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