428c5801d4b45cf19be318079ed23c8c55a7ccfd
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
26                                         buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
36
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
39
40 #include <linux/swapops.h>
41
42 /* possible outcome of pageout() */
43 typedef enum {
44         /* failed to write page out, page is locked */
45         PAGE_KEEP,
46         /* move page to the active list, page is locked */
47         PAGE_ACTIVATE,
48         /* page has been sent to the disk successfully, page is unlocked */
49         PAGE_SUCCESS,
50         /* page is clean and locked */
51         PAGE_CLEAN,
52 } pageout_t;
53
54 struct scan_control {
55         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
56         unsigned long nr_to_scan;
57
58         /* Incremented by the number of inactive pages that were scanned */
59         unsigned long nr_scanned;
60
61         /* Incremented by the number of pages reclaimed */
62         unsigned long nr_reclaimed;
63
64         unsigned long nr_mapped;        /* From page_state */
65
66         /* Ask shrink_caches, or shrink_zone to scan at this priority */
67         unsigned int priority;
68
69         /* This context's GFP mask */
70         gfp_t gfp_mask;
71
72         int may_writepage;
73
74         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
75          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
76          * In this context, it doesn't matter that we scan the
77          * whole list at once. */
78         int swap_cluster_max;
79 };
80
81 /*
82  * The list of shrinker callbacks used by to apply pressure to
83  * ageable caches.
84  */
85 struct shrinker {
86         shrinker_t              shrinker;
87         struct list_head        list;
88         int                     seeks;  /* seeks to recreate an obj */
89         long                    nr;     /* objs pending delete */
90 };
91
92 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
93
94 #ifdef ARCH_HAS_PREFETCH
95 #define prefetch_prev_lru_page(_page, _base, _field)                    \
96         do {                                                            \
97                 if ((_page)->lru.prev != _base) {                       \
98                         struct page *prev;                              \
99                                                                         \
100                         prev = lru_to_page(&(_page->lru));              \
101                         prefetch(&prev->_field);                        \
102                 }                                                       \
103         } while (0)
104 #else
105 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
106 #endif
107
108 #ifdef ARCH_HAS_PREFETCHW
109 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
110         do {                                                            \
111                 if ((_page)->lru.prev != _base) {                       \
112                         struct page *prev;                              \
113                                                                         \
114                         prev = lru_to_page(&(_page->lru));              \
115                         prefetchw(&prev->_field);                       \
116                 }                                                       \
117         } while (0)
118 #else
119 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
120 #endif
121
122 /*
123  * From 0 .. 100.  Higher means more swappy.
124  */
125 int vm_swappiness = 60;
126 static long total_memory;
127
128 static LIST_HEAD(shrinker_list);
129 static DECLARE_RWSEM(shrinker_rwsem);
130
131 /*
132  * Add a shrinker callback to be called from the vm
133  */
134 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
135 {
136         struct shrinker *shrinker;
137
138         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
139         if (shrinker) {
140                 shrinker->shrinker = theshrinker;
141                 shrinker->seeks = seeks;
142                 shrinker->nr = 0;
143                 down_write(&shrinker_rwsem);
144                 list_add_tail(&shrinker->list, &shrinker_list);
145                 up_write(&shrinker_rwsem);
146         }
147         return shrinker;
148 }
149 EXPORT_SYMBOL(set_shrinker);
150
151 /*
152  * Remove one
153  */
154 void remove_shrinker(struct shrinker *shrinker)
155 {
156         down_write(&shrinker_rwsem);
157         list_del(&shrinker->list);
158         up_write(&shrinker_rwsem);
159         kfree(shrinker);
160 }
161 EXPORT_SYMBOL(remove_shrinker);
162
163 #define SHRINK_BATCH 128
164 /*
165  * Call the shrink functions to age shrinkable caches
166  *
167  * Here we assume it costs one seek to replace a lru page and that it also
168  * takes a seek to recreate a cache object.  With this in mind we age equal
169  * percentages of the lru and ageable caches.  This should balance the seeks
170  * generated by these structures.
171  *
172  * If the vm encounted mapped pages on the LRU it increase the pressure on
173  * slab to avoid swapping.
174  *
175  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
176  *
177  * `lru_pages' represents the number of on-LRU pages in all the zones which
178  * are eligible for the caller's allocation attempt.  It is used for balancing
179  * slab reclaim versus page reclaim.
180  *
181  * Returns the number of slab objects which we shrunk.
182  */
183 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
184 {
185         struct shrinker *shrinker;
186         int ret = 0;
187
188         if (scanned == 0)
189                 scanned = SWAP_CLUSTER_MAX;
190
191         if (!down_read_trylock(&shrinker_rwsem))
192                 return 1;       /* Assume we'll be able to shrink next time */
193
194         list_for_each_entry(shrinker, &shrinker_list, list) {
195                 unsigned long long delta;
196                 unsigned long total_scan;
197                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
198
199                 delta = (4 * scanned) / shrinker->seeks;
200                 delta *= max_pass;
201                 do_div(delta, lru_pages + 1);
202                 shrinker->nr += delta;
203                 if (shrinker->nr < 0) {
204                         printk(KERN_ERR "%s: nr=%ld\n",
205                                         __FUNCTION__, shrinker->nr);
206                         shrinker->nr = max_pass;
207                 }
208
209                 /*
210                  * Avoid risking looping forever due to too large nr value:
211                  * never try to free more than twice the estimate number of
212                  * freeable entries.
213                  */
214                 if (shrinker->nr > max_pass * 2)
215                         shrinker->nr = max_pass * 2;
216
217                 total_scan = shrinker->nr;
218                 shrinker->nr = 0;
219
220                 while (total_scan >= SHRINK_BATCH) {
221                         long this_scan = SHRINK_BATCH;
222                         int shrink_ret;
223                         int nr_before;
224
225                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
226                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
227                         if (shrink_ret == -1)
228                                 break;
229                         if (shrink_ret < nr_before)
230                                 ret += nr_before - shrink_ret;
231                         mod_page_state(slabs_scanned, this_scan);
232                         total_scan -= this_scan;
233
234                         cond_resched();
235                 }
236
237                 shrinker->nr += total_scan;
238         }
239         up_read(&shrinker_rwsem);
240         return ret;
241 }
242
243 /* Called without lock on whether page is mapped, so answer is unstable */
244 static inline int page_mapping_inuse(struct page *page)
245 {
246         struct address_space *mapping;
247
248         /* Page is in somebody's page tables. */
249         if (page_mapped(page))
250                 return 1;
251
252         /* Be more reluctant to reclaim swapcache than pagecache */
253         if (PageSwapCache(page))
254                 return 1;
255
256         mapping = page_mapping(page);
257         if (!mapping)
258                 return 0;
259
260         /* File is mmap'd by somebody? */
261         return mapping_mapped(mapping);
262 }
263
264 static inline int is_page_cache_freeable(struct page *page)
265 {
266         return page_count(page) - !!PagePrivate(page) == 2;
267 }
268
269 static int may_write_to_queue(struct backing_dev_info *bdi)
270 {
271         if (current_is_kswapd())
272                 return 1;
273         if (current_is_pdflush())       /* This is unlikely, but why not... */
274                 return 1;
275         if (!bdi_write_congested(bdi))
276                 return 1;
277         if (bdi == current->backing_dev_info)
278                 return 1;
279         return 0;
280 }
281
282 /*
283  * We detected a synchronous write error writing a page out.  Probably
284  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
285  * fsync(), msync() or close().
286  *
287  * The tricky part is that after writepage we cannot touch the mapping: nothing
288  * prevents it from being freed up.  But we have a ref on the page and once
289  * that page is locked, the mapping is pinned.
290  *
291  * We're allowed to run sleeping lock_page() here because we know the caller has
292  * __GFP_FS.
293  */
294 static void handle_write_error(struct address_space *mapping,
295                                 struct page *page, int error)
296 {
297         lock_page(page);
298         if (page_mapping(page) == mapping) {
299                 if (error == -ENOSPC)
300                         set_bit(AS_ENOSPC, &mapping->flags);
301                 else
302                         set_bit(AS_EIO, &mapping->flags);
303         }
304         unlock_page(page);
305 }
306
307 /*
308  * pageout is called by shrink_list() for each dirty page. 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                         .nonblocking = 1,
356                         .for_reclaim = 1,
357                 };
358
359                 SetPageReclaim(page);
360                 res = mapping->a_ops->writepage(page, &wbc);
361                 if (res < 0)
362                         handle_write_error(mapping, page, res);
363                 if (res == AOP_WRITEPAGE_ACTIVATE) {
364                         ClearPageReclaim(page);
365                         return PAGE_ACTIVATE;
366                 }
367                 if (!PageWriteback(page)) {
368                         /* synchronous write or broken a_ops? */
369                         ClearPageReclaim(page);
370                 }
371
372                 return PAGE_SUCCESS;
373         }
374
375         return PAGE_CLEAN;
376 }
377
378 /*
379  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
380  */
381 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
382 {
383         LIST_HEAD(ret_pages);
384         struct pagevec freed_pvec;
385         int pgactivate = 0;
386         int reclaimed = 0;
387
388         cond_resched();
389
390         pagevec_init(&freed_pvec, 1);
391         while (!list_empty(page_list)) {
392                 struct address_space *mapping;
393                 struct page *page;
394                 int may_enter_fs;
395                 int referenced;
396
397                 cond_resched();
398
399                 page = lru_to_page(page_list);
400                 list_del(&page->lru);
401
402                 if (TestSetPageLocked(page))
403                         goto keep;
404
405                 BUG_ON(PageActive(page));
406
407                 sc->nr_scanned++;
408                 /* Double the slab pressure for mapped and swapcache pages */
409                 if (page_mapped(page) || PageSwapCache(page))
410                         sc->nr_scanned++;
411
412                 if (PageWriteback(page))
413                         goto keep_locked;
414
415                 referenced = page_referenced(page, 1);
416                 /* In active use or really unfreeable?  Activate it. */
417                 if (referenced && page_mapping_inuse(page))
418                         goto activate_locked;
419
420 #ifdef CONFIG_SWAP
421                 /*
422                  * Anonymous process memory has backing store?
423                  * Try to allocate it some swap space here.
424                  */
425                 if (PageAnon(page) && !PageSwapCache(page)) {
426                         if (!add_to_swap(page))
427                                 goto activate_locked;
428                 }
429 #endif /* CONFIG_SWAP */
430
431                 mapping = page_mapping(page);
432                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
433                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
434
435                 /*
436                  * The page is mapped into the page tables of one or more
437                  * processes. Try to unmap it here.
438                  */
439                 if (page_mapped(page) && mapping) {
440                         switch (try_to_unmap(page)) {
441                         case SWAP_FAIL:
442                                 goto activate_locked;
443                         case SWAP_AGAIN:
444                                 goto keep_locked;
445                         case SWAP_SUCCESS:
446                                 ; /* try to free the page below */
447                         }
448                 }
449
450                 if (PageDirty(page)) {
451                         if (referenced)
452                                 goto keep_locked;
453                         if (!may_enter_fs)
454                                 goto keep_locked;
455                         if (laptop_mode && !sc->may_writepage)
456                                 goto keep_locked;
457
458                         /* Page is dirty, try to write it out here */
459                         switch(pageout(page, mapping)) {
460                         case PAGE_KEEP:
461                                 goto keep_locked;
462                         case PAGE_ACTIVATE:
463                                 goto activate_locked;
464                         case PAGE_SUCCESS:
465                                 if (PageWriteback(page) || PageDirty(page))
466                                         goto keep;
467                                 /*
468                                  * A synchronous write - probably a ramdisk.  Go
469                                  * ahead and try to reclaim the page.
470                                  */
471                                 if (TestSetPageLocked(page))
472                                         goto keep;
473                                 if (PageDirty(page) || PageWriteback(page))
474                                         goto keep_locked;
475                                 mapping = page_mapping(page);
476                         case PAGE_CLEAN:
477                                 ; /* try to free the page below */
478                         }
479                 }
480
481                 /*
482                  * If the page has buffers, try to free the buffer mappings
483                  * associated with this page. If we succeed we try to free
484                  * the page as well.
485                  *
486                  * We do this even if the page is PageDirty().
487                  * try_to_release_page() does not perform I/O, but it is
488                  * possible for a page to have PageDirty set, but it is actually
489                  * clean (all its buffers are clean).  This happens if the
490                  * buffers were written out directly, with submit_bh(). ext3
491                  * will do this, as well as the blockdev mapping. 
492                  * try_to_release_page() will discover that cleanness and will
493                  * drop the buffers and mark the page clean - it can be freed.
494                  *
495                  * Rarely, pages can have buffers and no ->mapping.  These are
496                  * the pages which were not successfully invalidated in
497                  * truncate_complete_page().  We try to drop those buffers here
498                  * and if that worked, and the page is no longer mapped into
499                  * process address space (page_count == 1) it can be freed.
500                  * Otherwise, leave the page on the LRU so it is swappable.
501                  */
502                 if (PagePrivate(page)) {
503                         if (!try_to_release_page(page, sc->gfp_mask))
504                                 goto activate_locked;
505                         if (!mapping && page_count(page) == 1)
506                                 goto free_it;
507                 }
508
509                 if (!mapping)
510                         goto keep_locked;       /* truncate got there first */
511
512                 write_lock_irq(&mapping->tree_lock);
513
514                 /*
515                  * The non-racy check for busy page.  It is critical to check
516                  * PageDirty _after_ making sure that the page is freeable and
517                  * not in use by anybody.       (pagecache + us == 2)
518                  */
519                 if (unlikely(page_count(page) != 2))
520                         goto cannot_free;
521                 smp_rmb();
522                 if (unlikely(PageDirty(page)))
523                         goto cannot_free;
524
525 #ifdef CONFIG_SWAP
526                 if (PageSwapCache(page)) {
527                         swp_entry_t swap = { .val = page_private(page) };
528                         __delete_from_swap_cache(page);
529                         write_unlock_irq(&mapping->tree_lock);
530                         swap_free(swap);
531                         __put_page(page);       /* The pagecache ref */
532                         goto free_it;
533                 }
534 #endif /* CONFIG_SWAP */
535
536                 __remove_from_page_cache(page);
537                 write_unlock_irq(&mapping->tree_lock);
538                 __put_page(page);
539
540 free_it:
541                 unlock_page(page);
542                 reclaimed++;
543                 if (!pagevec_add(&freed_pvec, page))
544                         __pagevec_release_nonlru(&freed_pvec);
545                 continue;
546
547 cannot_free:
548                 write_unlock_irq(&mapping->tree_lock);
549                 goto keep_locked;
550
551 activate_locked:
552                 SetPageActive(page);
553                 pgactivate++;
554 keep_locked:
555                 unlock_page(page);
556 keep:
557                 list_add(&page->lru, &ret_pages);
558                 BUG_ON(PageLRU(page));
559         }
560         list_splice(&ret_pages, page_list);
561         if (pagevec_count(&freed_pvec))
562                 __pagevec_release_nonlru(&freed_pvec);
563         mod_page_state(pgactivate, pgactivate);
564         sc->nr_reclaimed += reclaimed;
565         return reclaimed;
566 }
567
568 /*
569  * zone->lru_lock is heavily contended.  Some of the functions that
570  * shrink the lists perform better by taking out a batch of pages
571  * and working on them outside the LRU lock.
572  *
573  * For pagecache intensive workloads, this function is the hottest
574  * spot in the kernel (apart from copy_*_user functions).
575  *
576  * Appropriate locks must be held before calling this function.
577  *
578  * @nr_to_scan: The number of pages to look through on the list.
579  * @src:        The LRU list to pull pages off.
580  * @dst:        The temp list to put pages on to.
581  * @scanned:    The number of pages that were scanned.
582  *
583  * returns how many pages were moved onto *@dst.
584  */
585 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
586                              struct list_head *dst, int *scanned)
587 {
588         int nr_taken = 0;
589         struct page *page;
590         int scan = 0;
591
592         while (scan++ < nr_to_scan && !list_empty(src)) {
593                 page = lru_to_page(src);
594                 prefetchw_prev_lru_page(page, src, flags);
595
596                 if (!TestClearPageLRU(page))
597                         BUG();
598                 list_del(&page->lru);
599                 if (get_page_testone(page)) {
600                         /*
601                          * It is being freed elsewhere
602                          */
603                         __put_page(page);
604                         SetPageLRU(page);
605                         list_add(&page->lru, src);
606                         continue;
607                 } else {
608                         list_add(&page->lru, dst);
609                         nr_taken++;
610                 }
611         }
612
613         *scanned = scan;
614         return nr_taken;
615 }
616
617 /*
618  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
619  */
620 static void shrink_cache(struct zone *zone, struct scan_control *sc)
621 {
622         LIST_HEAD(page_list);
623         struct pagevec pvec;
624         int max_scan = sc->nr_to_scan;
625
626         pagevec_init(&pvec, 1);
627
628         lru_add_drain();
629         spin_lock_irq(&zone->lru_lock);
630         while (max_scan > 0) {
631                 struct page *page;
632                 int nr_taken;
633                 int nr_scan;
634                 int nr_freed;
635
636                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
637                                              &zone->inactive_list,
638                                              &page_list, &nr_scan);
639                 zone->nr_inactive -= nr_taken;
640                 zone->pages_scanned += nr_scan;
641                 spin_unlock_irq(&zone->lru_lock);
642
643                 if (nr_taken == 0)
644                         goto done;
645
646                 max_scan -= nr_scan;
647                 nr_freed = shrink_list(&page_list, sc);
648
649                 local_irq_disable();
650                 if (current_is_kswapd()) {
651                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
652                         __mod_page_state(kswapd_steal, nr_freed);
653                 } else
654                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
655                 __mod_page_state_zone(zone, pgsteal, nr_freed);
656
657                 spin_lock(&zone->lru_lock);
658                 /*
659                  * Put back any unfreeable pages.
660                  */
661                 while (!list_empty(&page_list)) {
662                         page = lru_to_page(&page_list);
663                         if (TestSetPageLRU(page))
664                                 BUG();
665                         list_del(&page->lru);
666                         if (PageActive(page))
667                                 add_page_to_active_list(zone, page);
668                         else
669                                 add_page_to_inactive_list(zone, page);
670                         if (!pagevec_add(&pvec, page)) {
671                                 spin_unlock_irq(&zone->lru_lock);
672                                 __pagevec_release(&pvec);
673                                 spin_lock_irq(&zone->lru_lock);
674                         }
675                 }
676         }
677         spin_unlock_irq(&zone->lru_lock);
678 done:
679         pagevec_release(&pvec);
680 }
681
682 /*
683  * This moves pages from the active list to the inactive list.
684  *
685  * We move them the other way if the page is referenced by one or more
686  * processes, from rmap.
687  *
688  * If the pages are mostly unmapped, the processing is fast and it is
689  * appropriate to hold zone->lru_lock across the whole operation.  But if
690  * the pages are mapped, the processing is slow (page_referenced()) so we
691  * should drop zone->lru_lock around each page.  It's impossible to balance
692  * this, so instead we remove the pages from the LRU while processing them.
693  * It is safe to rely on PG_active against the non-LRU pages in here because
694  * nobody will play with that bit on a non-LRU page.
695  *
696  * The downside is that we have to touch page->_count against each page.
697  * But we had to alter page->flags anyway.
698  */
699 static void
700 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
701 {
702         int pgmoved;
703         int pgdeactivate = 0;
704         int pgscanned;
705         int nr_pages = sc->nr_to_scan;
706         LIST_HEAD(l_hold);      /* The pages which were snipped off */
707         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
708         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
709         struct page *page;
710         struct pagevec pvec;
711         int reclaim_mapped = 0;
712         long mapped_ratio;
713         long distress;
714         long swap_tendency;
715
716         lru_add_drain();
717         spin_lock_irq(&zone->lru_lock);
718         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
719                                     &l_hold, &pgscanned);
720         zone->pages_scanned += pgscanned;
721         zone->nr_active -= pgmoved;
722         spin_unlock_irq(&zone->lru_lock);
723
724         /*
725          * `distress' is a measure of how much trouble we're having reclaiming
726          * pages.  0 -> no problems.  100 -> great trouble.
727          */
728         distress = 100 >> zone->prev_priority;
729
730         /*
731          * The point of this algorithm is to decide when to start reclaiming
732          * mapped memory instead of just pagecache.  Work out how much memory
733          * is mapped.
734          */
735         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
736
737         /*
738          * Now decide how much we really want to unmap some pages.  The mapped
739          * ratio is downgraded - just because there's a lot of mapped memory
740          * doesn't necessarily mean that page reclaim isn't succeeding.
741          *
742          * The distress ratio is important - we don't want to start going oom.
743          *
744          * A 100% value of vm_swappiness overrides this algorithm altogether.
745          */
746         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
747
748         /*
749          * Now use this metric to decide whether to start moving mapped memory
750          * onto the inactive list.
751          */
752         if (swap_tendency >= 100)
753                 reclaim_mapped = 1;
754
755         while (!list_empty(&l_hold)) {
756                 cond_resched();
757                 page = lru_to_page(&l_hold);
758                 list_del(&page->lru);
759                 if (page_mapped(page)) {
760                         if (!reclaim_mapped ||
761                             (total_swap_pages == 0 && PageAnon(page)) ||
762                             page_referenced(page, 0)) {
763                                 list_add(&page->lru, &l_active);
764                                 continue;
765                         }
766                 }
767                 list_add(&page->lru, &l_inactive);
768         }
769
770         pagevec_init(&pvec, 1);
771         pgmoved = 0;
772         spin_lock_irq(&zone->lru_lock);
773         while (!list_empty(&l_inactive)) {
774                 page = lru_to_page(&l_inactive);
775                 prefetchw_prev_lru_page(page, &l_inactive, flags);
776                 if (TestSetPageLRU(page))
777                         BUG();
778                 if (!TestClearPageActive(page))
779                         BUG();
780                 list_move(&page->lru, &zone->inactive_list);
781                 pgmoved++;
782                 if (!pagevec_add(&pvec, page)) {
783                         zone->nr_inactive += pgmoved;
784                         spin_unlock_irq(&zone->lru_lock);
785                         pgdeactivate += pgmoved;
786                         pgmoved = 0;
787                         if (buffer_heads_over_limit)
788                                 pagevec_strip(&pvec);
789                         __pagevec_release(&pvec);
790                         spin_lock_irq(&zone->lru_lock);
791                 }
792         }
793         zone->nr_inactive += pgmoved;
794         pgdeactivate += pgmoved;
795         if (buffer_heads_over_limit) {
796                 spin_unlock_irq(&zone->lru_lock);
797                 pagevec_strip(&pvec);
798                 spin_lock_irq(&zone->lru_lock);
799         }
800
801         pgmoved = 0;
802         while (!list_empty(&l_active)) {
803                 page = lru_to_page(&l_active);
804                 prefetchw_prev_lru_page(page, &l_active, flags);
805                 if (TestSetPageLRU(page))
806                         BUG();
807                 BUG_ON(!PageActive(page));
808                 list_move(&page->lru, &zone->active_list);
809                 pgmoved++;
810                 if (!pagevec_add(&pvec, page)) {
811                         zone->nr_active += pgmoved;
812                         pgmoved = 0;
813                         spin_unlock_irq(&zone->lru_lock);
814                         __pagevec_release(&pvec);
815                         spin_lock_irq(&zone->lru_lock);
816                 }
817         }
818         zone->nr_active += pgmoved;
819         spin_unlock(&zone->lru_lock);
820
821         __mod_page_state_zone(zone, pgrefill, pgscanned);
822         __mod_page_state(pgdeactivate, pgdeactivate);
823         local_irq_enable();
824
825         pagevec_release(&pvec);
826 }
827
828 /*
829  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
830  */
831 static void
832 shrink_zone(struct zone *zone, struct scan_control *sc)
833 {
834         unsigned long nr_active;
835         unsigned long nr_inactive;
836
837         atomic_inc(&zone->reclaim_in_progress);
838
839         /*
840          * Add one to `nr_to_scan' just to make sure that the kernel will
841          * slowly sift through the active list.
842          */
843         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
844         nr_active = zone->nr_scan_active;
845         if (nr_active >= sc->swap_cluster_max)
846                 zone->nr_scan_active = 0;
847         else
848                 nr_active = 0;
849
850         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
851         nr_inactive = zone->nr_scan_inactive;
852         if (nr_inactive >= sc->swap_cluster_max)
853                 zone->nr_scan_inactive = 0;
854         else
855                 nr_inactive = 0;
856
857         while (nr_active || nr_inactive) {
858                 if (nr_active) {
859                         sc->nr_to_scan = min(nr_active,
860                                         (unsigned long)sc->swap_cluster_max);
861                         nr_active -= sc->nr_to_scan;
862                         refill_inactive_zone(zone, sc);
863                 }
864
865                 if (nr_inactive) {
866                         sc->nr_to_scan = min(nr_inactive,
867                                         (unsigned long)sc->swap_cluster_max);
868                         nr_inactive -= sc->nr_to_scan;
869                         shrink_cache(zone, sc);
870                 }
871         }
872
873         throttle_vm_writeout();
874
875         atomic_dec(&zone->reclaim_in_progress);
876 }
877
878 /*
879  * This is the direct reclaim path, for page-allocating processes.  We only
880  * try to reclaim pages from zones which will satisfy the caller's allocation
881  * request.
882  *
883  * We reclaim from a zone even if that zone is over pages_high.  Because:
884  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
885  *    allocation or
886  * b) The zones may be over pages_high but they must go *over* pages_high to
887  *    satisfy the `incremental min' zone defense algorithm.
888  *
889  * Returns the number of reclaimed pages.
890  *
891  * If a zone is deemed to be full of pinned pages then just give it a light
892  * scan then give up on it.
893  */
894 static void
895 shrink_caches(struct zone **zones, struct scan_control *sc)
896 {
897         int i;
898
899         for (i = 0; zones[i] != NULL; i++) {
900                 struct zone *zone = zones[i];
901
902                 if (!populated_zone(zone))
903                         continue;
904
905                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
906                         continue;
907
908                 zone->temp_priority = sc->priority;
909                 if (zone->prev_priority > sc->priority)
910                         zone->prev_priority = sc->priority;
911
912                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
913                         continue;       /* Let kswapd poll it */
914
915                 shrink_zone(zone, sc);
916         }
917 }
918  
919 /*
920  * This is the main entry point to direct page reclaim.
921  *
922  * If a full scan of the inactive list fails to free enough memory then we
923  * are "out of memory" and something needs to be killed.
924  *
925  * If the caller is !__GFP_FS then the probability of a failure is reasonably
926  * high - the zone may be full of dirty or under-writeback pages, which this
927  * caller can't do much about.  We kick pdflush and take explicit naps in the
928  * hope that some of these pages can be written.  But if the allocating task
929  * holds filesystem locks which prevent writeout this might not work, and the
930  * allocation attempt will fail.
931  */
932 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
933 {
934         int priority;
935         int ret = 0;
936         int total_scanned = 0, total_reclaimed = 0;
937         struct reclaim_state *reclaim_state = current->reclaim_state;
938         struct scan_control sc;
939         unsigned long lru_pages = 0;
940         int i;
941
942         sc.gfp_mask = gfp_mask;
943         sc.may_writepage = 0;
944
945         inc_page_state(allocstall);
946
947         for (i = 0; zones[i] != NULL; i++) {
948                 struct zone *zone = zones[i];
949
950                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
951                         continue;
952
953                 zone->temp_priority = DEF_PRIORITY;
954                 lru_pages += zone->nr_active + zone->nr_inactive;
955         }
956
957         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
958                 sc.nr_mapped = read_page_state(nr_mapped);
959                 sc.nr_scanned = 0;
960                 sc.nr_reclaimed = 0;
961                 sc.priority = priority;
962                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
963                 if (!priority)
964                         disable_swap_token();
965                 shrink_caches(zones, &sc);
966                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
967                 if (reclaim_state) {
968                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
969                         reclaim_state->reclaimed_slab = 0;
970                 }
971                 total_scanned += sc.nr_scanned;
972                 total_reclaimed += sc.nr_reclaimed;
973                 if (total_reclaimed >= sc.swap_cluster_max) {
974                         ret = 1;
975                         goto out;
976                 }
977
978                 /*
979                  * Try to write back as many pages as we just scanned.  This
980                  * tends to cause slow streaming writers to write data to the
981                  * disk smoothly, at the dirtying rate, which is nice.   But
982                  * that's undesirable in laptop mode, where we *want* lumpy
983                  * writeout.  So in laptop mode, write out the whole world.
984                  */
985                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
986                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
987                         sc.may_writepage = 1;
988                 }
989
990                 /* Take a nap, wait for some writeback to complete */
991                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
992                         blk_congestion_wait(WRITE, HZ/10);
993         }
994 out:
995         for (i = 0; zones[i] != 0; i++) {
996                 struct zone *zone = zones[i];
997
998                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
999                         continue;
1000
1001                 zone->prev_priority = zone->temp_priority;
1002         }
1003         return ret;
1004 }
1005
1006 /*
1007  * For kswapd, balance_pgdat() will work across all this node's zones until
1008  * they are all at pages_high.
1009  *
1010  * If `nr_pages' is non-zero then it is the number of pages which are to be
1011  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1012  * special.
1013  *
1014  * Returns the number of pages which were actually freed.
1015  *
1016  * There is special handling here for zones which are full of pinned pages.
1017  * This can happen if the pages are all mlocked, or if they are all used by
1018  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1019  * What we do is to detect the case where all pages in the zone have been
1020  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1021  * dead and from now on, only perform a short scan.  Basically we're polling
1022  * the zone for when the problem goes away.
1023  *
1024  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1025  * zones which have free_pages > pages_high, but once a zone is found to have
1026  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1027  * of the number of free pages in the lower zones.  This interoperates with
1028  * the page allocator fallback scheme to ensure that aging of pages is balanced
1029  * across the zones.
1030  */
1031 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1032 {
1033         int to_free = nr_pages;
1034         int all_zones_ok;
1035         int priority;
1036         int i;
1037         int total_scanned, total_reclaimed;
1038         struct reclaim_state *reclaim_state = current->reclaim_state;
1039         struct scan_control sc;
1040
1041 loop_again:
1042         total_scanned = 0;
1043         total_reclaimed = 0;
1044         sc.gfp_mask = GFP_KERNEL;
1045         sc.may_writepage = 0;
1046         sc.nr_mapped = read_page_state(nr_mapped);
1047
1048         inc_page_state(pageoutrun);
1049
1050         for (i = 0; i < pgdat->nr_zones; i++) {
1051                 struct zone *zone = pgdat->node_zones + i;
1052
1053                 zone->temp_priority = DEF_PRIORITY;
1054         }
1055
1056         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1057                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1058                 unsigned long lru_pages = 0;
1059
1060                 /* The swap token gets in the way of swapout... */
1061                 if (!priority)
1062                         disable_swap_token();
1063
1064                 all_zones_ok = 1;
1065
1066                 if (nr_pages == 0) {
1067                         /*
1068                          * Scan in the highmem->dma direction for the highest
1069                          * zone which needs scanning
1070                          */
1071                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1072                                 struct zone *zone = pgdat->node_zones + i;
1073
1074                                 if (!populated_zone(zone))
1075                                         continue;
1076
1077                                 if (zone->all_unreclaimable &&
1078                                                 priority != DEF_PRIORITY)
1079                                         continue;
1080
1081                                 if (!zone_watermark_ok(zone, order,
1082                                                 zone->pages_high, 0, 0)) {
1083                                         end_zone = i;
1084                                         goto scan;
1085                                 }
1086                         }
1087                         goto out;
1088                 } else {
1089                         end_zone = pgdat->nr_zones - 1;
1090                 }
1091 scan:
1092                 for (i = 0; i <= end_zone; i++) {
1093                         struct zone *zone = pgdat->node_zones + i;
1094
1095                         lru_pages += zone->nr_active + zone->nr_inactive;
1096                 }
1097
1098                 /*
1099                  * Now scan the zone in the dma->highmem direction, stopping
1100                  * at the last zone which needs scanning.
1101                  *
1102                  * We do this because the page allocator works in the opposite
1103                  * direction.  This prevents the page allocator from allocating
1104                  * pages behind kswapd's direction of progress, which would
1105                  * cause too much scanning of the lower zones.
1106                  */
1107                 for (i = 0; i <= end_zone; i++) {
1108                         struct zone *zone = pgdat->node_zones + i;
1109                         int nr_slab;
1110
1111                         if (!populated_zone(zone))
1112                                 continue;
1113
1114                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1115                                 continue;
1116
1117                         if (nr_pages == 0) {    /* Not software suspend */
1118                                 if (!zone_watermark_ok(zone, order,
1119                                                 zone->pages_high, end_zone, 0))
1120                                         all_zones_ok = 0;
1121                         }
1122                         zone->temp_priority = priority;
1123                         if (zone->prev_priority > priority)
1124                                 zone->prev_priority = priority;
1125                         sc.nr_scanned = 0;
1126                         sc.nr_reclaimed = 0;
1127                         sc.priority = priority;
1128                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1129                         atomic_inc(&zone->reclaim_in_progress);
1130                         shrink_zone(zone, &sc);
1131                         atomic_dec(&zone->reclaim_in_progress);
1132                         reclaim_state->reclaimed_slab = 0;
1133                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1134                                                 lru_pages);
1135                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1136                         total_reclaimed += sc.nr_reclaimed;
1137                         total_scanned += sc.nr_scanned;
1138                         if (zone->all_unreclaimable)
1139                                 continue;
1140                         if (nr_slab == 0 && zone->pages_scanned >=
1141                                     (zone->nr_active + zone->nr_inactive) * 4)
1142                                 zone->all_unreclaimable = 1;
1143                         /*
1144                          * If we've done a decent amount of scanning and
1145                          * the reclaim ratio is low, start doing writepage
1146                          * even in laptop mode
1147                          */
1148                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1149                             total_scanned > total_reclaimed+total_reclaimed/2)
1150                                 sc.may_writepage = 1;
1151                 }
1152                 if (nr_pages && to_free > total_reclaimed)
1153                         continue;       /* swsusp: need to do more work */
1154                 if (all_zones_ok)
1155                         break;          /* kswapd: all done */
1156                 /*
1157                  * OK, kswapd is getting into trouble.  Take a nap, then take
1158                  * another pass across the zones.
1159                  */
1160                 if (total_scanned && priority < DEF_PRIORITY - 2)
1161                         blk_congestion_wait(WRITE, HZ/10);
1162
1163                 /*
1164                  * We do this so kswapd doesn't build up large priorities for
1165                  * example when it is freeing in parallel with allocators. It
1166                  * matches the direct reclaim path behaviour in terms of impact
1167                  * on zone->*_priority.
1168                  */
1169                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1170                         break;
1171         }
1172 out:
1173         for (i = 0; i < pgdat->nr_zones; i++) {
1174                 struct zone *zone = pgdat->node_zones + i;
1175
1176                 zone->prev_priority = zone->temp_priority;
1177         }
1178         if (!all_zones_ok) {
1179                 cond_resched();
1180                 goto loop_again;
1181         }
1182
1183         return total_reclaimed;
1184 }
1185
1186 /*
1187  * The background pageout daemon, started as a kernel thread
1188  * from the init process. 
1189  *
1190  * This basically trickles out pages so that we have _some_
1191  * free memory available even if there is no other activity
1192  * that frees anything up. This is needed for things like routing
1193  * etc, where we otherwise might have all activity going on in
1194  * asynchronous contexts that cannot page things out.
1195  *
1196  * If there are applications that are active memory-allocators
1197  * (most normal use), this basically shouldn't matter.
1198  */
1199 static int kswapd(void *p)
1200 {
1201         unsigned long order;
1202         pg_data_t *pgdat = (pg_data_t*)p;
1203         struct task_struct *tsk = current;
1204         DEFINE_WAIT(wait);
1205         struct reclaim_state reclaim_state = {
1206                 .reclaimed_slab = 0,
1207         };
1208         cpumask_t cpumask;
1209
1210         daemonize("kswapd%d", pgdat->node_id);
1211         cpumask = node_to_cpumask(pgdat->node_id);
1212         if (!cpus_empty(cpumask))
1213                 set_cpus_allowed(tsk, cpumask);
1214         current->reclaim_state = &reclaim_state;
1215
1216         /*
1217          * Tell the memory management that we're a "memory allocator",
1218          * and that if we need more memory we should get access to it
1219          * regardless (see "__alloc_pages()"). "kswapd" should
1220          * never get caught in the normal page freeing logic.
1221          *
1222          * (Kswapd normally doesn't need memory anyway, but sometimes
1223          * you need a small amount of memory in order to be able to
1224          * page out something else, and this flag essentially protects
1225          * us from recursively trying to free more memory as we're
1226          * trying to free the first piece of memory in the first place).
1227          */
1228         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1229
1230         order = 0;
1231         for ( ; ; ) {
1232                 unsigned long new_order;
1233
1234                 try_to_freeze();
1235
1236                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1237                 new_order = pgdat->kswapd_max_order;
1238                 pgdat->kswapd_max_order = 0;
1239                 if (order < new_order) {
1240                         /*
1241                          * Don't sleep if someone wants a larger 'order'
1242                          * allocation
1243                          */
1244                         order = new_order;
1245                 } else {
1246                         schedule();
1247                         order = pgdat->kswapd_max_order;
1248                 }
1249                 finish_wait(&pgdat->kswapd_wait, &wait);
1250
1251                 balance_pgdat(pgdat, 0, order);
1252         }
1253         return 0;
1254 }
1255
1256 /*
1257  * A zone is low on free memory, so wake its kswapd task to service it.
1258  */
1259 void wakeup_kswapd(struct zone *zone, int order)
1260 {
1261         pg_data_t *pgdat;
1262
1263         if (!populated_zone(zone))
1264                 return;
1265
1266         pgdat = zone->zone_pgdat;
1267         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1268                 return;
1269         if (pgdat->kswapd_max_order < order)
1270                 pgdat->kswapd_max_order = order;
1271         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1272                 return;
1273         if (!waitqueue_active(&pgdat->kswapd_wait))
1274                 return;
1275         wake_up_interruptible(&pgdat->kswapd_wait);
1276 }
1277
1278 #ifdef CONFIG_PM
1279 /*
1280  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1281  * pages.
1282  */
1283 int shrink_all_memory(int nr_pages)
1284 {
1285         pg_data_t *pgdat;
1286         int nr_to_free = nr_pages;
1287         int ret = 0;
1288         struct reclaim_state reclaim_state = {
1289                 .reclaimed_slab = 0,
1290         };
1291
1292         current->reclaim_state = &reclaim_state;
1293         for_each_pgdat(pgdat) {
1294                 int freed;
1295                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1296                 ret += freed;
1297                 nr_to_free -= freed;
1298                 if (nr_to_free <= 0)
1299                         break;
1300         }
1301         current->reclaim_state = NULL;
1302         return ret;
1303 }
1304 #endif
1305
1306 #ifdef CONFIG_HOTPLUG_CPU
1307 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1308    not required for correctness.  So if the last cpu in a node goes
1309    away, we get changed to run anywhere: as the first one comes back,
1310    restore their cpu bindings. */
1311 static int __devinit cpu_callback(struct notifier_block *nfb,
1312                                   unsigned long action,
1313                                   void *hcpu)
1314 {
1315         pg_data_t *pgdat;
1316         cpumask_t mask;
1317
1318         if (action == CPU_ONLINE) {
1319                 for_each_pgdat(pgdat) {
1320                         mask = node_to_cpumask(pgdat->node_id);
1321                         if (any_online_cpu(mask) != NR_CPUS)
1322                                 /* One of our CPUs online: restore mask */
1323                                 set_cpus_allowed(pgdat->kswapd, mask);
1324                 }
1325         }
1326         return NOTIFY_OK;
1327 }
1328 #endif /* CONFIG_HOTPLUG_CPU */
1329
1330 static int __init kswapd_init(void)
1331 {
1332         pg_data_t *pgdat;
1333         swap_setup();
1334         for_each_pgdat(pgdat)
1335                 pgdat->kswapd
1336                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1337         total_memory = nr_free_pagecache_pages();
1338         hotcpu_notifier(cpu_callback, 0);
1339         return 0;
1340 }
1341
1342 module_init(kswapd_init)