online: restore mask */
1459                                 set_cpus_allowed(pgdat->kswapd, mask);
1460                 }
1461         }
1462         return NOTIFY_OK;
1463 }
1464 #endif /* CONFIG_HOTPLUG_CPU */
1465
1466 /*
1467  * This kswapd start function will be called by init and node-hot-add.
1468  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1469  */
1470 int kswapd_run(int nid)
1471 {
1472         pg_data_t *pgdat = NODE_DATA(nid);
1473         int ret = 0;
1474
1475         if (pgdat->kswapd)
1476                 return 0;
1477
1478         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1479         if (IS_ERR(pgdat->kswapd)) {
1480                 /* failure at boot is fatal */
1481                 BUG_ON(system_state == SYSTEM_BOOTING);
1482                 printk("Failed to start kswapd on node %d\n",nid);
1483                 ret = -1;
1484         }
1485         return ret;
1486 }
1487
1488 static int __init kswapd_init(void)
1489 {
1490         int nid;
1491
1492         swap_setup();
1493         for_each_online_node(nid)
1494                 kswapd_run(nid);
1495         hotcpu_notifier(cpu_callback, 0);
1496         return 0;
1497 }
1498
1499 module_init(kswapd_init)
1500
1501 #ifdef CONFIG_NUMA
1502 /*
1503  * Zone reclaim mode
1504  *
1505  * If non-zero call zone_reclaim when the number of free pages falls below
1506  * the watermarks.
1507  *
1508  * In the future we may add flags to the mode. However, the page allocator
1509  * should only have to check that zone_reclaim_mode != 0 before calling
1510  * zone_reclaim().
1511  */
1512 int zone_reclaim_mode __read_mostly;
1513
1514 #define RECLAIM_OFF 0
1515 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1516 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1517 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1518 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1519
1520 /*
1521  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1522  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1523  * a zone.
1524  */
1525 #define ZONE_RECLAIM_PRIORITY 4
1526
1527 /*
1528  * Try to free up some pages from this zone through reclaim.
1529  */
1530 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1531 {
1532         /* Minimum pages needed in order to stay on node */
1533         const unsigned long nr_pages = 1 << order;
1534         struct task_struct *p = current;
1535         struct reclaim_state reclaim_state;
1536         int priority;
1537         unsigned long nr_reclaimed = 0;
1538         struct scan_control sc = {
1539                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1540                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1541                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1542                                         SWAP_CLUSTER_MAX),
1543                 .gfp_mask = gfp_mask,
1544                 .swappiness = vm_swappiness,
1545         };
1546
1547         disable_swap_token();
1548         cond_resched();
1549         /*
1550          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1551          * and we also need to be able to write out pages for RECLAIM_WRITE
1552          * and RECLAIM_SWAP.
1553          */
1554         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1555         reclaim_state.reclaimed_slab = 0;
1556         p->reclaim_state = &reclaim_state;
1557
1558         /*
1559          * Free memory by calling shrink zone with increasing priorities
1560          * until we have enough memory freed.
1561          */
1562         priority = ZONE_RECLAIM_PRIORITY;
1563         do {
1564                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1565                 priority--;
1566         } while (priority >= 0 && nr_reclaimed < nr_pages);
1567
1568         if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1569                 /*
1570                  * shrink_slab() does not currently allow us to determine how
1571                  * many pages were freed in this zone. So we just shake the slab
1572                  * a bit and then go off node for this particular allocation
1573                  * despite possibly having freed enough memory to allocate in
1574                  * this zone.  If we freed local memory then the next
1575                  * allocations will be local again.
1576                  *
1577                  * shrink_slab will free memory on all zones and may take
1578                  * a long time.
1579                  */
1580                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1581         }
1582
1583         p->reclaim_state = NULL;
1584         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1585         return nr_reclaimed >= nr_pages;
1586 }
1587
1588 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1589 {
1590         cpumask_t mask;
1591         int node_id;
1592
1593         /*
1594          * Do not reclaim if there are not enough reclaimable pages in this
1595          * zone that would satify this allocations.
1596          *
1597          * All unmapped pagecache pages are reclaimable.
1598          *
1599          * Both counters may be temporarily off a bit so we use
1600          * SWAP_CLUSTER_MAX as the boundary. It may also be good to
1601          * leave a few frequently used unmapped pagecache pages around.
1602          */
1603         if (zone_page_state(zone, NR_FILE_PAGES) -
1604                 zone_page_state(zone, NR_FILE_MAPPED) < SWAP_CLUSTER_MAX)
1605                         return 0;
1606
1607         /*
1608          * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1609          * not have reclaimable pages and if we should not delay the allocation
1610          * then do not scan.
1611          */
1612         if (!(gfp_mask & __GFP_WAIT) ||
1613                 zone->all_unreclaimable ||
1614                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1615                 (current->flags & PF_MEMALLOC))
1616                         return 0;
1617
1618         /*
1619          * Only run zone reclaim on the local zone or on zones that do not
1620          * have associated processors. This will favor the local processor
1621          * over remote processors and spread off node memory allocations
1622          * as wide as possible.
1623          */
1624         node_id = zone->zone_pgdat->node_id;
1625         mask = node_to_cpumask(node_id);
1626         if (!cpus_empty(mask) && node_id != numa_node_id())
1627                 return 0;
1628         return __zone_reclaim(zone, gfp_mask, order);
1629 }
1630 #endif