• source navigation  • diff markup  • identifier search  • freetext search  • file search  • 


Version: 2.6.8   2.6.16   2.6.25   2.6.30   2.6.34  

Architecture: i386   arm   mips   ppc   alpha   m68k   sparc   sparc64  

linux/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/suspend.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/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         /* How many pages shrink_cache() should reclaim */
 67         int nr_to_reclaim;
 68 
 69         /* Ask shrink_caches, or shrink_zone to scan at this priority */
 70         unsigned int priority;
 71 
 72         /* This context's GFP mask */
 73         unsigned int gfp_mask;
 74 
 75         int may_writepage;
 76 };
 77 
 78 /*
 79  * The list of shrinker callbacks used by to apply pressure to
 80  * ageable caches.
 81  */
 82 struct shrinker {
 83         shrinker_t              shrinker;
 84         struct list_head        list;
 85         int                     seeks;  /* seeks to recreate an obj */
 86         long                    nr;     /* objs pending delete */
 87 };
 88 
 89 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
 90 
 91 #ifdef ARCH_HAS_PREFETCH
 92 #define prefetch_prev_lru_page(_page, _base, _field)                    \
 93         do {                                                            \
 94                 if ((_page)->lru.prev != _base) {                       \
 95                         struct page *prev;                              \
 96                                                                         \
 97                         prev = lru_to_page(&(_page->lru));              \
 98                         prefetch(&prev->_field);                        \
 99                 }                                                       \
100         } while (0)
101 #else
102 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
103 #endif
104 
105 #ifdef ARCH_HAS_PREFETCHW
106 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
107         do {                                                            \
108                 if ((_page)->lru.prev != _base) {                       \
109                         struct page *prev;                              \
110                                                                         \
111                         prev = lru_to_page(&(_page->lru));              \
112                         prefetchw(&prev->_field);                       \
113                 }                                                       \
114         } while (0)
115 #else
116 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
117 #endif
118 
119 /*
120  * From 0 .. 100.  Higher means more swappy.
121  */
122 int vm_swappiness = 60;
123 static long total_memory;
124 
125 static LIST_HEAD(shrinker_list);
126 static DECLARE_RWSEM(shrinker_rwsem);
127 
128 /*
129  * Add a shrinker callback to be called from the vm
130  */
131 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
132 {
133         struct shrinker *shrinker;
134 
135         shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
136         if (shrinker) {
137                 shrinker->shrinker = theshrinker;
138                 shrinker->seeks = seeks;
139                 shrinker->nr = 0;
140                 down_write(&shrinker_rwsem);
141                 list_add(&shrinker->list, &shrinker_list);
142                 up_write(&shrinker_rwsem);
143         }
144         return shrinker;
145 }
146 EXPORT_SYMBOL(set_shrinker);
147 
148 /*
149  * Remove one
150  */
151 void remove_shrinker(struct shrinker *shrinker)
152 {
153         down_write(&shrinker_rwsem);
154         list_del(&shrinker->list);
155         up_write(&shrinker_rwsem);
156         kfree(shrinker);
157 }
158 EXPORT_SYMBOL(remove_shrinker);
159 
160 #define SHRINK_BATCH 128
161 /*
162  * Call the shrink functions to age shrinkable caches
163  *
164  * Here we assume it costs one seek to replace a lru page and that it also
165  * takes a seek to recreate a cache object.  With this in mind we age equal
166  * percentages of the lru and ageable caches.  This should balance the seeks
167  * generated by these structures.
168  *
169  * If the vm encounted mapped pages on the LRU it increase the pressure on
170  * slab to avoid swapping.
171  *
172  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
173  *
174  * `lru_pages' represents the number of on-LRU pages in all the zones which
175  * are eligible for the caller's allocation attempt.  It is used for balancing
176  * slab reclaim versus page reclaim.
177  */
178 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
179                         unsigned long lru_pages)
180 {
181         struct shrinker *shrinker;
182 
183         if (scanned == 0)
184                 scanned = SWAP_CLUSTER_MAX;
185 
186         if (!down_read_trylock(&shrinker_rwsem))
187                 return 0;
188 
189         list_for_each_entry(shrinker, &shrinker_list, list) {
190                 unsigned long long delta;
191                 unsigned long total_scan;
192 
193                 delta = (4 * scanned) / shrinker->seeks;
194                 delta *= (*shrinker->shrinker)(0, gfp_mask);
195                 do_div(delta, lru_pages + 1);
196                 shrinker->nr += delta;
197                 if (shrinker->nr < 0)
198                         shrinker->nr = LONG_MAX;        /* It wrapped! */
199 
200                 total_scan = shrinker->nr;
201                 shrinker->nr = 0;
202 
203                 while (total_scan >= SHRINK_BATCH) {
204                         long this_scan = SHRINK_BATCH;
205                         int shrink_ret;
206 
207                         shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
208                         if (shrink_ret == -1)
209                                 break;
210                         mod_page_state(slabs_scanned, this_scan);
211                         total_scan -= this_scan;
212 
213                         cond_resched();
214                 }
215 
216                 shrinker->nr += total_scan;
217         }
218         up_read(&shrinker_rwsem);
219         return 0;
220 }
221 
222 /* Must be called with page's rmap lock held. */
223 static inline int page_mapping_inuse(struct page *page)
224 {
225         struct address_space *mapping;
226 
227         /* Page is in somebody's page tables. */
228         if (page_mapped(page))
229                 return 1;
230 
231         /* Be more reluctant to reclaim swapcache than pagecache */
232         if (PageSwapCache(page))
233                 return 1;
234 
235         mapping = page_mapping(page);
236         if (!mapping)
237                 return 0;
238 
239         /* File is mmap'd by somebody? */
240         return mapping_mapped(mapping);
241 }
242 
243 static inline int is_page_cache_freeable(struct page *page)
244 {
245         return page_count(page) - !!PagePrivate(page) == 2;
246 }
247 
248 static int may_write_to_queue(struct backing_dev_info *bdi)
249 {
250         if (current_is_kswapd())
251                 return 1;
252         if (current_is_pdflush())       /* This is unlikely, but why not... */
253                 return 1;
254         if (!bdi_write_congested(bdi))
255                 return 1;
256         if (bdi == current->backing_dev_info)
257                 return 1;
258         return 0;
259 }
260 
261 /*
262  * We detected a synchronous write error writing a page out.  Probably
263  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
264  * fsync(), msync() or close().
265  *
266  * The tricky part is that after writepage we cannot touch the mapping: nothing
267  * prevents it from being freed up.  But we have a ref on the page and once
268  * that page is locked, the mapping is pinned.
269  *
270  * We're allowed to run sleeping lock_page() here because we know the caller has
271  * __GFP_FS.
272  */
273 static void handle_write_error(struct address_space *mapping,
274                                 struct page *page, int error)
275 {
276         lock_page(page);
277         if (page_mapping(page) == mapping) {
278                 if (error == -ENOSPC)
279                         set_bit(AS_ENOSPC, &mapping->flags);
280                 else
281                         set_bit(AS_EIO, &mapping->flags);
282         }
283         unlock_page(page);
284 }
285 
286 /*
287  * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
288  */
289 static pageout_t pageout(struct page *page, struct address_space *mapping)
290 {
291         /*
292          * If the page is dirty, only perform writeback if that write
293          * will be non-blocking.  To prevent this allocation from being
294          * stalled by pagecache activity.  But note that there may be
295          * stalls if we need to run get_block().  We could test
296          * PagePrivate for that.
297          *
298          * If this process is currently in generic_file_write() against
299          * this page's queue, we can perform writeback even if that
300          * will block.
301          *
302          * If the page is swapcache, write it back even if that would
303          * block, for some throttling. This happens by accident, because
304          * swap_backing_dev_info is bust: it doesn't reflect the
305          * congestion state of the swapdevs.  Easy to fix, if needed.
306          * See swapfile.c:page_queue_congested().
307          */
308         if (!is_page_cache_freeable(page))
309                 return PAGE_KEEP;
310         if (!mapping)
311                 return PAGE_KEEP;
312         if (mapping->a_ops->writepage == NULL)
313                 return PAGE_ACTIVATE;
314         if (!may_write_to_queue(mapping->backing_dev_info))
315                 return PAGE_KEEP;
316 
317         if (clear_page_dirty_for_io(page)) {
318                 int res;
319                 struct writeback_control wbc = {
320                         .sync_mode = WB_SYNC_NONE,
321                         .nr_to_write = SWAP_CLUSTER_MAX,
322                         .nonblocking = 1,
323                         .for_reclaim = 1,
324                 };
325 
326                 SetPageReclaim(page);
327                 res = mapping->a_ops->writepage(page, &wbc);
328                 if (res < 0)
329                         handle_write_error(mapping, page, res);
330                 if (res == WRITEPAGE_ACTIVATE) {
331                         ClearPageReclaim(page);
332                         return PAGE_ACTIVATE;
333                 }
334                 if (!PageWriteback(page)) {
335                         /* synchronous write or broken a_ops? */
336                         ClearPageReclaim(page);
337                 }
338 
339                 return PAGE_SUCCESS;
340         }
341 
342         return PAGE_CLEAN;
343 }
344 
345 /*
346  * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
347  */
348 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
349 {
350         LIST_HEAD(ret_pages);
351         struct pagevec freed_pvec;
352         int pgactivate = 0;
353         int reclaimed = 0;
354 
355         cond_resched();
356 
357         pagevec_init(&freed_pvec, 1);
358         while (!list_empty(page_list)) {
359                 struct address_space *mapping;
360                 struct page *page;
361                 int may_enter_fs;
362                 int referenced;
363 
364                 page = lru_to_page(page_list);
365                 list_del(&page->lru);
366 
367                 if (TestSetPageLocked(page))
368                         goto keep;
369 
370                 BUG_ON(PageActive(page));
371 
372                 if (PageWriteback(page))
373                         goto keep_locked;
374 
375                 sc->nr_scanned++;
376                 /* Double the slab pressure for mapped and swapcache pages */
377                 if (page_mapped(page) || PageSwapCache(page))
378                         sc->nr_scanned++;
379 
380                 page_map_lock(page);
381                 referenced = page_referenced(page);
382                 if (referenced && page_mapping_inuse(page)) {
383                         /* In active use or really unfreeable.  Activate it. */
384                         page_map_unlock(page);
385                         goto activate_locked;
386                 }
387 
388 #ifdef CONFIG_SWAP
389                 /*
390                  * Anonymous process memory has backing store?
391                  * Try to allocate it some swap space here.
392                  *
393                  * XXX: implement swap clustering ?
394                  */
395                 if (PageAnon(page) && !PageSwapCache(page)) {
396                         page_map_unlock(page);
397                         if (!add_to_swap(page))
398                                 goto activate_locked;
399                         page_map_lock(page);
400                 }
401 #endif /* CONFIG_SWAP */
402 
403                 mapping = page_mapping(page);
404                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
405                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
406 
407                 /*
408                  * The page is mapped into the page tables of one or more
409                  * processes. Try to unmap it here.
410                  */
411                 if (page_mapped(page) && mapping) {
412                         switch (try_to_unmap(page)) {
413                         case SWAP_FAIL:
414                                 page_map_unlock(page);
415                                 goto activate_locked;
416                         case SWAP_AGAIN:
417                                 page_map_unlock(page);
418                                 goto keep_locked;
419                         case SWAP_SUCCESS:
420                                 ; /* try to free the page below */
421                         }
422                 }
423                 page_map_unlock(page);
424 
425                 if (PageDirty(page)) {
426                         if (referenced)
427                                 goto keep_locked;
428                         if (!may_enter_fs)
429                                 goto keep_locked;
430                         if (laptop_mode && !sc->may_writepage)
431                                 goto keep_locked;
432 
433                         /* Page is dirty, try to write it out here */
434                         switch(pageout(page, mapping)) {
435                         case PAGE_KEEP:
436                                 goto keep_locked;
437                         case PAGE_ACTIVATE:
438                                 goto activate_locked;
439                         case PAGE_SUCCESS:
440                                 if (PageWriteback(page) || PageDirty(page))
441                                         goto keep;
442                                 /*
443                                  * A synchronous write - probably a ramdisk.  Go
444                                  * ahead and try to reclaim the page.
445                                  */
446                                 if (TestSetPageLocked(page))
447                                         goto keep;
448                                 if (PageDirty(page) || PageWriteback(page))
449                                         goto keep_locked;
450                                 mapping = page_mapping(page);
451                         case PAGE_CLEAN:
452                                 ; /* try to free the page below */
453                         }
454                 }
455 
456                 /*
457                  * If the page has buffers, try to free the buffer mappings
458                  * associated with this page. If we succeed we try to free
459                  * the page as well.
460                  *
461                  * We do this even if the page is PageDirty().
462                  * try_to_release_page() does not perform I/O, but it is
463                  * possible for a page to have PageDirty set, but it is actually
464                  * clean (all its buffers are clean).  This happens if the
465                  * buffers were written out directly, with submit_bh(). ext3
466                  * will do this, as well as the blockdev mapping. 
467                  * try_to_release_page() will discover that cleanness and will
468                  * drop the buffers and mark the page clean - it can be freed.
469                  *
470                  * Rarely, pages can have buffers and no ->mapping.  These are
471                  * the pages which were not successfully invalidated in
472                  * truncate_complete_page().  We try to drop those buffers here
473                  * and if that worked, and the page is no longer mapped into
474                  * process address space (page_count == 1) it can be freed.
475                  * Otherwise, leave the page on the LRU so it is swappable.
476                  */
477                 if (PagePrivate(page)) {
478                         if (!try_to_release_page(page, sc->gfp_mask))
479                                 goto activate_locked;
480                         if (!mapping && page_count(page) == 1)
481                                 goto free_it;
482                 }
483 
484                 if (!mapping)
485                         goto keep_locked;       /* truncate got there first */
486 
487                 spin_lock_irq(&mapping->tree_lock);
488 
489                 /*
490                  * The non-racy check for busy page.  It is critical to check
491                  * PageDirty _after_ making sure that the page is freeable and
492                  * not in use by anybody.       (pagecache + us == 2)
493                  */
494                 if (page_count(page) != 2 || PageDirty(page)) {
495                         spin_unlock_irq(&mapping->tree_lock);
496                         goto keep_locked;
497                 }
498 
499 #ifdef CONFIG_SWAP
500                 if (PageSwapCache(page)) {
501                         swp_entry_t swap = { .val = page->private };
502                         __delete_from_swap_cache(page);
503                         spin_unlock_irq(&mapping->tree_lock);
504                         swap_free(swap);
505                         __put_page(page);       /* The pagecache ref */
506                         goto free_it;
507                 }
508 #endif /* CONFIG_SWAP */
509 
510                 __remove_from_page_cache(page);
511                 spin_unlock_irq(&mapping->tree_lock);
512                 __put_page(page);
513 
514 free_it:
515                 unlock_page(page);
516                 reclaimed++;
517                 if (!pagevec_add(&freed_pvec, page))
518                         __pagevec_release_nonlru(&freed_pvec);
519                 continue;
520 
521 activate_locked:
522                 SetPageActive(page);
523                 pgactivate++;
524 keep_locked:
525                 unlock_page(page);
526 keep:
527                 list_add(&page->lru, &ret_pages);
528                 BUG_ON(PageLRU(page));
529         }
530         list_splice(&ret_pages, page_list);
531         if (pagevec_count(&freed_pvec))
532                 __pagevec_release_nonlru(&freed_pvec);
533         mod_page_state(pgactivate, pgactivate);
534         sc->nr_reclaimed += reclaimed;
535         return reclaimed;
536 }
537 
538 /*
539  * zone->lru_lock is heavily contented.  We relieve it by quickly privatising
540  * a batch of pages and working on them outside the lock.  Any pages which were
541  * not freed will be added back to the LRU.
542  *
543  * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
544  *
545  * For pagecache intensive workloads, the first loop here is the hottest spot
546  * in the kernel (apart from the copy_*_user functions).
547  */
548 static void shrink_cache(struct zone *zone, struct scan_control *sc)
549 {
550         LIST_HEAD(page_list);
551         struct pagevec pvec;
552         int max_scan = sc->nr_to_scan;
553 
554         pagevec_init(&pvec, 1);
555 
556         lru_add_drain();
557         spin_lock_irq(&zone->lru_lock);
558         while (max_scan > 0) {
559                 struct page *page;
560                 int nr_taken = 0;
561                 int nr_scan = 0;
562                 int nr_freed;
563 
564                 while (nr_scan++ < SWAP_CLUSTER_MAX &&
565                                 !list_empty(&zone->inactive_list)) {
566                         page = lru_to_page(&zone->inactive_list);
567 
568                         prefetchw_prev_lru_page(page,
569                                                 &zone->inactive_list, flags);
570 
571                         if (!TestClearPageLRU(page))
572                                 BUG();
573                         list_del(&page->lru);
574                         if (get_page_testone(page)) {
575                                 /*
576                                  * It is being freed elsewhere
577                                  */
578                                 __put_page(page);
579                                 SetPageLRU(page);
580                                 list_add(&page->lru, &zone->inactive_list);
581                                 continue;
582                         }
583                         list_add(&page->lru, &page_list);
584                         nr_taken++;
585                 }
586                 zone->nr_inactive -= nr_taken;
587                 spin_unlock_irq(&zone->lru_lock);
588 
589                 if (nr_taken == 0)
590                         goto done;
591 
592                 max_scan -= nr_scan;
593                 if (current_is_kswapd())
594                         mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
595                 else
596                         mod_page_state_zone(zone, pgscan_direct, nr_scan);
597                 nr_freed = shrink_list(&page_list, sc);
598                 if (current_is_kswapd())
599                         mod_page_state(kswapd_steal, nr_freed);
600                 mod_page_state_zone(zone, pgsteal, nr_freed);
601                 sc->nr_to_reclaim -= nr_freed;
602 
603                 spin_lock_irq(&zone->lru_lock);
604                 /*
605                  * Put back any unfreeable pages.
606                  */
607                 while (!list_empty(&page_list)) {
608                         page = lru_to_page(&page_list);
609                         if (TestSetPageLRU(page))
610                                 BUG();
611                         list_del(&page->lru);
612                         if (PageActive(page))
613                                 add_page_to_active_list(zone, page);
614                         else
615                                 add_page_to_inactive_list(zone, page);
616                         if (!pagevec_add(&pvec, page)) {
617                                 spin_unlock_irq(&zone->lru_lock);
618                                 __pagevec_release(&pvec);
619                                 spin_lock_irq(&zone->lru_lock);
620                         }
621                 }
622         }
623         spin_unlock_irq(&zone->lru_lock);
624 done:
625         pagevec_release(&pvec);
626 }
627 
628 /*
629  * This moves pages from the active list to the inactive list.
630  *
631  * We move them the other way if the page is referenced by one or more
632  * processes, from rmap.
633  *
634  * If the pages are mostly unmapped, the processing is fast and it is
635  * appropriate to hold zone->lru_lock across the whole operation.  But if
636  * the pages are mapped, the processing is slow (page_referenced()) so we
637  * should drop zone->lru_lock around each page.  It's impossible to balance
638  * this, so instead we remove the pages from the LRU while processing them.
639  * It is safe to rely on PG_active against the non-LRU pages in here because
640  * nobody will play with that bit on a non-LRU page.
641  *
642  * The downside is that we have to touch page->_count against each page.
643  * But we had to alter page->flags anyway.
644  */
645 static void
646 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
647 {
648         int pgmoved;
649         int pgdeactivate = 0;
650         int pgscanned = 0;
651         int nr_pages = sc->nr_to_scan;
652         LIST_HEAD(l_hold);      /* The pages which were snipped off */
653         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
654         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
655         struct page *page;
656         struct pagevec pvec;
657         int reclaim_mapped = 0;
658         long mapped_ratio;
659         long distress;
660         long swap_tendency;
661 
662         lru_add_drain();
663         pgmoved = 0;
664         spin_lock_irq(&zone->lru_lock);
665         while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
666                 page = lru_to_page(&zone->active_list);
667                 prefetchw_prev_lru_page(page, &zone->active_list, flags);
668                 if (!TestClearPageLRU(page))
669                         BUG();
670                 list_del(&page->lru);
671                 if (get_page_testone(page)) {
672                         /*
673                          * It was already free!  release_pages() or put_page()
674                          * are about to remove it from the LRU and free it. So
675                          * put the refcount back and put the page back on the
676                          * LRU
677                          */
678                         __put_page(page);
679                         SetPageLRU(page);
680                         list_add(&page->lru, &zone->active_list);
681                 } else {
682                         list_add(&page->lru, &l_hold);
683                         pgmoved++;
684                 }
685                 pgscanned++;
686         }
687         zone->pages_scanned += pgscanned;
688         zone->nr_active -= pgmoved;
689         spin_unlock_irq(&zone->lru_lock);
690 
691         /*
692          * `distress' is a measure of how much trouble we're having reclaiming
693          * pages.  0 -> no problems.  100 -> great trouble.
694          */
695         distress = 100 >> zone->prev_priority;
696 
697         /*
698          * The point of this algorithm is to decide when to start reclaiming
699          * mapped memory instead of just pagecache.  Work out how much memory
700          * is mapped.
701          */
702         mapped_ratio = (sc->nr_mapped * 100) / total_memory;
703 
704         /*
705          * Now decide how much we really want to unmap some pages.  The mapped
706          * ratio is downgraded - just because there's a lot of mapped memory
707          * doesn't necessarily mean that page reclaim isn't succeeding.
708          *
709          * The distress ratio is important - we don't want to start going oom.
710          *
711          * A 100% value of vm_swappiness overrides this algorithm altogether.
712          */
713         swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
714 
715         /*
716          * Now use this metric to decide whether to start moving mapped memory
717          * onto the inactive list.
718          */
719         if (swap_tendency >= 100)
720                 reclaim_mapped = 1;
721 
722         while (!list_empty(&l_hold)) {
723                 page = lru_to_page(&l_hold);
724                 list_del(&page->lru);
725                 if (page_mapped(page)) {
726                         if (!reclaim_mapped) {
727                                 list_add(&page->lru, &l_active);
728                                 continue;
729                         }
730                         page_map_lock(page);
731                         if (page_referenced(page)) {
732                                 page_map_unlock(page);
733                                 list_add(&page->lru, &l_active);
734                                 continue;
735                         }
736                         page_map_unlock(page);
737                 }
738                 /*
739                  * FIXME: need to consider page_count(page) here if/when we
740                  * reap orphaned pages via the LRU (Daniel's locking stuff)
741                  */
742                 if (total_swap_pages == 0 && PageAnon(page)) {
743                         list_add(&page->lru, &l_active);
744                         continue;
745                 }
746                 list_add(&page->lru, &l_inactive);
747         }
748 
749         pagevec_init(&pvec, 1);
750         pgmoved = 0;
751         spin_lock_irq(&zone->lru_lock);
752         while (!list_empty(&l_inactive)) {
753                 page = lru_to_page(&l_inactive);
754                 prefetchw_prev_lru_page(page, &l_inactive, flags);
755                 if (TestSetPageLRU(page))
756                         BUG();
757                 if (!TestClearPageActive(page))
758                         BUG();
759                 list_move(&page->lru, &zone->inactive_list);
760                 pgmoved++;
761                 if (!pagevec_add(&pvec, page)) {
762                         zone->nr_inactive += pgmoved;
763                         spin_unlock_irq(&zone->lru_lock);
764                         pgdeactivate += pgmoved;
765                         pgmoved = 0;
766                         if (buffer_heads_over_limit)
767                                 pagevec_strip(&pvec);
768                         __pagevec_release(&pvec);
769                         spin_lock_irq(&zone->lru_lock);
770                 }
771         }
772         zone->nr_inactive += pgmoved;
773         pgdeactivate += pgmoved;
774         if (buffer_heads_over_limit) {
775                 spin_unlock_irq(&zone->lru_lock);
776                 pagevec_strip(&pvec);
777                 spin_lock_irq(&zone->lru_lock);
778         }
779 
780         pgmoved = 0;
781         while (!list_empty(&l_active)) {
782                 page = lru_to_page(&l_active);
783                 prefetchw_prev_lru_page(page, &l_active, flags);
784                 if (TestSetPageLRU(page))
785                         BUG();
786                 BUG_ON(!PageActive(page));
787                 list_move(&page->lru, &zone->active_list);
788                 pgmoved++;
789                 if (!pagevec_add(&pvec, page)) {
790                         zone->nr_active += pgmoved;
791                         pgmoved = 0;
792                         spin_unlock_irq(&zone->lru_lock);
793                         __pagevec_release(&pvec);
794                         spin_lock_irq(&zone->lru_lock);
795                 }
796         }
797         zone->nr_active += pgmoved;
798         spin_unlock_irq(&zone->lru_lock);
799         pagevec_release(&pvec);
800 
801         mod_page_state_zone(zone, pgrefill, pgscanned);
802         mod_page_state(pgdeactivate, pgdeactivate);
803 }
804 
805 /*
806  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
807  */
808 static void
809 shrink_zone(struct zone *zone, struct scan_control *sc)
810 {
811         unsigned long nr_active;
812         unsigned long nr_inactive;
813 
814         /*
815          * Add one to `nr_to_scan' just to make sure that the kernel will
816          * slowly sift through the active list.
817          */
818         zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
819         nr_active = zone->nr_scan_active;
820         if (nr_active >= SWAP_CLUSTER_MAX)
821                 zone->nr_scan_active = 0;
822         else
823                 nr_active = 0;
824 
825         zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
826         nr_inactive = zone->nr_scan_inactive;
827         if (nr_inactive >= SWAP_CLUSTER_MAX)
828                 zone->nr_scan_inactive = 0;
829         else
830                 nr_inactive = 0;
831 
832         sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
833 
834         while (nr_active || nr_inactive) {
835                 if (nr_active) {
836                         sc->nr_to_scan = min(nr_active,
837                                         (unsigned long)SWAP_CLUSTER_MAX);
838                         nr_active -= sc->nr_to_scan;
839                         refill_inactive_zone(zone, sc);
840                 }
841 
842                 if (nr_inactive) {
843                         sc->nr_to_scan = min(nr_inactive,
844                                         (unsigned long)SWAP_CLUSTER_MAX);
845                         nr_inactive -= sc->nr_to_scan;
846                         shrink_cache(zone, sc);
847                         if (sc->nr_to_reclaim <= 0)
848                                 break;
849                 }
850         }
851 }
852 
853 /*
854  * This is the direct reclaim path, for page-allocating processes.  We only
855  * try to reclaim pages from zones which will satisfy the caller's allocation
856  * request.
857  *
858  * We reclaim from a zone even if that zone is over pages_high.  Because:
859  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
860  *    allocation or
861  * b) The zones may be over pages_high but they must go *over* pages_high to
862  *    satisfy the `incremental min' zone defense algorithm.
863  *
864  * Returns the number of reclaimed pages.
865  *
866  * If a zone is deemed to be full of pinned pages then just give it a light
867  * scan then give up on it.
868  */
869 static void
870 shrink_caches(struct zone **zones, struct scan_control *sc)
871 {
872         int i;
873 
874         for (i = 0; zones[i] != NULL; i++) {
875                 struct zone *zone = zones[i];
876 
877                 if (zone->present_pages == 0)
878                         continue;
879 
880                 zone->temp_priority = sc->priority;
881                 if (zone->prev_priority > sc->priority)
882                         zone->prev_priority = sc->priority;
883 
884                 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
885                         continue;       /* Let kswapd poll it */
886 
887                 shrink_zone(zone, sc);
888         }
889 }
890  
891 /*
892  * This is the main entry point to direct page reclaim.
893  *
894  * If a full scan of the inactive list fails to free enough memory then we
895  * are "out of memory" and something needs to be killed.
896  *
897  * If the caller is !__GFP_FS then the probability of a failure is reasonably
898  * high - the zone may be full of dirty or under-writeback pages, which this
899  * caller can't do much about.  We kick pdflush and take explicit naps in the
900  * hope that some of these pages can be written.  But if the allocating task
901  * holds filesystem locks which prevent writeout this might not work, and the
902  * allocation attempt will fail.
903  */
904 int try_to_free_pages(struct zone **zones,
905                 unsigned int gfp_mask, unsigned int order)
906 {
907         int priority;
908         int ret = 0;
909         int total_scanned = 0, total_reclaimed = 0;
910         struct reclaim_state *reclaim_state = current->reclaim_state;
911         struct scan_control sc;
912         unsigned long lru_pages = 0;
913         int i;
914 
915         sc.gfp_mask = gfp_mask;
916         sc.may_writepage = 0;
917 
918         inc_page_state(allocstall);
919 
920         for (i = 0; zones[i] != NULL; i++) {
921                 struct zone *zone = zones[i];
922 
923                 zone->temp_priority = DEF_PRIORITY;
924                 lru_pages += zone->nr_active + zone->nr_inactive;
925         }
926 
927         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
928                 sc.nr_mapped = read_page_state(nr_mapped);
929                 sc.nr_scanned = 0;
930                 sc.nr_reclaimed = 0;
931                 sc.priority = priority;
932                 shrink_caches(zones, &sc);
933                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
934                 if (reclaim_state) {
935                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
936                         reclaim_state->reclaimed_slab = 0;
937                 }
938                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
939                         ret = 1;
940                         goto out;
941                 }
942                 total_scanned += sc.nr_scanned;
943                 total_reclaimed += sc.nr_reclaimed;
944 
945                 /*
946                  * Try to write back as many pages as we just scanned.  This
947                  * tends to cause slow streaming writers to write data to the
948                  * disk smoothly, at the dirtying rate, which is nice.   But
949                  * that's undesirable in laptop mode, where we *want* lumpy
950                  * writeout.  So in laptop mode, write out the whole world.
951                  */
952                 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
953                         wakeup_bdflush(laptop_mode ? 0 : total_scanned);
954                         sc.may_writepage = 1;
955                 }
956 
957                 /* Take a nap, wait for some writeback to complete */
958                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
959                         blk_congestion_wait(WRITE, HZ/10);
960         }
961         if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
962                 out_of_memory(gfp_mask);
963 out:
964         for (i = 0; zones[i] != 0; i++)
965                 zones[i]->prev_priority = zones[i]->temp_priority;
966         return ret;
967 }
968 
969 /*
970  * For kswapd, balance_pgdat() will work across all this node's zones until
971  * they are all at pages_high.
972  *
973  * If `nr_pages' is non-zero then it is the number of pages which are to be
974  * reclaimed, regardless of the zone occupancies.  This is a software suspend
975  * special.
976  *
977  * Returns the number of pages which were actually freed.
978  *
979  * There is special handling here for zones which are full of pinned pages.
980  * This can happen if the pages are all mlocked, or if they are all used by
981  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
982  * What we do is to detect the case where all pages in the zone have been
983  * scanned twice and there has been zero successful reclaim.  Mark the zone as
984  * dead and from now on, only perform a short scan.  Basically we're polling
985  * the zone for when the problem goes away.
986  *
987  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
988  * zones which have free_pages > pages_high, but once a zone is found to have
989  * free_pages <= pages_high, we scan that zone and the lower zones regardless
990  * of the number of free pages in the lower zones.  This interoperates with
991  * the page allocator fallback scheme to ensure that aging of pages is balanced
992  * across the zones.
993  */
994 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
995 {
996         int to_free = nr_pages;
997         int all_zones_ok;
998         int priority;
999         int i;
1000         int total_scanned, total_reclaimed;
1001         struct reclaim_state *reclaim_state = current->reclaim_state;
1002         struct scan_control sc;
1003 
1004 loop_again:
1005         total_scanned = 0;
1006         total_reclaimed = 0;
1007         sc.gfp_mask = GFP_KERNEL;
1008         sc.may_writepage = 0;
1009         sc.nr_mapped = read_page_state(nr_mapped);
1010 
1011         inc_page_state(pageoutrun);
1012 
1013         for (i = 0; i < pgdat->nr_zones; i++) {
1014                 struct zone *zone = pgdat->node_zones + i;
1015 
1016                 zone->temp_priority = DEF_PRIORITY;
1017         }
1018 
1019         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1020                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1021                 unsigned long lru_pages = 0;
1022 
1023                 all_zones_ok = 1;
1024 
1025                 if (nr_pages == 0) {
1026                         /*
1027                          * Scan in the highmem->dma direction for the highest
1028                          * zone which needs scanning
1029                          */
1030                         for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1031                                 struct zone *zone = pgdat->node_zones + i;
1032 
1033                                 if (zone->present_pages == 0)
1034                                         continue;
1035 
1036                                 if (zone->all_unreclaimable &&
1037                                                 priority != DEF_PRIORITY)
1038                                         continue;
1039 
1040                                 if (zone->free_pages <= zone->pages_high) {
1041                                         end_zone = i;
1042                                         goto scan;
1043                                 }
1044                         }
1045                         goto out;
1046                 } else {
1047                         end_zone = pgdat->nr_zones - 1;
1048                 }
1049 scan:
1050                 for (i = 0; i <= end_zone; i++) {
1051                         struct zone *zone = pgdat->node_zones + i;
1052 
1053                         lru_pages += zone->nr_active + zone->nr_inactive;
1054                 }
1055 
1056                 /*
1057                  * Now scan the zone in the dma->highmem direction, stopping
1058                  * at the last zone which needs scanning.
1059                  *
1060                  * We do this because the page allocator works in the opposite
1061                  * direction.  This prevents the page allocator from allocating
1062                  * pages behind kswapd's direction of progress, which would
1063                  * cause too much scanning of the lower zones.
1064                  */
1065                 for (i = 0; i <= end_zone; i++) {
1066                         struct zone *zone = pgdat->node_zones + i;
1067 
1068                         if (zone->present_pages == 0)
1069                                 continue;
1070 
1071                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1072                                 continue;
1073 
1074                         if (nr_pages == 0) {    /* Not software suspend */
1075                                 if (zone->free_pages <= zone->pages_high)
1076                                         all_zones_ok = 0;
1077                         }
1078                         zone->temp_priority = priority;
1079                         if (zone->prev_priority > priority)
1080                                 zone->prev_priority = priority;
1081                         sc.nr_scanned = 0;
1082                         sc.nr_reclaimed = 0;
1083                         sc.priority = priority;
1084                         shrink_zone(zone, &sc);
1085                         reclaim_state->reclaimed_slab = 0;
1086                         shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1087                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1088                         total_reclaimed += sc.nr_reclaimed;
1089                         if (zone->all_unreclaimable)
1090                                 continue;
1091                         if (zone->pages_scanned >= (zone->nr_active +
1092                                                         zone->nr_inactive) * 4)
1093                                 zone->all_unreclaimable = 1;
1094                         /*
1095                          * If we've done a decent amount of scanning and
1096                          * the reclaim ratio is low, start doing writepage
1097                          * even in laptop mode
1098                          */
1099                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1100                             total_scanned > total_reclaimed+total_reclaimed/2)
1101                                 sc.may_writepage = 1;
1102                 }
1103                 if (nr_pages && to_free > total_reclaimed)
1104                         continue;       /* swsusp: need to do more work */
1105                 if (all_zones_ok)
1106                         break;          /* kswapd: all done */
1107                 /*
1108                  * OK, kswapd is getting into trouble.  Take a nap, then take
1109                  * another pass across the zones.
1110                  */
1111                 if (total_scanned && priority < DEF_PRIORITY - 2)
1112                         blk_congestion_wait(WRITE, HZ/10);
1113 
1114                 /*
1115                  * We do this so kswapd doesn't build up large priorities for
1116                  * example when it is freeing in parallel with allocators. It
1117                  * matches the direct reclaim path behaviour in terms of impact
1118                  * on zone->*_priority.
1119                  */
1120                 if (total_reclaimed >= SWAP_CLUSTER_MAX)
1121                         break;
1122         }
1123 out:
1124         for (i = 0; i < pgdat->nr_zones; i++) {
1125                 struct zone *zone = pgdat->node_zones + i;
1126 
1127                 zone->prev_priority = zone->temp_priority;
1128         }
1129         if (!all_zones_ok) {
1130                 cond_resched();
1131                 goto loop_again;
1132         }
1133 
1134         return total_reclaimed;
1135 }
1136 
1137 /*
1138  * The background pageout daemon, started as a kernel thread
1139  * from the init process. 
1140  *
1141  * This basically trickles out pages so that we have _some_
1142  * free memory available even if there is no other activity
1143  * that frees anything up. This is needed for things like routing
1144  * etc, where we otherwise might have all activity going on in
1145  * asynchronous contexts that cannot page things out.
1146  *
1147  * If there are applications that are active memory-allocators
1148  * (most normal use), this basically shouldn't matter.
1149  */
1150 static int kswapd(void *p)
1151 {
1152         pg_data_t *pgdat = (pg_data_t*)p;
1153         struct task_struct *tsk = current;
1154         DEFINE_WAIT(wait);
1155         struct reclaim_state reclaim_state = {
1156                 .reclaimed_slab = 0,
1157         };
1158         cpumask_t cpumask;
1159 
1160         daemonize("kswapd%d", pgdat->node_id);
1161         cpumask = node_to_cpumask(pgdat->node_id);
1162         if (!cpus_empty(cpumask))
1163                 set_cpus_allowed(tsk, cpumask);
1164         current->reclaim_state = &reclaim_state;
1165 
1166         /*
1167          * Tell the memory management that we're a "memory allocator",
1168          * and that if we need more memory we should get access to it
1169          * regardless (see "__alloc_pages()"). "kswapd" should
1170          * never get caught in the normal page freeing logic.
1171          *
1172          * (Kswapd normally doesn't need memory anyway, but sometimes
1173          * you need a small amount of memory in order to be able to
1174          * page out something else, and this flag essentially protects
1175          * us from recursively trying to free more memory as we're
1176          * trying to free the first piece of memory in the first place).
1177          */
1178         tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1179 
1180         for ( ; ; ) {
1181                 if (current->flags & PF_FREEZE)
1182                         refrigerator(PF_FREEZE);
1183                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1184                 schedule();
1185                 finish_wait(&pgdat->kswapd_wait, &wait);
1186 
1187                 balance_pgdat(pgdat, 0);
1188         }
1189         return 0;
1190 }
1191 
1192 /*
1193  * A zone is low on free memory, so wake its kswapd task to service it.
1194  */
1195 void wakeup_kswapd(struct zone *zone)
1196 {
1197         if (zone->present_pages == 0)
1198                 return;
1199         if (zone->free_pages > zone->pages_low)
1200                 return;
1201         if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1202                 return;
1203         wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1204 }
1205 
1206 #ifdef CONFIG_PM
1207 /*
1208  * Try to free `nr_pages' of memory, system-wide.  Returns the number of freed
1209  * pages.
1210  */
1211 int shrink_all_memory(int nr_pages)
1212 {
1213         pg_data_t *pgdat;
1214         int nr_to_free = nr_pages;
1215         int ret = 0;
1216         struct reclaim_state reclaim_state = {
1217                 .reclaimed_slab = 0,
1218         };
1219 
1220         current->reclaim_state = &reclaim_state;
1221         for_each_pgdat(pgdat) {
1222                 int freed;
1223                 freed = balance_pgdat(pgdat, nr_to_free);
1224                 ret += freed;
1225                 nr_to_free -= freed;
1226                 if (nr_to_free <= 0)
1227                         break;
1228         }
1229         current->reclaim_state = NULL;
1230         return ret;
1231 }
1232 #endif
1233 
1234 #ifdef CONFIG_HOTPLUG_CPU
1235 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1236    not required for correctness.  So if the last cpu in a node goes
1237    away, we get changed to run anywhere: as the first one comes back,
1238    restore their cpu bindings. */
1239 static int __devinit cpu_callback(struct notifier_block *nfb,
1240                                   unsigned long action,
1241                                   void *hcpu)
1242 {
1243         pg_data_t *pgdat;
1244         cpumask_t mask;
1245 
1246         if (action == CPU_ONLINE) {
1247                 for_each_pgdat(pgdat) {
1248                         mask = node_to_cpumask(pgdat->node_id);
1249                         if (any_online_cpu(mask) != NR_CPUS)
1250                                 /* One of our CPUs online: restore mask */
1251                                 set_cpus_allowed(pgdat->kswapd, mask);
1252                 }
1253         }
1254         return NOTIFY_OK;
1255 }
1256 #endif /* CONFIG_HOTPLUG_CPU */
1257 
1258 static int __init kswapd_init(void)
1259 {
1260         pg_data_t *pgdat;
1261         swap_setup();
1262         for_each_pgdat(pgdat)
1263                 pgdat->kswapd
1264                 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1265         total_memory = nr_free_pagecache_pages();
1266         hotcpu_notifier(cpu_callback, 0);
1267         return 0;
1268 }
1269 
1270 module_init(kswapd_init)
1271 

This page was automatically generated by LXR 0.3.1. •  Linux is a registered trademark of Linus Torvalds