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