hw_random doc updates
[sfrench/cifs-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
42 #include <asm/tlbflush.h>
43 #include <asm/div64.h>
44
45 #include <linux/swapops.h>
46
47 #include "internal.h"
48
49 struct scan_control {
50         /* Incremented by the number of inactive pages that were scanned */
51         unsigned long nr_scanned;
52
53         /* This context's GFP mask */
54         gfp_t gfp_mask;
55
56         int may_writepage;
57
58         /* Can pages be swapped as part of reclaim? */
59         int may_swap;
60
61         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
62          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
63          * In this context, it doesn't matter that we scan the
64          * whole list at once. */
65         int swap_cluster_max;
66
67         int swappiness;
68
69         int all_unreclaimable;
70
71         int order;
72
73         /*
74          * Pages that have (or should have) IO pending.  If we run into
75          * a lot of these, we're better off waiting a little for IO to
76          * finish rather than scanning more pages in the VM.
77          */
78         int nr_io_pages;
79
80         /* Which cgroup do we reclaim from */
81         struct mem_cgroup *mem_cgroup;
82
83         /* Pluggable isolate pages callback */
84         unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
85                         unsigned long *scanned, int order, int mode,
86                         struct zone *z, struct mem_cgroup *mem_cont,
87                         int active);
88 };
89
90 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
91
92 #ifdef ARCH_HAS_PREFETCH
93 #define prefetch_prev_lru_page(_page, _base, _field)                    \
94         do {                                                            \
95                 if ((_page)->lru.prev != _base) {                       \
96                         struct page *prev;                              \
97                                                                         \
98                         prev = lru_to_page(&(_page->lru));              \
99                         prefetch(&prev->_field);                        \
100                 }                                                       \
101         } while (0)
102 #else
103 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
104 #endif
105
106 #ifdef ARCH_HAS_PREFETCHW
107 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
108         do {                                                            \
109                 if ((_page)->lru.prev != _base) {                       \
110                         struct page *prev;                              \
111                                                                         \
112                         prev = lru_to_page(&(_page->lru));              \
113                         prefetchw(&prev->_field);                       \
114                 }                                                       \
115         } while (0)
116 #else
117 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
118 #endif
119
120 /*
121  * From 0 .. 100.  Higher means more swappy.
122  */
123 int vm_swappiness = 60;
124 long vm_total_pages;    /* The total number of pages which the VM controls */
125
126 static LIST_HEAD(shrinker_list);
127 static DECLARE_RWSEM(shrinker_rwsem);
128
129 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
130 #define scan_global_lru(sc)     (!(sc)->mem_cgroup)
131 #else
132 #define scan_global_lru(sc)     (1)
133 #endif
134
135 /*
136  * Add a shrinker callback to be called from the vm
137  */
138 void register_shrinker(struct shrinker *shrinker)
139 {
140         shrinker->nr = 0;
141         down_write(&shrinker_rwsem);
142         list_add_tail(&shrinker->list, &shrinker_list);
143         up_write(&shrinker_rwsem);
144 }
145 EXPORT_SYMBOL(register_shrinker);
146
147 /*
148  * Remove one
149  */
150 void unregister_shrinker(struct shrinker *shrinker)
151 {
152         down_write(&shrinker_rwsem);
153         list_del(&shrinker->list);
154         up_write(&shrinker_rwsem);
155 }
156 EXPORT_SYMBOL(unregister_shrinker);
157
158 #define SHRINK_BATCH 128
159 /*
160  * Call the shrink functions to age shrinkable caches
161  *
162  * Here we assume it costs one seek to replace a lru page and that it also
163  * takes a seek to recreate a cache object.  With this in mind we age equal
164  * percentages of the lru and ageable caches.  This should balance the seeks
165  * generated by these structures.
166  *
167  * If the vm encountered mapped pages on the LRU it increase the pressure on
168  * slab to avoid swapping.
169  *
170  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
171  *
172  * `lru_pages' represents the number of on-LRU pages in all the zones which
173  * are eligible for the caller's allocation attempt.  It is used for balancing
174  * slab reclaim versus page reclaim.
175  *
176  * Returns the number of slab objects which we shrunk.
177  */
178 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
179                         unsigned long lru_pages)
180 {
181         struct shrinker *shrinker;
182         unsigned long ret = 0;
183
184         if (scanned == 0)
185                 scanned = SWAP_CLUSTER_MAX;
186
187         if (!down_read_trylock(&shrinker_rwsem))
188                 return 1;       /* Assume we'll be able to shrink next time */
189
190         list_for_each_entry(shrinker, &shrinker_list, list) {
191                 unsigned long long delta;
192                 unsigned long total_scan;
193                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
194
195                 delta = (4 * scanned) / shrinker->seeks;
196                 delta *= max_pass;
197                 do_div(delta, lru_pages + 1);
198                 shrinker->nr += delta;
199                 if (shrinker->nr < 0) {
200                         printk(KERN_ERR "%s: nr=%ld\n",
201                                         __FUNCTION__, shrinker->nr);
202                         shrinker->nr = max_pass;
203                 }
204
205                 /*
206                  * Avoid risking looping forever due to too large nr value:
207                  * never try to free more than twice the estimate number of
208                  * freeable entries.
209                  */
210                 if (shrinker->nr > max_pass * 2)
211                         shrinker->nr = max_pass * 2;
212
213                 total_scan = shrinker->nr;
214                 shrinker->nr = 0;
215
216                 while (total_scan >= SHRINK_BATCH) {
217                         long this_scan = SHRINK_BATCH;
218                         int shrink_ret;
219                         int nr_before;
220
221                         nr_before = (*shrinker->shrink)(0, gfp_mask);
222                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
223                         if (shrink_ret == -1)
224                                 break;
225                         if (shrink_ret < nr_before)
226                                 ret += nr_before - shrink_ret;
227                         count_vm_events(SLABS_SCANNED, this_scan);
228                         total_scan -= this_scan;
229
230                         cond_resched();
231                 }
232
233                 shrinker->nr += total_scan;
234         }
235         up_read(&shrinker_rwsem);
236         return ret;
237 }
238
239 /* Called without lock on whether page is mapped, so answer is unstable */
240 static inline int page_mapping_inuse(struct page *page)
241 {
242         struct address_space *mapping;
243
244         /* Page is in somebody's page tables. */
245         if (page_mapped(page))
246                 return 1;
247
248         /* Be more reluctant to reclaim swapcache than pagecache */
249         if (PageSwapCache(page))
250                 return 1;
251
252         mapping = page_mapping(page);
253         if (!mapping)
254                 return 0;
255
256         /* File is mmap'd by somebody? */
257         return mapping_mapped(mapping);
258 }
259
260 static inline int is_page_cache_freeable(struct page *page)
261 {
262         return page_count(page) - !!PagePrivate(page) == 2;
263 }
264
265 static int may_write_to_queue(struct backing_dev_info *bdi)
266 {
267         if (current->flags & PF_SWAPWRITE)
268                 return 1;
269         if (!bdi_write_congested(bdi))
270                 return 1;
271         if (bdi == current->backing_dev_info)
272                 return 1;
273         return 0;
274 }
275
276 /*
277  * We detected a synchronous write error writing a page out.  Probably
278  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
279  * fsync(), msync() or close().
280  *
281  * The tricky part is that after writepage we cannot touch the mapping: nothing
282  * prevents it from being freed up.  But we have a ref on the page and once
283  * that page is locked, the mapping is pinned.
284  *
285  * We're allowed to run sleeping lock_page() here because we know the caller has
286  * __GFP_FS.
287  */
288 static void handle_write_error(struct address_space *mapping,
289                                 struct page *page, int error)
290 {
291         lock_page(page);
292         if (page_mapping(page) == mapping)
293                 mapping_set_error(mapping, error);
294         unlock_page(page);
295 }
296
297 /* Request for sync pageout. */
298 enum pageout_io {
299         PAGEOUT_IO_ASYNC,
300         PAGEOUT_IO_SYNC,
301 };
302
303 /* possible outcome of pageout() */
304 typedef enum {
305         /* failed to write page out, page is locked */
306         PAGE_KEEP,
307         /* move page to the active list, page is locked */
308         PAGE_ACTIVATE,
309         /* page has been sent to the disk successfully, page is unlocked */
310         PAGE_SUCCESS,
311         /* page is clean and locked */
312         PAGE_CLEAN,
313 } pageout_t;
314
315 /*
316  * pageout is called by shrink_page_list() for each dirty page.
317  * Calls ->writepage().
318  */
319 static pageout_t pageout(struct page *page, struct address_space *mapping,
320                                                 enum pageout_io sync_writeback)
321 {
322         /*
323          * If the page is dirty, only perform writeback if that write
324          * will be non-blocking.  To prevent this allocation from being
325          * stalled by pagecache activity.  But note that there may be
326          * stalls if we need to run get_block().  We could test
327          * PagePrivate for that.
328          *
329          * If this process is currently in generic_file_write() against
330          * this page's queue, we can perform writeback even if that
331          * will block.
332          *
333          * If the page is swapcache, write it back even if that would
334          * block, for some throttling. This happens by accident, because
335          * swap_backing_dev_info is bust: it doesn't reflect the
336          * congestion state of the swapdevs.  Easy to fix, if needed.
337          * See swapfile.c:page_queue_congested().
338          */
339         if (!is_page_cache_freeable(page))
340                 return PAGE_KEEP;
341         if (!mapping) {
342                 /*
343                  * Some data journaling orphaned pages can have
344                  * page->mapping == NULL while being dirty with clean buffers.
345                  */
346                 if (PagePrivate(page)) {
347                         if (try_to_free_buffers(page)) {
348                                 ClearPageDirty(page);
349                                 printk("%s: orphaned page\n", __FUNCTION__);
350                                 return PAGE_CLEAN;
351                         }
352                 }
353                 return PAGE_KEEP;
354         }
355         if (mapping->a_ops->writepage == NULL)
356                 return PAGE_ACTIVATE;
357         if (!may_write_to_queue(mapping->backing_dev_info))
358                 return PAGE_KEEP;
359
360         if (clear_page_dirty_for_io(page)) {
361                 int res;
362                 struct writeback_control wbc = {
363                         .sync_mode = WB_SYNC_NONE,
364                         .nr_to_write = SWAP_CLUSTER_MAX,
365                         .range_start = 0,
366                         .range_end = LLONG_MAX,
367                         .nonblocking = 1,
368                         .for_reclaim = 1,
369                 };
370
371                 SetPageReclaim(page);
372                 res = mapping->a_ops->writepage(page, &wbc);
373                 if (res < 0)
374                         handle_write_error(mapping, page, res);
375                 if (res == AOP_WRITEPAGE_ACTIVATE) {
376                         ClearPageReclaim(page);
377                         return PAGE_ACTIVATE;
378                 }
379
380                 /*
381                  * Wait on writeback if requested to. This happens when
382                  * direct reclaiming a large contiguous area and the
383                  * first attempt to free a range of pages fails.
384                  */
385                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
386                         wait_on_page_writeback(page);
387
388                 if (!PageWriteback(page)) {
389                         /* synchronous write or broken a_ops? */
390                         ClearPageReclaim(page);
391                 }
392                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
393                 return PAGE_SUCCESS;
394         }
395
396         return PAGE_CLEAN;
397 }
398
399 /*
400  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
401  * someone else has a ref on the page, abort and return 0.  If it was
402  * successfully detached, return 1.  Assumes the caller has a single ref on
403  * this page.
404  */
405 int remove_mapping(struct address_space *mapping, struct page *page)
406 {
407         BUG_ON(!PageLocked(page));
408         BUG_ON(mapping != page_mapping(page));
409
410         write_lock_irq(&mapping->tree_lock);
411         /*
412          * The non racy check for a busy page.
413          *
414          * Must be careful with the order of the tests. When someone has
415          * a ref to the page, it may be possible that they dirty it then
416          * drop the reference. So if PageDirty is tested before page_count
417          * here, then the following race may occur:
418          *
419          * get_user_pages(&page);
420          * [user mapping goes away]
421          * write_to(page);
422          *                              !PageDirty(page)    [good]
423          * SetPageDirty(page);
424          * put_page(page);
425          *                              !page_count(page)   [good, discard it]
426          *
427          * [oops, our write_to data is lost]
428          *
429          * Reversing the order of the tests ensures such a situation cannot
430          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
431          * load is not satisfied before that of page->_count.
432          *
433          * Note that if SetPageDirty is always performed via set_page_dirty,
434          * and thus under tree_lock, then this ordering is not required.
435          */
436         if (unlikely(page_count(page) != 2))
437                 goto cannot_free;
438         smp_rmb();
439         if (unlikely(PageDirty(page)))
440                 goto cannot_free;
441
442         if (PageSwapCache(page)) {
443                 swp_entry_t swap = { .val = page_private(page) };
444                 __delete_from_swap_cache(page);
445                 write_unlock_irq(&mapping->tree_lock);
446                 swap_free(swap);
447                 __put_page(page);       /* The pagecache ref */
448                 return 1;
449         }
450
451         __remove_from_page_cache(page);
452         write_unlock_irq(&mapping->tree_lock);
453         __put_page(page);
454         return 1;
455
456 cannot_free:
457         write_unlock_irq(&mapping->tree_lock);
458         return 0;
459 }
460
461 /*
462  * shrink_page_list() returns the number of reclaimed pages
463  */
464 static unsigned long shrink_page_list(struct list_head *page_list,
465                                         struct scan_control *sc,
466                                         enum pageout_io sync_writeback)
467 {
468         LIST_HEAD(ret_pages);
469         struct pagevec freed_pvec;
470         int pgactivate = 0;
471         unsigned long nr_reclaimed = 0;
472
473         cond_resched();
474
475         pagevec_init(&freed_pvec, 1);
476         while (!list_empty(page_list)) {
477                 struct address_space *mapping;
478                 struct page *page;
479                 int may_enter_fs;
480                 int referenced;
481
482                 cond_resched();
483
484                 page = lru_to_page(page_list);
485                 list_del(&page->lru);
486
487                 if (TestSetPageLocked(page))
488                         goto keep;
489
490                 VM_BUG_ON(PageActive(page));
491
492                 sc->nr_scanned++;
493
494                 if (!sc->may_swap && page_mapped(page))
495                         goto keep_locked;
496
497                 /* Double the slab pressure for mapped and swapcache pages */
498                 if (page_mapped(page) || PageSwapCache(page))
499                         sc->nr_scanned++;
500
501                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
502                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
503
504                 if (PageWriteback(page)) {
505                         /*
506                          * Synchronous reclaim is performed in two passes,
507                          * first an asynchronous pass over the list to
508                          * start parallel writeback, and a second synchronous
509                          * pass to wait for the IO to complete.  Wait here
510                          * for any page for which writeback has already
511                          * started.
512                          */
513                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
514                                 wait_on_page_writeback(page);
515                         else {
516                                 sc->nr_io_pages++;
517                                 goto keep_locked;
518                         }
519                 }
520
521                 referenced = page_referenced(page, 1, sc->mem_cgroup);
522                 /* In active use or really unfreeable?  Activate it. */
523                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
524                                         referenced && page_mapping_inuse(page))
525                         goto activate_locked;
526
527 #ifdef CONFIG_SWAP
528                 /*
529                  * Anonymous process memory has backing store?
530                  * Try to allocate it some swap space here.
531                  */
532                 if (PageAnon(page) && !PageSwapCache(page))
533                         if (!add_to_swap(page, GFP_ATOMIC))
534                                 goto activate_locked;
535 #endif /* CONFIG_SWAP */
536
537                 mapping = page_mapping(page);
538
539                 /*
540                  * The page is mapped into the page tables of one or more
541                  * processes. Try to unmap it here.
542                  */
543                 if (page_mapped(page) && mapping) {
544                         switch (try_to_unmap(page, 0)) {
545                         case SWAP_FAIL:
546                                 goto activate_locked;
547                         case SWAP_AGAIN:
548                                 goto keep_locked;
549                         case SWAP_SUCCESS:
550                                 ; /* try to free the page below */
551                         }
552                 }
553
554                 if (PageDirty(page)) {
555                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
556                                 goto keep_locked;
557                         if (!may_enter_fs) {
558                                 sc->nr_io_pages++;
559                                 goto keep_locked;
560                         }
561                         if (!sc->may_writepage)
562                                 goto keep_locked;
563
564                         /* Page is dirty, try to write it out here */
565                         switch (pageout(page, mapping, sync_writeback)) {
566                         case PAGE_KEEP:
567                                 goto keep_locked;
568                         case PAGE_ACTIVATE:
569                                 goto activate_locked;
570                         case PAGE_SUCCESS:
571                                 if (PageWriteback(page) || PageDirty(page)) {
572                                         sc->nr_io_pages++;
573                                         goto keep;
574                                 }
575                                 /*
576                                  * A synchronous write - probably a ramdisk.  Go
577                                  * ahead and try to reclaim the page.
578                                  */
579                                 if (TestSetPageLocked(page))
580                                         goto keep;
581                                 if (PageDirty(page) || PageWriteback(page))
582                                         goto keep_locked;
583                                 mapping = page_mapping(page);
584                         case PAGE_CLEAN:
585                                 ; /* try to free the page below */
586                         }
587                 }
588
589                 /*
590                  * If the page has buffers, try to free the buffer mappings
591                  * associated with this page. If we succeed we try to free
592                  * the page as well.
593                  *
594                  * We do this even if the page is PageDirty().
595                  * try_to_release_page() does not perform I/O, but it is
596                  * possible for a page to have PageDirty set, but it is actually
597                  * clean (all its buffers are clean).  This happens if the
598                  * buffers were written out directly, with submit_bh(). ext3
599                  * will do this, as well as the blockdev mapping. 
600                  * try_to_release_page() will discover that cleanness and will
601                  * drop the buffers and mark the page clean - it can be freed.
602                  *
603                  * Rarely, pages can have buffers and no ->mapping.  These are
604                  * the pages which were not successfully invalidated in
605                  * truncate_complete_page().  We try to drop those buffers here
606                  * and if that worked, and the page is no longer mapped into
607                  * process address space (page_count == 1) it can be freed.
608                  * Otherwise, leave the page on the LRU so it is swappable.
609                  */
610                 if (PagePrivate(page)) {
611                         if (!try_to_release_page(page, sc->gfp_mask))
612                                 goto activate_locked;
613                         if (!mapping && page_count(page) == 1)
614                                 goto free_it;
615                 }
616
617                 if (!mapping || !remove_mapping(mapping, page))
618                         goto keep_locked;
619
620 free_it:
621                 unlock_page(page);
622                 nr_reclaimed++;
623                 if (!pagevec_add(&freed_pvec, page))
624                         __pagevec_release_nonlru(&freed_pvec);
625                 continue;
626
627 activate_locked:
628                 SetPageActive(page);
629                 pgactivate++;
630 keep_locked:
631                 unlock_page(page);
632 keep:
633                 list_add(&page->lru, &ret_pages);
634                 VM_BUG_ON(PageLRU(page));
635         }
636         list_splice(&ret_pages, page_list);
637         if (pagevec_count(&freed_pvec))
638                 __pagevec_release_nonlru(&freed_pvec);
639         count_vm_events(PGACTIVATE, pgactivate);
640         return nr_reclaimed;
641 }
642
643 /* LRU Isolation modes. */
644 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
645 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
646 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
647
648 /*
649  * Attempt to remove the specified page from its LRU.  Only take this page
650  * if it is of the appropriate PageActive status.  Pages which are being
651  * freed elsewhere are also ignored.
652  *
653  * page:        page to consider
654  * mode:        one of the LRU isolation modes defined above
655  *
656  * returns 0 on success, -ve errno on failure.
657  */
658 int __isolate_lru_page(struct page *page, int mode)
659 {
660         int ret = -EINVAL;
661
662         /* Only take pages on the LRU. */
663         if (!PageLRU(page))
664                 return ret;
665
666         /*
667          * When checking the active state, we need to be sure we are
668          * dealing with comparible boolean values.  Take the logical not
669          * of each.
670          */
671         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
672                 return ret;
673
674         ret = -EBUSY;
675         if (likely(get_page_unless_zero(page))) {
676                 /*
677                  * Be careful not to clear PageLRU until after we're
678                  * sure the page is not being freed elsewhere -- the
679                  * page release code relies on it.
680                  */
681                 ClearPageLRU(page);
682                 ret = 0;
683         }
684
685         return ret;
686 }
687
688 /*
689  * zone->lru_lock is heavily contended.  Some of the functions that
690  * shrink the lists perform better by taking out a batch of pages
691  * and working on them outside the LRU lock.
692  *
693  * For pagecache intensive workloads, this function is the hottest
694  * spot in the kernel (apart from copy_*_user functions).
695  *
696  * Appropriate locks must be held before calling this function.
697  *
698  * @nr_to_scan: The number of pages to look through on the list.
699  * @src:        The LRU list to pull pages off.
700  * @dst:        The temp list to put pages on to.
701  * @scanned:    The number of pages that were scanned.
702  * @order:      The caller's attempted allocation order
703  * @mode:       One of the LRU isolation modes
704  *
705  * returns how many pages were moved onto *@dst.
706  */
707 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
708                 struct list_head *src, struct list_head *dst,
709                 unsigned long *scanned, int order, int mode)
710 {
711         unsigned long nr_taken = 0;
712         unsigned long scan;
713
714         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
715                 struct page *page;
716                 unsigned long pfn;
717                 unsigned long end_pfn;
718                 unsigned long page_pfn;
719                 int zone_id;
720
721                 page = lru_to_page(src);
722                 prefetchw_prev_lru_page(page, src, flags);
723
724                 VM_BUG_ON(!PageLRU(page));
725
726                 switch (__isolate_lru_page(page, mode)) {
727                 case 0:
728                         list_move(&page->lru, dst);
729                         nr_taken++;
730                         break;
731
732                 case -EBUSY:
733                         /* else it is being freed elsewhere */
734                         list_move(&page->lru, src);
735                         continue;
736
737                 default:
738                         BUG();
739                 }
740
741                 if (!order)
742                         continue;
743
744                 /*
745                  * Attempt to take all pages in the order aligned region
746                  * surrounding the tag page.  Only take those pages of
747                  * the same active state as that tag page.  We may safely
748                  * round the target page pfn down to the requested order
749                  * as the mem_map is guarenteed valid out to MAX_ORDER,
750                  * where that page is in a different zone we will detect
751                  * it from its zone id and abort this block scan.
752                  */
753                 zone_id = page_zone_id(page);
754                 page_pfn = page_to_pfn(page);
755                 pfn = page_pfn & ~((1 << order) - 1);
756                 end_pfn = pfn + (1 << order);
757                 for (; pfn < end_pfn; pfn++) {
758                         struct page *cursor_page;
759
760                         /* The target page is in the block, ignore it. */
761                         if (unlikely(pfn == page_pfn))
762                                 continue;
763
764                         /* Avoid holes within the zone. */
765                         if (unlikely(!pfn_valid_within(pfn)))
766                                 break;
767
768                         cursor_page = pfn_to_page(pfn);
769                         /* Check that we have not crossed a zone boundary. */
770                         if (unlikely(page_zone_id(cursor_page) != zone_id))
771                                 continue;
772                         switch (__isolate_lru_page(cursor_page, mode)) {
773                         case 0:
774                                 list_move(&cursor_page->lru, dst);
775                                 nr_taken++;
776                                 scan++;
777                                 break;
778
779                         case -EBUSY:
780                                 /* else it is being freed elsewhere */
781                                 list_move(&cursor_page->lru, src);
782                         default:
783                                 break;
784                         }
785                 }
786         }
787
788         *scanned = scan;
789         return nr_taken;
790 }
791
792 static unsigned long isolate_pages_global(unsigned long nr,
793                                         struct list_head *dst,
794                                         unsigned long *scanned, int order,
795                                         int mode, struct zone *z,
796                                         struct mem_cgroup *mem_cont,
797                                         int active)
798 {
799         if (active)
800                 return isolate_lru_pages(nr, &z->active_list, dst,
801                                                 scanned, order, mode);
802         else
803                 return isolate_lru_pages(nr, &z->inactive_list, dst,
804                                                 scanned, order, mode);
805 }
806
807 /*
808  * clear_active_flags() is a helper for shrink_active_list(), clearing
809  * any active bits from the pages in the list.
810  */
811 static unsigned long clear_active_flags(struct list_head *page_list)
812 {
813         int nr_active = 0;
814         struct page *page;
815
816         list_for_each_entry(page, page_list, lru)
817                 if (PageActive(page)) {
818                         ClearPageActive(page);
819                         nr_active++;
820                 }
821
822         return nr_active;
823 }
824
825 /*
826  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
827  * of reclaimed pages
828  */
829 static unsigned long shrink_inactive_list(unsigned long max_scan,
830                                 struct zone *zone, struct scan_control *sc)
831 {
832         LIST_HEAD(page_list);
833         struct pagevec pvec;
834         unsigned long nr_scanned = 0;
835         unsigned long nr_reclaimed = 0;
836
837         pagevec_init(&pvec, 1);
838
839         lru_add_drain();
840         spin_lock_irq(&zone->lru_lock);
841         do {
842                 struct page *page;
843                 unsigned long nr_taken;
844                 unsigned long nr_scan;
845                 unsigned long nr_freed;
846                 unsigned long nr_active;
847
848                 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
849                              &page_list, &nr_scan, sc->order,
850                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
851                                              ISOLATE_BOTH : ISOLATE_INACTIVE,
852                                 zone, sc->mem_cgroup, 0);
853                 nr_active = clear_active_flags(&page_list);
854                 __count_vm_events(PGDEACTIVATE, nr_active);
855
856                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
857                 __mod_zone_page_state(zone, NR_INACTIVE,
858                                                 -(nr_taken - nr_active));
859                 if (scan_global_lru(sc))
860                         zone->pages_scanned += nr_scan;
861                 spin_unlock_irq(&zone->lru_lock);
862
863                 nr_scanned += nr_scan;
864                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
865
866                 /*
867                  * If we are direct reclaiming for contiguous pages and we do
868                  * not reclaim everything in the list, try again and wait
869                  * for IO to complete. This will stall high-order allocations
870                  * but that should be acceptable to the caller
871                  */
872                 if (nr_freed < nr_taken && !current_is_kswapd() &&
873                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
874                         congestion_wait(WRITE, HZ/10);
875
876                         /*
877                          * The attempt at page out may have made some
878                          * of the pages active, mark them inactive again.
879                          */
880                         nr_active = clear_active_flags(&page_list);
881                         count_vm_events(PGDEACTIVATE, nr_active);
882
883                         nr_freed += shrink_page_list(&page_list, sc,
884                                                         PAGEOUT_IO_SYNC);
885                 }
886
887                 nr_reclaimed += nr_freed;
888                 local_irq_disable();
889                 if (current_is_kswapd()) {
890                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
891                         __count_vm_events(KSWAPD_STEAL, nr_freed);
892                 } else if (scan_global_lru(sc))
893                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
894
895                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
896
897                 if (nr_taken == 0)
898                         goto done;
899
900                 spin_lock(&zone->lru_lock);
901                 /*
902                  * Put back any unfreeable pages.
903                  */
904                 while (!list_empty(&page_list)) {
905                         page = lru_to_page(&page_list);
906                         VM_BUG_ON(PageLRU(page));
907                         SetPageLRU(page);
908                         list_del(&page->lru);
909                         if (PageActive(page))
910                                 add_page_to_active_list(zone, page);
911                         else
912                                 add_page_to_inactive_list(zone, page);
913                         if (!pagevec_add(&pvec, page)) {
914                                 spin_unlock_irq(&zone->lru_lock);
915                                 __pagevec_release(&pvec);
916                                 spin_lock_irq(&zone->lru_lock);
917                         }
918                 }
919         } while (nr_scanned < max_scan);
920         spin_unlock(&zone->lru_lock);
921 done:
922         local_irq_enable();
923         pagevec_release(&pvec);
924         return nr_reclaimed;
925 }
926
927 /*
928  * We are about to scan this zone at a certain priority level.  If that priority
929  * level is smaller (ie: more urgent) than the previous priority, then note
930  * that priority level within the zone.  This is done so that when the next
931  * process comes in to scan this zone, it will immediately start out at this
932  * priority level rather than having to build up its own scanning priority.
933  * Here, this priority affects only the reclaim-mapped threshold.
934  */
935 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
936 {
937         if (priority < zone->prev_priority)
938                 zone->prev_priority = priority;
939 }
940
941 static inline int zone_is_near_oom(struct zone *zone)
942 {
943         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
944                                 + zone_page_state(zone, NR_INACTIVE))*3;
945 }
946
947 /*
948  * Determine we should try to reclaim mapped pages.
949  * This is called only when sc->mem_cgroup is NULL.
950  */
951 static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
952                                 int priority)
953 {
954         long mapped_ratio;
955         long distress;
956         long swap_tendency;
957         long imbalance;
958         int reclaim_mapped = 0;
959         int prev_priority;
960
961         if (scan_global_lru(sc) && zone_is_near_oom(zone))
962                 return 1;
963         /*
964          * `distress' is a measure of how much trouble we're having
965          * reclaiming pages.  0 -> no problems.  100 -> great trouble.
966          */
967         if (scan_global_lru(sc))
968                 prev_priority = zone->prev_priority;
969         else
970                 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
971
972         distress = 100 >> min(prev_priority, priority);
973
974         /*
975          * The point of this algorithm is to decide when to start
976          * reclaiming mapped memory instead of just pagecache.  Work out
977          * how much memory
978          * is mapped.
979          */
980         if (scan_global_lru(sc))
981                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
982                                 global_page_state(NR_ANON_PAGES)) * 100) /
983                                         vm_total_pages;
984         else
985                 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
986
987         /*
988          * Now decide how much we really want to unmap some pages.  The
989          * mapped ratio is downgraded - just because there's a lot of
990          * mapped memory doesn't necessarily mean that page reclaim
991          * isn't succeeding.
992          *
993          * The distress ratio is important - we don't want to start
994          * going oom.
995          *
996          * A 100% value of vm_swappiness overrides this algorithm
997          * altogether.
998          */
999         swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1000
1001         /*
1002          * If there's huge imbalance between active and inactive
1003          * (think active 100 times larger than inactive) we should
1004          * become more permissive, or the system will take too much
1005          * cpu before it start swapping during memory pressure.
1006          * Distress is about avoiding early-oom, this is about
1007          * making swappiness graceful despite setting it to low
1008          * values.
1009          *
1010          * Avoid div by zero with nr_inactive+1, and max resulting
1011          * value is vm_total_pages.
1012          */
1013         if (scan_global_lru(sc)) {
1014                 imbalance  = zone_page_state(zone, NR_ACTIVE);
1015                 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1016         } else
1017                 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1018
1019         /*
1020          * Reduce the effect of imbalance if swappiness is low,
1021          * this means for a swappiness very low, the imbalance
1022          * must be much higher than 100 for this logic to make
1023          * the difference.
1024          *
1025          * Max temporary value is vm_total_pages*100.
1026          */
1027         imbalance *= (vm_swappiness + 1);
1028         imbalance /= 100;
1029
1030         /*
1031          * If not much of the ram is mapped, makes the imbalance
1032          * less relevant, it's high priority we refill the inactive
1033          * list with mapped pages only in presence of high ratio of
1034          * mapped pages.
1035          *
1036          * Max temporary value is vm_total_pages*100.
1037          */
1038         imbalance *= mapped_ratio;
1039         imbalance /= 100;
1040
1041         /* apply imbalance feedback to swap_tendency */
1042         swap_tendency += imbalance;
1043
1044         /*
1045          * Now use this metric to decide whether to start moving mapped
1046          * memory onto the inactive list.
1047          */
1048         if (swap_tendency >= 100)
1049                 reclaim_mapped = 1;
1050
1051         return reclaim_mapped;
1052 }
1053
1054 /*
1055  * This moves pages from the active list to the inactive list.
1056  *
1057  * We move them the other way if the page is referenced by one or more
1058  * processes, from rmap.
1059  *
1060  * If the pages are mostly unmapped, the processing is fast and it is
1061  * appropriate to hold zone->lru_lock across the whole operation.  But if
1062  * the pages are mapped, the processing is slow (page_referenced()) so we
1063  * should drop zone->lru_lock around each page.  It's impossible to balance
1064  * this, so instead we remove the pages from the LRU while processing them.
1065  * It is safe to rely on PG_active against the non-LRU pages in here because
1066  * nobody will play with that bit on a non-LRU page.
1067  *
1068  * The downside is that we have to touch page->_count against each page.
1069  * But we had to alter page->flags anyway.
1070  */
1071
1072
1073 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1074                                 struct scan_control *sc, int priority)
1075 {
1076         unsigned long pgmoved;
1077         int pgdeactivate = 0;
1078         unsigned long pgscanned;
1079         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1080         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1081         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1082         struct page *page;
1083         struct pagevec pvec;
1084         int reclaim_mapped = 0;
1085
1086         if (sc->may_swap)
1087                 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
1088
1089         lru_add_drain();
1090         spin_lock_irq(&zone->lru_lock);
1091         pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1092                                         ISOLATE_ACTIVE, zone,
1093                                         sc->mem_cgroup, 1);
1094         /*
1095          * zone->pages_scanned is used for detect zone's oom
1096          * mem_cgroup remembers nr_scan by itself.
1097          */
1098         if (scan_global_lru(sc))
1099                 zone->pages_scanned += pgscanned;
1100
1101         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1102         spin_unlock_irq(&zone->lru_lock);
1103
1104         while (!list_empty(&l_hold)) {
1105                 cond_resched();
1106                 page = lru_to_page(&l_hold);
1107                 list_del(&page->lru);
1108                 if (page_mapped(page)) {
1109                         if (!reclaim_mapped ||
1110                             (total_swap_pages == 0 && PageAnon(page)) ||
1111                             page_referenced(page, 0, sc->mem_cgroup)) {
1112                                 list_add(&page->lru, &l_active);
1113                                 continue;
1114                         }
1115                 }
1116                 list_add(&page->lru, &l_inactive);
1117         }
1118
1119         pagevec_init(&pvec, 1);
1120         pgmoved = 0;
1121         spin_lock_irq(&zone->lru_lock);
1122         while (!list_empty(&l_inactive)) {
1123                 page = lru_to_page(&l_inactive);
1124                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1125                 VM_BUG_ON(PageLRU(page));
1126                 SetPageLRU(page);
1127                 VM_BUG_ON(!PageActive(page));
1128                 ClearPageActive(page);
1129
1130                 list_move(&page->lru, &zone->inactive_list);
1131                 mem_cgroup_move_lists(page, false);
1132                 pgmoved++;
1133                 if (!pagevec_add(&pvec, page)) {
1134                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1135                         spin_unlock_irq(&zone->lru_lock);
1136                         pgdeactivate += pgmoved;
1137                         pgmoved = 0;
1138                         if (buffer_heads_over_limit)
1139                                 pagevec_strip(&pvec);
1140                         __pagevec_release(&pvec);
1141                         spin_lock_irq(&zone->lru_lock);
1142                 }
1143         }
1144         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1145         pgdeactivate += pgmoved;
1146         if (buffer_heads_over_limit) {
1147                 spin_unlock_irq(&zone->lru_lock);
1148                 pagevec_strip(&pvec);
1149                 spin_lock_irq(&zone->lru_lock);
1150         }
1151
1152         pgmoved = 0;
1153         while (!list_empty(&l_active)) {
1154                 page = lru_to_page(&l_active);
1155                 prefetchw_prev_lru_page(page, &l_active, flags);
1156                 VM_BUG_ON(PageLRU(page));
1157                 SetPageLRU(page);
1158                 VM_BUG_ON(!PageActive(page));
1159
1160                 list_move(&page->lru, &zone->active_list);
1161                 mem_cgroup_move_lists(page, true);
1162                 pgmoved++;
1163                 if (!pagevec_add(&pvec, page)) {
1164                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1165                         pgmoved = 0;
1166                         spin_unlock_irq(&zone->lru_lock);
1167                         __pagevec_release(&pvec);
1168                         spin_lock_irq(&zone->lru_lock);
1169                 }
1170         }
1171         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1172
1173         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1174         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1175         spin_unlock_irq(&zone->lru_lock);
1176
1177         pagevec_release(&pvec);
1178 }
1179
1180 /*
1181  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1182  */
1183 static unsigned long shrink_zone(int priority, struct zone *zone,
1184                                 struct scan_control *sc)
1185 {
1186         unsigned long nr_active;
1187         unsigned long nr_inactive;
1188         unsigned long nr_to_scan;
1189         unsigned long nr_reclaimed = 0;
1190
1191         if (scan_global_lru(sc)) {
1192                 /*
1193                  * Add one to nr_to_scan just to make sure that the kernel
1194                  * will slowly sift through the active list.
1195                  */
1196                 zone->nr_scan_active +=
1197                         (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1198                 nr_active = zone->nr_scan_active;
1199                 zone->nr_scan_inactive +=
1200                         (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1201                 nr_inactive = zone->nr_scan_inactive;
1202                 if (nr_inactive >= sc->swap_cluster_max)
1203                         zone->nr_scan_inactive = 0;
1204                 else
1205                         nr_inactive = 0;
1206
1207                 if (nr_active >= sc->swap_cluster_max)
1208                         zone->nr_scan_active = 0;
1209                 else
1210                         nr_active = 0;
1211         } else {
1212                 /*
1213                  * This reclaim occurs not because zone memory shortage but
1214                  * because memory controller hits its limit.
1215                  * Then, don't modify zone reclaim related data.
1216                  */
1217                 nr_active = mem_cgroup_calc_reclaim_active(sc->mem_cgroup,
1218                                         zone, priority);
1219
1220                 nr_inactive = mem_cgroup_calc_reclaim_inactive(sc->mem_cgroup,
1221                                         zone, priority);
1222         }
1223
1224
1225         while (nr_active || nr_inactive) {
1226                 if (nr_active) {
1227                         nr_to_scan = min(nr_active,
1228                                         (unsigned long)sc->swap_cluster_max);
1229                         nr_active -= nr_to_scan;
1230                         shrink_active_list(nr_to_scan, zone, sc, priority);
1231                 }
1232
1233                 if (nr_inactive) {
1234                         nr_to_scan = min(nr_inactive,
1235                                         (unsigned long)sc->swap_cluster_max);
1236                         nr_inactive -= nr_to_scan;
1237                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1238                                                                 sc);
1239                 }
1240         }
1241
1242         throttle_vm_writeout(sc->gfp_mask);
1243         return nr_reclaimed;
1244 }
1245
1246 /*
1247  * This is the direct reclaim path, for page-allocating processes.  We only
1248  * try to reclaim pages from zones which will satisfy the caller's allocation
1249  * request.
1250  *
1251  * We reclaim from a zone even if that zone is over pages_high.  Because:
1252  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1253  *    allocation or
1254  * b) The zones may be over pages_high but they must go *over* pages_high to
1255  *    satisfy the `incremental min' zone defense algorithm.
1256  *
1257  * Returns the number of reclaimed pages.
1258  *
1259  * If a zone is deemed to be full of pinned pages then just give it a light
1260  * scan then give up on it.
1261  */
1262 static unsigned long shrink_zones(int priority, struct zone **zones,
1263                                         struct scan_control *sc)
1264 {
1265         unsigned long nr_reclaimed = 0;
1266         int i;
1267
1268
1269         sc->all_unreclaimable = 1;
1270         for (i = 0; zones[i] != NULL; i++) {
1271                 struct zone *zone = zones[i];
1272
1273                 if (!populated_zone(zone))
1274                         continue;
1275                 /*
1276                  * Take care memory controller reclaiming has small influence
1277                  * to global LRU.
1278                  */
1279                 if (scan_global_lru(sc)) {
1280                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1281                                 continue;
1282                         note_zone_scanning_priority(zone, priority);
1283
1284                         if (zone_is_all_unreclaimable(zone) &&
1285                                                 priority != DEF_PRIORITY)
1286                                 continue;       /* Let kswapd poll it */
1287                         sc->all_unreclaimable = 0;
1288                 } else {
1289                         /*
1290                          * Ignore cpuset limitation here. We just want to reduce
1291                          * # of used pages by us regardless of memory shortage.
1292                          */
1293                         sc->all_unreclaimable = 0;
1294                         mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1295                                                         priority);
1296                 }
1297
1298                 nr_reclaimed += shrink_zone(priority, zone, sc);
1299         }
1300
1301         return nr_reclaimed;
1302 }
1303  
1304 /*
1305  * This is the main entry point to direct page reclaim.
1306  *
1307  * If a full scan of the inactive list fails to free enough memory then we
1308  * are "out of memory" and something needs to be killed.
1309  *
1310  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1311  * high - the zone may be full of dirty or under-writeback pages, which this
1312  * caller can't do much about.  We kick pdflush and take explicit naps in the
1313  * hope that some of these pages can be written.  But if the allocating task
1314  * holds filesystem locks which prevent writeout this might not work, and the
1315  * allocation attempt will fail.
1316  */
1317 static unsigned long do_try_to_free_pages(struct zone **zones, gfp_t gfp_mask,
1318                                           struct scan_control *sc)
1319 {
1320         int priority;
1321         int ret = 0;
1322         unsigned long total_scanned = 0;
1323         unsigned long nr_reclaimed = 0;
1324         struct reclaim_state *reclaim_state = current->reclaim_state;
1325         unsigned long lru_pages = 0;
1326         int i;
1327
1328         if (scan_global_lru(sc))
1329                 count_vm_event(ALLOCSTALL);
1330         /*
1331          * mem_cgroup will not do shrink_slab.
1332          */
1333         if (scan_global_lru(sc)) {
1334                 for (i = 0; zones[i] != NULL; i++) {
1335                         struct zone *zone = zones[i];
1336
1337                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1338                                 continue;
1339
1340                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1341                                         + zone_page_state(zone, NR_INACTIVE);
1342                 }
1343         }
1344
1345         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1346                 sc->nr_scanned = 0;
1347                 sc->nr_io_pages = 0;
1348                 if (!priority)
1349                         disable_swap_token();
1350                 nr_reclaimed += shrink_zones(priority, zones, sc);
1351                 /*
1352                  * Don't shrink slabs when reclaiming memory from
1353                  * over limit cgroups
1354                  */
1355                 if (scan_global_lru(sc)) {
1356                         shrink_slab(sc->nr_scanned, gfp_mask, lru_pages);
1357                         if (reclaim_state) {
1358                                 nr_reclaimed += reclaim_state->reclaimed_slab;
1359                                 reclaim_state->reclaimed_slab = 0;
1360                         }
1361                 }
1362                 total_scanned += sc->nr_scanned;
1363                 if (nr_reclaimed >= sc->swap_cluster_max) {
1364                         ret = 1;
1365                         goto out;
1366                 }
1367
1368                 /*
1369                  * Try to write back as many pages as we just scanned.  This
1370                  * tends to cause slow streaming writers to write data to the
1371                  * disk smoothly, at the dirtying rate, which is nice.   But
1372                  * that's undesirable in laptop mode, where we *want* lumpy
1373                  * writeout.  So in laptop mode, write out the whole world.
1374                  */
1375                 if (total_scanned > sc->swap_cluster_max +
1376                                         sc->swap_cluster_max / 2) {
1377                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1378                         sc->may_writepage = 1;
1379                 }
1380
1381                 /* Take a nap, wait for some writeback to complete */
1382                 if (sc->nr_scanned && priority < DEF_PRIORITY - 2 &&
1383                                 sc->nr_io_pages > sc->swap_cluster_max)
1384                         congestion_wait(WRITE, HZ/10);
1385         }
1386         /* top priority shrink_caches still had more to do? don't OOM, then */
1387         if (!sc->all_unreclaimable && scan_global_lru(sc))
1388                 ret = 1;
1389 out:
1390         /*
1391          * Now that we've scanned all the zones at this priority level, note
1392          * that level within the zone so that the next thread which performs
1393          * scanning of this zone will immediately start out at this priority
1394          * level.  This affects only the decision whether or not to bring
1395          * mapped pages onto the inactive list.
1396          */
1397         if (priority < 0)
1398                 priority = 0;
1399
1400         if (scan_global_lru(sc)) {
1401                 for (i = 0; zones[i] != NULL; i++) {
1402                         struct zone *zone = zones[i];
1403
1404                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1405                                 continue;
1406
1407                         zone->prev_priority = priority;
1408                 }
1409         } else
1410                 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1411
1412         return ret;
1413 }
1414
1415 unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1416 {
1417         struct scan_control sc = {
1418                 .gfp_mask = gfp_mask,
1419                 .may_writepage = !laptop_mode,
1420                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1421                 .may_swap = 1,
1422                 .swappiness = vm_swappiness,
1423                 .order = order,
1424                 .mem_cgroup = NULL,
1425                 .isolate_pages = isolate_pages_global,
1426         };
1427
1428         return do_try_to_free_pages(zones, gfp_mask, &sc);
1429 }
1430
1431 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1432
1433 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1434                                                 gfp_t gfp_mask)
1435 {
1436         struct scan_control sc = {
1437                 .gfp_mask = gfp_mask,
1438                 .may_writepage = !laptop_mode,
1439                 .may_swap = 1,
1440                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1441                 .swappiness = vm_swappiness,
1442                 .order = 0,
1443                 .mem_cgroup = mem_cont,
1444                 .isolate_pages = mem_cgroup_isolate_pages,
1445         };
1446         struct zone **zones;
1447         int target_zone = gfp_zone(GFP_HIGHUSER_MOVABLE);
1448
1449         zones = NODE_DATA(numa_node_id())->node_zonelists[target_zone].zones;
1450         if (do_try_to_free_pages(zones, sc.gfp_mask, &sc))
1451                 return 1;
1452         return 0;
1453 }
1454 #endif
1455
1456 /*
1457  * For kswapd, balance_pgdat() will work across all this node's zones until
1458  * they are all at pages_high.
1459  *
1460  * Returns the number of pages which were actually freed.
1461  *
1462  * There is special handling here for zones which are full of pinned pages.
1463  * This can happen if the pages are all mlocked, or if they are all used by
1464  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1465  * What we do is to detect the case where all pages in the zone have been
1466  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1467  * dead and from now on, only perform a short scan.  Basically we're polling
1468  * the zone for when the problem goes away.
1469  *
1470  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1471  * zones which have free_pages > pages_high, but once a zone is found to have
1472  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1473  * of the number of free pages in the lower zones.  This interoperates with
1474  * the page allocator fallback scheme to ensure that aging of pages is balanced
1475  * across the zones.
1476  */
1477 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1478 {
1479         int all_zones_ok;
1480         int priority;
1481         int i;
1482         unsigned long total_scanned;
1483         unsigned long nr_reclaimed;
1484         struct reclaim_state *reclaim_state = current->reclaim_state;
1485         struct scan_control sc = {
1486                 .gfp_mask = GFP_KERNEL,
1487                 .may_swap = 1,
1488                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1489                 .swappiness = vm_swappiness,
1490                 .order = order,
1491                 .mem_cgroup = NULL,
1492                 .isolate_pages = isolate_pages_global,
1493         };
1494         /*
1495          * temp_priority is used to remember the scanning priority at which
1496          * this zone was successfully refilled to free_pages == pages_high.
1497          */
1498         int temp_priority[MAX_NR_ZONES];
1499
1500 loop_again:
1501         total_scanned = 0;
1502         nr_reclaimed = 0;
1503         sc.may_writepage = !laptop_mode;
1504         count_vm_event(PAGEOUTRUN);
1505
1506         for (i = 0; i < pgdat->nr_zones; i++)
1507                 temp_priority[i] = DEF_PRIORITY;
1508
1509         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1510                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1511                 unsigned long lru_pages = 0;
1512
1513                 /* The swap token gets in the way of swapout... */
1514                 if (!priority)
1515                         disable_swap_token();
1516
1517                 sc.nr_io_pages = 0;
1518                 all_zones_ok = 1;
1519
1520                 /*
1521                  * Scan in the highmem->dma direction for the highest
1522                  * zone which needs scanning
1523                  */
1524                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1525                         struct zone *zone = pgdat->node_zones + i;
1526
1527                         if (!populated_zone(zone))
1528                                 continue;
1529
1530                         if (zone_is_all_unreclaimable(zone) &&
1531                             priority != DEF_PRIORITY)
1532                                 continue;
1533
1534                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1535                                                0, 0)) {
1536                                 end_zone = i;
1537                                 break;
1538                         }
1539                 }
1540                 if (i < 0)
1541                         goto out;
1542
1543                 for (i = 0; i <= end_zone; i++) {
1544                         struct zone *zone = pgdat->node_zones + i;
1545
1546                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1547                                         + zone_page_state(zone, NR_INACTIVE);
1548                 }
1549
1550                 /*
1551                  * Now scan the zone in the dma->highmem direction, stopping
1552                  * at the last zone which needs scanning.
1553                  *
1554                  * We do this because the page allocator works in the opposite
1555                  * direction.  This prevents the page allocator from allocating
1556                  * pages behind kswapd's direction of progress, which would
1557                  * cause too much scanning of the lower zones.
1558                  */
1559                 for (i = 0; i <= end_zone; i++) {
1560                         struct zone *zone = pgdat->node_zones + i;
1561                         int nr_slab;
1562
1563                         if (!populated_zone(zone))
1564                                 continue;
1565
1566                         if (zone_is_all_unreclaimable(zone) &&
1567                                         priority != DEF_PRIORITY)
1568                                 continue;
1569
1570                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1571                                                end_zone, 0))
1572                                 all_zones_ok = 0;
1573                         temp_priority[i] = priority;
1574                         sc.nr_scanned = 0;
1575                         note_zone_scanning_priority(zone, priority);
1576                         /*
1577                          * We put equal pressure on every zone, unless one
1578                          * zone has way too many pages free already.
1579                          */
1580                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1581                                                 end_zone, 0))
1582                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1583                         reclaim_state->reclaimed_slab = 0;
1584                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1585                                                 lru_pages);
1586                         nr_reclaimed += reclaim_state->reclaimed_slab;
1587                         total_scanned += sc.nr_scanned;
1588                         if (zone_is_all_unreclaimable(zone))
1589                                 continue;
1590                         if (nr_slab == 0 && zone->pages_scanned >=
1591                                 (zone_page_state(zone, NR_ACTIVE)
1592                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1593                                         zone_set_flag(zone,
1594                                                       ZONE_ALL_UNRECLAIMABLE);
1595                         /*
1596                          * If we've done a decent amount of scanning and
1597                          * the reclaim ratio is low, start doing writepage
1598                          * even in laptop mode
1599                          */
1600                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1601                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1602                                 sc.may_writepage = 1;
1603                 }
1604                 if (all_zones_ok)
1605                         break;          /* kswapd: all done */
1606                 /*
1607                  * OK, kswapd is getting into trouble.  Take a nap, then take
1608                  * another pass across the zones.
1609                  */
1610                 if (total_scanned && priority < DEF_PRIORITY - 2 &&
1611                                         sc.nr_io_pages > sc.swap_cluster_max)
1612                         congestion_wait(WRITE, HZ/10);
1613
1614                 /*
1615                  * We do this so kswapd doesn't build up large priorities for
1616                  * example when it is freeing in parallel with allocators. It
1617                  * matches the direct reclaim path behaviour in terms of impact
1618                  * on zone->*_priority.
1619                  */
1620                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1621                         break;
1622         }
1623 out:
1624         /*
1625          * Note within each zone the priority level at which this zone was
1626          * brought into a happy state.  So that the next thread which scans this
1627          * zone will start out at that priority level.
1628          */
1629         for (i = 0; i < pgdat->nr_zones; i++) {
1630                 struct zone *zone = pgdat->node_zones + i;
1631
1632                 zone->prev_priority = temp_priority[i];
1633         }
1634         if (!all_zones_ok) {
1635                 cond_resched();
1636
1637                 try_to_freeze();
1638
1639                 goto loop_again;
1640         }
1641
1642         return nr_reclaimed;
1643 }
1644
1645 /*
1646  * The background pageout daemon, started as a kernel thread
1647  * from the init process. 
1648  *
1649  * This basically trickles out pages so that we have _some_
1650  * free memory available even if there is no other activity
1651  * that frees anything up. This is needed for things like routing
1652  * etc, where we otherwise might have all activity going on in
1653  * asynchronous contexts that cannot page things out.
1654  *
1655  * If there are applications that are active memory-allocators
1656  * (most normal use), this basically shouldn't matter.
1657  */
1658 static int kswapd(void *p)
1659 {
1660         unsigned long order;
1661         pg_data_t *pgdat = (pg_data_t*)p;
1662         struct task_struct *tsk = current;
1663         DEFINE_WAIT(wait);
1664         struct reclaim_state reclaim_state = {
1665                 .reclaimed_slab = 0,
1666         };
1667         cpumask_t cpumask;
1668
1669         cpumask = node_to_cpumask(pgdat->node_id);
1670         if (!cpus_empty(cpumask))
1671                 set_cpus_allowed(tsk, cpumask);
1672         current->reclaim_state = &reclaim_state;
1673
1674         /*
1675          * Tell the memory management that we're a "memory allocator",
1676          * and that if we need more memory we should get access to it
1677          * regardless (see "__alloc_pages()"). "kswapd" should
1678          * never get caught in the normal page freeing logic.
1679          *
1680          * (Kswapd normally doesn't need memory anyway, but sometimes
1681          * you need a small amount of memory in order to be able to
1682          * page out something else, and this flag essentially protects
1683          * us from recursively trying to free more memory as we're
1684          * trying to free the first piece of memory in the first place).
1685          */
1686         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1687         set_freezable();
1688
1689         order = 0;
1690         for ( ; ; ) {
1691                 unsigned long new_order;
1692
1693                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1694                 new_order = pgdat->kswapd_max_order;
1695                 pgdat->kswapd_max_order = 0;
1696                 if (order < new_order) {
1697                         /*
1698                          * Don't sleep if someone wants a larger 'order'
1699                          * allocation
1700                          */
1701                         order = new_order;
1702                 } else {
1703                         if (!freezing(current))
1704                                 schedule();
1705
1706                         order = pgdat->kswapd_max_order;
1707                 }
1708                 finish_wait(&pgdat->kswapd_wait, &wait);
1709
1710                 if (!try_to_freeze()) {
1711                         /* We can speed up thawing tasks if we don't call
1712                          * balance_pgdat after returning from the refrigerator
1713                          */
1714                         balance_pgdat(pgdat, order);
1715                 }
1716         }
1717         return 0;
1718 }
1719
1720 /*
1721  * A zone is low on free memory, so wake its kswapd task to service it.
1722  */
1723 void wakeup_kswapd(struct zone *zone, int order)
1724 {
1725         pg_data_t *pgdat;
1726
1727         if (!populated_zone(zone))
1728                 return;
1729
1730         pgdat = zone->zone_pgdat;
1731         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1732                 return;
1733         if (pgdat->kswapd_max_order < order)
1734                 pgdat->kswapd_max_order = order;
1735         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1736                 return;
1737         if (!waitqueue_active(&pgdat->kswapd_wait))
1738                 return;
1739         wake_up_interruptible(&pgdat->kswapd_wait);
1740 }
1741
1742 #ifdef CONFIG_PM
1743 /*
1744  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1745  * from LRU lists system-wide, for given pass and priority, and returns the
1746  * number of reclaimed pages
1747  *
1748  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1749  */
1750 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1751                                       int pass, struct scan_control *sc)
1752 {
1753         struct zone *zone;
1754         unsigned long nr_to_scan, ret = 0;
1755
1756         for_each_zone(zone) {
1757
1758                 if (!populated_zone(zone))
1759                         continue;
1760
1761                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1762                         continue;
1763
1764                 /* For pass = 0 we don't shrink the active list */
1765                 if (pass > 0) {
1766                         zone->nr_scan_active +=
1767                                 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1768                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1769                                 zone->nr_scan_active = 0;
1770                                 nr_to_scan = min(nr_pages,
1771                                         zone_page_state(zone, NR_ACTIVE));
1772                                 shrink_active_list(nr_to_scan, zone, sc, prio);
1773                         }
1774                 }
1775
1776                 zone->nr_scan_inactive +=
1777                         (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1778                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1779                         zone->nr_scan_inactive = 0;
1780                         nr_to_scan = min(nr_pages,
1781                                 zone_page_state(zone, NR_INACTIVE));
1782                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1783                         if (ret >= nr_pages)
1784                                 return ret;
1785                 }
1786         }
1787
1788         return ret;
1789 }
1790
1791 static unsigned long count_lru_pages(void)
1792 {
1793         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1794 }
1795
1796 /*
1797  * Try to free `nr_pages' of memory, system-wide, and return the number of
1798  * freed pages.
1799  *
1800  * Rather than trying to age LRUs the aim is to preserve the overall
1801  * LRU order by reclaiming preferentially
1802  * inactive > active > active referenced > active mapped
1803  */
1804 unsigned long shrink_all_memory(unsigned long nr_pages)
1805 {
1806         unsigned long lru_pages, nr_slab;
1807         unsigned long ret = 0;
1808         int pass;
1809         struct reclaim_state reclaim_state;
1810         struct scan_control sc = {
1811                 .gfp_mask = GFP_KERNEL,
1812                 .may_swap = 0,
1813                 .swap_cluster_max = nr_pages,
1814                 .may_writepage = 1,
1815                 .swappiness = vm_swappiness,
1816                 .isolate_pages = isolate_pages_global,
1817         };
1818
1819         current->reclaim_state = &reclaim_state;
1820
1821         lru_pages = count_lru_pages();
1822         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1823         /* If slab caches are huge, it's better to hit them first */
1824         while (nr_slab >= lru_pages) {
1825                 reclaim_state.reclaimed_slab = 0;
1826                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1827                 if (!reclaim_state.reclaimed_slab)
1828                         break;
1829
1830                 ret += reclaim_state.reclaimed_slab;
1831                 if (ret >= nr_pages)
1832                         goto out;
1833
1834                 nr_slab -= reclaim_state.reclaimed_slab;
1835         }
1836
1837         /*
1838          * We try to shrink LRUs in 5 passes:
1839          * 0 = Reclaim from inactive_list only
1840          * 1 = Reclaim from active list but don't reclaim mapped
1841          * 2 = 2nd pass of type 1
1842          * 3 = Reclaim mapped (normal reclaim)
1843          * 4 = 2nd pass of type 3
1844          */
1845         for (pass = 0; pass < 5; pass++) {
1846                 int prio;
1847
1848                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1849                 if (pass > 2) {
1850                         sc.may_swap = 1;
1851                         sc.swappiness = 100;
1852                 }
1853
1854                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1855                         unsigned long nr_to_scan = nr_pages - ret;
1856
1857                         sc.nr_scanned = 0;
1858                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1859                         if (ret >= nr_pages)
1860                                 goto out;
1861
1862                         reclaim_state.reclaimed_slab = 0;
1863                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1864                                         count_lru_pages());
1865                         ret += reclaim_state.reclaimed_slab;
1866                         if (ret >= nr_pages)
1867                                 goto out;
1868
1869                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1870                                 congestion_wait(WRITE, HZ / 10);
1871                 }
1872         }
1873
1874         /*
1875          * If ret = 0, we could not shrink LRUs, but there may be something
1876          * in slab caches
1877          */
1878         if (!ret) {
1879                 do {
1880                         reclaim_state.reclaimed_slab = 0;
1881                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1882                         ret += reclaim_state.reclaimed_slab;
1883                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1884         }
1885
1886 out:
1887         current->reclaim_state = NULL;
1888
1889         return ret;
1890 }
1891 #endif
1892
1893 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1894    not required for correctness.  So if the last cpu in a node goes
1895    away, we get changed to run anywhere: as the first one comes back,
1896    restore their cpu bindings. */
1897 static int __devinit cpu_callback(struct notifier_block *nfb,
1898                                   unsigned long action, void *hcpu)
1899 {
1900         pg_data_t *pgdat;
1901         cpumask_t mask;
1902         int nid;
1903
1904         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1905                 for_each_node_state(nid, N_HIGH_MEMORY) {
1906                         pgdat = NODE_DATA(nid);
1907                         mask = node_to_cpumask(pgdat->node_id);
1908                         if (any_online_cpu(mask) != NR_CPUS)
1909                                 /* One of our CPUs online: restore mask */
1910                                 set_cpus_allowed(pgdat->kswapd, mask);
1911                 }
1912         }
1913         return NOTIFY_OK;
1914 }
1915
1916 /*
1917  * This kswapd start function will be called by init and node-hot-add.
1918  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1919  */
1920 int kswapd_run(int nid)
1921 {
1922         pg_data_t *pgdat = NODE_DATA(nid);
1923         int ret = 0;
1924
1925         if (pgdat->kswapd)
1926                 return 0;
1927
1928         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1929         if (IS_ERR(pgdat->kswapd)) {
1930                 /* failure at boot is fatal */
1931                 BUG_ON(system_state == SYSTEM_BOOTING);
1932                 printk("Failed to start kswapd on node %d\n",nid);
1933                 ret = -1;
1934         }
1935         return ret;
1936 }
1937
1938 static int __init kswapd_init(void)
1939 {
1940         int nid;
1941
1942         swap_setup();
1943         for_each_node_state(nid, N_HIGH_MEMORY)
1944                 kswapd_run(nid);
1945         hotcpu_notifier(cpu_callback, 0);
1946         return 0;
1947 }
1948
1949 module_init(kswapd_init)
1950
1951 #ifdef CONFIG_NUMA
1952 /*
1953  * Zone reclaim mode
1954  *
1955  * If non-zero call zone_reclaim when the number of free pages falls below
1956  * the watermarks.
1957  */
1958 int zone_reclaim_mode __read_mostly;
1959
1960 #define RECLAIM_OFF 0
1961 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1962 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1963 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1964
1965 /*
1966  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1967  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1968  * a zone.
1969  */
1970 #define ZONE_RECLAIM_PRIORITY 4
1971
1972 /*
1973  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1974  * occur.
1975  */
1976 int sysctl_min_unmapped_ratio = 1;
1977
1978 /*
1979  * If the number of slab pages in a zone grows beyond this percentage then
1980  * slab reclaim needs to occur.
1981  */
1982 int sysctl_min_slab_ratio = 5;
1983
1984 /*
1985  * Try to free up some pages from this zone through reclaim.
1986  */
1987 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1988 {
1989         /* Minimum pages needed in order to stay on node */
1990         const unsigned long nr_pages = 1 << order;
1991         struct task_struct *p = current;
1992         struct reclaim_state reclaim_state;
1993         int priority;
1994         unsigned long nr_reclaimed = 0;
1995         struct scan_control sc = {
1996                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1997                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1998                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1999                                         SWAP_CLUSTER_MAX),
2000                 .gfp_mask = gfp_mask,
2001                 .swappiness = vm_swappiness,
2002                 .isolate_pages = isolate_pages_global,
2003         };
2004         unsigned long slab_reclaimable;
2005
2006         disable_swap_token();
2007         cond_resched();
2008         /*
2009          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2010          * and we also need to be able to write out pages for RECLAIM_WRITE
2011          * and RECLAIM_SWAP.
2012          */
2013         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2014         reclaim_state.reclaimed_slab = 0;
2015         p->reclaim_state = &reclaim_state;
2016
2017         if (zone_page_state(zone, NR_FILE_PAGES) -
2018                 zone_page_state(zone, NR_FILE_MAPPED) >
2019                 zone->min_unmapped_pages) {
2020                 /*
2021                  * Free memory by calling shrink zone with increasing
2022                  * priorities until we have enough memory freed.
2023                  */
2024                 priority = ZONE_RECLAIM_PRIORITY;
2025                 do {
2026                         note_zone_scanning_priority(zone, priority);
2027                         nr_reclaimed += shrink_zone(priority, zone, &sc);
2028                         priority--;
2029                 } while (priority >= 0 && nr_reclaimed < nr_pages);
2030         }
2031
2032         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2033         if (slab_reclaimable > zone->min_slab_pages) {
2034                 /*
2035                  * shrink_slab() does not currently allow us to determine how
2036                  * many pages were freed in this zone. So we take the current
2037                  * number of slab pages and shake the slab until it is reduced
2038                  * by the same nr_pages that we used for reclaiming unmapped
2039                  * pages.
2040                  *
2041                  * Note that shrink_slab will free memory on all zones and may
2042                  * take a long time.
2043                  */
2044                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2045                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2046                                 slab_reclaimable - nr_pages)
2047                         ;
2048
2049                 /*
2050                  * Update nr_reclaimed by the number of slab pages we
2051                  * reclaimed from this zone.
2052                  */
2053                 nr_reclaimed += slab_reclaimable -
2054                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2055         }
2056
2057         p->reclaim_state = NULL;
2058         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2059         return nr_reclaimed >= nr_pages;
2060 }
2061
2062 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2063 {
2064         int node_id;
2065         int ret;
2066
2067         /*
2068          * Zone reclaim reclaims unmapped file backed pages and
2069          * slab pages if we are over the defined limits.
2070          *
2071          * A small portion of unmapped file backed pages is needed for
2072          * file I/O otherwise pages read by file I/O will be immediately
2073          * thrown out if the zone is overallocated. So we do not reclaim
2074          * if less than a specified percentage of the zone is used by
2075          * unmapped file backed pages.
2076          */
2077         if (zone_page_state(zone, NR_FILE_PAGES) -
2078             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2079             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2080                         <= zone->min_slab_pages)
2081                 return 0;
2082
2083         if (zone_is_all_unreclaimable(zone))
2084                 return 0;
2085
2086         /*
2087          * Do not scan if the allocation should not be delayed.
2088          */
2089         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2090                         return 0;
2091
2092         /*
2093          * Only run zone reclaim on the local zone or on zones that do not
2094          * have associated processors. This will favor the local processor
2095          * over remote processors and spread off node memory allocations
2096          * as wide as possible.
2097          */
2098         node_id = zone_to_nid(zone);
2099         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2100                 return 0;
2101
2102         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2103                 return 0;
2104         ret = __zone_reclaim(zone, gfp_mask, order);
2105         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2106
2107         return ret;
2108 }
2109 #endif