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