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