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