9babfbc1ddc88ac99c21d91e9afc3d1194893681
[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 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42
43 #include <asm/tlbflush.h>
44 #include <asm/div64.h>
45
46 #include <linux/swapops.h>
47
48 #include "internal.h"
49
50 struct scan_control {
51         /* Incremented by the number of inactive pages that were scanned */
52         unsigned long nr_scanned;
53
54         /* This context's GFP mask */
55         gfp_t gfp_mask;
56
57         int may_writepage;
58
59         /* Can pages be swapped as part of reclaim? */
60         int may_swap;
61
62         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
63          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
64          * In this context, it doesn't matter that we scan the
65          * whole list at once. */
66         int swap_cluster_max;
67
68         int swappiness;
69
70         int all_unreclaimable;
71
72         int order;
73
74         /* Which cgroup do we reclaim from */
75         struct mem_cgroup *mem_cgroup;
76
77         /* Pluggable isolate pages callback */
78         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
79                         unsigned long *scanned, int order, int mode,
80                         struct zone *z, struct mem_cgroup *mem_cont,
81                         int active, int file);
82 };
83
84 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
85
86 #ifdef ARCH_HAS_PREFETCH
87 #define prefetch_prev_lru_page(_page, _base, _field)                    \
88         do {                                                            \
89                 if ((_page)->lru.prev != _base) {                       \
90                         struct page *prev;                              \
91                                                                         \
92                         prev = lru_to_page(&(_page->lru));              \
93                         prefetch(&prev->_field);                        \
94                 }                                                       \
95         } while (0)
96 #else
97 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
98 #endif
99
100 #ifdef ARCH_HAS_PREFETCHW
101 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
102         do {                                                            \
103                 if ((_page)->lru.prev != _base) {                       \
104                         struct page *prev;                              \
105                                                                         \
106                         prev = lru_to_page(&(_page->lru));              \
107                         prefetchw(&prev->_field);                       \
108                 }                                                       \
109         } while (0)
110 #else
111 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
112 #endif
113
114 /*
115  * From 0 .. 100.  Higher means more swappy.
116  */
117 int vm_swappiness = 60;
118 long vm_total_pages;    /* The total number of pages which the VM controls */
119
120 static LIST_HEAD(shrinker_list);
121 static DECLARE_RWSEM(shrinker_rwsem);
122
123 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
124 #define scan_global_lru(sc)     (!(sc)->mem_cgroup)
125 #else
126 #define scan_global_lru(sc)     (1)
127 #endif
128
129 /*
130  * Add a shrinker callback to be called from the vm
131  */
132 void register_shrinker(struct shrinker *shrinker)
133 {
134         shrinker->nr = 0;
135         down_write(&shrinker_rwsem);
136         list_add_tail(&shrinker->list, &shrinker_list);
137         up_write(&shrinker_rwsem);
138 }
139 EXPORT_SYMBOL(register_shrinker);
140
141 /*
142  * Remove one
143  */
144 void unregister_shrinker(struct shrinker *shrinker)
145 {
146         down_write(&shrinker_rwsem);
147         list_del(&shrinker->list);
148         up_write(&shrinker_rwsem);
149 }
150 EXPORT_SYMBOL(unregister_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 encountered 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->shrink)(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                                         __func__, 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->shrink)(0, gfp_mask);
216                         shrink_ret = (*shrinker->shrink)(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                 mapping_set_error(mapping, error);
288         unlock_page(page);
289 }
290
291 /* Request for sync pageout. */
292 enum pageout_io {
293         PAGEOUT_IO_ASYNC,
294         PAGEOUT_IO_SYNC,
295 };
296
297 /* possible outcome of pageout() */
298 typedef enum {
299         /* failed to write page out, page is locked */
300         PAGE_KEEP,
301         /* move page to the active list, page is locked */
302         PAGE_ACTIVATE,
303         /* page has been sent to the disk successfully, page is unlocked */
304         PAGE_SUCCESS,
305         /* page is clean and locked */
306         PAGE_CLEAN,
307 } pageout_t;
308
309 /*
310  * pageout is called by shrink_page_list() for each dirty page.
311  * Calls ->writepage().
312  */
313 static pageout_t pageout(struct page *page, struct address_space *mapping,
314                                                 enum pageout_io sync_writeback)
315 {
316         /*
317          * If the page is dirty, only perform writeback if that write
318          * will be non-blocking.  To prevent this allocation from being
319          * stalled by pagecache activity.  But note that there may be
320          * stalls if we need to run get_block().  We could test
321          * PagePrivate for that.
322          *
323          * If this process is currently in generic_file_write() against
324          * this page's queue, we can perform writeback even if that
325          * will block.
326          *
327          * If the page is swapcache, write it back even if that would
328          * block, for some throttling. This happens by accident, because
329          * swap_backing_dev_info is bust: it doesn't reflect the
330          * congestion state of the swapdevs.  Easy to fix, if needed.
331          * See swapfile.c:page_queue_congested().
332          */
333         if (!is_page_cache_freeable(page))
334                 return PAGE_KEEP;
335         if (!mapping) {
336                 /*
337                  * Some data journaling orphaned pages can have
338                  * page->mapping == NULL while being dirty with clean buffers.
339                  */
340                 if (PagePrivate(page)) {
341                         if (try_to_free_buffers(page)) {
342                                 ClearPageDirty(page);
343                                 printk("%s: orphaned page\n", __func__);
344                                 return PAGE_CLEAN;
345                         }
346                 }
347                 return PAGE_KEEP;
348         }
349         if (mapping->a_ops->writepage == NULL)
350                 return PAGE_ACTIVATE;
351         if (!may_write_to_queue(mapping->backing_dev_info))
352                 return PAGE_KEEP;
353
354         if (clear_page_dirty_for_io(page)) {
355                 int res;
356                 struct writeback_control wbc = {
357                         .sync_mode = WB_SYNC_NONE,
358                         .nr_to_write = SWAP_CLUSTER_MAX,
359                         .range_start = 0,
360                         .range_end = LLONG_MAX,
361                         .nonblocking = 1,
362                         .for_reclaim = 1,
363                 };
364
365                 SetPageReclaim(page);
366                 res = mapping->a_ops->writepage(page, &wbc);
367                 if (res < 0)
368                         handle_write_error(mapping, page, res);
369                 if (res == AOP_WRITEPAGE_ACTIVATE) {
370                         ClearPageReclaim(page);
371                         return PAGE_ACTIVATE;
372                 }
373
374                 /*
375                  * Wait on writeback if requested to. This happens when
376                  * direct reclaiming a large contiguous area and the
377                  * first attempt to free a range of pages fails.
378                  */
379                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
380                         wait_on_page_writeback(page);
381
382                 if (!PageWriteback(page)) {
383                         /* synchronous write or broken a_ops? */
384                         ClearPageReclaim(page);
385                 }
386                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
387                 return PAGE_SUCCESS;
388         }
389
390         return PAGE_CLEAN;
391 }
392
393 /*
394  * Same as remove_mapping, but if the page is removed from the mapping, it
395  * gets returned with a refcount of 0.
396  */
397 static int __remove_mapping(struct address_space *mapping, struct page *page)
398 {
399         BUG_ON(!PageLocked(page));
400         BUG_ON(mapping != page_mapping(page));
401
402         spin_lock_irq(&mapping->tree_lock);
403         /*
404          * The non racy check for a busy page.
405          *
406          * Must be careful with the order of the tests. When someone has
407          * a ref to the page, it may be possible that they dirty it then
408          * drop the reference. So if PageDirty is tested before page_count
409          * here, then the following race may occur:
410          *
411          * get_user_pages(&page);
412          * [user mapping goes away]
413          * write_to(page);
414          *                              !PageDirty(page)    [good]
415          * SetPageDirty(page);
416          * put_page(page);
417          *                              !page_count(page)   [good, discard it]
418          *
419          * [oops, our write_to data is lost]
420          *
421          * Reversing the order of the tests ensures such a situation cannot
422          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
423          * load is not satisfied before that of page->_count.
424          *
425          * Note that if SetPageDirty is always performed via set_page_dirty,
426          * and thus under tree_lock, then this ordering is not required.
427          */
428         if (!page_freeze_refs(page, 2))
429                 goto cannot_free;
430         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
431         if (unlikely(PageDirty(page))) {
432                 page_unfreeze_refs(page, 2);
433                 goto cannot_free;
434         }
435
436         if (PageSwapCache(page)) {
437                 swp_entry_t swap = { .val = page_private(page) };
438                 __delete_from_swap_cache(page);
439                 spin_unlock_irq(&mapping->tree_lock);
440                 swap_free(swap);
441         } else {
442                 __remove_from_page_cache(page);
443                 spin_unlock_irq(&mapping->tree_lock);
444         }
445
446         return 1;
447
448 cannot_free:
449         spin_unlock_irq(&mapping->tree_lock);
450         return 0;
451 }
452
453 /*
454  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
455  * someone else has a ref on the page, abort and return 0.  If it was
456  * successfully detached, return 1.  Assumes the caller has a single ref on
457  * this page.
458  */
459 int remove_mapping(struct address_space *mapping, struct page *page)
460 {
461         if (__remove_mapping(mapping, page)) {
462                 /*
463                  * Unfreezing the refcount with 1 rather than 2 effectively
464                  * drops the pagecache ref for us without requiring another
465                  * atomic operation.
466                  */
467                 page_unfreeze_refs(page, 1);
468                 return 1;
469         }
470         return 0;
471 }
472
473 /**
474  * putback_lru_page - put previously isolated page onto appropriate LRU list
475  * @page: page to be put back to appropriate lru list
476  *
477  * Add previously isolated @page to appropriate LRU list.
478  * Page may still be unevictable for other reasons.
479  *
480  * lru_lock must not be held, interrupts must be enabled.
481  */
482 #ifdef CONFIG_UNEVICTABLE_LRU
483 void putback_lru_page(struct page *page)
484 {
485         int lru;
486         int active = !!TestClearPageActive(page);
487         int was_unevictable = PageUnevictable(page);
488
489         VM_BUG_ON(PageLRU(page));
490
491 redo:
492         ClearPageUnevictable(page);
493
494         if (page_evictable(page, NULL)) {
495                 /*
496                  * For evictable pages, we can use the cache.
497                  * In event of a race, worst case is we end up with an
498                  * unevictable page on [in]active list.
499                  * We know how to handle that.
500                  */
501                 lru = active + page_is_file_cache(page);
502                 lru_cache_add_lru(page, lru);
503         } else {
504                 /*
505                  * Put unevictable pages directly on zone's unevictable
506                  * list.
507                  */
508                 lru = LRU_UNEVICTABLE;
509                 add_page_to_unevictable_list(page);
510         }
511         mem_cgroup_move_lists(page, lru);
512
513         /*
514          * page's status can change while we move it among lru. If an evictable
515          * page is on unevictable list, it never be freed. To avoid that,
516          * check after we added it to the list, again.
517          */
518         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
519                 if (!isolate_lru_page(page)) {
520                         put_page(page);
521                         goto redo;
522                 }
523                 /* This means someone else dropped this page from LRU
524                  * So, it will be freed or putback to LRU again. There is
525                  * nothing to do here.
526                  */
527         }
528
529         if (was_unevictable && lru != LRU_UNEVICTABLE)
530                 count_vm_event(UNEVICTABLE_PGRESCUED);
531         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
532                 count_vm_event(UNEVICTABLE_PGCULLED);
533
534         put_page(page);         /* drop ref from isolate */
535 }
536
537 #else /* CONFIG_UNEVICTABLE_LRU */
538
539 void putback_lru_page(struct page *page)
540 {
541         int lru;
542         VM_BUG_ON(PageLRU(page));
543
544         lru = !!TestClearPageActive(page) + page_is_file_cache(page);
545         lru_cache_add_lru(page, lru);
546         mem_cgroup_move_lists(page, lru);
547         put_page(page);
548 }
549 #endif /* CONFIG_UNEVICTABLE_LRU */
550
551
552 /*
553  * shrink_page_list() returns the number of reclaimed pages
554  */
555 static unsigned long shrink_page_list(struct list_head *page_list,
556                                         struct scan_control *sc,
557                                         enum pageout_io sync_writeback)
558 {
559         LIST_HEAD(ret_pages);
560         struct pagevec freed_pvec;
561         int pgactivate = 0;
562         unsigned long nr_reclaimed = 0;
563
564         cond_resched();
565
566         pagevec_init(&freed_pvec, 1);
567         while (!list_empty(page_list)) {
568                 struct address_space *mapping;
569                 struct page *page;
570                 int may_enter_fs;
571                 int referenced;
572
573                 cond_resched();
574
575                 page = lru_to_page(page_list);
576                 list_del(&page->lru);
577
578                 if (!trylock_page(page))
579                         goto keep;
580
581                 VM_BUG_ON(PageActive(page));
582
583                 sc->nr_scanned++;
584
585                 if (unlikely(!page_evictable(page, NULL))) {
586                         unlock_page(page);
587                         putback_lru_page(page);
588                         continue;
589                 }
590
591                 if (!sc->may_swap && page_mapped(page))
592                         goto keep_locked;
593
594                 /* Double the slab pressure for mapped and swapcache pages */
595                 if (page_mapped(page) || PageSwapCache(page))
596                         sc->nr_scanned++;
597
598                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
599                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
600
601                 if (PageWriteback(page)) {
602                         /*
603                          * Synchronous reclaim is performed in two passes,
604                          * first an asynchronous pass over the list to
605                          * start parallel writeback, and a second synchronous
606                          * pass to wait for the IO to complete.  Wait here
607                          * for any page for which writeback has already
608                          * started.
609                          */
610                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
611                                 wait_on_page_writeback(page);
612                         else
613                                 goto keep_locked;
614                 }
615
616                 referenced = page_referenced(page, 1, sc->mem_cgroup);
617                 /* In active use or really unfreeable?  Activate it. */
618                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
619                                         referenced && page_mapping_inuse(page))
620                         goto activate_locked;
621
622 #ifdef CONFIG_SWAP
623                 /*
624                  * Anonymous process memory has backing store?
625                  * Try to allocate it some swap space here.
626                  */
627                 if (PageAnon(page) && !PageSwapCache(page))
628                         if (!add_to_swap(page, GFP_ATOMIC))
629                                 goto activate_locked;
630 #endif /* CONFIG_SWAP */
631
632                 mapping = page_mapping(page);
633
634                 /*
635                  * The page is mapped into the page tables of one or more
636                  * processes. Try to unmap it here.
637                  */
638                 if (page_mapped(page) && mapping) {
639                         switch (try_to_unmap(page, 0)) {
640                         case SWAP_FAIL:
641                                 goto activate_locked;
642                         case SWAP_AGAIN:
643                                 goto keep_locked;
644                         case SWAP_SUCCESS:
645                                 ; /* try to free the page below */
646                         }
647                 }
648
649                 if (PageDirty(page)) {
650                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
651                                 goto keep_locked;
652                         if (!may_enter_fs)
653                                 goto keep_locked;
654                         if (!sc->may_writepage)
655                                 goto keep_locked;
656
657                         /* Page is dirty, try to write it out here */
658                         switch (pageout(page, mapping, sync_writeback)) {
659                         case PAGE_KEEP:
660                                 goto keep_locked;
661                         case PAGE_ACTIVATE:
662                                 goto activate_locked;
663                         case PAGE_SUCCESS:
664                                 if (PageWriteback(page) || PageDirty(page))
665                                         goto keep;
666                                 /*
667                                  * A synchronous write - probably a ramdisk.  Go
668                                  * ahead and try to reclaim the page.
669                                  */
670                                 if (!trylock_page(page))
671                                         goto keep;
672                                 if (PageDirty(page) || PageWriteback(page))
673                                         goto keep_locked;
674                                 mapping = page_mapping(page);
675                         case PAGE_CLEAN:
676                                 ; /* try to free the page below */
677                         }
678                 }
679
680                 /*
681                  * If the page has buffers, try to free the buffer mappings
682                  * associated with this page. If we succeed we try to free
683                  * the page as well.
684                  *
685                  * We do this even if the page is PageDirty().
686                  * try_to_release_page() does not perform I/O, but it is
687                  * possible for a page to have PageDirty set, but it is actually
688                  * clean (all its buffers are clean).  This happens if the
689                  * buffers were written out directly, with submit_bh(). ext3
690                  * will do this, as well as the blockdev mapping.
691                  * try_to_release_page() will discover that cleanness and will
692                  * drop the buffers and mark the page clean - it can be freed.
693                  *
694                  * Rarely, pages can have buffers and no ->mapping.  These are
695                  * the pages which were not successfully invalidated in
696                  * truncate_complete_page().  We try to drop those buffers here
697                  * and if that worked, and the page is no longer mapped into
698                  * process address space (page_count == 1) it can be freed.
699                  * Otherwise, leave the page on the LRU so it is swappable.
700                  */
701                 if (PagePrivate(page)) {
702                         if (!try_to_release_page(page, sc->gfp_mask))
703                                 goto activate_locked;
704                         if (!mapping && page_count(page) == 1) {
705                                 unlock_page(page);
706                                 if (put_page_testzero(page))
707                                         goto free_it;
708                                 else {
709                                         /*
710                                          * rare race with speculative reference.
711                                          * the speculative reference will free
712                                          * this page shortly, so we may
713                                          * increment nr_reclaimed here (and
714                                          * leave it off the LRU).
715                                          */
716                                         nr_reclaimed++;
717                                         continue;
718                                 }
719                         }
720                 }
721
722                 if (!mapping || !__remove_mapping(mapping, page))
723                         goto keep_locked;
724
725                 unlock_page(page);
726 free_it:
727                 nr_reclaimed++;
728                 if (!pagevec_add(&freed_pvec, page)) {
729                         __pagevec_free(&freed_pvec);
730                         pagevec_reinit(&freed_pvec);
731                 }
732                 continue;
733
734 activate_locked:
735                 /* Not a candidate for swapping, so reclaim swap space. */
736                 if (PageSwapCache(page) && vm_swap_full())
737                         remove_exclusive_swap_page_ref(page);
738                 VM_BUG_ON(PageActive(page));
739                 SetPageActive(page);
740                 pgactivate++;
741 keep_locked:
742                 unlock_page(page);
743 keep:
744                 list_add(&page->lru, &ret_pages);
745                 VM_BUG_ON(PageLRU(page));
746         }
747         list_splice(&ret_pages, page_list);
748         if (pagevec_count(&freed_pvec))
749                 __pagevec_free(&freed_pvec);
750         count_vm_events(PGACTIVATE, pgactivate);
751         return nr_reclaimed;
752 }
753
754 /* LRU Isolation modes. */
755 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
756 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
757 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
758
759 /*
760  * Attempt to remove the specified page from its LRU.  Only take this page
761  * if it is of the appropriate PageActive status.  Pages which are being
762  * freed elsewhere are also ignored.
763  *
764  * page:        page to consider
765  * mode:        one of the LRU isolation modes defined above
766  *
767  * returns 0 on success, -ve errno on failure.
768  */
769 int __isolate_lru_page(struct page *page, int mode, int file)
770 {
771         int ret = -EINVAL;
772
773         /* Only take pages on the LRU. */
774         if (!PageLRU(page))
775                 return ret;
776
777         /*
778          * When checking the active state, we need to be sure we are
779          * dealing with comparible boolean values.  Take the logical not
780          * of each.
781          */
782         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
783                 return ret;
784
785         if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
786                 return ret;
787
788         /*
789          * When this function is being called for lumpy reclaim, we
790          * initially look into all LRU pages, active, inactive and
791          * unevictable; only give shrink_page_list evictable pages.
792          */
793         if (PageUnevictable(page))
794                 return ret;
795
796         ret = -EBUSY;
797         if (likely(get_page_unless_zero(page))) {
798                 /*
799                  * Be careful not to clear PageLRU until after we're
800                  * sure the page is not being freed elsewhere -- the
801                  * page release code relies on it.
802                  */
803                 ClearPageLRU(page);
804                 ret = 0;
805         }
806
807         return ret;
808 }
809
810 /*
811  * zone->lru_lock is heavily contended.  Some of the functions that
812  * shrink the lists perform better by taking out a batch of pages
813  * and working on them outside the LRU lock.
814  *
815  * For pagecache intensive workloads, this function is the hottest
816  * spot in the kernel (apart from copy_*_user functions).
817  *
818  * Appropriate locks must be held before calling this function.
819  *
820  * @nr_to_scan: The number of pages to look through on the list.
821  * @src:        The LRU list to pull pages off.
822  * @dst:        The temp list to put pages on to.
823  * @scanned:    The number of pages that were scanned.
824  * @order:      The caller's attempted allocation order
825  * @mode:       One of the LRU isolation modes
826  * @file:       True [1] if isolating file [!anon] pages
827  *
828  * returns how many pages were moved onto *@dst.
829  */
830 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
831                 struct list_head *src, struct list_head *dst,
832                 unsigned long *scanned, int order, int mode, int file)
833 {
834         unsigned long nr_taken = 0;
835         unsigned long scan;
836
837         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
838                 struct page *page;
839                 unsigned long pfn;
840                 unsigned long end_pfn;
841                 unsigned long page_pfn;
842                 int zone_id;
843
844                 page = lru_to_page(src);
845                 prefetchw_prev_lru_page(page, src, flags);
846
847                 VM_BUG_ON(!PageLRU(page));
848
849                 switch (__isolate_lru_page(page, mode, file)) {
850                 case 0:
851                         list_move(&page->lru, dst);
852                         nr_taken++;
853                         break;
854
855                 case -EBUSY:
856                         /* else it is being freed elsewhere */
857                         list_move(&page->lru, src);
858                         continue;
859
860                 default:
861                         BUG();
862                 }
863
864                 if (!order)
865                         continue;
866
867                 /*
868                  * Attempt to take all pages in the order aligned region
869                  * surrounding the tag page.  Only take those pages of
870                  * the same active state as that tag page.  We may safely
871                  * round the target page pfn down to the requested order
872                  * as the mem_map is guarenteed valid out to MAX_ORDER,
873                  * where that page is in a different zone we will detect
874                  * it from its zone id and abort this block scan.
875                  */
876                 zone_id = page_zone_id(page);
877                 page_pfn = page_to_pfn(page);
878                 pfn = page_pfn & ~((1 << order) - 1);
879                 end_pfn = pfn + (1 << order);
880                 for (; pfn < end_pfn; pfn++) {
881                         struct page *cursor_page;
882
883                         /* The target page is in the block, ignore it. */
884                         if (unlikely(pfn == page_pfn))
885                                 continue;
886
887                         /* Avoid holes within the zone. */
888                         if (unlikely(!pfn_valid_within(pfn)))
889                                 break;
890
891                         cursor_page = pfn_to_page(pfn);
892
893                         /* Check that we have not crossed a zone boundary. */
894                         if (unlikely(page_zone_id(cursor_page) != zone_id))
895                                 continue;
896                         switch (__isolate_lru_page(cursor_page, mode, file)) {
897                         case 0:
898                                 list_move(&cursor_page->lru, dst);
899                                 nr_taken++;
900                                 scan++;
901                                 break;
902
903                         case -EBUSY:
904                                 /* else it is being freed elsewhere */
905                                 list_move(&cursor_page->lru, src);
906                         default:
907                                 break;  /* ! on LRU or wrong list */
908                         }
909                 }
910         }
911
912         *scanned = scan;
913         return nr_taken;
914 }
915
916 static unsigned long isolate_pages_global(unsigned long nr,
917                                         struct list_head *dst,
918                                         unsigned long *scanned, int order,
919                                         int mode, struct zone *z,
920                                         struct mem_cgroup *mem_cont,
921                                         int active, int file)
922 {
923         int lru = LRU_BASE;
924         if (active)
925                 lru += LRU_ACTIVE;
926         if (file)
927                 lru += LRU_FILE;
928         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
929                                                                 mode, !!file);
930 }
931
932 /*
933  * clear_active_flags() is a helper for shrink_active_list(), clearing
934  * any active bits from the pages in the list.
935  */
936 static unsigned long clear_active_flags(struct list_head *page_list,
937                                         unsigned int *count)
938 {
939         int nr_active = 0;
940         int lru;
941         struct page *page;
942
943         list_for_each_entry(page, page_list, lru) {
944                 lru = page_is_file_cache(page);
945                 if (PageActive(page)) {
946                         lru += LRU_ACTIVE;
947                         ClearPageActive(page);
948                         nr_active++;
949                 }
950                 count[lru]++;
951         }
952
953         return nr_active;
954 }
955
956 /**
957  * isolate_lru_page - tries to isolate a page from its LRU list
958  * @page: page to isolate from its LRU list
959  *
960  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
961  * vmstat statistic corresponding to whatever LRU list the page was on.
962  *
963  * Returns 0 if the page was removed from an LRU list.
964  * Returns -EBUSY if the page was not on an LRU list.
965  *
966  * The returned page will have PageLRU() cleared.  If it was found on
967  * the active list, it will have PageActive set.  If it was found on
968  * the unevictable list, it will have the PageUnevictable bit set. That flag
969  * may need to be cleared by the caller before letting the page go.
970  *
971  * The vmstat statistic corresponding to the list on which the page was
972  * found will be decremented.
973  *
974  * Restrictions:
975  * (1) Must be called with an elevated refcount on the page. This is a
976  *     fundamentnal difference from isolate_lru_pages (which is called
977  *     without a stable reference).
978  * (2) the lru_lock must not be held.
979  * (3) interrupts must be enabled.
980  */
981 int isolate_lru_page(struct page *page)
982 {
983         int ret = -EBUSY;
984
985         if (PageLRU(page)) {
986                 struct zone *zone = page_zone(page);
987
988                 spin_lock_irq(&zone->lru_lock);
989                 if (PageLRU(page) && get_page_unless_zero(page)) {
990                         int lru = page_lru(page);
991                         ret = 0;
992                         ClearPageLRU(page);
993
994                         del_page_from_lru_list(zone, page, lru);
995                 }
996                 spin_unlock_irq(&zone->lru_lock);
997         }
998         return ret;
999 }
1000
1001 /*
1002  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1003  * of reclaimed pages
1004  */
1005 static unsigned long shrink_inactive_list(unsigned long max_scan,
1006                         struct zone *zone, struct scan_control *sc,
1007                         int priority, int file)
1008 {
1009         LIST_HEAD(page_list);
1010         struct pagevec pvec;
1011         unsigned long nr_scanned = 0;
1012         unsigned long nr_reclaimed = 0;
1013
1014         pagevec_init(&pvec, 1);
1015
1016         lru_add_drain();
1017         spin_lock_irq(&zone->lru_lock);
1018         do {
1019                 struct page *page;
1020                 unsigned long nr_taken;
1021                 unsigned long nr_scan;
1022                 unsigned long nr_freed;
1023                 unsigned long nr_active;
1024                 unsigned int count[NR_LRU_LISTS] = { 0, };
1025                 int mode = ISOLATE_INACTIVE;
1026
1027                 /*
1028                  * If we need a large contiguous chunk of memory, or have
1029                  * trouble getting a small set of contiguous pages, we
1030                  * will reclaim both active and inactive pages.
1031                  *
1032                  * We use the same threshold as pageout congestion_wait below.
1033                  */
1034                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1035                         mode = ISOLATE_BOTH;
1036                 else if (sc->order && priority < DEF_PRIORITY - 2)
1037                         mode = ISOLATE_BOTH;
1038
1039                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1040                              &page_list, &nr_scan, sc->order, mode,
1041                                 zone, sc->mem_cgroup, 0, file);
1042                 nr_active = clear_active_flags(&page_list, count);
1043                 __count_vm_events(PGDEACTIVATE, nr_active);
1044
1045                 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1046                                                 -count[LRU_ACTIVE_FILE]);
1047                 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1048                                                 -count[LRU_INACTIVE_FILE]);
1049                 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1050                                                 -count[LRU_ACTIVE_ANON]);
1051                 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1052                                                 -count[LRU_INACTIVE_ANON]);
1053
1054                 if (scan_global_lru(sc)) {
1055                         zone->pages_scanned += nr_scan;
1056                         zone->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1057                         zone->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1058                         zone->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1059                         zone->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1060                 }
1061                 spin_unlock_irq(&zone->lru_lock);
1062
1063                 nr_scanned += nr_scan;
1064                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1065
1066                 /*
1067                  * If we are direct reclaiming for contiguous pages and we do
1068                  * not reclaim everything in the list, try again and wait
1069                  * for IO to complete. This will stall high-order allocations
1070                  * but that should be acceptable to the caller
1071                  */
1072                 if (nr_freed < nr_taken && !current_is_kswapd() &&
1073                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1074                         congestion_wait(WRITE, HZ/10);
1075
1076                         /*
1077                          * The attempt at page out may have made some
1078                          * of the pages active, mark them inactive again.
1079                          */
1080                         nr_active = clear_active_flags(&page_list, count);
1081                         count_vm_events(PGDEACTIVATE, nr_active);
1082
1083                         nr_freed += shrink_page_list(&page_list, sc,
1084                                                         PAGEOUT_IO_SYNC);
1085                 }
1086
1087                 nr_reclaimed += nr_freed;
1088                 local_irq_disable();
1089                 if (current_is_kswapd()) {
1090                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1091                         __count_vm_events(KSWAPD_STEAL, nr_freed);
1092                 } else if (scan_global_lru(sc))
1093                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1094
1095                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1096
1097                 if (nr_taken == 0)
1098                         goto done;
1099
1100                 spin_lock(&zone->lru_lock);
1101                 /*
1102                  * Put back any unfreeable pages.
1103                  */
1104                 while (!list_empty(&page_list)) {
1105                         int lru;
1106                         page = lru_to_page(&page_list);
1107                         VM_BUG_ON(PageLRU(page));
1108                         list_del(&page->lru);
1109                         if (unlikely(!page_evictable(page, NULL))) {
1110                                 spin_unlock_irq(&zone->lru_lock);
1111                                 putback_lru_page(page);
1112                                 spin_lock_irq(&zone->lru_lock);
1113                                 continue;
1114                         }
1115                         SetPageLRU(page);
1116                         lru = page_lru(page);
1117                         add_page_to_lru_list(zone, page, lru);
1118                         mem_cgroup_move_lists(page, lru);
1119                         if (PageActive(page) && scan_global_lru(sc)) {
1120                                 int file = !!page_is_file_cache(page);
1121                                 zone->recent_rotated[file]++;
1122                         }
1123                         if (!pagevec_add(&pvec, page)) {
1124                                 spin_unlock_irq(&zone->lru_lock);
1125                                 __pagevec_release(&pvec);
1126                                 spin_lock_irq(&zone->lru_lock);
1127                         }
1128                 }
1129         } while (nr_scanned < max_scan);
1130         spin_unlock(&zone->lru_lock);
1131 done:
1132         local_irq_enable();
1133         pagevec_release(&pvec);
1134         return nr_reclaimed;
1135 }
1136
1137 /*
1138  * We are about to scan this zone at a certain priority level.  If that priority
1139  * level is smaller (ie: more urgent) than the previous priority, then note
1140  * that priority level within the zone.  This is done so that when the next
1141  * process comes in to scan this zone, it will immediately start out at this
1142  * priority level rather than having to build up its own scanning priority.
1143  * Here, this priority affects only the reclaim-mapped threshold.
1144  */
1145 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1146 {
1147         if (priority < zone->prev_priority)
1148                 zone->prev_priority = priority;
1149 }
1150
1151 static inline int zone_is_near_oom(struct zone *zone)
1152 {
1153         return zone->pages_scanned >= (zone_lru_pages(zone) * 3);
1154 }
1155
1156 /*
1157  * This moves pages from the active list to the inactive list.
1158  *
1159  * We move them the other way if the page is referenced by one or more
1160  * processes, from rmap.
1161  *
1162  * If the pages are mostly unmapped, the processing is fast and it is
1163  * appropriate to hold zone->lru_lock across the whole operation.  But if
1164  * the pages are mapped, the processing is slow (page_referenced()) so we
1165  * should drop zone->lru_lock around each page.  It's impossible to balance
1166  * this, so instead we remove the pages from the LRU while processing them.
1167  * It is safe to rely on PG_active against the non-LRU pages in here because
1168  * nobody will play with that bit on a non-LRU page.
1169  *
1170  * The downside is that we have to touch page->_count against each page.
1171  * But we had to alter page->flags anyway.
1172  */
1173
1174
1175 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1176                         struct scan_control *sc, int priority, int file)
1177 {
1178         unsigned long pgmoved;
1179         int pgdeactivate = 0;
1180         unsigned long pgscanned;
1181         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1182         LIST_HEAD(l_inactive);
1183         struct page *page;
1184         struct pagevec pvec;
1185         enum lru_list lru;
1186
1187         lru_add_drain();
1188         spin_lock_irq(&zone->lru_lock);
1189         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1190                                         ISOLATE_ACTIVE, zone,
1191                                         sc->mem_cgroup, 1, file);
1192         /*
1193          * zone->pages_scanned is used for detect zone's oom
1194          * mem_cgroup remembers nr_scan by itself.
1195          */
1196         if (scan_global_lru(sc)) {
1197                 zone->pages_scanned += pgscanned;
1198                 zone->recent_scanned[!!file] += pgmoved;
1199         }
1200
1201         if (file)
1202                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1203         else
1204                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1205         spin_unlock_irq(&zone->lru_lock);
1206
1207         pgmoved = 0;
1208         while (!list_empty(&l_hold)) {
1209                 cond_resched();
1210                 page = lru_to_page(&l_hold);
1211                 list_del(&page->lru);
1212
1213                 if (unlikely(!page_evictable(page, NULL))) {
1214                         putback_lru_page(page);
1215                         continue;
1216                 }
1217
1218                 /* page_referenced clears PageReferenced */
1219                 if (page_mapping_inuse(page) &&
1220                     page_referenced(page, 0, sc->mem_cgroup))
1221                         pgmoved++;
1222
1223                 list_add(&page->lru, &l_inactive);
1224         }
1225
1226         /*
1227          * Count referenced pages from currently used mappings as
1228          * rotated, even though they are moved to the inactive list.
1229          * This helps balance scan pressure between file and anonymous
1230          * pages in get_scan_ratio.
1231          */
1232         zone->recent_rotated[!!file] += pgmoved;
1233
1234         /*
1235          * Move the pages to the [file or anon] inactive list.
1236          */
1237         pagevec_init(&pvec, 1);
1238
1239         pgmoved = 0;
1240         lru = LRU_BASE + file * LRU_FILE;
1241         spin_lock_irq(&zone->lru_lock);
1242         while (!list_empty(&l_inactive)) {
1243                 page = lru_to_page(&l_inactive);
1244                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1245                 VM_BUG_ON(PageLRU(page));
1246                 SetPageLRU(page);
1247                 VM_BUG_ON(!PageActive(page));
1248                 ClearPageActive(page);
1249
1250                 list_move(&page->lru, &zone->lru[lru].list);
1251                 mem_cgroup_move_lists(page, lru);
1252                 pgmoved++;
1253                 if (!pagevec_add(&pvec, page)) {
1254                         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1255                         spin_unlock_irq(&zone->lru_lock);
1256                         pgdeactivate += pgmoved;
1257                         pgmoved = 0;
1258                         if (buffer_heads_over_limit)
1259                                 pagevec_strip(&pvec);
1260                         __pagevec_release(&pvec);
1261                         spin_lock_irq(&zone->lru_lock);
1262                 }
1263         }
1264         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1265         pgdeactivate += pgmoved;
1266         if (buffer_heads_over_limit) {
1267                 spin_unlock_irq(&zone->lru_lock);
1268                 pagevec_strip(&pvec);
1269                 spin_lock_irq(&zone->lru_lock);
1270         }
1271         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1272         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1273         spin_unlock_irq(&zone->lru_lock);
1274         if (vm_swap_full())
1275                 pagevec_swap_free(&pvec);
1276
1277         pagevec_release(&pvec);
1278 }
1279
1280 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1281         struct zone *zone, struct scan_control *sc, int priority)
1282 {
1283         int file = is_file_lru(lru);
1284
1285         if (lru == LRU_ACTIVE_FILE) {
1286                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1287                 return 0;
1288         }
1289
1290         if (lru == LRU_ACTIVE_ANON &&
1291             (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1292                 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1293                 return 0;
1294         }
1295         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1296 }
1297
1298 /*
1299  * Determine how aggressively the anon and file LRU lists should be
1300  * scanned.  The relative value of each set of LRU lists is determined
1301  * by looking at the fraction of the pages scanned we did rotate back
1302  * onto the active list instead of evict.
1303  *
1304  * percent[0] specifies how much pressure to put on ram/swap backed
1305  * memory, while percent[1] determines pressure on the file LRUs.
1306  */
1307 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1308                                         unsigned long *percent)
1309 {
1310         unsigned long anon, file, free;
1311         unsigned long anon_prio, file_prio;
1312         unsigned long ap, fp;
1313
1314         anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
1315                 zone_page_state(zone, NR_INACTIVE_ANON);
1316         file  = zone_page_state(zone, NR_ACTIVE_FILE) +
1317                 zone_page_state(zone, NR_INACTIVE_FILE);
1318         free  = zone_page_state(zone, NR_FREE_PAGES);
1319
1320         /* If we have no swap space, do not bother scanning anon pages. */
1321         if (nr_swap_pages <= 0) {
1322                 percent[0] = 0;
1323                 percent[1] = 100;
1324                 return;
1325         }
1326
1327         /* If we have very few page cache pages, force-scan anon pages. */
1328         if (unlikely(file + free <= zone->pages_high)) {
1329                 percent[0] = 100;
1330                 percent[1] = 0;
1331                 return;
1332         }
1333
1334         /*
1335          * OK, so we have swap space and a fair amount of page cache
1336          * pages.  We use the recently rotated / recently scanned
1337          * ratios to determine how valuable each cache is.
1338          *
1339          * Because workloads change over time (and to avoid overflow)
1340          * we keep these statistics as a floating average, which ends
1341          * up weighing recent references more than old ones.
1342          *
1343          * anon in [0], file in [1]
1344          */
1345         if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1346                 spin_lock_irq(&zone->lru_lock);
1347                 zone->recent_scanned[0] /= 2;
1348                 zone->recent_rotated[0] /= 2;
1349                 spin_unlock_irq(&zone->lru_lock);
1350         }
1351
1352         if (unlikely(zone->recent_scanned[1] > file / 4)) {
1353                 spin_lock_irq(&zone->lru_lock);
1354                 zone->recent_scanned[1] /= 2;
1355                 zone->recent_rotated[1] /= 2;
1356                 spin_unlock_irq(&zone->lru_lock);
1357         }
1358
1359         /*
1360          * With swappiness at 100, anonymous and file have the same priority.
1361          * This scanning priority is essentially the inverse of IO cost.
1362          */
1363         anon_prio = sc->swappiness;
1364         file_prio = 200 - sc->swappiness;
1365
1366         /*
1367          *                  anon       recent_rotated[0]
1368          * %anon = 100 * ----------- / ----------------- * IO cost
1369          *               anon + file      rotate_sum
1370          */
1371         ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1372         ap /= zone->recent_rotated[0] + 1;
1373
1374         fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1375         fp /= zone->recent_rotated[1] + 1;
1376
1377         /* Normalize to percentages */
1378         percent[0] = 100 * ap / (ap + fp + 1);
1379         percent[1] = 100 - percent[0];
1380 }
1381
1382
1383 /*
1384  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1385  */
1386 static unsigned long shrink_zone(int priority, struct zone *zone,
1387                                 struct scan_control *sc)
1388 {
1389         unsigned long nr[NR_LRU_LISTS];
1390         unsigned long nr_to_scan;
1391         unsigned long nr_reclaimed = 0;
1392         unsigned long percent[2];       /* anon @ 0; file @ 1 */
1393         enum lru_list l;
1394
1395         get_scan_ratio(zone, sc, percent);
1396
1397         for_each_evictable_lru(l) {
1398                 if (scan_global_lru(sc)) {
1399                         int file = is_file_lru(l);
1400                         int scan;
1401                         /*
1402                          * Add one to nr_to_scan just to make sure that the
1403                          * kernel will slowly sift through each list.
1404                          */
1405                         scan = zone_page_state(zone, NR_LRU_BASE + l);
1406                         if (priority) {
1407                                 scan >>= priority;
1408                                 scan = (scan * percent[file]) / 100;
1409                         }
1410                         zone->lru[l].nr_scan += scan + 1;
1411                         nr[l] = zone->lru[l].nr_scan;
1412                         if (nr[l] >= sc->swap_cluster_max)
1413                                 zone->lru[l].nr_scan = 0;
1414                         else
1415                                 nr[l] = 0;
1416                 } else {
1417                         /*
1418                          * This reclaim occurs not because zone memory shortage
1419                          * but because memory controller hits its limit.
1420                          * Don't modify zone reclaim related data.
1421                          */
1422                         nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1423                                                                 priority, l);
1424                 }
1425         }
1426
1427         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1428                                         nr[LRU_INACTIVE_FILE]) {
1429                 for_each_evictable_lru(l) {
1430                         if (nr[l]) {
1431                                 nr_to_scan = min(nr[l],
1432                                         (unsigned long)sc->swap_cluster_max);
1433                                 nr[l] -= nr_to_scan;
1434
1435                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1436                                                         zone, sc, priority);
1437                         }
1438                 }
1439         }
1440
1441         /*
1442          * Even if we did not try to evict anon pages at all, we want to
1443          * rebalance the anon lru active/inactive ratio.
1444          */
1445         if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1446                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1447         else if (!scan_global_lru(sc))
1448                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1449
1450         throttle_vm_writeout(sc->gfp_mask);
1451         return nr_reclaimed;
1452 }
1453
1454 /*
1455  * This is the direct reclaim path, for page-allocating processes.  We only
1456  * try to reclaim pages from zones which will satisfy the caller's allocation
1457  * request.
1458  *
1459  * We reclaim from a zone even if that zone is over pages_high.  Because:
1460  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1461  *    allocation or
1462  * b) The zones may be over pages_high but they must go *over* pages_high to
1463  *    satisfy the `incremental min' zone defense algorithm.
1464  *
1465  * Returns the number of reclaimed pages.
1466  *
1467  * If a zone is deemed to be full of pinned pages then just give it a light
1468  * scan then give up on it.
1469  */
1470 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1471                                         struct scan_control *sc)
1472 {
1473         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1474         unsigned long nr_reclaimed = 0;
1475         struct zoneref *z;
1476         struct zone *zone;
1477
1478         sc->all_unreclaimable = 1;
1479         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1480                 if (!populated_zone(zone))
1481                         continue;
1482                 /*
1483                  * Take care memory controller reclaiming has small influence
1484                  * to global LRU.
1485                  */
1486                 if (scan_global_lru(sc)) {
1487                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1488                                 continue;
1489                         note_zone_scanning_priority(zone, priority);
1490
1491                         if (zone_is_all_unreclaimable(zone) &&
1492                                                 priority != DEF_PRIORITY)
1493                                 continue;       /* Let kswapd poll it */
1494                         sc->all_unreclaimable = 0;
1495                 } else {
1496                         /*
1497                          * Ignore cpuset limitation here. We just want to reduce
1498                          * # of used pages by us regardless of memory shortage.
1499                          */
1500                         sc->all_unreclaimable = 0;
1501                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1502                                                         priority);
1503                 }
1504
1505                 nr_reclaimed += shrink_zone(priority, zone, sc);
1506         }
1507
1508         return nr_reclaimed;
1509 }
1510
1511 /*
1512  * This is the main entry point to direct page reclaim.
1513  *
1514  * If a full scan of the inactive list fails to free enough memory then we
1515  * are "out of memory" and something needs to be killed.
1516  *
1517  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1518  * high - the zone may be full of dirty or under-writeback pages, which this
1519  * caller can't do much about.  We kick pdflush and take explicit naps in the
1520  * hope that some of these pages can be written.  But if the allocating task
1521  * holds filesystem locks which prevent writeout this might not work, and the
1522  * allocation attempt will fail.
1523  *
1524  * returns:     0, if no pages reclaimed
1525  *              else, the number of pages reclaimed
1526  */
1527 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1528                                         struct scan_control *sc)
1529 {
1530         int priority;
1531         unsigned long ret = 0;
1532         unsigned long total_scanned = 0;
1533         unsigned long nr_reclaimed = 0;
1534         struct reclaim_state *reclaim_state = current->reclaim_state;
1535         unsigned long lru_pages = 0;
1536         struct zoneref *z;
1537         struct zone *zone;
1538         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1539
1540         delayacct_freepages_start();
1541
1542         if (scan_global_lru(sc))
1543                 count_vm_event(ALLOCSTALL);
1544         /*
1545          * mem_cgroup will not do shrink_slab.
1546          */
1547         if (scan_global_lru(sc)) {
1548                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1549
1550                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1551                                 continue;
1552
1553                         lru_pages += zone_lru_pages(zone);
1554                 }
1555         }
1556
1557         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1558                 sc->nr_scanned = 0;
1559                 if (!priority)
1560                         disable_swap_token();
1561                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1562                 /*
1563                  * Don't shrink slabs when reclaiming memory from
1564                  * over limit cgroups
1565                  */
1566                 if (scan_global_lru(sc)) {
1567                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1568                         if (reclaim_state) {
1569                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1570                                 reclaim_state->reclaimed_slab = 0;
1571                         }
1572                 }
1573                 total_scanned += sc->nr_scanned;
1574                 if (nr_reclaimed >= sc->swap_cluster_max) {
1575                         ret = nr_reclaimed;
1576                         goto out;
1577                 }
1578
1579                 /*
1580                  * Try to write back as many pages as we just scanned.  This
1581                  * tends to cause slow streaming writers to write data to the
1582                  * disk smoothly, at the dirtying rate, which is nice.   But
1583                  * that's undesirable in laptop mode, where we *want* lumpy
1584                  * writeout.  So in laptop mode, write out the whole world.
1585                  */
1586                 if (total_scanned > sc->swap_cluster_max +
1587                                         sc->swap_cluster_max / 2) {
1588                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1589                         sc->may_writepage = 1;
1590                 }
1591
1592                 /* Take a nap, wait for some writeback to complete */
1593                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1594                         congestion_wait(WRITE, HZ/10);
1595         }
1596         /* top priority shrink_zones still had more to do? don't OOM, then */
1597         if (!sc->all_unreclaimable && scan_global_lru(sc))
1598                 ret = nr_reclaimed;
1599 out:
1600         /*
1601          * Now that we've scanned all the zones at this priority level, note
1602          * that level within the zone so that the next thread which performs
1603          * scanning of this zone will immediately start out at this priority
1604          * level.  This affects only the decision whether or not to bring
1605          * mapped pages onto the inactive list.
1606          */
1607         if (priority < 0)
1608                 priority = 0;
1609
1610         if (scan_global_lru(sc)) {
1611                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1612
1613                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1614                                 continue;
1615
1616                         zone->prev_priority = priority;
1617                 }
1618         } else
1619                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1620
1621         delayacct_freepages_end();
1622
1623         return ret;
1624 }
1625
1626 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1627                                                                 gfp_t gfp_mask)
1628 {
1629         struct scan_control sc = {
1630                 .gfp_mask = gfp_mask,
1631                 .may_writepage = !laptop_mode,
1632                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1633                 .may_swap = 1,
1634                 .swappiness = vm_swappiness,
1635                 .order = order,
1636                 .mem_cgroup = NULL,
1637                 .isolate_pages = isolate_pages_global,
1638         };
1639
1640         return do_try_to_free_pages(zonelist, &sc);
1641 }
1642
1643 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1644
1645 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1646                                                 gfp_t gfp_mask)
1647 {
1648         struct scan_control sc = {
1649                 .may_writepage = !laptop_mode,
1650                 .may_swap = 1,
1651                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1652                 .swappiness = vm_swappiness,
1653                 .order = 0,
1654                 .mem_cgroup = mem_cont,
1655                 .isolate_pages = mem_cgroup_isolate_pages,
1656         };
1657         struct zonelist *zonelist;
1658
1659         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1660                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1661         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1662         return do_try_to_free_pages(zonelist, &sc);
1663 }
1664 #endif
1665
1666 /*
1667  * For kswapd, balance_pgdat() will work across all this node's zones until
1668  * they are all at pages_high.
1669  *
1670  * Returns the number of pages which were actually freed.
1671  *
1672  * There is special handling here for zones which are full of pinned pages.
1673  * This can happen if the pages are all mlocked, or if they are all used by
1674  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1675  * What we do is to detect the case where all pages in the zone have been
1676  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1677  * dead and from now on, only perform a short scan.  Basically we're polling
1678  * the zone for when the problem goes away.
1679  *
1680  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1681  * zones which have free_pages > pages_high, but once a zone is found to have
1682  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1683  * of the number of free pages in the lower zones.  This interoperates with
1684  * the page allocator fallback scheme to ensure that aging of pages is balanced
1685  * across the zones.
1686  */
1687 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1688 {
1689         int all_zones_ok;
1690         int priority;
1691         int i;
1692         unsigned long total_scanned;
1693         unsigned long nr_reclaimed;
1694         struct reclaim_state *reclaim_state = current->reclaim_state;
1695         struct scan_control sc = {
1696                 .gfp_mask = GFP_KERNEL,
1697                 .may_swap = 1,
1698                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1699                 .swappiness = vm_swappiness,
1700                 .order = order,
1701                 .mem_cgroup = NULL,
1702                 .isolate_pages = isolate_pages_global,
1703         };
1704         /*
1705          * temp_priority is used to remember the scanning priority at which
1706          * this zone was successfully refilled to free_pages == pages_high.
1707          */
1708         int temp_priority[MAX_NR_ZONES];
1709
1710 loop_again:
1711         total_scanned = 0;
1712         nr_reclaimed = 0;
1713         sc.may_writepage = !laptop_mode;
1714         count_vm_event(PAGEOUTRUN);
1715
1716         for (i = 0; i < pgdat->nr_zones; i++)
1717                 temp_priority[i] = DEF_PRIORITY;
1718
1719         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1720                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1721                 unsigned long lru_pages = 0;
1722
1723                 /* The swap token gets in the way of swapout... */
1724                 if (!priority)
1725                         disable_swap_token();
1726
1727                 all_zones_ok = 1;
1728
1729                 /*
1730                  * Scan in the highmem->dma direction for the highest
1731                  * zone which needs scanning
1732                  */
1733                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1734                         struct zone *zone = pgdat->node_zones + i;
1735
1736                         if (!populated_zone(zone))
1737                                 continue;
1738
1739                         if (zone_is_all_unreclaimable(zone) &&
1740                             priority != DEF_PRIORITY)
1741                                 continue;
1742
1743                         /*
1744                          * Do some background aging of the anon list, to give
1745                          * pages a chance to be referenced before reclaiming.
1746                          */
1747                         if (inactive_anon_is_low(zone))
1748                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1749                                                         &sc, priority, 0);
1750
1751                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1752                                                0, 0)) {
1753                                 end_zone = i;
1754                                 break;
1755                         }
1756                 }
1757                 if (i < 0)
1758                         goto out;
1759
1760                 for (i = 0; i <= end_zone; i++) {
1761                         struct zone *zone = pgdat->node_zones + i;
1762
1763                         lru_pages += zone_lru_pages(zone);
1764                 }
1765
1766                 /*
1767                  * Now scan the zone in the dma->highmem direction, stopping
1768                  * at the last zone which needs scanning.
1769                  *
1770                  * We do this because the page allocator works in the opposite
1771                  * direction.  This prevents the page allocator from allocating
1772                  * pages behind kswapd's direction of progress, which would
1773                  * cause too much scanning of the lower zones.
1774                  */
1775                 for (i = 0; i <= end_zone; i++) {
1776                         struct zone *zone = pgdat->node_zones + i;
1777                         int nr_slab;
1778
1779                         if (!populated_zone(zone))
1780                                 continue;
1781
1782                         if (zone_is_all_unreclaimable(zone) &&
1783                                         priority != DEF_PRIORITY)
1784                                 continue;
1785
1786                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1787                                                end_zone, 0))
1788                                 all_zones_ok = 0;
1789                         temp_priority[i] = priority;
1790                         sc.nr_scanned = 0;
1791                         note_zone_scanning_priority(zone, priority);
1792                         /*
1793                          * We put equal pressure on every zone, unless one
1794                          * zone has way too many pages free already.
1795                          */
1796                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1797                                                 end_zone, 0))
1798                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1799                         reclaim_state->reclaimed_slab = 0;
1800                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1801                                                 lru_pages);
1802                         nr_reclaimed += reclaim_state->reclaimed_slab;
1803                         total_scanned += sc.nr_scanned;
1804                         if (zone_is_all_unreclaimable(zone))
1805                                 continue;
1806                         if (nr_slab == 0 && zone->pages_scanned >=
1807                                                 (zone_lru_pages(zone) * 6))
1808                                         zone_set_flag(zone,
1809                                                       ZONE_ALL_UNRECLAIMABLE);
1810                         /*
1811                          * If we've done a decent amount of scanning and
1812                          * the reclaim ratio is low, start doing writepage
1813                          * even in laptop mode
1814                          */
1815                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1816                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1817                                 sc.may_writepage = 1;
1818                 }
1819                 if (all_zones_ok)
1820                         break;          /* kswapd: all done */
1821                 /*
1822                  * OK, kswapd is getting into trouble.  Take a nap, then take
1823                  * another pass across the zones.
1824                  */
1825                 if (total_scanned && priority < DEF_PRIORITY - 2)
1826                         congestion_wait(WRITE, HZ/10);
1827
1828                 /*
1829                  * We do this so kswapd doesn't build up large priorities for
1830                  * example when it is freeing in parallel with allocators. It
1831                  * matches the direct reclaim path behaviour in terms of impact
1832                  * on zone->*_priority.
1833                  */
1834                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1835                         break;
1836         }
1837 out:
1838         /*
1839          * Note within each zone the priority level at which this zone was
1840          * brought into a happy state.  So that the next thread which scans this
1841          * zone will start out at that priority level.
1842          */
1843         for (i = 0; i < pgdat->nr_zones; i++) {
1844                 struct zone *zone = pgdat->node_zones + i;
1845
1846                 zone->prev_priority = temp_priority[i];
1847         }
1848         if (!all_zones_ok) {
1849                 cond_resched();
1850
1851                 try_to_freeze();
1852
1853                 goto loop_again;
1854         }
1855
1856         return nr_reclaimed;
1857 }
1858
1859 /*
1860  * The background pageout daemon, started as a kernel thread
1861  * from the init process.
1862  *
1863  * This basically trickles out pages so that we have _some_
1864  * free memory available even if there is no other activity
1865  * that frees anything up. This is needed for things like routing
1866  * etc, where we otherwise might have all activity going on in
1867  * asynchronous contexts that cannot page things out.
1868  *
1869  * If there are applications that are active memory-allocators
1870  * (most normal use), this basically shouldn't matter.
1871  */
1872 static int kswapd(void *p)
1873 {
1874         unsigned long order;
1875         pg_data_t *pgdat = (pg_data_t*)p;
1876         struct task_struct *tsk = current;
1877         DEFINE_WAIT(wait);
1878         struct reclaim_state reclaim_state = {
1879                 .reclaimed_slab = 0,
1880         };
1881         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1882
1883         if (!cpus_empty(*cpumask))
1884                 set_cpus_allowed_ptr(tsk, cpumask);
1885         current->reclaim_state = &reclaim_state;
1886
1887         /*
1888          * Tell the memory management that we're a "memory allocator",
1889          * and that if we need more memory we should get access to it
1890          * regardless (see "__alloc_pages()"). "kswapd" should
1891          * never get caught in the normal page freeing logic.
1892          *
1893          * (Kswapd normally doesn't need memory anyway, but sometimes
1894          * you need a small amount of memory in order to be able to
1895          * page out something else, and this flag essentially protects
1896          * us from recursively trying to free more memory as we're
1897          * trying to free the first piece of memory in the first place).
1898          */
1899         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1900         set_freezable();
1901
1902         order = 0;
1903         for ( ; ; ) {
1904                 unsigned long new_order;
1905
1906                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1907                 new_order = pgdat->kswapd_max_order;
1908                 pgdat->kswapd_max_order = 0;
1909                 if (order < new_order) {
1910                         /*
1911                          * Don't sleep if someone wants a larger 'order'
1912                          * allocation
1913                          */
1914                         order = new_order;
1915                 } else {
1916                         if (!freezing(current))
1917                                 schedule();
1918
1919                         order = pgdat->kswapd_max_order;
1920                 }
1921                 finish_wait(&pgdat->kswapd_wait, &wait);
1922
1923                 if (!try_to_freeze()) {
1924                         /* We can speed up thawing tasks if we don't call
1925                          * balance_pgdat after returning from the refrigerator
1926                          */
1927                         balance_pgdat(pgdat, order);
1928                 }
1929         }
1930         return 0;
1931 }
1932
1933 /*
1934  * A zone is low on free memory, so wake its kswapd task to service it.
1935  */
1936 void wakeup_kswapd(struct zone *zone, int order)
1937 {
1938         pg_data_t *pgdat;
1939
1940         if (!populated_zone(zone))
1941                 return;
1942
1943         pgdat = zone->zone_pgdat;
1944         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1945                 return;
1946         if (pgdat->kswapd_max_order < order)
1947                 pgdat->kswapd_max_order = order;
1948         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1949                 return;
1950         if (!waitqueue_active(&pgdat->kswapd_wait))
1951                 return;
1952         wake_up_interruptible(&pgdat->kswapd_wait);
1953 }
1954
1955 unsigned long global_lru_pages(void)
1956 {
1957         return global_page_state(NR_ACTIVE_ANON)
1958                 + global_page_state(NR_ACTIVE_FILE)
1959                 + global_page_state(NR_INACTIVE_ANON)
1960                 + global_page_state(NR_INACTIVE_FILE);
1961 }
1962
1963 #ifdef CONFIG_PM
1964 /*
1965  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1966  * from LRU lists system-wide, for given pass and priority, and returns the
1967  * number of reclaimed pages
1968  *
1969  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1970  */
1971 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1972                                       int pass, struct scan_control *sc)
1973 {
1974         struct zone *zone;
1975         unsigned long nr_to_scan, ret = 0;
1976         enum lru_list l;
1977
1978         for_each_zone(zone) {
1979
1980                 if (!populated_zone(zone))
1981                         continue;
1982
1983                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1984                         continue;
1985
1986                 for_each_evictable_lru(l) {
1987                         /* For pass = 0, we don't shrink the active list */
1988                         if (pass == 0 &&
1989                                 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
1990                                 continue;
1991
1992                         zone->lru[l].nr_scan +=
1993                                 (zone_page_state(zone, NR_LRU_BASE + l)
1994                                                                 >> prio) + 1;
1995                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
1996                                 zone->lru[l].nr_scan = 0;
1997                                 nr_to_scan = min(nr_pages,
1998                                         zone_page_state(zone,
1999                                                         NR_LRU_BASE + l));
2000                                 ret += shrink_list(l, nr_to_scan, zone,
2001                                                                 sc, prio);
2002                                 if (ret >= nr_pages)
2003                                         return ret;
2004                         }
2005                 }
2006         }
2007
2008         return ret;
2009 }
2010
2011 /*
2012  * Try to free `nr_pages' of memory, system-wide, and return the number of
2013  * freed pages.
2014  *
2015  * Rather than trying to age LRUs the aim is to preserve the overall
2016  * LRU order by reclaiming preferentially
2017  * inactive > active > active referenced > active mapped
2018  */
2019 unsigned long shrink_all_memory(unsigned long nr_pages)
2020 {
2021         unsigned long lru_pages, nr_slab;
2022         unsigned long ret = 0;
2023         int pass;
2024         struct reclaim_state reclaim_state;
2025         struct scan_control sc = {
2026                 .gfp_mask = GFP_KERNEL,
2027                 .may_swap = 0,
2028                 .swap_cluster_max = nr_pages,
2029                 .may_writepage = 1,
2030                 .swappiness = vm_swappiness,
2031                 .isolate_pages = isolate_pages_global,
2032         };
2033
2034         current->reclaim_state = &reclaim_state;
2035
2036         lru_pages = global_lru_pages();
2037         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2038         /* If slab caches are huge, it's better to hit them first */
2039         while (nr_slab >= lru_pages) {
2040                 reclaim_state.reclaimed_slab = 0;
2041                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2042                 if (!reclaim_state.reclaimed_slab)
2043                         break;
2044
2045                 ret += reclaim_state.reclaimed_slab;
2046                 if (ret >= nr_pages)
2047                         goto out;
2048
2049                 nr_slab -= reclaim_state.reclaimed_slab;
2050         }
2051
2052         /*
2053          * We try to shrink LRUs in 5 passes:
2054          * 0 = Reclaim from inactive_list only
2055          * 1 = Reclaim from active list but don't reclaim mapped
2056          * 2 = 2nd pass of type 1
2057          * 3 = Reclaim mapped (normal reclaim)
2058          * 4 = 2nd pass of type 3
2059          */
2060         for (pass = 0; pass < 5; pass++) {
2061                 int prio;
2062
2063                 /* Force reclaiming mapped pages in the passes #3 and #4 */
2064                 if (pass > 2) {
2065                         sc.may_swap = 1;
2066                         sc.swappiness = 100;
2067                 }
2068
2069                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2070                         unsigned long nr_to_scan = nr_pages - ret;
2071
2072                         sc.nr_scanned = 0;
2073                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2074                         if (ret >= nr_pages)
2075                                 goto out;
2076
2077                         reclaim_state.reclaimed_slab = 0;
2078                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
2079                                         global_lru_pages());
2080                         ret += reclaim_state.reclaimed_slab;
2081                         if (ret >= nr_pages)
2082                                 goto out;
2083
2084                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2085                                 congestion_wait(WRITE, HZ / 10);
2086                 }
2087         }
2088
2089         /*
2090          * If ret = 0, we could not shrink LRUs, but there may be something
2091          * in slab caches
2092          */
2093         if (!ret) {
2094                 do {
2095                         reclaim_state.reclaimed_slab = 0;
2096                         shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2097                         ret += reclaim_state.reclaimed_slab;
2098                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2099         }
2100
2101 out:
2102         current->reclaim_state = NULL;
2103
2104         return ret;
2105 }
2106 #endif
2107
2108 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2109    not required for correctness.  So if the last cpu in a node goes
2110    away, we get changed to run anywhere: as the first one comes back,
2111    restore their cpu bindings. */
2112 static int __devinit cpu_callback(struct notifier_block *nfb,
2113                                   unsigned long action, void *hcpu)
2114 {
2115         int nid;
2116
2117         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2118                 for_each_node_state(nid, N_HIGH_MEMORY) {
2119                         pg_data_t *pgdat = NODE_DATA(nid);
2120                         node_to_cpumask_ptr(mask, pgdat->node_id);
2121
2122                         if (any_online_cpu(*mask) < nr_cpu_ids)
2123                                 /* One of our CPUs online: restore mask */
2124                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2125                 }
2126         }
2127         return NOTIFY_OK;
2128 }
2129
2130 /*
2131  * This kswapd start function will be called by init and node-hot-add.
2132  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2133  */
2134 int kswapd_run(int nid)
2135 {
2136         pg_data_t *pgdat = NODE_DATA(nid);
2137         int ret = 0;
2138
2139         if (pgdat->kswapd)
2140                 return 0;
2141
2142         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2143         if (IS_ERR(pgdat->kswapd)) {
2144                 /* failure at boot is fatal */
2145                 BUG_ON(system_state == SYSTEM_BOOTING);
2146                 printk("Failed to start kswapd on node %d\n",nid);
2147                 ret = -1;
2148         }
2149         return ret;
2150 }
2151
2152 static int __init kswapd_init(void)
2153 {
2154         int nid;
2155
2156         swap_setup();
2157         for_each_node_state(nid, N_HIGH_MEMORY)
2158                 kswapd_run(nid);
2159         hotcpu_notifier(cpu_callback, 0);
2160         return 0;
2161 }
2162
2163 module_init(kswapd_init)
2164
2165 #ifdef CONFIG_NUMA
2166 /*
2167  * Zone reclaim mode
2168  *
2169  * If non-zero call zone_reclaim when the number of free pages falls below
2170  * the watermarks.
2171  */
2172 int zone_reclaim_mode __read_mostly;
2173
2174 #define RECLAIM_OFF 0
2175 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2176 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2177 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2178
2179 /*
2180  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2181  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2182  * a zone.
2183  */
2184 #define ZONE_RECLAIM_PRIORITY 4
2185
2186 /*
2187  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2188  * occur.
2189  */
2190 int sysctl_min_unmapped_ratio = 1;
2191
2192 /*
2193  * If the number of slab pages in a zone grows beyond this percentage then
2194  * slab reclaim needs to occur.
2195  */
2196 int sysctl_min_slab_ratio = 5;
2197
2198 /*
2199  * Try to free up some pages from this zone through reclaim.
2200  */
2201 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2202 {
2203         /* Minimum pages needed in order to stay on node */
2204         const unsigned long nr_pages = 1 << order;
2205         struct task_struct *p = current;
2206         struct reclaim_state reclaim_state;
2207         int priority;
2208         unsigned long nr_reclaimed = 0;
2209         struct scan_control sc = {
2210                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2211                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2212                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2213                                         SWAP_CLUSTER_MAX),
2214                 .gfp_mask = gfp_mask,
2215                 .swappiness = vm_swappiness,
2216                 .isolate_pages = isolate_pages_global,
2217         };
2218         unsigned long slab_reclaimable;
2219
2220         disable_swap_token();
2221         cond_resched();
2222         /*
2223          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2224          * and we also need to be able to write out pages for RECLAIM_WRITE
2225          * and RECLAIM_SWAP.
2226          */
2227         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2228         reclaim_state.reclaimed_slab = 0;
2229         p->reclaim_state = &reclaim_state;
2230
2231         if (zone_page_state(zone, NR_FILE_PAGES) -
2232                 zone_page_state(zone, NR_FILE_MAPPED) >
2233                 zone->min_unmapped_pages) {
2234                 /*
2235                  * Free memory by calling shrink zone with increasing
2236                  * priorities until we have enough memory freed.
2237                  */
2238                 priority = ZONE_RECLAIM_PRIORITY;
2239                 do {
2240                         note_zone_scanning_priority(zone, priority);
2241                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2242                         priority--;
2243                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2244         }
2245
2246         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2247         if (slab_reclaimable > zone->min_slab_pages) {
2248                 /*
2249                  * shrink_slab() does not currently allow us to determine how
2250                  * many pages were freed in this zone. So we take the current
2251                  * number of slab pages and shake the slab until it is reduced
2252                  * by the same nr_pages that we used for reclaiming unmapped
2253                  * pages.
2254                  *
2255                  * Note that shrink_slab will free memory on all zones and may
2256                  * take a long time.
2257                  */
2258                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2259                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2260                                 slab_reclaimable - nr_pages)
2261                         ;
2262
2263                 /*
2264                  * Update nr_reclaimed by the number of slab pages we
2265                  * reclaimed from this zone.
2266                  */
2267                 nr_reclaimed += slab_reclaimable -
2268                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2269         }
2270
2271         p->reclaim_state = NULL;
2272         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2273         return nr_reclaimed >= nr_pages;
2274 }
2275
2276 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2277 {
2278         int node_id;
2279         int ret;
2280
2281         /*
2282          * Zone reclaim reclaims unmapped file backed pages and
2283          * slab pages if we are over the defined limits.
2284          *
2285          * A small portion of unmapped file backed pages is needed for
2286          * file I/O otherwise pages read by file I/O will be immediately
2287          * thrown out if the zone is overallocated. So we do not reclaim
2288          * if less than a specified percentage of the zone is used by
2289          * unmapped file backed pages.
2290          */
2291         if (zone_page_state(zone, NR_FILE_PAGES) -
2292             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2293             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2294                         <= zone->min_slab_pages)
2295                 return 0;
2296
2297         if (zone_is_all_unreclaimable(zone))
2298                 return 0;
2299
2300         /*
2301          * Do not scan if the allocation should not be delayed.
2302          */
2303         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2304                         return 0;
2305
2306         /*
2307          * Only run zone reclaim on the local zone or on zones that do not
2308          * have associated processors. This will favor the local processor
2309          * over remote processors and spread off node memory allocations
2310          * as wide as possible.
2311          */
2312         node_id = zone_to_nid(zone);
2313         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2314                 return 0;
2315
2316         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2317                 return 0;
2318         ret = __zone_reclaim(zone, gfp_mask, order);
2319         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2320
2321         return ret;
2322 }
2323 #endif
2324
2325 #ifdef CONFIG_UNEVICTABLE_LRU
2326 /*
2327  * page_evictable - test whether a page is evictable
2328  * @page: the page to test
2329  * @vma: the VMA in which the page is or will be mapped, may be NULL
2330  *
2331  * Test whether page is evictable--i.e., should be placed on active/inactive
2332  * lists vs unevictable list.
2333  *
2334  * Reasons page might not be evictable:
2335  * (1) page's mapping marked unevictable
2336  *
2337  * TODO - later patches
2338  */
2339 int page_evictable(struct page *page, struct vm_area_struct *vma)
2340 {
2341
2342         if (mapping_unevictable(page_mapping(page)))
2343                 return 0;
2344
2345         /* TODO:  test page [!]evictable conditions */
2346
2347         return 1;
2348 }
2349 #endif