46be8a02280e70f1526e518a820340977a3afc69
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
26                                         buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
36 #include <linux/delay.h>
37
38 #include <asm/tlbflush.h>
39 #include <asm/div64.h>
40
41 #include <linux/swapops.h>
42
43 #include "internal.h"
44
45 struct scan_control {
46         /* Incremented by the number of inactive pages that were scanned */
47         unsigned long nr_scanned;
48
49         unsigned long nr_mapped;        /* From page_state */
50
51         /* This context's GFP mask */
52         gfp_t gfp_mask;
53
54         int may_writepage;
55
56         /* Can pages be swapped as part of reclaim? */
57         int may_swap;
58
59         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
60          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
61          * In this context, it doesn't matter that we scan the
62          * whole list at once. */
63         int swap_cluster_max;
64
65         int swappiness;
66 };
67
68 /*
69  * The list of shrinker callbacks used by to apply pressure to
70  * ageable caches.
71  */
72 struct shrinker {
73         shrinker_t              shrinker;
74         struct list_head        list;
75         int                     seeks;  /* seeks to recreate an obj */
76         long                    nr;     /* objs pending delete */
77 };
78
79 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
80
81 #ifdef ARCH_HAS_PREFETCH
82 #define prefetch_prev_lru_page(_page, _base, _field)                    \
83         do {                                                            \
84                 if ((_page)->lru.prev != _base) {                       \
85                         struct page *prev;                              \
86                                                                         \
87                         prev = lru_to_page(&(_page->lru));              \
88                         prefetch(&prev->_field);                        \
89                 }                                                       \
90         } while (0)
91 #else
92 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
93 #endif
94
95 #ifdef ARCH_HAS_PREFETCHW
96 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
97         do {                                                            \
98                 if ((_page)->lru.prev != _base) {                       \
99                         struct page *prev;                              \
100                                                                         \
101                         prev = lru_to_page(&(_page->lru));              \
102                         prefetchw(&prev->_field);                       \
103                 }                                                       \
104         } while (0)
105 #else
106 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
107 #endif
108
109 /*
110  * From 0 .. 100.  Higher means more swappy.
111  */
112 int vm_swappiness = 60;
113 static long total_memory;
114
115 static LIST_HEAD(shrinker_list);
116 static DECLARE_RWSEM(shrinker_rwsem);
117
118 /*
119  * Add a shrinker callback to be called from the vm
120  */
121 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
122 {
123         struct shrinker *shrinker;
124
125         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
126         if (shrinker) {
127                 shrinker->shrinker = theshrinker;
128                 shrinker->seeks = seeks;
129                 shrinker->nr = 0;
130                 down_write(&shrinker_rwsem);
131                 list_add_tail(&shrinker->list, &shrinker_list);
132                 up_write(&shrinker_rwsem);
133         }
134         return shrinker;
135 }
136 EXPORT_SYMBOL(set_shrinker);
137
138 /*
139  * Remove one
140  */
141 void remove_shrinker(struct shrinker *shrinker)
142 {
143         down_write(&shrinker_rwsem);
144         list_del(&shrinker->list);
145         up_write(&shrinker_rwsem);
146         kfree(shrinker);
147 }
148 EXPORT_SYMBOL(remove_shrinker);
149
150 #define SHRINK_BATCH 128
151 /*
152  * Call the shrink functions to age shrinkable caches
153  *
154  * Here we assume it costs one seek to replace a lru page and that it also
155  * takes a seek to recreate a cache object.  With this in mind we age equal
156  * percentages of the lru and ageable caches.  This should balance the seeks
157  * generated by these structures.
158  *
159  * If the vm encounted mapped pages on the LRU it increase the pressure on
160  * slab to avoid swapping.
161  *
162  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
163  *
164  * `lru_pages' represents the number of on-LRU pages in all the zones which
165  * are eligible for the caller's allocation attempt.  It is used for balancing
166  * slab reclaim versus page reclaim.
167  *
168  * Returns the number of slab objects which we shrunk.
169  */
170 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
171                         unsigned long lru_pages)
172 {
173         struct shrinker *shrinker;
174         unsigned long ret = 0;
175
176         if (scanned == 0)
177                 scanned = SWAP_CLUSTER_MAX;
178
179         if (!down_read_trylock(&shrinker_rwsem))
180                 return 1;       /* Assume we'll be able to shrink next time */
181
182         list_for_each_entry(shrinker, &shrinker_list, list) {
183                 unsigned long long delta;
184                 unsigned long total_scan;
185                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
186
187                 delta = (4 * scanned) / shrinker->seeks;
188                 delta *= max_pass;
189                 do_div(delta, lru_pages + 1);
190                 shrinker->nr += delta;
191                 if (shrinker->nr < 0) {
192                         printk(KERN_ERR "%s: nr=%ld\n",
193                                         __FUNCTION__, shrinker->nr);
194                         shrinker->nr = max_pass;
195                 }
196
197                 /*
198                  * Avoid risking looping forever due to too large nr value:
199                  * never try to free more than twice the estimate number of
200                  * freeable entries.
201                  */
202                 if (shrinker->nr > max_pass * 2)
203                         shrinker->nr = max_pass * 2;
204
205                 total_scan = shrinker->nr;
206                 shrinker->nr = 0;
207
208                 while (total_scan >= SHRINK_BATCH) {
209                         long this_scan = SHRINK_BATCH;
210                         int shrink_ret;
211                         int nr_before;
212
213                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
214                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
215                         if (shrink_ret == -1)
216                                 break;
217                         if (shrink_ret < nr_before)
218                                 ret += nr_before - shrink_ret;
219                         mod_page_state(slabs_scanned, this_scan);
220                         total_scan -= this_scan;
221
222                         cond_resched();
223                 }
224
225                 shrinker->nr += total_scan;
226         }
227         up_read(&shrinker_rwsem);
228         return ret;
229 }
230
231 /* Called without lock on whether page is mapped, so answer is unstable */
232 static inline int page_mapping_inuse(struct page *page)
233 {
234         struct address_space *mapping;
235
236         /* Page is in somebody's page tables. */
237         if (page_mapped(page))
238                 return 1;
239
240         /* Be more reluctant to reclaim swapcache than pagecache */
241         if (PageSwapCache(page))
242                 return 1;
243
244         mapping = page_mapping(page);
245         if (!mapping)
246                 return 0;
247
248         /* File is mmap'd by somebody? */
249         return mapping_mapped(mapping);
250 }
251
252 static inline int is_page_cache_freeable(struct page *page)
253 {
254         return page_count(page) - !!PagePrivate(page) == 2;
255 }
256
257 static int may_write_to_queue(struct backing_dev_info *bdi)
258 {
259         if (current->flags & PF_SWAPWRITE)
260                 return 1;
261         if (!bdi_write_congested(bdi))
262                 return 1;
263         if (bdi == current->backing_dev_info)
264                 return 1;
265         return 0;
266 }
267
268 /*
269  * We detected a synchronous write error writing a page out.  Probably
270  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
271  * fsync(), msync() or close().
272  *
273  * The tricky part is that after writepage we cannot touch the mapping: nothing
274  * prevents it from being freed up.  But we have a ref on the page and once
275  * that page is locked, the mapping is pinned.
276  *
277  * We're allowed to run sleeping lock_page() here because we know the caller has
278  * __GFP_FS.
279  */
280 static void handle_write_error(struct address_space *mapping,
281                                 struct page *page, int error)
282 {
283         lock_page(page);
284         if (page_mapping(page) == mapping) {
285                 if (error == -ENOSPC)
286                         set_bit(AS_ENOSPC, &mapping->flags);
287                 else
288                         set_bit(AS_EIO, &mapping->flags);
289         }
290         unlock_page(page);
291 }
292
293 /*
294  * pageout is called by shrink_page_list() for each dirty page.
295  * Calls ->writepage().
296  */
297 pageout_t pageout(struct page *page, struct address_space *mapping)
298 {
299         /*
300          * If the page is dirty, only perform writeback if that write
301          * will be non-blocking.  To prevent this allocation from being
302          * stalled by pagecache activity.  But note that there may be
303          * stalls if we need to run get_block().  We could test
304          * PagePrivate for that.
305          *
306          * If this process is currently in generic_file_write() against
307          * this page's queue, we can perform writeback even if that
308          * will block.
309          *
310          * If the page is swapcache, write it back even if that would
311          * block, for some throttling. This happens by accident, because
312          * swap_backing_dev_info is bust: it doesn't reflect the
313          * congestion state of the swapdevs.  Easy to fix, if needed.
314          * See swapfile.c:page_queue_congested().
315          */
316         if (!is_page_cache_freeable(page))
317                 return PAGE_KEEP;
318         if (!mapping) {
319                 /*
320                  * Some data journaling orphaned pages can have
321                  * page->mapping == NULL while being dirty with clean buffers.
322                  */
323                 if (PagePrivate(page)) {
324                         if (try_to_free_buffers(page)) {
325                                 ClearPageDirty(page);
326                                 printk("%s: orphaned page\n", __FUNCTION__);
327                                 return PAGE_CLEAN;
328                         }
329                 }
330                 return PAGE_KEEP;
331         }
332         if (mapping->a_ops->writepage == NULL)
333                 return PAGE_ACTIVATE;
334         if (!may_write_to_queue(mapping->backing_dev_info))
335                 return PAGE_KEEP;
336
337         if (clear_page_dirty_for_io(page)) {
338                 int res;
339                 struct writeback_control wbc = {
340                         .sync_mode = WB_SYNC_NONE,
341                         .nr_to_write = SWAP_CLUSTER_MAX,
342                         .nonblocking = 1,
343                         .for_reclaim = 1,
344                 };
345
346                 SetPageReclaim(page);
347                 res = mapping->a_ops->writepage(page, &wbc);
348                 if (res < 0)
349                         handle_write_error(mapping, page, res);
350                 if (res == AOP_WRITEPAGE_ACTIVATE) {
351                         ClearPageReclaim(page);
352                         return PAGE_ACTIVATE;
353                 }
354                 if (!PageWriteback(page)) {
355                         /* synchronous write or broken a_ops? */
356                         ClearPageReclaim(page);
357                 }
358
359                 return PAGE_SUCCESS;
360         }
361
362         return PAGE_CLEAN;
363 }
364
365 int remove_mapping(struct address_space *mapping, struct page *page)
366 {
367         if (!mapping)
368                 return 0;               /* truncate got there first */
369
370         write_lock_irq(&mapping->tree_lock);
371
372         /*
373          * The non-racy check for busy page.  It is critical to check
374          * PageDirty _after_ making sure that the page is freeable and
375          * not in use by anybody.       (pagecache + us == 2)
376          */
377         if (unlikely(page_count(page) != 2))
378                 goto cannot_free;
379         smp_rmb();
380         if (unlikely(PageDirty(page)))
381                 goto cannot_free;
382
383         if (PageSwapCache(page)) {
384                 swp_entry_t swap = { .val = page_private(page) };
385                 __delete_from_swap_cache(page);
386                 write_unlock_irq(&mapping->tree_lock);
387                 swap_free(swap);
388                 __put_page(page);       /* The pagecache ref */
389                 return 1;
390         }
391
392         __remove_from_page_cache(page);
393         write_unlock_irq(&mapping->tree_lock);
394         __put_page(page);
395         return 1;
396
397 cannot_free:
398         write_unlock_irq(&mapping->tree_lock);
399         return 0;
400 }
401
402 /*
403  * shrink_page_list() returns the number of reclaimed pages
404  */
405 static unsigned long shrink_page_list(struct list_head *page_list,
406                                         struct scan_control *sc)
407 {
408         LIST_HEAD(ret_pages);
409         struct pagevec freed_pvec;
410         int pgactivate = 0;
411         unsigned long nr_reclaimed = 0;
412
413         cond_resched();
414
415         pagevec_init(&freed_pvec, 1);
416         while (!list_empty(page_list)) {
417                 struct address_space *mapping;
418                 struct page *page;
419                 int may_enter_fs;
420                 int referenced;
421
422                 cond_resched();
423
424                 page = lru_to_page(page_list);
425                 list_del(&page->lru);
426
427                 if (TestSetPageLocked(page))
428                         goto keep;
429
430                 BUG_ON(PageActive(page));
431
432                 sc->nr_scanned++;
433
434                 if (!sc->may_swap && page_mapped(page))
435                         goto keep_locked;
436
437                 /* Double the slab pressure for mapped and swapcache pages */
438                 if (page_mapped(page) || PageSwapCache(page))
439                         sc->nr_scanned++;
440
441                 if (PageWriteback(page))
442                         goto keep_locked;
443
444                 referenced = page_referenced(page, 1);
445                 /* In active use or really unfreeable?  Activate it. */
446                 if (referenced && page_mapping_inuse(page))
447                         goto activate_locked;
448
449 #ifdef CONFIG_SWAP
450                 /*
451                  * Anonymous process memory has backing store?
452                  * Try to allocate it some swap space here.
453                  */
454                 if (PageAnon(page) && !PageSwapCache(page))
455                         if (!add_to_swap(page, GFP_ATOMIC))
456                                 goto activate_locked;
457 #endif /* CONFIG_SWAP */
458
459                 mapping = page_mapping(page);
460                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
461                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
462
463                 /*
464                  * The page is mapped into the page tables of one or more
465                  * processes. Try to unmap it here.
466                  */
467                 if (page_mapped(page) && mapping) {
468                         switch (try_to_unmap(page, 0)) {
469                         case SWAP_FAIL:
470                                 goto activate_locked;
471                         case SWAP_AGAIN:
472                                 goto keep_locked;
473                         case SWAP_SUCCESS:
474                                 ; /* try to free the page below */
475                         }
476                 }
477
478                 if (PageDirty(page)) {
479                         if (referenced)
480                                 goto keep_locked;
481                         if (!may_enter_fs)
482                                 goto keep_locked;
483                         if (!sc->may_writepage)
484                                 goto keep_locked;
485
486                         /* Page is dirty, try to write it out here */
487                         switch(pageout(page, mapping)) {
488                         case PAGE_KEEP:
489                                 goto keep_locked;
490                         case PAGE_ACTIVATE:
491                                 goto activate_locked;
492                         case PAGE_SUCCESS:
493                                 if (PageWriteback(page) || PageDirty(page))
494                                         goto keep;
495                                 /*
496                                  * A synchronous write - probably a ramdisk.  Go
497                                  * ahead and try to reclaim the page.
498                                  */
499                                 if (TestSetPageLocked(page))
500                                         goto keep;
501                                 if (PageDirty(page) || PageWriteback(page))
502                                         goto keep_locked;
503                                 mapping = page_mapping(page);
504                         case PAGE_CLEAN:
505                                 ; /* try to free the page below */
506                         }
507                 }
508
509                 /*
510                  * If the page has buffers, try to free the buffer mappings
511                  * associated with this page. If we succeed we try to free
512                  * the page as well.
513                  *
514                  * We do this even if the page is PageDirty().
515                  * try_to_release_page() does not perform I/O, but it is
516                  * possible for a page to have PageDirty set, but it is actually
517                  * clean (all its buffers are clean).  This happens if the
518                  * buffers were written out directly, with submit_bh(). ext3
519                  * will do this, as well as the blockdev mapping. 
520                  * try_to_release_page() will discover that cleanness and will
521                  * drop the buffers and mark the page clean - it can be freed.
522                  *
523                  * Rarely, pages can have buffers and no ->mapping.  These are
524                  * the pages which were not successfully invalidated in
525                  * truncate_complete_page().  We try to drop those buffers here
526                  * and if that worked, and the page is no longer mapped into
527                  * process address space (page_count == 1) it can be freed.
528                  * Otherwise, leave the page on the LRU so it is swappable.
529                  */
530                 if (PagePrivate(page)) {
531                         if (!try_to_release_page(page, sc->gfp_mask))
532                                 goto activate_locked;
533                         if (!mapping && page_count(page) == 1)
534                                 goto free_it;
535                 }
536
537                 if (!remove_mapping(mapping, page))
538                         goto keep_locked;
539
540 free_it:
541                 unlock_page(page);
542                 nr_reclaimed++;
543                 if (!pagevec_add(&freed_pvec, page))
544                         __pagevec_release_nonlru(&freed_pvec);
545                 continue;
546
547 activate_locked:
548                 SetPageActive(page);
549                 pgactivate++;
550 keep_locked:
551                 unlock_page(page);
552 keep:
553                 list_add(&page->lru, &ret_pages);
554                 BUG_ON(PageLRU(page));
555         }
556         list_splice(&ret_pages, page_list);
557         if (pagevec_count(&freed_pvec))
558                 __pagevec_release_nonlru(&freed_pvec);
559         mod_page_state(pgactivate, pgactivate);
560         return nr_reclaimed;
561 }
562
563 /*
564  * zone->lru_lock is heavily contended.  Some of the functions that
565  * shrink the lists perform better by taking out a batch of pages
566  * and working on them outside the LRU lock.
567  *
568  * For pagecache intensive workloads, this function is the hottest
569  * spot in the kernel (apart from copy_*_user functions).
570  *
571  * Appropriate locks must be held before calling this function.
572  *
573  * @nr_to_scan: The number of pages to look through on the list.
574  * @src:        The LRU list to pull pages off.
575  * @dst:        The temp list to put pages on to.
576  * @scanned:    The number of pages that were scanned.
577  *
578  * returns how many pages were moved onto *@dst.
579  */
580 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
581                 struct list_head *src, struct list_head *dst,
582                 unsigned long *scanned)
583 {
584         unsigned long nr_taken = 0;
585         struct page *page;
586         unsigned long scan;
587
588         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
589                 struct list_head *target;
590                 page = lru_to_page(src);
591                 prefetchw_prev_lru_page(page, src, flags);
592
593                 BUG_ON(!PageLRU(page));
594
595                 list_del(&page->lru);
596                 target = src;
597                 if (likely(get_page_unless_zero(page))) {
598                         /*
599                          * Be careful not to clear PageLRU until after we're
600                          * sure the page is not being freed elsewhere -- the
601                          * page release code relies on it.
602                          */
603                         ClearPageLRU(page);
604                         target = dst;
605                         nr_taken++;
606                 } /* else it is being freed elsewhere */
607
608                 list_add(&page->lru, target);
609         }
610
611         *scanned = scan;
612         return nr_taken;
613 }
614
615 /*
616  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
617  * of reclaimed pages
618  */
619 static unsigned long shrink_inactive_list(unsigned long max_scan,
620                                 struct zone *zone, struct scan_control *sc)
621 {
622         LIST_HEAD(page_list);
623         struct pagevec pvec;
624         unsigned long nr_scanned = 0;
625         unsigned long nr_reclaimed = 0;
626
627         pagevec_init(&pvec, 1);
628
629         lru_add_drain();
630         spin_lock_irq(&zone->lru_lock);
631         do {
632                 struct page *page;
633                 unsigned long nr_taken;
634                 unsigned long nr_scan;
635                 unsigned long nr_freed;
636
637                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
638                                              &zone->inactive_list,
639                                              &page_list, &nr_scan);
640                 zone->nr_inactive -= nr_taken;
641                 zone->pages_scanned += nr_scan;
642                 spin_unlock_irq(&zone->lru_lock);
643
644                 nr_scanned += nr_scan;
645                 nr_freed = shrink_page_list(&page_list, sc);
646                 nr_reclaimed += nr_freed;
647                 local_irq_disable();
648                 if (current_is_kswapd()) {
649                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
650                         __mod_page_state(kswapd_steal, nr_freed);
651                 } else
652                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
653                 __mod_page_state_zone(zone, pgsteal, nr_freed);
654
655                 if (nr_taken == 0)
656                         goto done;
657
658                 spin_lock(&zone->lru_lock);
659                 /*
660                  * Put back any unfreeable pages.
661                  */
662                 while (!list_empty(&page_list)) {
663                         page = lru_to_page(&page_list);
664                         BUG_ON(PageLRU(page));
665                         SetPageLRU(page);
666                         list_del(&page->lru);
667                         if (PageActive(page))
668                                 add_page_to_active_list(zone, page);
669                         else
670                                 add_page_to_inactive_list(zone, page);
671                         if (!pagevec_add(&pvec, page)) {
672                                 spin_unlock_irq(&zone->lru_lock);
673                                 __pagevec_release(&pvec);
674                                 spin_lock_irq(&zone->lru_lock);
675                         }
676                 }
677         } while (nr_scanned < max_scan);
678         spin_unlock(&zone->lru_lock);
679 done:
680         local_irq_enable();
681         pagevec_release(&pvec);
682         return nr_reclaimed;
683 }
684
685 /*
686  * This moves pages from the active list to the inactive list.
687  *
688  * We move them the other way if the page is referenced by one or more
689  * processes, from rmap.
690  *
691  * If the pages are mostly unmapped, the processing is fast and it is
692  * appropriate to hold zone->lru_lock across the whole operation.  But if
693  * the pages are mapped, the processing is slow (page_referenced()) so we
694  * should drop zone->lru_lock around each page.  It's impossible to balance
695  * this, so instead we remove the pages from the LRU while processing them.
696  * It is safe to rely on PG_active against the non-LRU pages in here because
697  * nobody will play with that bit on a non-LRU page.
698  *
699  * The downside is that we have to touch page->_count against each page.
700  * But we had to alter page->flags anyway.
701  */
702 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
703                                 struct scan_control *sc)
704 {
705         unsigned long pgmoved;
706         int pgdeactivate = 0;
707         unsigned long pgscanned;
708         LIST_HEAD(l_hold);      /* The pages which were snipped off */
709         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
710         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
711         struct page *page;
712         struct pagevec pvec;
713         int reclaim_mapped = 0;
714
715         if (sc->may_swap) {
716                 long mapped_ratio;
717                 long distress;
718                 long swap_tendency;
719
720                 /*
721                  * `distress' is a measure of how much trouble we're having
722                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
723                  */
724                 distress = 100 >> zone->prev_priority;
725
726                 /*
727                  * The point of this algorithm is to decide when to start
728                  * reclaiming mapped memory instead of just pagecache.  Work out
729                  * how much memory
730                  * is mapped.
731                  */
732                 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
733
734                 /*
735                  * Now decide how much we really want to unmap some pages.  The
736                  * mapped ratio is downgraded - just because there's a lot of
737                  * mapped memory doesn't necessarily mean that page reclaim
738                  * isn't succeeding.
739                  *
740                  * The distress ratio is important - we don't want to start
741                  * going oom.
742                  *
743                  * A 100% value of vm_swappiness overrides this algorithm
744                  * altogether.
745                  */
746                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
747
748                 /*
749                  * Now use this metric to decide whether to start moving mapped
750                  * memory onto the inactive list.
751                  */
752                 if (swap_tendency >= 100)
753                         reclaim_mapped = 1;
754         }
755
756         lru_add_drain();
757         spin_lock_irq(&zone->lru_lock);
758         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
759                                     &l_hold, &pgscanned);
760         zone->pages_scanned += pgscanned;
761         zone->nr_active -= pgmoved;
762         spin_unlock_irq(&zone->lru_lock);
763
764         while (!list_empty(&l_hold)) {
765                 cond_resched();
766                 page = lru_to_page(&l_hold);
767                 list_del(&page->lru);
768                 if (page_mapped(page)) {
769                         if (!reclaim_mapped ||
770                             (total_swap_pages == 0 && PageAnon(page)) ||
771                             page_referenced(page, 0)) {
772                                 list_add(&page->lru, &l_active);
773                                 continue;
774                         }
775                 }
776                 list_add(&page->lru, &l_inactive);
777         }
778
779         pagevec_init(&pvec, 1);
780         pgmoved = 0;
781         spin_lock_irq(&zone->lru_lock);
782         while (!list_empty(&l_inactive)) {
783                 page = lru_to_page(&l_inactive);
784                 prefetchw_prev_lru_page(page, &l_inactive, flags);
785                 BUG_ON(PageLRU(page));
786                 SetPageLRU(page);
787                 BUG_ON(!PageActive(page));
788                 ClearPageActive(page);
789
790                 list_move(&page->lru, &zone->inactive_list);
791                 pgmoved++;
792                 if (!pagevec_add(&pvec, page)) {
793                         zone->nr_inactive += pgmoved;
794                         spin_unlock_irq(&zone->lru_lock);
795                         pgdeactivate += pgmoved;
796                         pgmoved = 0;
797                         if (buffer_heads_over_limit)
798                                 pagevec_strip(&pvec);
799                         __pagevec_release(&pvec);
800                         spin_lock_irq(&zone->lru_lock);
801                 }
802         }
803         zone->nr_inactive += pgmoved;
804         pgdeactivate += pgmoved;
805         if (buffer_heads_over_limit) {
806                 spin_unlock_irq(&zone->lru_lock);
807                 pagevec_strip(&pvec);
808                 spin_lock_irq(&zone->lru_lock);
809         }
810
811         pgmoved = 0;
812         while (!list_empty(&l_active)) {
813                 page = lru_to_page(&l_active);
814                 prefetchw_prev_lru_page(page, &l_active, flags);
815                 BUG_ON(PageLRU(page));
816                 SetPageLRU(page);
817                 BUG_ON(!PageActive(page));
818                 list_move(&page->lru, &zone->active_list);
819                 pgmoved++;
820                 if (!pagevec_add(&pvec, page)) {
821                         zone->nr_active += pgmoved;
822                         pgmoved = 0;
823                         spin_unlock_irq(&zone->lru_lock);
824                         __pagevec_release(&pvec);
825                         spin_lock_irq(&zone->lru_lock);
826                 }
827         }
828         zone->nr_active += pgmoved;
829         spin_unlock(&zone->lru_lock);
830
831         __mod_page_state_zone(zone, pgrefill, pgscanned);
832         __mod_page_state(pgdeactivate, pgdeactivate);
833         local_irq_enable();
834
835         pagevec_release(&pvec);
836 }
837
838 /*
839  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
840  */
841 static unsigned long shrink_zone(int priority, struct zone *zone,
842                                 struct scan_control *sc)
843 {
844         unsigned long nr_active;
845         unsigned long nr_inactive;
846         unsigned long nr_to_scan;
847         unsigned long nr_reclaimed = 0;
848
849         atomic_inc(&zone->reclaim_in_progress);
850
851         /*
852          * Add one to `nr_to_scan' just to make sure that the kernel will
853          * slowly sift through the active list.
854          */
855         zone->nr_scan_active += (zone->nr_active >> priority) + 1;
856         nr_active = zone->nr_scan_active;
857         if (nr_active >= sc->swap_cluster_max)
858                 zone->nr_scan_active = 0;
859         else
860                 nr_active = 0;
861
862         zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1;
863         nr_inactive = zone->nr_scan_inactive;
864         if (nr_inactive >= sc->swap_cluster_max)
865                 zone->nr_scan_inactive = 0;
866         else
867                 nr_inactive = 0;
868
869         while (nr_active || nr_inactive) {
870                 if (nr_active) {
871                         nr_to_scan = min(nr_active,
872                                         (unsigned long)sc->swap_cluster_max);
873                         nr_active -= nr_to_scan;
874                         shrink_active_list(nr_to_scan, zone, sc);
875                 }
876
877                 if (nr_inactive) {
878                         nr_to_scan = min(nr_inactive,
879                                         (unsigned long)sc->swap_cluster_max);
880                         nr_inactive -= nr_to_scan;
881                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
882                                                                 sc);
883                 }
884         }
885
886         throttle_vm_writeout();
887
888         atomic_dec(&zone->reclaim_in_progress);
889         return nr_reclaimed;
890 }
891
892 /*
893  * This is the direct reclaim path, for page-allocating processes.  We only
894  * try to reclaim pages from zones which will satisfy the caller's allocation
895  * request.
896  *
897  * We reclaim from a zone even if that zone is over pages_high.  Because:
898  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
899  *    allocation or
900  * b) The zones may be over pages_high but they must go *over* pages_high to
901  *    satisfy the `incremental min' zone defense algorithm.
902  *
903  * Returns the number of reclaimed pages.
904  *
905  * If a zone is deemed to be full of pinned pages then just give it a light
906  * scan then give up on it.
907  */
908 static unsigned long shrink_zones(int priority, struct zone **zones,
909                                         struct scan_control *sc)
910 {
911         unsigned long nr_reclaimed = 0;
912         int i;
913
914         for (i = 0; zones[i] != NULL; i++) {
915                 struct zone *zone = zones[i];
916
917                 if (!populated_zone(zone))
918                         continue;
919
920                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
921                         continue;
922
923                 zone->temp_priority = priority;
924                 if (zone->prev_priority > priority)
925                         zone->prev_priority = priority;
926
927                 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
928                         continue;       /* Let kswapd poll it */
929
930                 nr_reclaimed += shrink_zone(priority, zone, sc);
931         }
932         return nr_reclaimed;
933 }
934  
935 /*
936  * This is the main entry point to direct page reclaim.
937  *
938  * If a full scan of the inactive list fails to free enough memory then we
939  * are "out of memory" and something needs to be killed.
940  *
941  * If the caller is !__GFP_FS then the probability of a failure is reasonably
942  * high - the zone may be full of dirty or under-writeback pages, which this
943  * caller can't do much about.  We kick pdflush and take explicit naps in the
944  * hope that some of these pages can be written.  But if the allocating task
945  * holds filesystem locks which prevent writeout this might not work, and the
946  * allocation attempt will fail.
947  */
948 unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
949 {
950         int priority;
951         int ret = 0;
952         unsigned long total_scanned = 0;
953         unsigned long nr_reclaimed = 0;
954         struct reclaim_state *reclaim_state = current->reclaim_state;
955         unsigned long lru_pages = 0;
956         int i;
957         struct scan_control sc = {
958                 .gfp_mask = gfp_mask,
959                 .may_writepage = !laptop_mode,
960                 .swap_cluster_max = SWAP_CLUSTER_MAX,
961                 .may_swap = 1,
962                 .swappiness = vm_swappiness,
963         };
964
965         inc_page_state(allocstall);
966
967         for (i = 0; zones[i] != NULL; i++) {
968                 struct zone *zone = zones[i];
969
970                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
971                         continue;
972
973                 zone->temp_priority = DEF_PRIORITY;
974                 lru_pages += zone->nr_active + zone->nr_inactive;
975         }
976
977         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
978                 sc.nr_mapped = read_page_state(nr_mapped);
979                 sc.nr_scanned = 0;
980                 if (!priority)
981                         disable_swap_token();
982                 nr_reclaimed += shrink_zones(priority, zones, &sc);
983                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
984                 if (reclaim_state) {
985                         nr_reclaimed += reclaim_state->reclaimed_slab;
986                         reclaim_state->reclaimed_slab = 0;
987                 }
988                 total_scanned += sc.nr_scanned;
989                 if (nr_reclaimed >= sc.swap_cluster_max) {
990                         ret = 1;
991                         goto out;
992                 }
993
994                 /*
995                  * Try to write back as many pages as we just scanned.  This
996                  * tends to cause slow streaming writers to write data to the
997                  * disk smoothly, at the dirtying rate, which is nice.   But
998                  * that's undesirable in laptop mode, where we *want* lumpy
999                  * writeout.  So in laptop mode, write out the whole world.
1000                  */
1001                 if (total_scanned > sc.swap_cluster_max +
1002                                         sc.swap_cluster_max / 2) {
1003                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1004                         sc.may_writepage = 1;
1005                 }
1006
1007                 /* Take a nap, wait for some writeback to complete */
1008                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1009                         blk_congestion_wait(WRITE, HZ/10);
1010         }
1011 out:
1012         for (i = 0; zones[i] != 0; i++) {
1013                 struct zone *zone = zones[i];
1014
1015                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1016                         continue;
1017
1018                 zone->prev_priority = zone->temp_priority;
1019         }
1020         return ret;
1021 }
1022
1023 /*
1024  * For kswapd, balance_pgdat() will work across all this node's zones until
1025  * they are all at pages_high.
1026  *
1027  * Returns the number of pages which were actually freed.
1028  *
1029  * There is special handling here for zones which are full of pinned pages.
1030  * This can happen if the pages are all mlocked, or if they are all used by
1031  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1032  * What we do is to detect the case where all pages in the zone have been
1033  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1034  * dead and from now on, only perform a short scan.  Basically we're polling
1035  * the zone for when the problem goes away.
1036  *
1037  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1038  * zones which have free_pages > pages_high, but once a zone is found to have
1039  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1040  * of the number of free pages in the lower zones.  This interoperates with
1041  * the page allocator fallback scheme to ensure that aging of pages is balanced
1042  * across the zones.
1043  */
1044 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1045 {
1046         int all_zones_ok;
1047         int priority;
1048         int i;
1049         unsigned long total_scanned;
1050         unsigned long nr_reclaimed;
1051         struct reclaim_state *reclaim_state = current->reclaim_state;
1052         struct scan_control sc = {
1053                 .gfp_mask = GFP_KERNEL,
1054                 .may_swap = 1,
1055                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1056                 .swappiness = vm_swappiness,
1057         };
1058
1059 loop_again:
1060         total_scanned = 0;
1061         nr_reclaimed = 0;
1062         sc.may_writepage = !laptop_mode;
1063         sc.nr_mapped = read_page_state(nr_mapped);
1064
1065         inc_page_state(pageoutrun);
1066
1067         for (i = 0; i < pgdat->nr_zones; i++) {
1068                 struct zone *zone = pgdat->node_zones + i;
1069
1070                 zone->temp_priority = DEF_PRIORITY;
1071         }
1072
1073         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1074                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1075                 unsigned long lru_pages = 0;
1076
1077                 /* The swap token gets in the way of swapout... */
1078                 if (!priority)
1079                         disable_swap_token();
1080
1081                 all_zones_ok = 1;
1082
1083                 /*
1084                  * Scan in the highmem->dma direction for the highest
1085                  * zone which needs scanning
1086                  */
1087                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1088                         struct zone *zone = pgdat->node_zones + i;
1089
1090                         if (!populated_zone(zone))
1091                                 continue;
1092
1093                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1094                                 continue;
1095
1096                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1097                                                0, 0)) {
1098                                 end_zone = i;
1099                                 goto scan;
1100                         }
1101                 }
1102                 goto out;
1103 scan:
1104                 for (i = 0; i <= end_zone; i++) {
1105                         struct zone *zone = pgdat->node_zones + i;
1106
1107                         lru_pages += zone->nr_active + zone->nr_inactive;
1108                 }
1109
1110                 /*
1111                  * Now scan the zone in the dma->highmem direction, stopping
1112                  * at the last zone which needs scanning.
1113                  *
1114                  * We do this because the page allocator works in the opposite
1115                  * direction.  This prevents the page allocator from allocating
1116                  * pages behind kswapd's direction of progress, which would
1117                  * cause too much scanning of the lower zones.
1118                  */
1119                 for (i = 0; i <= end_zone; i++) {
1120                         struct zone *zone = pgdat->node_zones + i;
1121                         int nr_slab;
1122
1123                         if (!populated_zone(zone))
1124                                 continue;
1125
1126                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1127                                 continue;
1128
1129                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1130                                                end_zone, 0))
1131                                 all_zones_ok = 0;
1132                         zone->temp_priority = priority;
1133                         if (zone->prev_priority > priority)
1134                                 zone->prev_priority = priority;
1135                         sc.nr_scanned = 0;
1136                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1137                         reclaim_state->reclaimed_slab = 0;
1138                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1139                                                 lru_pages);
1140                         nr_reclaimed += reclaim_state->reclaimed_slab;
1141                         total_scanned += sc.nr_scanned;
1142                         if (zone->all_unreclaimable)
1143                                 continue;
1144                         if (nr_slab == 0 && zone->pages_scanned >=
1145                                     (zone->nr_active + zone->nr_inactive) * 4)
1146                                 zone->all_unreclaimable = 1;
1147                         /*
1148                          * If we've done a decent amount of scanning and
1149                          * the reclaim ratio is low, start doing writepage
1150                          * even in laptop mode
1151                          */
1152                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1153                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1154                                 sc.may_writepage = 1;
1155                 }
1156                 if (all_zones_ok)
1157                         break;          /* kswapd: all done */
1158                 /*
1159                  * OK, kswapd is getting into trouble.  Take a nap, then take
1160                  * another pass across the zones.
1161                  */
1162                 if (total_scanned && priority < DEF_PRIORITY - 2)
1163                         blk_congestion_wait(WRITE, HZ/10);
1164
1165                 /*
1166                  * We do this so kswapd doesn't build up large priorities for
1167                  * example when it is freeing in parallel with allocators. It
1168                  * matches the direct reclaim path behaviour in terms of impact
1169                  * on zone->*_priority.
1170                  */
1171                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1172                         break;
1173         }
1174 out:
1175         for (i = 0; i < pgdat->nr_zones; i++) {
1176                 struct zone *zone = pgdat->node_zones + i;
1177
1178                 zone->prev_priority = zone->temp_priority;
1179         }
1180         if (!all_zones_ok) {
1181                 cond_resched();
1182                 goto loop_again;
1183         }
1184
1185         return nr_reclaimed;
1186 }
1187
1188 /*
1189  * The background pageout daemon, started as a kernel thread
1190  * from the init process. 
1191  *
1192  * This basically trickles out pages so that we have _some_
1193  * free memory available even if there is no other activity
1194  * that frees anything up. This is needed for things like routing
1195  * etc, where we otherwise might have all activity going on in
1196  * asynchronous contexts that cannot page things out.
1197  *
1198  * If there are applications that are active memory-allocators
1199  * (most normal use), this basically shouldn't matter.
1200  */
1201 static int kswapd(void *p)
1202 {
1203         unsigned long order;
1204         pg_data_t *pgdat = (pg_data_t*)p;
1205         struct task_struct *tsk = current;
1206         DEFINE_WAIT(wait);
1207         struct reclaim_state reclaim_state = {
1208                 .reclaimed_slab = 0,
1209         };
1210         cpumask_t cpumask;
1211
1212         daemonize("kswapd%d", pgdat->node_id);
1213         cpumask = node_to_cpumask(pgdat->node_id);
1214         if (!cpus_empty(cpumask))
1215                 set_cpus_allowed(tsk, cpumask);
1216         current->reclaim_state = &reclaim_state;
1217
1218         /*
1219          * Tell the memory management that we're a "memory allocator",
1220          * and that if we need more memory we should get access to it
1221          * regardless (see "__alloc_pages()"). "kswapd" should
1222          * never get caught in the normal page freeing logic.
1223          *
1224          * (Kswapd normally doesn't need memory anyway, but sometimes
1225          * you need a small amount of memory in order to be able to
1226          * page out something else, and this flag essentially protects
1227          * us from recursively trying to free more memory as we're
1228          * trying to free the first piece of memory in the first place).
1229          */
1230         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1231
1232         order = 0;
1233         for ( ; ; ) {
1234                 unsigned long new_order;
1235
1236                 try_to_freeze();
1237
1238                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1239                 new_order = pgdat->kswapd_max_order;
1240                 pgdat->kswapd_max_order = 0;
1241                 if (order < new_order) {
1242                         /*
1243                          * Don't sleep if someone wants a larger 'order'
1244                          * allocation
1245                          */
1246                         order = new_order;
1247                 } else {
1248                         schedule();
1249                         order = pgdat->kswapd_max_order;
1250                 }
1251                 finish_wait(&pgdat->kswapd_wait, &wait);
1252
1253                 balance_pgdat(pgdat, order);
1254         }
1255         return 0;
1256 }
1257
1258 /*
1259  * A zone is low on free memory, so wake its kswapd task to service it.
1260  */
1261 void wakeup_kswapd(struct zone *zone, int order)
1262 {
1263         pg_data_t *pgdat;
1264
1265         if (!populated_zone(zone))
1266                 return;
1267
1268         pgdat = zone->zone_pgdat;
1269         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1270                 return;
1271         if (pgdat->kswapd_max_order < order)
1272                 pgdat->kswapd_max_order = order;
1273         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1274                 return;
1275         if (!waitqueue_active(&pgdat->kswapd_wait))
1276                 return;
1277         wake_up_interruptible(&pgdat->kswapd_wait);
1278 }
1279
1280 #ifdef CONFIG_PM
1281 /*
1282  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1283  * from LRU lists system-wide, for given pass and priority, and returns the
1284  * number of reclaimed pages
1285  *
1286  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1287  */
1288 static unsigned long shrink_all_zones(unsigned long nr_pages, int pass,
1289                                       int prio, struct scan_control *sc)
1290 {
1291         struct zone *zone;
1292         unsigned long nr_to_scan, ret = 0;
1293
1294         for_each_zone(zone) {
1295
1296                 if (!populated_zone(zone))
1297                         continue;
1298
1299                 if (zone->all_unreclaimable && prio != DEF_PRIORITY)
1300                         continue;
1301
1302                 /* For pass = 0 we don't shrink the active list */
1303                 if (pass > 0) {
1304                         zone->nr_scan_active += (zone->nr_active >> prio) + 1;
1305                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1306                                 zone->nr_scan_active = 0;
1307                                 nr_to_scan = min(nr_pages, zone->nr_active);
1308                                 shrink_active_list(nr_to_scan, zone, sc);
1309                         }
1310                 }
1311
1312                 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1;
1313                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1314                         zone->nr_scan_inactive = 0;
1315                         nr_to_scan = min(nr_pages, zone->nr_inactive);
1316                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1317                         if (ret >= nr_pages)
1318                                 return ret;
1319                 }
1320         }
1321
1322         return ret;
1323 }
1324
1325 /*
1326  * Try to free `nr_pages' of memory, system-wide, and return the number of
1327  * freed pages.
1328  *
1329  * Rather than trying to age LRUs the aim is to preserve the overall
1330  * LRU order by reclaiming preferentially
1331  * inactive > active > active referenced > active mapped
1332  */
1333 unsigned long shrink_all_memory(unsigned long nr_pages)
1334 {
1335         unsigned long lru_pages, nr_slab;
1336         unsigned long ret = 0;
1337         int pass;
1338         struct reclaim_state reclaim_state;
1339         struct zone *zone;
1340         struct scan_control sc = {
1341                 .gfp_mask = GFP_KERNEL,
1342                 .may_swap = 0,
1343                 .swap_cluster_max = nr_pages,
1344                 .may_writepage = 1,
1345                 .swappiness = vm_swappiness,
1346         };
1347
1348         current->reclaim_state = &reclaim_state;
1349
1350         lru_pages = 0;
1351         for_each_zone(zone)
1352                 lru_pages += zone->nr_active + zone->nr_inactive;
1353
1354         nr_slab = read_page_state(nr_slab);
1355         /* If slab caches are huge, it's better to hit them first */
1356         while (nr_slab >= lru_pages) {
1357                 reclaim_state.reclaimed_slab = 0;
1358                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1359                 if (!reclaim_state.reclaimed_slab)
1360                         break;
1361
1362                 ret += reclaim_state.reclaimed_slab;
1363                 if (ret >= nr_pages)
1364                         goto out;
1365
1366                 nr_slab -= reclaim_state.reclaimed_slab;
1367         }
1368
1369         /*
1370          * We try to shrink LRUs in 5 passes:
1371          * 0 = Reclaim from inactive_list only
1372          * 1 = Reclaim from active list but don't reclaim mapped
1373          * 2 = 2nd pass of type 1
1374          * 3 = Reclaim mapped (normal reclaim)
1375          * 4 = 2nd pass of type 3
1376          */
1377         for (pass = 0; pass < 5; pass++) {
1378                 int prio;
1379
1380                 /* Needed for shrinking slab caches later on */
1381                 if (!lru_pages)
1382                         for_each_zone(zone) {
1383                                 lru_pages += zone->nr_active;
1384                                 lru_pages += zone->nr_inactive;
1385                         }
1386
1387                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1388                 if (pass > 2) {
1389                         sc.may_swap = 1;
1390                         sc.swappiness = 100;
1391                 }
1392
1393                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1394                         unsigned long nr_to_scan = nr_pages - ret;
1395
1396                         sc.nr_mapped = read_page_state(nr_mapped);
1397                         sc.nr_scanned = 0;
1398
1399                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1400                         if (ret >= nr_pages)
1401                                 goto out;
1402
1403                         reclaim_state.reclaimed_slab = 0;
1404                         shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages);
1405                         ret += reclaim_state.reclaimed_slab;
1406                         if (ret >= nr_pages)
1407                                 goto out;
1408
1409                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1410                                 blk_congestion_wait(WRITE, HZ / 10);
1411                 }
1412
1413                 lru_pages = 0;
1414         }
1415
1416         /*
1417          * If ret = 0, we could not shrink LRUs, but there may be something
1418          * in slab caches
1419          */
1420         if (!ret)
1421                 do {
1422                         reclaim_state.reclaimed_slab = 0;
1423                         shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1424                         ret += reclaim_state.reclaimed_slab;
1425                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1426
1427 out:
1428         current->reclaim_state = NULL;
1429
1430         return ret;
1431 }
1432 #endif
1433
1434 #ifdef CONFIG_HOTPLUG_CPU
1435 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1436    not required for correctness.  So if the last cpu in a node goes
1437    away, we get changed to run anywhere: as the first one comes back,
1438    restore their cpu bindings. */
1439 static int cpu_callback(struct notifier_block *nfb,
1440                                   unsigned long action, void *hcpu)
1441 {
1442         pg_data_t *pgdat;
1443         cpumask_t mask;
1444
1445         if (action == CPU_ONLINE) {
1446                 for_each_online_pgdat(pgdat) {
1447                         mask = node_to_cpumask(pgdat->node_id);
1448                         if (any_online_cpu(mask) != NR_CPUS)
1449                                 /* One of our CPUs online: restore mask */
1450                                 set_cpus_allowed(pgdat->kswapd, mask);
1451                 }
1452         }
1453         return NOTIFY_OK;
1454 }
1455 #endif /* CONFIG_HOTPLUG_CPU */
1456
1457 static int __init kswapd_init(void)
1458 {
1459         pg_data_t *pgdat;
1460
1461         swap_setup();
1462         for_each_online_pgdat(pgdat) {
1463                 pid_t pid;
1464
1465                 pid = kernel_thread(kswapd, pgdat, CLONE_KERNEL);
1466                 BUG_ON(pid < 0);
1467                 read_lock(&tasklist_lock);
1468                 pgdat->kswapd = find_task_by_pid(pid);
1469                 read_unlock(&tasklist_lock);
1470         }
1471         total_memory = nr_free_pagecache_pages();
1472         hotcpu_notifier(cpu_callback, 0);
1473         return 0;
1474 }
1475
1476 module_init(kswapd_init)
1477
1478 #ifdef CONFIG_NUMA
1479 /*
1480  * Zone reclaim mode
1481  *
1482  * If non-zero call zone_reclaim when the number of free pages falls below
1483  * the watermarks.
1484  *
1485  * In the future we may add flags to the mode. However, the page allocator
1486  * should only have to check that zone_reclaim_mode != 0 before calling
1487  * zone_reclaim().
1488  */
1489 int zone_reclaim_mode __read_mostly;
1490
1491 #define RECLAIM_OFF 0
1492 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1493 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1494 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1495 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1496
1497 /*
1498  * Mininum time between zone reclaim scans
1499  */
1500 int zone_reclaim_interval __read_mostly = 30*HZ;
1501
1502 /*
1503  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1504  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1505  * a zone.
1506  */
1507 #define ZONE_RECLAIM_PRIORITY 4
1508
1509 /*
1510  * Try to free up some pages from this zone through reclaim.
1511  */
1512 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1513 {
1514         /* Minimum pages needed in order to stay on node */
1515         const unsigned long nr_pages = 1 << order;
1516         struct task_struct *p = current;
1517         struct reclaim_state reclaim_state;
1518         int priority;
1519         unsigned long nr_reclaimed = 0;
1520         struct scan_control sc = {
1521                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1522                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1523                 .nr_mapped = read_page_state(nr_mapped),
1524                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1525                                         SWAP_CLUSTER_MAX),
1526                 .gfp_mask = gfp_mask,
1527                 .swappiness = vm_swappiness,
1528         };
1529
1530         disable_swap_token();
1531         cond_resched();
1532         /*
1533          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1534          * and we also need to be able to write out pages for RECLAIM_WRITE
1535          * and RECLAIM_SWAP.
1536          */
1537         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1538         reclaim_state.reclaimed_slab = 0;
1539         p->reclaim_state = &reclaim_state;
1540
1541         /*
1542          * Free memory by calling shrink zone with increasing priorities
1543          * until we have enough memory freed.
1544          */
1545         priority = ZONE_RECLAIM_PRIORITY;
1546         do {
1547                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1548                 priority--;
1549         } while (priority >= 0 && nr_reclaimed < nr_pages);
1550
1551         if (nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1552                 /*
1553                  * shrink_slab() does not currently allow us to determine how
1554                  * many pages were freed in this zone. So we just shake the slab
1555                  * a bit and then go off node for this particular allocation
1556                  * despite possibly having freed enough memory to allocate in
1557                  * this zone.  If we freed local memory then the next
1558                  * allocations will be local again.
1559                  *
1560                  * shrink_slab will free memory on all zones and may take
1561                  * a long time.
1562                  */
1563                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1564         }
1565
1566         p->reclaim_state = NULL;
1567         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1568
1569         if (nr_reclaimed == 0) {
1570                 /*
1571                  * We were unable to reclaim enough pages to stay on node.  We
1572                  * now allow off node accesses for a certain time period before
1573                  * trying again to reclaim pages from the local zone.
1574                  */
1575                 zone->last_unsuccessful_zone_reclaim = jiffies;
1576         }
1577
1578         return nr_reclaimed >= nr_pages;
1579 }
1580
1581 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1582 {
1583         cpumask_t mask;
1584         int node_id;
1585
1586         /*
1587          * Do not reclaim if there was a recent unsuccessful attempt at zone
1588          * reclaim.  In that case we let allocations go off node for the
1589          * zone_reclaim_interval.  Otherwise we would scan for each off-node
1590          * page allocation.
1591          */
1592         if (time_before(jiffies,
1593                 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1594                         return 0;
1595
1596         /*
1597          * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1598          * not have reclaimable pages and if we should not delay the allocation
1599          * then do not scan.
1600          */
1601         if (!(gfp_mask & __GFP_WAIT) ||
1602                 zone->all_unreclaimable ||
1603                 atomic_read(&zone->reclaim_in_progress) > 0 ||
1604                 (current->flags & PF_MEMALLOC))
1605                         return 0;
1606
1607         /*
1608          * Only run zone reclaim on the local zone or on zones that do not
1609          * have associated processors. This will favor the local processor
1610          * over remote processors and spread off node memory allocations
1611          * as wide as possible.
1612          */
1613         node_id = zone->zone_pgdat->node_id;
1614         mask = node_to_cpumask(node_id);
1615         if (!cpus_empty(mask) && node_id != numa_node_id())
1616                 return 0;
1617         return __zone_reclaim(zone, gfp_mask, order);
1618 }
1619 #endif