46fdaa546b8d0493acb74b2b428693d46e0b14b3
[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);
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                 SetPageActive(page);
651                 pgactivate++;
652 keep_locked:
653                 unlock_page(page);
654 keep:
655                 list_add(&page->lru, &ret_pages);
656                 VM_BUG_ON(PageLRU(page));
657         }
658         list_splice(&ret_pages, page_list);
659         if (pagevec_count(&freed_pvec))
660                 __pagevec_free(&freed_pvec);
661         count_vm_events(PGACTIVATE, pgactivate);
662         return nr_reclaimed;
663 }
664
665 /* LRU Isolation modes. */
666 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
667 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
668 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
669
670 /*
671  * Attempt to remove the specified page from its LRU.  Only take this page
672  * if it is of the appropriate PageActive status.  Pages which are being
673  * freed elsewhere are also ignored.
674  *
675  * page:        page to consider
676  * mode:        one of the LRU isolation modes defined above
677  *
678  * returns 0 on success, -ve errno on failure.
679  */
680 int __isolate_lru_page(struct page *page, int mode)
681 {
682         int ret = -EINVAL;
683
684         /* Only take pages on the LRU. */
685         if (!PageLRU(page))
686                 return ret;
687
688         /*
689          * When checking the active state, we need to be sure we are
690          * dealing with comparible boolean values.  Take the logical not
691          * of each.
692          */
693         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
694                 return ret;
695
696         ret = -EBUSY;
697         if (likely(get_page_unless_zero(page))) {
698                 /*
699                  * Be careful not to clear PageLRU until after we're
700                  * sure the page is not being freed elsewhere -- the
701                  * page release code relies on it.
702                  */
703                 ClearPageLRU(page);
704                 ret = 0;
705         }
706
707         return ret;
708 }
709
710 /*
711  * zone->lru_lock is heavily contended.  Some of the functions that
712  * shrink the lists perform better by taking out a batch of pages
713  * and working on them outside the LRU lock.
714  *
715  * For pagecache intensive workloads, this function is the hottest
716  * spot in the kernel (apart from copy_*_user functions).
717  *
718  * Appropriate locks must be held before calling this function.
719  *
720  * @nr_to_scan: The number of pages to look through on the list.
721  * @src:        The LRU list to pull pages off.
722  * @dst:        The temp list to put pages on to.
723  * @scanned:    The number of pages that were scanned.
724  * @order:      The caller's attempted allocation order
725  * @mode:       One of the LRU isolation modes
726  *
727  * returns how many pages were moved onto *@dst.
728  */
729 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
730                 struct list_head *src, struct list_head *dst,
731                 unsigned long *scanned, int order, int mode)
732 {
733         unsigned long nr_taken = 0;
734         unsigned long scan;
735
736         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
737                 struct page *page;
738                 unsigned long pfn;
739                 unsigned long end_pfn;
740                 unsigned long page_pfn;
741                 int zone_id;
742
743                 page = lru_to_page(src);
744                 prefetchw_prev_lru_page(page, src, flags);
745
746                 VM_BUG_ON(!PageLRU(page));
747
748                 switch (__isolate_lru_page(page, mode)) {
749                 case 0:
750                         list_move(&page->lru, dst);
751                         nr_taken++;
752                         break;
753
754                 case -EBUSY:
755                         /* else it is being freed elsewhere */
756                         list_move(&page->lru, src);
757                         continue;
758
759                 default:
760                         BUG();
761                 }
762
763                 if (!order)
764                         continue;
765
766                 /*
767                  * Attempt to take all pages in the order aligned region
768                  * surrounding the tag page.  Only take those pages of
769                  * the same active state as that tag page.  We may safely
770                  * round the target page pfn down to the requested order
771                  * as the mem_map is guarenteed valid out to MAX_ORDER,
772                  * where that page is in a different zone we will detect
773                  * it from its zone id and abort this block scan.
774                  */
775                 zone_id = page_zone_id(page);
776                 page_pfn = page_to_pfn(page);
777                 pfn = page_pfn & ~((1 << order) - 1);
778                 end_pfn = pfn + (1 << order);
779                 for (; pfn < end_pfn; pfn++) {
780                         struct page *cursor_page;
781
782                         /* The target page is in the block, ignore it. */
783                         if (unlikely(pfn == page_pfn))
784                                 continue;
785
786                         /* Avoid holes within the zone. */
787                         if (unlikely(!pfn_valid_within(pfn)))
788                                 break;
789
790                         cursor_page = pfn_to_page(pfn);
791                         /* Check that we have not crossed a zone boundary. */
792                         if (unlikely(page_zone_id(cursor_page) != zone_id))
793                                 continue;
794                         switch (__isolate_lru_page(cursor_page, mode)) {
795                         case 0:
796                                 list_move(&cursor_page->lru, dst);
797                                 nr_taken++;
798                                 scan++;
799                                 break;
800
801                         case -EBUSY:
802                                 /* else it is being freed elsewhere */
803                                 list_move(&cursor_page->lru, src);
804                         default:
805                                 break;
806                         }
807                 }
808         }
809
810         *scanned = scan;
811         return nr_taken;
812 }
813
814 static unsigned long isolate_pages_global(unsigned long nr,
815                                         struct list_head *dst,
816                                         unsigned long *scanned, int order,
817                                         int mode, struct zone *z,
818                                         struct mem_cgroup *mem_cont,
819                                         int active)
820 {
821         if (active)
822                 return isolate_lru_pages(nr, &z->lru[LRU_ACTIVE].list, dst,
823                                                 scanned, order, mode);
824         else
825                 return isolate_lru_pages(nr, &z->lru[LRU_INACTIVE].list, dst,
826                                                 scanned, order, mode);
827 }
828
829 /*
830  * clear_active_flags() is a helper for shrink_active_list(), clearing
831  * any active bits from the pages in the list.
832  */
833 static unsigned long clear_active_flags(struct list_head *page_list)
834 {
835         int nr_active = 0;
836         struct page *page;
837
838         list_for_each_entry(page, page_list, lru)
839                 if (PageActive(page)) {
840                         ClearPageActive(page);
841                         nr_active++;
842                 }
843
844         return nr_active;
845 }
846
847 /**
848  * isolate_lru_page - tries to isolate a page from its LRU list
849  * @page: page to isolate from its LRU list
850  *
851  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
852  * vmstat statistic corresponding to whatever LRU list the page was on.
853  *
854  * Returns 0 if the page was removed from an LRU list.
855  * Returns -EBUSY if the page was not on an LRU list.
856  *
857  * The returned page will have PageLRU() cleared.  If it was found on
858  * the active list, it will have PageActive set.  That flag may need
859  * to be cleared by the caller before letting the page go.
860  *
861  * The vmstat statistic corresponding to the list on which the page was
862  * found will be decremented.
863  *
864  * Restrictions:
865  * (1) Must be called with an elevated refcount on the page. This is a
866  *     fundamentnal difference from isolate_lru_pages (which is called
867  *     without a stable reference).
868  * (2) the lru_lock must not be held.
869  * (3) interrupts must be enabled.
870  */
871 int isolate_lru_page(struct page *page)
872 {
873         int ret = -EBUSY;
874
875         if (PageLRU(page)) {
876                 struct zone *zone = page_zone(page);
877
878                 spin_lock_irq(&zone->lru_lock);
879                 if (PageLRU(page) && get_page_unless_zero(page)) {
880                         ret = 0;
881                         ClearPageLRU(page);
882                         if (PageActive(page))
883                                 del_page_from_active_list(zone, page);
884                         else
885                                 del_page_from_inactive_list(zone, page);
886                 }
887                 spin_unlock_irq(&zone->lru_lock);
888         }
889         return ret;
890 }
891
892 /*
893  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
894  * of reclaimed pages
895  */
896 static unsigned long shrink_inactive_list(unsigned long max_scan,
897                                 struct zone *zone, struct scan_control *sc)
898 {
899         LIST_HEAD(page_list);
900         struct pagevec pvec;
901         unsigned long nr_scanned = 0;
902         unsigned long nr_reclaimed = 0;
903
904         pagevec_init(&pvec, 1);
905
906         lru_add_drain();
907         spin_lock_irq(&zone->lru_lock);
908         do {
909                 struct page *page;
910                 unsigned long nr_taken;
911                 unsigned long nr_scan;
912                 unsigned long nr_freed;
913                 unsigned long nr_active;
914
915                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
916                              &page_list, &nr_scan, sc->order,
917                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
918                                              ISOLATE_BOTH : ISOLATE_INACTIVE,
919                                 zone, sc->mem_cgroup, 0);
920                 nr_active = clear_active_flags(&page_list);
921                 __count_vm_events(PGDEACTIVATE, nr_active);
922
923                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
924                 __mod_zone_page_state(zone, NR_INACTIVE,
925                                                 -(nr_taken - nr_active));
926                 if (scan_global_lru(sc))
927                         zone->pages_scanned += nr_scan;
928                 spin_unlock_irq(&zone->lru_lock);
929
930                 nr_scanned += nr_scan;
931                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
932
933                 /*
934                  * If we are direct reclaiming for contiguous pages and we do
935                  * not reclaim everything in the list, try again and wait
936                  * for IO to complete. This will stall high-order allocations
937                  * but that should be acceptable to the caller
938                  */
939                 if (nr_freed < nr_taken && !current_is_kswapd() &&
940                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
941                         congestion_wait(WRITE, HZ/10);
942
943                         /*
944                          * The attempt at page out may have made some
945                          * of the pages active, mark them inactive again.
946                          */
947                         nr_active = clear_active_flags(&page_list);
948                         count_vm_events(PGDEACTIVATE, nr_active);
949
950                         nr_freed += shrink_page_list(&page_list, sc,
951                                                         PAGEOUT_IO_SYNC);
952                 }
953
954                 nr_reclaimed += nr_freed;
955                 local_irq_disable();
956                 if (current_is_kswapd()) {
957                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
958                         __count_vm_events(KSWAPD_STEAL, nr_freed);
959                 } else if (scan_global_lru(sc))
960                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
961
962                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
963
964                 if (nr_taken == 0)
965                         goto done;
966
967                 spin_lock(&zone->lru_lock);
968                 /*
969                  * Put back any unfreeable pages.
970                  */
971                 while (!list_empty(&page_list)) {
972                         page = lru_to_page(&page_list);
973                         VM_BUG_ON(PageLRU(page));
974                         SetPageLRU(page);
975                         list_del(&page->lru);
976                         add_page_to_lru_list(zone, page, page_lru(page));
977                         if (!pagevec_add(&pvec, page)) {
978                                 spin_unlock_irq(&zone->lru_lock);
979                                 __pagevec_release(&pvec);
980                                 spin_lock_irq(&zone->lru_lock);
981                         }
982                 }
983         } while (nr_scanned < max_scan);
984         spin_unlock(&zone->lru_lock);
985 done:
986         local_irq_enable();
987         pagevec_release(&pvec);
988         return nr_reclaimed;
989 }
990
991 /*
992  * We are about to scan this zone at a certain priority level.  If that priority
993  * level is smaller (ie: more urgent) than the previous priority, then note
994  * that priority level within the zone.  This is done so that when the next
995  * process comes in to scan this zone, it will immediately start out at this
996  * priority level rather than having to build up its own scanning priority.
997  * Here, this priority affects only the reclaim-mapped threshold.
998  */
999 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1000 {
1001         if (priority < zone->prev_priority)
1002                 zone->prev_priority = priority;
1003 }
1004
1005 static inline int zone_is_near_oom(struct zone *zone)
1006 {
1007         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
1008                                 + zone_page_state(zone, NR_INACTIVE))*3;
1009 }
1010
1011 /*
1012  * Determine we should try to reclaim mapped pages.
1013  * This is called only when sc->mem_cgroup is NULL.
1014  */
1015 static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
1016                                 int priority)
1017 {
1018         long mapped_ratio;
1019         long distress;
1020         long swap_tendency;
1021         long imbalance;
1022         int reclaim_mapped = 0;
1023         int prev_priority;
1024
1025         if (scan_global_lru(sc) && zone_is_near_oom(zone))
1026                 return 1;
1027         /*
1028          * `distress' is a measure of how much trouble we're having
1029          * reclaiming pages.  0 -> no problems.  100 -> great trouble.
1030          */
1031         if (scan_global_lru(sc))
1032                 prev_priority = zone->prev_priority;
1033         else
1034                 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
1035
1036         distress = 100 >> min(prev_priority, priority);
1037
1038         /*
1039          * The point of this algorithm is to decide when to start
1040          * reclaiming mapped memory instead of just pagecache.  Work out
1041          * how much memory
1042          * is mapped.
1043          */
1044         if (scan_global_lru(sc))
1045                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
1046                                 global_page_state(NR_ANON_PAGES)) * 100) /
1047                                         vm_total_pages;
1048         else
1049                 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
1050
1051         /*
1052          * Now decide how much we really want to unmap some pages.  The
1053          * mapped ratio is downgraded - just because there's a lot of
1054          * mapped memory doesn't necessarily mean that page reclaim
1055          * isn't succeeding.
1056          *
1057          * The distress ratio is important - we don't want to start
1058          * going oom.
1059          *
1060          * A 100% value of vm_swappiness overrides this algorithm
1061          * altogether.
1062          */
1063         swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1064
1065         /*
1066          * If there's huge imbalance between active and inactive
1067          * (think active 100 times larger than inactive) we should
1068          * become more permissive, or the system will take too much
1069          * cpu before it start swapping during memory pressure.
1070          * Distress is about avoiding early-oom, this is about
1071          * making swappiness graceful despite setting it to low
1072          * values.
1073          *
1074          * Avoid div by zero with nr_inactive+1, and max resulting
1075          * value is vm_total_pages.
1076          */
1077         if (scan_global_lru(sc)) {
1078                 imbalance  = zone_page_state(zone, NR_ACTIVE);
1079                 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1080         } else
1081                 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1082
1083         /*
1084          * Reduce the effect of imbalance if swappiness is low,
1085          * this means for a swappiness very low, the imbalance
1086          * must be much higher than 100 for this logic to make
1087          * the difference.
1088          *
1089          * Max temporary value is vm_total_pages*100.
1090          */
1091         imbalance *= (vm_swappiness + 1);
1092         imbalance /= 100;
1093
1094         /*
1095          * If not much of the ram is mapped, makes the imbalance
1096          * less relevant, it's high priority we refill the inactive
1097          * list with mapped pages only in presence of high ratio of
1098          * mapped pages.
1099          *
1100          * Max temporary value is vm_total_pages*100.
1101          */
1102         imbalance *= mapped_ratio;
1103         imbalance /= 100;
1104
1105         /* apply imbalance feedback to swap_tendency */
1106         swap_tendency += imbalance;
1107
1108         /*
1109          * Now use this metric to decide whether to start moving mapped
1110          * memory onto the inactive list.
1111          */
1112         if (swap_tendency >= 100)
1113                 reclaim_mapped = 1;
1114
1115         return reclaim_mapped;
1116 }
1117
1118 /*
1119  * This moves pages from the active list to the inactive list.
1120  *
1121  * We move them the other way if the page is referenced by one or more
1122  * processes, from rmap.
1123  *
1124  * If the pages are mostly unmapped, the processing is fast and it is
1125  * appropriate to hold zone->lru_lock across the whole operation.  But if
1126  * the pages are mapped, the processing is slow (page_referenced()) so we
1127  * should drop zone->lru_lock around each page.  It's impossible to balance
1128  * this, so instead we remove the pages from the LRU while processing them.
1129  * It is safe to rely on PG_active against the non-LRU pages in here because
1130  * nobody will play with that bit on a non-LRU page.
1131  *
1132  * The downside is that we have to touch page->_count against each page.
1133  * But we had to alter page->flags anyway.
1134  */
1135
1136
1137 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1138                                 struct scan_control *sc, int priority)
1139 {
1140         unsigned long pgmoved;
1141         int pgdeactivate = 0;
1142         unsigned long pgscanned;
1143         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1144         LIST_HEAD(l_active);
1145         LIST_HEAD(l_inactive);
1146         struct page *page;
1147         struct pagevec pvec;
1148         int reclaim_mapped = 0;
1149
1150         if (sc->may_swap)
1151                 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
1152
1153         lru_add_drain();
1154         spin_lock_irq(&zone->lru_lock);
1155         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1156                                         ISOLATE_ACTIVE, zone,
1157                                         sc->mem_cgroup, 1);
1158         /*
1159          * zone->pages_scanned is used for detect zone's oom
1160          * mem_cgroup remembers nr_scan by itself.
1161          */
1162         if (scan_global_lru(sc))
1163                 zone->pages_scanned += pgscanned;
1164
1165         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1166         spin_unlock_irq(&zone->lru_lock);
1167
1168         while (!list_empty(&l_hold)) {
1169                 cond_resched();
1170                 page = lru_to_page(&l_hold);
1171                 list_del(&page->lru);
1172                 if (page_mapped(page)) {
1173                         if (!reclaim_mapped ||
1174                             (total_swap_pages == 0 && PageAnon(page)) ||
1175                             page_referenced(page, 0, sc->mem_cgroup)) {
1176                                 list_add(&page->lru, &l_active);
1177                                 continue;
1178                         }
1179                 }
1180                 list_add(&page->lru, &l_inactive);
1181         }
1182
1183         pagevec_init(&pvec, 1);
1184         pgmoved = 0;
1185         spin_lock_irq(&zone->lru_lock);
1186         while (!list_empty(&l_inactive)) {
1187                 page = lru_to_page(&l_inactive);
1188                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1189                 VM_BUG_ON(PageLRU(page));
1190                 SetPageLRU(page);
1191                 VM_BUG_ON(!PageActive(page));
1192                 ClearPageActive(page);
1193
1194                 list_move(&page->lru, &zone->lru[LRU_INACTIVE].list);
1195                 mem_cgroup_move_lists(page, false);
1196                 pgmoved++;
1197                 if (!pagevec_add(&pvec, page)) {
1198                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1199                         spin_unlock_irq(&zone->lru_lock);
1200                         pgdeactivate += pgmoved;
1201                         pgmoved = 0;
1202                         if (buffer_heads_over_limit)
1203                                 pagevec_strip(&pvec);
1204                         __pagevec_release(&pvec);
1205                         spin_lock_irq(&zone->lru_lock);
1206                 }
1207         }
1208         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1209         pgdeactivate += pgmoved;
1210         if (buffer_heads_over_limit) {
1211                 spin_unlock_irq(&zone->lru_lock);
1212                 pagevec_strip(&pvec);
1213                 spin_lock_irq(&zone->lru_lock);
1214         }
1215
1216         pgmoved = 0;
1217         while (!list_empty(&l_active)) {
1218                 page = lru_to_page(&l_active);
1219                 prefetchw_prev_lru_page(page, &l_active, flags);
1220                 VM_BUG_ON(PageLRU(page));
1221                 SetPageLRU(page);
1222                 VM_BUG_ON(!PageActive(page));
1223
1224                 list_move(&page->lru, &zone->lru[LRU_ACTIVE].list);
1225                 mem_cgroup_move_lists(page, true);
1226                 pgmoved++;
1227                 if (!pagevec_add(&pvec, page)) {
1228                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1229                         pgmoved = 0;
1230                         spin_unlock_irq(&zone->lru_lock);
1231                         __pagevec_release(&pvec);
1232                         spin_lock_irq(&zone->lru_lock);
1233                 }
1234         }
1235         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1236
1237         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1238         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1239         spin_unlock_irq(&zone->lru_lock);
1240
1241         pagevec_release(&pvec);
1242 }
1243
1244 static unsigned long shrink_list(enum lru_list l, unsigned long nr_to_scan,
1245         struct zone *zone, struct scan_control *sc, int priority)
1246 {
1247         if (l == LRU_ACTIVE) {
1248                 shrink_active_list(nr_to_scan, zone, sc, priority);
1249                 return 0;
1250         }
1251         return shrink_inactive_list(nr_to_scan, zone, sc);
1252 }
1253
1254 /*
1255  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1256  */
1257 static unsigned long shrink_zone(int priority, struct zone *zone,
1258                                 struct scan_control *sc)
1259 {
1260         unsigned long nr[NR_LRU_LISTS];
1261         unsigned long nr_to_scan;
1262         unsigned long nr_reclaimed = 0;
1263         enum lru_list l;
1264
1265         if (scan_global_lru(sc)) {
1266                 /*
1267                  * Add one to nr_to_scan just to make sure that the kernel
1268                  * will slowly sift through the active list.
1269                  */
1270                 for_each_lru(l) {
1271                         zone->lru[l].nr_scan += (zone_page_state(zone,
1272                                         NR_LRU_BASE + l)  >> priority) + 1;
1273                         nr[l] = zone->lru[l].nr_scan;
1274                         if (nr[l] >= sc->swap_cluster_max)
1275                                 zone->lru[l].nr_scan = 0;
1276                         else
1277                                 nr[l] = 0;
1278                 }
1279         } else {
1280                 /*
1281                  * This reclaim occurs not because zone memory shortage but
1282                  * because memory controller hits its limit.
1283                  * Then, don't modify zone reclaim related data.
1284                  */
1285                 nr[LRU_ACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1286                                         zone, priority, LRU_ACTIVE);
1287
1288                 nr[LRU_INACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1289                                         zone, priority, LRU_INACTIVE);
1290         }
1291
1292         while (nr[LRU_ACTIVE] || nr[LRU_INACTIVE]) {
1293                 for_each_lru(l) {
1294                         if (nr[l]) {
1295                                 nr_to_scan = min(nr[l],
1296                                         (unsigned long)sc->swap_cluster_max);
1297                                 nr[l] -= nr_to_scan;
1298
1299                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1300                                                         zone, sc, priority);
1301                         }
1302                 }
1303         }
1304
1305         throttle_vm_writeout(sc->gfp_mask);
1306         return nr_reclaimed;
1307 }
1308
1309 /*
1310  * This is the direct reclaim path, for page-allocating processes.  We only
1311  * try to reclaim pages from zones which will satisfy the caller's allocation
1312  * request.
1313  *
1314  * We reclaim from a zone even if that zone is over pages_high.  Because:
1315  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1316  *    allocation or
1317  * b) The zones may be over pages_high but they must go *over* pages_high to
1318  *    satisfy the `incremental min' zone defense algorithm.
1319  *
1320  * Returns the number of reclaimed pages.
1321  *
1322  * If a zone is deemed to be full of pinned pages then just give it a light
1323  * scan then give up on it.
1324  */
1325 static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
1326                                         struct scan_control *sc)
1327 {
1328         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1329         unsigned long nr_reclaimed = 0;
1330         struct zoneref *z;
1331         struct zone *zone;
1332
1333         sc->all_unreclaimable = 1;
1334         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1335                 if (!populated_zone(zone))
1336                         continue;
1337                 /*
1338                  * Take care memory controller reclaiming has small influence
1339                  * to global LRU.
1340                  */
1341                 if (scan_global_lru(sc)) {
1342                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1343                                 continue;
1344                         note_zone_scanning_priority(zone, priority);
1345
1346                         if (zone_is_all_unreclaimable(zone) &&
1347                                                 priority != DEF_PRIORITY)
1348                                 continue;       /* Let kswapd poll it */
1349                         sc->all_unreclaimable = 0;
1350                 } else {
1351                         /*
1352                          * Ignore cpuset limitation here. We just want to reduce
1353                          * # of used pages by us regardless of memory shortage.
1354                          */
1355                         sc->all_unreclaimable = 0;
1356                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1357                                                         priority);
1358                 }
1359
1360                 nr_reclaimed += shrink_zone(priority, zone, sc);
1361         }
1362
1363         return nr_reclaimed;
1364 }
1365  
1366 /*
1367  * This is the main entry point to direct page reclaim.
1368  *
1369  * If a full scan of the inactive list fails to free enough memory then we
1370  * are "out of memory" and something needs to be killed.
1371  *
1372  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1373  * high - the zone may be full of dirty or under-writeback pages, which this
1374  * caller can't do much about.  We kick pdflush and take explicit naps in the
1375  * hope that some of these pages can be written.  But if the allocating task
1376  * holds filesystem locks which prevent writeout this might not work, and the
1377  * allocation attempt will fail.
1378  *
1379  * returns:     0, if no pages reclaimed
1380  *              else, the number of pages reclaimed
1381  */
1382 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1383                                         struct scan_control *sc)
1384 {
1385         int priority;
1386         unsigned long ret = 0;
1387         unsigned long total_scanned = 0;
1388         unsigned long nr_reclaimed = 0;
1389         struct reclaim_state *reclaim_state = current->reclaim_state;
1390         unsigned long lru_pages = 0;
1391         struct zoneref *z;
1392         struct zone *zone;
1393         enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1394
1395         delayacct_freepages_start();
1396
1397         if (scan_global_lru(sc))
1398                 count_vm_event(ALLOCSTALL);
1399         /*
1400          * mem_cgroup will not do shrink_slab.
1401          */
1402         if (scan_global_lru(sc)) {
1403                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1404
1405                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1406                                 continue;
1407
1408                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1409                                         + zone_page_state(zone, NR_INACTIVE);
1410                 }
1411         }
1412
1413         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1414                 sc->nr_scanned = 0;
1415                 if (!priority)
1416                         disable_swap_token();
1417                 nr_reclaimed += shrink_zones(priority, zonelist, sc);
1418                 /*
1419                  * Don't shrink slabs when reclaiming memory from
1420                  * over limit cgroups
1421                  */
1422                 if (scan_global_lru(sc)) {
1423                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1424                         if (reclaim_state) {
1425                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1426                                 reclaim_state->reclaimed_slab = 0;
1427                         }
1428                 }
1429                 total_scanned += sc->nr_scanned;
1430                 if (nr_reclaimed >= sc->swap_cluster_max) {
1431                         ret = nr_reclaimed;
1432                         goto out;
1433                 }
1434
1435                 /*
1436                  * Try to write back as many pages as we just scanned.  This
1437                  * tends to cause slow streaming writers to write data to the
1438                  * disk smoothly, at the dirtying rate, which is nice.   But
1439                  * that's undesirable in laptop mode, where we *want* lumpy
1440                  * writeout.  So in laptop mode, write out the whole world.
1441                  */
1442                 if (total_scanned > sc->swap_cluster_max +
1443                                         sc->swap_cluster_max / 2) {
1444                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1445                         sc->may_writepage = 1;
1446                 }
1447
1448                 /* Take a nap, wait for some writeback to complete */
1449                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1450                         congestion_wait(WRITE, HZ/10);
1451         }
1452         /* top priority shrink_zones still had more to do? don't OOM, then */
1453         if (!sc->all_unreclaimable && scan_global_lru(sc))
1454                 ret = nr_reclaimed;
1455 out:
1456         /*
1457          * Now that we've scanned all the zones at this priority level, note
1458          * that level within the zone so that the next thread which performs
1459          * scanning of this zone will immediately start out at this priority
1460          * level.  This affects only the decision whether or not to bring
1461          * mapped pages onto the inactive list.
1462          */
1463         if (priority < 0)
1464                 priority = 0;
1465
1466         if (scan_global_lru(sc)) {
1467                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1468
1469                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1470                                 continue;
1471
1472                         zone->prev_priority = priority;
1473                 }
1474         } else
1475                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1476
1477         delayacct_freepages_end();
1478
1479         return ret;
1480 }
1481
1482 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1483                                                                 gfp_t gfp_mask)
1484 {
1485         struct scan_control sc = {
1486                 .gfp_mask = gfp_mask,
1487                 .may_writepage = !laptop_mode,
1488                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1489                 .may_swap = 1,
1490                 .swappiness = vm_swappiness,
1491                 .order = order,
1492                 .mem_cgroup = NULL,
1493                 .isolate_pages = isolate_pages_global,
1494         };
1495
1496         return do_try_to_free_pages(zonelist, &sc);
1497 }
1498
1499 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1500
1501 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1502                                                 gfp_t gfp_mask)
1503 {
1504         struct scan_control sc = {
1505                 .may_writepage = !laptop_mode,
1506                 .may_swap = 1,
1507                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1508                 .swappiness = vm_swappiness,
1509                 .order = 0,
1510                 .mem_cgroup = mem_cont,
1511                 .isolate_pages = mem_cgroup_isolate_pages,
1512         };
1513         struct zonelist *zonelist;
1514
1515         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1516                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1517         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1518         return do_try_to_free_pages(zonelist, &sc);
1519 }
1520 #endif
1521
1522 /*
1523  * For kswapd, balance_pgdat() will work across all this node's zones until
1524  * they are all at pages_high.
1525  *
1526  * Returns the number of pages which were actually freed.
1527  *
1528  * There is special handling here for zones which are full of pinned pages.
1529  * This can happen if the pages are all mlocked, or if they are all used by
1530  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1531  * What we do is to detect the case where all pages in the zone have been
1532  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1533  * dead and from now on, only perform a short scan.  Basically we're polling
1534  * the zone for when the problem goes away.
1535  *
1536  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1537  * zones which have free_pages > pages_high, but once a zone is found to have
1538  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1539  * of the number of free pages in the lower zones.  This interoperates with
1540  * the page allocator fallback scheme to ensure that aging of pages is balanced
1541  * across the zones.
1542  */
1543 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1544 {
1545         int all_zones_ok;
1546         int priority;
1547         int i;
1548         unsigned long total_scanned;
1549         unsigned long nr_reclaimed;
1550         struct reclaim_state *reclaim_state = current->reclaim_state;
1551         struct scan_control sc = {
1552                 .gfp_mask = GFP_KERNEL,
1553                 .may_swap = 1,
1554                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1555                 .swappiness = vm_swappiness,
1556                 .order = order,
1557                 .mem_cgroup = NULL,
1558                 .isolate_pages = isolate_pages_global,
1559         };
1560         /*
1561          * temp_priority is used to remember the scanning priority at which
1562          * this zone was successfully refilled to free_pages == pages_high.
1563          */
1564         int temp_priority[MAX_NR_ZONES];
1565
1566 loop_again:
1567         total_scanned = 0;
1568         nr_reclaimed = 0;
1569         sc.may_writepage = !laptop_mode;
1570         count_vm_event(PAGEOUTRUN);
1571
1572         for (i = 0; i < pgdat->nr_zones; i++)
1573                 temp_priority[i] = DEF_PRIORITY;
1574
1575         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1576                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1577                 unsigned long lru_pages = 0;
1578
1579                 /* The swap token gets in the way of swapout... */
1580                 if (!priority)
1581                         disable_swap_token();
1582
1583                 all_zones_ok = 1;
1584
1585                 /*
1586                  * Scan in the highmem->dma direction for the highest
1587                  * zone which needs scanning
1588                  */
1589                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1590                         struct zone *zone = pgdat->node_zones + i;
1591
1592                         if (!populated_zone(zone))
1593                                 continue;
1594
1595                         if (zone_is_all_unreclaimable(zone) &&
1596                             priority != DEF_PRIORITY)
1597                                 continue;
1598
1599                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1600                                                0, 0)) {
1601                                 end_zone = i;
1602                                 break;
1603                         }
1604                 }
1605                 if (i < 0)
1606                         goto out;
1607
1608                 for (i = 0; i <= end_zone; i++) {
1609                         struct zone *zone = pgdat->node_zones + i;
1610
1611                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1612                                         + zone_page_state(zone, NR_INACTIVE);
1613                 }
1614
1615                 /*
1616                  * Now scan the zone in the dma->highmem direction, stopping
1617                  * at the last zone which needs scanning.
1618                  *
1619                  * We do this because the page allocator works in the opposite
1620                  * direction.  This prevents the page allocator from allocating
1621                  * pages behind kswapd's direction of progress, which would
1622                  * cause too much scanning of the lower zones.
1623                  */
1624                 for (i = 0; i <= end_zone; i++) {
1625                         struct zone *zone = pgdat->node_zones + i;
1626                         int nr_slab;
1627
1628                         if (!populated_zone(zone))
1629                                 continue;
1630
1631                         if (zone_is_all_unreclaimable(zone) &&
1632                                         priority != DEF_PRIORITY)
1633                                 continue;
1634
1635                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1636                                                end_zone, 0))
1637                                 all_zones_ok = 0;
1638                         temp_priority[i] = priority;
1639                         sc.nr_scanned = 0;
1640                         note_zone_scanning_priority(zone, priority);
1641                         /*
1642                          * We put equal pressure on every zone, unless one
1643                          * zone has way too many pages free already.
1644                          */
1645                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1646                                                 end_zone, 0))
1647                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1648                         reclaim_state->reclaimed_slab = 0;
1649                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1650                                                 lru_pages);
1651                         nr_reclaimed += reclaim_state->reclaimed_slab;
1652                         total_scanned += sc.nr_scanned;
1653                         if (zone_is_all_unreclaimable(zone))
1654                                 continue;
1655                         if (nr_slab == 0 && zone->pages_scanned >=
1656                                 (zone_page_state(zone, NR_ACTIVE)
1657                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1658                                         zone_set_flag(zone,
1659                                                       ZONE_ALL_UNRECLAIMABLE);
1660                         /*
1661                          * If we've done a decent amount of scanning and
1662                          * the reclaim ratio is low, start doing writepage
1663                          * even in laptop mode
1664                          */
1665                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1666                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1667                                 sc.may_writepage = 1;
1668                 }
1669                 if (all_zones_ok)
1670                         break;          /* kswapd: all done */
1671                 /*
1672                  * OK, kswapd is getting into trouble.  Take a nap, then take
1673                  * another pass across the zones.
1674                  */
1675                 if (total_scanned && priority < DEF_PRIORITY - 2)
1676                         congestion_wait(WRITE, HZ/10);
1677
1678                 /*
1679                  * We do this so kswapd doesn't build up large priorities for
1680                  * example when it is freeing in parallel with allocators. It
1681                  * matches the direct reclaim path behaviour in terms of impact
1682                  * on zone->*_priority.
1683                  */
1684                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1685                         break;
1686         }
1687 out:
1688         /*
1689          * Note within each zone the priority level at which this zone was
1690          * brought into a happy state.  So that the next thread which scans this
1691          * zone will start out at that priority level.
1692          */
1693         for (i = 0; i < pgdat->nr_zones; i++) {
1694                 struct zone *zone = pgdat->node_zones + i;
1695
1696                 zone->prev_priority = temp_priority[i];
1697         }
1698         if (!all_zones_ok) {
1699                 cond_resched();
1700
1701                 try_to_freeze();
1702
1703                 goto loop_again;
1704         }
1705
1706         return nr_reclaimed;
1707 }
1708
1709 /*
1710  * The background pageout daemon, started as a kernel thread
1711  * from the init process. 
1712  *
1713  * This basically trickles out pages so that we have _some_
1714  * free memory available even if there is no other activity
1715  * that frees anything up. This is needed for things like routing
1716  * etc, where we otherwise might have all activity going on in
1717  * asynchronous contexts that cannot page things out.
1718  *
1719  * If there are applications that are active memory-allocators
1720  * (most normal use), this basically shouldn't matter.
1721  */
1722 static int kswapd(void *p)
1723 {
1724         unsigned long order;
1725         pg_data_t *pgdat = (pg_data_t*)p;
1726         struct task_struct *tsk = current;
1727         DEFINE_WAIT(wait);
1728         struct reclaim_state reclaim_state = {
1729                 .reclaimed_slab = 0,
1730         };
1731         node_to_cpumask_ptr(cpumask, pgdat->node_id);
1732
1733         if (!cpus_empty(*cpumask))
1734                 set_cpus_allowed_ptr(tsk, cpumask);
1735         current->reclaim_state = &reclaim_state;
1736
1737         /*
1738          * Tell the memory management that we're a "memory allocator",
1739          * and that if we need more memory we should get access to it
1740          * regardless (see "__alloc_pages()"). "kswapd" should
1741          * never get caught in the normal page freeing logic.
1742          *
1743          * (Kswapd normally doesn't need memory anyway, but sometimes
1744          * you need a small amount of memory in order to be able to
1745          * page out something else, and this flag essentially protects
1746          * us from recursively trying to free more memory as we're
1747          * trying to free the first piece of memory in the first place).
1748          */
1749         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1750         set_freezable();
1751
1752         order = 0;
1753         for ( ; ; ) {
1754                 unsigned long new_order;
1755
1756                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1757                 new_order = pgdat->kswapd_max_order;
1758                 pgdat->kswapd_max_order = 0;
1759                 if (order < new_order) {
1760                         /*
1761                          * Don't sleep if someone wants a larger 'order'
1762                          * allocation
1763                          */
1764                         order = new_order;
1765                 } else {
1766                         if (!freezing(current))
1767                                 schedule();
1768
1769                         order = pgdat->kswapd_max_order;
1770                 }
1771                 finish_wait(&pgdat->kswapd_wait, &wait);
1772
1773                 if (!try_to_freeze()) {
1774                         /* We can speed up thawing tasks if we don't call
1775                          * balance_pgdat after returning from the refrigerator
1776                          */
1777                         balance_pgdat(pgdat, order);
1778                 }
1779         }
1780         return 0;
1781 }
1782
1783 /*
1784  * A zone is low on free memory, so wake its kswapd task to service it.
1785  */
1786 void wakeup_kswapd(struct zone *zone, int order)
1787 {
1788         pg_data_t *pgdat;
1789
1790         if (!populated_zone(zone))
1791                 return;
1792
1793         pgdat = zone->zone_pgdat;
1794         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1795                 return;
1796         if (pgdat->kswapd_max_order < order)
1797                 pgdat->kswapd_max_order = order;
1798         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1799                 return;
1800         if (!waitqueue_active(&pgdat->kswapd_wait))
1801                 return;
1802         wake_up_interruptible(&pgdat->kswapd_wait);
1803 }
1804
1805 #ifdef CONFIG_PM
1806 /*
1807  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1808  * from LRU lists system-wide, for given pass and priority, and returns the
1809  * number of reclaimed pages
1810  *
1811  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1812  */
1813 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1814                                       int pass, struct scan_control *sc)
1815 {
1816         struct zone *zone;
1817         unsigned long nr_to_scan, ret = 0;
1818         enum lru_list l;
1819
1820         for_each_zone(zone) {
1821
1822                 if (!populated_zone(zone))
1823                         continue;
1824
1825                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1826                         continue;
1827
1828                 for_each_lru(l) {
1829                         /* For pass = 0 we don't shrink the active list */
1830                         if (pass == 0 && l == LRU_ACTIVE)
1831                                 continue;
1832
1833                         zone->lru[l].nr_scan +=
1834                                 (zone_page_state(zone, NR_LRU_BASE + l)
1835                                                                 >> prio) + 1;
1836                         if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
1837                                 zone->lru[l].nr_scan = 0;
1838                                 nr_to_scan = min(nr_pages,
1839                                         zone_page_state(zone,
1840                                                         NR_LRU_BASE + l));
1841                                 ret += shrink_list(l, nr_to_scan, zone,
1842                                                                 sc, prio);
1843                                 if (ret >= nr_pages)
1844                                         return ret;
1845                         }
1846                 }
1847         }
1848
1849         return ret;
1850 }
1851
1852 static unsigned long count_lru_pages(void)
1853 {
1854         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1855 }
1856
1857 /*
1858  * Try to free `nr_pages' of memory, system-wide, and return the number of
1859  * freed pages.
1860  *
1861  * Rather than trying to age LRUs the aim is to preserve the overall
1862  * LRU order by reclaiming preferentially
1863  * inactive > active > active referenced > active mapped
1864  */
1865 unsigned long shrink_all_memory(unsigned long nr_pages)
1866 {
1867         unsigned long lru_pages, nr_slab;
1868         unsigned long ret = 0;
1869         int pass;
1870         struct reclaim_state reclaim_state;
1871         struct scan_control sc = {
1872                 .gfp_mask = GFP_KERNEL,
1873                 .may_swap = 0,
1874                 .swap_cluster_max = nr_pages,
1875                 .may_writepage = 1,
1876                 .swappiness = vm_swappiness,
1877                 .isolate_pages = isolate_pages_global,
1878         };
1879
1880         current->reclaim_state = &reclaim_state;
1881
1882         lru_pages = count_lru_pages();
1883         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1884         /* If slab caches are huge, it's better to hit them first */
1885         while (nr_slab >= lru_pages) {
1886                 reclaim_state.reclaimed_slab = 0;
1887                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1888                 if (!reclaim_state.reclaimed_slab)
1889                         break;
1890
1891                 ret += reclaim_state.reclaimed_slab;
1892                 if (ret >= nr_pages)
1893                         goto out;
1894
1895                 nr_slab -= reclaim_state.reclaimed_slab;
1896         }
1897
1898         /*
1899          * We try to shrink LRUs in 5 passes:
1900          * 0 = Reclaim from inactive_list only
1901          * 1 = Reclaim from active list but don't reclaim mapped
1902          * 2 = 2nd pass of type 1
1903          * 3 = Reclaim mapped (normal reclaim)
1904          * 4 = 2nd pass of type 3
1905          */
1906         for (pass = 0; pass < 5; pass++) {
1907                 int prio;
1908
1909                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1910                 if (pass > 2) {
1911                         sc.may_swap = 1;
1912                         sc.swappiness = 100;
1913                 }
1914
1915                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1916                         unsigned long nr_to_scan = nr_pages - ret;
1917
1918                         sc.nr_scanned = 0;
1919                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1920                         if (ret >= nr_pages)
1921                                 goto out;
1922
1923                         reclaim_state.reclaimed_slab = 0;
1924                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1925                                         count_lru_pages());
1926                         ret += reclaim_state.reclaimed_slab;
1927                         if (ret >= nr_pages)
1928                                 goto out;
1929
1930                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1931                                 congestion_wait(WRITE, HZ / 10);
1932                 }
1933         }
1934
1935         /*
1936          * If ret = 0, we could not shrink LRUs, but there may be something
1937          * in slab caches
1938          */
1939         if (!ret) {
1940                 do {
1941                         reclaim_state.reclaimed_slab = 0;
1942                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1943                         ret += reclaim_state.reclaimed_slab;
1944                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1945         }
1946
1947 out:
1948         current->reclaim_state = NULL;
1949
1950         return ret;
1951 }
1952 #endif
1953
1954 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1955    not required for correctness.  So if the last cpu in a node goes
1956    away, we get changed to run anywhere: as the first one comes back,
1957    restore their cpu bindings. */
1958 static int __devinit cpu_callback(struct notifier_block *nfb,
1959                                   unsigned long action, void *hcpu)
1960 {
1961         int nid;
1962
1963         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1964                 for_each_node_state(nid, N_HIGH_MEMORY) {
1965                         pg_data_t *pgdat = NODE_DATA(nid);
1966                         node_to_cpumask_ptr(mask, pgdat->node_id);
1967
1968                         if (any_online_cpu(*mask) < nr_cpu_ids)
1969                                 /* One of our CPUs online: restore mask */
1970                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1971                 }
1972         }
1973         return NOTIFY_OK;
1974 }
1975
1976 /*
1977  * This kswapd start function will be called by init and node-hot-add.
1978  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1979  */
1980 int kswapd_run(int nid)
1981 {
1982         pg_data_t *pgdat = NODE_DATA(nid);
1983         int ret = 0;
1984
1985         if (pgdat->kswapd)
1986                 return 0;
1987
1988         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1989         if (IS_ERR(pgdat->kswapd)) {
1990                 /* failure at boot is fatal */
1991                 BUG_ON(system_state == SYSTEM_BOOTING);
1992                 printk("Failed to start kswapd on node %d\n",nid);
1993                 ret = -1;
1994         }
1995         return ret;
1996 }
1997
1998 static int __init kswapd_init(void)
1999 {
2000         int nid;
2001
2002         swap_setup();
2003         for_each_node_state(nid, N_HIGH_MEMORY)
2004                 kswapd_run(nid);
2005         hotcpu_notifier(cpu_callback, 0);
2006         return 0;
2007 }
2008
2009 module_init(kswapd_init)
2010
2011 #ifdef CONFIG_NUMA
2012 /*
2013  * Zone reclaim mode
2014  *
2015  * If non-zero call zone_reclaim when the number of free pages falls below
2016  * the watermarks.
2017  */
2018 int zone_reclaim_mode __read_mostly;
2019
2020 #define RECLAIM_OFF 0
2021 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2022 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2023 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2024
2025 /*
2026  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2027  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2028  * a zone.
2029  */
2030 #define ZONE_RECLAIM_PRIORITY 4
2031
2032 /*
2033  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2034  * occur.
2035  */
2036 int sysctl_min_unmapped_ratio = 1;
2037
2038 /*
2039  * If the number of slab pages in a zone grows beyond this percentage then
2040  * slab reclaim needs to occur.
2041  */
2042 int sysctl_min_slab_ratio = 5;
2043
2044 /*
2045  * Try to free up some pages from this zone through reclaim.
2046  */
2047 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2048 {
2049         /* Minimum pages needed in order to stay on node */
2050         const unsigned long nr_pages = 1 << order;
2051         struct task_struct *p = current;
2052         struct reclaim_state reclaim_state;
2053         int priority;
2054         unsigned long nr_reclaimed = 0;
2055         struct scan_control sc = {
2056                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2057                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2058                 .swap_cluster_max = max_t(unsigned long, nr_pages,
2059                                         SWAP_CLUSTER_MAX),
2060                 .gfp_mask = gfp_mask,
2061                 .swappiness = vm_swappiness,
2062                 .isolate_pages = isolate_pages_global,
2063         };
2064         unsigned long slab_reclaimable;
2065
2066         disable_swap_token();
2067         cond_resched();
2068         /*
2069          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2070          * and we also need to be able to write out pages for RECLAIM_WRITE
2071          * and RECLAIM_SWAP.
2072          */
2073         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2074         reclaim_state.reclaimed_slab = 0;
2075         p->reclaim_state = &reclaim_state;
2076
2077         if (zone_page_state(zone, NR_FILE_PAGES) -
2078                 zone_page_state(zone, NR_FILE_MAPPED) >
2079                 zone->min_unmapped_pages) {
2080                 /*
2081                  * Free memory by calling shrink zone with increasing
2082                  * priorities until we have enough memory freed.
2083                  */
2084                 priority = ZONE_RECLAIM_PRIORITY;
2085                 do {
2086                         note_zone_scanning_priority(zone, priority);
2087                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2088                         priority--;
2089                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2090         }
2091
2092         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2093         if (slab_reclaimable > zone->min_slab_pages) {
2094                 /*
2095                  * shrink_slab() does not currently allow us to determine how
2096                  * many pages were freed in this zone. So we take the current
2097                  * number of slab pages and shake the slab until it is reduced
2098                  * by the same nr_pages that we used for reclaiming unmapped
2099                  * pages.
2100                  *
2101                  * Note that shrink_slab will free memory on all zones and may
2102                  * take a long time.
2103                  */
2104                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2105                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2106                                 slab_reclaimable - nr_pages)
2107                         ;
2108
2109                 /*
2110                  * Update nr_reclaimed by the number of slab pages we
2111                  * reclaimed from this zone.
2112                  */
2113                 nr_reclaimed += slab_reclaimable -
2114                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2115         }
2116
2117         p->reclaim_state = NULL;
2118         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2119         return nr_reclaimed >= nr_pages;
2120 }
2121
2122 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2123 {
2124         int node_id;
2125         int ret;
2126
2127         /*
2128          * Zone reclaim reclaims unmapped file backed pages and
2129          * slab pages if we are over the defined limits.
2130          *
2131          * A small portion of unmapped file backed pages is needed for
2132          * file I/O otherwise pages read by file I/O will be immediately
2133          * thrown out if the zone is overallocated. So we do not reclaim
2134          * if less than a specified percentage of the zone is used by
2135          * unmapped file backed pages.
2136          */
2137         if (zone_page_state(zone, NR_FILE_PAGES) -
2138             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2139             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2140                         <= zone->min_slab_pages)
2141                 return 0;
2142
2143         if (zone_is_all_unreclaimable(zone))
2144                 return 0;
2145
2146         /*
2147          * Do not scan if the allocation should not be delayed.
2148          */
2149         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2150                         return 0;
2151
2152         /*
2153          * Only run zone reclaim on the local zone or on zones that do not
2154          * have associated processors. This will favor the local processor
2155          * over remote processors and spread off node memory allocations
2156          * as wide as possible.
2157          */
2158         node_id = zone_to_nid(zone);
2159         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2160                 return 0;
2161
2162         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2163                 return 0;
2164         ret = __zone_reclaim(zone, gfp_mask, order);
2165         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2166
2167         return ret;
2168 }
2169 #endif