Merge ../linux-2.6
[linux-2.6] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/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
447                 if (!sc->may_swap && page_mapped(page))
448                         goto keep_locked;
449
450                 /* Double the slab pressure for mapped and swapcache pages */
451                 if (page_mapped(page) || PageSwapCache(page))
452                         sc->nr_scanned++;
453
454                 if (PageWriteback(page))
455                         goto keep_locked;
456
457                 referenced = page_referenced(page, 1);
458                 /* In active use or really unfreeable?  Activate it. */
459                 if (referenced && page_mapping_inuse(page))
460                         goto activate_locked;
461
462 #ifdef CONFIG_SWAP
463                 /*
464                  * Anonymous process memory has backing store?
465                  * Try to allocate it some swap space here.
466                  */
467                 if (PageAnon(page) && !PageSwapCache(page)) {
468                         if (!sc->may_swap)
469                                 goto keep_locked;
470                         if (!add_to_swap(page, GFP_ATOMIC))
471                                 goto activate_locked;
472                 }
473 #endif /* CONFIG_SWAP */
474
475                 mapping = page_mapping(page);
476                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
477                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
478
479                 /*
480                  * The page is mapped into the page tables of one or more
481                  * processes. Try to unmap it here.
482                  */
483                 if (page_mapped(page) && mapping) {
484                         /*
485                          * No unmapping if we do not swap
486                          */
487                         if (!sc->may_swap)
488                                 goto keep_locked;
489
490                         switch (try_to_unmap(page, 0)) {
491                         case SWAP_FAIL:
492                                 goto activate_locked;
493                         case SWAP_AGAIN:
494                                 goto keep_locked;
495                         case SWAP_SUCCESS:
496                                 ; /* try to free the page below */
497                         }
498                 }
499
500                 if (PageDirty(page)) {
501                         if (referenced)
502                                 goto keep_locked;
503                         if (!may_enter_fs)
504                                 goto keep_locked;
505                         if (!sc->may_writepage)
506                                 goto keep_locked;
507
508                         /* Page is dirty, try to write it out here */
509                         switch(pageout(page, mapping)) {
510                         case PAGE_KEEP:
511                                 goto keep_locked;
512                         case PAGE_ACTIVATE:
513                                 goto activate_locked;
514                         case PAGE_SUCCESS:
515                                 if (PageWriteback(page) || PageDirty(page))
516                                         goto keep;
517                                 /*
518                                  * A synchronous write - probably a ramdisk.  Go
519                                  * ahead and try to reclaim the page.
520                                  */
521                                 if (TestSetPageLocked(page))
522                                         goto keep;
523                                 if (PageDirty(page) || PageWriteback(page))
524                                         goto keep_locked;
525                                 mapping = page_mapping(page);
526                         case PAGE_CLEAN:
527                                 ; /* try to free the page below */
528                         }
529                 }
530
531                 /*
532                  * If the page has buffers, try to free the buffer mappings
533                  * associated with this page. If we succeed we try to free
534                  * the page as well.
535                  *
536                  * We do this even if the page is PageDirty().
537                  * try_to_release_page() does not perform I/O, but it is
538                  * possible for a page to have PageDirty set, but it is actually
539                  * clean (all its buffers are clean).  This happens if the
540                  * buffers were written out directly, with submit_bh(). ext3
541                  * will do this, as well as the blockdev mapping. 
542                  * try_to_release_page() will discover that cleanness and will
543                  * drop the buffers and mark the page clean - it can be freed.
544                  *
545                  * Rarely, pages can have buffers and no ->mapping.  These are
546                  * the pages which were not successfully invalidated in
547                  * truncate_complete_page().  We try to drop those buffers here
548                  * and if that worked, and the page is no longer mapped into
549                  * process address space (page_count == 1) it can be freed.
550                  * Otherwise, leave the page on the LRU so it is swappable.
551                  */
552                 if (PagePrivate(page)) {
553                         if (!try_to_release_page(page, sc->gfp_mask))
554                                 goto activate_locked;
555                         if (!mapping && page_count(page) == 1)
556                                 goto free_it;
557                 }
558
559                 if (!remove_mapping(mapping, page))
560                         goto keep_locked;
561
562 free_it:
563                 unlock_page(page);
564                 reclaimed++;
565                 if (!pagevec_add(&freed_pvec, page))
566                         __pagevec_release_nonlru(&freed_pvec);
567                 continue;
568
569 activate_locked:
570                 SetPageActive(page);
571                 pgactivate++;
572 keep_locked:
573                 unlock_page(page);
574 keep:
575                 list_add(&page->lru, &ret_pages);
576                 BUG_ON(PageLRU(page));
577         }
578         list_splice(&ret_pages, page_list);
579         if (pagevec_count(&freed_pvec))
580                 __pagevec_release_nonlru(&freed_pvec);
581         mod_page_state(pgactivate, pgactivate);
582         sc->nr_reclaimed += reclaimed;
583         return reclaimed;
584 }
585
586 #ifdef CONFIG_MIGRATION
587 static inline void move_to_lru(struct page *page)
588 {
589         list_del(&page->lru);
590         if (PageActive(page)) {
591                 /*
592                  * lru_cache_add_active checks that
593                  * the PG_active bit is off.
594                  */
595                 ClearPageActive(page);
596                 lru_cache_add_active(page);
597         } else {
598                 lru_cache_add(page);
599         }
600         put_page(page);
601 }
602
603 /*
604  * Add isolated pages on the list back to the LRU.
605  *
606  * returns the number of pages put back.
607  */
608 int putback_lru_pages(struct list_head *l)
609 {
610         struct page *page;
611         struct page *page2;
612         int count = 0;
613
614         list_for_each_entry_safe(page, page2, l, lru) {
615                 move_to_lru(page);
616                 count++;
617         }
618         return count;
619 }
620
621 /*
622  * Non migratable page
623  */
624 int fail_migrate_page(struct page *newpage, struct page *page)
625 {
626         return -EIO;
627 }
628 EXPORT_SYMBOL(fail_migrate_page);
629
630 /*
631  * swapout a single page
632  * page is locked upon entry, unlocked on exit
633  */
634 static int swap_page(struct page *page)
635 {
636         struct address_space *mapping = page_mapping(page);
637
638         if (page_mapped(page) && mapping)
639                 if (try_to_unmap(page, 1) != SWAP_SUCCESS)
640                         goto unlock_retry;
641
642         if (PageDirty(page)) {
643                 /* Page is dirty, try to write it out here */
644                 switch(pageout(page, mapping)) {
645                 case PAGE_KEEP:
646                 case PAGE_ACTIVATE:
647                         goto unlock_retry;
648
649                 case PAGE_SUCCESS:
650                         goto retry;
651
652                 case PAGE_CLEAN:
653                         ; /* try to free the page below */
654                 }
655         }
656
657         if (PagePrivate(page)) {
658                 if (!try_to_release_page(page, GFP_KERNEL) ||
659                     (!mapping && page_count(page) == 1))
660                         goto unlock_retry;
661         }
662
663         if (remove_mapping(mapping, page)) {
664                 /* Success */
665                 unlock_page(page);
666                 return 0;
667         }
668
669 unlock_retry:
670         unlock_page(page);
671
672 retry:
673         return -EAGAIN;
674 }
675 EXPORT_SYMBOL(swap_page);
676
677 /*
678  * Page migration was first developed in the context of the memory hotplug
679  * project. The main authors of the migration code are:
680  *
681  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
682  * Hirokazu Takahashi <taka@valinux.co.jp>
683  * Dave Hansen <haveblue@us.ibm.com>
684  * Christoph Lameter <clameter@sgi.com>
685  */
686
687 /*
688  * Remove references for a page and establish the new page with the correct
689  * basic settings to be able to stop accesses to the page.
690  */
691 int migrate_page_remove_references(struct page *newpage,
692                                 struct page *page, int nr_refs)
693 {
694         struct address_space *mapping = page_mapping(page);
695         struct page **radix_pointer;
696
697         /*
698          * Avoid doing any of the following work if the page count
699          * indicates that the page is in use or truncate has removed
700          * the page.
701          */
702         if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
703                 return -EAGAIN;
704
705         /*
706          * Establish swap ptes for anonymous pages or destroy pte
707          * maps for files.
708          *
709          * In order to reestablish file backed mappings the fault handlers
710          * will take the radix tree_lock which may then be used to stop
711          * processses from accessing this page until the new page is ready.
712          *
713          * A process accessing via a swap pte (an anonymous page) will take a
714          * page_lock on the old page which will block the process until the
715          * migration attempt is complete. At that time the PageSwapCache bit
716          * will be examined. If the page was migrated then the PageSwapCache
717          * bit will be clear and the operation to retrieve the page will be
718          * retried which will find the new page in the radix tree. Then a new
719          * direct mapping may be generated based on the radix tree contents.
720          *
721          * If the page was not migrated then the PageSwapCache bit
722          * is still set and the operation may continue.
723          */
724         if (try_to_unmap(page, 1) == SWAP_FAIL)
725                 /* A vma has VM_LOCKED set -> Permanent failure */
726                 return -EPERM;
727
728         /*
729          * Give up if we were unable to remove all mappings.
730          */
731         if (page_mapcount(page))
732                 return -EAGAIN;
733
734         write_lock_irq(&mapping->tree_lock);
735
736         radix_pointer = (struct page **)radix_tree_lookup_slot(
737                                                 &mapping->page_tree,
738                                                 page_index(page));
739
740         if (!page_mapping(page) || page_count(page) != nr_refs ||
741                         *radix_pointer != page) {
742                 write_unlock_irq(&mapping->tree_lock);
743                 return -EAGAIN;
744         }
745
746         /*
747          * Now we know that no one else is looking at the page.
748          *
749          * Certain minimal information about a page must be available
750          * in order for other subsystems to properly handle the page if they
751          * find it through the radix tree update before we are finished
752          * copying the page.
753          */
754         get_page(newpage);
755         newpage->index = page->index;
756         newpage->mapping = page->mapping;
757         if (PageSwapCache(page)) {
758                 SetPageSwapCache(newpage);
759                 set_page_private(newpage, page_private(page));
760         }
761
762         *radix_pointer = newpage;
763         __put_page(page);
764         write_unlock_irq(&mapping->tree_lock);
765
766         return 0;
767 }
768 EXPORT_SYMBOL(migrate_page_remove_references);
769
770 /*
771  * Copy the page to its new location
772  */
773 void migrate_page_copy(struct page *newpage, struct page *page)
774 {
775         copy_highpage(newpage, page);
776
777         if (PageError(page))
778                 SetPageError(newpage);
779         if (PageReferenced(page))
780                 SetPageReferenced(newpage);
781         if (PageUptodate(page))
782                 SetPageUptodate(newpage);
783         if (PageActive(page))
784                 SetPageActive(newpage);
785         if (PageChecked(page))
786                 SetPageChecked(newpage);
787         if (PageMappedToDisk(page))
788                 SetPageMappedToDisk(newpage);
789
790         if (PageDirty(page)) {
791                 clear_page_dirty_for_io(page);
792                 set_page_dirty(newpage);
793         }
794
795         ClearPageSwapCache(page);
796         ClearPageActive(page);
797         ClearPagePrivate(page);
798         set_page_private(page, 0);
799         page->mapping = NULL;
800
801         /*
802          * If any waiters have accumulated on the new page then
803          * wake them up.
804          */
805         if (PageWriteback(newpage))
806                 end_page_writeback(newpage);
807 }
808 EXPORT_SYMBOL(migrate_page_copy);
809
810 /*
811  * Common logic to directly migrate a single page suitable for
812  * pages that do not use PagePrivate.
813  *
814  * Pages are locked upon entry and exit.
815  */
816 int migrate_page(struct page *newpage, struct page *page)
817 {
818         int rc;
819
820         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
821
822         rc = migrate_page_remove_references(newpage, page, 2);
823
824         if (rc)
825                 return rc;
826
827         migrate_page_copy(newpage, page);
828
829         /*
830          * Remove auxiliary swap entries and replace
831          * them with real ptes.
832          *
833          * Note that a real pte entry will allow processes that are not
834          * waiting on the page lock to use the new page via the page tables
835          * before the new page is unlocked.
836          */
837         remove_from_swap(newpage);
838         return 0;
839 }
840 EXPORT_SYMBOL(migrate_page);
841
842 /*
843  * migrate_pages
844  *
845  * Two lists are passed to this function. The first list
846  * contains the pages isolated from the LRU to be migrated.
847  * The second list contains new pages that the pages isolated
848  * can be moved to. If the second list is NULL then all
849  * pages are swapped out.
850  *
851  * The function returns after 10 attempts or if no pages
852  * are movable anymore because to has become empty
853  * or no retryable pages exist anymore.
854  *
855  * Return: Number of pages not migrated when "to" ran empty.
856  */
857 int migrate_pages(struct list_head *from, struct list_head *to,
858                   struct list_head *moved, struct list_head *failed)
859 {
860         int retry;
861         int nr_failed = 0;
862         int pass = 0;
863         struct page *page;
864         struct page *page2;
865         int swapwrite = current->flags & PF_SWAPWRITE;
866         int rc;
867
868         if (!swapwrite)
869                 current->flags |= PF_SWAPWRITE;
870
871 redo:
872         retry = 0;
873
874         list_for_each_entry_safe(page, page2, from, lru) {
875                 struct page *newpage = NULL;
876                 struct address_space *mapping;
877
878                 cond_resched();
879
880                 rc = 0;
881                 if (page_count(page) == 1)
882                         /* page was freed from under us. So we are done. */
883                         goto next;
884
885                 if (to && list_empty(to))
886                         break;
887
888                 /*
889                  * Skip locked pages during the first two passes to give the
890                  * functions holding the lock time to release the page. Later we
891                  * use lock_page() to have a higher chance of acquiring the
892                  * lock.
893                  */
894                 rc = -EAGAIN;
895                 if (pass > 2)
896                         lock_page(page);
897                 else
898                         if (TestSetPageLocked(page))
899                                 goto next;
900
901                 /*
902                  * Only wait on writeback if we have already done a pass where
903                  * we we may have triggered writeouts for lots of pages.
904                  */
905                 if (pass > 0) {
906                         wait_on_page_writeback(page);
907                 } else {
908                         if (PageWriteback(page))
909                                 goto unlock_page;
910                 }
911
912                 /*
913                  * Anonymous pages must have swap cache references otherwise
914                  * the information contained in the page maps cannot be
915                  * preserved.
916                  */
917                 if (PageAnon(page) && !PageSwapCache(page)) {
918                         if (!add_to_swap(page, GFP_KERNEL)) {
919                                 rc = -ENOMEM;
920                                 goto unlock_page;
921                         }
922                 }
923
924                 if (!to) {
925                         rc = swap_page(page);
926                         goto next;
927                 }
928
929                 newpage = lru_to_page(to);
930                 lock_page(newpage);
931
932                 /*
933                  * Pages are properly locked and writeback is complete.
934                  * Try to migrate the page.
935                  */
936                 mapping = page_mapping(page);
937                 if (!mapping)
938                         goto unlock_both;
939
940                 if (mapping->a_ops->migratepage) {
941                         /*
942                          * Most pages have a mapping and most filesystems
943                          * should provide a migration function. Anonymous
944                          * pages are part of swap space which also has its
945                          * own migration function. This is the most common
946                          * path for page migration.
947                          */
948                         rc = mapping->a_ops->migratepage(newpage, page);
949                         goto unlock_both;
950                 }
951
952                 /*
953                  * Default handling if a filesystem does not provide
954                  * a migration function. We can only migrate clean
955                  * pages so try to write out any dirty pages first.
956                  */
957                 if (PageDirty(page)) {
958                         switch (pageout(page, mapping)) {
959                         case PAGE_KEEP:
960                         case PAGE_ACTIVATE:
961                                 goto unlock_both;
962
963                         case PAGE_SUCCESS:
964                                 unlock_page(newpage);
965                                 goto next;
966
967                         case PAGE_CLEAN:
968                                 ; /* try to migrate the page below */
969                         }
970                 }
971
972                 /*
973                  * Buffers are managed in a filesystem specific way.
974                  * We must have no buffers or drop them.
975                  */
976                 if (!page_has_buffers(page) ||
977                     try_to_release_page(page, GFP_KERNEL)) {
978                         rc = migrate_page(newpage, page);
979                         goto unlock_both;
980                 }
981
982                 /*
983                  * On early passes with mapped pages simply
984                  * retry. There may be a lock held for some
985                  * buffers that may go away. Later
986                  * swap them out.
987                  */
988                 if (pass > 4) {
989                         /*
990                          * Persistently unable to drop buffers..... As a
991                          * measure of last resort we fall back to
992                          * swap_page().
993                          */
994                         unlock_page(newpage);
995                         newpage = NULL;
996                         rc = swap_page(page);
997                         goto next;
998                 }
999
1000 unlock_both:
1001                 unlock_page(newpage);
1002
1003 unlock_page:
1004                 unlock_page(page);
1005
1006 next:
1007                 if (rc == -EAGAIN) {
1008                         retry++;
1009                 } else if (rc) {
1010                         /* Permanent failure */
1011                         list_move(&page->lru, failed);
1012                         nr_failed++;
1013                 } else {
1014                         if (newpage) {
1015                                 /* Successful migration. Return page to LRU */
1016                                 move_to_lru(newpage);
1017                         }
1018                         list_move(&page->lru, moved);
1019                 }
1020         }
1021         if (retry && pass++ < 10)
1022                 goto redo;
1023
1024         if (!swapwrite)
1025                 current->flags &= ~PF_SWAPWRITE;
1026
1027         return nr_failed + retry;
1028 }
1029
1030 /*
1031  * Isolate one page from the LRU lists and put it on the
1032  * indicated list with elevated refcount.
1033  *
1034  * Result:
1035  *  0 = page not on LRU list
1036  *  1 = page removed from LRU list and added to the specified list.
1037  */
1038 int isolate_lru_page(struct page *page)
1039 {
1040         int ret = 0;
1041
1042         if (PageLRU(page)) {
1043                 struct zone *zone = page_zone(page);
1044                 spin_lock_irq(&zone->lru_lock);
1045                 if (TestClearPageLRU(page)) {
1046                         ret = 1;
1047                         get_page(page);
1048                         if (PageActive(page))
1049                                 del_page_from_active_list(zone, page);
1050                         else
1051                                 del_page_from_inactive_list(zone, page);
1052                 }
1053                 spin_unlock_irq(&zone->lru_lock);
1054         }
1055
1056         return ret;
1057 }
1058 #endif
1059
1060 /*
1061  * zone->lru_lock is heavily contended.  Some of the functions that
1062  * shrink the lists perform better by taking out a batch of pages
1063  * and working on them outside the LRU lock.
1064  *
1065  * For pagecache intensive workloads, this function is the hottest
1066  * spot in the kernel (apart from copy_*_user functions).
1067  *
1068  * Appropriate locks must be held before calling this function.
1069  *
1070  * @nr_to_scan: The number of pages to look through on the list.
1071  * @src:        The LRU list to pull pages off.
1072  * @dst:        The temp list to put pages on to.
1073  * @scanned:    The number of pages that were scanned.
1074  *
1075  * returns how many pages were moved onto *@dst.
1076  */
1077 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
1078                              struct list_head *dst, int *scanned)
1079 {
1080         int nr_taken = 0;
1081         struct page *page;
1082         int scan = 0;
1083
1084         while (scan++ < nr_to_scan && !list_empty(src)) {
1085                 page = lru_to_page(src);
1086                 prefetchw_prev_lru_page(page, src, flags);
1087
1088                 if (!TestClearPageLRU(page))
1089                         BUG();
1090                 list_del(&page->lru);
1091                 if (get_page_testone(page)) {
1092                         /*
1093                          * It is being freed elsewhere
1094                          */
1095                         __put_page(page);
1096                         SetPageLRU(page);
1097                         list_add(&page->lru, src);
1098                         continue;
1099                 } else {
1100                         list_add(&page->lru, dst);
1101                         nr_taken++;
1102                 }
1103         }
1104
1105         *scanned = scan;
1106         return nr_taken;
1107 }
1108
1109 /*
1110  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1111  */
1112 static void shrink_cache(struct zone *zone, struct scan_control *sc)
1113 {
1114         LIST_HEAD(page_list);
1115         struct pagevec pvec;
1116         int max_scan = sc->nr_to_scan;
1117
1118         pagevec_init(&pvec, 1);
1119
1120         lru_add_drain();
1121         spin_lock_irq(&zone->lru_lock);
1122         while (max_scan > 0) {
1123                 struct page *page;
1124                 int nr_taken;
1125                 int nr_scan;
1126                 int nr_freed;
1127
1128                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
1129                                              &zone->inactive_list,
1130                                              &page_list, &nr_scan);
1131                 zone->nr_inactive -= nr_taken;
1132                 zone->pages_scanned += nr_scan;
1133                 spin_unlock_irq(&zone->lru_lock);
1134
1135                 if (nr_taken == 0)
1136                         goto done;
1137
1138                 max_scan -= nr_scan;
1139                 nr_freed = shrink_list(&page_list, sc);
1140
1141                 local_irq_disable();
1142                 if (current_is_kswapd()) {
1143                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
1144                         __mod_page_state(kswapd_steal, nr_freed);
1145                 } else
1146                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
1147                 __mod_page_state_zone(zone, pgsteal, nr_freed);
1148
1149                 spin_lock(&zone->lru_lock);
1150                 /*
1151                  * Put back any unfreeable pages.
1152                  */
1153                 while (!list_empty(&page_list)) {
1154                         page = lru_to_page(&page_list);
1155                         if (TestSetPageLRU(page))
1156                                 BUG();
1157                         list_del(&page->lru);
1158                         if (PageActive(page))
1159                                 add_page_to_active_list(zone, page);
1160                         else
1161                                 add_page_to_inactive_list(zone, page);
1162                         if (!pagevec_add(&pvec, page)) {
1163                                 spin_unlock_irq(&zone->lru_lock);
1164                                 __pagevec_release(&pvec);
1165                                 spin_lock_irq(&zone->lru_lock);
1166                         }
1167                 }
1168         }
1169         spin_unlock_irq(&zone->lru_lock);
1170 done:
1171         pagevec_release(&pvec);
1172 }
1173
1174 /*
1175  * This moves pages from the active list to the inactive list.
1176  *
1177  * We move them the other way if the page is referenced by one or more
1178  * processes, from rmap.
1179  *
1180  * If the pages are mostly unmapped, the processing is fast and it is
1181  * appropriate to hold zone->lru_lock across the whole operation.  But if
1182  * the pages are mapped, the processing is slow (page_referenced()) so we
1183  * should drop zone->lru_lock around each page.  It's impossible to balance
1184  * this, so instead we remove the pages from the LRU while processing them.
1185  * It is safe to rely on PG_active against the non-LRU pages in here because
1186  * nobody will play with that bit on a non-LRU page.
1187  *
1188  * The downside is that we have to touch page->_count against each page.
1189  * But we had to alter page->flags anyway.
1190  */
1191 static void
1192 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
1193 {
1194         int pgmoved;
1195         int pgdeactivate = 0;
1196         int pgscanned;
1197         int nr_pages = sc->nr_to_scan;
1198         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1199         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1200         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1201         struct page *page;
1202         struct pagevec pvec;
1203         int reclaim_mapped = 0;
1204
1205         if (unlikely(sc->may_swap)) {
1206                 long mapped_ratio;
1207                 long distress;
1208                 long swap_tendency;
1209
1210                 /*
1211                  * `distress' is a measure of how much trouble we're having
1212                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
1213                  */
1214                 distress = 100 >> zone->prev_priority;
1215
1216                 /*
1217                  * The point of this algorithm is to decide when to start
1218                  * reclaiming mapped memory instead of just pagecache.  Work out
1219                  * how much memory
1220                  * is mapped.
1221                  */
1222                 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
1223
1224                 /*
1225                  * Now decide how much we really want to unmap some pages.  The
1226                  * mapped ratio is downgraded - just because there's a lot of
1227                  * mapped memory doesn't necessarily mean that page reclaim
1228                  * isn't succeeding.
1229                  *
1230                  * The distress ratio is important - we don't want to start
1231                  * going oom.
1232                  *
1233                  * A 100% value of vm_swappiness overrides this algorithm
1234                  * altogether.
1235                  */
1236                 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
1237
1238                 /*
1239                  * Now use this metric to decide whether to start moving mapped
1240                  * memory onto the inactive list.
1241                  */
1242                 if (swap_tendency >= 100)
1243                         reclaim_mapped = 1;
1244         }
1245
1246         lru_add_drain();
1247         spin_lock_irq(&zone->lru_lock);
1248         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1249                                     &l_hold, &pgscanned);
1250         zone->pages_scanned += pgscanned;
1251         zone->nr_active -= pgmoved;
1252         spin_unlock_irq(&zone->lru_lock);
1253
1254         while (!list_empty(&l_hold)) {
1255                 cond_resched();
1256                 page = lru_to_page(&l_hold);
1257                 list_del(&page->lru);
1258                 if (page_mapped(page)) {
1259                         if (!reclaim_mapped ||
1260                             (total_swap_pages == 0 && PageAnon(page)) ||
1261                             page_referenced(page, 0)) {
1262                                 list_add(&page->lru, &l_active);
1263                                 continue;
1264                         }
1265                 }
1266                 list_add(&page->lru, &l_inactive);
1267         }
1268
1269         pagevec_init(&pvec, 1);
1270         pgmoved = 0;
1271         spin_lock_irq(&zone->lru_lock);
1272         while (!list_empty(&l_inactive)) {
1273                 page = lru_to_page(&l_inactive);
1274                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1275                 if (TestSetPageLRU(page))
1276                         BUG();
1277                 if (!TestClearPageActive(page))
1278                         BUG();
1279                 list_move(&page->lru, &zone->inactive_list);
1280                 pgmoved++;
1281                 if (!pagevec_add(&pvec, page)) {
1282                         zone->nr_inactive += pgmoved;
1283                         spin_unlock_irq(&zone->lru_lock);
1284                         pgdeactivate += pgmoved;
1285                         pgmoved = 0;
1286                         if (buffer_heads_over_limit)
1287                                 pagevec_strip(&pvec);
1288                         __pagevec_release(&pvec);
1289                         spin_lock_irq(&zone->lru_lock);
1290                 }
1291         }
1292         zone->nr_inactive += pgmoved;
1293         pgdeactivate += pgmoved;
1294         if (buffer_heads_over_limit) {
1295                 spin_unlock_irq(&zone->lru_lock);
1296                 pagevec_strip(&pvec);
1297                 spin_lock_irq(&zone->lru_lock);
1298         }
1299
1300         pgmoved = 0;
1301         while (!list_empty(&l_active)) {
1302                 page = lru_to_page(&l_active);
1303                 prefetchw_prev_lru_page(page, &l_active, flags);
1304                 if (TestSetPageLRU(page))
1305                         BUG();
1306                 BUG_ON(!PageActive(page));
1307                 list_move(&page->lru, &zone->active_list);
1308                 pgmoved++;
1309                 if (!pagevec_add(&pvec, page)) {
1310                         zone->nr_active += pgmoved;
1311                         pgmoved = 0;
1312                         spin_unlock_irq(&zone->lru_lock);
1313                         __pagevec_release(&pvec);
1314                         spin_lock_irq(&zone->lru_lock);
1315                 }
1316         }
1317         zone->nr_active += pgmoved;
1318         spin_unlock(&zone->lru_lock);
1319
1320         __mod_page_state_zone(zone, pgrefill, pgscanned);
1321         __mod_page_state(pgdeactivate, pgdeactivate);
1322         local_irq_enable();
1323
1324         pagevec_release(&pvec);
1325 }
1326
1327 /*
1328  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1329  */
1330 static void
1331 shrink_zone(struct zone *zone, struct scan_control *sc)
1332 {
1333         unsigned long nr_active;
1334         unsigned long nr_inactive;
1335
1336         atomic_inc(&zone->reclaim_in_progress);
1337
1338         /*
1339          * Add one to `nr_to_scan' just to make sure that the kernel will
1340          * slowly sift through the active list.
1341          */
1342         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1343         nr_active = zone->nr_scan_active;
1344         if (nr_active >= sc->swap_cluster_max)
1345                 zone->nr_scan_active = 0;
1346         else
1347                 nr_active = 0;
1348
1349         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1350         nr_inactive = zone->nr_scan_inactive;
1351         if (nr_inactive >= sc->swap_cluster_max)
1352                 zone->nr_scan_inactive = 0;
1353         else
1354                 nr_inactive = 0;
1355
1356         while (nr_active || nr_inactive) {
1357                 if (nr_active) {
1358                         sc->nr_to_scan = min(nr_active,
1359                                         (unsigned long)sc->swap_cluster_max);
1360                         nr_active -= sc->nr_to_scan;
1361                         refill_inactive_zone(zone, sc);
1362                 }
1363
1364                 if (nr_inactive) {
1365                         sc->nr_to_scan = min(nr_inactive,
1366                                         (unsigned long)sc->swap_cluster_max);
1367                         nr_inactive -= sc->nr_to_scan;
1368                         shrink_cache(zone, sc);
1369                 }
1370         }
1371
1372         throttle_vm_writeout();
1373
1374         atomic_dec(&zone->reclaim_in_progress);
1375 }
1376
1377 /*
1378  * This is the direct reclaim path, for page-allocating processes.  We only
1379  * try to reclaim pages from zones which will satisfy the caller's allocation
1380  * request.
1381  *
1382  * We reclaim from a zone even if that zone is over pages_high.  Because:
1383  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1384  *    allocation or
1385  * b) The zones may be over pages_high but they must go *over* pages_high to
1386  *    satisfy the `incremental min' zone defense algorithm.
1387  *
1388  * Returns the number of reclaimed pages.
1389  *
1390  * If a zone is deemed to be full of pinned pages then just give it a light
1391  * scan then give up on it.
1392  */
1393 static void
1394 shrink_caches(struct zone **zones, struct scan_control *sc)
1395 {
1396         int i;
1397
1398         for (i = 0; zones[i] != NULL; i++) {
1399                 struct zone *zone = zones[i];
1400
1401                 if (!populated_zone(zone))
1402                         continue;
1403
1404                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1405                         continue;
1406
1407                 zone->temp_priority = sc->priority;
1408                 if (zone->prev_priority > sc->priority)
1409                         zone->prev_priority = sc->priority;
1410
1411                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1412                         continue;       /* Let kswapd poll it */
1413
1414                 shrink_zone(zone, sc);
1415         }
1416 }
1417  
1418 /*
1419  * This is the main entry point to direct page reclaim.
1420  *
1421  * If a full scan of the inactive list fails to free enough memory then we
1422  * are "out of memory" and something needs to be killed.
1423  *
1424  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1425  * high - the zone may be full of dirty or under-writeback pages, which this
1426  * caller can't do much about.  We kick pdflush and take explicit naps in the
1427  * hope that some of these pages can be written.  But if the allocating task
1428  * holds filesystem locks which prevent writeout this might not work, and the
1429  * allocation attempt will fail.
1430  */
1431 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1432 {
1433         int priority;
1434         int ret = 0;
1435         int total_scanned = 0, total_reclaimed = 0;
1436         struct reclaim_state *reclaim_state = current->reclaim_state;
1437         struct scan_control sc;
1438         unsigned long lru_pages = 0;
1439         int i;
1440
1441         sc.gfp_mask = gfp_mask;
1442         sc.may_writepage = !laptop_mode;
1443         sc.may_swap = 1;
1444
1445         inc_page_state(allocstall);
1446
1447         for (i = 0; zones[i] != NULL; i++) {
1448                 struct zone *zone = zones[i];
1449
1450                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1451                         continue;
1452
1453                 zone->temp_priority = DEF_PRIORITY;
1454                 lru_pages += zone->nr_active + zone->nr_inactive;
1455         }
1456
1457         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1458                 sc.nr_mapped = read_page_state(nr_mapped);
1459                 sc.nr_scanned = 0;
1460                 sc.nr_reclaimed = 0;
1461                 sc.priority = priority;
1462                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1463                 if (!priority)
1464                         disable_swap_token();
1465                 shrink_caches(zones, &sc);
1466                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1467                 if (reclaim_state) {
1468                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1469                         reclaim_state->reclaimed_slab = 0;
1470                 }
1471                 total_scanned += sc.nr_scanned;
1472                 total_reclaimed += sc.nr_reclaimed;
1473                 if (total_reclaimed >= sc.swap_cluster_max) {
1474                         ret = 1;
1475                         goto out;
1476                 }
1477
1478                 /*
1479                  * Try to write back as many pages as we just scanned.  This
1480                  * tends to cause slow streaming writers to write data to the
1481                  * disk smoothly, at the dirtying rate, which is nice.   But
1482                  * that's undesirable in laptop mode, where we *want* lumpy
1483                  * writeout.  So in laptop mode, write out the whole world.
1484                  */
1485                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1486                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1487                         sc.may_writepage = 1;
1488                 }
1489
1490                 /* Take a nap, wait for some writeback to complete */
1491                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1492                         blk_congestion_wait(WRITE, HZ/10);
1493         }
1494 out:
1495         for (i = 0; zones[i] != 0; i++) {
1496                 struct zone *zone = zones[i];
1497
1498                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1499                         continue;
1500
1501                 zone->prev_priority = zone->temp_priority;
1502         }
1503         return ret;
1504 }
1505
1506 /*
1507  * For kswapd, balance_pgdat() will work across all this node's zones until
1508  * they are all at pages_high.
1509  *
1510  * If `nr_pages' is non-zero then it is the number of pages which are to be
1511  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1512  * special.
1513  *
1514  * Returns the number of pages which were actually freed.
1515  *
1516  * There is special handling here for zones which are full of pinned pages.
1517  * This can happen if the pages are all mlocked, or if they are all used by
1518  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1519  * What we do is to detect the case where all pages in the zone have been
1520  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1521  * dead and from now on, only perform a short scan.  Basically we're polling
1522  * the zone for when the problem goes away.
1523  *
1524  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1525  * zones which have free_pages > pages_high, but once a zone is found to have
1526  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1527  * of the number of free pages in the lower zones.  This interoperates with
1528  * the page allocator fallback scheme to ensure that aging of pages is balanced
1529  * across the zones.
1530  */
1531 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1532 {
1533         int to_free = nr_pages;
1534         int all_zones_ok;
1535         int priority;
1536         int i;
1537         int total_scanned, total_reclaimed;
1538         struct reclaim_state *reclaim_state = current->reclaim_state;
1539         struct scan_control sc;
1540
1541 loop_again:
1542         total_scanned = 0;
1543         total_reclaimed = 0;
1544         sc.gfp_mask = GFP_KERNEL;
1545         sc.may_writepage = !laptop_mode;
1546         sc.may_swap = 1;
1547         sc.nr_mapped = read_page_state(nr_mapped);
1548
1549         inc_page_state(pageoutrun);
1550
1551         for (i = 0; i < pgdat->nr_zones; i++) {
1552                 struct zone *zone = pgdat->node_zones + i;
1553
1554                 zone->temp_priority = DEF_PRIORITY;
1555         }
1556
1557         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1558                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1559                 unsigned long lru_pages = 0;
1560
1561                 /* The swap token gets in the way of swapout... */
1562                 if (!priority)
1563                         disable_swap_token();
1564
1565                 all_zones_ok = 1;
1566
1567                 if (nr_pages == 0) {
1568                         /*
1569                          * Scan in the highmem->dma direction for the highest
1570                          * zone which needs scanning
1571                          */
1572                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1573                                 struct zone *zone = pgdat->node_zones + i;
1574
1575                                 if (!populated_zone(zone))
1576                                         continue;
1577
1578                                 if (zone->all_unreclaimable &&
1579                                                 priority != DEF_PRIORITY)
1580                                         continue;
1581
1582                                 if (!zone_watermark_ok(zone, order,
1583                                                 zone->pages_high, 0, 0)) {
1584                                         end_zone = i;
1585                                         goto scan;
1586                                 }
1587                         }
1588                         goto out;
1589                 } else {
1590                         end_zone = pgdat->nr_zones - 1;
1591                 }
1592 scan:
1593                 for (i = 0; i <= end_zone; i++) {
1594                         struct zone *zone = pgdat->node_zones + i;
1595
1596                         lru_pages += zone->nr_active + zone->nr_inactive;
1597                 }
1598
1599                 /*
1600                  * Now scan the zone in the dma->highmem direction, stopping
1601                  * at the last zone which needs scanning.
1602                  *
1603                  * We do this because the page allocator works in the opposite
1604                  * direction.  This prevents the page allocator from allocating
1605                  * pages behind kswapd's direction of progress, which would
1606                  * cause too much scanning of the lower zones.
1607                  */
1608                 for (i = 0; i <= end_zone; i++) {
1609                         struct zone *zone = pgdat->node_zones + i;
1610                         int nr_slab;
1611
1612                         if (!populated_zone(zone))
1613                                 continue;
1614
1615                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1616                                 continue;
1617
1618                         if (nr_pages == 0) {    /* Not software suspend */
1619                                 if (!zone_watermark_ok(zone, order,
1620                                                 zone->pages_high, end_zone, 0))
1621                                         all_zones_ok = 0;
1622                         }
1623                         zone->temp_priority = priority;
1624                         if (zone->prev_priority > priority)
1625                                 zone->prev_priority = priority;
1626                         sc.nr_scanned = 0;
1627                         sc.nr_reclaimed = 0;
1628                         sc.priority = priority;
1629                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1630                         shrink_zone(zone, &sc);
1631                         reclaim_state->reclaimed_slab = 0;
1632                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1633                                                 lru_pages);
1634                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1635                         total_reclaimed += sc.nr_reclaimed;
1636                         total_scanned += sc.nr_scanned;
1637                         if (zone->all_unreclaimable)
1638                                 continue;
1639                         if (nr_slab == 0 && zone->pages_scanned >=
1640                                     (zone->nr_active + zone->nr_inactive) * 4)
1641                                 zone->all_unreclaimable = 1;
1642                         /*
1643                          * If we've done a decent amount of scanning and
1644                          * the reclaim ratio is low, start doing writepage
1645                          * even in laptop mode
1646                          */
1647                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1648                             total_scanned > total_reclaimed+total_reclaimed/2)
1649                                 sc.may_writepage = 1;
1650                 }
1651                 if (nr_pages && to_free > total_reclaimed)
1652                         continue;       /* swsusp: need to do more work */
1653                 if (all_zones_ok)
1654                         break;          /* kswapd: all done */
1655                 /*
1656                  * OK, kswapd is getting into trouble.  Take a nap, then take
1657                  * another pass across the zones.
1658                  */
1659                 if (total_scanned && priority < DEF_PRIORITY - 2)
1660                         blk_congestion_wait(WRITE, HZ/10);
1661
1662                 /*
1663                  * We do this so kswapd doesn't build up large priorities for
1664                  * example when it is freeing in parallel with allocators. It
1665                  * matches the direct reclaim path behaviour in terms of impact
1666                  * on zone->*_priority.
1667                  */
1668                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1669                         break;
1670         }
1671 out:
1672         for (i = 0; i < pgdat->nr_zones; i++) {
1673                 struct zone *zone = pgdat->node_zones + i;
1674
1675                 zone->prev_priority = zone->temp_priority;
1676         }
1677         if (!all_zones_ok) {
1678                 cond_resched();
1679                 goto loop_again;
1680         }
1681
1682         return total_reclaimed;
1683 }
1684
1685 /*
1686  * The background pageout daemon, started as a kernel thread
1687  * from the init process. 
1688  *
1689  * This basically trickles out pages so that we have _some_
1690  * free memory available even if there is no other activity
1691  * that frees anything up. This is needed for things like routing
1692  * etc, where we otherwise might have all activity going on in
1693  * asynchronous contexts that cannot page things out.
1694  *
1695  * If there are applications that are active memory-allocators
1696  * (most normal use), this basically shouldn't matter.
1697  */
1698 static int kswapd(void *p)
1699 {
1700         unsigned long order;
1701         pg_data_t *pgdat = (pg_data_t*)p;
1702         struct task_struct *tsk = current;
1703         DEFINE_WAIT(wait);
1704         struct reclaim_state reclaim_state = {
1705                 .reclaimed_slab = 0,
1706         };
1707         cpumask_t cpumask;
1708
1709         daemonize("kswapd%d", pgdat->node_id);
1710         cpumask = node_to_cpumask(pgdat->node_id);
1711         if (!cpus_empty(cpumask))
1712                 set_cpus_allowed(tsk, cpumask);
1713         current->reclaim_state = &reclaim_state;
1714
1715         /*
1716          * Tell the memory management that we're a "memory allocator",
1717          * and that if we need more memory we should get access to it
1718          * regardless (see "__alloc_pages()"). "kswapd" should
1719          * never get caught in the normal page freeing logic.
1720          *
1721          * (Kswapd normally doesn't need memory anyway, but sometimes
1722          * you need a small amount of memory in order to be able to
1723          * page out something else, and this flag essentially protects
1724          * us from recursively trying to free more memory as we're
1725          * trying to free the first piece of memory in the first place).
1726          */
1727         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1728
1729         order = 0;
1730         for ( ; ; ) {
1731                 unsigned long new_order;
1732
1733                 try_to_freeze();
1734
1735                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1736                 new_order = pgdat->kswapd_max_order;
1737                 pgdat->kswapd_max_order = 0;
1738                 if (order < new_order) {
1739                         /*
1740                          * Don't sleep if someone wants a larger 'order'
1741                          * allocation
1742                          */
1743                         order = new_order;
1744                 } else {
1745                         schedule();
1746                         order = pgdat->kswapd_max_order;
1747                 }
1748                 finish_wait(&pgdat->kswapd_wait, &wait);
1749
1750                 balance_pgdat(pgdat, 0, order);
1751         }
1752         return 0;
1753 }
1754
1755 /*
1756  * A zone is low on free memory, so wake its kswapd task to service it.
1757  */
1758 void wakeup_kswapd(struct zone *zone, int order)
1759 {
1760         pg_data_t *pgdat;
1761
1762         if (!populated_zone(zone))
1763                 return;
1764
1765         pgdat = zone->zone_pgdat;
1766         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1767                 return;
1768         if (pgdat->kswapd_max_order < order)
1769                 pgdat->kswapd_max_order = order;
1770         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1771                 return;
1772         if (!waitqueue_active(&pgdat->kswapd_wait))
1773                 return;
1774         wake_up_interruptible(&pgdat->kswapd_wait);
1775 }
1776
1777 #ifdef CONFIG_PM
1778 /*
1779  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1780  * pages.
1781  */
1782 int shrink_all_memory(int nr_pages)
1783 {
1784         pg_data_t *pgdat;
1785         int nr_to_free = nr_pages;
1786         int ret = 0;
1787         struct reclaim_state reclaim_state = {
1788                 .reclaimed_slab = 0,
1789         };
1790
1791         current->reclaim_state = &reclaim_state;
1792         for_each_pgdat(pgdat) {
1793                 int freed;
1794                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1795                 ret += freed;
1796                 nr_to_free -= freed;
1797                 if (nr_to_free <= 0)
1798                         break;
1799         }
1800         current->reclaim_state = NULL;
1801         return ret;
1802 }
1803 #endif
1804
1805 #ifdef CONFIG_HOTPLUG_CPU
1806 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1807    not required for correctness.  So if the last cpu in a node goes
1808    away, we get changed to run anywhere: as the first one comes back,
1809    restore their cpu bindings. */
1810 static int __devinit cpu_callback(struct notifier_block *nfb,
1811                                   unsigned long action,
1812                                   void *hcpu)
1813 {
1814         pg_data_t *pgdat;
1815         cpumask_t mask;
1816
1817         if (action == CPU_ONLINE) {
1818                 for_each_pgdat(pgdat) {
1819                         mask = node_to_cpumask(pgdat->node_id);
1820                         if (any_online_cpu(mask) != NR_CPUS)
1821                                 /* One of our CPUs online: restore mask */
1822                                 set_cpus_allowed(pgdat->kswapd, mask);
1823                 }
1824         }
1825         return NOTIFY_OK;
1826 }
1827 #endif /* CONFIG_HOTPLUG_CPU */
1828
1829 static int __init kswapd_init(void)
1830 {
1831         pg_data_t *pgdat;
1832         swap_setup();
1833         for_each_pgdat(pgdat)
1834                 pgdat->kswapd
1835                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1836         total_memory = nr_free_pagecache_pages();
1837         hotcpu_notifier(cpu_callback, 0);
1838         return 0;
1839 }
1840
1841 module_init(kswapd_init)
1842
1843 #ifdef CONFIG_NUMA
1844 /*
1845  * Zone reclaim mode
1846  *
1847  * If non-zero call zone_reclaim when the number of free pages falls below
1848  * the watermarks.
1849  *
1850  * In the future we may add flags to the mode. However, the page allocator
1851  * should only have to check that zone_reclaim_mode != 0 before calling
1852  * zone_reclaim().
1853  */
1854 int zone_reclaim_mode __read_mostly;
1855
1856 #define RECLAIM_OFF 0
1857 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1858 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1859 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1860 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1861
1862 /*
1863  * Mininum time between zone reclaim scans
1864  */
1865 int zone_reclaim_interval __read_mostly = 30*HZ;
1866
1867 /*
1868  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1869  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1870  * a zone.
1871  */
1872 #define ZONE_RECLAIM_PRIORITY 4
1873
1874 /*
1875  * Try to free up some pages from this zone through reclaim.
1876  */
1877 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1878 {
1879         int nr_pages;
1880         struct task_struct *p = current;
1881         struct reclaim_state reclaim_state;
1882         struct scan_control sc;
1883         cpumask_t mask;
1884         int node_id;
1885
1886         if (time_before(jiffies,
1887                 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1888                         return 0;
1889
1890         if (!(gfp_mask & __GFP_WAIT) ||
1891                 zone->all_unreclaimable ||
1892                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1893                 (p->flags & PF_MEMALLOC))
1894                         return 0;
1895
1896         node_id = zone->zone_pgdat->node_id;
1897         mask = node_to_cpumask(node_id);
1898         if (!cpus_empty(mask) && node_id != numa_node_id())
1899                 return 0;
1900
1901         sc.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE);
1902         sc.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP);
1903         sc.nr_scanned = 0;
1904         sc.nr_reclaimed = 0;
1905         sc.priority = ZONE_RECLAIM_PRIORITY + 1;
1906         sc.nr_mapped = read_page_state(nr_mapped);
1907         sc.gfp_mask = gfp_mask;
1908
1909         disable_swap_token();
1910
1911         nr_pages = 1 << order;
1912         if (nr_pages > SWAP_CLUSTER_MAX)
1913                 sc.swap_cluster_max = nr_pages;
1914         else
1915                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1916
1917         cond_resched();
1918         /*
1919          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1920          * and we also need to be able to write out pages for RECLAIM_WRITE
1921          * and RECLAIM_SWAP.
1922          */
1923         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1924         reclaim_state.reclaimed_slab = 0;
1925         p->reclaim_state = &reclaim_state;
1926
1927         /*
1928          * Free memory by calling shrink zone with increasing priorities
1929          * until we have enough memory freed.
1930          */
1931         do {
1932                 sc.priority--;
1933                 shrink_zone(zone, &sc);
1934
1935         } while (sc.nr_reclaimed < nr_pages && sc.priority > 0);
1936
1937         if (sc.nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1938                 /*
1939                  * shrink_slab does not currently allow us to determine
1940                  * how many pages were freed in the zone. So we just
1941                  * shake the slab and then go offnode for a single allocation.
1942                  *
1943                  * shrink_slab will free memory on all zones and may take
1944                  * a long time.
1945                  */
1946                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1947         }
1948
1949         p->reclaim_state = NULL;
1950         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1951
1952         if (sc.nr_reclaimed == 0)
1953                 zone->last_unsuccessful_zone_reclaim = jiffies;
1954
1955         return sc.nr_reclaimed >= nr_pages;
1956 }
1957 #endif
1958