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