Merge master.kernel.org:/home/rmk/linux-2.6-mmc
[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
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         /* Can pages be swapped as part of reclaim? */
75         int may_swap;
76
77         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
78          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
79          * In this context, it doesn't matter that we scan the
80          * whole list at once. */
81         int swap_cluster_max;
82 };
83
84 /*
85  * The list of shrinker callbacks used by to apply pressure to
86  * ageable caches.
87  */
88 struct shrinker {
89         shrinker_t              shrinker;
90         struct list_head        list;
91         int                     seeks;  /* seeks to recreate an obj */
92         long                    nr;     /* objs pending delete */
93 };
94
95 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
96
97 #ifdef ARCH_HAS_PREFETCH
98 #define prefetch_prev_lru_page(_page, _base, _field)                    \
99         do {                                                            \
100                 if ((_page)->lru.prev != _base) {                       \
101                         struct page *prev;                              \
102                                                                         \
103                         prev = lru_to_page(&(_page->lru));              \
104                         prefetch(&prev->_field);                        \
105                 }                                                       \
106         } while (0)
107 #else
108 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
109 #endif
110
111 #ifdef ARCH_HAS_PREFETCHW
112 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
113         do {                                                            \
114                 if ((_page)->lru.prev != _base) {                       \
115                         struct page *prev;                              \
116                                                                         \
117                         prev = lru_to_page(&(_page->lru));              \
118                         prefetchw(&prev->_field);                       \
119                 }                                                       \
120         } while (0)
121 #else
122 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
123 #endif
124
125 /*
126  * From 0 .. 100.  Higher means more swappy.
127  */
128 int vm_swappiness = 60;
129 static long total_memory;
130
131 static LIST_HEAD(shrinker_list);
132 static DECLARE_RWSEM(shrinker_rwsem);
133
134 /*
135  * Add a shrinker callback to be called from the vm
136  */
137 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
138 {
139         struct shrinker *shrinker;
140
141         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
142         if (shrinker) {
143                 shrinker->shrinker = theshrinker;
144                 shrinker->seeks = seeks;
145                 shrinker->nr = 0;
146                 down_write(&shrinker_rwsem);
147                 list_add_tail(&shrinker->list, &shrinker_list);
148                 up_write(&shrinker_rwsem);
149         }
150         return shrinker;
151 }
152 EXPORT_SYMBOL(set_shrinker);
153
154 /*
155  * Remove one
156  */
157 void remove_shrinker(struct shrinker *shrinker)
158 {
159         down_write(&shrinker_rwsem);
160         list_del(&shrinker->list);
161         up_write(&shrinker_rwsem);
162         kfree(shrinker);
163 }
164 EXPORT_SYMBOL(remove_shrinker);
165
166 #define SHRINK_BATCH 128
167 /*
168  * Call the shrink functions to age shrinkable caches
169  *
170  * Here we assume it costs one seek to replace a lru page and that it also
171  * takes a seek to recreate a cache object.  With this in mind we age equal
172  * percentages of the lru and ageable caches.  This should balance the seeks
173  * generated by these structures.
174  *
175  * If the vm encounted mapped pages on the LRU it increase the pressure on
176  * slab to avoid swapping.
177  *
178  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
179  *
180  * `lru_pages' represents the number of on-LRU pages in all the zones which
181  * are eligible for the caller's allocation attempt.  It is used for balancing
182  * slab reclaim versus page reclaim.
183  *
184  * Returns the number of slab objects which we shrunk.
185  */
186 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
187 {
188         struct shrinker *shrinker;
189         int ret = 0;
190
191         if (scanned == 0)
192                 scanned = SWAP_CLUSTER_MAX;
193
194         if (!down_read_trylock(&shrinker_rwsem))
195                 return 1;       /* Assume we'll be able to shrink next time */
196
197         list_for_each_entry(shrinker, &shrinker_list, list) {
198                 unsigned long long delta;
199                 unsigned long total_scan;
200                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
201
202                 delta = (4 * scanned) / shrinker->seeks;
203                 delta *= max_pass;
204                 do_div(delta, lru_pages + 1);
205                 shrinker->nr += delta;
206                 if (shrinker->nr < 0) {
207                         printk(KERN_ERR "%s: nr=%ld\n",
208                                         __FUNCTION__, shrinker->nr);
209                         shrinker->nr = max_pass;
210                 }
211
212                 /*
213                  * Avoid risking looping forever due to too large nr value:
214                  * never try to free more than twice the estimate number of
215                  * freeable entries.
216                  */
217                 if (shrinker->nr > max_pass * 2)
218                         shrinker->nr = max_pass * 2;
219
220                 total_scan = shrinker->nr;
221                 shrinker->nr = 0;
222
223                 while (total_scan >= SHRINK_BATCH) {
224                         long this_scan = SHRINK_BATCH;
225                         int shrink_ret;
226                         int nr_before;
227
228                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
229                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
230                         if (shrink_ret == -1)
231                                 break;
232                         if (shrink_ret < nr_before)
233                                 ret += nr_before - shrink_ret;
234                         mod_page_state(slabs_scanned, this_scan);
235                         total_scan -= this_scan;
236
237                         cond_resched();
238                 }
239
240                 shrinker->nr += total_scan;
241         }
242         up_read(&shrinker_rwsem);
243         return ret;
244 }
245
246 /* Called without lock on whether page is mapped, so answer is unstable */
247 static inline int page_mapping_inuse(struct page *page)
248 {
249         struct address_space *mapping;
250
251         /* Page is in somebody's page tables. */
252         if (page_mapped(page))
253                 return 1;
254
255         /* Be more reluctant to reclaim swapcache than pagecache */
256         if (PageSwapCache(page))
257                 return 1;
258
259         mapping = page_mapping(page);
260         if (!mapping)
261                 return 0;
262
263         /* File is mmap'd by somebody? */
264         return mapping_mapped(mapping);
265 }
266
267 static inline int is_page_cache_freeable(struct page *page)
268 {
269         return page_count(page) - !!PagePrivate(page) == 2;
270 }
271
272 static int may_write_to_queue(struct backing_dev_info *bdi)
273 {
274         if (current->flags & PF_SWAPWRITE)
275                 return 1;
276         if (!bdi_write_congested(bdi))
277                 return 1;
278         if (bdi == current->backing_dev_info)
279                 return 1;
280         return 0;
281 }
282
283 /*
284  * We detected a synchronous write error writing a page out.  Probably
285  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
286  * fsync(), msync() or close().
287  *
288  * The tricky part is that after writepage we cannot touch the mapping: nothing
289  * prevents it from being freed up.  But we have a ref on the page and once
290  * that page is locked, the mapping is pinned.
291  *
292  * We're allowed to run sleeping lock_page() here because we know the caller has
293  * __GFP_FS.
294  */
295 static void handle_write_error(struct address_space *mapping,
296                                 struct page *page, int error)
297 {
298         lock_page(page);
299         if (page_mapping(page) == mapping) {
300                 if (error == -ENOSPC)
301                         set_bit(AS_ENOSPC, &mapping->flags);
302                 else
303                         set_bit(AS_EIO, &mapping->flags);
304         }
305         unlock_page(page);
306 }
307
308 /*
309  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
310  */
311 static pageout_t pageout(struct page *page, struct address_space *mapping)
312 {
313         /*
314          * If the page is dirty, only perform writeback if that write
315          * will be non-blocking.  To prevent this allocation from being
316          * stalled by pagecache activity.  But note that there may be
317          * stalls if we need to run get_block().  We could test
318          * PagePrivate for that.
319          *
320          * If this process is currently in generic_file_write() against
321          * this page's queue, we can perform writeback even if that
322          * will block.
323          *
324          * If the page is swapcache, write it back even if that would
325          * block, for some throttling. This happens by accident, because
326          * swap_backing_dev_info is bust: it doesn't reflect the
327          * congestion state of the swapdevs.  Easy to fix, if needed.
328          * See swapfile.c:page_queue_congested().
329          */
330         if (!is_page_cache_freeable(page))
331                 return PAGE_KEEP;
332         if (!mapping) {
333                 /*
334                  * Some data journaling orphaned pages can have
335                  * page->mapping == NULL while being dirty with clean buffers.
336                  */
337                 if (PagePrivate(page)) {
338                         if (try_to_free_buffers(page)) {
339                                 ClearPageDirty(page);
340                                 printk("%s: orphaned page\n", __FUNCTION__);
341                                 return PAGE_CLEAN;
342                         }
343                 }
344                 return PAGE_KEEP;
345         }
346         if (mapping->a_ops->writepage == NULL)
347                 return PAGE_ACTIVATE;
348         if (!may_write_to_queue(mapping->backing_dev_info))
349                 return PAGE_KEEP;
350
351         if (clear_page_dirty_for_io(page)) {
352                 int res;
353                 struct writeback_control wbc = {
354                         .sync_mode = WB_SYNC_NONE,
355                         .nr_to_write = SWAP_CLUSTER_MAX,
356                         .nonblocking = 1,
357                         .for_reclaim = 1,
358                 };
359
360                 SetPageReclaim(page);
361                 res = mapping->a_ops->writepage(page, &wbc);
362                 if (res < 0)
363                         handle_write_error(mapping, page, res);
364                 if (res == AOP_WRITEPAGE_ACTIVATE) {
365                         ClearPageReclaim(page);
366                         return PAGE_ACTIVATE;
367                 }
368                 if (!PageWriteback(page)) {
369                         /* synchronous write or broken a_ops? */
370                         ClearPageReclaim(page);
371                 }
372
373                 return PAGE_SUCCESS;
374         }
375
376         return PAGE_CLEAN;
377 }
378
379 static int remove_mapping(struct address_space *mapping, struct page *page)
380 {
381         if (!mapping)
382                 return 0;               /* truncate got there first */
383
384         write_lock_irq(&mapping->tree_lock);
385
386         /*
387          * The non-racy check for busy page.  It is critical to check
388          * PageDirty _after_ making sure that the page is freeable and
389          * not in use by anybody.       (pagecache + us == 2)
390          */
391         if (unlikely(page_count(page) != 2))
392                 goto cannot_free;
393         smp_rmb();
394         if (unlikely(PageDirty(page)))
395                 goto cannot_free;
396
397         if (PageSwapCache(page)) {
398                 swp_entry_t swap = { .val = page_private(page) };
399                 __delete_from_swap_cache(page);
400                 write_unlock_irq(&mapping->tree_lock);
401                 swap_free(swap);
402                 __put_page(page);       /* The pagecache ref */
403                 return 1;
404         }
405
406         __remove_from_page_cache(page);
407         write_unlock_irq(&mapping->tree_lock);
408         __put_page(page);
409         return 1;
410
411 cannot_free:
412         write_unlock_irq(&mapping->tree_lock);
413         return 0;
414 }
415
416 /*
417  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
418  */
419 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
420 {
421         LIST_HEAD(ret_pages);
422         struct pagevec freed_pvec;
423         int pgactivate = 0;
424         int reclaimed = 0;
425
426         cond_resched();
427
428         pagevec_init(&freed_pvec, 1);
429         while (!list_empty(page_list)) {
430                 struct address_space *mapping;
431                 struct page *page;
432                 int may_enter_fs;
433                 int referenced;
434
435                 cond_resched();
436
437                 page = lru_to_page(page_list);
438                 list_del(&page->lru);
439
440                 if (TestSetPageLocked(page))
441                         goto keep;
442
443                 BUG_ON(PageActive(page));
444
445                 sc->nr_scanned++;
446                 /* Double the slab pressure for mapped and swapcache pages */
447                 if (page_mapped(page) || PageSwapCache(page))
448                         sc->nr_scanned++;
449
450                 if (PageWriteback(page))
451                         goto keep_locked;
452
453                 referenced = page_referenced(page, 1);
454                 /* In active use or really unfreeable?  Activate it. */
455                 if (referenced && page_mapping_inuse(page))
456                         goto activate_locked;
457
458 #ifdef CONFIG_SWAP
459                 /*
460                  * Anonymous process memory has backing store?
461                  * Try to allocate it some swap space here.
462                  */
463                 if (PageAnon(page) && !PageSwapCache(page)) {
464                         if (!sc->may_swap)
465                                 goto keep_locked;
466                         if (!add_to_swap(page, GFP_ATOMIC))
467                                 goto activate_locked;
468                 }
469 #endif /* CONFIG_SWAP */
470
471                 mapping = page_mapping(page);
472                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
473                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
474
475                 /*
476                  * The page is mapped into the page tables of one or more
477                  * processes. Try to unmap it here.
478                  */
479                 if (page_mapped(page) && mapping) {
480                         /*
481                          * No unmapping if we do not swap
482                          */
483                         if (!sc->may_swap)
484                                 goto keep_locked;
485
486                         switch (try_to_unmap(page, 0)) {
487                         case SWAP_FAIL:
488                                 goto activate_locked;
489                         case SWAP_AGAIN:
490                                 goto keep_locked;
491                         case SWAP_SUCCESS:
492                                 ; /* try to free the page below */
493                         }
494                 }
495
496                 if (PageDirty(page)) {
497                         if (referenced)
498                                 goto keep_locked;
499                         if (!may_enter_fs)
500                                 goto keep_locked;
501                         if (!sc->may_writepage)
502                                 goto keep_locked;
503
504                         /* Page is dirty, try to write it out here */
505                         switch(pageout(page, mapping)) {
506                         case PAGE_KEEP:
507                                 goto keep_locked;
508                         case PAGE_ACTIVATE:
509                                 goto activate_locked;
510                         case PAGE_SUCCESS:
511                                 if (PageWriteback(page) || PageDirty(page))
512                                         goto keep;
513                                 /*
514                                  * A synchronous write - probably a ramdisk.  Go
515                                  * ahead and try to reclaim the page.
516                                  */
517                                 if (TestSetPageLocked(page))
518                                         goto keep;
519                                 if (PageDirty(page) || PageWriteback(page))
520                                         goto keep_locked;
521                                 mapping = page_mapping(page);
522                         case PAGE_CLEAN:
523                                 ; /* try to free the page below */
524                         }
525                 }
526
527                 /*
528                  * If the page has buffers, try to free the buffer mappings
529                  * associated with this page. If we succeed we try to free
530                  * the page as well.
531                  *
532                  * We do this even if the page is PageDirty().
533                  * try_to_release_page() does not perform I/O, but it is
534                  * possible for a page to have PageDirty set, but it is actually
535                  * clean (all its buffers are clean).  This happens if the
536                  * buffers were written out directly, with submit_bh(). ext3
537                  * will do this, as well as the blockdev mapping. 
538                  * try_to_release_page() will discover that cleanness and will
539                  * drop the buffers and mark the page clean - it can be freed.
540                  *
541                  * Rarely, pages can have buffers and no ->mapping.  These are
542                  * the pages which were not successfully invalidated in
543                  * truncate_complete_page().  We try to drop those buffers here
544                  * and if that worked, and the page is no longer mapped into
545                  * process address space (page_count == 1) it can be freed.
546                  * Otherwise, leave the page on the LRU so it is swappable.
547                  */
548                 if (PagePrivate(page)) {
549                         if (!try_to_release_page(page, sc->gfp_mask))
550                                 goto activate_locked;
551                         if (!mapping && page_count(page) == 1)
552                                 goto free_it;
553                 }
554
555                 if (!remove_mapping(mapping, page))
556                         goto keep_locked;
557
558 free_it:
559                 unlock_page(page);
560                 reclaimed++;
561                 if (!pagevec_add(&freed_pvec, page))
562                         __pagevec_release_nonlru(&freed_pvec);
563                 continue;
564
565 activate_locked:
566                 SetPageActive(page);
567                 pgactivate++;
568 keep_locked:
569                 unlock_page(page);
570 keep:
571                 list_add(&page->lru, &ret_pages);
572                 BUG_ON(PageLRU(page));
573         }
574         list_splice(&ret_pages, page_list);
575         if (pagevec_count(&freed_pvec))
576                 __pagevec_release_nonlru(&freed_pvec);
577         mod_page_state(pgactivate, pgactivate);
578         sc->nr_reclaimed += reclaimed;
579         return reclaimed;
580 }
581
582 #ifdef CONFIG_MIGRATION
583 static inline void move_to_lru(struct page *page)
584 {
585         list_del(&page->lru);
586         if (PageActive(page)) {
587                 /*
588                  * lru_cache_add_active checks that
589                  * the PG_active bit is off.
590                  */
591                 ClearPageActive(page);
592                 lru_cache_add_active(page);
593         } else {
594                 lru_cache_add(page);
595         }
596         put_page(page);
597 }
598
599 /*
600  * Add isolated pages on the list back to the LRU.
601  *
602  * returns the number of pages put back.
603  */
604 int putback_lru_pages(struct list_head *l)
605 {
606         struct page *page;
607         struct page *page2;
608         int count = 0;
609
610         list_for_each_entry_safe(page, page2, l, lru) {
611                 move_to_lru(page);
612                 count++;
613         }
614         return count;
615 }
616
617 /*
618  * Non migratable page
619  */
620 int fail_migrate_page(struct page *newpage, struct page *page)
621 {
622         return -EIO;
623 }
624 EXPORT_SYMBOL(fail_migrate_page);
625
626 /*
627  * swapout a single page
628  * page is locked upon entry, unlocked on exit
629  */
630 static int swap_page(struct page *page)
631 {
632         struct address_space *mapping = page_mapping(page);
633
634         if (page_mapped(page) && mapping)
635                 if (try_to_unmap(page, 0) != SWAP_SUCCESS)
636                         goto unlock_retry;
637
638         if (PageDirty(page)) {
639                 /* Page is dirty, try to write it out here */
640                 switch(pageout(page, mapping)) {
641                 case PAGE_KEEP:
642                 case PAGE_ACTIVATE:
643                         goto unlock_retry;
644
645                 case PAGE_SUCCESS:
646                         goto retry;
647
648                 case PAGE_CLEAN:
649                         ; /* try to free the page below */
650                 }
651         }
652
653         if (PagePrivate(page)) {
654                 if (!try_to_release_page(page, GFP_KERNEL) ||
655                     (!mapping && page_count(page) == 1))
656                         goto unlock_retry;
657         }
658
659         if (remove_mapping(mapping, page)) {
660                 /* Success */
661                 unlock_page(page);
662                 return 0;
663         }
664
665 unlock_retry:
666         unlock_page(page);
667
668 retry:
669         return -EAGAIN;
670 }
671 EXPORT_SYMBOL(swap_page);
672
673 /*
674  * Page migration was first developed in the context of the memory hotplug
675  * project. The main authors of the migration code are:
676  *
677  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
678  * Hirokazu Takahashi <taka@valinux.co.jp>
679  * Dave Hansen <haveblue@us.ibm.com>
680  * Christoph Lameter <clameter@sgi.com>
681  */
682
683 /*
684  * Remove references for a page and establish the new page with the correct
685  * basic settings to be able to stop accesses to the page.
686  */
687 int migrate_page_remove_references(struct page *newpage,
688                                 struct page *page, int nr_refs)
689 {
690         struct address_space *mapping = page_mapping(page);
691         struct page **radix_pointer;
692
693         /*
694          * Avoid doing any of the following work if the page count
695          * indicates that the page is in use or truncate has removed
696          * the page.
697          */
698         if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
699                 return 1;
700
701         /*
702          * Establish swap ptes for anonymous pages or destroy pte
703          * maps for files.
704          *
705          * In order to reestablish file backed mappings the fault handlers
706          * will take the radix tree_lock which may then be used to stop
707          * processses from accessing this page until the new page is ready.
708          *
709          * A process accessing via a swap pte (an anonymous page) will take a
710          * page_lock on the old page which will block the process until the
711          * migration attempt is complete. At that time the PageSwapCache bit
712          * will be examined. If the page was migrated then the PageSwapCache
713          * bit will be clear and the operation to retrieve the page will be
714          * retried which will find the new page in the radix tree. Then a new
715          * direct mapping may be generated based on the radix tree contents.
716          *
717          * If the page was not migrated then the PageSwapCache bit
718          * is still set and the operation may continue.
719          */
720         try_to_unmap(page, 1);
721
722         /*
723          * Give up if we were unable to remove all mappings.
724          */
725         if (page_mapcount(page))
726                 return 1;
727
728         write_lock_irq(&mapping->tree_lock);
729
730         radix_pointer = (struct page **)radix_tree_lookup_slot(
731                                                 &mapping->page_tree,
732                                                 page_index(page));
733
734         if (!page_mapping(page) || page_count(page) != nr_refs ||
735                         *radix_pointer != page) {
736                 write_unlock_irq(&mapping->tree_lock);
737                 return 1;
738         }
739
740         /*
741          * Now we know that no one else is looking at the page.
742          *
743          * Certain minimal information about a page must be available
744          * in order for other subsystems to properly handle the page if they
745          * find it through the radix tree update before we are finished
746          * copying the page.
747          */
748         get_page(newpage);
749         newpage->index = page->index;
750         newpage->mapping = page->mapping;
751         if (PageSwapCache(page)) {
752                 SetPageSwapCache(newpage);
753                 set_page_private(newpage, page_private(page));
754         }
755
756         *radix_pointer = newpage;
757         __put_page(page);
758         write_unlock_irq(&mapping->tree_lock);
759
760         return 0;
761 }
762 EXPORT_SYMBOL(migrate_page_remove_references);
763
764 /*
765  * Copy the page to its new location
766  */
767 void migrate_page_copy(struct page *newpage, struct page *page)
768 {
769         copy_highpage(newpage, page);
770
771         if (PageError(page))
772                 SetPageError(newpage);
773         if (PageReferenced(page))
774                 SetPageReferenced(newpage);
775         if (PageUptodate(page))
776                 SetPageUptodate(newpage);
777         if (PageActive(page))
778                 SetPageActive(newpage);
779         if (PageChecked(page))
780                 SetPageChecked(newpage);
781         if (PageMappedToDisk(page))
782                 SetPageMappedToDisk(newpage);
783
784         if (PageDirty(page)) {
785                 clear_page_dirty_for_io(page);
786                 set_page_dirty(newpage);
787         }
788
789         ClearPageSwapCache(page);
790         ClearPageActive(page);
791         ClearPagePrivate(page);
792         set_page_private(page, 0);
793         page->mapping = NULL;
794
795         /*
796          * If any waiters have accumulated on the new page then
797          * wake them up.
798          */
799         if (PageWriteback(newpage))
800                 end_page_writeback(newpage);
801 }
802 EXPORT_SYMBOL(migrate_page_copy);
803
804 /*
805  * Common logic to directly migrate a single page suitable for
806  * pages that do not use PagePrivate.
807  *
808  * Pages are locked upon entry and exit.
809  */
810 int migrate_page(struct page *newpage, struct page *page)
811 {
812         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
813
814         if (migrate_page_remove_references(newpage, page, 2))
815                 return -EAGAIN;
816
817         migrate_page_copy(newpage, page);
818
819         /*
820          * Remove auxiliary swap entries and replace
821          * them with real ptes.
822          *
823          * Note that a real pte entry will allow processes that are not
824          * waiting on the page lock to use the new page via the page tables
825          * before the new page is unlocked.
826          */
827         remove_from_swap(newpage);
828         return 0;
829 }
830 EXPORT_SYMBOL(migrate_page);
831
832 /*
833  * migrate_pages
834  *
835  * Two lists are passed to this function. The first list
836  * contains the pages isolated from the LRU to be migrated.
837  * The second list contains new pages that the pages isolated
838  * can be moved to. If the second list is NULL then all
839  * pages are swapped out.
840  *
841  * The function returns after 10 attempts or if no pages
842  * are movable anymore because t has become empty
843  * or no retryable pages exist anymore.
844  *
845  * Return: Number of pages not migrated when "to" ran empty.
846  */
847 int migrate_pages(struct list_head *from, struct list_head *to,
848                   struct list_head *moved, struct list_head *failed)
849 {
850         int retry;
851         int nr_failed = 0;
852         int pass = 0;
853         struct page *page;
854         struct page *page2;
855         int swapwrite = current->flags & PF_SWAPWRITE;
856         int rc;
857
858         if (!swapwrite)
859                 current->flags |= PF_SWAPWRITE;
860
861 redo:
862         retry = 0;
863
864         list_for_each_entry_safe(page, page2, from, lru) {
865                 struct page *newpage = NULL;
866                 struct address_space *mapping;
867
868                 cond_resched();
869
870                 rc = 0;
871                 if (page_count(page) == 1)
872                         /* page was freed from under us. So we are done. */
873                         goto next;
874
875                 if (to && list_empty(to))
876                         break;
877
878                 /*
879                  * Skip locked pages during the first two passes to give the
880                  * functions holding the lock time to release the page. Later we
881                  * use lock_page() to have a higher chance of acquiring the
882                  * lock.
883                  */
884                 rc = -EAGAIN;
885                 if (pass > 2)
886                         lock_page(page);
887                 else
888                         if (TestSetPageLocked(page))
889                                 goto next;
890
891                 /*
892                  * Only wait on writeback if we have already done a pass where
893                  * we we may have triggered writeouts for lots of pages.
894                  */
895                 if (pass > 0) {
896                         wait_on_page_writeback(page);
897                 } else {
898                         if (PageWriteback(page))
899                                 goto unlock_page;
900                 }
901
902                 /*
903                  * Anonymous pages must have swap cache references otherwise
904                  * the information contained in the page maps cannot be
905                  * preserved.
906                  */
907                 if (PageAnon(page) && !PageSwapCache(page)) {
908                         if (!add_to_swap(page, GFP_KERNEL)) {
909                                 rc = -ENOMEM;
910                                 goto unlock_page;
911                         }
912                 }
913
914                 if (!to) {
915                         rc = swap_page(page);
916                         goto next;
917                 }
918
919                 newpage = lru_to_page(to);
920                 lock_page(newpage);
921
922                 /*
923                  * Pages are properly locked and writeback is complete.
924                  * Try to migrate the page.
925                  */
926                 mapping = page_mapping(page);
927                 if (!mapping)
928                         goto unlock_both;
929
930                 if (mapping->a_ops->migratepage) {
931                         rc = mapping->a_ops->migratepage(newpage, page);
932                         goto unlock_both;
933                 }
934
935                 /*
936                  * Trigger writeout if page is dirty
937                  */
938                 if (PageDirty(page)) {
939                         switch (pageout(page, mapping)) {
940                         case PAGE_KEEP:
941                         case PAGE_ACTIVATE:
942                                 goto unlock_both;
943
944                         case PAGE_SUCCESS:
945                                 unlock_page(newpage);
946                                 goto next;
947
948                         case PAGE_CLEAN:
949                                 ; /* try to migrate the page below */
950                         }
951                 }
952                 /*
953                  * If we have no buffer or can release the buffer
954                  * then do a simple migration.
955                  */
956                 if (!page_has_buffers(page) ||
957                     try_to_release_page(page, GFP_KERNEL)) {
958                         rc = migrate_page(newpage, page);
959                         goto unlock_both;
960                 }
961
962                 /*
963                  * On early passes with mapped pages simply
964                  * retry. There may be a lock held for some
965                  * buffers that may go away. Later
966                  * swap them out.
967                  */
968                 if (pass > 4) {
969                         unlock_page(newpage);
970                         newpage = NULL;
971                         rc = swap_page(page);
972                         goto next;
973                 }
974
975 unlock_both:
976                 unlock_page(newpage);
977
978 unlock_page:
979                 unlock_page(page);
980
981 next:
982                 if (rc == -EAGAIN) {
983                         retry++;
984                 } else if (rc) {
985                         /* Permanent failure */
986                         list_move(&page->lru, failed);
987                         nr_failed++;
988                 } else {
989                         if (newpage) {
990                                 /* Successful migration. Return page to LRU */
991                                 move_to_lru(newpage);
992                         }
993                         list_move(&page->lru, moved);
994                 }
995         }
996         if (retry && pass++ < 10)
997                 goto redo;
998
999         if (!swapwrite)
1000                 current->flags &= ~PF_SWAPWRITE;
1001
1002         return nr_failed + retry;
1003 }
1004
1005 /*
1006  * Isolate one page from the LRU lists and put it on the
1007  * indicated list with elevated refcount.
1008  *
1009  * Result:
1010  *  0 = page not on LRU list
1011  *  1 = page removed from LRU list and added to the specified list.
1012  */
1013 int isolate_lru_page(struct page *page)
1014 {
1015         int ret = 0;
1016
1017         if (PageLRU(page)) {
1018                 struct zone *zone = page_zone(page);
1019                 spin_lock_irq(&zone->lru_lock);
1020                 if (TestClearPageLRU(page)) {
1021                         ret = 1;
1022                         get_page(page);
1023                         if (PageActive(page))
1024                                 del_page_from_active_list(zone, page);
1025                         else
1026                                 del_page_from_inactive_list(zone, page);
1027                 }
1028                 spin_unlock_irq(&zone->lru_lock);
1029         }
1030
1031         return ret;
1032 }
1033 #endif
1034
1035 /*
1036  * zone->lru_lock is heavily contended.  Some of the functions that
1037  * shrink the lists perform better by taking out a batch of pages
1038  * and working on them outside the LRU lock.
1039  *
1040  * For pagecache intensive workloads, this function is the hottest
1041  * spot in the kernel (apart from copy_*_user functions).
1042  *
1043  * Appropriate locks must be held before calling this function.
1044  *
1045  * @nr_to_scan: The number of pages to look through on the list.
1046  * @src:        The LRU list to pull pages off.
1047  * @dst:        The temp list to put pages on to.
1048  * @scanned:    The number of pages that were scanned.
1049  *
1050  * returns how many pages were moved onto *@dst.
1051  */
1052 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
1053                              struct list_head *dst, int *scanned)
1054 {
1055         int nr_taken = 0;
1056         struct page *page;
1057         int scan = 0;
1058
1059         while (scan++ < nr_to_scan && !list_empty(src)) {
1060                 page = lru_to_page(src);
1061                 prefetchw_prev_lru_page(page, src, flags);
1062
1063                 if (!TestClearPageLRU(page))
1064                         BUG();
1065                 list_del(&page->lru);
1066                 if (get_page_testone(page)) {
1067                         /*
1068                          * It is being freed elsewhere
1069                          */
1070                         __put_page(page);
1071                         SetPageLRU(page);
1072                         list_add(&page->lru, src);
1073                         continue;
1074                 } else {
1075                         list_add(&page->lru, dst);
1076                         nr_taken++;
1077                 }
1078         }
1079
1080         *scanned = scan;
1081         return nr_taken;
1082 }
1083
1084 /*
1085  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1086  */
1087 static void shrink_cache(struct zone *zone, struct scan_control *sc)
1088 {
1089         LIST_HEAD(page_list);
1090         struct pagevec pvec;
1091         int max_scan = sc->nr_to_scan;
1092
1093         pagevec_init(&pvec, 1);
1094
1095         lru_add_drain();
1096         spin_lock_irq(&zone->lru_lock);
1097         while (max_scan > 0) {
1098                 struct page *page;
1099                 int nr_taken;
1100                 int nr_scan;
1101                 int nr_freed;
1102
1103                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
1104                                              &zone->inactive_list,
1105                                              &page_list, &nr_scan);
1106                 zone->nr_inactive -= nr_taken;
1107                 zone->pages_scanned += nr_scan;
1108                 spin_unlock_irq(&zone->lru_lock);
1109
1110                 if (nr_taken == 0)
1111                         goto done;
1112
1113                 max_scan -= nr_scan;
1114                 nr_freed = shrink_list(&page_list, sc);
1115
1116                 local_irq_disable();
1117                 if (current_is_kswapd()) {
1118                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
1119                         __mod_page_state(kswapd_steal, nr_freed);
1120                 } else
1121                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
1122                 __mod_page_state_zone(zone, pgsteal, nr_freed);
1123
1124                 spin_lock(&zone->lru_lock);
1125                 /*
1126                  * Put back any unfreeable pages.
1127                  */
1128                 while (!list_empty(&page_list)) {
1129                         page = lru_to_page(&page_list);
1130                         if (TestSetPageLRU(page))
1131                                 BUG();
1132                         list_del(&page->lru);
1133                         if (PageActive(page))
1134                                 add_page_to_active_list(zone, page);
1135                         else
1136                                 add_page_to_inactive_list(zone, page);
1137                         if (!pagevec_add(&pvec, page)) {
1138                                 spin_unlock_irq(&zone->lru_lock);
1139                                 __pagevec_release(&pvec);
1140                                 spin_lock_irq(&zone->lru_lock);
1141                         }
1142                 }
1143         }
1144         spin_unlock_irq(&zone->lru_lock);
1145 done:
1146         pagevec_release(&pvec);
1147 }
1148
1149 /*
1150  * This moves pages from the active list to the inactive list.
1151  *
1152  * We move them the other way if the page is referenced by one or more
1153  * processes, from rmap.
1154  *
1155  * If the pages are mostly unmapped, the processing is fast and it is
1156  * appropriate to hold zone->lru_lock across the whole operation.  But if
1157  * the pages are mapped, the processing is slow (page_referenced()) so we
1158  * should drop zone->lru_lock around each page.  It's impossible to balance
1159  * this, so instead we remove the pages from the LRU while processing them.
1160  * It is safe to rely on PG_active against the non-LRU pages in here because
1161  * nobody will play with that bit on a non-LRU page.
1162  *
1163  * The downside is that we have to touch page->_count against each page.
1164  * But we had to alter page->flags anyway.
1165  */
1166 static void
1167 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
1168 {
1169         int pgmoved;
1170         int pgdeactivate = 0;
1171         int pgscanned;
1172         int nr_pages = sc->nr_to_scan;
1173         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1174         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1175         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1176         struct page *page;
1177         struct pagevec pvec;
1178         int reclaim_mapped = 0;
1179         long mapped_ratio;
1180         long distress;
1181         long swap_tendency;
1182
1183         lru_add_drain();
1184         spin_lock_irq(&zone->lru_lock);
1185         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1186                                     &l_hold, &pgscanned);
1187         zone->pages_scanned += pgscanned;
1188         zone->nr_active -= pgmoved;
1189         spin_unlock_irq(&zone->lru_lock);
1190
1191         /*
1192          * `distress' is a measure of how much trouble we're having reclaiming
1193          * pages.  0 -> no problems.  100 -> great trouble.
1194          */
1195         distress = 100 >> zone->prev_priority;
1196
1197         /*
1198          * The point of this algorithm is to decide when to start reclaiming
1199          * mapped memory instead of just pagecache.  Work out how much memory
1200          * is mapped.
1201          */
1202         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
1203
1204         /*
1205          * Now decide how much we really want to unmap some pages.  The mapped
1206          * ratio is downgraded - just because there's a lot of mapped memory
1207          * doesn't necessarily mean that page reclaim isn't succeeding.
1208          *
1209          * The distress ratio is important - we don't want to start going oom.
1210          *
1211          * A 100% value of vm_swappiness overrides this algorithm altogether.
1212          */
1213         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
1214
1215         /*
1216          * Now use this metric to decide whether to start moving mapped memory
1217          * onto the inactive list.
1218          */
1219         if (swap_tendency >= 100)
1220                 reclaim_mapped = 1;
1221
1222         while (!list_empty(&l_hold)) {
1223                 cond_resched();
1224                 page = lru_to_page(&l_hold);
1225                 list_del(&page->lru);
1226                 if (page_mapped(page)) {
1227                         if (!reclaim_mapped ||
1228                             (total_swap_pages == 0 && PageAnon(page)) ||
1229                             page_referenced(page, 0)) {
1230                                 list_add(&page->lru, &l_active);
1231                                 continue;
1232                         }
1233                 }
1234                 list_add(&page->lru, &l_inactive);
1235         }
1236
1237         pagevec_init(&pvec, 1);
1238         pgmoved = 0;
1239         spin_lock_irq(&zone->lru_lock);
1240         while (!list_empty(&l_inactive)) {
1241                 page = lru_to_page(&l_inactive);
1242                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1243                 if (TestSetPageLRU(page))
1244                         BUG();
1245                 if (!TestClearPageActive(page))
1246                         BUG();
1247                 list_move(&page->lru, &zone->inactive_list);
1248                 pgmoved++;
1249                 if (!pagevec_add(&pvec, page)) {
1250                         zone->nr_inactive += pgmoved;
1251                         spin_unlock_irq(&zone->lru_lock);
1252                         pgdeactivate += pgmoved;
1253                         pgmoved = 0;
1254                         if (buffer_heads_over_limit)
1255                                 pagevec_strip(&pvec);
1256                         __pagevec_release(&pvec);
1257                         spin_lock_irq(&zone->lru_lock);
1258                 }
1259         }
1260         zone->nr_inactive += pgmoved;
1261         pgdeactivate += pgmoved;
1262         if (buffer_heads_over_limit) {
1263                 spin_unlock_irq(&zone->lru_lock);
1264                 pagevec_strip(&pvec);
1265                 spin_lock_irq(&zone->lru_lock);
1266         }
1267
1268         pgmoved = 0;
1269         while (!list_empty(&l_active)) {
1270                 page = lru_to_page(&l_active);
1271                 prefetchw_prev_lru_page(page, &l_active, flags);
1272                 if (TestSetPageLRU(page))
1273                         BUG();
1274                 BUG_ON(!PageActive(page));
1275                 list_move(&page->lru, &zone->active_list);
1276                 pgmoved++;
1277                 if (!pagevec_add(&pvec, page)) {
1278                         zone->nr_active += pgmoved;
1279                         pgmoved = 0;
1280                         spin_unlock_irq(&zone->lru_lock);
1281                         __pagevec_release(&pvec);
1282                         spin_lock_irq(&zone->lru_lock);
1283                 }
1284         }
1285         zone->nr_active += pgmoved;
1286         spin_unlock(&zone->lru_lock);
1287
1288         __mod_page_state_zone(zone, pgrefill, pgscanned);
1289         __mod_page_state(pgdeactivate, pgdeactivate);
1290         local_irq_enable();
1291
1292         pagevec_release(&pvec);
1293 }
1294
1295 /*
1296  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1297  */
1298 static void
1299 shrink_zone(struct zone *zone, struct scan_control *sc)
1300 {
1301         unsigned long nr_active;
1302         unsigned long nr_inactive;
1303
1304         atomic_inc(&zone->reclaim_in_progress);
1305
1306         /*
1307          * Add one to `nr_to_scan' just to make sure that the kernel will
1308          * slowly sift through the active list.
1309          */
1310         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1311         nr_active = zone->nr_scan_active;
1312         if (nr_active >= sc->swap_cluster_max)
1313                 zone->nr_scan_active = 0;
1314         else
1315                 nr_active = 0;
1316
1317         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1318         nr_inactive = zone->nr_scan_inactive;
1319         if (nr_inactive >= sc->swap_cluster_max)
1320                 zone->nr_scan_inactive = 0;
1321         else
1322                 nr_inactive = 0;
1323
1324         while (nr_active || nr_inactive) {
1325                 if (nr_active) {
1326                         sc->nr_to_scan = min(nr_active,
1327                                         (unsigned long)sc->swap_cluster_max);
1328                         nr_active -= sc->nr_to_scan;
1329                         refill_inactive_zone(zone, sc);
1330                 }
1331
1332                 if (nr_inactive) {
1333                         sc->nr_to_scan = min(nr_inactive,
1334                                         (unsigned long)sc->swap_cluster_max);
1335                         nr_inactive -= sc->nr_to_scan;
1336                         shrink_cache(zone, sc);
1337                 }
1338         }
1339
1340         throttle_vm_writeout();
1341
1342         atomic_dec(&zone->reclaim_in_progress);
1343 }
1344
1345 /*
1346  * This is the direct reclaim path, for page-allocating processes.  We only
1347  * try to reclaim pages from zones which will satisfy the caller's allocation
1348  * request.
1349  *
1350  * We reclaim from a zone even if that zone is over pages_high.  Because:
1351  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1352  *    allocation or
1353  * b) The zones may be over pages_high but they must go *over* pages_high to
1354  *    satisfy the `incremental min' zone defense algorithm.
1355  *
1356  * Returns the number of reclaimed pages.
1357  *
1358  * If a zone is deemed to be full of pinned pages then just give it a light
1359  * scan then give up on it.
1360  */
1361 static void
1362 shrink_caches(struct zone **zones, struct scan_control *sc)
1363 {
1364         int i;
1365
1366         for (i = 0; zones[i] != NULL; i++) {
1367                 struct zone *zone = zones[i];
1368
1369                 if (!populated_zone(zone))
1370                         continue;
1371
1372                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1373                         continue;
1374
1375                 zone->temp_priority = sc->priority;
1376                 if (zone->prev_priority > sc->priority)
1377                         zone->prev_priority = sc->priority;
1378
1379                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1380                         continue;       /* Let kswapd poll it */
1381
1382                 shrink_zone(zone, sc);
1383         }
1384 }
1385  
1386 /*
1387  * This is the main entry point to direct page reclaim.
1388  *
1389  * If a full scan of the inactive list fails to free enough memory then we
1390  * are "out of memory" and something needs to be killed.
1391  *
1392  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1393  * high - the zone may be full of dirty or under-writeback pages, which this
1394  * caller can't do much about.  We kick pdflush and take explicit naps in the
1395  * hope that some of these pages can be written.  But if the allocating task
1396  * holds filesystem locks which prevent writeout this might not work, and the
1397  * allocation attempt will fail.
1398  */
1399 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1400 {
1401         int priority;
1402         int ret = 0;
1403         int total_scanned = 0, total_reclaimed = 0;
1404         struct reclaim_state *reclaim_state = current->reclaim_state;
1405         struct scan_control sc;
1406         unsigned long lru_pages = 0;
1407         int i;
1408
1409         sc.gfp_mask = gfp_mask;
1410         sc.may_writepage = !laptop_mode;
1411         sc.may_swap = 1;
1412
1413         inc_page_state(allocstall);
1414
1415         for (i = 0; zones[i] != NULL; i++) {
1416                 struct zone *zone = zones[i];
1417
1418                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1419                         continue;
1420
1421                 zone->temp_priority = DEF_PRIORITY;
1422                 lru_pages += zone->nr_active + zone->nr_inactive;
1423         }
1424
1425         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1426                 sc.nr_mapped = read_page_state(nr_mapped);
1427                 sc.nr_scanned = 0;
1428                 sc.nr_reclaimed = 0;
1429                 sc.priority = priority;
1430                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1431                 if (!priority)
1432                         disable_swap_token();
1433                 shrink_caches(zones, &sc);
1434                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1435                 if (reclaim_state) {
1436                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1437                         reclaim_state->reclaimed_slab = 0;
1438                 }
1439                 total_scanned += sc.nr_scanned;
1440                 total_reclaimed += sc.nr_reclaimed;
1441                 if (total_reclaimed >= sc.swap_cluster_max) {
1442                         ret = 1;
1443                         goto out;
1444                 }
1445
1446                 /*
1447                  * Try to write back as many pages as we just scanned.  This
1448                  * tends to cause slow streaming writers to write data to the
1449                  * disk smoothly, at the dirtying rate, which is nice.   But
1450                  * that's undesirable in laptop mode, where we *want* lumpy
1451                  * writeout.  So in laptop mode, write out the whole world.
1452                  */
1453                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1454                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1455                         sc.may_writepage = 1;
1456                 }
1457
1458                 /* Take a nap, wait for some writeback to complete */
1459                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1460                         blk_congestion_wait(WRITE, HZ/10);
1461         }
1462 out:
1463         for (i = 0; zones[i] != 0; i++) {
1464                 struct zone *zone = zones[i];
1465
1466                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1467                         continue;
1468
1469                 zone->prev_priority = zone->temp_priority;
1470         }
1471         return ret;
1472 }
1473
1474 /*
1475  * For kswapd, balance_pgdat() will work across all this node's zones until
1476  * they are all at pages_high.
1477  *
1478  * If `nr_pages' is non-zero then it is the number of pages which are to be
1479  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1480  * special.
1481  *
1482  * Returns the number of pages which were actually freed.
1483  *
1484  * There is special handling here for zones which are full of pinned pages.
1485  * This can happen if the pages are all mlocked, or if they are all used by
1486  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1487  * What we do is to detect the case where all pages in the zone have been
1488  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1489  * dead and from now on, only perform a short scan.  Basically we're polling
1490  * the zone for when the problem goes away.
1491  *
1492  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1493  * zones which have free_pages > pages_high, but once a zone is found to have
1494  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1495  * of the number of free pages in the lower zones.  This interoperates with
1496  * the page allocator fallback scheme to ensure that aging of pages is balanced
1497  * across the zones.
1498  */
1499 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1500 {
1501         int to_free = nr_pages;
1502         int all_zones_ok;
1503         int priority;
1504         int i;
1505         int total_scanned, total_reclaimed;
1506         struct reclaim_state *reclaim_state = current->reclaim_state;
1507         struct scan_control sc;
1508
1509 loop_again:
1510         total_scanned = 0;
1511         total_reclaimed = 0;
1512         sc.gfp_mask = GFP_KERNEL;
1513         sc.may_writepage = !laptop_mode;
1514         sc.may_swap = 1;
1515         sc.nr_mapped = read_page_state(nr_mapped);
1516
1517         inc_page_state(pageoutrun);
1518
1519         for (i = 0; i < pgdat->nr_zones; i++) {
1520                 struct zone *zone = pgdat->node_zones + i;
1521
1522                 zone->temp_priority = DEF_PRIORITY;
1523         }
1524
1525         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1526                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1527                 unsigned long lru_pages = 0;
1528
1529                 /* The swap token gets in the way of swapout... */
1530                 if (!priority)
1531                         disable_swap_token();
1532
1533                 all_zones_ok = 1;
1534
1535                 if (nr_pages == 0) {
1536                         /*
1537                          * Scan in the highmem->dma direction for the highest
1538                          * zone which needs scanning
1539                          */
1540                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1541                                 struct zone *zone = pgdat->node_zones + i;
1542
1543                                 if (!populated_zone(zone))
1544                                         continue;
1545
1546                                 if (zone->all_unreclaimable &&
1547                                                 priority != DEF_PRIORITY)
1548                                         continue;
1549
1550                                 if (!zone_watermark_ok(zone, order,
1551                                                 zone->pages_high, 0, 0)) {
1552                                         end_zone = i;
1553                                         goto scan;
1554                                 }
1555                         }
1556                         goto out;
1557                 } else {
1558                         end_zone = pgdat->nr_zones - 1;
1559                 }
1560 scan:
1561                 for (i = 0; i <= end_zone; i++) {
1562                         struct zone *zone = pgdat->node_zones + i;
1563
1564                         lru_pages += zone->nr_active + zone->nr_inactive;
1565                 }
1566
1567                 /*
1568                  * Now scan the zone in the dma->highmem direction, stopping
1569                  * at the last zone which needs scanning.
1570                  *
1571                  * We do this because the page allocator works in the opposite
1572                  * direction.  This prevents the page allocator from allocating
1573                  * pages behind kswapd's direction of progress, which would
1574                  * cause too much scanning of the lower zones.
1575                  */
1576                 for (i = 0; i <= end_zone; i++) {
1577                         struct zone *zone = pgdat->node_zones + i;
1578                         int nr_slab;
1579
1580                         if (!populated_zone(zone))
1581                                 continue;
1582
1583                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1584                                 continue;
1585
1586                         if (nr_pages == 0) {    /* Not software suspend */
1587                                 if (!zone_watermark_ok(zone, order,
1588                                                 zone->pages_high, end_zone, 0))
1589                                         all_zones_ok = 0;
1590                         }
1591                         zone->temp_priority = priority;
1592                         if (zone->prev_priority > priority)
1593                                 zone->prev_priority = priority;
1594                         sc.nr_scanned = 0;
1595                         sc.nr_reclaimed = 0;
1596                         sc.priority = priority;
1597                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1598                         atomic_inc(&zone->reclaim_in_progress);
1599                         shrink_zone(zone, &sc);
1600                         atomic_dec(&zone->reclaim_in_progress);
1601                         reclaim_state->reclaimed_slab = 0;
1602                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1603                                                 lru_pages);
1604                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1605                         total_reclaimed += sc.nr_reclaimed;
1606                         total_scanned += sc.nr_scanned;
1607                         if (zone->all_unreclaimable)
1608                                 continue;
1609                         if (nr_slab == 0 && zone->pages_scanned >=
1610                                     (zone->nr_active + zone->nr_inactive) * 4)
1611                                 zone->all_unreclaimable = 1;
1612                         /*
1613                          * If we've done a decent amount of scanning and
1614                          * the reclaim ratio is low, start doing writepage
1615                          * even in laptop mode
1616                          */
1617                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1618                             total_scanned > total_reclaimed+total_reclaimed/2)
1619                                 sc.may_writepage = 1;
1620                 }
1621                 if (nr_pages && to_free > total_reclaimed)
1622                         continue;       /* swsusp: need to do more work */
1623                 if (all_zones_ok)
1624                         break;          /* kswapd: all done */
1625                 /*
1626                  * OK, kswapd is getting into trouble.  Take a nap, then take
1627                  * another pass across the zones.
1628                  */
1629                 if (total_scanned && priority < DEF_PRIORITY - 2)
1630                         blk_congestion_wait(WRITE, HZ/10);
1631
1632                 /*
1633                  * We do this so kswapd doesn't build up large priorities for
1634                  * example when it is freeing in parallel with allocators. It
1635                  * matches the direct reclaim path behaviour in terms of impact
1636                  * on zone->*_priority.
1637                  */
1638                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1639                         break;
1640         }
1641 out:
1642         for (i = 0; i < pgdat->nr_zones; i++) {
1643                 struct zone *zone = pgdat->node_zones + i;
1644
1645                 zone->prev_priority = zone->temp_priority;
1646         }
1647         if (!all_zones_ok) {
1648                 cond_resched();
1649                 goto loop_again;
1650         }
1651
1652         return total_reclaimed;
1653 }
1654
1655 /*
1656  * The background pageout daemon, started as a kernel thread
1657  * from the init process. 
1658  *
1659  * This basically trickles out pages so that we have _some_
1660  * free memory available even if there is no other activity
1661  * that frees anything up. This is needed for things like routing
1662  * etc, where we otherwise might have all activity going on in
1663  * asynchronous contexts that cannot page things out.
1664  *
1665  * If there are applications that are active memory-allocators
1666  * (most normal use), this basically shouldn't matter.
1667  */
1668 static int kswapd(void *p)
1669 {
1670         unsigned long order;
1671         pg_data_t *pgdat = (pg_data_t*)p;
1672         struct task_struct *tsk = current;
1673         DEFINE_WAIT(wait);
1674         struct reclaim_state reclaim_state = {
1675                 .reclaimed_slab = 0,
1676         };
1677         cpumask_t cpumask;
1678
1679         daemonize("kswapd%d", pgdat->node_id);
1680         cpumask = node_to_cpumask(pgdat->node_id);
1681         if (!cpus_empty(cpumask))
1682                 set_cpus_allowed(tsk, cpumask);
1683         current->reclaim_state = &reclaim_state;
1684
1685         /*
1686          * Tell the memory management that we're a "memory allocator",
1687          * and that if we need more memory we should get access to it
1688          * regardless (see "__alloc_pages()"). "kswapd" should
1689          * never get caught in the normal page freeing logic.
1690          *
1691          * (Kswapd normally doesn't need memory anyway, but sometimes
1692          * you need a small amount of memory in order to be able to
1693          * page out something else, and this flag essentially protects
1694          * us from recursively trying to free more memory as we're
1695          * trying to free the first piece of memory in the first place).
1696          */
1697         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1698
1699         order = 0;
1700         for ( ; ; ) {
1701                 unsigned long new_order;
1702
1703                 try_to_freeze();
1704
1705                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1706                 new_order = pgdat->kswapd_max_order;
1707                 pgdat->kswapd_max_order = 0;
1708                 if (order < new_order) {
1709                         /*
1710                          * Don't sleep if someone wants a larger 'order'
1711                          * allocation
1712                          */
1713                         order = new_order;
1714                 } else {
1715                         schedule();
1716                         order = pgdat->kswapd_max_order;
1717                 }
1718                 finish_wait(&pgdat->kswapd_wait, &wait);
1719
1720                 balance_pgdat(pgdat, 0, order);
1721         }
1722         return 0;
1723 }
1724
1725 /*
1726  * A zone is low on free memory, so wake its kswapd task to service it.
1727  */
1728 void wakeup_kswapd(struct zone *zone, int order)
1729 {
1730         pg_data_t *pgdat;
1731
1732         if (!populated_zone(zone))
1733                 return;
1734
1735         pgdat = zone->zone_pgdat;
1736         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1737                 return;
1738         if (pgdat->kswapd_max_order < order)
1739                 pgdat->kswapd_max_order = order;
1740         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1741                 return;
1742         if (!waitqueue_active(&pgdat->kswapd_wait))
1743                 return;
1744         wake_up_interruptible(&pgdat->kswapd_wait);
1745 }
1746
1747 #ifdef CONFIG_PM
1748 /*
1749  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1750  * pages.
1751  */
1752 int shrink_all_memory(int nr_pages)
1753 {
1754         pg_data_t *pgdat;
1755         int nr_to_free = nr_pages;
1756         int ret = 0;
1757         struct reclaim_state reclaim_state = {
1758                 .reclaimed_slab = 0,
1759         };
1760
1761         current->reclaim_state = &reclaim_state;
1762         for_each_pgdat(pgdat) {
1763                 int freed;
1764                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1765                 ret += freed;
1766                 nr_to_free -= freed;
1767                 if (nr_to_free <= 0)
1768                         break;
1769         }
1770         current->reclaim_state = NULL;
1771         return ret;
1772 }
1773 #endif
1774
1775 #ifdef CONFIG_HOTPLUG_CPU
1776 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1777    not required for correctness.  So if the last cpu in a node goes
1778    away, we get changed to run anywhere: as the first one comes back,
1779    restore their cpu bindings. */
1780 static int __devinit cpu_callback(struct notifier_block *nfb,
1781                                   unsigned long action,
1782                                   void *hcpu)
1783 {
1784         pg_data_t *pgdat;
1785         cpumask_t mask;
1786
1787         if (action == CPU_ONLINE) {
1788                 for_each_pgdat(pgdat) {
1789                         mask = node_to_cpumask(pgdat->node_id);
1790                         if (any_online_cpu(mask) != NR_CPUS)
1791                                 /* One of our CPUs online: restore mask */
1792                                 set_cpus_allowed(pgdat->kswapd, mask);
1793                 }
1794         }
1795         return NOTIFY_OK;
1796 }
1797 #endif /* CONFIG_HOTPLUG_CPU */
1798
1799 static int __init kswapd_init(void)
1800 {
1801         pg_data_t *pgdat;
1802         swap_setup();
1803         for_each_pgdat(pgdat)
1804                 pgdat->kswapd
1805                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1806         total_memory = nr_free_pagecache_pages();
1807         hotcpu_notifier(cpu_callback, 0);
1808         return 0;
1809 }
1810
1811 module_init(kswapd_init)
1812
1813 #ifdef CONFIG_NUMA
1814 /*
1815  * Zone reclaim mode
1816  *
1817  * If non-zero call zone_reclaim when the number of free pages falls below
1818  * the watermarks.
1819  *
1820  * In the future we may add flags to the mode. However, the page allocator
1821  * should only have to check that zone_reclaim_mode != 0 before calling
1822  * zone_reclaim().
1823  */
1824 int zone_reclaim_mode __read_mostly;
1825
1826 #define RECLAIM_OFF 0
1827 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1828 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1829 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1830 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1831
1832 /*
1833  * Mininum time between zone reclaim scans
1834  */
1835 int zone_reclaim_interval __read_mostly = 30*HZ;
1836
1837 /*
1838  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1839  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1840  * a zone.
1841  */
1842 #define ZONE_RECLAIM_PRIORITY 4
1843
1844 /*
1845  * Try to free up some pages from this zone through reclaim.
1846  */
1847 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1848 {
1849         int nr_pages;
1850         struct task_struct *p = current;
1851         struct reclaim_state reclaim_state;
1852         struct scan_control sc;
1853         cpumask_t mask;
1854         int node_id;
1855
1856         if (time_before(jiffies,
1857                 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1858                         return 0;
1859
1860         if (!(gfp_mask & __GFP_WAIT) ||
1861                 zone->all_unreclaimable ||
1862                 atomic_read(&zone->reclaim_in_progress) > 0)
1863                         return 0;
1864
1865         node_id = zone->zone_pgdat->node_id;
1866         mask = node_to_cpumask(node_id);
1867         if (!cpus_empty(mask) && node_id != numa_node_id())
1868                 return 0;
1869
1870         sc.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE);
1871         sc.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP);
1872         sc.nr_scanned = 0;
1873         sc.nr_reclaimed = 0;
1874         sc.priority = ZONE_RECLAIM_PRIORITY + 1;
1875         sc.nr_mapped = read_page_state(nr_mapped);
1876         sc.gfp_mask = gfp_mask;
1877
1878         disable_swap_token();
1879
1880         nr_pages = 1 << order;
1881         if (nr_pages > SWAP_CLUSTER_MAX)
1882                 sc.swap_cluster_max = nr_pages;
1883         else
1884                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1885
1886         cond_resched();
1887         p->flags |= PF_MEMALLOC;
1888         reclaim_state.reclaimed_slab = 0;
1889         p->reclaim_state = &reclaim_state;
1890
1891         /*
1892          * Free memory by calling shrink zone with increasing priorities
1893          * until we have enough memory freed.
1894          */
1895         do {
1896                 sc.priority--;
1897                 shrink_zone(zone, &sc);
1898
1899         } while (sc.nr_reclaimed < nr_pages && sc.priority > 0);
1900
1901         if (sc.nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1902                 /*
1903                  * shrink_slab does not currently allow us to determine
1904                  * how many pages were freed in the zone. So we just
1905                  * shake the slab and then go offnode for a single allocation.
1906                  *
1907                  * shrink_slab will free memory on all zones and may take
1908                  * a long time.
1909                  */
1910                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1911                 sc.nr_reclaimed = 1;    /* Avoid getting the off node timeout */
1912         }
1913
1914         p->reclaim_state = NULL;
1915         current->flags &= ~PF_MEMALLOC;
1916
1917         if (sc.nr_reclaimed == 0)
1918                 zone->last_unsuccessful_zone_reclaim = jiffies;
1919
1920         return sc.nr_reclaimed >= nr_pages;
1921 }
1922 #endif
1923