Merge master.kernel.org:/pub/scm/linux/kernel/git/lethal/sh-2.6
[linux-2.6] / 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/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
40 #include <asm/tlbflush.h>
41 #include <asm/div64.h>
42
43 #include <linux/swapops.h>
44
45 #include "internal.h"
46
47 struct scan_control {
48         /* Incremented by the number of inactive pages that were scanned */
49         unsigned long nr_scanned;
50
51         /* This context's GFP mask */
52         gfp_t gfp_mask;
53
54         int may_writepage;
55
56         /* Can pages be swapped as part of reclaim? */
57         int may_swap;
58
59         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
60          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
61          * In this context, it doesn't matter that we scan the
62          * whole list at once. */
63         int swap_cluster_max;
64
65         int swappiness;
66
67         int all_unreclaimable;
68 };
69
70 /*
71  * The list of shrinker callbacks used by to apply pressure to
72  * ageable caches.
73  */
74 struct shrinker {
75         shrinker_t              shrinker;
76         struct list_head        list;
77         int                     seeks;  /* seeks to recreate an obj */
78         long                    nr;     /* objs pending delete */
79 };
80
81 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
82
83 #ifdef ARCH_HAS_PREFETCH
84 #define prefetch_prev_lru_page(_page, _base, _field)                    \
85         do {                                                            \
86                 if ((_page)->lru.prev != _base) {                       \
87                         struct page *prev;                              \
88                                                                         \
89                         prev = lru_to_page(&(_page->lru));              \
90                         prefetch(&prev->_field);                        \
91                 }                                                       \
92         } while (0)
93 #else
94 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
95 #endif
96
97 #ifdef ARCH_HAS_PREFETCHW
98 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
99         do {                                                            \
100                 if ((_page)->lru.prev != _base) {                       \
101                         struct page *prev;                              \
102                                                                         \
103                         prev = lru_to_page(&(_page->lru));              \
104                         prefetchw(&prev->_field);                       \
105                 }                                                       \
106         } while (0)
107 #else
108 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
109 #endif
110
111 /*
112  * From 0 .. 100.  Higher means more swappy.
113  */
114 int vm_swappiness = 60;
115 long vm_total_pages;    /* The total number of pages which the VM controls */
116
117 static LIST_HEAD(shrinker_list);
118 static DECLARE_RWSEM(shrinker_rwsem);
119
120 /*
121  * Add a shrinker callback to be called from the vm
122  */
123 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
124 {
125         struct shrinker *shrinker;
126
127         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
128         if (shrinker) {
129                 shrinker->shrinker = theshrinker;
130                 shrinker->seeks = seeks;
131                 shrinker->nr = 0;
132                 down_write(&shrinker_rwsem);
133                 list_add_tail(&shrinker->list, &shrinker_list);
134                 up_write(&shrinker_rwsem);
135         }
136         return shrinker;
137 }
138 EXPORT_SYMBOL(set_shrinker);
139
140 /*
141  * Remove one
142  */
143 void remove_shrinker(struct shrinker *shrinker)
144 {
145         down_write(&shrinker_rwsem);
146         list_del(&shrinker->list);
147         up_write(&shrinker_rwsem);
148         kfree(shrinker);
149 }
150 EXPORT_SYMBOL(remove_shrinker);
151
152 #define SHRINK_BATCH 128
153 /*
154  * Call the shrink functions to age shrinkable caches
155  *
156  * Here we assume it costs one seek to replace a lru page and that it also
157  * takes a seek to recreate a cache object.  With this in mind we age equal
158  * percentages of the lru and ageable caches.  This should balance the seeks
159  * generated by these structures.
160  *
161  * If the vm encounted mapped pages on the LRU it increase the pressure on
162  * slab to avoid swapping.
163  *
164  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
165  *
166  * `lru_pages' represents the number of on-LRU pages in all the zones which
167  * are eligible for the caller's allocation attempt.  It is used for balancing
168  * slab reclaim versus page reclaim.
169  *
170  * Returns the number of slab objects which we shrunk.
171  */
172 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
173                         unsigned long lru_pages)
174 {
175         struct shrinker *shrinker;
176         unsigned long ret = 0;
177
178         if (scanned == 0)
179                 scanned = SWAP_CLUSTER_MAX;
180
181         if (!down_read_trylock(&shrinker_rwsem))
182                 return 1;       /* Assume we'll be able to shrink next time */
183
184         list_for_each_entry(shrinker, &shrinker_list, list) {
185                 unsigned long long delta;
186                 unsigned long total_scan;
187                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
188
189                 delta = (4 * scanned) / shrinker->seeks;
190                 delta *= max_pass;
191                 do_div(delta, lru_pages + 1);
192                 shrinker->nr += delta;
193                 if (shrinker->nr < 0) {
194                         printk(KERN_ERR "%s: nr=%ld\n",
195                                         __FUNCTION__, shrinker->nr);
196                         shrinker->nr = max_pass;
197                 }
198
199                 /*
200                  * Avoid risking looping forever due to too large nr value:
201                  * never try to free more than twice the estimate number of
202                  * freeable entries.
203                  */
204                 if (shrinker->nr > max_pass * 2)
205                         shrinker->nr = max_pass * 2;
206
207                 total_scan = shrinker->nr;
208                 shrinker->nr = 0;
209
210                 while (total_scan >= SHRINK_BATCH) {
211                         long this_scan = SHRINK_BATCH;
212                         int shrink_ret;
213                         int nr_before;
214
215                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
216                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
217                         if (shrink_ret == -1)
218                                 break;
219                         if (shrink_ret < nr_before)
220                                 ret += nr_before - shrink_ret;
221                         count_vm_events(SLABS_SCANNED, this_scan);
222                         total_scan -= this_scan;
223
224                         cond_resched();
225                 }
226
227                 shrinker->nr += total_scan;
228         }
229         up_read(&shrinker_rwsem);
230         return ret;
231 }
232
233 /* Called without lock on whether page is mapped, so answer is unstable */
234 static inline int page_mapping_inuse(struct page *page)
235 {
236         struct address_space *mapping;
237
238         /* Page is in somebody's page tables. */
239         if (page_mapped(page))
240                 return 1;
241
242         /* Be more reluctant to reclaim swapcache than pagecache */
243         if (PageSwapCache(page))
244                 return 1;
245
246         mapping = page_mapping(page);
247         if (!mapping)
248                 return 0;
249
250         /* File is mmap'd by somebody? */
251         return mapping_mapped(mapping);
252 }
253
254 static inline int is_page_cache_freeable(struct page *page)
255 {
256         return page_count(page) - !!PagePrivate(page) == 2;
257 }
258
259 static int may_write_to_queue(struct backing_dev_info *bdi)
260 {
261         if (current->flags & PF_SWAPWRITE)
262                 return 1;
263         if (!bdi_write_congested(bdi))
264                 return 1;
265         if (bdi == current->backing_dev_info)
266                 return 1;
267         return 0;
268 }
269
270 /*
271  * We detected a synchronous write error writing a page out.  Probably
272  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
273  * fsync(), msync() or close().
274  *
275  * The tricky part is that after writepage we cannot touch the mapping: nothing
276  * prevents it from being freed up.  But we have a ref on the page and once
277  * that page is locked, the mapping is pinned.
278  *
279  * We're allowed to run sleeping lock_page() here because we know the caller has
280  * __GFP_FS.
281  */
282 static void handle_write_error(struct address_space *mapping,
283                                 struct page *page, int error)
284 {
285         lock_page(page);
286         if (page_mapping(page) == mapping) {
287                 if (error == -ENOSPC)
288                         set_bit(AS_ENOSPC, &mapping->flags);
289                 else
290                         set_bit(AS_EIO, &mapping->flags);
291         }
292         unlock_page(page);
293 }
294
295 /* possible outcome of pageout() */
296 typedef enum {
297         /* failed to write page out, page is locked */
298         PAGE_KEEP,
299         /* move page to the active list, page is locked */
300         PAGE_ACTIVATE,
301         /* page has been sent to the disk successfully, page is unlocked */
302         PAGE_SUCCESS,
303         /* page is clean and locked */
304         PAGE_CLEAN,
305 } pageout_t;
306
307 /*
308  * pageout is called by shrink_page_list() for each dirty page.
309  * Calls ->writepage().
310  */
311 static pageout_t pageout(struct page *page, struct address_space *mapping)
312 {
313         /*
314          * If the page is dirty, only perform writeback if that write
315          * will be non-blocking.  To prevent this allocation from being
316          * stalled by pagecache activity.  But note that there may be
317          * stalls if we need to run get_block().  We could test
318          * PagePrivate for that.
319          *
320          * If this process is currently in generic_file_write() against
321          * this page's queue, we can perform writeback even if that
322          * will block.
323          *
324          * If the page is swapcache, write it back even if that would
325          * block, for some throttling. This happens by accident, because
326          * swap_backing_dev_info is bust: it doesn't reflect the
327          * congestion state of the swapdevs.  Easy to fix, if needed.
328          * See swapfile.c:page_queue_congested().
329          */
330         if (!is_page_cache_freeable(page))
331                 return PAGE_KEEP;
332         if (!mapping) {
333                 /*
334                  * Some data journaling orphaned pages can have
335                  * page->mapping == NULL while being dirty with clean buffers.
336                  */
337                 if (PagePrivate(page)) {
338                         if (try_to_free_buffers(page)) {
339                                 ClearPageDirty(page);
340                                 printk("%s: orphaned page\n", __FUNCTION__);
341                                 return PAGE_CLEAN;
342                         }
343                 }
344                 return PAGE_KEEP;
345         }
346         if (mapping->a_ops->writepage == NULL)
347                 return PAGE_ACTIVATE;
348         if (!may_write_to_queue(mapping->backing_dev_info))
349                 return PAGE_KEEP;
350
351         if (clear_page_dirty_for_io(page)) {
352                 int res;
353                 struct writeback_control wbc = {
354                         .sync_mode = WB_SYNC_NONE,
355                         .nr_to_write = SWAP_CLUSTER_MAX,
356                         .range_start = 0,
357                         .range_end = LLONG_MAX,
358                         .nonblocking = 1,
359                         .for_reclaim = 1,
360                 };
361
362                 SetPageReclaim(page);
363                 res = mapping->a_ops->writepage(page, &wbc);
364                 if (res < 0)
365                         handle_write_error(mapping, page, res);
366                 if (res == AOP_WRITEPAGE_ACTIVATE) {
367                         ClearPageReclaim(page);
368                         return PAGE_ACTIVATE;
369                 }
370                 if (!PageWriteback(page)) {
371                         /* synchronous write or broken a_ops? */
372                         ClearPageReclaim(page);
373                 }
374                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
375                 return PAGE_SUCCESS;
376         }
377
378         return PAGE_CLEAN;
379 }
380
381 int remove_mapping(struct address_space *mapping, struct page *page)
382 {
383         BUG_ON(!PageLocked(page));
384         BUG_ON(mapping != page_mapping(page));
385
386         write_lock_irq(&mapping->tree_lock);
387         /*
388          * The non racy check for a busy page.
389          *
390          * Must be careful with the order of the tests. When someone has
391          * a ref to the page, it may be possible that they dirty it then
392          * drop the reference. So if PageDirty is tested before page_count
393          * here, then the following race may occur:
394          *
395          * get_user_pages(&page);
396          * [user mapping goes away]
397          * write_to(page);
398          *                              !PageDirty(page)    [good]
399          * SetPageDirty(page);
400          * put_page(page);
401          *                              !page_count(page)   [good, discard it]
402          *
403          * [oops, our write_to data is lost]
404          *
405          * Reversing the order of the tests ensures such a situation cannot
406          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
407          * load is not satisfied before that of page->_count.
408          *
409          * Note that if SetPageDirty is always performed via set_page_dirty,
410          * and thus under tree_lock, then this ordering is not required.
411          */
412         if (unlikely(page_count(page) != 2))
413                 goto cannot_free;
414         smp_rmb();
415         if (unlikely(PageDirty(page)))
416                 goto cannot_free;
417
418         if (PageSwapCache(page)) {
419                 swp_entry_t swap = { .val = page_private(page) };
420                 __delete_from_swap_cache(page);
421                 write_unlock_irq(&mapping->tree_lock);
422                 swap_free(swap);
423                 __put_page(page);       /* The pagecache ref */
424                 return 1;
425         }
426
427         __remove_from_page_cache(page);
428         write_unlock_irq(&mapping->tree_lock);
429         __put_page(page);
430         return 1;
431
432 cannot_free:
433         write_unlock_irq(&mapping->tree_lock);
434         return 0;
435 }
436
437 /*
438  * shrink_page_list() returns the number of reclaimed pages
439  */
440 static unsigned long shrink_page_list(struct list_head *page_list,
441                                         struct scan_control *sc)
442 {
443         LIST_HEAD(ret_pages);
444         struct pagevec freed_pvec;
445         int pgactivate = 0;
446         unsigned long nr_reclaimed = 0;
447
448         cond_resched();
449
450         pagevec_init(&freed_pvec, 1);
451         while (!list_empty(page_list)) {
452                 struct address_space *mapping;
453                 struct page *page;
454                 int may_enter_fs;
455                 int referenced;
456
457                 cond_resched();
458
459                 page = lru_to_page(page_list);
460                 list_del(&page->lru);
461
462                 if (TestSetPageLocked(page))
463                         goto keep;
464
465                 VM_BUG_ON(PageActive(page));
466
467                 sc->nr_scanned++;
468
469                 if (!sc->may_swap && page_mapped(page))
470                         goto keep_locked;
471
472                 /* Double the slab pressure for mapped and swapcache pages */
473                 if (page_mapped(page) || PageSwapCache(page))
474                         sc->nr_scanned++;
475
476                 if (PageWriteback(page))
477                         goto keep_locked;
478
479                 referenced = page_referenced(page, 1);
480                 /* In active use or really unfreeable?  Activate it. */
481                 if (referenced && page_mapping_inuse(page))
482                         goto activate_locked;
483
484 #ifdef CONFIG_SWAP
485                 /*
486                  * Anonymous process memory has backing store?
487                  * Try to allocate it some swap space here.
488                  */
489                 if (PageAnon(page) && !PageSwapCache(page))
490                         if (!add_to_swap(page, GFP_ATOMIC))
491                                 goto activate_locked;
492 #endif /* CONFIG_SWAP */
493
494                 mapping = page_mapping(page);
495                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
496                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
497
498                 /*
499                  * The page is mapped into the page tables of one or more
500                  * processes. Try to unmap it here.
501                  */
502                 if (page_mapped(page) && mapping) {
503                         switch (try_to_unmap(page, 0)) {
504                         case SWAP_FAIL:
505                                 goto activate_locked;
506                         case SWAP_AGAIN:
507                                 goto keep_locked;
508                         case SWAP_SUCCESS:
509                                 ; /* try to free the page below */
510                         }
511                 }
512
513                 if (PageDirty(page)) {
514                         if (referenced)
515                                 goto keep_locked;
516                         if (!may_enter_fs)
517                                 goto keep_locked;
518                         if (!sc->may_writepage)
519                                 goto keep_locked;
520
521                         /* Page is dirty, try to write it out here */
522                         switch(pageout(page, mapping)) {
523                         case PAGE_KEEP:
524                                 goto keep_locked;
525                         case PAGE_ACTIVATE:
526                                 goto activate_locked;
527                         case PAGE_SUCCESS:
528                                 if (PageWriteback(page) || PageDirty(page))
529                                         goto keep;
530                                 /*
531                                  * A synchronous write - probably a ramdisk.  Go
532                                  * ahead and try to reclaim the page.
533                                  */
534                                 if (TestSetPageLocked(page))
535                                         goto keep;
536                                 if (PageDirty(page) || PageWriteback(page))
537                                         goto keep_locked;
538                                 mapping = page_mapping(page);
539                         case PAGE_CLEAN:
540                                 ; /* try to free the page below */
541                         }
542                 }
543
544                 /*
545                  * If the page has buffers, try to free the buffer mappings
546                  * associated with this page. If we succeed we try to free
547                  * the page as well.
548                  *
549                  * We do this even if the page is PageDirty().
550                  * try_to_release_page() does not perform I/O, but it is
551                  * possible for a page to have PageDirty set, but it is actually
552                  * clean (all its buffers are clean).  This happens if the
553                  * buffers were written out directly, with submit_bh(). ext3
554                  * will do this, as well as the blockdev mapping. 
555                  * try_to_release_page() will discover that cleanness and will
556                  * drop the buffers and mark the page clean - it can be freed.
557                  *
558                  * Rarely, pages can have buffers and no ->mapping.  These are
559                  * the pages which were not successfully invalidated in
560                  * truncate_complete_page().  We try to drop those buffers here
561                  * and if that worked, and the page is no longer mapped into
562                  * process address space (page_count == 1) it can be freed.
563                  * Otherwise, leave the page on the LRU so it is swappable.
564                  */
565                 if (PagePrivate(page)) {
566                         if (!try_to_release_page(page, sc->gfp_mask))
567                                 goto activate_locked;
568                         if (!mapping && page_count(page) == 1)
569                                 goto free_it;
570                 }
571
572                 if (!mapping || !remove_mapping(mapping, page))
573                         goto keep_locked;
574
575 free_it:
576                 unlock_page(page);
577                 nr_reclaimed++;
578                 if (!pagevec_add(&freed_pvec, page))
579                         __pagevec_release_nonlru(&freed_pvec);
580                 continue;
581
582 activate_locked:
583                 SetPageActive(page);
584                 pgactivate++;
585 keep_locked:
586                 unlock_page(page);
587 keep:
588                 list_add(&page->lru, &ret_pages);
589                 VM_BUG_ON(PageLRU(page));
590         }
591         list_splice(&ret_pages, page_list);
592         if (pagevec_count(&freed_pvec))
593                 __pagevec_release_nonlru(&freed_pvec);
594         count_vm_events(PGACTIVATE, pgactivate);
595         return nr_reclaimed;
596 }
597
598 /*
599  * zone->lru_lock is heavily contended.  Some of the functions that
600  * shrink the lists perform better by taking out a batch of pages
601  * and working on them outside the LRU lock.
602  *
603  * For pagecache intensive workloads, this function is the hottest
604  * spot in the kernel (apart from copy_*_user functions).
605  *
606  * Appropriate locks must be held before calling this function.
607  *
608  * @nr_to_scan: The number of pages to look through on the list.
609  * @src:        The LRU list to pull pages off.
610  * @dst:        The temp list to put pages on to.
611  * @scanned:    The number of pages that were scanned.
612  *
613  * returns how many pages were moved onto *@dst.
614  */
615 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
616                 struct list_head *src, struct list_head *dst,
617                 unsigned long *scanned)
618 {
619         unsigned long nr_taken = 0;
620         struct page *page;
621         unsigned long scan;
622
623         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
624                 struct list_head *target;
625                 page = lru_to_page(src);
626                 prefetchw_prev_lru_page(page, src, flags);
627
628                 VM_BUG_ON(!PageLRU(page));
629
630                 list_del(&page->lru);
631                 target = src;
632                 if (likely(get_page_unless_zero(page))) {
633                         /*
634                          * Be careful not to clear PageLRU until after we're
635                          * sure the page is not being freed elsewhere -- the
636                          * page release code relies on it.
637                          */
638                         ClearPageLRU(page);
639                         target = dst;
640                         nr_taken++;
641                 } /* else it is being freed elsewhere */
642
643                 list_add(&page->lru, target);
644         }
645
646         *scanned = scan;
647         return nr_taken;
648 }
649
650 /*
651  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
652  * of reclaimed pages
653  */
654 static unsigned long shrink_inactive_list(unsigned long max_scan,
655                                 struct zone *zone, struct scan_control *sc)
656 {
657         LIST_HEAD(page_list);
658         struct pagevec pvec;
659         unsigned long nr_scanned = 0;
660         unsigned long nr_reclaimed = 0;
661
662         pagevec_init(&pvec, 1);
663
664         lru_add_drain();
665         spin_lock_irq(&zone->lru_lock);
666         do {
667                 struct page *page;
668                 unsigned long nr_taken;
669                 unsigned long nr_scan;
670                 unsigned long nr_freed;
671
672                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
673                                              &zone->inactive_list,
674                                              &page_list, &nr_scan);
675                 zone->nr_inactive -= nr_taken;
676                 zone->pages_scanned += nr_scan;
677                 spin_unlock_irq(&zone->lru_lock);
678
679                 nr_scanned += nr_scan;
680                 nr_freed = shrink_page_list(&page_list, sc);
681                 nr_reclaimed += nr_freed;
682                 local_irq_disable();
683                 if (current_is_kswapd()) {
684                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
685                         __count_vm_events(KSWAPD_STEAL, nr_freed);
686                 } else
687                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
688                 __count_vm_events(PGACTIVATE, nr_freed);
689
690                 if (nr_taken == 0)
691                         goto done;
692
693                 spin_lock(&zone->lru_lock);
694                 /*
695                  * Put back any unfreeable pages.
696                  */
697                 while (!list_empty(&page_list)) {
698                         page = lru_to_page(&page_list);
699                         VM_BUG_ON(PageLRU(page));
700                         SetPageLRU(page);
701                         list_del(&page->lru);
702                         if (PageActive(page))
703                                 add_page_to_active_list(zone, page);
704                         else
705                                 add_page_to_inactive_list(zone, page);
706                         if (!pagevec_add(&pvec, page)) {
707                                 spin_unlock_irq(&zone->lru_lock);
708                                 __pagevec_release(&pvec);
709                                 spin_lock_irq(&zone->lru_lock);
710                         }
711                 }
712         } while (nr_scanned < max_scan);
713         spin_unlock(&zone->lru_lock);
714 done:
715         local_irq_enable();
716         pagevec_release(&pvec);
717         return nr_reclaimed;
718 }
719
720 static inline int zone_is_near_oom(struct zone *zone)
721 {
722         return zone->pages_scanned >= (zone->nr_active + zone->nr_inactive)*3;
723 }
724
725 /*
726  * This moves pages from the active list to the inactive list.
727  *
728  * We move them the other way if the page is referenced by one or more
729  * processes, from rmap.
730  *
731  * If the pages are mostly unmapped, the processing is fast and it is
732  * appropriate to hold zone->lru_lock across the whole operation.  But if
733  * the pages are mapped, the processing is slow (page_referenced()) so we
734  * should drop zone->lru_lock around each page.  It's impossible to balance
735  * this, so instead we remove the pages from the LRU while processing them.
736  * It is safe to rely on PG_active against the non-LRU pages in here because
737  * nobody will play with that bit on a non-LRU page.
738  *
739  * The downside is that we have to touch page->_count against each page.
740  * But we had to alter page->flags anyway.
741  */
742 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
743                                 struct scan_control *sc)
744 {
745         unsigned long pgmoved;
746         int pgdeactivate = 0;
747         unsigned long pgscanned;
748         LIST_HEAD(l_hold);      /* The pages which were snipped off */
749         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
750         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
751         struct page *page;
752         struct pagevec pvec;
753         int reclaim_mapped = 0;
754
755         if (sc->may_swap) {
756                 long mapped_ratio;
757                 long distress;
758                 long swap_tendency;
759
760                 if (zone_is_near_oom(zone))
761                         goto force_reclaim_mapped;
762
763                 /*
764                  * `distress' is a measure of how much trouble we're having
765                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
766                  */
767                 distress = 100 >> zone->prev_priority;
768
769                 /*
770                  * The point of this algorithm is to decide when to start
771                  * reclaiming mapped memory instead of just pagecache.  Work out
772                  * how much memory
773                  * is mapped.
774                  */
775                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
776                                 global_page_state(NR_ANON_PAGES)) * 100) /
777                                         vm_total_pages;
778
779                 /*
780                  * Now decide how much we really want to unmap some pages.  The
781                  * mapped ratio is downgraded - just because there's a lot of
782                  * mapped memory doesn't necessarily mean that page reclaim
783                  * isn't succeeding.
784                  *
785                  * The distress ratio is important - we don't want to start
786                  * going oom.
787                  *
788                  * A 100% value of vm_swappiness overrides this algorithm
789                  * altogether.
790                  */
791                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
792
793                 /*
794                  * Now use this metric to decide whether to start moving mapped
795                  * memory onto the inactive list.
796                  */
797                 if (swap_tendency >= 100)
798 force_reclaim_mapped:
799                         reclaim_mapped = 1;
800         }
801
802         lru_add_drain();
803         spin_lock_irq(&zone->lru_lock);
804         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
805                                     &l_hold, &pgscanned);
806         zone->pages_scanned += pgscanned;
807         zone->nr_active -= pgmoved;
808         spin_unlock_irq(&zone->lru_lock);
809
810         while (!list_empty(&l_hold)) {
811                 cond_resched();
812                 page = lru_to_page(&l_hold);
813                 list_del(&page->lru);
814                 if (page_mapped(page)) {
815                         if (!reclaim_mapped ||
816                             (total_swap_pages == 0 && PageAnon(page)) ||
817                             page_referenced(page, 0)) {
818                                 list_add(&page->lru, &l_active);
819                                 continue;
820                         }
821                 }
822                 list_add(&page->lru, &l_inactive);
823         }
824
825         pagevec_init(&pvec, 1);
826         pgmoved = 0;
827         spin_lock_irq(&zone->lru_lock);
828         while (!list_empty(&l_inactive)) {
829                 page = lru_to_page(&l_inactive);
830                 prefetchw_prev_lru_page(page, &l_inactive, flags);
831                 VM_BUG_ON(PageLRU(page));
832                 SetPageLRU(page);
833                 VM_BUG_ON(!PageActive(page));
834                 ClearPageActive(page);
835
836                 list_move(&page->lru, &zone->inactive_list);
837                 pgmoved++;
838                 if (!pagevec_add(&pvec, page)) {
839                         zone->nr_inactive += pgmoved;
840                         spin_unlock_irq(&zone->lru_lock);
841                         pgdeactivate += pgmoved;
842                         pgmoved = 0;
843                         if (buffer_heads_over_limit)
844                                 pagevec_strip(&pvec);
845                         __pagevec_release(&pvec);
846                         spin_lock_irq(&zone->lru_lock);
847                 }
848         }
849         zone->nr_inactive += pgmoved;
850         pgdeactivate += pgmoved;
851         if (buffer_heads_over_limit) {
852                 spin_unlock_irq(&zone->lru_lock);
853                 pagevec_strip(&pvec);
854                 spin_lock_irq(&zone->lru_lock);
855         }
856
857         pgmoved = 0;
858         while (!list_empty(&l_active)) {
859                 page = lru_to_page(&l_active);
860                 prefetchw_prev_lru_page(page, &l_active, flags);
861                 VM_BUG_ON(PageLRU(page));
862                 SetPageLRU(page);
863                 VM_BUG_ON(!PageActive(page));
864                 list_move(&page->lru, &zone->active_list);
865                 pgmoved++;
866                 if (!pagevec_add(&pvec, page)) {
867                         zone->nr_active += pgmoved;
868                         pgmoved = 0;
869                         spin_unlock_irq(&zone->lru_lock);
870                         __pagevec_release(&pvec);
871                         spin_lock_irq(&zone->lru_lock);
872                 }
873         }
874         zone->nr_active += pgmoved;
875
876         __count_zone_vm_events(PGREFILL, zone, pgscanned);
877         __count_vm_events(PGDEACTIVATE, pgdeactivate);
878         spin_unlock_irq(&zone->lru_lock);
879
880         pagevec_release(&pvec);
881 }
882
883 /*
884  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
885  */
886 static unsigned long shrink_zone(int priority, struct zone *zone,
887                                 struct scan_control *sc)
888 {
889         unsigned long nr_active;
890         unsigned long nr_inactive;
891         unsigned long nr_to_scan;
892         unsigned long nr_reclaimed = 0;
893
894         atomic_inc(&zone->reclaim_in_progress);
895
896         /*
897          * Add one to `nr_to_scan' just to make sure that the kernel will
898          * slowly sift through the active list.
899          */
900         zone->nr_scan_active += (zone->nr_active >> priority) + 1;
901         nr_active = zone->nr_scan_active;
902         if (nr_active >= sc->swap_cluster_max)
903                 zone->nr_scan_active = 0;
904         else
905                 nr_active = 0;
906
907         zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
908         nr_inactive = zone->nr_scan_inactive;
909         if (nr_inactive >= sc->swap_cluster_max)
910                 zone->nr_scan_inactive = 0;
911         else
912                 nr_inactive = 0;
913
914         while (nr_active || nr_inactive) {
915                 if (nr_active) {
916                         nr_to_scan = min(nr_active,
917                                         (unsigned long)sc->swap_cluster_max);
918                         nr_active -= nr_to_scan;
919                         shrink_active_list(nr_to_scan, zone, sc);
920                 }
921
922                 if (nr_inactive) {
923                         nr_to_scan = min(nr_inactive,
924                                         (unsigned long)sc->swap_cluster_max);
925                         nr_inactive -= nr_to_scan;
926                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
927                                                                 sc);
928                 }
929         }
930
931         throttle_vm_writeout();
932
933         atomic_dec(&zone->reclaim_in_progress);
934         return nr_reclaimed;
935 }
936
937 /*
938  * This is the direct reclaim path, for page-allocating processes.  We only
939  * try to reclaim pages from zones which will satisfy the caller's allocation
940  * request.
941  *
942  * We reclaim from a zone even if that zone is over pages_high.  Because:
943  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
944  *    allocation or
945  * b) The zones may be over pages_high but they must go *over* pages_high to
946  *    satisfy the `incremental min' zone defense algorithm.
947  *
948  * Returns the number of reclaimed pages.
949  *
950  * If a zone is deemed to be full of pinned pages then just give it a light
951  * scan then give up on it.
952  */
953 static unsigned long shrink_zones(int priority, struct zone **zones,
954                                         struct scan_control *sc)
955 {
956         unsigned long nr_reclaimed = 0;
957         int i;
958
959         sc->all_unreclaimable = 1;
960         for (i = 0; zones[i] != NULL; i++) {
961                 struct zone *zone = zones[i];
962
963                 if (!populated_zone(zone))
964                         continue;
965
966                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
967                         continue;
968
969                 zone->temp_priority = priority;
970                 if (zone->prev_priority > priority)
971                         zone->prev_priority = priority;
972
973                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
974                         continue;       /* Let kswapd poll it */
975
976                 sc->all_unreclaimable = 0;
977
978                 nr_reclaimed += shrink_zone(priority, zone, sc);
979         }
980         return nr_reclaimed;
981 }
982  
983 /*
984  * This is the main entry point to direct page reclaim.
985  *
986  * If a full scan of the inactive list fails to free enough memory then we
987  * are "out of memory" and something needs to be killed.
988  *
989  * If the caller is !__GFP_FS then the probability of a failure is reasonably
990  * high - the zone may be full of dirty or under-writeback pages, which this
991  * caller can't do much about.  We kick pdflush and take explicit naps in the
992  * hope that some of these pages can be written.  But if the allocating task
993  * holds filesystem locks which prevent writeout this might not work, and the
994  * allocation attempt will fail.
995  */
996 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
997 {
998         int priority;
999         int ret = 0;
1000         unsigned long total_scanned = 0;
1001         unsigned long nr_reclaimed = 0;
1002         struct reclaim_state *reclaim_state = current->reclaim_state;
1003         unsigned long lru_pages = 0;
1004         int i;
1005         struct scan_control sc = {
1006                 .gfp_mask = gfp_mask,
1007                 .may_writepage = !laptop_mode,
1008                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1009                 .may_swap = 1,
1010                 .swappiness = vm_swappiness,
1011         };
1012
1013         count_vm_event(ALLOCSTALL);
1014
1015         for (i = 0; zones[i] != NULL; i++) {
1016                 struct zone *zone = zones[i];
1017
1018                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1019                         continue;
1020
1021                 zone->temp_priority = DEF_PRIORITY;
1022                 lru_pages += zone->nr_active + zone->nr_inactive;
1023         }
1024
1025         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1026                 sc.nr_scanned = 0;
1027                 if (!priority)
1028                         disable_swap_token();
1029                 nr_reclaimed += shrink_zones(priority, zones, &sc);
1030                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1031                 if (reclaim_state) {
1032                         nr_reclaimed += reclaim_state->reclaimed_slab;
1033                         reclaim_state->reclaimed_slab = 0;
1034                 }
1035                 total_scanned += sc.nr_scanned;
1036                 if (nr_reclaimed >= sc.swap_cluster_max) {
1037                         ret = 1;
1038                         goto out;
1039                 }
1040
1041                 /*
1042                  * Try to write back as many pages as we just scanned.  This
1043                  * tends to cause slow streaming writers to write data to the
1044                  * disk smoothly, at the dirtying rate, which is nice.   But
1045                  * that's undesirable in laptop mode, where we *want* lumpy
1046                  * writeout.  So in laptop mode, write out the whole world.
1047                  */
1048                 if (total_scanned > sc.swap_cluster_max +
1049                                         sc.swap_cluster_max / 2) {
1050                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1051                         sc.may_writepage = 1;
1052                 }
1053
1054                 /* Take a nap, wait for some writeback to complete */
1055                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1056                         blk_congestion_wait(WRITE, HZ/10);
1057         }
1058         /* top priority shrink_caches still had more to do? don't OOM, then */
1059         if (!sc.all_unreclaimable)
1060                 ret = 1;
1061 out:
1062         for (i = 0; zones[i] != 0; i++) {
1063                 struct zone *zone = zones[i];
1064
1065                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1066                         continue;
1067
1068                 zone->prev_priority = zone->temp_priority;
1069         }
1070         return ret;
1071 }
1072
1073 /*
1074  * For kswapd, balance_pgdat() will work across all this node's zones until
1075  * they are all at pages_high.
1076  *
1077  * Returns the number of pages which were actually freed.
1078  *
1079  * There is special handling here for zones which are full of pinned pages.
1080  * This can happen if the pages are all mlocked, or if they are all used by
1081  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1082  * What we do is to detect the case where all pages in the zone have been
1083  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1084  * dead and from now on, only perform a short scan.  Basically we're polling
1085  * the zone for when the problem goes away.
1086  *
1087  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1088  * zones which have free_pages > pages_high, but once a zone is found to have
1089  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1090  * of the number of free pages in the lower zones.  This interoperates with
1091  * the page allocator fallback scheme to ensure that aging of pages is balanced
1092  * across the zones.
1093  */
1094 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1095 {
1096         int all_zones_ok;
1097         int priority;
1098         int i;
1099         unsigned long total_scanned;
1100         unsigned long nr_reclaimed;
1101         struct reclaim_state *reclaim_state = current->reclaim_state;
1102         struct scan_control sc = {
1103                 .gfp_mask = GFP_KERNEL,
1104                 .may_swap = 1,
1105                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1106                 .swappiness = vm_swappiness,
1107         };
1108
1109 loop_again:
1110         total_scanned = 0;
1111         nr_reclaimed = 0;
1112         sc.may_writepage = !laptop_mode;
1113         count_vm_event(PAGEOUTRUN);
1114
1115         for (i = 0; i < pgdat->nr_zones; i++) {
1116                 struct zone *zone = pgdat->node_zones + i;
1117
1118                 zone->temp_priority = DEF_PRIORITY;
1119         }
1120
1121         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1122                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1123                 unsigned long lru_pages = 0;
1124
1125                 /* The swap token gets in the way of swapout... */
1126                 if (!priority)
1127                         disable_swap_token();
1128
1129                 all_zones_ok = 1;
1130
1131                 /*
1132                  * Scan in the highmem->dma direction for the highest
1133                  * zone which needs scanning
1134                  */
1135                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1136                         struct zone *zone = pgdat->node_zones + i;
1137
1138                         if (!populated_zone(zone))
1139                                 continue;
1140
1141                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1142                                 continue;
1143
1144                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1145                                                0, 0)) {
1146                                 end_zone = i;
1147                                 goto scan;
1148                         }
1149                 }
1150                 goto out;
1151 scan:
1152                 for (i = 0; i <= end_zone; i++) {
1153                         struct zone *zone = pgdat->node_zones + i;
1154
1155                         lru_pages += zone->nr_active + zone->nr_inactive;
1156                 }
1157
1158                 /*
1159                  * Now scan the zone in the dma->highmem direction, stopping
1160                  * at the last zone which needs scanning.
1161                  *
1162                  * We do this because the page allocator works in the opposite
1163                  * direction.  This prevents the page allocator from allocating
1164                  * pages behind kswapd's direction of progress, which would
1165                  * cause too much scanning of the lower zones.
1166                  */
1167                 for (i = 0; i <= end_zone; i++) {
1168                         struct zone *zone = pgdat->node_zones + i;
1169                         int nr_slab;
1170
1171                         if (!populated_zone(zone))
1172                                 continue;
1173
1174                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1175                                 continue;
1176
1177                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1178                                                end_zone, 0))
1179                                 all_zones_ok = 0;
1180                         zone->temp_priority = priority;
1181                         if (zone->prev_priority > priority)
1182                                 zone->prev_priority = priority;
1183                         sc.nr_scanned = 0;
1184                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1185                         reclaim_state->reclaimed_slab = 0;
1186                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1187                                                 lru_pages);
1188                         nr_reclaimed += reclaim_state->reclaimed_slab;
1189                         total_scanned += sc.nr_scanned;
1190                         if (zone->all_unreclaimable)
1191                                 continue;
1192                         if (nr_slab == 0 && zone->pages_scanned >=
1193                                     (zone->nr_active + zone->nr_inactive) * 6)
1194                                 zone->all_unreclaimable = 1;
1195                         /*
1196                          * If we've done a decent amount of scanning and
1197                          * the reclaim ratio is low, start doing writepage
1198                          * even in laptop mode
1199                          */
1200                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1201                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1202                                 sc.may_writepage = 1;
1203                 }
1204                 if (all_zones_ok)
1205                         break;          /* kswapd: all done */
1206                 /*
1207                  * OK, kswapd is getting into trouble.  Take a nap, then take
1208                  * another pass across the zones.
1209                  */
1210                 if (total_scanned && priority < DEF_PRIORITY - 2)
1211                         blk_congestion_wait(WRITE, HZ/10);
1212
1213                 /*
1214                  * We do this so kswapd doesn't build up large priorities for
1215                  * example when it is freeing in parallel with allocators. It
1216                  * matches the direct reclaim path behaviour in terms of impact
1217                  * on zone->*_priority.
1218                  */
1219                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1220                         break;
1221         }
1222 out:
1223         for (i = 0; i < pgdat->nr_zones; i++) {
1224                 struct zone *zone = pgdat->node_zones + i;
1225
1226                 zone->prev_priority = zone->temp_priority;
1227         }
1228         if (!all_zones_ok) {
1229                 cond_resched();
1230                 goto loop_again;
1231         }
1232
1233         return nr_reclaimed;
1234 }
1235
1236 /*
1237  * The background pageout daemon, started as a kernel thread
1238  * from the init process. 
1239  *
1240  * This basically trickles out pages so that we have _some_
1241  * free memory available even if there is no other activity
1242  * that frees anything up. This is needed for things like routing
1243  * etc, where we otherwise might have all activity going on in
1244  * asynchronous contexts that cannot page things out.
1245  *
1246  * If there are applications that are active memory-allocators
1247  * (most normal use), this basically shouldn't matter.
1248  */
1249 static int kswapd(void *p)
1250 {
1251         unsigned long order;
1252         pg_data_t *pgdat = (pg_data_t*)p;
1253         struct task_struct *tsk = current;
1254         DEFINE_WAIT(wait);
1255         struct reclaim_state reclaim_state = {
1256                 .reclaimed_slab = 0,
1257         };
1258         cpumask_t cpumask;
1259
1260         cpumask = node_to_cpumask(pgdat->node_id);
1261         if (!cpus_empty(cpumask))
1262                 set_cpus_allowed(tsk, cpumask);
1263         current->reclaim_state = &reclaim_state;
1264
1265         /*
1266          * Tell the memory management that we're a "memory allocator",
1267          * and that if we need more memory we should get access to it
1268          * regardless (see "__alloc_pages()"). "kswapd" should
1269          * never get caught in the normal page freeing logic.
1270          *
1271          * (Kswapd normally doesn't need memory anyway, but sometimes
1272          * you need a small amount of memory in order to be able to
1273          * page out something else, and this flag essentially protects
1274          * us from recursively trying to free more memory as we're
1275          * trying to free the first piece of memory in the first place).
1276          */
1277         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1278
1279         order = 0;
1280         for ( ; ; ) {
1281                 unsigned long new_order;
1282
1283                 try_to_freeze();
1284
1285                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1286                 new_order = pgdat->kswapd_max_order;
1287                 pgdat->kswapd_max_order = 0;
1288                 if (order < new_order) {
1289                         /*
1290                          * Don't sleep if someone wants a larger 'order'
1291                          * allocation
1292                          */
1293                         order = new_order;
1294                 } else {
1295                         schedule();
1296                         order = pgdat->kswapd_max_order;
1297                 }
1298                 finish_wait(&pgdat->kswapd_wait, &wait);
1299
1300                 balance_pgdat(pgdat, order);
1301         }
1302         return 0;
1303 }
1304
1305 /*
1306  * A zone is low on free memory, so wake its kswapd task to service it.
1307  */
1308 void wakeup_kswapd(struct zone *zone, int order)
1309 {
1310         pg_data_t *pgdat;
1311
1312         if (!populated_zone(zone))
1313                 return;
1314
1315         pgdat = zone->zone_pgdat;
1316         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1317                 return;
1318         if (pgdat->kswapd_max_order < order)
1319                 pgdat->kswapd_max_order = order;
1320         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1321                 return;
1322         if (!waitqueue_active(&pgdat->kswapd_wait))
1323                 return;
1324         wake_up_interruptible(&pgdat->kswapd_wait);
1325 }
1326
1327 #ifdef CONFIG_PM
1328 /*
1329  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1330  * from LRU lists system-wide, for given pass and priority, and returns the
1331  * number of reclaimed pages
1332  *
1333  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1334  */
1335 static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1336                                       int prio, struct scan_control *sc)
1337 {
1338         struct zone *zone;
1339         unsigned long nr_to_scan, ret = 0;
1340
1341         for_each_zone(zone) {
1342
1343                 if (!populated_zone(zone))
1344                         continue;
1345
1346                 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1347                         continue;
1348
1349                 /* For pass = 0 we don't shrink the active list */
1350                 if (pass > 0) {
1351                         zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1352                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1353                                 zone->nr_scan_active = 0;
1354                                 nr_to_scan = min(nr_pages, zone->nr_active);
1355                                 shrink_active_list(nr_to_scan, zone, sc);
1356                         }
1357                 }
1358
1359                 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1360                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1361                         zone->nr_scan_inactive = 0;
1362                         nr_to_scan = min(nr_pages, zone->nr_inactive);
1363                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1364                         if (ret >= nr_pages)
1365                                 return ret;
1366                 }
1367         }
1368
1369         return ret;
1370 }
1371
1372 /*
1373  * Try to free `nr_pages' of memory, system-wide, and return the number of
1374  * freed pages.
1375  *
1376  * Rather than trying to age LRUs the aim is to preserve the overall
1377  * LRU order by reclaiming preferentially
1378  * inactive > active > active referenced > active mapped
1379  */
1380 unsigned long shrink_all_memory(unsigned long nr_pages)
1381 {
1382         unsigned long lru_pages, nr_slab;
1383         unsigned long ret = 0;
1384         int pass;
1385         struct reclaim_state reclaim_state;
1386         struct zone *zone;
1387         struct scan_control sc = {
1388                 .gfp_mask = GFP_KERNEL,
1389                 .may_swap = 0,
1390                 .swap_cluster_max = nr_pages,
1391                 .may_writepage = 1,
1392                 .swappiness = vm_swappiness,
1393         };
1394
1395         current->reclaim_state = &reclaim_state;
1396
1397         lru_pages = 0;
1398         for_each_zone(zone)
1399                 lru_pages += zone->nr_active + zone->nr_inactive;
1400
1401         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1402         /* If slab caches are huge, it's better to hit them first */
1403         while (nr_slab >= lru_pages) {
1404                 reclaim_state.reclaimed_slab = 0;
1405                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1406                 if (!reclaim_state.reclaimed_slab)
1407                         break;
1408
1409                 ret += reclaim_state.reclaimed_slab;
1410                 if (ret >= nr_pages)
1411                         goto out;
1412
1413                 nr_slab -= reclaim_state.reclaimed_slab;
1414         }
1415
1416         /*
1417          * We try to shrink LRUs in 5 passes:
1418          * 0 = Reclaim from inactive_list only
1419          * 1 = Reclaim from active list but don't reclaim mapped
1420          * 2 = 2nd pass of type 1
1421          * 3 = Reclaim mapped (normal reclaim)
1422          * 4 = 2nd pass of type 3
1423          */
1424         for (pass = 0; pass < 5; pass++) {
1425                 int prio;
1426
1427                 /* Needed for shrinking slab caches later on */
1428                 if (!lru_pages)
1429                         for_each_zone(zone) {
1430                                 lru_pages += zone->nr_active;
1431                                 lru_pages += zone->nr_inactive;
1432                         }
1433
1434                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1435                 if (pass > 2) {
1436                         sc.may_swap = 1;
1437                         sc.swappiness = 100;
1438                 }
1439
1440                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1441                         unsigned long nr_to_scan = nr_pages - ret;
1442
1443                         sc.nr_scanned = 0;
1444                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1445                         if (ret >= nr_pages)
1446                                 goto out;
1447
1448                         reclaim_state.reclaimed_slab = 0;
1449                         shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1450                         ret += reclaim_state.reclaimed_slab;
1451                         if (ret >= nr_pages)
1452                                 goto out;
1453
1454                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1455                                 blk_congestion_wait(WRITE, HZ / 10);
1456                 }
1457
1458                 lru_pages = 0;
1459         }
1460
1461         /*
1462          * If ret = 0, we could not shrink LRUs, but there may be something
1463          * in slab caches
1464          */
1465         if (!ret)
1466                 do {
1467                         reclaim_state.reclaimed_slab = 0;
1468                         shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1469                         ret += reclaim_state.reclaimed_slab;
1470                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1471
1472 out:
1473         current->reclaim_state = NULL;
1474
1475         return ret;
1476 }
1477 #endif
1478
1479 #ifdef CONFIG_HOTPLUG_CPU
1480 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1481    not required for correctness.  So if the last cpu in a node goes
1482    away, we get changed to run anywhere: as the first one comes back,
1483    restore their cpu bindings. */
1484 static int __devinit cpu_callback(struct notifier_block *nfb,
1485                                   unsigned long action, void *hcpu)
1486 {
1487         pg_data_t *pgdat;
1488         cpumask_t mask;
1489
1490         if (action == CPU_ONLINE) {
1491                 for_each_online_pgdat(pgdat) {
1492                         mask = node_to_cpumask(pgdat->node_id);
1493                         if (any_online_cpu(mask) != NR_CPUS)
1494                                 /* One of our CPUs online: restore mask */
1495                                 set_cpus_allowed(pgdat->kswapd, mask);
1496                 }
1497         }
1498         return NOTIFY_OK;
1499 }
1500 #endif /* CONFIG_HOTPLUG_CPU */
1501
1502 /*
1503  * This kswapd start function will be called by init and node-hot-add.
1504  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1505  */
1506 int kswapd_run(int nid)
1507 {
1508         pg_data_t *pgdat = NODE_DATA(nid);
1509         int ret = 0;
1510
1511         if (pgdat->kswapd)
1512                 return 0;
1513
1514         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1515         if (IS_ERR(pgdat->kswapd)) {
1516                 /* failure at boot is fatal */
1517                 BUG_ON(system_state == SYSTEM_BOOTING);
1518                 printk("Failed to start kswapd on node %d\n",nid);
1519                 ret = -1;
1520         }
1521         return ret;
1522 }
1523
1524 static int __init kswapd_init(void)
1525 {
1526         int nid;
1527
1528         swap_setup();
1529         for_each_online_node(nid)
1530                 kswapd_run(nid);
1531         hotcpu_notifier(cpu_callback, 0);
1532         return 0;
1533 }
1534
1535 module_init(kswapd_init)
1536
1537 #ifdef CONFIG_NUMA
1538 /*
1539  * Zone reclaim mode
1540  *
1541  * If non-zero call zone_reclaim when the number of free pages falls below
1542  * the watermarks.
1543  */
1544 int zone_reclaim_mode __read_mostly;
1545
1546 #define RECLAIM_OFF 0
1547 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1548 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1549 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1550
1551 /*
1552  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1553  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1554  * a zone.
1555  */
1556 #define ZONE_RECLAIM_PRIORITY 4
1557
1558 /*
1559  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1560  * occur.
1561  */
1562 int sysctl_min_unmapped_ratio = 1;
1563
1564 /*
1565  * If the number of slab pages in a zone grows beyond this percentage then
1566  * slab reclaim needs to occur.
1567  */
1568 int sysctl_min_slab_ratio = 5;
1569
1570 /*
1571  * Try to free up some pages from this zone through reclaim.
1572  */
1573 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1574 {
1575         /* Minimum pages needed in order to stay on node */
1576         const unsigned long nr_pages = 1 << order;
1577         struct task_struct *p = current;
1578         struct reclaim_state reclaim_state;
1579         int priority;
1580         unsigned long nr_reclaimed = 0;
1581         struct scan_control sc = {
1582                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1583                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1584                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1585                                         SWAP_CLUSTER_MAX),
1586                 .gfp_mask = gfp_mask,
1587                 .swappiness = vm_swappiness,
1588         };
1589         unsigned long slab_reclaimable;
1590
1591         disable_swap_token();
1592         cond_resched();
1593         /*
1594          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1595          * and we also need to be able to write out pages for RECLAIM_WRITE
1596          * and RECLAIM_SWAP.
1597          */
1598         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1599         reclaim_state.reclaimed_slab = 0;
1600         p->reclaim_state = &reclaim_state;
1601
1602         if (zone_page_state(zone, NR_FILE_PAGES) -
1603                 zone_page_state(zone, NR_FILE_MAPPED) >
1604                 zone->min_unmapped_pages) {
1605                 /*
1606                  * Free memory by calling shrink zone with increasing
1607                  * priorities until we have enough memory freed.
1608                  */
1609                 priority = ZONE_RECLAIM_PRIORITY;
1610                 do {
1611                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1612                         priority--;
1613                 } while (priority >= 0 && nr_reclaimed < nr_pages);
1614         }
1615
1616         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1617         if (slab_reclaimable > zone->min_slab_pages) {
1618                 /*
1619                  * shrink_slab() does not currently allow us to determine how
1620                  * many pages were freed in this zone. So we take the current
1621                  * number of slab pages and shake the slab until it is reduced
1622                  * by the same nr_pages that we used for reclaiming unmapped
1623                  * pages.
1624                  *
1625                  * Note that shrink_slab will free memory on all zones and may
1626                  * take a long time.
1627                  */
1628                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
1629                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1630                                 slab_reclaimable - nr_pages)
1631                         ;
1632
1633                 /*
1634                  * Update nr_reclaimed by the number of slab pages we
1635                  * reclaimed from this zone.
1636                  */
1637                 nr_reclaimed += slab_reclaimable -
1638                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1639         }
1640
1641         p->reclaim_state = NULL;
1642         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1643         return nr_reclaimed >= nr_pages;
1644 }
1645
1646 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1647 {
1648         cpumask_t mask;
1649         int node_id;
1650
1651         /*
1652          * Zone reclaim reclaims unmapped file backed pages and
1653          * slab pages if we are over the defined limits.
1654          *
1655          * A small portion of unmapped file backed pages is needed for
1656          * file I/O otherwise pages read by file I/O will be immediately
1657          * thrown out if the zone is overallocated. So we do not reclaim
1658          * if less than a specified percentage of the zone is used by
1659          * unmapped file backed pages.
1660          */
1661         if (zone_page_state(zone, NR_FILE_PAGES) -
1662             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1663             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1664                         <= zone->min_slab_pages)
1665                 return 0;
1666
1667         /*
1668          * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1669          * not have reclaimable pages and if we should not delay the allocation
1670          * then do not scan.
1671          */
1672         if (!(gfp_mask & __GFP_WAIT) ||
1673                 zone->all_unreclaimable ||
1674                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1675                 (current->flags & PF_MEMALLOC))
1676                         return 0;
1677
1678         /*
1679          * Only run zone reclaim on the local zone or on zones that do not
1680          * have associated processors. This will favor the local processor
1681          * over remote processors and spread off node memory allocations
1682          * as wide as possible.
1683          */
1684         node_id = zone_to_nid(zone);
1685         mask = node_to_cpumask(node_id);
1686         if (!cpus_empty(mask) && node_id != numa_node_id())
1687                 return 0;
1688         return __zone_reclaim(zone, gfp_mask, order);
1689 }
1690 #endif