[PATCH] Direct Migration V9: remove_from_swap() to remove swap ptes
[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/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h>  /* for try_to_release_page(),
26                                         buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
36
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
39
40 #include <linux/swapops.h>
41
42 /* possible outcome of pageout() */
43 typedef enum {
44         /* failed to write page out, page is locked */
45         PAGE_KEEP,
46         /* move page to the active list, page is locked */
47         PAGE_ACTIVATE,
48         /* page has been sent to the disk successfully, page is unlocked */
49         PAGE_SUCCESS,
50         /* page is clean and locked */
51         PAGE_CLEAN,
52 } pageout_t;
53
54 struct scan_control {
55         /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
56         unsigned long nr_to_scan;
57
58         /* Incremented by the number of inactive pages that were scanned */
59         unsigned long nr_scanned;
60
61         /* Incremented by the number of pages reclaimed */
62         unsigned long nr_reclaimed;
63
64         unsigned long nr_mapped;        /* From page_state */
65
66         /* Ask shrink_caches, or shrink_zone to scan at this priority */
67         unsigned int priority;
68
69         /* This context's GFP mask */
70         gfp_t gfp_mask;
71
72         int may_writepage;
73
74         /* Can pages be swapped as part of reclaim? */
75         int may_swap;
76
77         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
78          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
79          * In this context, it doesn't matter that we scan the
80          * whole list at once. */
81         int swap_cluster_max;
82 };
83
84 /*
85  * The list of shrinker callbacks used by to apply pressure to
86  * ageable caches.
87  */
88 struct shrinker {
89         shrinker_t              shrinker;
90         struct list_head        list;
91         int                     seeks;  /* seeks to recreate an obj */
92         long                    nr;     /* objs pending delete */
93 };
94
95 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
96
97 #ifdef ARCH_HAS_PREFETCH
98 #define prefetch_prev_lru_page(_page, _base, _field)                    \
99         do {                                                            \
100                 if ((_page)->lru.prev != _base) {                       \
101                         struct page *prev;                              \
102                                                                         \
103                         prev = lru_to_page(&(_page->lru));              \
104                         prefetch(&prev->_field);                        \
105                 }                                                       \
106         } while (0)
107 #else
108 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
109 #endif
110
111 #ifdef ARCH_HAS_PREFETCHW
112 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
113         do {                                                            \
114                 if ((_page)->lru.prev != _base) {                       \
115                         struct page *prev;                              \
116                                                                         \
117                         prev = lru_to_page(&(_page->lru));              \
118                         prefetchw(&prev->_field);                       \
119                 }                                                       \
120         } while (0)
121 #else
122 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
123 #endif
124
125 /*
126  * From 0 .. 100.  Higher means more swappy.
127  */
128 int vm_swappiness = 60;
129 static long total_memory;
130
131 static LIST_HEAD(shrinker_list);
132 static DECLARE_RWSEM(shrinker_rwsem);
133
134 /*
135  * Add a shrinker callback to be called from the vm
136  */
137 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
138 {
139         struct shrinker *shrinker;
140
141         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
142         if (shrinker) {
143                 shrinker->shrinker = theshrinker;
144                 shrinker->seeks = seeks;
145                 shrinker->nr = 0;
146                 down_write(&shrinker_rwsem);
147                 list_add_tail(&shrinker->list, &shrinker_list);
148                 up_write(&shrinker_rwsem);
149         }
150         return shrinker;
151 }
152 EXPORT_SYMBOL(set_shrinker);
153
154 /*
155  * Remove one
156  */
157 void remove_shrinker(struct shrinker *shrinker)
158 {
159         down_write(&shrinker_rwsem);
160         list_del(&shrinker->list);
161         up_write(&shrinker_rwsem);
162         kfree(shrinker);
163 }
164 EXPORT_SYMBOL(remove_shrinker);
165
166 #define SHRINK_BATCH 128
167 /*
168  * Call the shrink functions to age shrinkable caches
169  *
170  * Here we assume it costs one seek to replace a lru page and that it also
171  * takes a seek to recreate a cache object.  With this in mind we age equal
172  * percentages of the lru and ageable caches.  This should balance the seeks
173  * generated by these structures.
174  *
175  * If the vm encounted mapped pages on the LRU it increase the pressure on
176  * slab to avoid swapping.
177  *
178  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
179  *
180  * `lru_pages' represents the number of on-LRU pages in all the zones which
181  * are eligible for the caller's allocation attempt.  It is used for balancing
182  * slab reclaim versus page reclaim.
183  *
184  * Returns the number of slab objects which we shrunk.
185  */
186 int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages)
187 {
188         struct shrinker *shrinker;
189         int ret = 0;
190
191         if (scanned == 0)
192                 scanned = SWAP_CLUSTER_MAX;
193
194         if (!down_read_trylock(&shrinker_rwsem))
195                 return 1;       /* Assume we'll be able to shrink next time */
196
197         list_for_each_entry(shrinker, &shrinker_list, list) {
198                 unsigned long long delta;
199                 unsigned long total_scan;
200                 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask);
201
202                 delta = (4 * scanned) / shrinker->seeks;
203                 delta *= max_pass;
204                 do_div(delta, lru_pages + 1);
205                 shrinker->nr += delta;
206                 if (shrinker->nr < 0) {
207                         printk(KERN_ERR "%s: nr=%ld\n",
208                                         __FUNCTION__, shrinker->nr);
209                         shrinker->nr = max_pass;
210                 }
211
212                 /*
213                  * Avoid risking looping forever due to too large nr value:
214                  * never try to free more than twice the estimate number of
215                  * freeable entries.
216                  */
217                 if (shrinker->nr > max_pass * 2)
218                         shrinker->nr = max_pass * 2;
219
220                 total_scan = shrinker->nr;
221                 shrinker->nr = 0;
222
223                 while (total_scan >= SHRINK_BATCH) {
224                         long this_scan = SHRINK_BATCH;
225                         int shrink_ret;
226                         int nr_before;
227
228                         nr_before = (*shrinker->shrinker)(0, gfp_mask);
229                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
230                         if (shrink_ret == -1)
231                                 break;
232                         if (shrink_ret < nr_before)
233                                 ret += nr_before - shrink_ret;
234                         mod_page_state(slabs_scanned, this_scan);
235                         total_scan -= this_scan;
236
237                         cond_resched();
238                 }
239
240                 shrinker->nr += total_scan;
241         }
242         up_read(&shrinker_rwsem);
243         return ret;
244 }
245
246 /* Called without lock on whether page is mapped, so answer is unstable */
247 static inline int page_mapping_inuse(struct page *page)
248 {
249         struct address_space *mapping;
250
251         /* Page is in somebody's page tables. */
252         if (page_mapped(page))
253                 return 1;
254
255         /* Be more reluctant to reclaim swapcache than pagecache */
256         if (PageSwapCache(page))
257                 return 1;
258
259         mapping = page_mapping(page);
260         if (!mapping)
261                 return 0;
262
263         /* File is mmap'd by somebody? */
264         return mapping_mapped(mapping);
265 }
266
267 static inline int is_page_cache_freeable(struct page *page)
268 {
269         return page_count(page) - !!PagePrivate(page) == 2;
270 }
271
272 static int may_write_to_queue(struct backing_dev_info *bdi)
273 {
274         if (current->flags & PF_SWAPWRITE)
275                 return 1;
276         if (!bdi_write_congested(bdi))
277                 return 1;
278         if (bdi == current->backing_dev_info)
279                 return 1;
280         return 0;
281 }
282
283 /*
284  * We detected a synchronous write error writing a page out.  Probably
285  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
286  * fsync(), msync() or close().
287  *
288  * The tricky part is that after writepage we cannot touch the mapping: nothing
289  * prevents it from being freed up.  But we have a ref on the page and once
290  * that page is locked, the mapping is pinned.
291  *
292  * We're allowed to run sleeping lock_page() here because we know the caller has
293  * __GFP_FS.
294  */
295 static void handle_write_error(struct address_space *mapping,
296                                 struct page *page, int error)
297 {
298         lock_page(page);
299         if (page_mapping(page) == mapping) {
300                 if (error == -ENOSPC)
301                         set_bit(AS_ENOSPC, &mapping->flags);
302                 else
303                         set_bit(AS_EIO, &mapping->flags);
304         }
305         unlock_page(page);
306 }
307
308 /*
309  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
310  */
311 static pageout_t pageout(struct page *page, struct address_space *mapping)
312 {
313         /*
314          * If the page is dirty, only perform writeback if that write
315          * will be non-blocking.  To prevent this allocation from being
316          * stalled by pagecache activity.  But note that there may be
317          * stalls if we need to run get_block().  We could test
318          * PagePrivate for that.
319          *
320          * If this process is currently in generic_file_write() against
321          * this page's queue, we can perform writeback even if that
322          * will block.
323          *
324          * If the page is swapcache, write it back even if that would
325          * block, for some throttling. This happens by accident, because
326          * swap_backing_dev_info is bust: it doesn't reflect the
327          * congestion state of the swapdevs.  Easy to fix, if needed.
328          * See swapfile.c:page_queue_congested().
329          */
330         if (!is_page_cache_freeable(page))
331                 return PAGE_KEEP;
332         if (!mapping) {
333                 /*
334                  * Some data journaling orphaned pages can have
335                  * page->mapping == NULL while being dirty with clean buffers.
336                  */
337                 if (PagePrivate(page)) {
338                         if (try_to_free_buffers(page)) {
339                                 ClearPageDirty(page);
340                                 printk("%s: orphaned page\n", __FUNCTION__);
341                                 return PAGE_CLEAN;
342                         }
343                 }
344                 return PAGE_KEEP;
345         }
346         if (mapping->a_ops->writepage == NULL)
347                 return PAGE_ACTIVATE;
348         if (!may_write_to_queue(mapping->backing_dev_info))
349                 return PAGE_KEEP;
350
351         if (clear_page_dirty_for_io(page)) {
352                 int res;
353                 struct writeback_control wbc = {
354                         .sync_mode = WB_SYNC_NONE,
355                         .nr_to_write = SWAP_CLUSTER_MAX,
356                         .nonblocking = 1,
357                         .for_reclaim = 1,
358                 };
359
360                 SetPageReclaim(page);
361                 res = mapping->a_ops->writepage(page, &wbc);
362                 if (res < 0)
363                         handle_write_error(mapping, page, res);
364                 if (res == AOP_WRITEPAGE_ACTIVATE) {
365                         ClearPageReclaim(page);
366                         return PAGE_ACTIVATE;
367                 }
368                 if (!PageWriteback(page)) {
369                         /* synchronous write or broken a_ops? */
370                         ClearPageReclaim(page);
371                 }
372
373                 return PAGE_SUCCESS;
374         }
375
376         return PAGE_CLEAN;
377 }
378
379 static int remove_mapping(struct address_space *mapping, struct page *page)
380 {
381         if (!mapping)
382                 return 0;               /* truncate got there first */
383
384         write_lock_irq(&mapping->tree_lock);
385
386         /*
387          * The non-racy check for busy page.  It is critical to check
388          * PageDirty _after_ making sure that the page is freeable and
389          * not in use by anybody.       (pagecache + us == 2)
390          */
391         if (unlikely(page_count(page) != 2))
392                 goto cannot_free;
393         smp_rmb();
394         if (unlikely(PageDirty(page)))
395                 goto cannot_free;
396
397         if (PageSwapCache(page)) {
398                 swp_entry_t swap = { .val = page_private(page) };
399                 __delete_from_swap_cache(page);
400                 write_unlock_irq(&mapping->tree_lock);
401                 swap_free(swap);
402                 __put_page(page);       /* The pagecache ref */
403                 return 1;
404         }
405
406         __remove_from_page_cache(page);
407         write_unlock_irq(&mapping->tree_lock);
408         __put_page(page);
409         return 1;
410
411 cannot_free:
412         write_unlock_irq(&mapping->tree_lock);
413         return 0;
414 }
415
416 /*
417  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
418  */
419 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
420 {
421         LIST_HEAD(ret_pages);
422         struct pagevec freed_pvec;
423         int pgactivate = 0;
424         int reclaimed = 0;
425
426         cond_resched();
427
428         pagevec_init(&freed_pvec, 1);
429         while (!list_empty(page_list)) {
430                 struct address_space *mapping;
431                 struct page *page;
432                 int may_enter_fs;
433                 int referenced;
434
435                 cond_resched();
436
437                 page = lru_to_page(page_list);
438                 list_del(&page->lru);
439
440                 if (TestSetPageLocked(page))
441                         goto keep;
442
443                 BUG_ON(PageActive(page));
444
445                 sc->nr_scanned++;
446                 /* Double the slab pressure for mapped and swapcache pages */
447                 if (page_mapped(page) || PageSwapCache(page))
448                         sc->nr_scanned++;
449
450                 if (PageWriteback(page))
451                         goto keep_locked;
452
453                 referenced = page_referenced(page, 1);
454                 /* In active use or really unfreeable?  Activate it. */
455                 if (referenced && page_mapping_inuse(page))
456                         goto activate_locked;
457
458 #ifdef CONFIG_SWAP
459                 /*
460                  * Anonymous process memory has backing store?
461                  * Try to allocate it some swap space here.
462                  */
463                 if (PageAnon(page) && !PageSwapCache(page)) {
464                         if (!sc->may_swap)
465                                 goto keep_locked;
466                         if (!add_to_swap(page, GFP_ATOMIC))
467                                 goto activate_locked;
468                 }
469 #endif /* CONFIG_SWAP */
470
471                 mapping = page_mapping(page);
472                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
473                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
474
475                 /*
476                  * The page is mapped into the page tables of one or more
477                  * processes. Try to unmap it here.
478                  */
479                 if (page_mapped(page) && mapping) {
480                         /*
481                          * No unmapping if we do not swap
482                          */
483                         if (!sc->may_swap)
484                                 goto keep_locked;
485
486                         switch (try_to_unmap(page, 0)) {
487                         case SWAP_FAIL:
488                                 goto activate_locked;
489                         case SWAP_AGAIN:
490                                 goto keep_locked;
491                         case SWAP_SUCCESS:
492                                 ; /* try to free the page below */
493                         }
494                 }
495
496                 if (PageDirty(page)) {
497                         if (referenced)
498                                 goto keep_locked;
499                         if (!may_enter_fs)
500                                 goto keep_locked;
501                         if (!sc->may_writepage)
502                                 goto keep_locked;
503
504                         /* Page is dirty, try to write it out here */
505                         switch(pageout(page, mapping)) {
506                         case PAGE_KEEP:
507                                 goto keep_locked;
508                         case PAGE_ACTIVATE:
509                                 goto activate_locked;
510                         case PAGE_SUCCESS:
511                                 if (PageWriteback(page) || PageDirty(page))
512                                         goto keep;
513                                 /*
514                                  * A synchronous write - probably a ramdisk.  Go
515                                  * ahead and try to reclaim the page.
516                                  */
517                                 if (TestSetPageLocked(page))
518                                         goto keep;
519                                 if (PageDirty(page) || PageWriteback(page))
520                                         goto keep_locked;
521                                 mapping = page_mapping(page);
522                         case PAGE_CLEAN:
523                                 ; /* try to free the page below */
524                         }
525                 }
526
527                 /*
528                  * If the page has buffers, try to free the buffer mappings
529                  * associated with this page. If we succeed we try to free
530                  * the page as well.
531                  *
532                  * We do this even if the page is PageDirty().
533                  * try_to_release_page() does not perform I/O, but it is
534                  * possible for a page to have PageDirty set, but it is actually
535                  * clean (all its buffers are clean).  This happens if the
536                  * buffers were written out directly, with submit_bh(). ext3
537                  * will do this, as well as the blockdev mapping. 
538                  * try_to_release_page() will discover that cleanness and will
539                  * drop the buffers and mark the page clean - it can be freed.
540                  *
541                  * Rarely, pages can have buffers and no ->mapping.  These are
542                  * the pages which were not successfully invalidated in
543                  * truncate_complete_page().  We try to drop those buffers here
544                  * and if that worked, and the page is no longer mapped into
545                  * process address space (page_count == 1) it can be freed.
546                  * Otherwise, leave the page on the LRU so it is swappable.
547                  */
548                 if (PagePrivate(page)) {
549                         if (!try_to_release_page(page, sc->gfp_mask))
550                                 goto activate_locked;
551                         if (!mapping && page_count(page) == 1)
552                                 goto free_it;
553                 }
554
555                 if (!remove_mapping(mapping, page))
556                         goto keep_locked;
557
558 free_it:
559                 unlock_page(page);
560                 reclaimed++;
561                 if (!pagevec_add(&freed_pvec, page))
562                         __pagevec_release_nonlru(&freed_pvec);
563                 continue;
564
565 activate_locked:
566                 SetPageActive(page);
567                 pgactivate++;
568 keep_locked:
569                 unlock_page(page);
570 keep:
571                 list_add(&page->lru, &ret_pages);
572                 BUG_ON(PageLRU(page));
573         }
574         list_splice(&ret_pages, page_list);
575         if (pagevec_count(&freed_pvec))
576                 __pagevec_release_nonlru(&freed_pvec);
577         mod_page_state(pgactivate, pgactivate);
578         sc->nr_reclaimed += reclaimed;
579         return reclaimed;
580 }
581
582 #ifdef CONFIG_MIGRATION
583 static inline void move_to_lru(struct page *page)
584 {
585         list_del(&page->lru);
586         if (PageActive(page)) {
587                 /*
588                  * lru_cache_add_active checks that
589                  * the PG_active bit is off.
590                  */
591                 ClearPageActive(page);
592                 lru_cache_add_active(page);
593         } else {
594                 lru_cache_add(page);
595         }
596         put_page(page);
597 }
598
599 /*
600  * Add isolated pages on the list back to the LRU.
601  *
602  * returns the number of pages put back.
603  */
604 int putback_lru_pages(struct list_head *l)
605 {
606         struct page *page;
607         struct page *page2;
608         int count = 0;
609
610         list_for_each_entry_safe(page, page2, l, lru) {
611                 move_to_lru(page);
612                 count++;
613         }
614         return count;
615 }
616
617 /*
618  * swapout a single page
619  * page is locked upon entry, unlocked on exit
620  */
621 static int swap_page(struct page *page)
622 {
623         struct address_space *mapping = page_mapping(page);
624
625         if (page_mapped(page) && mapping)
626                 if (try_to_unmap(page, 0) != SWAP_SUCCESS)
627                         goto unlock_retry;
628
629         if (PageDirty(page)) {
630                 /* Page is dirty, try to write it out here */
631                 switch(pageout(page, mapping)) {
632                 case PAGE_KEEP:
633                 case PAGE_ACTIVATE:
634                         goto unlock_retry;
635
636                 case PAGE_SUCCESS:
637                         goto retry;
638
639                 case PAGE_CLEAN:
640                         ; /* try to free the page below */
641                 }
642         }
643
644         if (PagePrivate(page)) {
645                 if (!try_to_release_page(page, GFP_KERNEL) ||
646                     (!mapping && page_count(page) == 1))
647                         goto unlock_retry;
648         }
649
650         if (remove_mapping(mapping, page)) {
651                 /* Success */
652                 unlock_page(page);
653                 return 0;
654         }
655
656 unlock_retry:
657         unlock_page(page);
658
659 retry:
660         return -EAGAIN;
661 }
662
663 /*
664  * Page migration was first developed in the context of the memory hotplug
665  * project. The main authors of the migration code are:
666  *
667  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
668  * Hirokazu Takahashi <taka@valinux.co.jp>
669  * Dave Hansen <haveblue@us.ibm.com>
670  * Christoph Lameter <clameter@sgi.com>
671  */
672
673 /*
674  * Remove references for a page and establish the new page with the correct
675  * basic settings to be able to stop accesses to the page.
676  */
677 static int migrate_page_remove_references(struct page *newpage,
678                                 struct page *page, int nr_refs)
679 {
680         struct address_space *mapping = page_mapping(page);
681         struct page **radix_pointer;
682
683         /*
684          * Avoid doing any of the following work if the page count
685          * indicates that the page is in use or truncate has removed
686          * the page.
687          */
688         if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
689                 return 1;
690
691         /*
692          * Establish swap ptes for anonymous pages or destroy pte
693          * maps for files.
694          *
695          * In order to reestablish file backed mappings the fault handlers
696          * will take the radix tree_lock which may then be used to stop
697          * processses from accessing this page until the new page is ready.
698          *
699          * A process accessing via a swap pte (an anonymous page) will take a
700          * page_lock on the old page which will block the process until the
701          * migration attempt is complete. At that time the PageSwapCache bit
702          * will be examined. If the page was migrated then the PageSwapCache
703          * bit will be clear and the operation to retrieve the page will be
704          * retried which will find the new page in the radix tree. Then a new
705          * direct mapping may be generated based on the radix tree contents.
706          *
707          * If the page was not migrated then the PageSwapCache bit
708          * is still set and the operation may continue.
709          */
710         try_to_unmap(page, 1);
711
712         /*
713          * Give up if we were unable to remove all mappings.
714          */
715         if (page_mapcount(page))
716                 return 1;
717
718         write_lock_irq(&mapping->tree_lock);
719
720         radix_pointer = (struct page **)radix_tree_lookup_slot(
721                                                 &mapping->page_tree,
722                                                 page_index(page));
723
724         if (!page_mapping(page) || page_count(page) != nr_refs ||
725                         *radix_pointer != page) {
726                 write_unlock_irq(&mapping->tree_lock);
727                 return 1;
728         }
729
730         /*
731          * Now we know that no one else is looking at the page.
732          *
733          * Certain minimal information about a page must be available
734          * in order for other subsystems to properly handle the page if they
735          * find it through the radix tree update before we are finished
736          * copying the page.
737          */
738         get_page(newpage);
739         newpage->index = page->index;
740         newpage->mapping = page->mapping;
741         if (PageSwapCache(page)) {
742                 SetPageSwapCache(newpage);
743                 set_page_private(newpage, page_private(page));
744         }
745
746         *radix_pointer = newpage;
747         __put_page(page);
748         write_unlock_irq(&mapping->tree_lock);
749
750         return 0;
751 }
752
753 /*
754  * Copy the page to its new location
755  */
756 void migrate_page_copy(struct page *newpage, struct page *page)
757 {
758         copy_highpage(newpage, page);
759
760         if (PageError(page))
761                 SetPageError(newpage);
762         if (PageReferenced(page))
763                 SetPageReferenced(newpage);
764         if (PageUptodate(page))
765                 SetPageUptodate(newpage);
766         if (PageActive(page))
767                 SetPageActive(newpage);
768         if (PageChecked(page))
769                 SetPageChecked(newpage);
770         if (PageMappedToDisk(page))
771                 SetPageMappedToDisk(newpage);
772
773         if (PageDirty(page)) {
774                 clear_page_dirty_for_io(page);
775                 set_page_dirty(newpage);
776         }
777
778         ClearPageSwapCache(page);
779         ClearPageActive(page);
780         ClearPagePrivate(page);
781         set_page_private(page, 0);
782         page->mapping = NULL;
783
784         /*
785          * If any waiters have accumulated on the new page then
786          * wake them up.
787          */
788         if (PageWriteback(newpage))
789                 end_page_writeback(newpage);
790 }
791
792 /*
793  * Common logic to directly migrate a single page suitable for
794  * pages that do not use PagePrivate.
795  *
796  * Pages are locked upon entry and exit.
797  */
798 int migrate_page(struct page *newpage, struct page *page)
799 {
800         BUG_ON(PageWriteback(page));    /* Writeback must be complete */
801
802         if (migrate_page_remove_references(newpage, page, 2))
803                 return -EAGAIN;
804
805         migrate_page_copy(newpage, page);
806
807         /*
808          * Remove auxiliary swap entries and replace
809          * them with real ptes.
810          *
811          * Note that a real pte entry will allow processes that are not
812          * waiting on the page lock to use the new page via the page tables
813          * before the new page is unlocked.
814          */
815         remove_from_swap(newpage);
816         return 0;
817 }
818
819 /*
820  * migrate_pages
821  *
822  * Two lists are passed to this function. The first list
823  * contains the pages isolated from the LRU to be migrated.
824  * The second list contains new pages that the pages isolated
825  * can be moved to. If the second list is NULL then all
826  * pages are swapped out.
827  *
828  * The function returns after 10 attempts or if no pages
829  * are movable anymore because t has become empty
830  * or no retryable pages exist anymore.
831  *
832  * Return: Number of pages not migrated when "to" ran empty.
833  */
834 int migrate_pages(struct list_head *from, struct list_head *to,
835                   struct list_head *moved, struct list_head *failed)
836 {
837         int retry;
838         int nr_failed = 0;
839         int pass = 0;
840         struct page *page;
841         struct page *page2;
842         int swapwrite = current->flags & PF_SWAPWRITE;
843         int rc;
844
845         if (!swapwrite)
846                 current->flags |= PF_SWAPWRITE;
847
848 redo:
849         retry = 0;
850
851         list_for_each_entry_safe(page, page2, from, lru) {
852                 struct page *newpage = NULL;
853                 struct address_space *mapping;
854
855                 cond_resched();
856
857                 rc = 0;
858                 if (page_count(page) == 1)
859                         /* page was freed from under us. So we are done. */
860                         goto next;
861
862                 if (to && list_empty(to))
863                         break;
864
865                 /*
866                  * Skip locked pages during the first two passes to give the
867                  * functions holding the lock time to release the page. Later we
868                  * use lock_page() to have a higher chance of acquiring the
869                  * lock.
870                  */
871                 rc = -EAGAIN;
872                 if (pass > 2)
873                         lock_page(page);
874                 else
875                         if (TestSetPageLocked(page))
876                                 goto next;
877
878                 /*
879                  * Only wait on writeback if we have already done a pass where
880                  * we we may have triggered writeouts for lots of pages.
881                  */
882                 if (pass > 0) {
883                         wait_on_page_writeback(page);
884                 } else {
885                         if (PageWriteback(page))
886                                 goto unlock_page;
887                 }
888
889                 /*
890                  * Anonymous pages must have swap cache references otherwise
891                  * the information contained in the page maps cannot be
892                  * preserved.
893                  */
894                 if (PageAnon(page) && !PageSwapCache(page)) {
895                         if (!add_to_swap(page, GFP_KERNEL)) {
896                                 rc = -ENOMEM;
897                                 goto unlock_page;
898                         }
899                 }
900
901                 if (!to) {
902                         rc = swap_page(page);
903                         goto next;
904                 }
905
906                 newpage = lru_to_page(to);
907                 lock_page(newpage);
908
909                 /*
910                  * Pages are properly locked and writeback is complete.
911                  * Try to migrate the page.
912                  */
913                 mapping = page_mapping(page);
914                 if (!mapping)
915                         goto unlock_both;
916
917                 /*
918                  * Trigger writeout if page is dirty
919                  */
920                 if (PageDirty(page)) {
921                         switch (pageout(page, mapping)) {
922                         case PAGE_KEEP:
923                         case PAGE_ACTIVATE:
924                                 goto unlock_both;
925
926                         case PAGE_SUCCESS:
927                                 unlock_page(newpage);
928                                 goto next;
929
930                         case PAGE_CLEAN:
931                                 ; /* try to migrate the page below */
932                         }
933                 }
934                 /*
935                  * If we have no buffer or can release the buffer
936                  * then do a simple migration.
937                  */
938                 if (!page_has_buffers(page) ||
939                     try_to_release_page(page, GFP_KERNEL)) {
940                         rc = migrate_page(newpage, page);
941                         goto unlock_both;
942                 }
943
944                 /*
945                  * On early passes with mapped pages simply
946                  * retry. There may be a lock held for some
947                  * buffers that may go away. Later
948                  * swap them out.
949                  */
950                 if (pass > 4) {
951                         unlock_page(newpage);
952                         newpage = NULL;
953                         rc = swap_page(page);
954                         goto next;
955                 }
956
957 unlock_both:
958                 unlock_page(newpage);
959
960 unlock_page:
961                 unlock_page(page);
962
963 next:
964                 if (rc == -EAGAIN) {
965                         retry++;
966                 } else if (rc) {
967                         /* Permanent failure */
968                         list_move(&page->lru, failed);
969                         nr_failed++;
970                 } else {
971                         if (newpage) {
972                                 /* Successful migration. Return page to LRU */
973                                 move_to_lru(newpage);
974                         }
975                         list_move(&page->lru, moved);
976                 }
977         }
978         if (retry && pass++ < 10)
979                 goto redo;
980
981         if (!swapwrite)
982                 current->flags &= ~PF_SWAPWRITE;
983
984         return nr_failed + retry;
985 }
986
987 /*
988  * Isolate one page from the LRU lists and put it on the
989  * indicated list with elevated refcount.
990  *
991  * Result:
992  *  0 = page not on LRU list
993  *  1 = page removed from LRU list and added to the specified list.
994  */
995 int isolate_lru_page(struct page *page)
996 {
997         int ret = 0;
998
999         if (PageLRU(page)) {
1000                 struct zone *zone = page_zone(page);
1001                 spin_lock_irq(&zone->lru_lock);
1002                 if (TestClearPageLRU(page)) {
1003                         ret = 1;
1004                         get_page(page);
1005                         if (PageActive(page))
1006                                 del_page_from_active_list(zone, page);
1007                         else
1008                                 del_page_from_inactive_list(zone, page);
1009                 }
1010                 spin_unlock_irq(&zone->lru_lock);
1011         }
1012
1013         return ret;
1014 }
1015 #endif
1016
1017 /*
1018  * zone->lru_lock is heavily contended.  Some of the functions that
1019  * shrink the lists perform better by taking out a batch of pages
1020  * and working on them outside the LRU lock.
1021  *
1022  * For pagecache intensive workloads, this function is the hottest
1023  * spot in the kernel (apart from copy_*_user functions).
1024  *
1025  * Appropriate locks must be held before calling this function.
1026  *
1027  * @nr_to_scan: The number of pages to look through on the list.
1028  * @src:        The LRU list to pull pages off.
1029  * @dst:        The temp list to put pages on to.
1030  * @scanned:    The number of pages that were scanned.
1031  *
1032  * returns how many pages were moved onto *@dst.
1033  */
1034 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
1035                              struct list_head *dst, int *scanned)
1036 {
1037         int nr_taken = 0;
1038         struct page *page;
1039         int scan = 0;
1040
1041         while (scan++ < nr_to_scan && !list_empty(src)) {
1042                 page = lru_to_page(src);
1043                 prefetchw_prev_lru_page(page, src, flags);
1044
1045                 if (!TestClearPageLRU(page))
1046                         BUG();
1047                 list_del(&page->lru);
1048                 if (get_page_testone(page)) {
1049                         /*
1050                          * It is being freed elsewhere
1051                          */
1052                         __put_page(page);
1053                         SetPageLRU(page);
1054                         list_add(&page->lru, src);
1055                         continue;
1056                 } else {
1057                         list_add(&page->lru, dst);
1058                         nr_taken++;
1059                 }
1060         }
1061
1062         *scanned = scan;
1063         return nr_taken;
1064 }
1065
1066 /*
1067  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1068  */
1069 static void shrink_cache(struct zone *zone, struct scan_control *sc)
1070 {
1071         LIST_HEAD(page_list);
1072         struct pagevec pvec;
1073         int max_scan = sc->nr_to_scan;
1074
1075         pagevec_init(&pvec, 1);
1076
1077         lru_add_drain();
1078         spin_lock_irq(&zone->lru_lock);
1079         while (max_scan > 0) {
1080                 struct page *page;
1081                 int nr_taken;
1082                 int nr_scan;
1083                 int nr_freed;
1084
1085                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
1086                                              &zone->inactive_list,
1087                                              &page_list, &nr_scan);
1088                 zone->nr_inactive -= nr_taken;
1089                 zone->pages_scanned += nr_scan;
1090                 spin_unlock_irq(&zone->lru_lock);
1091
1092                 if (nr_taken == 0)
1093                         goto done;
1094
1095                 max_scan -= nr_scan;
1096                 nr_freed = shrink_list(&page_list, sc);
1097
1098                 local_irq_disable();
1099                 if (current_is_kswapd()) {
1100                         __mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
1101                         __mod_page_state(kswapd_steal, nr_freed);
1102                 } else
1103                         __mod_page_state_zone(zone, pgscan_direct, nr_scan);
1104                 __mod_page_state_zone(zone, pgsteal, nr_freed);
1105
1106                 spin_lock(&zone->lru_lock);
1107                 /*
1108                  * Put back any unfreeable pages.
1109                  */
1110                 while (!list_empty(&page_list)) {
1111                         page = lru_to_page(&page_list);
1112                         if (TestSetPageLRU(page))
1113                                 BUG();
1114                         list_del(&page->lru);
1115                         if (PageActive(page))
1116                                 add_page_to_active_list(zone, page);
1117                         else
1118                                 add_page_to_inactive_list(zone, page);
1119                         if (!pagevec_add(&pvec, page)) {
1120                                 spin_unlock_irq(&zone->lru_lock);
1121                                 __pagevec_release(&pvec);
1122                                 spin_lock_irq(&zone->lru_lock);
1123                         }
1124                 }
1125         }
1126         spin_unlock_irq(&zone->lru_lock);
1127 done:
1128         pagevec_release(&pvec);
1129 }
1130
1131 /*
1132  * This moves pages from the active list to the inactive list.
1133  *
1134  * We move them the other way if the page is referenced by one or more
1135  * processes, from rmap.
1136  *
1137  * If the pages are mostly unmapped, the processing is fast and it is
1138  * appropriate to hold zone->lru_lock across the whole operation.  But if
1139  * the pages are mapped, the processing is slow (page_referenced()) so we
1140  * should drop zone->lru_lock around each page.  It's impossible to balance
1141  * this, so instead we remove the pages from the LRU while processing them.
1142  * It is safe to rely on PG_active against the non-LRU pages in here because
1143  * nobody will play with that bit on a non-LRU page.
1144  *
1145  * The downside is that we have to touch page->_count against each page.
1146  * But we had to alter page->flags anyway.
1147  */
1148 static void
1149 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
1150 {
1151         int pgmoved;
1152         int pgdeactivate = 0;
1153         int pgscanned;
1154         int nr_pages = sc->nr_to_scan;
1155         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1156         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
1157         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
1158         struct page *page;
1159         struct pagevec pvec;
1160         int reclaim_mapped = 0;
1161         long mapped_ratio;
1162         long distress;
1163         long swap_tendency;
1164
1165         lru_add_drain();
1166         spin_lock_irq(&zone->lru_lock);
1167         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1168                                     &l_hold, &pgscanned);
1169         zone->pages_scanned += pgscanned;
1170         zone->nr_active -= pgmoved;
1171         spin_unlock_irq(&zone->lru_lock);
1172
1173         /*
1174          * `distress' is a measure of how much trouble we're having reclaiming
1175          * pages.  0 -> no problems.  100 -> great trouble.
1176          */
1177         distress = 100 >> zone->prev_priority;
1178
1179         /*
1180          * The point of this algorithm is to decide when to start reclaiming
1181          * mapped memory instead of just pagecache.  Work out how much memory
1182          * is mapped.
1183          */
1184         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
1185
1186         /*
1187          * Now decide how much we really want to unmap some pages.  The mapped
1188          * ratio is downgraded - just because there's a lot of mapped memory
1189          * doesn't necessarily mean that page reclaim isn't succeeding.
1190          *
1191          * The distress ratio is important - we don't want to start going oom.
1192          *
1193          * A 100% value of vm_swappiness overrides this algorithm altogether.
1194          */
1195         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
1196
1197         /*
1198          * Now use this metric to decide whether to start moving mapped memory
1199          * onto the inactive list.
1200          */
1201         if (swap_tendency >= 100)
1202                 reclaim_mapped = 1;
1203
1204         while (!list_empty(&l_hold)) {
1205                 cond_resched();
1206                 page = lru_to_page(&l_hold);
1207                 list_del(&page->lru);
1208                 if (page_mapped(page)) {
1209                         if (!reclaim_mapped ||
1210                             (total_swap_pages == 0 && PageAnon(page)) ||
1211                             page_referenced(page, 0)) {
1212                                 list_add(&page->lru, &l_active);
1213                                 continue;
1214                         }
1215                 }
1216                 list_add(&page->lru, &l_inactive);
1217         }
1218
1219         pagevec_init(&pvec, 1);
1220         pgmoved = 0;
1221         spin_lock_irq(&zone->lru_lock);
1222         while (!list_empty(&l_inactive)) {
1223                 page = lru_to_page(&l_inactive);
1224                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1225                 if (TestSetPageLRU(page))
1226                         BUG();
1227                 if (!TestClearPageActive(page))
1228                         BUG();
1229                 list_move(&page->lru, &zone->inactive_list);
1230                 pgmoved++;
1231                 if (!pagevec_add(&pvec, page)) {
1232                         zone->nr_inactive += pgmoved;
1233                         spin_unlock_irq(&zone->lru_lock);
1234                         pgdeactivate += pgmoved;
1235                         pgmoved = 0;
1236                         if (buffer_heads_over_limit)
1237                                 pagevec_strip(&pvec);
1238                         __pagevec_release(&pvec);
1239                         spin_lock_irq(&zone->lru_lock);
1240                 }
1241         }
1242         zone->nr_inactive += pgmoved;
1243         pgdeactivate += pgmoved;
1244         if (buffer_heads_over_limit) {
1245                 spin_unlock_irq(&zone->lru_lock);
1246                 pagevec_strip(&pvec);
1247                 spin_lock_irq(&zone->lru_lock);
1248         }
1249
1250         pgmoved = 0;
1251         while (!list_empty(&l_active)) {
1252                 page = lru_to_page(&l_active);
1253                 prefetchw_prev_lru_page(page, &l_active, flags);
1254                 if (TestSetPageLRU(page))
1255                         BUG();
1256                 BUG_ON(!PageActive(page));
1257                 list_move(&page->lru, &zone->active_list);
1258                 pgmoved++;
1259                 if (!pagevec_add(&pvec, page)) {
1260                         zone->nr_active += pgmoved;
1261                         pgmoved = 0;
1262                         spin_unlock_irq(&zone->lru_lock);
1263                         __pagevec_release(&pvec);
1264                         spin_lock_irq(&zone->lru_lock);
1265                 }
1266         }
1267         zone->nr_active += pgmoved;
1268         spin_unlock(&zone->lru_lock);
1269
1270         __mod_page_state_zone(zone, pgrefill, pgscanned);
1271         __mod_page_state(pgdeactivate, pgdeactivate);
1272         local_irq_enable();
1273
1274         pagevec_release(&pvec);
1275 }
1276
1277 /*
1278  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1279  */
1280 static void
1281 shrink_zone(struct zone *zone, struct scan_control *sc)
1282 {
1283         unsigned long nr_active;
1284         unsigned long nr_inactive;
1285
1286         atomic_inc(&zone->reclaim_in_progress);
1287
1288         /*
1289          * Add one to `nr_to_scan' just to make sure that the kernel will
1290          * slowly sift through the active list.
1291          */
1292         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
1293         nr_active = zone->nr_scan_active;
1294         if (nr_active >= sc->swap_cluster_max)
1295                 zone->nr_scan_active = 0;
1296         else
1297                 nr_active = 0;
1298
1299         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
1300         nr_inactive = zone->nr_scan_inactive;
1301         if (nr_inactive >= sc->swap_cluster_max)
1302                 zone->nr_scan_inactive = 0;
1303         else
1304                 nr_inactive = 0;
1305
1306         while (nr_active || nr_inactive) {
1307                 if (nr_active) {
1308                         sc->nr_to_scan = min(nr_active,
1309                                         (unsigned long)sc->swap_cluster_max);
1310                         nr_active -= sc->nr_to_scan;
1311                         refill_inactive_zone(zone, sc);
1312                 }
1313
1314                 if (nr_inactive) {
1315                         sc->nr_to_scan = min(nr_inactive,
1316                                         (unsigned long)sc->swap_cluster_max);
1317                         nr_inactive -= sc->nr_to_scan;
1318                         shrink_cache(zone, sc);
1319                 }
1320         }
1321
1322         throttle_vm_writeout();
1323
1324         atomic_dec(&zone->reclaim_in_progress);
1325 }
1326
1327 /*
1328  * This is the direct reclaim path, for page-allocating processes.  We only
1329  * try to reclaim pages from zones which will satisfy the caller's allocation
1330  * request.
1331  *
1332  * We reclaim from a zone even if that zone is over pages_high.  Because:
1333  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1334  *    allocation or
1335  * b) The zones may be over pages_high but they must go *over* pages_high to
1336  *    satisfy the `incremental min' zone defense algorithm.
1337  *
1338  * Returns the number of reclaimed pages.
1339  *
1340  * If a zone is deemed to be full of pinned pages then just give it a light
1341  * scan then give up on it.
1342  */
1343 static void
1344 shrink_caches(struct zone **zones, struct scan_control *sc)
1345 {
1346         int i;
1347
1348         for (i = 0; zones[i] != NULL; i++) {
1349                 struct zone *zone = zones[i];
1350
1351                 if (!populated_zone(zone))
1352                         continue;
1353
1354                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1355                         continue;
1356
1357                 zone->temp_priority = sc->priority;
1358                 if (zone->prev_priority > sc->priority)
1359                         zone->prev_priority = sc->priority;
1360
1361                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
1362                         continue;       /* Let kswapd poll it */
1363
1364                 shrink_zone(zone, sc);
1365         }
1366 }
1367  
1368 /*
1369  * This is the main entry point to direct page reclaim.
1370  *
1371  * If a full scan of the inactive list fails to free enough memory then we
1372  * are "out of memory" and something needs to be killed.
1373  *
1374  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1375  * high - the zone may be full of dirty or under-writeback pages, which this
1376  * caller can't do much about.  We kick pdflush and take explicit naps in the
1377  * hope that some of these pages can be written.  But if the allocating task
1378  * holds filesystem locks which prevent writeout this might not work, and the
1379  * allocation attempt will fail.
1380  */
1381 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
1382 {
1383         int priority;
1384         int ret = 0;
1385         int total_scanned = 0, total_reclaimed = 0;
1386         struct reclaim_state *reclaim_state = current->reclaim_state;
1387         struct scan_control sc;
1388         unsigned long lru_pages = 0;
1389         int i;
1390
1391         sc.gfp_mask = gfp_mask;
1392         sc.may_writepage = !laptop_mode;
1393         sc.may_swap = 1;
1394
1395         inc_page_state(allocstall);
1396
1397         for (i = 0; zones[i] != NULL; i++) {
1398                 struct zone *zone = zones[i];
1399
1400                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1401                         continue;
1402
1403                 zone->temp_priority = DEF_PRIORITY;
1404                 lru_pages += zone->nr_active + zone->nr_inactive;
1405         }
1406
1407         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1408                 sc.nr_mapped = read_page_state(nr_mapped);
1409                 sc.nr_scanned = 0;
1410                 sc.nr_reclaimed = 0;
1411                 sc.priority = priority;
1412                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1413                 if (!priority)
1414                         disable_swap_token();
1415                 shrink_caches(zones, &sc);
1416                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1417                 if (reclaim_state) {
1418                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1419                         reclaim_state->reclaimed_slab = 0;
1420                 }
1421                 total_scanned += sc.nr_scanned;
1422                 total_reclaimed += sc.nr_reclaimed;
1423                 if (total_reclaimed >= sc.swap_cluster_max) {
1424                         ret = 1;
1425                         goto out;
1426                 }
1427
1428                 /*
1429                  * Try to write back as many pages as we just scanned.  This
1430                  * tends to cause slow streaming writers to write data to the
1431                  * disk smoothly, at the dirtying rate, which is nice.   But
1432                  * that's undesirable in laptop mode, where we *want* lumpy
1433                  * writeout.  So in laptop mode, write out the whole world.
1434                  */
1435                 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
1436                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1437                         sc.may_writepage = 1;
1438                 }
1439
1440                 /* Take a nap, wait for some writeback to complete */
1441                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1442                         blk_congestion_wait(WRITE, HZ/10);
1443         }
1444 out:
1445         for (i = 0; zones[i] != 0; i++) {
1446                 struct zone *zone = zones[i];
1447
1448                 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1449                         continue;
1450
1451                 zone->prev_priority = zone->temp_priority;
1452         }
1453         return ret;
1454 }
1455
1456 /*
1457  * For kswapd, balance_pgdat() will work across all this node's zones until
1458  * they are all at pages_high.
1459  *
1460  * If `nr_pages' is non-zero then it is the number of pages which are to be
1461  * reclaimed, regardless of the zone occupancies.  This is a software suspend
1462  * special.
1463  *
1464  * Returns the number of pages which were actually freed.
1465  *
1466  * There is special handling here for zones which are full of pinned pages.
1467  * This can happen if the pages are all mlocked, or if they are all used by
1468  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1469  * What we do is to detect the case where all pages in the zone have been
1470  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1471  * dead and from now on, only perform a short scan.  Basically we're polling
1472  * the zone for when the problem goes away.
1473  *
1474  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1475  * zones which have free_pages > pages_high, but once a zone is found to have
1476  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1477  * of the number of free pages in the lower zones.  This interoperates with
1478  * the page allocator fallback scheme to ensure that aging of pages is balanced
1479  * across the zones.
1480  */
1481 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1482 {
1483         int to_free = nr_pages;
1484         int all_zones_ok;
1485         int priority;
1486         int i;
1487         int total_scanned, total_reclaimed;
1488         struct reclaim_state *reclaim_state = current->reclaim_state;
1489         struct scan_control sc;
1490
1491 loop_again:
1492         total_scanned = 0;
1493         total_reclaimed = 0;
1494         sc.gfp_mask = GFP_KERNEL;
1495         sc.may_writepage = !laptop_mode;
1496         sc.may_swap = 1;
1497         sc.nr_mapped = read_page_state(nr_mapped);
1498
1499         inc_page_state(pageoutrun);
1500
1501         for (i = 0; i < pgdat->nr_zones; i++) {
1502                 struct zone *zone = pgdat->node_zones + i;
1503
1504                 zone->temp_priority = DEF_PRIORITY;
1505         }
1506
1507         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1508                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1509                 unsigned long lru_pages = 0;
1510
1511                 /* The swap token gets in the way of swapout... */
1512                 if (!priority)
1513                         disable_swap_token();
1514
1515                 all_zones_ok = 1;
1516
1517                 if (nr_pages == 0) {
1518                         /*
1519                          * Scan in the highmem->dma direction for the highest
1520                          * zone which needs scanning
1521                          */
1522                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1523                                 struct zone *zone = pgdat->node_zones + i;
1524
1525                                 if (!populated_zone(zone))
1526                                         continue;
1527
1528                                 if (zone->all_unreclaimable &&
1529                                                 priority != DEF_PRIORITY)
1530                                         continue;
1531
1532                                 if (!zone_watermark_ok(zone, order,
1533                                                 zone->pages_high, 0, 0)) {
1534                                         end_zone = i;
1535                                         goto scan;
1536                                 }
1537                         }
1538                         goto out;
1539                 } else {
1540                         end_zone = pgdat->nr_zones - 1;
1541                 }
1542 scan:
1543                 for (i = 0; i <= end_zone; i++) {
1544                         struct zone *zone = pgdat->node_zones + i;
1545
1546                         lru_pages += zone->nr_active + zone->nr_inactive;
1547                 }
1548
1549                 /*
1550                  * Now scan the zone in the dma->highmem direction, stopping
1551                  * at the last zone which needs scanning.
1552                  *
1553                  * We do this because the page allocator works in the opposite
1554                  * direction.  This prevents the page allocator from allocating
1555                  * pages behind kswapd's direction of progress, which would
1556                  * cause too much scanning of the lower zones.
1557                  */
1558                 for (i = 0; i <= end_zone; i++) {
1559                         struct zone *zone = pgdat->node_zones + i;
1560                         int nr_slab;
1561
1562                         if (!populated_zone(zone))
1563                                 continue;
1564
1565                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1566                                 continue;
1567
1568                         if (nr_pages == 0) {    /* Not software suspend */
1569                                 if (!zone_watermark_ok(zone, order,
1570                                                 zone->pages_high, end_zone, 0))
1571                                         all_zones_ok = 0;
1572                         }
1573                         zone->temp_priority = priority;
1574                         if (zone->prev_priority > priority)
1575                                 zone->prev_priority = priority;
1576                         sc.nr_scanned = 0;
1577                         sc.nr_reclaimed = 0;
1578                         sc.priority = priority;
1579                         sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1580                         atomic_inc(&zone->reclaim_in_progress);
1581                         shrink_zone(zone, &sc);
1582                         atomic_dec(&zone->reclaim_in_progress);
1583                         reclaim_state->reclaimed_slab = 0;
1584                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1585                                                 lru_pages);
1586                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1587                         total_reclaimed += sc.nr_reclaimed;
1588                         total_scanned += sc.nr_scanned;
1589                         if (zone->all_unreclaimable)
1590                                 continue;
1591                         if (nr_slab == 0 && zone->pages_scanned >=
1592                                     (zone->nr_active + zone->nr_inactive) * 4)
1593                                 zone->all_unreclaimable = 1;
1594                         /*
1595                          * If we've done a decent amount of scanning and
1596                          * the reclaim ratio is low, start doing writepage
1597                          * even in laptop mode
1598                          */
1599                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1600                             total_scanned > total_reclaimed+total_reclaimed/2)
1601                                 sc.may_writepage = 1;
1602                 }
1603                 if (nr_pages && to_free > total_reclaimed)
1604                         continue;       /* swsusp: need to do more work */
1605                 if (all_zones_ok)
1606                         break;          /* kswapd: all done */
1607                 /*
1608                  * OK, kswapd is getting into trouble.  Take a nap, then take
1609                  * another pass across the zones.
1610                  */
1611                 if (total_scanned && priority < DEF_PRIORITY - 2)
1612                         blk_congestion_wait(WRITE, HZ/10);
1613
1614                 /*
1615                  * We do this so kswapd doesn't build up large priorities for
1616                  * example when it is freeing in parallel with allocators. It
1617                  * matches the direct reclaim path behaviour in terms of impact
1618                  * on zone->*_priority.
1619                  */
1620                 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1621                         break;
1622         }
1623 out:
1624         for (i = 0; i < pgdat->nr_zones; i++) {
1625                 struct zone *zone = pgdat->node_zones + i;
1626
1627                 zone->prev_priority = zone->temp_priority;
1628         }
1629         if (!all_zones_ok) {
1630                 cond_resched();
1631                 goto loop_again;
1632         }
1633
1634         return total_reclaimed;
1635 }
1636
1637 /*
1638  * The background pageout daemon, started as a kernel thread
1639  * from the init process. 
1640  *
1641  * This basically trickles out pages so that we have _some_
1642  * free memory available even if there is no other activity
1643  * that frees anything up. This is needed for things like routing
1644  * etc, where we otherwise might have all activity going on in
1645  * asynchronous contexts that cannot page things out.
1646  *
1647  * If there are applications that are active memory-allocators
1648  * (most normal use), this basically shouldn't matter.
1649  */
1650 static int kswapd(void *p)
1651 {
1652         unsigned long order;
1653         pg_data_t *pgdat = (pg_data_t*)p;
1654         struct task_struct *tsk = current;
1655         DEFINE_WAIT(wait);
1656         struct reclaim_state reclaim_state = {
1657                 .reclaimed_slab = 0,
1658         };
1659         cpumask_t cpumask;
1660
1661         daemonize("kswapd%d", pgdat->node_id);
1662         cpumask = node_to_cpumask(pgdat->node_id);
1663         if (!cpus_empty(cpumask))
1664                 set_cpus_allowed(tsk, cpumask);
1665         current->reclaim_state = &reclaim_state;
1666
1667         /*
1668          * Tell the memory management that we're a "memory allocator",
1669          * and that if we need more memory we should get access to it
1670          * regardless (see "__alloc_pages()"). "kswapd" should
1671          * never get caught in the normal page freeing logic.
1672          *
1673          * (Kswapd normally doesn't need memory anyway, but sometimes
1674          * you need a small amount of memory in order to be able to
1675          * page out something else, and this flag essentially protects
1676          * us from recursively trying to free more memory as we're
1677          * trying to free the first piece of memory in the first place).
1678          */
1679         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1680
1681         order = 0;
1682         for ( ; ; ) {
1683                 unsigned long new_order;
1684
1685                 try_to_freeze();
1686
1687                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1688                 new_order = pgdat->kswapd_max_order;
1689                 pgdat->kswapd_max_order = 0;
1690                 if (order < new_order) {
1691                         /*
1692                          * Don't sleep if someone wants a larger 'order'
1693                          * allocation
1694                          */
1695                         order = new_order;
1696                 } else {
1697                         schedule();
1698                         order = pgdat->kswapd_max_order;
1699                 }
1700                 finish_wait(&pgdat->kswapd_wait, &wait);
1701
1702                 balance_pgdat(pgdat, 0, order);
1703         }
1704         return 0;
1705 }
1706
1707 /*
1708  * A zone is low on free memory, so wake its kswapd task to service it.
1709  */
1710 void wakeup_kswapd(struct zone *zone, int order)
1711 {
1712         pg_data_t *pgdat;
1713
1714         if (!populated_zone(zone))
1715                 return;
1716
1717         pgdat = zone->zone_pgdat;
1718         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1719                 return;
1720         if (pgdat->kswapd_max_order < order)
1721                 pgdat->kswapd_max_order = order;
1722         if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1723                 return;
1724         if (!waitqueue_active(&pgdat->kswapd_wait))
1725                 return;
1726         wake_up_interruptible(&pgdat->kswapd_wait);
1727 }
1728
1729 #ifdef CONFIG_PM
1730 /*
1731  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1732  * pages.
1733  */
1734 int shrink_all_memory(int nr_pages)
1735 {
1736         pg_data_t *pgdat;
1737         int nr_to_free = nr_pages;
1738         int ret = 0;
1739         struct reclaim_state reclaim_state = {
1740                 .reclaimed_slab = 0,
1741         };
1742
1743         current->reclaim_state = &reclaim_state;
1744         for_each_pgdat(pgdat) {
1745                 int freed;
1746                 freed = balance_pgdat(pgdat, nr_to_free, 0);
1747                 ret += freed;
1748                 nr_to_free -= freed;
1749                 if (nr_to_free <= 0)
1750                         break;
1751         }
1752         current->reclaim_state = NULL;
1753         return ret;
1754 }
1755 #endif
1756
1757 #ifdef CONFIG_HOTPLUG_CPU
1758 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1759    not required for correctness.  So if the last cpu in a node goes
1760    away, we get changed to run anywhere: as the first one comes back,
1761    restore their cpu bindings. */
1762 static int __devinit cpu_callback(struct notifier_block *nfb,
1763                                   unsigned long action,
1764                                   void *hcpu)
1765 {
1766         pg_data_t *pgdat;
1767         cpumask_t mask;
1768
1769         if (action == CPU_ONLINE) {
1770                 for_each_pgdat(pgdat) {
1771                         mask = node_to_cpumask(pgdat->node_id);
1772                         if (any_online_cpu(mask) != NR_CPUS)
1773                                 /* One of our CPUs online: restore mask */
1774                                 set_cpus_allowed(pgdat->kswapd, mask);
1775                 }
1776         }
1777         return NOTIFY_OK;
1778 }
1779 #endif /* CONFIG_HOTPLUG_CPU */
1780
1781 static int __init kswapd_init(void)
1782 {
1783         pg_data_t *pgdat;
1784         swap_setup();
1785         for_each_pgdat(pgdat)
1786                 pgdat->kswapd
1787                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1788         total_memory = nr_free_pagecache_pages();
1789         hotcpu_notifier(cpu_callback, 0);
1790         return 0;
1791 }
1792
1793 module_init(kswapd_init)
1794
1795 #ifdef CONFIG_NUMA
1796 /*
1797  * Zone reclaim mode
1798  *
1799  * If non-zero call zone_reclaim when the number of free pages falls below
1800  * the watermarks.
1801  *
1802  * In the future we may add flags to the mode. However, the page allocator
1803  * should only have to check that zone_reclaim_mode != 0 before calling
1804  * zone_reclaim().
1805  */
1806 int zone_reclaim_mode __read_mostly;
1807
1808 #define RECLAIM_OFF 0
1809 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1810 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1811 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1812 #define RECLAIM_SLAB (1<<3)     /* Do a global slab shrink if the zone is out of memory */
1813
1814 /*
1815  * Mininum time between zone reclaim scans
1816  */
1817 int zone_reclaim_interval __read_mostly = 30*HZ;
1818
1819 /*
1820  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1821  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1822  * a zone.
1823  */
1824 #define ZONE_RECLAIM_PRIORITY 4
1825
1826 /*
1827  * Try to free up some pages from this zone through reclaim.
1828  */
1829 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1830 {
1831         int nr_pages;
1832         struct task_struct *p = current;
1833         struct reclaim_state reclaim_state;
1834         struct scan_control sc;
1835         cpumask_t mask;
1836         int node_id;
1837
1838         if (time_before(jiffies,
1839                 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval))
1840                         return 0;
1841
1842         if (!(gfp_mask & __GFP_WAIT) ||
1843                 zone->all_unreclaimable ||
1844                 atomic_read(&zone->reclaim_in_progress) > 0)
1845                         return 0;
1846
1847         node_id = zone->zone_pgdat->node_id;
1848         mask = node_to_cpumask(node_id);
1849         if (!cpus_empty(mask) && node_id != numa_node_id())
1850                 return 0;
1851
1852         sc.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE);
1853         sc.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP);
1854         sc.nr_scanned = 0;
1855         sc.nr_reclaimed = 0;
1856         sc.priority = ZONE_RECLAIM_PRIORITY + 1;
1857         sc.nr_mapped = read_page_state(nr_mapped);
1858         sc.gfp_mask = gfp_mask;
1859
1860         disable_swap_token();
1861
1862         nr_pages = 1 << order;
1863         if (nr_pages > SWAP_CLUSTER_MAX)
1864                 sc.swap_cluster_max = nr_pages;
1865         else
1866                 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1867
1868         cond_resched();
1869         p->flags |= PF_MEMALLOC;
1870         reclaim_state.reclaimed_slab = 0;
1871         p->reclaim_state = &reclaim_state;
1872
1873         /*
1874          * Free memory by calling shrink zone with increasing priorities
1875          * until we have enough memory freed.
1876          */
1877         do {
1878                 sc.priority--;
1879                 shrink_zone(zone, &sc);
1880
1881         } while (sc.nr_reclaimed < nr_pages && sc.priority > 0);
1882
1883         if (sc.nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) {
1884                 /*
1885                  * shrink_slab does not currently allow us to determine
1886                  * how many pages were freed in the zone. So we just
1887                  * shake the slab and then go offnode for a single allocation.
1888                  *
1889                  * shrink_slab will free memory on all zones and may take
1890                  * a long time.
1891                  */
1892                 shrink_slab(sc.nr_scanned, gfp_mask, order);
1893                 sc.nr_reclaimed = 1;    /* Avoid getting the off node timeout */
1894         }
1895
1896         p->reclaim_state = NULL;
1897         current->flags &= ~PF_MEMALLOC;
1898
1899         if (sc.nr_reclaimed == 0)
1900                 zone->last_unsuccessful_zone_reclaim = jiffies;
1901
1902         return sc.nr_reclaimed >= nr_pages;
1903 }
1904 #endif
1905