tcp: enforce tcp_min_snd_mss in tcp_mtu_probing()
[pandora-kernel.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/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108
109         /*
110          * Nodemask of nodes allowed by the caller. If NULL, all nodes
111          * are scanned.
112          */
113         nodemask_t      *nodemask;
114 };
115
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field)                    \
120         do {                                                            \
121                 if ((_page)->lru.prev != _base) {                       \
122                         struct page *prev;                              \
123                                                                         \
124                         prev = lru_to_page(&(_page->lru));              \
125                         prefetch(&prev->_field);                        \
126                 }                                                       \
127         } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
131
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
134         do {                                                            \
135                 if ((_page)->lru.prev != _base) {                       \
136                         struct page *prev;                              \
137                                                                         \
138                         prev = lru_to_page(&(_page->lru));              \
139                         prefetchw(&prev->_field);                       \
140                 }                                                       \
141         } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
145
146 /*
147  * From 0 .. 100.  Higher means more swappy.
148  */
149 int vm_swappiness = 60;
150 long vm_total_pages;    /* The total number of pages which the VM controls */
151
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
154
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
160
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162                                                   struct scan_control *sc)
163 {
164         if (!scanning_global_lru(sc))
165                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
167         return &zone->reclaim_stat;
168 }
169
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171                                 struct scan_control *sc, enum lru_list lru)
172 {
173         if (!scanning_global_lru(sc))
174                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175                                 zone_to_nid(zone), zone_idx(zone), BIT(lru));
176
177         return zone_page_state(zone, NR_LRU_BASE + lru);
178 }
179
180
181 /*
182  * Add a shrinker callback to be called from the vm
183  */
184 void register_shrinker(struct shrinker *shrinker)
185 {
186         atomic_long_set(&shrinker->nr_in_batch, 0);
187         down_write(&shrinker_rwsem);
188         list_add_tail(&shrinker->list, &shrinker_list);
189         up_write(&shrinker_rwsem);
190 }
191 EXPORT_SYMBOL(register_shrinker);
192
193 /*
194  * Remove one
195  */
196 void unregister_shrinker(struct shrinker *shrinker)
197 {
198         down_write(&shrinker_rwsem);
199         list_del(&shrinker->list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(unregister_shrinker);
203
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205                                      struct shrink_control *sc,
206                                      unsigned long nr_to_scan)
207 {
208         sc->nr_to_scan = nr_to_scan;
209         return (*shrinker->shrink)(shrinker, sc);
210 }
211
212 #define SHRINK_BATCH 128
213 /*
214  * Call the shrink functions to age shrinkable caches
215  *
216  * Here we assume it costs one seek to replace a lru page and that it also
217  * takes a seek to recreate a cache object.  With this in mind we age equal
218  * percentages of the lru and ageable caches.  This should balance the seeks
219  * generated by these structures.
220  *
221  * If the vm encountered mapped pages on the LRU it increase the pressure on
222  * slab to avoid swapping.
223  *
224  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
225  *
226  * `lru_pages' represents the number of on-LRU pages in all the zones which
227  * are eligible for the caller's allocation attempt.  It is used for balancing
228  * slab reclaim versus page reclaim.
229  *
230  * Returns the number of slab objects which we shrunk.
231  */
232 unsigned long shrink_slab(struct shrink_control *shrink,
233                           unsigned long nr_pages_scanned,
234                           unsigned long lru_pages)
235 {
236         struct shrinker *shrinker;
237         unsigned long ret = 0;
238
239         if (nr_pages_scanned == 0)
240                 nr_pages_scanned = SWAP_CLUSTER_MAX;
241
242         if (!down_read_trylock(&shrinker_rwsem)) {
243                 /* Assume we'll be able to shrink next time */
244                 ret = 1;
245                 goto out;
246         }
247
248         list_for_each_entry(shrinker, &shrinker_list, list) {
249                 unsigned long long delta;
250                 long total_scan;
251                 long max_pass;
252                 int shrink_ret = 0;
253                 long nr;
254                 long new_nr;
255                 long batch_size = shrinker->batch ? shrinker->batch
256                                                   : SHRINK_BATCH;
257
258                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
259                 if (max_pass <= 0)
260                         continue;
261
262                 /*
263                  * copy the current shrinker scan count into a local variable
264                  * and zero it so that other concurrent shrinker invocations
265                  * don't also do this scanning work.
266                  */
267                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
268
269                 total_scan = nr;
270                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
271                 delta *= max_pass;
272                 do_div(delta, lru_pages + 1);
273                 total_scan += delta;
274                 if (total_scan < 0) {
275                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
276                                "delete nr=%ld\n",
277                                shrinker->shrink, total_scan);
278                         total_scan = max_pass;
279                 }
280
281                 /*
282                  * We need to avoid excessive windup on filesystem shrinkers
283                  * due to large numbers of GFP_NOFS allocations causing the
284                  * shrinkers to return -1 all the time. This results in a large
285                  * nr being built up so when a shrink that can do some work
286                  * comes along it empties the entire cache due to nr >>>
287                  * max_pass.  This is bad for sustaining a working set in
288                  * memory.
289                  *
290                  * Hence only allow the shrinker to scan the entire cache when
291                  * a large delta change is calculated directly.
292                  */
293                 if (delta < max_pass / 4)
294                         total_scan = min(total_scan, max_pass / 2);
295
296                 /*
297                  * Avoid risking looping forever due to too large nr value:
298                  * never try to free more than twice the estimate number of
299                  * freeable entries.
300                  */
301                 if (total_scan > max_pass * 2)
302                         total_scan = max_pass * 2;
303
304                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305                                         nr_pages_scanned, lru_pages,
306                                         max_pass, delta, total_scan);
307
308                 while (total_scan >= batch_size) {
309                         int nr_before;
310
311                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
313                                                         batch_size);
314                         if (shrink_ret == -1)
315                                 break;
316                         if (shrink_ret < nr_before)
317                                 ret += nr_before - shrink_ret;
318                         count_vm_events(SLABS_SCANNED, batch_size);
319                         total_scan -= batch_size;
320
321                         cond_resched();
322                 }
323
324                 /*
325                  * move the unused scan count back into the shrinker in a
326                  * manner that handles concurrent updates. If we exhausted the
327                  * scan, there is no need to do an update.
328                  */
329                 if (total_scan > 0)
330                         new_nr = atomic_long_add_return(total_scan,
331                                         &shrinker->nr_in_batch);
332                 else
333                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
334
335                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
336         }
337         up_read(&shrinker_rwsem);
338 out:
339         cond_resched();
340         return ret;
341 }
342
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
344                                    bool sync)
345 {
346         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
347
348         /*
349          * Initially assume we are entering either lumpy reclaim or
350          * reclaim/compaction.Depending on the order, we will either set the
351          * sync mode or just reclaim order-0 pages later.
352          */
353         if (COMPACTION_BUILD)
354                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
355         else
356                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
357
358         /*
359          * Avoid using lumpy reclaim or reclaim/compaction if possible by
360          * restricting when its set to either costly allocations or when
361          * under memory pressure
362          */
363         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364                 sc->reclaim_mode |= syncmode;
365         else if (sc->order && priority < DEF_PRIORITY - 2)
366                 sc->reclaim_mode |= syncmode;
367         else
368                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
369 }
370
371 static void reset_reclaim_mode(struct scan_control *sc)
372 {
373         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
374 }
375
376 static inline int is_page_cache_freeable(struct page *page)
377 {
378         /*
379          * A freeable page cache page is referenced only by the caller
380          * that isolated the page, the page cache radix tree and
381          * optional buffer heads at page->private.
382          */
383         return page_count(page) - page_has_private(page) == 2;
384 }
385
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387                               struct scan_control *sc)
388 {
389         if (current->flags & PF_SWAPWRITE)
390                 return 1;
391         if (!bdi_write_congested(bdi))
392                 return 1;
393         if (bdi == current->backing_dev_info)
394                 return 1;
395
396         /* lumpy reclaim for hugepage often need a lot of write */
397         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398                 return 1;
399         return 0;
400 }
401
402 /*
403  * We detected a synchronous write error writing a page out.  Probably
404  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
405  * fsync(), msync() or close().
406  *
407  * The tricky part is that after writepage we cannot touch the mapping: nothing
408  * prevents it from being freed up.  But we have a ref on the page and once
409  * that page is locked, the mapping is pinned.
410  *
411  * We're allowed to run sleeping lock_page() here because we know the caller has
412  * __GFP_FS.
413  */
414 static void handle_write_error(struct address_space *mapping,
415                                 struct page *page, int error)
416 {
417         lock_page(page);
418         if (page_mapping(page) == mapping)
419                 mapping_set_error(mapping, error);
420         unlock_page(page);
421 }
422
423 /* possible outcome of pageout() */
424 typedef enum {
425         /* failed to write page out, page is locked */
426         PAGE_KEEP,
427         /* move page to the active list, page is locked */
428         PAGE_ACTIVATE,
429         /* page has been sent to the disk successfully, page is unlocked */
430         PAGE_SUCCESS,
431         /* page is clean and locked */
432         PAGE_CLEAN,
433 } pageout_t;
434
435 /*
436  * pageout is called by shrink_page_list() for each dirty page.
437  * Calls ->writepage().
438  */
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440                          struct scan_control *sc)
441 {
442         /*
443          * If the page is dirty, only perform writeback if that write
444          * will be non-blocking.  To prevent this allocation from being
445          * stalled by pagecache activity.  But note that there may be
446          * stalls if we need to run get_block().  We could test
447          * PagePrivate for that.
448          *
449          * If this process is currently in __generic_file_aio_write() against
450          * this page's queue, we can perform writeback even if that
451          * will block.
452          *
453          * If the page is swapcache, write it back even if that would
454          * block, for some throttling. This happens by accident, because
455          * swap_backing_dev_info is bust: it doesn't reflect the
456          * congestion state of the swapdevs.  Easy to fix, if needed.
457          */
458         if (!is_page_cache_freeable(page))
459                 return PAGE_KEEP;
460         if (!mapping) {
461                 /*
462                  * Some data journaling orphaned pages can have
463                  * page->mapping == NULL while being dirty with clean buffers.
464                  */
465                 if (page_has_private(page)) {
466                         if (try_to_free_buffers(page)) {
467                                 ClearPageDirty(page);
468                                 printk("%s: orphaned page\n", __func__);
469                                 return PAGE_CLEAN;
470                         }
471                 }
472                 return PAGE_KEEP;
473         }
474         if (mapping->a_ops->writepage == NULL)
475                 return PAGE_ACTIVATE;
476         if (!may_write_to_queue(mapping->backing_dev_info, sc))
477                 return PAGE_KEEP;
478
479         if (clear_page_dirty_for_io(page)) {
480                 int res;
481                 struct writeback_control wbc = {
482                         .sync_mode = WB_SYNC_NONE,
483                         .nr_to_write = SWAP_CLUSTER_MAX,
484                         .range_start = 0,
485                         .range_end = LLONG_MAX,
486                         .for_reclaim = 1,
487                 };
488
489                 SetPageReclaim(page);
490                 res = mapping->a_ops->writepage(page, &wbc);
491                 if (res < 0)
492                         handle_write_error(mapping, page, res);
493                 if (res == AOP_WRITEPAGE_ACTIVATE) {
494                         ClearPageReclaim(page);
495                         return PAGE_ACTIVATE;
496                 }
497
498                 if (!PageWriteback(page)) {
499                         /* synchronous write or broken a_ops? */
500                         ClearPageReclaim(page);
501                 }
502                 trace_mm_vmscan_writepage(page,
503                         trace_reclaim_flags(page, sc->reclaim_mode));
504                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
505                 return PAGE_SUCCESS;
506         }
507
508         return PAGE_CLEAN;
509 }
510
511 /*
512  * Same as remove_mapping, but if the page is removed from the mapping, it
513  * gets returned with a refcount of 0.
514  */
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
516 {
517         BUG_ON(!PageLocked(page));
518         BUG_ON(mapping != page_mapping(page));
519
520         spin_lock_irq(&mapping->tree_lock);
521         /*
522          * The non racy check for a busy page.
523          *
524          * Must be careful with the order of the tests. When someone has
525          * a ref to the page, it may be possible that they dirty it then
526          * drop the reference. So if PageDirty is tested before page_count
527          * here, then the following race may occur:
528          *
529          * get_user_pages(&page);
530          * [user mapping goes away]
531          * write_to(page);
532          *                              !PageDirty(page)    [good]
533          * SetPageDirty(page);
534          * put_page(page);
535          *                              !page_count(page)   [good, discard it]
536          *
537          * [oops, our write_to data is lost]
538          *
539          * Reversing the order of the tests ensures such a situation cannot
540          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541          * load is not satisfied before that of page->_count.
542          *
543          * Note that if SetPageDirty is always performed via set_page_dirty,
544          * and thus under tree_lock, then this ordering is not required.
545          */
546         if (!page_freeze_refs(page, 2))
547                 goto cannot_free;
548         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549         if (unlikely(PageDirty(page))) {
550                 page_unfreeze_refs(page, 2);
551                 goto cannot_free;
552         }
553
554         if (PageSwapCache(page)) {
555                 swp_entry_t swap = { .val = page_private(page) };
556                 __delete_from_swap_cache(page);
557                 spin_unlock_irq(&mapping->tree_lock);
558                 swapcache_free(swap, page);
559         } else {
560                 void (*freepage)(struct page *);
561
562                 freepage = mapping->a_ops->freepage;
563
564                 __delete_from_page_cache(page);
565                 spin_unlock_irq(&mapping->tree_lock);
566                 mem_cgroup_uncharge_cache_page(page);
567
568                 if (freepage != NULL)
569                         freepage(page);
570         }
571
572         return 1;
573
574 cannot_free:
575         spin_unlock_irq(&mapping->tree_lock);
576         return 0;
577 }
578
579 /*
580  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
581  * someone else has a ref on the page, abort and return 0.  If it was
582  * successfully detached, return 1.  Assumes the caller has a single ref on
583  * this page.
584  */
585 int remove_mapping(struct address_space *mapping, struct page *page)
586 {
587         if (__remove_mapping(mapping, page)) {
588                 /*
589                  * Unfreezing the refcount with 1 rather than 2 effectively
590                  * drops the pagecache ref for us without requiring another
591                  * atomic operation.
592                  */
593                 page_unfreeze_refs(page, 1);
594                 return 1;
595         }
596         return 0;
597 }
598
599 /**
600  * putback_lru_page - put previously isolated page onto appropriate LRU list
601  * @page: page to be put back to appropriate lru list
602  *
603  * Add previously isolated @page to appropriate LRU list.
604  * Page may still be unevictable for other reasons.
605  *
606  * lru_lock must not be held, interrupts must be enabled.
607  */
608 void putback_lru_page(struct page *page)
609 {
610         int lru;
611         int active = !!TestClearPageActive(page);
612         int was_unevictable = PageUnevictable(page);
613
614         VM_BUG_ON(PageLRU(page));
615
616 redo:
617         ClearPageUnevictable(page);
618
619         if (page_evictable(page, NULL)) {
620                 /*
621                  * For evictable pages, we can use the cache.
622                  * In event of a race, worst case is we end up with an
623                  * unevictable page on [in]active list.
624                  * We know how to handle that.
625                  */
626                 lru = active + page_lru_base_type(page);
627                 lru_cache_add_lru(page, lru);
628         } else {
629                 /*
630                  * Put unevictable pages directly on zone's unevictable
631                  * list.
632                  */
633                 lru = LRU_UNEVICTABLE;
634                 add_page_to_unevictable_list(page);
635                 /*
636                  * When racing with an mlock or AS_UNEVICTABLE clearing
637                  * (page is unlocked) make sure that if the other thread
638                  * does not observe our setting of PG_lru and fails
639                  * isolation/check_move_unevictable_pages,
640                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641                  * the page back to the evictable list.
642                  *
643                  * The other side is TestClearPageMlocked() or shmem_lock().
644                  */
645                 smp_mb();
646         }
647
648         /*
649          * page's status can change while we move it among lru. If an evictable
650          * page is on unevictable list, it never be freed. To avoid that,
651          * check after we added it to the list, again.
652          */
653         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654                 if (!isolate_lru_page(page)) {
655                         put_page(page);
656                         goto redo;
657                 }
658                 /* This means someone else dropped this page from LRU
659                  * So, it will be freed or putback to LRU again. There is
660                  * nothing to do here.
661                  */
662         }
663
664         if (was_unevictable && lru != LRU_UNEVICTABLE)
665                 count_vm_event(UNEVICTABLE_PGRESCUED);
666         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667                 count_vm_event(UNEVICTABLE_PGCULLED);
668
669         put_page(page);         /* drop ref from isolate */
670 }
671
672 enum page_references {
673         PAGEREF_RECLAIM,
674         PAGEREF_RECLAIM_CLEAN,
675         PAGEREF_KEEP,
676         PAGEREF_ACTIVATE,
677 };
678
679 static enum page_references page_check_references(struct page *page,
680                                                   struct scan_control *sc)
681 {
682         int referenced_ptes, referenced_page;
683         unsigned long vm_flags;
684
685         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686         referenced_page = TestClearPageReferenced(page);
687
688         /* Lumpy reclaim - ignore references */
689         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690                 return PAGEREF_RECLAIM;
691
692         /*
693          * Mlock lost the isolation race with us.  Let try_to_unmap()
694          * move the page to the unevictable list.
695          */
696         if (vm_flags & VM_LOCKED)
697                 return PAGEREF_RECLAIM;
698
699         if (referenced_ptes) {
700                 if (PageSwapBacked(page))
701                         return PAGEREF_ACTIVATE;
702                 /*
703                  * All mapped pages start out with page table
704                  * references from the instantiating fault, so we need
705                  * to look twice if a mapped file page is used more
706                  * than once.
707                  *
708                  * Mark it and spare it for another trip around the
709                  * inactive list.  Another page table reference will
710                  * lead to its activation.
711                  *
712                  * Note: the mark is set for activated pages as well
713                  * so that recently deactivated but used pages are
714                  * quickly recovered.
715                  */
716                 SetPageReferenced(page);
717
718                 if (referenced_page || referenced_ptes > 1)
719                         return PAGEREF_ACTIVATE;
720
721                 /*
722                  * Activate file-backed executable pages after first usage.
723                  */
724                 if (vm_flags & VM_EXEC)
725                         return PAGEREF_ACTIVATE;
726
727                 return PAGEREF_KEEP;
728         }
729
730         /* Reclaim if clean, defer dirty pages to writeback */
731         if (referenced_page && !PageSwapBacked(page))
732                 return PAGEREF_RECLAIM_CLEAN;
733
734         return PAGEREF_RECLAIM;
735 }
736
737 /*
738  * shrink_page_list() returns the number of reclaimed pages
739  */
740 static unsigned long shrink_page_list(struct list_head *page_list,
741                                       struct zone *zone,
742                                       struct scan_control *sc,
743                                       int priority,
744                                       unsigned long *ret_nr_dirty,
745                                       unsigned long *ret_nr_writeback)
746 {
747         LIST_HEAD(ret_pages);
748         LIST_HEAD(free_pages);
749         int pgactivate = 0;
750         unsigned long nr_dirty = 0;
751         unsigned long nr_congested = 0;
752         unsigned long nr_reclaimed = 0;
753         unsigned long nr_writeback = 0;
754
755         cond_resched();
756
757         while (!list_empty(page_list)) {
758                 enum page_references references;
759                 struct address_space *mapping;
760                 struct page *page;
761                 int may_enter_fs;
762
763                 cond_resched();
764
765                 page = lru_to_page(page_list);
766                 list_del(&page->lru);
767
768                 if (!trylock_page(page))
769                         goto keep;
770
771                 VM_BUG_ON(PageActive(page));
772                 VM_BUG_ON(page_zone(page) != zone);
773
774                 sc->nr_scanned++;
775
776                 if (unlikely(!page_evictable(page, NULL)))
777                         goto cull_mlocked;
778
779                 if (!sc->may_unmap && page_mapped(page))
780                         goto keep_locked;
781
782                 /* Double the slab pressure for mapped and swapcache pages */
783                 if (page_mapped(page) || PageSwapCache(page))
784                         sc->nr_scanned++;
785
786                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
787                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
788
789                 if (PageWriteback(page)) {
790                         nr_writeback++;
791                         /*
792                          * Synchronous reclaim cannot queue pages for
793                          * writeback due to the possibility of stack overflow
794                          * but if it encounters a page under writeback, wait
795                          * for the IO to complete.
796                          */
797                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
798                             may_enter_fs)
799                                 wait_on_page_writeback(page);
800                         else {
801                                 unlock_page(page);
802                                 goto keep_lumpy;
803                         }
804                 }
805
806                 references = page_check_references(page, sc);
807                 switch (references) {
808                 case PAGEREF_ACTIVATE:
809                         goto activate_locked;
810                 case PAGEREF_KEEP:
811                         goto keep_locked;
812                 case PAGEREF_RECLAIM:
813                 case PAGEREF_RECLAIM_CLEAN:
814                         ; /* try to reclaim the page below */
815                 }
816
817                 /*
818                  * Anonymous process memory has backing store?
819                  * Try to allocate it some swap space here.
820                  */
821                 if (PageAnon(page) && !PageSwapCache(page)) {
822                         if (!(sc->gfp_mask & __GFP_IO))
823                                 goto keep_locked;
824                         if (!add_to_swap(page))
825                                 goto activate_locked;
826                         may_enter_fs = 1;
827                 }
828
829                 mapping = page_mapping(page);
830
831                 /*
832                  * The page is mapped into the page tables of one or more
833                  * processes. Try to unmap it here.
834                  */
835                 if (page_mapped(page) && mapping) {
836                         switch (try_to_unmap(page, TTU_UNMAP)) {
837                         case SWAP_FAIL:
838                                 goto activate_locked;
839                         case SWAP_AGAIN:
840                                 goto keep_locked;
841                         case SWAP_MLOCK:
842                                 goto cull_mlocked;
843                         case SWAP_SUCCESS:
844                                 ; /* try to free the page below */
845                         }
846                 }
847
848                 if (PageDirty(page)) {
849                         nr_dirty++;
850
851                         /*
852                          * Only kswapd can writeback filesystem pages to
853                          * avoid risk of stack overflow but do not writeback
854                          * unless under significant pressure.
855                          */
856                         if (page_is_file_cache(page) &&
857                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
858                                 /*
859                                  * Immediately reclaim when written back.
860                                  * Similar in principal to deactivate_page()
861                                  * except we already have the page isolated
862                                  * and know it's dirty
863                                  */
864                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
865                                 SetPageReclaim(page);
866
867                                 goto keep_locked;
868                         }
869
870                         if (references == PAGEREF_RECLAIM_CLEAN)
871                                 goto keep_locked;
872                         if (!may_enter_fs)
873                                 goto keep_locked;
874                         if (!sc->may_writepage)
875                                 goto keep_locked;
876
877                         /* Page is dirty, try to write it out here */
878                         switch (pageout(page, mapping, sc)) {
879                         case PAGE_KEEP:
880                                 nr_congested++;
881                                 goto keep_locked;
882                         case PAGE_ACTIVATE:
883                                 goto activate_locked;
884                         case PAGE_SUCCESS:
885                                 if (PageWriteback(page))
886                                         goto keep_lumpy;
887                                 if (PageDirty(page))
888                                         goto keep;
889
890                                 /*
891                                  * A synchronous write - probably a ramdisk.  Go
892                                  * ahead and try to reclaim the page.
893                                  */
894                                 if (!trylock_page(page))
895                                         goto keep;
896                                 if (PageDirty(page) || PageWriteback(page))
897                                         goto keep_locked;
898                                 mapping = page_mapping(page);
899                         case PAGE_CLEAN:
900                                 ; /* try to free the page below */
901                         }
902                 }
903
904                 /*
905                  * If the page has buffers, try to free the buffer mappings
906                  * associated with this page. If we succeed we try to free
907                  * the page as well.
908                  *
909                  * We do this even if the page is PageDirty().
910                  * try_to_release_page() does not perform I/O, but it is
911                  * possible for a page to have PageDirty set, but it is actually
912                  * clean (all its buffers are clean).  This happens if the
913                  * buffers were written out directly, with submit_bh(). ext3
914                  * will do this, as well as the blockdev mapping.
915                  * try_to_release_page() will discover that cleanness and will
916                  * drop the buffers and mark the page clean - it can be freed.
917                  *
918                  * Rarely, pages can have buffers and no ->mapping.  These are
919                  * the pages which were not successfully invalidated in
920                  * truncate_complete_page().  We try to drop those buffers here
921                  * and if that worked, and the page is no longer mapped into
922                  * process address space (page_count == 1) it can be freed.
923                  * Otherwise, leave the page on the LRU so it is swappable.
924                  */
925                 if (page_has_private(page)) {
926                         if (!try_to_release_page(page, sc->gfp_mask))
927                                 goto activate_locked;
928                         if (!mapping && page_count(page) == 1) {
929                                 unlock_page(page);
930                                 if (put_page_testzero(page))
931                                         goto free_it;
932                                 else {
933                                         /*
934                                          * rare race with speculative reference.
935                                          * the speculative reference will free
936                                          * this page shortly, so we may
937                                          * increment nr_reclaimed here (and
938                                          * leave it off the LRU).
939                                          */
940                                         nr_reclaimed++;
941                                         continue;
942                                 }
943                         }
944                 }
945
946                 if (!mapping || !__remove_mapping(mapping, page))
947                         goto keep_locked;
948
949                 /*
950                  * At this point, we have no other references and there is
951                  * no way to pick any more up (removed from LRU, removed
952                  * from pagecache). Can use non-atomic bitops now (and
953                  * we obviously don't have to worry about waking up a process
954                  * waiting on the page lock, because there are no references.
955                  */
956                 __clear_page_locked(page);
957 free_it:
958                 nr_reclaimed++;
959
960                 /*
961                  * Is there need to periodically free_page_list? It would
962                  * appear not as the counts should be low
963                  */
964                 list_add(&page->lru, &free_pages);
965                 continue;
966
967 cull_mlocked:
968                 if (PageSwapCache(page))
969                         try_to_free_swap(page);
970                 unlock_page(page);
971                 putback_lru_page(page);
972                 reset_reclaim_mode(sc);
973                 continue;
974
975 activate_locked:
976                 /* Not a candidate for swapping, so reclaim swap space. */
977                 if (PageSwapCache(page) && vm_swap_full())
978                         try_to_free_swap(page);
979                 VM_BUG_ON(PageActive(page));
980                 SetPageActive(page);
981                 pgactivate++;
982 keep_locked:
983                 unlock_page(page);
984 keep:
985                 reset_reclaim_mode(sc);
986 keep_lumpy:
987                 list_add(&page->lru, &ret_pages);
988                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
989         }
990
991         /*
992          * Tag a zone as congested if all the dirty pages encountered were
993          * backed by a congested BDI. In this case, reclaimers should just
994          * back off and wait for congestion to clear because further reclaim
995          * will encounter the same problem
996          */
997         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
998                 zone_set_flag(zone, ZONE_CONGESTED);
999
1000         free_hot_cold_page_list(&free_pages, 1);
1001
1002         list_splice(&ret_pages, page_list);
1003         count_vm_events(PGACTIVATE, pgactivate);
1004         *ret_nr_dirty += nr_dirty;
1005         *ret_nr_writeback += nr_writeback;
1006         return nr_reclaimed;
1007 }
1008
1009 /*
1010  * Attempt to remove the specified page from its LRU.  Only take this page
1011  * if it is of the appropriate PageActive status.  Pages which are being
1012  * freed elsewhere are also ignored.
1013  *
1014  * page:        page to consider
1015  * mode:        one of the LRU isolation modes defined above
1016  *
1017  * returns 0 on success, -ve errno on failure.
1018  */
1019 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1020 {
1021         bool all_lru_mode;
1022         int ret = -EINVAL;
1023
1024         /* Only take pages on the LRU. */
1025         if (!PageLRU(page))
1026                 return ret;
1027
1028         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1029                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1030
1031         /*
1032          * When checking the active state, we need to be sure we are
1033          * dealing with comparible boolean values.  Take the logical not
1034          * of each.
1035          */
1036         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1037                 return ret;
1038
1039         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1040                 return ret;
1041
1042         /*
1043          * When this function is being called for lumpy reclaim, we
1044          * initially look into all LRU pages, active, inactive and
1045          * unevictable; only give shrink_page_list evictable pages.
1046          */
1047         if (PageUnevictable(page))
1048                 return ret;
1049
1050         ret = -EBUSY;
1051
1052         /*
1053          * To minimise LRU disruption, the caller can indicate that it only
1054          * wants to isolate pages it will be able to operate on without
1055          * blocking - clean pages for the most part.
1056          *
1057          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1058          * is used by reclaim when it is cannot write to backing storage
1059          *
1060          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1061          * that it is possible to migrate without blocking
1062          */
1063         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1064                 /* All the caller can do on PageWriteback is block */
1065                 if (PageWriteback(page))
1066                         return ret;
1067
1068                 if (PageDirty(page)) {
1069                         struct address_space *mapping;
1070                         bool migrate_dirty;
1071
1072                         /* ISOLATE_CLEAN means only clean pages */
1073                         if (mode & ISOLATE_CLEAN)
1074                                 return ret;
1075
1076                         /*
1077                          * Only pages without mappings or that have a
1078                          * ->migratepage callback are possible to migrate
1079                          * without blocking. However, we can be racing with
1080                          * truncation so it's necessary to lock the page
1081                          * to stabilise the mapping as truncation holds
1082                          * the page lock until after the page is removed
1083                          * from the page cache.
1084                          */
1085                         if (!trylock_page(page))
1086                                 return ret;
1087
1088                         mapping = page_mapping(page);
1089                         migrate_dirty = mapping && mapping->a_ops->migratepage;
1090                         unlock_page(page);
1091                         if (!migrate_dirty)
1092                                 return ret;
1093                 }
1094         }
1095
1096         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1097                 return ret;
1098
1099         if (likely(get_page_unless_zero(page))) {
1100                 /*
1101                  * Be careful not to clear PageLRU until after we're
1102                  * sure the page is not being freed elsewhere -- the
1103                  * page release code relies on it.
1104                  */
1105                 ClearPageLRU(page);
1106                 ret = 0;
1107         }
1108
1109         return ret;
1110 }
1111
1112 /*
1113  * zone->lru_lock is heavily contended.  Some of the functions that
1114  * shrink the lists perform better by taking out a batch of pages
1115  * and working on them outside the LRU lock.
1116  *
1117  * For pagecache intensive workloads, this function is the hottest
1118  * spot in the kernel (apart from copy_*_user functions).
1119  *
1120  * Appropriate locks must be held before calling this function.
1121  *
1122  * @nr_to_scan: The number of pages to look through on the list.
1123  * @src:        The LRU list to pull pages off.
1124  * @dst:        The temp list to put pages on to.
1125  * @scanned:    The number of pages that were scanned.
1126  * @order:      The caller's attempted allocation order
1127  * @mode:       One of the LRU isolation modes
1128  * @file:       True [1] if isolating file [!anon] pages
1129  *
1130  * returns how many pages were moved onto *@dst.
1131  */
1132 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1133                 struct list_head *src, struct list_head *dst,
1134                 unsigned long *scanned, int order, isolate_mode_t mode,
1135                 int file)
1136 {
1137         unsigned long nr_taken = 0;
1138         unsigned long nr_lumpy_taken = 0;
1139         unsigned long nr_lumpy_dirty = 0;
1140         unsigned long nr_lumpy_failed = 0;
1141         unsigned long scan;
1142
1143         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1144                 struct page *page;
1145                 unsigned long pfn;
1146                 unsigned long end_pfn;
1147                 unsigned long page_pfn;
1148                 int zone_id;
1149
1150                 page = lru_to_page(src);
1151                 prefetchw_prev_lru_page(page, src, flags);
1152
1153                 VM_BUG_ON(!PageLRU(page));
1154
1155                 switch (__isolate_lru_page(page, mode, file)) {
1156                 case 0:
1157                         list_move(&page->lru, dst);
1158                         mem_cgroup_del_lru(page);
1159                         nr_taken += hpage_nr_pages(page);
1160                         break;
1161
1162                 case -EBUSY:
1163                         /* else it is being freed elsewhere */
1164                         list_move(&page->lru, src);
1165                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1166                         continue;
1167
1168                 default:
1169                         BUG();
1170                 }
1171
1172                 if (!order)
1173                         continue;
1174
1175                 /*
1176                  * Attempt to take all pages in the order aligned region
1177                  * surrounding the tag page.  Only take those pages of
1178                  * the same active state as that tag page.  We may safely
1179                  * round the target page pfn down to the requested order
1180                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1181                  * where that page is in a different zone we will detect
1182                  * it from its zone id and abort this block scan.
1183                  */
1184                 zone_id = page_zone_id(page);
1185                 page_pfn = page_to_pfn(page);
1186                 pfn = page_pfn & ~((1 << order) - 1);
1187                 end_pfn = pfn + (1 << order);
1188                 for (; pfn < end_pfn; pfn++) {
1189                         struct page *cursor_page;
1190
1191                         /* The target page is in the block, ignore it. */
1192                         if (unlikely(pfn == page_pfn))
1193                                 continue;
1194
1195                         /* Avoid holes within the zone. */
1196                         if (unlikely(!pfn_valid_within(pfn)))
1197                                 break;
1198
1199                         cursor_page = pfn_to_page(pfn);
1200
1201                         /* Check that we have not crossed a zone boundary. */
1202                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1203                                 break;
1204
1205                         /*
1206                          * If we don't have enough swap space, reclaiming of
1207                          * anon page which don't already have a swap slot is
1208                          * pointless.
1209                          */
1210                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1211                             !PageSwapCache(cursor_page))
1212                                 break;
1213
1214                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1215                                 list_move(&cursor_page->lru, dst);
1216                                 mem_cgroup_del_lru(cursor_page);
1217                                 nr_taken += hpage_nr_pages(page);
1218                                 nr_lumpy_taken++;
1219                                 if (PageDirty(cursor_page))
1220                                         nr_lumpy_dirty++;
1221                                 scan++;
1222                         } else {
1223                                 /*
1224                                  * Check if the page is freed already.
1225                                  *
1226                                  * We can't use page_count() as that
1227                                  * requires compound_head and we don't
1228                                  * have a pin on the page here. If a
1229                                  * page is tail, we may or may not
1230                                  * have isolated the head, so assume
1231                                  * it's not free, it'd be tricky to
1232                                  * track the head status without a
1233                                  * page pin.
1234                                  */
1235                                 if (!PageTail(cursor_page) &&
1236                                     !atomic_read(&cursor_page->_count))
1237                                         continue;
1238                                 break;
1239                         }
1240                 }
1241
1242                 /* If we break out of the loop above, lumpy reclaim failed */
1243                 if (pfn < end_pfn)
1244                         nr_lumpy_failed++;
1245         }
1246
1247         *scanned = scan;
1248
1249         trace_mm_vmscan_lru_isolate(order,
1250                         nr_to_scan, scan,
1251                         nr_taken,
1252                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1253                         mode);
1254         return nr_taken;
1255 }
1256
1257 static unsigned long isolate_pages_global(unsigned long nr,
1258                                         struct list_head *dst,
1259                                         unsigned long *scanned, int order,
1260                                         isolate_mode_t mode,
1261                                         struct zone *z, int active, int file)
1262 {
1263         int lru = LRU_BASE;
1264         if (active)
1265                 lru += LRU_ACTIVE;
1266         if (file)
1267                 lru += LRU_FILE;
1268         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1269                                                                 mode, file);
1270 }
1271
1272 /*
1273  * clear_active_flags() is a helper for shrink_active_list(), clearing
1274  * any active bits from the pages in the list.
1275  */
1276 static unsigned long clear_active_flags(struct list_head *page_list,
1277                                         unsigned int *count)
1278 {
1279         int nr_active = 0;
1280         int lru;
1281         struct page *page;
1282
1283         list_for_each_entry(page, page_list, lru) {
1284                 int numpages = hpage_nr_pages(page);
1285                 lru = page_lru_base_type(page);
1286                 if (PageActive(page)) {
1287                         lru += LRU_ACTIVE;
1288                         ClearPageActive(page);
1289                         nr_active += numpages;
1290                 }
1291                 if (count)
1292                         count[lru] += numpages;
1293         }
1294
1295         return nr_active;
1296 }
1297
1298 /**
1299  * isolate_lru_page - tries to isolate a page from its LRU list
1300  * @page: page to isolate from its LRU list
1301  *
1302  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1303  * vmstat statistic corresponding to whatever LRU list the page was on.
1304  *
1305  * Returns 0 if the page was removed from an LRU list.
1306  * Returns -EBUSY if the page was not on an LRU list.
1307  *
1308  * The returned page will have PageLRU() cleared.  If it was found on
1309  * the active list, it will have PageActive set.  If it was found on
1310  * the unevictable list, it will have the PageUnevictable bit set. That flag
1311  * may need to be cleared by the caller before letting the page go.
1312  *
1313  * The vmstat statistic corresponding to the list on which the page was
1314  * found will be decremented.
1315  *
1316  * Restrictions:
1317  * (1) Must be called with an elevated refcount on the page. This is a
1318  *     fundamentnal difference from isolate_lru_pages (which is called
1319  *     without a stable reference).
1320  * (2) the lru_lock must not be held.
1321  * (3) interrupts must be enabled.
1322  */
1323 int isolate_lru_page(struct page *page)
1324 {
1325         int ret = -EBUSY;
1326
1327         VM_BUG_ON(!page_count(page));
1328
1329         if (PageLRU(page)) {
1330                 struct zone *zone = page_zone(page);
1331
1332                 spin_lock_irq(&zone->lru_lock);
1333                 if (PageLRU(page)) {
1334                         int lru = page_lru(page);
1335                         ret = 0;
1336                         get_page(page);
1337                         ClearPageLRU(page);
1338
1339                         del_page_from_lru_list(zone, page, lru);
1340                 }
1341                 spin_unlock_irq(&zone->lru_lock);
1342         }
1343         return ret;
1344 }
1345
1346 /*
1347  * Are there way too many processes in the direct reclaim path already?
1348  */
1349 static int too_many_isolated(struct zone *zone, int file,
1350                 struct scan_control *sc)
1351 {
1352         unsigned long inactive, isolated;
1353
1354         if (current_is_kswapd())
1355                 return 0;
1356
1357         if (!scanning_global_lru(sc))
1358                 return 0;
1359
1360         if (file) {
1361                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1362                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1363         } else {
1364                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1365                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1366         }
1367
1368         return isolated > inactive;
1369 }
1370
1371 /*
1372  * TODO: Try merging with migrations version of putback_lru_pages
1373  */
1374 static noinline_for_stack void
1375 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1376                                 unsigned long nr_anon, unsigned long nr_file,
1377                                 struct list_head *page_list)
1378 {
1379         struct page *page;
1380         struct pagevec pvec;
1381         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1382
1383         pagevec_init(&pvec, 1);
1384
1385         /*
1386          * Put back any unfreeable pages.
1387          */
1388         spin_lock(&zone->lru_lock);
1389         while (!list_empty(page_list)) {
1390                 int lru;
1391                 page = lru_to_page(page_list);
1392                 VM_BUG_ON(PageLRU(page));
1393                 list_del(&page->lru);
1394                 if (unlikely(!page_evictable(page, NULL))) {
1395                         spin_unlock_irq(&zone->lru_lock);
1396                         putback_lru_page(page);
1397                         spin_lock_irq(&zone->lru_lock);
1398                         continue;
1399                 }
1400                 SetPageLRU(page);
1401                 lru = page_lru(page);
1402                 add_page_to_lru_list(zone, page, lru);
1403                 if (is_active_lru(lru)) {
1404                         int file = is_file_lru(lru);
1405                         int numpages = hpage_nr_pages(page);
1406                         reclaim_stat->recent_rotated[file] += numpages;
1407                 }
1408                 if (!pagevec_add(&pvec, page)) {
1409                         spin_unlock_irq(&zone->lru_lock);
1410                         __pagevec_release(&pvec);
1411                         spin_lock_irq(&zone->lru_lock);
1412                 }
1413         }
1414         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1415         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1416
1417         spin_unlock_irq(&zone->lru_lock);
1418         pagevec_release(&pvec);
1419 }
1420
1421 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1422                                         struct scan_control *sc,
1423                                         unsigned long *nr_anon,
1424                                         unsigned long *nr_file,
1425                                         struct list_head *isolated_list)
1426 {
1427         unsigned long nr_active;
1428         unsigned int count[NR_LRU_LISTS] = { 0, };
1429         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1430
1431         nr_active = clear_active_flags(isolated_list, count);
1432         __count_vm_events(PGDEACTIVATE, nr_active);
1433
1434         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1435                               -count[LRU_ACTIVE_FILE]);
1436         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1437                               -count[LRU_INACTIVE_FILE]);
1438         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1439                               -count[LRU_ACTIVE_ANON]);
1440         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1441                               -count[LRU_INACTIVE_ANON]);
1442
1443         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1444         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1445         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1446         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1447
1448         reclaim_stat->recent_scanned[0] += *nr_anon;
1449         reclaim_stat->recent_scanned[1] += *nr_file;
1450 }
1451
1452 /*
1453  * Returns true if a direct reclaim should wait on pages under writeback.
1454  *
1455  * If we are direct reclaiming for contiguous pages and we do not reclaim
1456  * everything in the list, try again and wait for writeback IO to complete.
1457  * This will stall high-order allocations noticeably. Only do that when really
1458  * need to free the pages under high memory pressure.
1459  */
1460 static inline bool should_reclaim_stall(unsigned long nr_taken,
1461                                         unsigned long nr_freed,
1462                                         int priority,
1463                                         struct scan_control *sc)
1464 {
1465         int lumpy_stall_priority;
1466
1467         /* kswapd should not stall on sync IO */
1468         if (current_is_kswapd())
1469                 return false;
1470
1471         /* Only stall on lumpy reclaim */
1472         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1473                 return false;
1474
1475         /* If we have reclaimed everything on the isolated list, no stall */
1476         if (nr_freed == nr_taken)
1477                 return false;
1478
1479         /*
1480          * For high-order allocations, there are two stall thresholds.
1481          * High-cost allocations stall immediately where as lower
1482          * order allocations such as stacks require the scanning
1483          * priority to be much higher before stalling.
1484          */
1485         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1486                 lumpy_stall_priority = DEF_PRIORITY;
1487         else
1488                 lumpy_stall_priority = DEF_PRIORITY / 3;
1489
1490         return priority <= lumpy_stall_priority;
1491 }
1492
1493 /*
1494  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1495  * of reclaimed pages
1496  */
1497 static noinline_for_stack unsigned long
1498 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1499                         struct scan_control *sc, int priority, int file)
1500 {
1501         LIST_HEAD(page_list);
1502         unsigned long nr_scanned;
1503         unsigned long nr_reclaimed = 0;
1504         unsigned long nr_taken;
1505         unsigned long nr_anon;
1506         unsigned long nr_file;
1507         unsigned long nr_dirty = 0;
1508         unsigned long nr_writeback = 0;
1509         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1510
1511         while (unlikely(too_many_isolated(zone, file, sc))) {
1512                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1513
1514                 /* We are about to die and free our memory. Return now. */
1515                 if (fatal_signal_pending(current))
1516                         return SWAP_CLUSTER_MAX;
1517         }
1518
1519         set_reclaim_mode(priority, sc, false);
1520         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1521                 reclaim_mode |= ISOLATE_ACTIVE;
1522
1523         lru_add_drain();
1524
1525         if (!sc->may_unmap)
1526                 reclaim_mode |= ISOLATE_UNMAPPED;
1527         if (!sc->may_writepage)
1528                 reclaim_mode |= ISOLATE_CLEAN;
1529
1530         spin_lock_irq(&zone->lru_lock);
1531
1532         if (scanning_global_lru(sc)) {
1533                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1534                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1535                 zone->pages_scanned += nr_scanned;
1536                 if (current_is_kswapd())
1537                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1538                                                nr_scanned);
1539                 else
1540                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1541                                                nr_scanned);
1542         } else {
1543                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1544                         &nr_scanned, sc->order, reclaim_mode, zone,
1545                         sc->mem_cgroup, 0, file);
1546                 /*
1547                  * mem_cgroup_isolate_pages() keeps track of
1548                  * scanned pages on its own.
1549                  */
1550         }
1551
1552         if (nr_taken == 0) {
1553                 spin_unlock_irq(&zone->lru_lock);
1554                 return 0;
1555         }
1556
1557         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1558
1559         spin_unlock_irq(&zone->lru_lock);
1560
1561         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1562                                                 &nr_dirty, &nr_writeback);
1563
1564         /* Check if we should syncronously wait for writeback */
1565         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1566                 set_reclaim_mode(priority, sc, true);
1567                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1568                                         priority, &nr_dirty, &nr_writeback);
1569         }
1570
1571         local_irq_disable();
1572         if (current_is_kswapd())
1573                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1574         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1575
1576         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1577
1578         /*
1579          * If reclaim is isolating dirty pages under writeback, it implies
1580          * that the long-lived page allocation rate is exceeding the page
1581          * laundering rate. Either the global limits are not being effective
1582          * at throttling processes due to the page distribution throughout
1583          * zones or there is heavy usage of a slow backing device. The
1584          * only option is to throttle from reclaim context which is not ideal
1585          * as there is no guarantee the dirtying process is throttled in the
1586          * same way balance_dirty_pages() manages.
1587          *
1588          * This scales the number of dirty pages that must be under writeback
1589          * before throttling depending on priority. It is a simple backoff
1590          * function that has the most effect in the range DEF_PRIORITY to
1591          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1592          * in trouble and reclaim is considered to be in trouble.
1593          *
1594          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1595          * DEF_PRIORITY-1  50% must be PageWriteback
1596          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1597          * ...
1598          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1599          *                     isolated page is PageWriteback
1600          */
1601         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1602                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1603
1604         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1605                 zone_idx(zone),
1606                 nr_scanned, nr_reclaimed,
1607                 priority,
1608                 trace_shrink_flags(file, sc->reclaim_mode));
1609         return nr_reclaimed;
1610 }
1611
1612 /*
1613  * This moves pages from the active list to the inactive list.
1614  *
1615  * We move them the other way if the page is referenced by one or more
1616  * processes, from rmap.
1617  *
1618  * If the pages are mostly unmapped, the processing is fast and it is
1619  * appropriate to hold zone->lru_lock across the whole operation.  But if
1620  * the pages are mapped, the processing is slow (page_referenced()) so we
1621  * should drop zone->lru_lock around each page.  It's impossible to balance
1622  * this, so instead we remove the pages from the LRU while processing them.
1623  * It is safe to rely on PG_active against the non-LRU pages in here because
1624  * nobody will play with that bit on a non-LRU page.
1625  *
1626  * The downside is that we have to touch page->_count against each page.
1627  * But we had to alter page->flags anyway.
1628  */
1629
1630 static void move_active_pages_to_lru(struct zone *zone,
1631                                      struct list_head *list,
1632                                      enum lru_list lru)
1633 {
1634         unsigned long pgmoved = 0;
1635         struct pagevec pvec;
1636         struct page *page;
1637
1638         pagevec_init(&pvec, 1);
1639
1640         while (!list_empty(list)) {
1641                 page = lru_to_page(list);
1642
1643                 VM_BUG_ON(PageLRU(page));
1644                 SetPageLRU(page);
1645
1646                 list_move(&page->lru, &zone->lru[lru].list);
1647                 mem_cgroup_add_lru_list(page, lru);
1648                 pgmoved += hpage_nr_pages(page);
1649
1650                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1651                         spin_unlock_irq(&zone->lru_lock);
1652                         if (buffer_heads_over_limit)
1653                                 pagevec_strip(&pvec);
1654                         __pagevec_release(&pvec);
1655                         spin_lock_irq(&zone->lru_lock);
1656                 }
1657         }
1658         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1659         if (!is_active_lru(lru))
1660                 __count_vm_events(PGDEACTIVATE, pgmoved);
1661 }
1662
1663 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1664                         struct scan_control *sc, int priority, int file)
1665 {
1666         unsigned long nr_taken;
1667         unsigned long pgscanned;
1668         unsigned long vm_flags;
1669         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1670         LIST_HEAD(l_active);
1671         LIST_HEAD(l_inactive);
1672         struct page *page;
1673         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1674         unsigned long nr_rotated = 0;
1675         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1676
1677         lru_add_drain();
1678
1679         if (!sc->may_unmap)
1680                 reclaim_mode |= ISOLATE_UNMAPPED;
1681         if (!sc->may_writepage)
1682                 reclaim_mode |= ISOLATE_CLEAN;
1683
1684         spin_lock_irq(&zone->lru_lock);
1685         if (scanning_global_lru(sc)) {
1686                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1687                                                 &pgscanned, sc->order,
1688                                                 reclaim_mode, zone,
1689                                                 1, file);
1690                 zone->pages_scanned += pgscanned;
1691         } else {
1692                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1693                                                 &pgscanned, sc->order,
1694                                                 reclaim_mode, zone,
1695                                                 sc->mem_cgroup, 1, file);
1696                 /*
1697                  * mem_cgroup_isolate_pages() keeps track of
1698                  * scanned pages on its own.
1699                  */
1700         }
1701
1702         reclaim_stat->recent_scanned[file] += nr_taken;
1703
1704         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1705         if (file)
1706                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1707         else
1708                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1709         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1710         spin_unlock_irq(&zone->lru_lock);
1711
1712         while (!list_empty(&l_hold)) {
1713                 cond_resched();
1714                 page = lru_to_page(&l_hold);
1715                 list_del(&page->lru);
1716
1717                 if (unlikely(!page_evictable(page, NULL))) {
1718                         putback_lru_page(page);
1719                         continue;
1720                 }
1721
1722                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1723                         nr_rotated += hpage_nr_pages(page);
1724                         /*
1725                          * Identify referenced, file-backed active pages and
1726                          * give them one more trip around the active list. So
1727                          * that executable code get better chances to stay in
1728                          * memory under moderate memory pressure.  Anon pages
1729                          * are not likely to be evicted by use-once streaming
1730                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1731                          * so we ignore them here.
1732                          */
1733                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1734                                 list_add(&page->lru, &l_active);
1735                                 continue;
1736                         }
1737                 }
1738
1739                 ClearPageActive(page);  /* we are de-activating */
1740                 list_add(&page->lru, &l_inactive);
1741         }
1742
1743         /*
1744          * Move pages back to the lru list.
1745          */
1746         spin_lock_irq(&zone->lru_lock);
1747         /*
1748          * Count referenced pages from currently used mappings as rotated,
1749          * even though only some of them are actually re-activated.  This
1750          * helps balance scan pressure between file and anonymous pages in
1751          * get_scan_ratio.
1752          */
1753         reclaim_stat->recent_rotated[file] += nr_rotated;
1754
1755         move_active_pages_to_lru(zone, &l_active,
1756                                                 LRU_ACTIVE + file * LRU_FILE);
1757         move_active_pages_to_lru(zone, &l_inactive,
1758                                                 LRU_BASE   + file * LRU_FILE);
1759         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1760         spin_unlock_irq(&zone->lru_lock);
1761 }
1762
1763 #ifdef CONFIG_SWAP
1764 static int inactive_anon_is_low_global(struct zone *zone)
1765 {
1766         unsigned long active, inactive;
1767
1768         active = zone_page_state(zone, NR_ACTIVE_ANON);
1769         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1770
1771         if (inactive * zone->inactive_ratio < active)
1772                 return 1;
1773
1774         return 0;
1775 }
1776
1777 /**
1778  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1779  * @zone: zone to check
1780  * @sc:   scan control of this context
1781  *
1782  * Returns true if the zone does not have enough inactive anon pages,
1783  * meaning some active anon pages need to be deactivated.
1784  */
1785 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1786 {
1787         int low;
1788
1789         /*
1790          * If we don't have swap space, anonymous page deactivation
1791          * is pointless.
1792          */
1793         if (!total_swap_pages)
1794                 return 0;
1795
1796         if (scanning_global_lru(sc))
1797                 low = inactive_anon_is_low_global(zone);
1798         else
1799                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1800         return low;
1801 }
1802 #else
1803 static inline int inactive_anon_is_low(struct zone *zone,
1804                                         struct scan_control *sc)
1805 {
1806         return 0;
1807 }
1808 #endif
1809
1810 static int inactive_file_is_low_global(struct zone *zone)
1811 {
1812         unsigned long active, inactive;
1813
1814         active = zone_page_state(zone, NR_ACTIVE_FILE);
1815         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1816
1817         return (active > inactive);
1818 }
1819
1820 /**
1821  * inactive_file_is_low - check if file pages need to be deactivated
1822  * @zone: zone to check
1823  * @sc:   scan control of this context
1824  *
1825  * When the system is doing streaming IO, memory pressure here
1826  * ensures that active file pages get deactivated, until more
1827  * than half of the file pages are on the inactive list.
1828  *
1829  * Once we get to that situation, protect the system's working
1830  * set from being evicted by disabling active file page aging.
1831  *
1832  * This uses a different ratio than the anonymous pages, because
1833  * the page cache uses a use-once replacement algorithm.
1834  */
1835 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1836 {
1837         int low;
1838
1839         if (scanning_global_lru(sc))
1840                 low = inactive_file_is_low_global(zone);
1841         else
1842                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1843         return low;
1844 }
1845
1846 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1847                                 int file)
1848 {
1849         if (file)
1850                 return inactive_file_is_low(zone, sc);
1851         else
1852                 return inactive_anon_is_low(zone, sc);
1853 }
1854
1855 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1856         struct zone *zone, struct scan_control *sc, int priority)
1857 {
1858         int file = is_file_lru(lru);
1859
1860         if (is_active_lru(lru)) {
1861                 if (inactive_list_is_low(zone, sc, file))
1862                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1863                 return 0;
1864         }
1865
1866         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1867 }
1868
1869 static int vmscan_swappiness(struct scan_control *sc)
1870 {
1871         if (scanning_global_lru(sc))
1872                 return vm_swappiness;
1873         return mem_cgroup_swappiness(sc->mem_cgroup);
1874 }
1875
1876 /*
1877  * Determine how aggressively the anon and file LRU lists should be
1878  * scanned.  The relative value of each set of LRU lists is determined
1879  * by looking at the fraction of the pages scanned we did rotate back
1880  * onto the active list instead of evict.
1881  *
1882  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1883  */
1884 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1885                                         unsigned long *nr, int priority)
1886 {
1887         unsigned long anon, file, free;
1888         unsigned long anon_prio, file_prio;
1889         unsigned long ap, fp;
1890         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1891         u64 fraction[2], denominator;
1892         enum lru_list l;
1893         int noswap = 0;
1894         bool force_scan = false;
1895
1896         /*
1897          * If the zone or memcg is small, nr[l] can be 0.  This
1898          * results in no scanning on this priority and a potential
1899          * priority drop.  Global direct reclaim can go to the next
1900          * zone and tends to have no problems. Global kswapd is for
1901          * zone balancing and it needs to scan a minimum amount. When
1902          * reclaiming for a memcg, a priority drop can cause high
1903          * latencies, so it's better to scan a minimum amount there as
1904          * well.
1905          */
1906         if (scanning_global_lru(sc) && current_is_kswapd() &&
1907             zone->all_unreclaimable)
1908                 force_scan = true;
1909         if (!scanning_global_lru(sc))
1910                 force_scan = true;
1911
1912         /* If we have no swap space, do not bother scanning anon pages. */
1913         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1914                 noswap = 1;
1915                 fraction[0] = 0;
1916                 fraction[1] = 1;
1917                 denominator = 1;
1918                 goto out;
1919         }
1920
1921         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1922                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1923         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1924                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1925
1926         if (scanning_global_lru(sc)) {
1927                 free  = zone_page_state(zone, NR_FREE_PAGES);
1928                 /* If we have very few page cache pages,
1929                    force-scan anon pages. */
1930                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1931                         fraction[0] = 1;
1932                         fraction[1] = 0;
1933                         denominator = 1;
1934                         goto out;
1935                 }
1936         }
1937
1938         /*
1939          * With swappiness at 100, anonymous and file have the same priority.
1940          * This scanning priority is essentially the inverse of IO cost.
1941          */
1942         anon_prio = vmscan_swappiness(sc);
1943         file_prio = 200 - vmscan_swappiness(sc);
1944
1945         /*
1946          * OK, so we have swap space and a fair amount of page cache
1947          * pages.  We use the recently rotated / recently scanned
1948          * ratios to determine how valuable each cache is.
1949          *
1950          * Because workloads change over time (and to avoid overflow)
1951          * we keep these statistics as a floating average, which ends
1952          * up weighing recent references more than old ones.
1953          *
1954          * anon in [0], file in [1]
1955          */
1956         spin_lock_irq(&zone->lru_lock);
1957         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1958                 reclaim_stat->recent_scanned[0] /= 2;
1959                 reclaim_stat->recent_rotated[0] /= 2;
1960         }
1961
1962         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1963                 reclaim_stat->recent_scanned[1] /= 2;
1964                 reclaim_stat->recent_rotated[1] /= 2;
1965         }
1966
1967         /*
1968          * The amount of pressure on anon vs file pages is inversely
1969          * proportional to the fraction of recently scanned pages on
1970          * each list that were recently referenced and in active use.
1971          */
1972         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1973         ap /= reclaim_stat->recent_rotated[0] + 1;
1974
1975         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1976         fp /= reclaim_stat->recent_rotated[1] + 1;
1977         spin_unlock_irq(&zone->lru_lock);
1978
1979         fraction[0] = ap;
1980         fraction[1] = fp;
1981         denominator = ap + fp + 1;
1982 out:
1983         for_each_evictable_lru(l) {
1984                 int file = is_file_lru(l);
1985                 unsigned long scan;
1986
1987                 scan = zone_nr_lru_pages(zone, sc, l);
1988                 if (priority || noswap || !vmscan_swappiness(sc)) {
1989                         scan >>= priority;
1990                         if (!scan && force_scan)
1991                                 scan = SWAP_CLUSTER_MAX;
1992                         scan = div64_u64(scan * fraction[file], denominator);
1993                 }
1994                 nr[l] = scan;
1995         }
1996 }
1997
1998 /*
1999  * Reclaim/compaction depends on a number of pages being freed. To avoid
2000  * disruption to the system, a small number of order-0 pages continue to be
2001  * rotated and reclaimed in the normal fashion. However, by the time we get
2002  * back to the allocator and call try_to_compact_zone(), we ensure that
2003  * there are enough free pages for it to be likely successful
2004  */
2005 static inline bool should_continue_reclaim(struct zone *zone,
2006                                         unsigned long nr_reclaimed,
2007                                         unsigned long nr_scanned,
2008                                         struct scan_control *sc)
2009 {
2010         unsigned long pages_for_compaction;
2011         unsigned long inactive_lru_pages;
2012
2013         /* If not in reclaim/compaction mode, stop */
2014         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2015                 return false;
2016
2017         /* Consider stopping depending on scan and reclaim activity */
2018         if (sc->gfp_mask & __GFP_REPEAT) {
2019                 /*
2020                  * For __GFP_REPEAT allocations, stop reclaiming if the
2021                  * full LRU list has been scanned and we are still failing
2022                  * to reclaim pages. This full LRU scan is potentially
2023                  * expensive but a __GFP_REPEAT caller really wants to succeed
2024                  */
2025                 if (!nr_reclaimed && !nr_scanned)
2026                         return false;
2027         } else {
2028                 /*
2029                  * For non-__GFP_REPEAT allocations which can presumably
2030                  * fail without consequence, stop if we failed to reclaim
2031                  * any pages from the last SWAP_CLUSTER_MAX number of
2032                  * pages that were scanned. This will return to the
2033                  * caller faster at the risk reclaim/compaction and
2034                  * the resulting allocation attempt fails
2035                  */
2036                 if (!nr_reclaimed)
2037                         return false;
2038         }
2039
2040         /*
2041          * If we have not reclaimed enough pages for compaction and the
2042          * inactive lists are large enough, continue reclaiming
2043          */
2044         pages_for_compaction = (2UL << sc->order);
2045         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2046         if (nr_swap_pages > 0)
2047                 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2048         if (sc->nr_reclaimed < pages_for_compaction &&
2049                         inactive_lru_pages > pages_for_compaction)
2050                 return true;
2051
2052         /* If compaction would go ahead or the allocation would succeed, stop */
2053         switch (compaction_suitable(zone, sc->order)) {
2054         case COMPACT_PARTIAL:
2055         case COMPACT_CONTINUE:
2056                 return false;
2057         default:
2058                 return true;
2059         }
2060 }
2061
2062 /*
2063  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2064  */
2065 static void shrink_zone(int priority, struct zone *zone,
2066                                 struct scan_control *sc)
2067 {
2068         unsigned long nr[NR_LRU_LISTS];
2069         unsigned long nr_to_scan;
2070         enum lru_list l;
2071         unsigned long nr_reclaimed, nr_scanned;
2072         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2073         struct blk_plug plug;
2074
2075 restart:
2076         nr_reclaimed = 0;
2077         nr_scanned = sc->nr_scanned;
2078         get_scan_count(zone, sc, nr, priority);
2079
2080         blk_start_plug(&plug);
2081         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2082                                         nr[LRU_INACTIVE_FILE]) {
2083                 for_each_evictable_lru(l) {
2084                         if (nr[l]) {
2085                                 nr_to_scan = min_t(unsigned long,
2086                                                    nr[l], SWAP_CLUSTER_MAX);
2087                                 nr[l] -= nr_to_scan;
2088
2089                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2090                                                             zone, sc, priority);
2091                         }
2092                 }
2093                 /*
2094                  * On large memory systems, scan >> priority can become
2095                  * really large. This is fine for the starting priority;
2096                  * we want to put equal scanning pressure on each zone.
2097                  * However, if the VM has a harder time of freeing pages,
2098                  * with multiple processes reclaiming pages, the total
2099                  * freeing target can get unreasonably large.
2100                  */
2101                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2102                         break;
2103         }
2104         blk_finish_plug(&plug);
2105         sc->nr_reclaimed += nr_reclaimed;
2106
2107         /*
2108          * Even if we did not try to evict anon pages at all, we want to
2109          * rebalance the anon lru active/inactive ratio.
2110          */
2111         if (inactive_anon_is_low(zone, sc))
2112                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2113
2114         /* reclaim/compaction might need reclaim to continue */
2115         if (should_continue_reclaim(zone, nr_reclaimed,
2116                                         sc->nr_scanned - nr_scanned, sc))
2117                 goto restart;
2118
2119         throttle_vm_writeout(sc->gfp_mask);
2120 }
2121
2122 /* Returns true if compaction should go ahead for a high-order request */
2123 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2124 {
2125         unsigned long balance_gap, watermark;
2126         bool watermark_ok;
2127
2128         /* Do not consider compaction for orders reclaim is meant to satisfy */
2129         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2130                 return false;
2131
2132         /*
2133          * Compaction takes time to run and there are potentially other
2134          * callers using the pages just freed. Continue reclaiming until
2135          * there is a buffer of free pages available to give compaction
2136          * a reasonable chance of completing and allocating the page
2137          */
2138         balance_gap = min(low_wmark_pages(zone),
2139                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2140                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2141         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2142         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2143
2144         /*
2145          * If compaction is deferred, reclaim up to a point where
2146          * compaction will have a chance of success when re-enabled
2147          */
2148         if (compaction_deferred(zone))
2149                 return watermark_ok;
2150
2151         /* If compaction is not ready to start, keep reclaiming */
2152         if (!compaction_suitable(zone, sc->order))
2153                 return false;
2154
2155         return watermark_ok;
2156 }
2157
2158 /*
2159  * This is the direct reclaim path, for page-allocating processes.  We only
2160  * try to reclaim pages from zones which will satisfy the caller's allocation
2161  * request.
2162  *
2163  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2164  * Because:
2165  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2166  *    allocation or
2167  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2168  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2169  *    zone defense algorithm.
2170  *
2171  * If a zone is deemed to be full of pinned pages then just give it a light
2172  * scan then give up on it.
2173  *
2174  * This function returns true if a zone is being reclaimed for a costly
2175  * high-order allocation and compaction is ready to begin. This indicates to
2176  * the caller that it should consider retrying the allocation instead of
2177  * further reclaim.
2178  */
2179 static bool shrink_zones(int priority, struct zonelist *zonelist,
2180                                         struct scan_control *sc)
2181 {
2182         struct zoneref *z;
2183         struct zone *zone;
2184         unsigned long nr_soft_reclaimed;
2185         unsigned long nr_soft_scanned;
2186         bool aborted_reclaim = false;
2187
2188         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2189                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2190                 if (!populated_zone(zone))
2191                         continue;
2192                 /*
2193                  * Take care memory controller reclaiming has small influence
2194                  * to global LRU.
2195                  */
2196                 if (scanning_global_lru(sc)) {
2197                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2198                                 continue;
2199                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2200                                 continue;       /* Let kswapd poll it */
2201                         if (COMPACTION_BUILD) {
2202                                 /*
2203                                  * If we already have plenty of memory free for
2204                                  * compaction in this zone, don't free any more.
2205                                  * Even though compaction is invoked for any
2206                                  * non-zero order, only frequent costly order
2207                                  * reclamation is disruptive enough to become a
2208                                  * noticable problem, like transparent huge page
2209                                  * allocations.
2210                                  */
2211                                 if (compaction_ready(zone, sc)) {
2212                                         aborted_reclaim = true;
2213                                         continue;
2214                                 }
2215                         }
2216                         /*
2217                          * This steals pages from memory cgroups over softlimit
2218                          * and returns the number of reclaimed pages and
2219                          * scanned pages. This works for global memory pressure
2220                          * and balancing, not for a memcg's limit.
2221                          */
2222                         nr_soft_scanned = 0;
2223                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2224                                                 sc->order, sc->gfp_mask,
2225                                                 &nr_soft_scanned);
2226                         sc->nr_reclaimed += nr_soft_reclaimed;
2227                         sc->nr_scanned += nr_soft_scanned;
2228                         /* need some check for avoid more shrink_zone() */
2229                 }
2230
2231                 shrink_zone(priority, zone, sc);
2232         }
2233
2234         return aborted_reclaim;
2235 }
2236
2237 static bool zone_reclaimable(struct zone *zone)
2238 {
2239         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2240 }
2241
2242 /* All zones in zonelist are unreclaimable? */
2243 static bool all_unreclaimable(struct zonelist *zonelist,
2244                 struct scan_control *sc)
2245 {
2246         struct zoneref *z;
2247         struct zone *zone;
2248
2249         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2250                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2251                 if (!populated_zone(zone))
2252                         continue;
2253                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2254                         continue;
2255                 if (!zone->all_unreclaimable)
2256                         return false;
2257         }
2258
2259         return true;
2260 }
2261
2262 /*
2263  * This is the main entry point to direct page reclaim.
2264  *
2265  * If a full scan of the inactive list fails to free enough memory then we
2266  * are "out of memory" and something needs to be killed.
2267  *
2268  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2269  * high - the zone may be full of dirty or under-writeback pages, which this
2270  * caller can't do much about.  We kick the writeback threads and take explicit
2271  * naps in the hope that some of these pages can be written.  But if the
2272  * allocating task holds filesystem locks which prevent writeout this might not
2273  * work, and the allocation attempt will fail.
2274  *
2275  * returns:     0, if no pages reclaimed
2276  *              else, the number of pages reclaimed
2277  */
2278 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2279                                         struct scan_control *sc,
2280                                         struct shrink_control *shrink)
2281 {
2282         int priority;
2283         unsigned long total_scanned = 0;
2284         struct reclaim_state *reclaim_state = current->reclaim_state;
2285         struct zoneref *z;
2286         struct zone *zone;
2287         unsigned long writeback_threshold;
2288         bool aborted_reclaim;
2289
2290         delayacct_freepages_start();
2291
2292         if (scanning_global_lru(sc))
2293                 count_vm_event(ALLOCSTALL);
2294
2295         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2296                 sc->nr_scanned = 0;
2297                 if (!priority)
2298                         disable_swap_token(sc->mem_cgroup);
2299                 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2300
2301                 /*
2302                  * Don't shrink slabs when reclaiming memory from
2303                  * over limit cgroups
2304                  */
2305                 if (scanning_global_lru(sc)) {
2306                         unsigned long lru_pages = 0;
2307                         for_each_zone_zonelist(zone, z, zonelist,
2308                                         gfp_zone(sc->gfp_mask)) {
2309                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2310                                         continue;
2311
2312                                 lru_pages += zone_reclaimable_pages(zone);
2313                         }
2314
2315                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2316                         if (reclaim_state) {
2317                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2318                                 reclaim_state->reclaimed_slab = 0;
2319                         }
2320                 }
2321                 total_scanned += sc->nr_scanned;
2322                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2323                         goto out;
2324
2325                 /*
2326                  * Try to write back as many pages as we just scanned.  This
2327                  * tends to cause slow streaming writers to write data to the
2328                  * disk smoothly, at the dirtying rate, which is nice.   But
2329                  * that's undesirable in laptop mode, where we *want* lumpy
2330                  * writeout.  So in laptop mode, write out the whole world.
2331                  */
2332                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2333                 if (total_scanned > writeback_threshold) {
2334                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2335                                                 WB_REASON_TRY_TO_FREE_PAGES);
2336                         sc->may_writepage = 1;
2337                 }
2338
2339                 /* Take a nap, wait for some writeback to complete */
2340                 if (!sc->hibernation_mode && sc->nr_scanned &&
2341                     priority < DEF_PRIORITY - 2) {
2342                         struct zone *preferred_zone;
2343
2344                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2345                                                 &cpuset_current_mems_allowed,
2346                                                 &preferred_zone);
2347                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2348                 }
2349         }
2350
2351 out:
2352         delayacct_freepages_end();
2353
2354         if (sc->nr_reclaimed)
2355                 return sc->nr_reclaimed;
2356
2357         /*
2358          * As hibernation is going on, kswapd is freezed so that it can't mark
2359          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2360          * check.
2361          */
2362         if (oom_killer_disabled)
2363                 return 0;
2364
2365         /* Aborted reclaim to try compaction? don't OOM, then */
2366         if (aborted_reclaim)
2367                 return 1;
2368
2369         /* top priority shrink_zones still had more to do? don't OOM, then */
2370         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2371                 return 1;
2372
2373         return 0;
2374 }
2375
2376 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2377                                 gfp_t gfp_mask, nodemask_t *nodemask)
2378 {
2379         unsigned long nr_reclaimed;
2380         struct scan_control sc = {
2381                 .gfp_mask = gfp_mask,
2382                 .may_writepage = !laptop_mode,
2383                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2384                 .may_unmap = 1,
2385                 .may_swap = 1,
2386                 .order = order,
2387                 .mem_cgroup = NULL,
2388                 .nodemask = nodemask,
2389         };
2390         struct shrink_control shrink = {
2391                 .gfp_mask = sc.gfp_mask,
2392         };
2393
2394         trace_mm_vmscan_direct_reclaim_begin(order,
2395                                 sc.may_writepage,
2396                                 gfp_mask);
2397
2398         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2399
2400         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2401
2402         return nr_reclaimed;
2403 }
2404
2405 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2406
2407 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2408                                                 gfp_t gfp_mask, bool noswap,
2409                                                 struct zone *zone,
2410                                                 unsigned long *nr_scanned)
2411 {
2412         struct scan_control sc = {
2413                 .nr_scanned = 0,
2414                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2415                 .may_writepage = !laptop_mode,
2416                 .may_unmap = 1,
2417                 .may_swap = !noswap,
2418                 .order = 0,
2419                 .mem_cgroup = mem,
2420         };
2421
2422         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2423                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2424
2425         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2426                                                       sc.may_writepage,
2427                                                       sc.gfp_mask);
2428
2429         /*
2430          * NOTE: Although we can get the priority field, using it
2431          * here is not a good idea, since it limits the pages we can scan.
2432          * if we don't reclaim here, the shrink_zone from balance_pgdat
2433          * will pick up pages from other mem cgroup's as well. We hack
2434          * the priority and make it zero.
2435          */
2436         shrink_zone(0, zone, &sc);
2437
2438         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2439
2440         *nr_scanned = sc.nr_scanned;
2441         return sc.nr_reclaimed;
2442 }
2443
2444 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2445                                            gfp_t gfp_mask,
2446                                            bool noswap)
2447 {
2448         struct zonelist *zonelist;
2449         unsigned long nr_reclaimed;
2450         int nid;
2451         struct scan_control sc = {
2452                 .may_writepage = !laptop_mode,
2453                 .may_unmap = 1,
2454                 .may_swap = !noswap,
2455                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2456                 .order = 0,
2457                 .mem_cgroup = mem_cont,
2458                 .nodemask = NULL, /* we don't care the placement */
2459                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2460                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2461         };
2462         struct shrink_control shrink = {
2463                 .gfp_mask = sc.gfp_mask,
2464         };
2465
2466         /*
2467          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2468          * take care of from where we get pages. So the node where we start the
2469          * scan does not need to be the current node.
2470          */
2471         nid = mem_cgroup_select_victim_node(mem_cont);
2472
2473         zonelist = NODE_DATA(nid)->node_zonelists;
2474
2475         trace_mm_vmscan_memcg_reclaim_begin(0,
2476                                             sc.may_writepage,
2477                                             sc.gfp_mask);
2478
2479         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2480
2481         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2482
2483         return nr_reclaimed;
2484 }
2485 #endif
2486
2487 /*
2488  * pgdat_balanced is used when checking if a node is balanced for high-order
2489  * allocations. Only zones that meet watermarks and are in a zone allowed
2490  * by the callers classzone_idx are added to balanced_pages. The total of
2491  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2492  * for the node to be considered balanced. Forcing all zones to be balanced
2493  * for high orders can cause excessive reclaim when there are imbalanced zones.
2494  * The choice of 25% is due to
2495  *   o a 16M DMA zone that is balanced will not balance a zone on any
2496  *     reasonable sized machine
2497  *   o On all other machines, the top zone must be at least a reasonable
2498  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2499  *     would need to be at least 256M for it to be balance a whole node.
2500  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2501  *     to balance a node on its own. These seemed like reasonable ratios.
2502  */
2503 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2504                                                 int classzone_idx)
2505 {
2506         unsigned long present_pages = 0;
2507         int i;
2508
2509         for (i = 0; i <= classzone_idx; i++)
2510                 present_pages += pgdat->node_zones[i].present_pages;
2511
2512         /* A special case here: if zone has no page, we think it's balanced */
2513         return balanced_pages >= (present_pages >> 2);
2514 }
2515
2516 /* is kswapd sleeping prematurely? */
2517 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2518                                         int classzone_idx)
2519 {
2520         int i;
2521         unsigned long balanced = 0;
2522         bool all_zones_ok = true;
2523
2524         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2525         if (remaining)
2526                 return true;
2527
2528         /* Check the watermark levels */
2529         for (i = 0; i <= classzone_idx; i++) {
2530                 struct zone *zone = pgdat->node_zones + i;
2531
2532                 if (!populated_zone(zone))
2533                         continue;
2534
2535                 /*
2536                  * balance_pgdat() skips over all_unreclaimable after
2537                  * DEF_PRIORITY. Effectively, it considers them balanced so
2538                  * they must be considered balanced here as well if kswapd
2539                  * is to sleep
2540                  */
2541                 if (zone->all_unreclaimable) {
2542                         balanced += zone->present_pages;
2543                         continue;
2544                 }
2545
2546                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2547                                                         i, 0))
2548                         all_zones_ok = false;
2549                 else
2550                         balanced += zone->present_pages;
2551         }
2552
2553         /*
2554          * For high-order requests, the balanced zones must contain at least
2555          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2556          * must be balanced
2557          */
2558         if (order)
2559                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2560         else
2561                 return !all_zones_ok;
2562 }
2563
2564 /*
2565  * For kswapd, balance_pgdat() will work across all this node's zones until
2566  * they are all at high_wmark_pages(zone).
2567  *
2568  * Returns the final order kswapd was reclaiming at
2569  *
2570  * There is special handling here for zones which are full of pinned pages.
2571  * This can happen if the pages are all mlocked, or if they are all used by
2572  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2573  * What we do is to detect the case where all pages in the zone have been
2574  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2575  * dead and from now on, only perform a short scan.  Basically we're polling
2576  * the zone for when the problem goes away.
2577  *
2578  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2579  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2580  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2581  * lower zones regardless of the number of free pages in the lower zones. This
2582  * interoperates with the page allocator fallback scheme to ensure that aging
2583  * of pages is balanced across the zones.
2584  */
2585 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2586                                                         int *classzone_idx)
2587 {
2588         int all_zones_ok;
2589         unsigned long balanced;
2590         int priority;
2591         int i;
2592         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2593         unsigned long total_scanned;
2594         struct reclaim_state *reclaim_state = current->reclaim_state;
2595         unsigned long nr_soft_reclaimed;
2596         unsigned long nr_soft_scanned;
2597         struct scan_control sc = {
2598                 .gfp_mask = GFP_KERNEL,
2599                 .may_unmap = 1,
2600                 .may_swap = 1,
2601                 /*
2602                  * kswapd doesn't want to be bailed out while reclaim. because
2603                  * we want to put equal scanning pressure on each zone.
2604                  */
2605                 .nr_to_reclaim = ULONG_MAX,
2606                 .order = order,
2607                 .mem_cgroup = NULL,
2608         };
2609         struct shrink_control shrink = {
2610                 .gfp_mask = sc.gfp_mask,
2611         };
2612 loop_again:
2613         total_scanned = 0;
2614         sc.nr_reclaimed = 0;
2615         sc.may_writepage = !laptop_mode;
2616         count_vm_event(PAGEOUTRUN);
2617
2618         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2619                 unsigned long lru_pages = 0;
2620                 int has_under_min_watermark_zone = 0;
2621
2622                 /* The swap token gets in the way of swapout... */
2623                 if (!priority)
2624                         disable_swap_token(NULL);
2625
2626                 all_zones_ok = 1;
2627                 balanced = 0;
2628
2629                 /*
2630                  * Scan in the highmem->dma direction for the highest
2631                  * zone which needs scanning
2632                  */
2633                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2634                         struct zone *zone = pgdat->node_zones + i;
2635
2636                         if (!populated_zone(zone))
2637                                 continue;
2638
2639                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2640                                 continue;
2641
2642                         /*
2643                          * Do some background aging of the anon list, to give
2644                          * pages a chance to be referenced before reclaiming.
2645                          */
2646                         if (inactive_anon_is_low(zone, &sc))
2647                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2648                                                         &sc, priority, 0);
2649
2650                         if (!zone_watermark_ok_safe(zone, order,
2651                                         high_wmark_pages(zone), 0, 0)) {
2652                                 end_zone = i;
2653                                 break;
2654                         } else {
2655                                 /* If balanced, clear the congested flag */
2656                                 zone_clear_flag(zone, ZONE_CONGESTED);
2657                         }
2658                 }
2659                 if (i < 0)
2660                         goto out;
2661
2662                 for (i = 0; i <= end_zone; i++) {
2663                         struct zone *zone = pgdat->node_zones + i;
2664
2665                         lru_pages += zone_reclaimable_pages(zone);
2666                 }
2667
2668                 /*
2669                  * Now scan the zone in the dma->highmem direction, stopping
2670                  * at the last zone which needs scanning.
2671                  *
2672                  * We do this because the page allocator works in the opposite
2673                  * direction.  This prevents the page allocator from allocating
2674                  * pages behind kswapd's direction of progress, which would
2675                  * cause too much scanning of the lower zones.
2676                  */
2677                 for (i = 0; i <= end_zone; i++) {
2678                         struct zone *zone = pgdat->node_zones + i;
2679                         int nr_slab;
2680                         unsigned long balance_gap;
2681
2682                         if (!populated_zone(zone))
2683                                 continue;
2684
2685                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2686                                 continue;
2687
2688                         sc.nr_scanned = 0;
2689
2690                         nr_soft_scanned = 0;
2691                         /*
2692                          * Call soft limit reclaim before calling shrink_zone.
2693                          */
2694                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2695                                                         order, sc.gfp_mask,
2696                                                         &nr_soft_scanned);
2697                         sc.nr_reclaimed += nr_soft_reclaimed;
2698                         total_scanned += nr_soft_scanned;
2699
2700                         /*
2701                          * We put equal pressure on every zone, unless
2702                          * one zone has way too many pages free
2703                          * already. The "too many pages" is defined
2704                          * as the high wmark plus a "gap" where the
2705                          * gap is either the low watermark or 1%
2706                          * of the zone, whichever is smaller.
2707                          */
2708                         balance_gap = min(low_wmark_pages(zone),
2709                                 (zone->present_pages +
2710                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2711                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2712                         if (!zone_watermark_ok_safe(zone, order,
2713                                         high_wmark_pages(zone) + balance_gap,
2714                                         end_zone, 0)) {
2715                                 shrink_zone(priority, zone, &sc);
2716
2717                                 reclaim_state->reclaimed_slab = 0;
2718                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2719                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2720                                 total_scanned += sc.nr_scanned;
2721
2722                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2723                                         zone->all_unreclaimable = 1;
2724                         }
2725
2726                         /*
2727                          * If we've done a decent amount of scanning and
2728                          * the reclaim ratio is low, start doing writepage
2729                          * even in laptop mode
2730                          */
2731                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2732                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2733                                 sc.may_writepage = 1;
2734
2735                         if (zone->all_unreclaimable) {
2736                                 if (end_zone && end_zone == i)
2737                                         end_zone--;
2738                                 continue;
2739                         }
2740
2741                         if (!zone_watermark_ok_safe(zone, order,
2742                                         high_wmark_pages(zone), end_zone, 0)) {
2743                                 all_zones_ok = 0;
2744                                 /*
2745                                  * We are still under min water mark.  This
2746                                  * means that we have a GFP_ATOMIC allocation
2747                                  * failure risk. Hurry up!
2748                                  */
2749                                 if (!zone_watermark_ok_safe(zone, order,
2750                                             min_wmark_pages(zone), end_zone, 0))
2751                                         has_under_min_watermark_zone = 1;
2752                         } else {
2753                                 /*
2754                                  * If a zone reaches its high watermark,
2755                                  * consider it to be no longer congested. It's
2756                                  * possible there are dirty pages backed by
2757                                  * congested BDIs but as pressure is relieved,
2758                                  * spectulatively avoid congestion waits
2759                                  */
2760                                 zone_clear_flag(zone, ZONE_CONGESTED);
2761                                 if (i <= *classzone_idx)
2762                                         balanced += zone->present_pages;
2763                         }
2764
2765                 }
2766                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2767                         break;          /* kswapd: all done */
2768                 /*
2769                  * OK, kswapd is getting into trouble.  Take a nap, then take
2770                  * another pass across the zones.
2771                  */
2772                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2773                         if (has_under_min_watermark_zone)
2774                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2775                         else
2776                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2777                 }
2778
2779                 /*
2780                  * We do this so kswapd doesn't build up large priorities for
2781                  * example when it is freeing in parallel with allocators. It
2782                  * matches the direct reclaim path behaviour in terms of impact
2783                  * on zone->*_priority.
2784                  */
2785                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2786                         break;
2787         }
2788 out:
2789
2790         /*
2791          * order-0: All zones must meet high watermark for a balanced node
2792          * high-order: Balanced zones must make up at least 25% of the node
2793          *             for the node to be balanced
2794          */
2795         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2796                 cond_resched();
2797
2798                 try_to_freeze();
2799
2800                 /*
2801                  * Fragmentation may mean that the system cannot be
2802                  * rebalanced for high-order allocations in all zones.
2803                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2804                  * it means the zones have been fully scanned and are still
2805                  * not balanced. For high-order allocations, there is
2806                  * little point trying all over again as kswapd may
2807                  * infinite loop.
2808                  *
2809                  * Instead, recheck all watermarks at order-0 as they
2810                  * are the most important. If watermarks are ok, kswapd will go
2811                  * back to sleep. High-order users can still perform direct
2812                  * reclaim if they wish.
2813                  */
2814                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2815                         order = sc.order = 0;
2816
2817                 goto loop_again;
2818         }
2819
2820         /*
2821          * If kswapd was reclaiming at a higher order, it has the option of
2822          * sleeping without all zones being balanced. Before it does, it must
2823          * ensure that the watermarks for order-0 on *all* zones are met and
2824          * that the congestion flags are cleared. The congestion flag must
2825          * be cleared as kswapd is the only mechanism that clears the flag
2826          * and it is potentially going to sleep here.
2827          */
2828         if (order) {
2829                 for (i = 0; i <= end_zone; i++) {
2830                         struct zone *zone = pgdat->node_zones + i;
2831
2832                         if (!populated_zone(zone))
2833                                 continue;
2834
2835                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2836                                 continue;
2837
2838                         /* Confirm the zone is balanced for order-0 */
2839                         if (!zone_watermark_ok(zone, 0,
2840                                         high_wmark_pages(zone), 0, 0)) {
2841                                 order = sc.order = 0;
2842                                 goto loop_again;
2843                         }
2844
2845                         /* If balanced, clear the congested flag */
2846                         zone_clear_flag(zone, ZONE_CONGESTED);
2847                         if (i <= *classzone_idx)
2848                                 balanced += zone->present_pages;
2849                 }
2850         }
2851
2852         /*
2853          * Return the order we were reclaiming at so sleeping_prematurely()
2854          * makes a decision on the order we were last reclaiming at. However,
2855          * if another caller entered the allocator slow path while kswapd
2856          * was awake, order will remain at the higher level
2857          */
2858         *classzone_idx = end_zone;
2859         return order;
2860 }
2861
2862 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2863 {
2864         long remaining = 0;
2865         DEFINE_WAIT(wait);
2866
2867         if (freezing(current) || kthread_should_stop())
2868                 return;
2869
2870         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2871
2872         /* Try to sleep for a short interval */
2873         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2874                 remaining = schedule_timeout(HZ/10);
2875                 finish_wait(&pgdat->kswapd_wait, &wait);
2876                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2877         }
2878
2879         /*
2880          * After a short sleep, check if it was a premature sleep. If not, then
2881          * go fully to sleep until explicitly woken up.
2882          */
2883         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2884                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2885
2886                 /*
2887                  * vmstat counters are not perfectly accurate and the estimated
2888                  * value for counters such as NR_FREE_PAGES can deviate from the
2889                  * true value by nr_online_cpus * threshold. To avoid the zone
2890                  * watermarks being breached while under pressure, we reduce the
2891                  * per-cpu vmstat threshold while kswapd is awake and restore
2892                  * them before going back to sleep.
2893                  */
2894                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2895
2896                 if (!kthread_should_stop())
2897                         schedule();
2898
2899                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2900         } else {
2901                 if (remaining)
2902                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2903                 else
2904                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2905         }
2906         finish_wait(&pgdat->kswapd_wait, &wait);
2907 }
2908
2909 /*
2910  * The background pageout daemon, started as a kernel thread
2911  * from the init process.
2912  *
2913  * This basically trickles out pages so that we have _some_
2914  * free memory available even if there is no other activity
2915  * that frees anything up. This is needed for things like routing
2916  * etc, where we otherwise might have all activity going on in
2917  * asynchronous contexts that cannot page things out.
2918  *
2919  * If there are applications that are active memory-allocators
2920  * (most normal use), this basically shouldn't matter.
2921  */
2922 static int kswapd(void *p)
2923 {
2924         unsigned long order, new_order;
2925         unsigned balanced_order;
2926         int classzone_idx, new_classzone_idx;
2927         int balanced_classzone_idx;
2928         pg_data_t *pgdat = (pg_data_t*)p;
2929         struct task_struct *tsk = current;
2930
2931         struct reclaim_state reclaim_state = {
2932                 .reclaimed_slab = 0,
2933         };
2934         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2935
2936         lockdep_set_current_reclaim_state(GFP_KERNEL);
2937
2938         if (!cpumask_empty(cpumask))
2939                 set_cpus_allowed_ptr(tsk, cpumask);
2940         current->reclaim_state = &reclaim_state;
2941
2942         /*
2943          * Tell the memory management that we're a "memory allocator",
2944          * and that if we need more memory we should get access to it
2945          * regardless (see "__alloc_pages()"). "kswapd" should
2946          * never get caught in the normal page freeing logic.
2947          *
2948          * (Kswapd normally doesn't need memory anyway, but sometimes
2949          * you need a small amount of memory in order to be able to
2950          * page out something else, and this flag essentially protects
2951          * us from recursively trying to free more memory as we're
2952          * trying to free the first piece of memory in the first place).
2953          */
2954         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2955         set_freezable();
2956
2957         order = new_order = 0;
2958         balanced_order = 0;
2959         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2960         balanced_classzone_idx = classzone_idx;
2961         for ( ; ; ) {
2962                 int ret;
2963
2964                 /*
2965                  * If the last balance_pgdat was unsuccessful it's unlikely a
2966                  * new request of a similar or harder type will succeed soon
2967                  * so consider going to sleep on the basis we reclaimed at
2968                  */
2969                 if (balanced_classzone_idx >= new_classzone_idx &&
2970                                         balanced_order == new_order) {
2971                         new_order = pgdat->kswapd_max_order;
2972                         new_classzone_idx = pgdat->classzone_idx;
2973                         pgdat->kswapd_max_order =  0;
2974                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2975                 }
2976
2977                 if (order < new_order || classzone_idx > new_classzone_idx) {
2978                         /*
2979                          * Don't sleep if someone wants a larger 'order'
2980                          * allocation or has tigher zone constraints
2981                          */
2982                         order = new_order;
2983                         classzone_idx = new_classzone_idx;
2984                 } else {
2985                         kswapd_try_to_sleep(pgdat, balanced_order,
2986                                                 balanced_classzone_idx);
2987                         order = pgdat->kswapd_max_order;
2988                         classzone_idx = pgdat->classzone_idx;
2989                         new_order = order;
2990                         new_classzone_idx = classzone_idx;
2991                         pgdat->kswapd_max_order = 0;
2992                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2993                 }
2994
2995                 ret = try_to_freeze();
2996                 if (kthread_should_stop())
2997                         break;
2998
2999                 /*
3000                  * We can speed up thawing tasks if we don't call balance_pgdat
3001                  * after returning from the refrigerator
3002                  */
3003                 if (!ret) {
3004                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3005                         balanced_classzone_idx = classzone_idx;
3006                         balanced_order = balance_pgdat(pgdat, order,
3007                                                 &balanced_classzone_idx);
3008                 }
3009         }
3010
3011         tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3012         current->reclaim_state = NULL;
3013         lockdep_clear_current_reclaim_state();
3014
3015         return 0;
3016 }
3017
3018 /*
3019  * A zone is low on free memory, so wake its kswapd task to service it.
3020  */
3021 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3022 {
3023         pg_data_t *pgdat;
3024
3025         if (!populated_zone(zone))
3026                 return;
3027
3028         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3029                 return;
3030         pgdat = zone->zone_pgdat;
3031         if (pgdat->kswapd_max_order < order) {
3032                 pgdat->kswapd_max_order = order;
3033                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3034         }
3035         if (!waitqueue_active(&pgdat->kswapd_wait))
3036                 return;
3037         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3038                 return;
3039
3040         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3041         wake_up_interruptible(&pgdat->kswapd_wait);
3042 }
3043
3044 /*
3045  * The reclaimable count would be mostly accurate.
3046  * The less reclaimable pages may be
3047  * - mlocked pages, which will be moved to unevictable list when encountered
3048  * - mapped pages, which may require several travels to be reclaimed
3049  * - dirty pages, which is not "instantly" reclaimable
3050  */
3051 unsigned long global_reclaimable_pages(void)
3052 {
3053         int nr;
3054
3055         nr = global_page_state(NR_ACTIVE_FILE) +
3056              global_page_state(NR_INACTIVE_FILE);
3057
3058         if (nr_swap_pages > 0)
3059                 nr += global_page_state(NR_ACTIVE_ANON) +
3060                       global_page_state(NR_INACTIVE_ANON);
3061
3062         return nr;
3063 }
3064
3065 unsigned long zone_reclaimable_pages(struct zone *zone)
3066 {
3067         int nr;
3068
3069         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3070              zone_page_state(zone, NR_INACTIVE_FILE);
3071
3072         if (nr_swap_pages > 0)
3073                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3074                       zone_page_state(zone, NR_INACTIVE_ANON);
3075
3076         return nr;
3077 }
3078
3079 #ifdef CONFIG_HIBERNATION
3080 /*
3081  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3082  * freed pages.
3083  *
3084  * Rather than trying to age LRUs the aim is to preserve the overall
3085  * LRU order by reclaiming preferentially
3086  * inactive > active > active referenced > active mapped
3087  */
3088 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3089 {
3090         struct reclaim_state reclaim_state;
3091         struct scan_control sc = {
3092                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3093                 .may_swap = 1,
3094                 .may_unmap = 1,
3095                 .may_writepage = 1,
3096                 .nr_to_reclaim = nr_to_reclaim,
3097                 .hibernation_mode = 1,