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