mm: clean up and kernelify shrinker registration
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
43
44 #include <linux/swapops.h>
45
46 #include "internal.h"
47
48 struct scan_control {
49         /* Incremented by the number of inactive pages that were scanned */
50         unsigned long nr_scanned;
51
52         /* This context's GFP mask */
53         gfp_t gfp_mask;
54
55         int may_writepage;
56
57         /* Can pages be swapped as part of reclaim? */
58         int may_swap;
59
60         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
61          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
62          * In this context, it doesn't matter that we scan the
63          * whole list at once. */
64         int swap_cluster_max;
65
66         int swappiness;
67
68         int all_unreclaimable;
69
70         int order;
71 };
72
73 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
74
75 #ifdef ARCH_HAS_PREFETCH
76 #define prefetch_prev_lru_page(_page, _base, _field)                    \
77         do {                                                            \
78                 if ((_page)->lru.prev != _base) {                       \
79                         struct page *prev;                              \
80                                                                         \
81                         prev = lru_to_page(&(_page->lru));              \
82                         prefetch(&prev->_field);                        \
83                 }                                                       \
84         } while (0)
85 #else
86 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
87 #endif
88
89 #ifdef ARCH_HAS_PREFETCHW
90 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
91         do {                                                            \
92                 if ((_page)->lru.prev != _base) {                       \
93                         struct page *prev;                              \
94                                                                         \
95                         prev = lru_to_page(&(_page->lru));              \
96                         prefetchw(&prev->_field);                       \
97                 }                                                       \
98         } while (0)
99 #else
100 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
101 #endif
102
103 /*
104  * From 0 .. 100.  Higher means more swappy.
105  */
106 int vm_swappiness = 60;
107 long vm_total_pages;    /* The total number of pages which the VM controls */
108
109 static LIST_HEAD(shrinker_list);
110 static DECLARE_RWSEM(shrinker_rwsem);
111
112 /*
113  * Add a shrinker callback to be called from the vm
114  */
115 void register_shrinker(struct shrinker *shrinker)
116 {
117         shrinker->nr = 0;
118         down_write(&shrinker_rwsem);
119         list_add_tail(&shrinker->list, &shrinker_list);
120         up_write(&shrinker_rwsem);
121 }
122 EXPORT_SYMBOL(register_shrinker);
123
124 /*
125  * Remove one
126  */
127 void unregister_shrinker(struct shrinker *shrinker)
128 {
129         down_write(&shrinker_rwsem);
130         list_del(&shrinker->list);
131         up_write(&shrinker_rwsem);
132 }
133 EXPORT_SYMBOL(unregister_shrinker);
134
135 #define SHRINK_BATCH 128
136 /*
137  * Call the shrink functions to age shrinkable caches
138  *
139  * Here we assume it costs one seek to replace a lru page and that it also
140  * takes a seek to recreate a cache object.  With this in mind we age equal
141  * percentages of the lru and ageable caches.  This should balance the seeks
142  * generated by these structures.
143  *
144  * If the vm encounted mapped pages on the LRU it increase the pressure on
145  * slab to avoid swapping.
146  *
147  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
148  *
149  * `lru_pages' represents the number of on-LRU pages in all the zones which
150  * are eligible for the caller's allocation attempt.  It is used for balancing
151  * slab reclaim versus page reclaim.
152  *
153  * Returns the number of slab objects which we shrunk.
154  */
155 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
156                         unsigned long lru_pages)
157 {
158         struct shrinker *shrinker;
159         unsigned long ret = 0;
160
161         if (scanned == 0)
162                 scanned = SWAP_CLUSTER_MAX;
163
164         if (!down_read_trylock(&shrinker_rwsem))
165                 return 1;       /* Assume we'll be able to shrink next time */
166
167         list_for_each_entry(shrinker, &shrinker_list, list) {
168                 unsigned long long delta;
169                 unsigned long total_scan;
170                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
171
172                 delta = (4 * scanned) / shrinker->seeks;
173                 delta *= max_pass;
174                 do_div(delta, lru_pages + 1);
175                 shrinker->nr += delta;
176                 if (shrinker->nr < 0) {
177                         printk(KERN_ERR "%s: nr=%ld\n",
178                                         __FUNCTION__, shrinker->nr);
179                         shrinker->nr = max_pass;
180                 }
181
182                 /*
183                  * Avoid risking looping forever due to too large nr value:
184                  * never try to free more than twice the estimate number of
185                  * freeable entries.
186                  */
187                 if (shrinker->nr > max_pass * 2)
188                         shrinker->nr = max_pass * 2;
189
190                 total_scan = shrinker->nr;
191                 shrinker->nr = 0;
192
193                 while (total_scan >= SHRINK_BATCH) {
194                         long this_scan = SHRINK_BATCH;
195                         int shrink_ret;
196                         int nr_before;
197
198                         nr_before = (*shrinker->shrink)(0, gfp_mask);
199                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
200                         if (shrink_ret == -1)
201                                 break;
202                         if (shrink_ret < nr_before)
203                                 ret += nr_before - shrink_ret;
204                         count_vm_events(SLABS_SCANNED, this_scan);
205                         total_scan -= this_scan;
206
207                         cond_resched();
208                 }
209
210                 shrinker->nr += total_scan;
211         }
212         up_read(&shrinker_rwsem);
213         return ret;
214 }
215
216 /* Called without lock on whether page is mapped, so answer is unstable */
217 static inline int page_mapping_inuse(struct page *page)
218 {
219         struct address_space *mapping;
220
221         /* Page is in somebody's page tables. */
222         if (page_mapped(page))
223                 return 1;
224
225         /* Be more reluctant to reclaim swapcache than pagecache */
226         if (PageSwapCache(page))
227                 return 1;
228
229         mapping = page_mapping(page);
230         if (!mapping)
231                 return 0;
232
233         /* File is mmap'd by somebody? */
234         return mapping_mapped(mapping);
235 }
236
237 static inline int is_page_cache_freeable(struct page *page)
238 {
239         return page_count(page) - !!PagePrivate(page) == 2;
240 }
241
242 static int may_write_to_queue(struct backing_dev_info *bdi)
243 {
244         if (current->flags & PF_SWAPWRITE)
245                 return 1;
246         if (!bdi_write_congested(bdi))
247                 return 1;
248         if (bdi == current->backing_dev_info)
249                 return 1;
250         return 0;
251 }
252
253 /*
254  * We detected a synchronous write error writing a page out.  Probably
255  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
256  * fsync(), msync() or close().
257  *
258  * The tricky part is that after writepage we cannot touch the mapping: nothing
259  * prevents it from being freed up.  But we have a ref on the page and once
260  * that page is locked, the mapping is pinned.
261  *
262  * We're allowed to run sleeping lock_page() here because we know the caller has
263  * __GFP_FS.
264  */
265 static void handle_write_error(struct address_space *mapping,
266                                 struct page *page, int error)
267 {
268         lock_page(page);
269         if (page_mapping(page) == mapping)
270                 mapping_set_error(mapping, error);
271         unlock_page(page);
272 }
273
274 /* possible outcome of pageout() */
275 typedef enum {
276         /* failed to write page out, page is locked */
277         PAGE_KEEP,
278         /* move page to the active list, page is locked */
279         PAGE_ACTIVATE,
280         /* page has been sent to the disk successfully, page is unlocked */
281         PAGE_SUCCESS,
282         /* page is clean and locked */
283         PAGE_CLEAN,
284 } pageout_t;
285
286 /*
287  * pageout is called by shrink_page_list() for each dirty page.
288  * Calls ->writepage().
289  */
290 static pageout_t pageout(struct page *page, struct address_space *mapping)
291 {
292         /*
293          * If the page is dirty, only perform writeback if that write
294          * will be non-blocking.  To prevent this allocation from being
295          * stalled by pagecache activity.  But note that there may be
296          * stalls if we need to run get_block().  We could test
297          * PagePrivate for that.
298          *
299          * If this process is currently in generic_file_write() against
300          * this page's queue, we can perform writeback even if that
301          * will block.
302          *
303          * If the page is swapcache, write it back even if that would
304          * block, for some throttling. This happens by accident, because
305          * swap_backing_dev_info is bust: it doesn't reflect the
306          * congestion state of the swapdevs.  Easy to fix, if needed.
307          * See swapfile.c:page_queue_congested().
308          */
309         if (!is_page_cache_freeable(page))
310                 return PAGE_KEEP;
311         if (!mapping) {
312                 /*
313                  * Some data journaling orphaned pages can have
314                  * page->mapping == NULL while being dirty with clean buffers.
315                  */
316                 if (PagePrivate(page)) {
317                         if (try_to_free_buffers(page)) {
318                                 ClearPageDirty(page);
319                                 printk("%s: orphaned page\n", __FUNCTION__);
320                                 return PAGE_CLEAN;
321                         }
322                 }
323                 return PAGE_KEEP;
324         }
325         if (mapping->a_ops->writepage == NULL)
326                 return PAGE_ACTIVATE;
327         if (!may_write_to_queue(mapping->backing_dev_info))
328                 return PAGE_KEEP;
329
330         if (clear_page_dirty_for_io(page)) {
331                 int res;
332                 struct writeback_control wbc = {
333                         .sync_mode = WB_SYNC_NONE,
334                         .nr_to_write = SWAP_CLUSTER_MAX,
335                         .range_start = 0,
336                         .range_end = LLONG_MAX,
337                         .nonblocking = 1,
338                         .for_reclaim = 1,
339                 };
340
341                 SetPageReclaim(page);
342                 res = mapping->a_ops->writepage(page, &wbc);
343                 if (res < 0)
344                         handle_write_error(mapping, page, res);
345                 if (res == AOP_WRITEPAGE_ACTIVATE) {
346                         ClearPageReclaim(page);
347                         return PAGE_ACTIVATE;
348                 }
349                 if (!PageWriteback(page)) {
350                         /* synchronous write or broken a_ops? */
351                         ClearPageReclaim(page);
352                 }
353                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
354                 return PAGE_SUCCESS;
355         }
356
357         return PAGE_CLEAN;
358 }
359
360 /*
361  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
362  * someone else has a ref on the page, abort and return 0.  If it was
363  * successfully detached, return 1.  Assumes the caller has a single ref on
364  * this page.
365  */
366 int remove_mapping(struct address_space *mapping, struct page *page)
367 {
368         BUG_ON(!PageLocked(page));
369         BUG_ON(mapping != page_mapping(page));
370
371         write_lock_irq(&mapping->tree_lock);
372         /*
373          * The non racy check for a busy page.
374          *
375          * Must be careful with the order of the tests. When someone has
376          * a ref to the page, it may be possible that they dirty it then
377          * drop the reference. So if PageDirty is tested before page_count
378          * here, then the following race may occur:
379          *
380          * get_user_pages(&page);
381          * [user mapping goes away]
382          * write_to(page);
383          *                              !PageDirty(page)    [good]
384          * SetPageDirty(page);
385          * put_page(page);
386          *                              !page_count(page)   [good, discard it]
387          *
388          * [oops, our write_to data is lost]
389          *
390          * Reversing the order of the tests ensures such a situation cannot
391          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
392          * load is not satisfied before that of page->_count.
393          *
394          * Note that if SetPageDirty is always performed via set_page_dirty,
395          * and thus under tree_lock, then this ordering is not required.
396          */
397         if (unlikely(page_count(page) != 2))
398                 goto cannot_free;
399         smp_rmb();
400         if (unlikely(PageDirty(page)))
401                 goto cannot_free;
402
403         if (PageSwapCache(page)) {
404                 swp_entry_t swap = { .val = page_private(page) };
405                 __delete_from_swap_cache(page);
406                 write_unlock_irq(&mapping->tree_lock);
407                 swap_free(swap);
408                 __put_page(page);       /* The pagecache ref */
409                 return 1;
410         }
411
412         __remove_from_page_cache(page);
413         write_unlock_irq(&mapping->tree_lock);
414         __put_page(page);
415         return 1;
416
417 cannot_free:
418         write_unlock_irq(&mapping->tree_lock);
419         return 0;
420 }
421
422 /*
423  * shrink_page_list() returns the number of reclaimed pages
424  */
425 static unsigned long shrink_page_list(struct list_head *page_list,
426                                         struct scan_control *sc)
427 {
428         LIST_HEAD(ret_pages);
429         struct pagevec freed_pvec;
430         int pgactivate = 0;
431         unsigned long nr_reclaimed = 0;
432
433         cond_resched();
434
435         pagevec_init(&freed_pvec, 1);
436         while (!list_empty(page_list)) {
437                 struct address_space *mapping;
438                 struct page *page;
439                 int may_enter_fs;
440                 int referenced;
441
442                 cond_resched();
443
444                 page = lru_to_page(page_list);
445                 list_del(&page->lru);
446
447                 if (TestSetPageLocked(page))
448                         goto keep;
449
450                 VM_BUG_ON(PageActive(page));
451
452                 sc->nr_scanned++;
453
454                 if (!sc->may_swap && page_mapped(page))
455                         goto keep_locked;
456
457                 /* Double the slab pressure for mapped and swapcache pages */
458                 if (page_mapped(page) || PageSwapCache(page))
459                         sc->nr_scanned++;
460
461                 if (PageWriteback(page))
462                         goto keep_locked;
463
464                 referenced = page_referenced(page, 1);
465                 /* In active use or really unfreeable?  Activate it. */
466                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
467                                         referenced && page_mapping_inuse(page))
468                         goto activate_locked;
469
470 #ifdef CONFIG_SWAP
471                 /*
472                  * Anonymous process memory has backing store?
473                  * Try to allocate it some swap space here.
474                  */
475                 if (PageAnon(page) && !PageSwapCache(page))
476                         if (!add_to_swap(page, GFP_ATOMIC))
477                                 goto activate_locked;
478 #endif /* CONFIG_SWAP */
479
480                 mapping = page_mapping(page);
481                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
482                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
483
484                 /*
485                  * The page is mapped into the page tables of one or more
486                  * processes. Try to unmap it here.
487                  */
488                 if (page_mapped(page) && mapping) {
489                         switch (try_to_unmap(page, 0)) {
490                         case SWAP_FAIL:
491                                 goto activate_locked;
492                         case SWAP_AGAIN:
493                                 goto keep_locked;
494                         case SWAP_SUCCESS:
495                                 ; /* try to free the page below */
496                         }
497                 }
498
499                 if (PageDirty(page)) {
500                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
501                                 goto keep_locked;
502                         if (!may_enter_fs)
503                                 goto keep_locked;
504                         if (!sc->may_writepage)
505                                 goto keep_locked;
506
507                         /* Page is dirty, try to write it out here */
508                         switch(pageout(page, mapping)) {
509                         case PAGE_KEEP:
510                                 goto keep_locked;
511                         case PAGE_ACTIVATE:
512                                 goto activate_locked;
513                         case PAGE_SUCCESS:
514                                 if (PageWriteback(page) || PageDirty(page))
515                                         goto keep;
516                                 /*
517                                  * A synchronous write - probably a ramdisk.  Go
518                                  * ahead and try to reclaim the page.
519                                  */
520                                 if (TestSetPageLocked(page))
521                                         goto keep;
522                                 if (PageDirty(page) || PageWriteback(page))
523                                         goto keep_locked;
524                                 mapping = page_mapping(page);
525                         case PAGE_CLEAN:
526                                 ; /* try to free the page below */
527                         }
528                 }
529
530                 /*
531                  * If the page has buffers, try to free the buffer mappings
532                  * associated with this page. If we succeed we try to free
533                  * the page as well.
534                  *
535                  * We do this even if the page is PageDirty().
536                  * try_to_release_page() does not perform I/O, but it is
537                  * possible for a page to have PageDirty set, but it is actually
538                  * clean (all its buffers are clean).  This happens if the
539                  * buffers were written out directly, with submit_bh(). ext3
540                  * will do this, as well as the blockdev mapping. 
541                  * try_to_release_page() will discover that cleanness and will
542                  * drop the buffers and mark the page clean - it can be freed.
543                  *
544                  * Rarely, pages can have buffers and no ->mapping.  These are
545                  * the pages which were not successfully invalidated in
546                  * truncate_complete_page().  We try to drop those buffers here
547                  * and if that worked, and the page is no longer mapped into
548                  * process address space (page_count == 1) it can be freed.
549                  * Otherwise, leave the page on the LRU so it is swappable.
550                  */
551                 if (PagePrivate(page)) {
552                         if (!try_to_release_page(page, sc->gfp_mask))
553                                 goto activate_locked;
554                         if (!mapping && page_count(page) == 1)
555                                 goto free_it;
556                 }
557
558                 if (!mapping || !remove_mapping(mapping, page))
559                         goto keep_locked;
560
561 free_it:
562                 unlock_page(page);
563                 nr_reclaimed++;
564                 if (!pagevec_add(&freed_pvec, page))
565                         __pagevec_release_nonlru(&freed_pvec);
566                 continue;
567
568 activate_locked:
569                 SetPageActive(page);
570                 pgactivate++;
571 keep_locked:
572                 unlock_page(page);
573 keep:
574                 list_add(&page->lru, &ret_pages);
575                 VM_BUG_ON(PageLRU(page));
576         }
577         list_splice(&ret_pages, page_list);
578         if (pagevec_count(&freed_pvec))
579                 __pagevec_release_nonlru(&freed_pvec);
580         count_vm_events(PGACTIVATE, pgactivate);
581         return nr_reclaimed;
582 }
583
584 /* LRU Isolation modes. */
585 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
586 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
587 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
588
589 /*
590  * Attempt to remove the specified page from its LRU.  Only take this page
591  * if it is of the appropriate PageActive status.  Pages which are being
592  * freed elsewhere are also ignored.
593  *
594  * page:        page to consider
595  * mode:        one of the LRU isolation modes defined above
596  *
597  * returns 0 on success, -ve errno on failure.
598  */
599 static int __isolate_lru_page(struct page *page, int mode)
600 {
601         int ret = -EINVAL;
602
603         /* Only take pages on the LRU. */
604         if (!PageLRU(page))
605                 return ret;
606
607         /*
608          * When checking the active state, we need to be sure we are
609          * dealing with comparible boolean values.  Take the logical not
610          * of each.
611          */
612         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
613                 return ret;
614
615         ret = -EBUSY;
616         if (likely(get_page_unless_zero(page))) {
617                 /*
618                  * Be careful not to clear PageLRU until after we're
619                  * sure the page is not being freed elsewhere -- the
620                  * page release code relies on it.
621                  */
622                 ClearPageLRU(page);
623                 ret = 0;
624         }
625
626         return ret;
627 }
628
629 /*
630  * zone->lru_lock is heavily contended.  Some of the functions that
631  * shrink the lists perform better by taking out a batch of pages
632  * and working on them outside the LRU lock.
633  *
634  * For pagecache intensive workloads, this function is the hottest
635  * spot in the kernel (apart from copy_*_user functions).
636  *
637  * Appropriate locks must be held before calling this function.
638  *
639  * @nr_to_scan: The number of pages to look through on the list.
640  * @src:        The LRU list to pull pages off.
641  * @dst:        The temp list to put pages on to.
642  * @scanned:    The number of pages that were scanned.
643  * @order:      The caller's attempted allocation order
644  * @mode:       One of the LRU isolation modes
645  *
646  * returns how many pages were moved onto *@dst.
647  */
648 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
649                 struct list_head *src, struct list_head *dst,
650                 unsigned long *scanned, int order, int mode)
651 {
652         unsigned long nr_taken = 0;
653         unsigned long scan;
654
655         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
656                 struct page *page;
657                 unsigned long pfn;
658                 unsigned long end_pfn;
659                 unsigned long page_pfn;
660                 int zone_id;
661
662                 page = lru_to_page(src);
663                 prefetchw_prev_lru_page(page, src, flags);
664
665                 VM_BUG_ON(!PageLRU(page));
666
667                 switch (__isolate_lru_page(page, mode)) {
668                 case 0:
669                         list_move(&page->lru, dst);
670                         nr_taken++;
671                         break;
672
673                 case -EBUSY:
674                         /* else it is being freed elsewhere */
675                         list_move(&page->lru, src);
676                         continue;
677
678                 default:
679                         BUG();
680                 }
681
682                 if (!order)
683                         continue;
684
685                 /*
686                  * Attempt to take all pages in the order aligned region
687                  * surrounding the tag page.  Only take those pages of
688                  * the same active state as that tag page.  We may safely
689                  * round the target page pfn down to the requested order
690                  * as the mem_map is guarenteed valid out to MAX_ORDER,
691                  * where that page is in a different zone we will detect
692                  * it from its zone id and abort this block scan.
693                  */
694                 zone_id = page_zone_id(page);
695                 page_pfn = page_to_pfn(page);
696                 pfn = page_pfn & ~((1 << order) - 1);
697                 end_pfn = pfn + (1 << order);
698                 for (; pfn < end_pfn; pfn++) {
699                         struct page *cursor_page;
700
701                         /* The target page is in the block, ignore it. */
702                         if (unlikely(pfn == page_pfn))
703                                 continue;
704
705                         /* Avoid holes within the zone. */
706                         if (unlikely(!pfn_valid_within(pfn)))
707                                 break;
708
709                         cursor_page = pfn_to_page(pfn);
710                         /* Check that we have not crossed a zone boundary. */
711                         if (unlikely(page_zone_id(cursor_page) != zone_id))
712                                 continue;
713                         switch (__isolate_lru_page(cursor_page, mode)) {
714                         case 0:
715                                 list_move(&cursor_page->lru, dst);
716                                 nr_taken++;
717                                 scan++;
718                                 break;
719
720                         case -EBUSY:
721                                 /* else it is being freed elsewhere */
722                                 list_move(&cursor_page->lru, src);
723                         default:
724                                 break;
725                         }
726                 }
727         }
728
729         *scanned = scan;
730         return nr_taken;
731 }
732
733 /*
734  * clear_active_flags() is a helper for shrink_active_list(), clearing
735  * any active bits from the pages in the list.
736  */
737 static unsigned long clear_active_flags(struct list_head *page_list)
738 {
739         int nr_active = 0;
740         struct page *page;
741
742         list_for_each_entry(page, page_list, lru)
743                 if (PageActive(page)) {
744                         ClearPageActive(page);
745                         nr_active++;
746                 }
747
748         return nr_active;
749 }
750
751 /*
752  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
753  * of reclaimed pages
754  */
755 static unsigned long shrink_inactive_list(unsigned long max_scan,
756                                 struct zone *zone, struct scan_control *sc)
757 {
758         LIST_HEAD(page_list);
759         struct pagevec pvec;
760         unsigned long nr_scanned = 0;
761         unsigned long nr_reclaimed = 0;
762
763         pagevec_init(&pvec, 1);
764
765         lru_add_drain();
766         spin_lock_irq(&zone->lru_lock);
767         do {
768                 struct page *page;
769                 unsigned long nr_taken;
770                 unsigned long nr_scan;
771                 unsigned long nr_freed;
772                 unsigned long nr_active;
773
774                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
775                              &zone->inactive_list,
776                              &page_list, &nr_scan, sc->order,
777                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
778                                              ISOLATE_BOTH : ISOLATE_INACTIVE);
779                 nr_active = clear_active_flags(&page_list);
780
781                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
782                 __mod_zone_page_state(zone, NR_INACTIVE,
783                                                 -(nr_taken - nr_active));
784                 zone->pages_scanned += nr_scan;
785                 spin_unlock_irq(&zone->lru_lock);
786
787                 nr_scanned += nr_scan;
788                 nr_freed = shrink_page_list(&page_list, sc);
789                 nr_reclaimed += nr_freed;
790                 local_irq_disable();
791                 if (current_is_kswapd()) {
792                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
793                         __count_vm_events(KSWAPD_STEAL, nr_freed);
794                 } else
795                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
796                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
797
798                 if (nr_taken == 0)
799                         goto done;
800
801                 spin_lock(&zone->lru_lock);
802                 /*
803                  * Put back any unfreeable pages.
804                  */
805                 while (!list_empty(&page_list)) {
806                         page = lru_to_page(&page_list);
807                         VM_BUG_ON(PageLRU(page));
808                         SetPageLRU(page);
809                         list_del(&page->lru);
810                         if (PageActive(page))
811                                 add_page_to_active_list(zone, page);
812                         else
813                                 add_page_to_inactive_list(zone, page);
814                         if (!pagevec_add(&pvec, page)) {
815                                 spin_unlock_irq(&zone->lru_lock);
816                                 __pagevec_release(&pvec);
817                                 spin_lock_irq(&zone->lru_lock);
818                         }
819                 }
820         } while (nr_scanned < max_scan);
821         spin_unlock(&zone->lru_lock);
822 done:
823         local_irq_enable();
824         pagevec_release(&pvec);
825         return nr_reclaimed;
826 }
827
828 /*
829  * We are about to scan this zone at a certain priority level.  If that priority
830  * level is smaller (ie: more urgent) than the previous priority, then note
831  * that priority level within the zone.  This is done so that when the next
832  * process comes in to scan this zone, it will immediately start out at this
833  * priority level rather than having to build up its own scanning priority.
834  * Here, this priority affects only the reclaim-mapped threshold.
835  */
836 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
837 {
838         if (priority < zone->prev_priority)
839                 zone->prev_priority = priority;
840 }
841
842 static inline int zone_is_near_oom(struct zone *zone)
843 {
844         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
845                                 + zone_page_state(zone, NR_INACTIVE))*3;
846 }
847
848 /*
849  * This moves pages from the active list to the inactive list.
850  *
851  * We move them the other way if the page is referenced by one or more
852  * processes, from rmap.
853  *
854  * If the pages are mostly unmapped, the processing is fast and it is
855  * appropriate to hold zone->lru_lock across the whole operation.  But if
856  * the pages are mapped, the processing is slow (page_referenced()) so we
857  * should drop zone->lru_lock around each page.  It's impossible to balance
858  * this, so instead we remove the pages from the LRU while processing them.
859  * It is safe to rely on PG_active against the non-LRU pages in here because
860  * nobody will play with that bit on a non-LRU page.
861  *
862  * The downside is that we have to touch page->_count against each page.
863  * But we had to alter page->flags anyway.
864  */
865 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
866                                 struct scan_control *sc, int priority)
867 {
868         unsigned long pgmoved;
869         int pgdeactivate = 0;
870         unsigned long pgscanned;
871         LIST_HEAD(l_hold);      /* The pages which were snipped off */
872         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
873         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
874         struct page *page;
875         struct pagevec pvec;
876         int reclaim_mapped = 0;
877
878         if (sc->may_swap) {
879                 long mapped_ratio;
880                 long distress;
881                 long swap_tendency;
882
883                 if (zone_is_near_oom(zone))
884                         goto force_reclaim_mapped;
885
886                 /*
887                  * `distress' is a measure of how much trouble we're having
888                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
889                  */
890                 distress = 100 >> min(zone->prev_priority, priority);
891
892                 /*
893                  * The point of this algorithm is to decide when to start
894                  * reclaiming mapped memory instead of just pagecache.  Work out
895                  * how much memory
896                  * is mapped.
897                  */
898                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
899                                 global_page_state(NR_ANON_PAGES)) * 100) /
900                                         vm_total_pages;
901
902                 /*
903                  * Now decide how much we really want to unmap some pages.  The
904                  * mapped ratio is downgraded - just because there's a lot of
905                  * mapped memory doesn't necessarily mean that page reclaim
906                  * isn't succeeding.
907                  *
908                  * The distress ratio is important - we don't want to start
909                  * going oom.
910                  *
911                  * A 100% value of vm_swappiness overrides this algorithm
912                  * altogether.
913                  */
914                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
915
916                 /*
917                  * Now use this metric to decide whether to start moving mapped
918                  * memory onto the inactive list.
919                  */
920                 if (swap_tendency >= 100)
921 force_reclaim_mapped:
922                         reclaim_mapped = 1;
923         }
924
925         lru_add_drain();
926         spin_lock_irq(&zone->lru_lock);
927         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
928                             &l_hold, &pgscanned, sc->order, ISOLATE_ACTIVE);
929         zone->pages_scanned += pgscanned;
930         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
931         spin_unlock_irq(&zone->lru_lock);
932
933         while (!list_empty(&l_hold)) {
934                 cond_resched();
935                 page = lru_to_page(&l_hold);
936                 list_del(&page->lru);
937                 if (page_mapped(page)) {
938                         if (!reclaim_mapped ||
939                             (total_swap_pages == 0 && PageAnon(page)) ||
940                             page_referenced(page, 0)) {
941                                 list_add(&page->lru, &l_active);
942                                 continue;
943                         }
944                 }
945                 list_add(&page->lru, &l_inactive);
946         }
947
948         pagevec_init(&pvec, 1);
949         pgmoved = 0;
950         spin_lock_irq(&zone->lru_lock);
951         while (!list_empty(&l_inactive)) {
952                 page = lru_to_page(&l_inactive);
953                 prefetchw_prev_lru_page(page, &l_inactive, flags);
954                 VM_BUG_ON(PageLRU(page));
955                 SetPageLRU(page);
956                 VM_BUG_ON(!PageActive(page));
957                 ClearPageActive(page);
958
959                 list_move(&page->lru, &zone->inactive_list);
960                 pgmoved++;
961                 if (!pagevec_add(&pvec, page)) {
962                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
963                         spin_unlock_irq(&zone->lru_lock);
964                         pgdeactivate += pgmoved;
965                         pgmoved = 0;
966                         if (buffer_heads_over_limit)
967                                 pagevec_strip(&pvec);
968                         __pagevec_release(&pvec);
969                         spin_lock_irq(&zone->lru_lock);
970                 }
971         }
972         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
973         pgdeactivate += pgmoved;
974         if (buffer_heads_over_limit) {
975                 spin_unlock_irq(&zone->lru_lock);
976                 pagevec_strip(&pvec);
977                 spin_lock_irq(&zone->lru_lock);
978         }
979
980         pgmoved = 0;
981         while (!list_empty(&l_active)) {
982                 page = lru_to_page(&l_active);
983                 prefetchw_prev_lru_page(page, &l_active, flags);
984                 VM_BUG_ON(PageLRU(page));
985                 SetPageLRU(page);
986                 VM_BUG_ON(!PageActive(page));
987                 list_move(&page->lru, &zone->active_list);
988                 pgmoved++;
989                 if (!pagevec_add(&pvec, page)) {
990                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
991                         pgmoved = 0;
992                         spin_unlock_irq(&zone->lru_lock);
993                         __pagevec_release(&pvec);
994                         spin_lock_irq(&zone->lru_lock);
995                 }
996         }
997         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
998
999         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1000         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1001         spin_unlock_irq(&zone->lru_lock);
1002
1003         pagevec_release(&pvec);
1004 }
1005
1006 /*
1007  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1008  */
1009 static unsigned long shrink_zone(int priority, struct zone *zone,
1010                                 struct scan_control *sc)
1011 {
1012         unsigned long nr_active;
1013         unsigned long nr_inactive;
1014         unsigned long nr_to_scan;
1015         unsigned long nr_reclaimed = 0;
1016
1017         atomic_inc(&zone->reclaim_in_progress);
1018
1019         /*
1020          * Add one to `nr_to_scan' just to make sure that the kernel will
1021          * slowly sift through the active list.
1022          */
1023         zone->nr_scan_active +=
1024                 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1025         nr_active = zone->nr_scan_active;
1026         if (nr_active >= sc->swap_cluster_max)
1027                 zone->nr_scan_active = 0;
1028         else
1029                 nr_active = 0;
1030
1031         zone->nr_scan_inactive +=
1032                 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1033         nr_inactive = zone->nr_scan_inactive;
1034         if (nr_inactive >= sc->swap_cluster_max)
1035                 zone->nr_scan_inactive = 0;
1036         else
1037                 nr_inactive = 0;
1038
1039         while (nr_active || nr_inactive) {
1040                 if (nr_active) {
1041                         nr_to_scan = min(nr_active,
1042                                         (unsigned long)sc->swap_cluster_max);
1043                         nr_active -= nr_to_scan;
1044                         shrink_active_list(nr_to_scan, zone, sc, priority);
1045                 }
1046
1047                 if (nr_inactive) {
1048                         nr_to_scan = min(nr_inactive,
1049                                         (unsigned long)sc->swap_cluster_max);
1050                         nr_inactive -= nr_to_scan;
1051                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1052                                                                 sc);
1053                 }
1054         }
1055
1056         throttle_vm_writeout(sc->gfp_mask);
1057
1058         atomic_dec(&zone->reclaim_in_progress);
1059         return nr_reclaimed;
1060 }
1061
1062 /*
1063  * This is the direct reclaim path, for page-allocating processes.  We only
1064  * try to reclaim pages from zones which will satisfy the caller's allocation
1065  * request.
1066  *
1067  * We reclaim from a zone even if that zone is over pages_high.  Because:
1068  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1069  *    allocation or
1070  * b) The zones may be over pages_high but they must go *over* pages_high to
1071  *    satisfy the `incremental min' zone defense algorithm.
1072  *
1073  * Returns the number of reclaimed pages.
1074  *
1075  * If a zone is deemed to be full of pinned pages then just give it a light
1076  * scan then give up on it.
1077  */
1078 static unsigned long shrink_zones(int priority, struct zone **zones,
1079                                         struct scan_control *sc)
1080 {
1081         unsigned long nr_reclaimed = 0;
1082         int i;
1083
1084         sc->all_unreclaimable = 1;
1085         for (i = 0; zones[i] != NULL; i++) {
1086                 struct zone *zone = zones[i];
1087
1088                 if (!populated_zone(zone))
1089                         continue;
1090
1091                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1092                         continue;
1093
1094                 note_zone_scanning_priority(zone, priority);
1095
1096                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1097                         continue;       /* Let kswapd poll it */
1098
1099                 sc->all_unreclaimable = 0;
1100
1101                 nr_reclaimed += shrink_zone(priority, zone, sc);
1102         }
1103         return nr_reclaimed;
1104 }
1105  
1106 /*
1107  * This is the main entry point to direct page reclaim.
1108  *
1109  * If a full scan of the inactive list fails to free enough memory then we
1110  * are "out of memory" and something needs to be killed.
1111  *
1112  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1113  * high - the zone may be full of dirty or under-writeback pages, which this
1114  * caller can't do much about.  We kick pdflush and take explicit naps in the
1115  * hope that some of these pages can be written.  But if the allocating task
1116  * holds filesystem locks which prevent writeout this might not work, and the
1117  * allocation attempt will fail.
1118  */
1119 unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1120 {
1121         int priority;
1122         int ret = 0;
1123         unsigned long total_scanned = 0;
1124         unsigned long nr_reclaimed = 0;
1125         struct reclaim_state *reclaim_state = current->reclaim_state;
1126         unsigned long lru_pages = 0;
1127         int i;
1128         struct scan_control sc = {
1129                 .gfp_mask = gfp_mask,
1130                 .may_writepage = !laptop_mode,
1131                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1132                 .may_swap = 1,
1133                 .swappiness = vm_swappiness,
1134                 .order = order,
1135         };
1136
1137         count_vm_event(ALLOCSTALL);
1138
1139         for (i = 0; zones[i] != NULL; i++) {
1140                 struct zone *zone = zones[i];
1141
1142                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1143                         continue;
1144
1145                 lru_pages += zone_page_state(zone, NR_ACTIVE)
1146                                 + zone_page_state(zone, NR_INACTIVE);
1147         }
1148
1149         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1150                 sc.nr_scanned = 0;
1151                 if (!priority)
1152                         disable_swap_token();
1153                 nr_reclaimed += shrink_zones(priority, zones, &sc);
1154                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1155                 if (reclaim_state) {
1156                         nr_reclaimed += reclaim_state->reclaimed_slab;
1157                         reclaim_state->reclaimed_slab = 0;
1158                 }
1159                 total_scanned += sc.nr_scanned;
1160                 if (nr_reclaimed >= sc.swap_cluster_max) {
1161                         ret = 1;
1162                         goto out;
1163                 }
1164
1165                 /*
1166                  * Try to write back as many pages as we just scanned.  This
1167                  * tends to cause slow streaming writers to write data to the
1168                  * disk smoothly, at the dirtying rate, which is nice.   But
1169                  * that's undesirable in laptop mode, where we *want* lumpy
1170                  * writeout.  So in laptop mode, write out the whole world.
1171                  */
1172                 if (total_scanned > sc.swap_cluster_max +
1173                                         sc.swap_cluster_max / 2) {
1174                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1175                         sc.may_writepage = 1;
1176                 }
1177
1178                 /* Take a nap, wait for some writeback to complete */
1179                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1180                         congestion_wait(WRITE, HZ/10);
1181         }
1182         /* top priority shrink_caches still had more to do? don't OOM, then */
1183         if (!sc.all_unreclaimable)
1184                 ret = 1;
1185 out:
1186         /*
1187          * Now that we've scanned all the zones at this priority level, note
1188          * that level within the zone so that the next thread which performs
1189          * scanning of this zone will immediately start out at this priority
1190          * level.  This affects only the decision whether or not to bring
1191          * mapped pages onto the inactive list.
1192          */
1193         if (priority < 0)
1194                 priority = 0;
1195         for (i = 0; zones[i] != 0; i++) {
1196                 struct zone *zone = zones[i];
1197
1198                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1199                         continue;
1200
1201                 zone->prev_priority = priority;
1202         }
1203         return ret;
1204 }
1205
1206 /*
1207  * For kswapd, balance_pgdat() will work across all this node's zones until
1208  * they are all at pages_high.
1209  *
1210  * Returns the number of pages which were actually freed.
1211  *
1212  * There is special handling here for zones which are full of pinned pages.
1213  * This can happen if the pages are all mlocked, or if they are all used by
1214  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1215  * What we do is to detect the case where all pages in the zone have been
1216  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1217  * dead and from now on, only perform a short scan.  Basically we're polling
1218  * the zone for when the problem goes away.
1219  *
1220  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1221  * zones which have free_pages > pages_high, but once a zone is found to have
1222  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1223  * of the number of free pages in the lower zones.  This interoperates with
1224  * the page allocator fallback scheme to ensure that aging of pages is balanced
1225  * across the zones.
1226  */
1227 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1228 {
1229         int all_zones_ok;
1230         int priority;
1231         int i;
1232         unsigned long total_scanned;
1233         unsigned long nr_reclaimed;
1234         struct reclaim_state *reclaim_state = current->reclaim_state;
1235         struct scan_control sc = {
1236                 .gfp_mask = GFP_KERNEL,
1237                 .may_swap = 1,
1238                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1239                 .swappiness = vm_swappiness,
1240                 .order = order,
1241         };
1242         /*
1243          * temp_priority is used to remember the scanning priority at which
1244          * this zone was successfully refilled to free_pages == pages_high.
1245          */
1246         int temp_priority[MAX_NR_ZONES];
1247
1248 loop_again:
1249         total_scanned = 0;
1250         nr_reclaimed = 0;
1251         sc.may_writepage = !laptop_mode;
1252         count_vm_event(PAGEOUTRUN);
1253
1254         for (i = 0; i < pgdat->nr_zones; i++)
1255                 temp_priority[i] = DEF_PRIORITY;
1256
1257         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1258                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1259                 unsigned long lru_pages = 0;
1260
1261                 /* The swap token gets in the way of swapout... */
1262                 if (!priority)
1263                         disable_swap_token();
1264
1265                 all_zones_ok = 1;
1266
1267                 /*
1268                  * Scan in the highmem->dma direction for the highest
1269                  * zone which needs scanning
1270                  */
1271                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1272                         struct zone *zone = pgdat->node_zones + i;
1273
1274                         if (!populated_zone(zone))
1275                                 continue;
1276
1277                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1278                                 continue;
1279
1280                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1281                                                0, 0)) {
1282                                 end_zone = i;
1283                                 break;
1284                         }
1285                 }
1286                 if (i < 0)
1287                         goto out;
1288
1289                 for (i = 0; i <= end_zone; i++) {
1290                         struct zone *zone = pgdat->node_zones + i;
1291
1292                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1293                                         + zone_page_state(zone, NR_INACTIVE);
1294                 }
1295
1296                 /*
1297                  * Now scan the zone in the dma->highmem direction, stopping
1298                  * at the last zone which needs scanning.
1299                  *
1300                  * We do this because the page allocator works in the opposite
1301                  * direction.  This prevents the page allocator from allocating
1302                  * pages behind kswapd's direction of progress, which would
1303                  * cause too much scanning of the lower zones.
1304                  */
1305                 for (i = 0; i <= end_zone; i++) {
1306                         struct zone *zone = pgdat->node_zones + i;
1307                         int nr_slab;
1308
1309                         if (!populated_zone(zone))
1310                                 continue;
1311
1312                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1313                                 continue;
1314
1315                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1316                                                end_zone, 0))
1317                                 all_zones_ok = 0;
1318                         temp_priority[i] = priority;
1319                         sc.nr_scanned = 0;
1320                         note_zone_scanning_priority(zone, priority);
1321                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1322                         reclaim_state->reclaimed_slab = 0;
1323                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1324                                                 lru_pages);
1325                         nr_reclaimed += reclaim_state->reclaimed_slab;
1326                         total_scanned += sc.nr_scanned;
1327                         if (zone->all_unreclaimable)
1328                                 continue;
1329                         if (nr_slab == 0 && zone->pages_scanned >=
1330                                 (zone_page_state(zone, NR_ACTIVE)
1331                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1332                                         zone->all_unreclaimable = 1;
1333                         /*
1334                          * If we've done a decent amount of scanning and
1335                          * the reclaim ratio is low, start doing writepage
1336                          * even in laptop mode
1337                          */
1338                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1339                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1340                                 sc.may_writepage = 1;
1341                 }
1342                 if (all_zones_ok)
1343                         break;          /* kswapd: all done */
1344                 /*
1345                  * OK, kswapd is getting into trouble.  Take a nap, then take
1346                  * another pass across the zones.
1347                  */
1348                 if (total_scanned && priority < DEF_PRIORITY - 2)
1349                         congestion_wait(WRITE, HZ/10);
1350
1351                 /*
1352                  * We do this so kswapd doesn't build up large priorities for
1353                  * example when it is freeing in parallel with allocators. It
1354                  * matches the direct reclaim path behaviour in terms of impact
1355                  * on zone->*_priority.
1356                  */
1357                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1358                         break;
1359         }
1360 out:
1361         /*
1362          * Note within each zone the priority level at which this zone was
1363          * brought into a happy state.  So that the next thread which scans this
1364          * zone will start out at that priority level.
1365          */
1366         for (i = 0; i < pgdat->nr_zones; i++) {
1367                 struct zone *zone = pgdat->node_zones + i;
1368
1369                 zone->prev_priority = temp_priority[i];
1370         }
1371         if (!all_zones_ok) {
1372                 cond_resched();
1373
1374                 try_to_freeze();
1375
1376                 goto loop_again;
1377         }
1378
1379         return nr_reclaimed;
1380 }
1381
1382 /*
1383  * The background pageout daemon, started as a kernel thread
1384  * from the init process. 
1385  *
1386  * This basically trickles out pages so that we have _some_
1387  * free memory available even if there is no other activity
1388  * that frees anything up. This is needed for things like routing
1389  * etc, where we otherwise might have all activity going on in
1390  * asynchronous contexts that cannot page things out.
1391  *
1392  * If there are applications that are active memory-allocators
1393  * (most normal use), this basically shouldn't matter.
1394  */
1395 static int kswapd(void *p)
1396 {
1397         unsigned long order;
1398         pg_data_t *pgdat = (pg_data_t*)p;
1399         struct task_struct *tsk = current;
1400         DEFINE_WAIT(wait);
1401         struct reclaim_state reclaim_state = {
1402                 .reclaimed_slab = 0,
1403         };
1404         cpumask_t cpumask;
1405
1406         cpumask = node_to_cpumask(pgdat->node_id);
1407         if (!cpus_empty(cpumask))
1408                 set_cpus_allowed(tsk, cpumask);
1409         current->reclaim_state = &reclaim_state;
1410
1411         /*
1412          * Tell the memory management that we're a "memory allocator",
1413          * and that if we need more memory we should get access to it
1414          * regardless (see "__alloc_pages()"). "kswapd" should
1415          * never get caught in the normal page freeing logic.
1416          *
1417          * (Kswapd normally doesn't need memory anyway, but sometimes
1418          * you need a small amount of memory in order to be able to
1419          * page out something else, and this flag essentially protects
1420          * us from recursively trying to free more memory as we're
1421          * trying to free the first piece of memory in the first place).
1422          */
1423         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1424
1425         order = 0;
1426         for ( ; ; ) {
1427                 unsigned long new_order;
1428
1429                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1430                 new_order = pgdat->kswapd_max_order;
1431                 pgdat->kswapd_max_order = 0;
1432                 if (order < new_order) {
1433                         /*
1434                          * Don't sleep if someone wants a larger 'order'
1435                          * allocation
1436                          */
1437                         order = new_order;
1438                 } else {
1439                         if (!freezing(current))
1440                                 schedule();
1441
1442                         order = pgdat->kswapd_max_order;
1443                 }
1444                 finish_wait(&pgdat->kswapd_wait, &wait);
1445
1446                 if (!try_to_freeze()) {
1447                         /* We can speed up thawing tasks if we don't call
1448                          * balance_pgdat after returning from the refrigerator
1449                          */
1450                         balance_pgdat(pgdat, order);
1451                 }
1452         }
1453         return 0;
1454 }
1455
1456 /*
1457  * A zone is low on free memory, so wake its kswapd task to service it.
1458  */
1459 void wakeup_kswapd(struct zone *zone, int order)
1460 {
1461         pg_data_t *pgdat;
1462
1463         if (!populated_zone(zone))
1464                 return;
1465
1466         pgdat = zone->zone_pgdat;
1467         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1468                 return;
1469         if (pgdat->kswapd_max_order < order)
1470                 pgdat->kswapd_max_order = order;
1471         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1472                 return;
1473         if (!waitqueue_active(&pgdat->kswapd_wait))
1474                 return;
1475         wake_up_interruptible(&pgdat->kswapd_wait);
1476 }
1477
1478 #ifdef CONFIG_PM
1479 /*
1480  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1481  * from LRU lists system-wide, for given pass and priority, and returns the
1482  * number of reclaimed pages
1483  *
1484  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1485  */
1486 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1487                                       int pass, struct scan_control *sc)
1488 {
1489         struct zone *zone;
1490         unsigned long nr_to_scan, ret = 0;
1491
1492         for_each_zone(zone) {
1493
1494                 if (!populated_zone(zone))
1495                         continue;
1496
1497                 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1498                         continue;
1499
1500                 /* For pass = 0 we don't shrink the active list */
1501                 if (pass > 0) {
1502                         zone->nr_scan_active +=
1503                                 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1504                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1505                                 zone->nr_scan_active = 0;
1506                                 nr_to_scan = min(nr_pages,
1507                                         zone_page_state(zone, NR_ACTIVE));
1508                                 shrink_active_list(nr_to_scan, zone, sc, prio);
1509                         }
1510                 }
1511
1512                 zone->nr_scan_inactive +=
1513                         (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1514                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1515                         zone->nr_scan_inactive = 0;
1516                         nr_to_scan = min(nr_pages,
1517                                 zone_page_state(zone, NR_INACTIVE));
1518                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1519                         if (ret >= nr_pages)
1520                                 return ret;
1521                 }
1522         }
1523
1524         return ret;
1525 }
1526
1527 static unsigned long count_lru_pages(void)
1528 {
1529         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1530 }
1531
1532 /*
1533  * Try to free `nr_pages' of memory, system-wide, and return the number of
1534  * freed pages.
1535  *
1536  * Rather than trying to age LRUs the aim is to preserve the overall
1537  * LRU order by reclaiming preferentially
1538  * inactive > active > active referenced > active mapped
1539  */
1540 unsigned long shrink_all_memory(unsigned long nr_pages)
1541 {
1542         unsigned long lru_pages, nr_slab;
1543         unsigned long ret = 0;
1544         int pass;
1545         struct reclaim_state reclaim_state;
1546         struct scan_control sc = {
1547                 .gfp_mask = GFP_KERNEL,
1548                 .may_swap = 0,
1549                 .swap_cluster_max = nr_pages,
1550                 .may_writepage = 1,
1551                 .swappiness = vm_swappiness,
1552         };
1553
1554         current->reclaim_state = &reclaim_state;
1555
1556         lru_pages = count_lru_pages();
1557         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1558         /* If slab caches are huge, it's better to hit them first */
1559         while (nr_slab >= lru_pages) {
1560                 reclaim_state.reclaimed_slab = 0;
1561                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1562                 if (!reclaim_state.reclaimed_slab)
1563                         break;
1564
1565                 ret += reclaim_state.reclaimed_slab;
1566                 if (ret >= nr_pages)
1567                         goto out;
1568
1569                 nr_slab -= reclaim_state.reclaimed_slab;
1570         }
1571
1572         /*
1573          * We try to shrink LRUs in 5 passes:
1574          * 0 = Reclaim from inactive_list only
1575          * 1 = Reclaim from active list but don't reclaim mapped
1576          * 2 = 2nd pass of type 1
1577          * 3 = Reclaim mapped (normal reclaim)
1578          * 4 = 2nd pass of type 3
1579          */
1580         for (pass = 0; pass < 5; pass++) {
1581                 int prio;
1582
1583                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1584                 if (pass > 2) {
1585                         sc.may_swap = 1;
1586                         sc.swappiness = 100;
1587                 }
1588
1589                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1590                         unsigned long nr_to_scan = nr_pages - ret;
1591
1592                         sc.nr_scanned = 0;
1593                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1594                         if (ret >= nr_pages)
1595                                 goto out;
1596
1597                         reclaim_state.reclaimed_slab = 0;
1598                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1599                                         count_lru_pages());
1600                         ret += reclaim_state.reclaimed_slab;
1601                         if (ret >= nr_pages)
1602                                 goto out;
1603
1604                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1605                                 congestion_wait(WRITE, HZ / 10);
1606                 }
1607         }
1608
1609         /*
1610          * If ret = 0, we could not shrink LRUs, but there may be something
1611          * in slab caches
1612          */
1613         if (!ret) {
1614                 do {
1615                         reclaim_state.reclaimed_slab = 0;
1616                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1617                         ret += reclaim_state.reclaimed_slab;
1618                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1619         }
1620
1621 out:
1622         current->reclaim_state = NULL;
1623
1624         return ret;
1625 }
1626 #endif
1627
1628 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1629    not required for correctness.  So if the last cpu in a node goes
1630    away, we get changed to run anywhere: as the first one comes back,
1631    restore their cpu bindings. */
1632 static int __devinit cpu_callback(struct notifier_block *nfb,
1633                                   unsigned long action, void *hcpu)
1634 {
1635         pg_data_t *pgdat;
1636         cpumask_t mask;
1637
1638         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1639                 for_each_online_pgdat(pgdat) {
1640                         mask = node_to_cpumask(pgdat->node_id);
1641                         if (any_online_cpu(mask) != NR_CPUS)
1642                                 /* One of our CPUs online: restore mask */
1643                                 set_cpus_allowed(pgdat->kswapd, mask);
1644                 }
1645         }
1646         return NOTIFY_OK;
1647 }
1648
1649 /*
1650  * This kswapd start function will be called by init and node-hot-add.
1651  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1652  */
1653 int kswapd_run(int nid)
1654 {
1655         pg_data_t *pgdat = NODE_DATA(nid);
1656         int ret = 0;
1657
1658         if (pgdat->kswapd)
1659                 return 0;
1660
1661         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1662         if (IS_ERR(pgdat->kswapd)) {
1663                 /* failure at boot is fatal */
1664                 BUG_ON(system_state == SYSTEM_BOOTING);
1665                 printk("Failed to start kswapd on node %d\n",nid);
1666                 ret = -1;
1667         }
1668         return ret;
1669 }
1670
1671 static int __init kswapd_init(void)
1672 {
1673         int nid;
1674
1675         swap_setup();
1676         for_each_online_node(nid)
1677                 kswapd_run(nid);
1678         hotcpu_notifier(cpu_callback, 0);
1679         return 0;
1680 }
1681
1682 module_init(kswapd_init)
1683
1684 #ifdef CONFIG_NUMA
1685 /*
1686  * Zone reclaim mode
1687  *
1688  * If non-zero call zone_reclaim when the number of free pages falls below
1689  * the watermarks.
1690  */
1691 int zone_reclaim_mode __read_mostly;
1692
1693 #define RECLAIM_OFF 0
1694 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1695 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1696 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1697
1698 /*
1699  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1700  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1701  * a zone.
1702  */
1703 #define ZONE_RECLAIM_PRIORITY 4
1704
1705 /*
1706  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1707  * occur.
1708  */
1709 int sysctl_min_unmapped_ratio = 1;
1710
1711 /*
1712  * If the number of slab pages in a zone grows beyond this percentage then
1713  * slab reclaim needs to occur.
1714  */
1715 int sysctl_min_slab_ratio = 5;
1716
1717 /*
1718  * Try to free up some pages from this zone through reclaim.
1719  */
1720 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1721 {
1722         /* Minimum pages needed in order to stay on node */
1723         const unsigned long nr_pages = 1 << order;
1724         struct task_struct *p = current;
1725         struct reclaim_state reclaim_state;
1726         int priority;
1727         unsigned long nr_reclaimed = 0;
1728         struct scan_control sc = {
1729                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1730                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1731                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1732                                         SWAP_CLUSTER_MAX),
1733                 .gfp_mask = gfp_mask,
1734                 .swappiness = vm_swappiness,
1735         };
1736         unsigned long slab_reclaimable;
1737
1738         disable_swap_token();
1739         cond_resched();
1740         /*
1741          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1742          * and we also need to be able to write out pages for RECLAIM_WRITE
1743          * and RECLAIM_SWAP.
1744          */
1745         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1746         reclaim_state.reclaimed_slab = 0;
1747         p->reclaim_state = &reclaim_state;
1748
1749         if (zone_page_state(zone, NR_FILE_PAGES) -
1750                 zone_page_state(zone, NR_FILE_MAPPED) >
1751                 zone->min_unmapped_pages) {
1752                 /*
1753                  * Free memory by calling shrink zone with increasing
1754                  * priorities until we have enough memory freed.
1755                  */
1756                 priority = ZONE_RECLAIM_PRIORITY;
1757                 do {
1758                         note_zone_scanning_priority(zone, priority);
1759                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1760                         priority--;
1761                 } while (priority >= 0 && nr_reclaimed < nr_pages);
1762         }
1763
1764         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1765         if (slab_reclaimable > zone->min_slab_pages) {
1766                 /*
1767                  * shrink_slab() does not currently allow us to determine how
1768                  * many pages were freed in this zone. So we take the current
1769                  * number of slab pages and shake the slab until it is reduced
1770                  * by the same nr_pages that we used for reclaiming unmapped
1771                  * pages.
1772                  *
1773                  * Note that shrink_slab will free memory on all zones and may
1774                  * take a long time.
1775                  */
1776                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
1777                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1778                                 slab_reclaimable - nr_pages)
1779                         ;
1780
1781                 /*
1782                  * Update nr_reclaimed by the number of slab pages we
1783                  * reclaimed from this zone.
1784                  */
1785                 nr_reclaimed += slab_reclaimable -
1786                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1787         }
1788
1789         p->reclaim_state = NULL;
1790         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1791         return nr_reclaimed >= nr_pages;
1792 }
1793
1794 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1795 {
1796         cpumask_t mask;
1797         int node_id;
1798
1799         /*
1800          * Zone reclaim reclaims unmapped file backed pages and
1801          * slab pages if we are over the defined limits.
1802          *
1803          * A small portion of unmapped file backed pages is needed for
1804          * file I/O otherwise pages read by file I/O will be immediately
1805          * thrown out if the zone is overallocated. So we do not reclaim
1806          * if less than a specified percentage of the zone is used by
1807          * unmapped file backed pages.
1808          */
1809         if (zone_page_state(zone, NR_FILE_PAGES) -
1810             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1811             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1812                         <= zone->min_slab_pages)
1813                 return 0;
1814
1815         /*
1816          * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1817          * not have reclaimable pages and if we should not delay the allocation
1818          * then do not scan.
1819          */
1820         if (!(gfp_mask & __GFP_WAIT) ||
1821                 zone->all_unreclaimable ||
1822                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1823                 (current->flags & PF_MEMALLOC))
1824                         return 0;
1825
1826         /*
1827          * Only run zone reclaim on the local zone or on zones that do not
1828          * have associated processors. This will favor the local processor
1829          * over remote processors and spread off node memory allocations
1830          * as wide as possible.
1831          */
1832         node_id = zone_to_nid(zone);
1833         mask = node_to_cpumask(node_id);
1834         if (!cpus_empty(mask) && node_id != numa_node_id())
1835                 return 0;
1836         return __zone_reclaim(zone, gfp_mask, order);
1837 }
1838 #endif