Merge tag 'kvm-3.11-2' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[platform/adaptation/renesas_rcar/renesas_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/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h>  /* for try_to_release_page(),
28                                         buffer_heads_over_limit */
29 #include <linux/mm_inline.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 struct scan_control {
58         /* Incremented by the number of inactive pages that were scanned */
59         unsigned long nr_scanned;
60
61         /* Number of pages freed so far during a call to shrink_zones() */
62         unsigned long nr_reclaimed;
63
64         /* How many pages shrink_list() should reclaim */
65         unsigned long nr_to_reclaim;
66
67         unsigned long hibernation_mode;
68
69         /* This context's GFP mask */
70         gfp_t gfp_mask;
71
72         int may_writepage;
73
74         /* Can mapped pages be reclaimed? */
75         int may_unmap;
76
77         /* Can pages be swapped as part of reclaim? */
78         int may_swap;
79
80         int order;
81
82         /* Scan (total_size >> priority) pages at once */
83         int priority;
84
85         /*
86          * The memory cgroup that hit its limit and as a result is the
87          * primary target of this reclaim invocation.
88          */
89         struct mem_cgroup *target_mem_cgroup;
90
91         /*
92          * Nodemask of nodes allowed by the caller. If NULL, all nodes
93          * are scanned.
94          */
95         nodemask_t      *nodemask;
96 };
97
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field)                    \
102         do {                                                            \
103                 if ((_page)->lru.prev != _base) {                       \
104                         struct page *prev;                              \
105                                                                         \
106                         prev = lru_to_page(&(_page->lru));              \
107                         prefetch(&prev->_field);                        \
108                 }                                                       \
109         } while (0)
110 #else
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
112 #endif
113
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
116         do {                                                            \
117                 if ((_page)->lru.prev != _base) {                       \
118                         struct page *prev;                              \
119                                                                         \
120                         prev = lru_to_page(&(_page->lru));              \
121                         prefetchw(&prev->_field);                       \
122                 }                                                       \
123         } while (0)
124 #else
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
126 #endif
127
128 /*
129  * From 0 .. 100.  Higher means more swappy.
130  */
131 int vm_swappiness = 60;
132 unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
133
134 static LIST_HEAD(shrinker_list);
135 static DECLARE_RWSEM(shrinker_rwsem);
136
137 #ifdef CONFIG_MEMCG
138 static bool global_reclaim(struct scan_control *sc)
139 {
140         return !sc->target_mem_cgroup;
141 }
142 #else
143 static bool global_reclaim(struct scan_control *sc)
144 {
145         return true;
146 }
147 #endif
148
149 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
150 {
151         if (!mem_cgroup_disabled())
152                 return mem_cgroup_get_lru_size(lruvec, lru);
153
154         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
155 }
156
157 /*
158  * Add a shrinker callback to be called from the vm
159  */
160 void register_shrinker(struct shrinker *shrinker)
161 {
162         atomic_long_set(&shrinker->nr_in_batch, 0);
163         down_write(&shrinker_rwsem);
164         list_add_tail(&shrinker->list, &shrinker_list);
165         up_write(&shrinker_rwsem);
166 }
167 EXPORT_SYMBOL(register_shrinker);
168
169 /*
170  * Remove one
171  */
172 void unregister_shrinker(struct shrinker *shrinker)
173 {
174         down_write(&shrinker_rwsem);
175         list_del(&shrinker->list);
176         up_write(&shrinker_rwsem);
177 }
178 EXPORT_SYMBOL(unregister_shrinker);
179
180 static inline int do_shrinker_shrink(struct shrinker *shrinker,
181                                      struct shrink_control *sc,
182                                      unsigned long nr_to_scan)
183 {
184         sc->nr_to_scan = nr_to_scan;
185         return (*shrinker->shrink)(shrinker, sc);
186 }
187
188 #define SHRINK_BATCH 128
189 /*
190  * Call the shrink functions to age shrinkable caches
191  *
192  * Here we assume it costs one seek to replace a lru page and that it also
193  * takes a seek to recreate a cache object.  With this in mind we age equal
194  * percentages of the lru and ageable caches.  This should balance the seeks
195  * generated by these structures.
196  *
197  * If the vm encountered mapped pages on the LRU it increase the pressure on
198  * slab to avoid swapping.
199  *
200  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
201  *
202  * `lru_pages' represents the number of on-LRU pages in all the zones which
203  * are eligible for the caller's allocation attempt.  It is used for balancing
204  * slab reclaim versus page reclaim.
205  *
206  * Returns the number of slab objects which we shrunk.
207  */
208 unsigned long shrink_slab(struct shrink_control *shrink,
209                           unsigned long nr_pages_scanned,
210                           unsigned long lru_pages)
211 {
212         struct shrinker *shrinker;
213         unsigned long ret = 0;
214
215         if (nr_pages_scanned == 0)
216                 nr_pages_scanned = SWAP_CLUSTER_MAX;
217
218         if (!down_read_trylock(&shrinker_rwsem)) {
219                 /* Assume we'll be able to shrink next time */
220                 ret = 1;
221                 goto out;
222         }
223
224         list_for_each_entry(shrinker, &shrinker_list, list) {
225                 unsigned long long delta;
226                 long total_scan;
227                 long max_pass;
228                 int shrink_ret = 0;
229                 long nr;
230                 long new_nr;
231                 long batch_size = shrinker->batch ? shrinker->batch
232                                                   : SHRINK_BATCH;
233
234                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
235                 if (max_pass <= 0)
236                         continue;
237
238                 /*
239                  * copy the current shrinker scan count into a local variable
240                  * and zero it so that other concurrent shrinker invocations
241                  * don't also do this scanning work.
242                  */
243                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
244
245                 total_scan = nr;
246                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
247                 delta *= max_pass;
248                 do_div(delta, lru_pages + 1);
249                 total_scan += delta;
250                 if (total_scan < 0) {
251                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
252                                "delete nr=%ld\n",
253                                shrinker->shrink, total_scan);
254                         total_scan = max_pass;
255                 }
256
257                 /*
258                  * We need to avoid excessive windup on filesystem shrinkers
259                  * due to large numbers of GFP_NOFS allocations causing the
260                  * shrinkers to return -1 all the time. This results in a large
261                  * nr being built up so when a shrink that can do some work
262                  * comes along it empties the entire cache due to nr >>>
263                  * max_pass.  This is bad for sustaining a working set in
264                  * memory.
265                  *
266                  * Hence only allow the shrinker to scan the entire cache when
267                  * a large delta change is calculated directly.
268                  */
269                 if (delta < max_pass / 4)
270                         total_scan = min(total_scan, max_pass / 2);
271
272                 /*
273                  * Avoid risking looping forever due to too large nr value:
274                  * never try to free more than twice the estimate number of
275                  * freeable entries.
276                  */
277                 if (total_scan > max_pass * 2)
278                         total_scan = max_pass * 2;
279
280                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
281                                         nr_pages_scanned, lru_pages,
282                                         max_pass, delta, total_scan);
283
284                 while (total_scan >= batch_size) {
285                         int nr_before;
286
287                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
288                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
289                                                         batch_size);
290                         if (shrink_ret == -1)
291                                 break;
292                         if (shrink_ret < nr_before)
293                                 ret += nr_before - shrink_ret;
294                         count_vm_events(SLABS_SCANNED, batch_size);
295                         total_scan -= batch_size;
296
297                         cond_resched();
298                 }
299
300                 /*
301                  * move the unused scan count back into the shrinker in a
302                  * manner that handles concurrent updates. If we exhausted the
303                  * scan, there is no need to do an update.
304                  */
305                 if (total_scan > 0)
306                         new_nr = atomic_long_add_return(total_scan,
307                                         &shrinker->nr_in_batch);
308                 else
309                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
310
311                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
312         }
313         up_read(&shrinker_rwsem);
314 out:
315         cond_resched();
316         return ret;
317 }
318
319 static inline int is_page_cache_freeable(struct page *page)
320 {
321         /*
322          * A freeable page cache page is referenced only by the caller
323          * that isolated the page, the page cache radix tree and
324          * optional buffer heads at page->private.
325          */
326         return page_count(page) - page_has_private(page) == 2;
327 }
328
329 static int may_write_to_queue(struct backing_dev_info *bdi,
330                               struct scan_control *sc)
331 {
332         if (current->flags & PF_SWAPWRITE)
333                 return 1;
334         if (!bdi_write_congested(bdi))
335                 return 1;
336         if (bdi == current->backing_dev_info)
337                 return 1;
338         return 0;
339 }
340
341 /*
342  * We detected a synchronous write error writing a page out.  Probably
343  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
344  * fsync(), msync() or close().
345  *
346  * The tricky part is that after writepage we cannot touch the mapping: nothing
347  * prevents it from being freed up.  But we have a ref on the page and once
348  * that page is locked, the mapping is pinned.
349  *
350  * We're allowed to run sleeping lock_page() here because we know the caller has
351  * __GFP_FS.
352  */
353 static void handle_write_error(struct address_space *mapping,
354                                 struct page *page, int error)
355 {
356         lock_page(page);
357         if (page_mapping(page) == mapping)
358                 mapping_set_error(mapping, error);
359         unlock_page(page);
360 }
361
362 /* possible outcome of pageout() */
363 typedef enum {
364         /* failed to write page out, page is locked */
365         PAGE_KEEP,
366         /* move page to the active list, page is locked */
367         PAGE_ACTIVATE,
368         /* page has been sent to the disk successfully, page is unlocked */
369         PAGE_SUCCESS,
370         /* page is clean and locked */
371         PAGE_CLEAN,
372 } pageout_t;
373
374 /*
375  * pageout is called by shrink_page_list() for each dirty page.
376  * Calls ->writepage().
377  */
378 static pageout_t pageout(struct page *page, struct address_space *mapping,
379                          struct scan_control *sc)
380 {
381         /*
382          * If the page is dirty, only perform writeback if that write
383          * will be non-blocking.  To prevent this allocation from being
384          * stalled by pagecache activity.  But note that there may be
385          * stalls if we need to run get_block().  We could test
386          * PagePrivate for that.
387          *
388          * If this process is currently in __generic_file_aio_write() against
389          * this page's queue, we can perform writeback even if that
390          * will block.
391          *
392          * If the page is swapcache, write it back even if that would
393          * block, for some throttling. This happens by accident, because
394          * swap_backing_dev_info is bust: it doesn't reflect the
395          * congestion state of the swapdevs.  Easy to fix, if needed.
396          */
397         if (!is_page_cache_freeable(page))
398                 return PAGE_KEEP;
399         if (!mapping) {
400                 /*
401                  * Some data journaling orphaned pages can have
402                  * page->mapping == NULL while being dirty with clean buffers.
403                  */
404                 if (page_has_private(page)) {
405                         if (try_to_free_buffers(page)) {
406                                 ClearPageDirty(page);
407                                 printk("%s: orphaned page\n", __func__);
408                                 return PAGE_CLEAN;
409                         }
410                 }
411                 return PAGE_KEEP;
412         }
413         if (mapping->a_ops->writepage == NULL)
414                 return PAGE_ACTIVATE;
415         if (!may_write_to_queue(mapping->backing_dev_info, sc))
416                 return PAGE_KEEP;
417
418         if (clear_page_dirty_for_io(page)) {
419                 int res;
420                 struct writeback_control wbc = {
421                         .sync_mode = WB_SYNC_NONE,
422                         .nr_to_write = SWAP_CLUSTER_MAX,
423                         .range_start = 0,
424                         .range_end = LLONG_MAX,
425                         .for_reclaim = 1,
426                 };
427
428                 SetPageReclaim(page);
429                 res = mapping->a_ops->writepage(page, &wbc);
430                 if (res < 0)
431                         handle_write_error(mapping, page, res);
432                 if (res == AOP_WRITEPAGE_ACTIVATE) {
433                         ClearPageReclaim(page);
434                         return PAGE_ACTIVATE;
435                 }
436
437                 if (!PageWriteback(page)) {
438                         /* synchronous write or broken a_ops? */
439                         ClearPageReclaim(page);
440                 }
441                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
442                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
443                 return PAGE_SUCCESS;
444         }
445
446         return PAGE_CLEAN;
447 }
448
449 /*
450  * Same as remove_mapping, but if the page is removed from the mapping, it
451  * gets returned with a refcount of 0.
452  */
453 static int __remove_mapping(struct address_space *mapping, struct page *page)
454 {
455         BUG_ON(!PageLocked(page));
456         BUG_ON(mapping != page_mapping(page));
457
458         spin_lock_irq(&mapping->tree_lock);
459         /*
460          * The non racy check for a busy page.
461          *
462          * Must be careful with the order of the tests. When someone has
463          * a ref to the page, it may be possible that they dirty it then
464          * drop the reference. So if PageDirty is tested before page_count
465          * here, then the following race may occur:
466          *
467          * get_user_pages(&page);
468          * [user mapping goes away]
469          * write_to(page);
470          *                              !PageDirty(page)    [good]
471          * SetPageDirty(page);
472          * put_page(page);
473          *                              !page_count(page)   [good, discard it]
474          *
475          * [oops, our write_to data is lost]
476          *
477          * Reversing the order of the tests ensures such a situation cannot
478          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
479          * load is not satisfied before that of page->_count.
480          *
481          * Note that if SetPageDirty is always performed via set_page_dirty,
482          * and thus under tree_lock, then this ordering is not required.
483          */
484         if (!page_freeze_refs(page, 2))
485                 goto cannot_free;
486         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
487         if (unlikely(PageDirty(page))) {
488                 page_unfreeze_refs(page, 2);
489                 goto cannot_free;
490         }
491
492         if (PageSwapCache(page)) {
493                 swp_entry_t swap = { .val = page_private(page) };
494                 __delete_from_swap_cache(page);
495                 spin_unlock_irq(&mapping->tree_lock);
496                 swapcache_free(swap, page);
497         } else {
498                 void (*freepage)(struct page *);
499
500                 freepage = mapping->a_ops->freepage;
501
502                 __delete_from_page_cache(page);
503                 spin_unlock_irq(&mapping->tree_lock);
504                 mem_cgroup_uncharge_cache_page(page);
505
506                 if (freepage != NULL)
507                         freepage(page);
508         }
509
510         return 1;
511
512 cannot_free:
513         spin_unlock_irq(&mapping->tree_lock);
514         return 0;
515 }
516
517 /*
518  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
519  * someone else has a ref on the page, abort and return 0.  If it was
520  * successfully detached, return 1.  Assumes the caller has a single ref on
521  * this page.
522  */
523 int remove_mapping(struct address_space *mapping, struct page *page)
524 {
525         if (__remove_mapping(mapping, page)) {
526                 /*
527                  * Unfreezing the refcount with 1 rather than 2 effectively
528                  * drops the pagecache ref for us without requiring another
529                  * atomic operation.
530                  */
531                 page_unfreeze_refs(page, 1);
532                 return 1;
533         }
534         return 0;
535 }
536
537 /**
538  * putback_lru_page - put previously isolated page onto appropriate LRU list
539  * @page: page to be put back to appropriate lru list
540  *
541  * Add previously isolated @page to appropriate LRU list.
542  * Page may still be unevictable for other reasons.
543  *
544  * lru_lock must not be held, interrupts must be enabled.
545  */
546 void putback_lru_page(struct page *page)
547 {
548         int lru;
549         int was_unevictable = PageUnevictable(page);
550
551         VM_BUG_ON(PageLRU(page));
552
553 redo:
554         ClearPageUnevictable(page);
555
556         if (page_evictable(page)) {
557                 /*
558                  * For evictable pages, we can use the cache.
559                  * In event of a race, worst case is we end up with an
560                  * unevictable page on [in]active list.
561                  * We know how to handle that.
562                  */
563                 lru = page_lru_base_type(page);
564                 lru_cache_add(page);
565         } else {
566                 /*
567                  * Put unevictable pages directly on zone's unevictable
568                  * list.
569                  */
570                 lru = LRU_UNEVICTABLE;
571                 add_page_to_unevictable_list(page);
572                 /*
573                  * When racing with an mlock or AS_UNEVICTABLE clearing
574                  * (page is unlocked) make sure that if the other thread
575                  * does not observe our setting of PG_lru and fails
576                  * isolation/check_move_unevictable_pages,
577                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578                  * the page back to the evictable list.
579                  *
580                  * The other side is TestClearPageMlocked() or shmem_lock().
581                  */
582                 smp_mb();
583         }
584
585         /*
586          * page's status can change while we move it among lru. If an evictable
587          * page is on unevictable list, it never be freed. To avoid that,
588          * check after we added it to the list, again.
589          */
590         if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
591                 if (!isolate_lru_page(page)) {
592                         put_page(page);
593                         goto redo;
594                 }
595                 /* This means someone else dropped this page from LRU
596                  * So, it will be freed or putback to LRU again. There is
597                  * nothing to do here.
598                  */
599         }
600
601         if (was_unevictable && lru != LRU_UNEVICTABLE)
602                 count_vm_event(UNEVICTABLE_PGRESCUED);
603         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604                 count_vm_event(UNEVICTABLE_PGCULLED);
605
606         put_page(page);         /* drop ref from isolate */
607 }
608
609 enum page_references {
610         PAGEREF_RECLAIM,
611         PAGEREF_RECLAIM_CLEAN,
612         PAGEREF_KEEP,
613         PAGEREF_ACTIVATE,
614 };
615
616 static enum page_references page_check_references(struct page *page,
617                                                   struct scan_control *sc)
618 {
619         int referenced_ptes, referenced_page;
620         unsigned long vm_flags;
621
622         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623                                           &vm_flags);
624         referenced_page = TestClearPageReferenced(page);
625
626         /*
627          * Mlock lost the isolation race with us.  Let try_to_unmap()
628          * move the page to the unevictable list.
629          */
630         if (vm_flags & VM_LOCKED)
631                 return PAGEREF_RECLAIM;
632
633         if (referenced_ptes) {
634                 if (PageSwapBacked(page))
635                         return PAGEREF_ACTIVATE;
636                 /*
637                  * All mapped pages start out with page table
638                  * references from the instantiating fault, so we need
639                  * to look twice if a mapped file page is used more
640                  * than once.
641                  *
642                  * Mark it and spare it for another trip around the
643                  * inactive list.  Another page table reference will
644                  * lead to its activation.
645                  *
646                  * Note: the mark is set for activated pages as well
647                  * so that recently deactivated but used pages are
648                  * quickly recovered.
649                  */
650                 SetPageReferenced(page);
651
652                 if (referenced_page || referenced_ptes > 1)
653                         return PAGEREF_ACTIVATE;
654
655                 /*
656                  * Activate file-backed executable pages after first usage.
657                  */
658                 if (vm_flags & VM_EXEC)
659                         return PAGEREF_ACTIVATE;
660
661                 return PAGEREF_KEEP;
662         }
663
664         /* Reclaim if clean, defer dirty pages to writeback */
665         if (referenced_page && !PageSwapBacked(page))
666                 return PAGEREF_RECLAIM_CLEAN;
667
668         return PAGEREF_RECLAIM;
669 }
670
671 /* Check if a page is dirty or under writeback */
672 static void page_check_dirty_writeback(struct page *page,
673                                        bool *dirty, bool *writeback)
674 {
675         struct address_space *mapping;
676
677         /*
678          * Anonymous pages are not handled by flushers and must be written
679          * from reclaim context. Do not stall reclaim based on them
680          */
681         if (!page_is_file_cache(page)) {
682                 *dirty = false;
683                 *writeback = false;
684                 return;
685         }
686
687         /* By default assume that the page flags are accurate */
688         *dirty = PageDirty(page);
689         *writeback = PageWriteback(page);
690
691         /* Verify dirty/writeback state if the filesystem supports it */
692         if (!page_has_private(page))
693                 return;
694
695         mapping = page_mapping(page);
696         if (mapping && mapping->a_ops->is_dirty_writeback)
697                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
698 }
699
700 /*
701  * shrink_page_list() returns the number of reclaimed pages
702  */
703 static unsigned long shrink_page_list(struct list_head *page_list,
704                                       struct zone *zone,
705                                       struct scan_control *sc,
706                                       enum ttu_flags ttu_flags,
707                                       unsigned long *ret_nr_dirty,
708                                       unsigned long *ret_nr_unqueued_dirty,
709                                       unsigned long *ret_nr_congested,
710                                       unsigned long *ret_nr_writeback,
711                                       unsigned long *ret_nr_immediate,
712                                       bool force_reclaim)
713 {
714         LIST_HEAD(ret_pages);
715         LIST_HEAD(free_pages);
716         int pgactivate = 0;
717         unsigned long nr_unqueued_dirty = 0;
718         unsigned long nr_dirty = 0;
719         unsigned long nr_congested = 0;
720         unsigned long nr_reclaimed = 0;
721         unsigned long nr_writeback = 0;
722         unsigned long nr_immediate = 0;
723
724         cond_resched();
725
726         mem_cgroup_uncharge_start();
727         while (!list_empty(page_list)) {
728                 struct address_space *mapping;
729                 struct page *page;
730                 int may_enter_fs;
731                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
732                 bool dirty, writeback;
733
734                 cond_resched();
735
736                 page = lru_to_page(page_list);
737                 list_del(&page->lru);
738
739                 if (!trylock_page(page))
740                         goto keep;
741
742                 VM_BUG_ON(PageActive(page));
743                 VM_BUG_ON(page_zone(page) != zone);
744
745                 sc->nr_scanned++;
746
747                 if (unlikely(!page_evictable(page)))
748                         goto cull_mlocked;
749
750                 if (!sc->may_unmap && page_mapped(page))
751                         goto keep_locked;
752
753                 /* Double the slab pressure for mapped and swapcache pages */
754                 if (page_mapped(page) || PageSwapCache(page))
755                         sc->nr_scanned++;
756
757                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
758                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
759
760                 /*
761                  * The number of dirty pages determines if a zone is marked
762                  * reclaim_congested which affects wait_iff_congested. kswapd
763                  * will stall and start writing pages if the tail of the LRU
764                  * is all dirty unqueued pages.
765                  */
766                 page_check_dirty_writeback(page, &dirty, &writeback);
767                 if (dirty || writeback)
768                         nr_dirty++;
769
770                 if (dirty && !writeback)
771                         nr_unqueued_dirty++;
772
773                 /*
774                  * Treat this page as congested if the underlying BDI is or if
775                  * pages are cycling through the LRU so quickly that the
776                  * pages marked for immediate reclaim are making it to the
777                  * end of the LRU a second time.
778                  */
779                 mapping = page_mapping(page);
780                 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
781                     (writeback && PageReclaim(page)))
782                         nr_congested++;
783
784                 /*
785                  * If a page at the tail of the LRU is under writeback, there
786                  * are three cases to consider.
787                  *
788                  * 1) If reclaim is encountering an excessive number of pages
789                  *    under writeback and this page is both under writeback and
790                  *    PageReclaim then it indicates that pages are being queued
791                  *    for IO but are being recycled through the LRU before the
792                  *    IO can complete. Waiting on the page itself risks an
793                  *    indefinite stall if it is impossible to writeback the
794                  *    page due to IO error or disconnected storage so instead
795                  *    note that the LRU is being scanned too quickly and the
796                  *    caller can stall after page list has been processed.
797                  *
798                  * 2) Global reclaim encounters a page, memcg encounters a
799                  *    page that is not marked for immediate reclaim or
800                  *    the caller does not have __GFP_IO. In this case mark
801                  *    the page for immediate reclaim and continue scanning.
802                  *
803                  *    __GFP_IO is checked  because a loop driver thread might
804                  *    enter reclaim, and deadlock if it waits on a page for
805                  *    which it is needed to do the write (loop masks off
806                  *    __GFP_IO|__GFP_FS for this reason); but more thought
807                  *    would probably show more reasons.
808                  *
809                  *    Don't require __GFP_FS, since we're not going into the
810                  *    FS, just waiting on its writeback completion. Worryingly,
811                  *    ext4 gfs2 and xfs allocate pages with
812                  *    grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
813                  *    may_enter_fs here is liable to OOM on them.
814                  *
815                  * 3) memcg encounters a page that is not already marked
816                  *    PageReclaim. memcg does not have any dirty pages
817                  *    throttling so we could easily OOM just because too many
818                  *    pages are in writeback and there is nothing else to
819                  *    reclaim. Wait for the writeback to complete.
820                  */
821                 if (PageWriteback(page)) {
822                         /* Case 1 above */
823                         if (current_is_kswapd() &&
824                             PageReclaim(page) &&
825                             zone_is_reclaim_writeback(zone)) {
826                                 nr_immediate++;
827                                 goto keep_locked;
828
829                         /* Case 2 above */
830                         } else if (global_reclaim(sc) ||
831                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
832                                 /*
833                                  * This is slightly racy - end_page_writeback()
834                                  * might have just cleared PageReclaim, then
835                                  * setting PageReclaim here end up interpreted
836                                  * as PageReadahead - but that does not matter
837                                  * enough to care.  What we do want is for this
838                                  * page to have PageReclaim set next time memcg
839                                  * reclaim reaches the tests above, so it will
840                                  * then wait_on_page_writeback() to avoid OOM;
841                                  * and it's also appropriate in global reclaim.
842                                  */
843                                 SetPageReclaim(page);
844                                 nr_writeback++;
845
846                                 goto keep_locked;
847
848                         /* Case 3 above */
849                         } else {
850                                 wait_on_page_writeback(page);
851                         }
852                 }
853
854                 if (!force_reclaim)
855                         references = page_check_references(page, sc);
856
857                 switch (references) {
858                 case PAGEREF_ACTIVATE:
859                         goto activate_locked;
860                 case PAGEREF_KEEP:
861                         goto keep_locked;
862                 case PAGEREF_RECLAIM:
863                 case PAGEREF_RECLAIM_CLEAN:
864                         ; /* try to reclaim the page below */
865                 }
866
867                 /*
868                  * Anonymous process memory has backing store?
869                  * Try to allocate it some swap space here.
870                  */
871                 if (PageAnon(page) && !PageSwapCache(page)) {
872                         if (!(sc->gfp_mask & __GFP_IO))
873                                 goto keep_locked;
874                         if (!add_to_swap(page, page_list))
875                                 goto activate_locked;
876                         may_enter_fs = 1;
877
878                         /* Adding to swap updated mapping */
879                         mapping = page_mapping(page);
880                 }
881
882                 /*
883                  * The page is mapped into the page tables of one or more
884                  * processes. Try to unmap it here.
885                  */
886                 if (page_mapped(page) && mapping) {
887                         switch (try_to_unmap(page, ttu_flags)) {
888                         case SWAP_FAIL:
889                                 goto activate_locked;
890                         case SWAP_AGAIN:
891                                 goto keep_locked;
892                         case SWAP_MLOCK:
893                                 goto cull_mlocked;
894                         case SWAP_SUCCESS:
895                                 ; /* try to free the page below */
896                         }
897                 }
898
899                 if (PageDirty(page)) {
900                         /*
901                          * Only kswapd can writeback filesystem pages to
902                          * avoid risk of stack overflow but only writeback
903                          * if many dirty pages have been encountered.
904                          */
905                         if (page_is_file_cache(page) &&
906                                         (!current_is_kswapd() ||
907                                          !zone_is_reclaim_dirty(zone))) {
908                                 /*
909                                  * Immediately reclaim when written back.
910                                  * Similar in principal to deactivate_page()
911                                  * except we already have the page isolated
912                                  * and know it's dirty
913                                  */
914                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
915                                 SetPageReclaim(page);
916
917                                 goto keep_locked;
918                         }
919
920                         if (references == PAGEREF_RECLAIM_CLEAN)
921                                 goto keep_locked;
922                         if (!may_enter_fs)
923                                 goto keep_locked;
924                         if (!sc->may_writepage)
925                                 goto keep_locked;
926
927                         /* Page is dirty, try to write it out here */
928                         switch (pageout(page, mapping, sc)) {
929                         case PAGE_KEEP:
930                                 goto keep_locked;
931                         case PAGE_ACTIVATE:
932                                 goto activate_locked;
933                         case PAGE_SUCCESS:
934                                 if (PageWriteback(page))
935                                         goto keep;
936                                 if (PageDirty(page))
937                                         goto keep;
938
939                                 /*
940                                  * A synchronous write - probably a ramdisk.  Go
941                                  * ahead and try to reclaim the page.
942                                  */
943                                 if (!trylock_page(page))
944                                         goto keep;
945                                 if (PageDirty(page) || PageWriteback(page))
946                                         goto keep_locked;
947                                 mapping = page_mapping(page);
948                         case PAGE_CLEAN:
949                                 ; /* try to free the page below */
950                         }
951                 }
952
953                 /*
954                  * If the page has buffers, try to free the buffer mappings
955                  * associated with this page. If we succeed we try to free
956                  * the page as well.
957                  *
958                  * We do this even if the page is PageDirty().
959                  * try_to_release_page() does not perform I/O, but it is
960                  * possible for a page to have PageDirty set, but it is actually
961                  * clean (all its buffers are clean).  This happens if the
962                  * buffers were written out directly, with submit_bh(). ext3
963                  * will do this, as well as the blockdev mapping.
964                  * try_to_release_page() will discover that cleanness and will
965                  * drop the buffers and mark the page clean - it can be freed.
966                  *
967                  * Rarely, pages can have buffers and no ->mapping.  These are
968                  * the pages which were not successfully invalidated in
969                  * truncate_complete_page().  We try to drop those buffers here
970                  * and if that worked, and the page is no longer mapped into
971                  * process address space (page_count == 1) it can be freed.
972                  * Otherwise, leave the page on the LRU so it is swappable.
973                  */
974                 if (page_has_private(page)) {
975                         if (!try_to_release_page(page, sc->gfp_mask))
976                                 goto activate_locked;
977                         if (!mapping && page_count(page) == 1) {
978                                 unlock_page(page);
979                                 if (put_page_testzero(page))
980                                         goto free_it;
981                                 else {
982                                         /*
983                                          * rare race with speculative reference.
984                                          * the speculative reference will free
985                                          * this page shortly, so we may
986                                          * increment nr_reclaimed here (and
987                                          * leave it off the LRU).
988                                          */
989                                         nr_reclaimed++;
990                                         continue;
991                                 }
992                         }
993                 }
994
995                 if (!mapping || !__remove_mapping(mapping, page))
996                         goto keep_locked;
997
998                 /*
999                  * At this point, we have no other references and there is
1000                  * no way to pick any more up (removed from LRU, removed
1001                  * from pagecache). Can use non-atomic bitops now (and
1002                  * we obviously don't have to worry about waking up a process
1003                  * waiting on the page lock, because there are no references.
1004                  */
1005                 __clear_page_locked(page);
1006 free_it:
1007                 nr_reclaimed++;
1008
1009                 /*
1010                  * Is there need to periodically free_page_list? It would
1011                  * appear not as the counts should be low
1012                  */
1013                 list_add(&page->lru, &free_pages);
1014                 continue;
1015
1016 cull_mlocked:
1017                 if (PageSwapCache(page))
1018                         try_to_free_swap(page);
1019                 unlock_page(page);
1020                 putback_lru_page(page);
1021                 continue;
1022
1023 activate_locked:
1024                 /* Not a candidate for swapping, so reclaim swap space. */
1025                 if (PageSwapCache(page) && vm_swap_full())
1026                         try_to_free_swap(page);
1027                 VM_BUG_ON(PageActive(page));
1028                 SetPageActive(page);
1029                 pgactivate++;
1030 keep_locked:
1031                 unlock_page(page);
1032 keep:
1033                 list_add(&page->lru, &ret_pages);
1034                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1035         }
1036
1037         free_hot_cold_page_list(&free_pages, 1);
1038
1039         list_splice(&ret_pages, page_list);
1040         count_vm_events(PGACTIVATE, pgactivate);
1041         mem_cgroup_uncharge_end();
1042         *ret_nr_dirty += nr_dirty;
1043         *ret_nr_congested += nr_congested;
1044         *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1045         *ret_nr_writeback += nr_writeback;
1046         *ret_nr_immediate += nr_immediate;
1047         return nr_reclaimed;
1048 }
1049
1050 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1051                                             struct list_head *page_list)
1052 {
1053         struct scan_control sc = {
1054                 .gfp_mask = GFP_KERNEL,
1055                 .priority = DEF_PRIORITY,
1056                 .may_unmap = 1,
1057         };
1058         unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1059         struct page *page, *next;
1060         LIST_HEAD(clean_pages);
1061
1062         list_for_each_entry_safe(page, next, page_list, lru) {
1063                 if (page_is_file_cache(page) && !PageDirty(page)) {
1064                         ClearPageActive(page);
1065                         list_move(&page->lru, &clean_pages);
1066                 }
1067         }
1068
1069         ret = shrink_page_list(&clean_pages, zone, &sc,
1070                         TTU_UNMAP|TTU_IGNORE_ACCESS,
1071                         &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1072         list_splice(&clean_pages, page_list);
1073         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1074         return ret;
1075 }
1076
1077 /*
1078  * Attempt to remove the specified page from its LRU.  Only take this page
1079  * if it is of the appropriate PageActive status.  Pages which are being
1080  * freed elsewhere are also ignored.
1081  *
1082  * page:        page to consider
1083  * mode:        one of the LRU isolation modes defined above
1084  *
1085  * returns 0 on success, -ve errno on failure.
1086  */
1087 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1088 {
1089         int ret = -EINVAL;
1090
1091         /* Only take pages on the LRU. */
1092         if (!PageLRU(page))
1093                 return ret;
1094
1095         /* Compaction should not handle unevictable pages but CMA can do so */
1096         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1097                 return ret;
1098
1099         ret = -EBUSY;
1100
1101         /*
1102          * To minimise LRU disruption, the caller can indicate that it only
1103          * wants to isolate pages it will be able to operate on without
1104          * blocking - clean pages for the most part.
1105          *
1106          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1107          * is used by reclaim when it is cannot write to backing storage
1108          *
1109          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1110          * that it is possible to migrate without blocking
1111          */
1112         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1113                 /* All the caller can do on PageWriteback is block */
1114                 if (PageWriteback(page))
1115                         return ret;
1116
1117                 if (PageDirty(page)) {
1118                         struct address_space *mapping;
1119
1120                         /* ISOLATE_CLEAN means only clean pages */
1121                         if (mode & ISOLATE_CLEAN)
1122                                 return ret;
1123
1124                         /*
1125                          * Only pages without mappings or that have a
1126                          * ->migratepage callback are possible to migrate
1127                          * without blocking
1128                          */
1129                         mapping = page_mapping(page);
1130                         if (mapping && !mapping->a_ops->migratepage)
1131                                 return ret;
1132                 }
1133         }
1134
1135         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1136                 return ret;
1137
1138         if (likely(get_page_unless_zero(page))) {
1139                 /*
1140                  * Be careful not to clear PageLRU until after we're
1141                  * sure the page is not being freed elsewhere -- the
1142                  * page release code relies on it.
1143                  */
1144                 ClearPageLRU(page);
1145                 ret = 0;
1146         }
1147
1148         return ret;
1149 }
1150
1151 /*
1152  * zone->lru_lock is heavily contended.  Some of the functions that
1153  * shrink the lists perform better by taking out a batch of pages
1154  * and working on them outside the LRU lock.
1155  *
1156  * For pagecache intensive workloads, this function is the hottest
1157  * spot in the kernel (apart from copy_*_user functions).
1158  *
1159  * Appropriate locks must be held before calling this function.
1160  *
1161  * @nr_to_scan: The number of pages to look through on the list.
1162  * @lruvec:     The LRU vector to pull pages from.
1163  * @dst:        The temp list to put pages on to.
1164  * @nr_scanned: The number of pages that were scanned.
1165  * @sc:         The scan_control struct for this reclaim session
1166  * @mode:       One of the LRU isolation modes
1167  * @lru:        LRU list id for isolating
1168  *
1169  * returns how many pages were moved onto *@dst.
1170  */
1171 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1172                 struct lruvec *lruvec, struct list_head *dst,
1173                 unsigned long *nr_scanned, struct scan_control *sc,
1174                 isolate_mode_t mode, enum lru_list lru)
1175 {
1176         struct list_head *src = &lruvec->lists[lru];
1177         unsigned long nr_taken = 0;
1178         unsigned long scan;
1179
1180         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1181                 struct page *page;
1182                 int nr_pages;
1183
1184                 page = lru_to_page(src);
1185                 prefetchw_prev_lru_page(page, src, flags);
1186
1187                 VM_BUG_ON(!PageLRU(page));
1188
1189                 switch (__isolate_lru_page(page, mode)) {
1190                 case 0:
1191                         nr_pages = hpage_nr_pages(page);
1192                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1193                         list_move(&page->lru, dst);
1194                         nr_taken += nr_pages;
1195                         break;
1196
1197                 case -EBUSY:
1198                         /* else it is being freed elsewhere */
1199                         list_move(&page->lru, src);
1200                         continue;
1201
1202                 default:
1203                         BUG();
1204                 }
1205         }
1206
1207         *nr_scanned = scan;
1208         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1209                                     nr_taken, mode, is_file_lru(lru));
1210         return nr_taken;
1211 }
1212
1213 /**
1214  * isolate_lru_page - tries to isolate a page from its LRU list
1215  * @page: page to isolate from its LRU list
1216  *
1217  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1218  * vmstat statistic corresponding to whatever LRU list the page was on.
1219  *
1220  * Returns 0 if the page was removed from an LRU list.
1221  * Returns -EBUSY if the page was not on an LRU list.
1222  *
1223  * The returned page will have PageLRU() cleared.  If it was found on
1224  * the active list, it will have PageActive set.  If it was found on
1225  * the unevictable list, it will have the PageUnevictable bit set. That flag
1226  * may need to be cleared by the caller before letting the page go.
1227  *
1228  * The vmstat statistic corresponding to the list on which the page was
1229  * found will be decremented.
1230  *
1231  * Restrictions:
1232  * (1) Must be called with an elevated refcount on the page. This is a
1233  *     fundamentnal difference from isolate_lru_pages (which is called
1234  *     without a stable reference).
1235  * (2) the lru_lock must not be held.
1236  * (3) interrupts must be enabled.
1237  */
1238 int isolate_lru_page(struct page *page)
1239 {
1240         int ret = -EBUSY;
1241
1242         VM_BUG_ON(!page_count(page));
1243
1244         if (PageLRU(page)) {
1245                 struct zone *zone = page_zone(page);
1246                 struct lruvec *lruvec;
1247
1248                 spin_lock_irq(&zone->lru_lock);
1249                 lruvec = mem_cgroup_page_lruvec(page, zone);
1250                 if (PageLRU(page)) {
1251                         int lru = page_lru(page);
1252                         get_page(page);
1253                         ClearPageLRU(page);
1254                         del_page_from_lru_list(page, lruvec, lru);
1255                         ret = 0;
1256                 }
1257                 spin_unlock_irq(&zone->lru_lock);
1258         }
1259         return ret;
1260 }
1261
1262 /*
1263  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1264  * then get resheduled. When there are massive number of tasks doing page
1265  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1266  * the LRU list will go small and be scanned faster than necessary, leading to
1267  * unnecessary swapping, thrashing and OOM.
1268  */
1269 static int too_many_isolated(struct zone *zone, int file,
1270                 struct scan_control *sc)
1271 {
1272         unsigned long inactive, isolated;
1273
1274         if (current_is_kswapd())
1275                 return 0;
1276
1277         if (!global_reclaim(sc))
1278                 return 0;
1279
1280         if (file) {
1281                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1282                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1283         } else {
1284                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1285                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1286         }
1287
1288         /*
1289          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1290          * won't get blocked by normal direct-reclaimers, forming a circular
1291          * deadlock.
1292          */
1293         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1294                 inactive >>= 3;
1295
1296         return isolated > inactive;
1297 }
1298
1299 static noinline_for_stack void
1300 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1301 {
1302         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1303         struct zone *zone = lruvec_zone(lruvec);
1304         LIST_HEAD(pages_to_free);
1305
1306         /*
1307          * Put back any unfreeable pages.
1308          */
1309         while (!list_empty(page_list)) {
1310                 struct page *page = lru_to_page(page_list);
1311                 int lru;
1312
1313                 VM_BUG_ON(PageLRU(page));
1314                 list_del(&page->lru);
1315                 if (unlikely(!page_evictable(page))) {
1316                         spin_unlock_irq(&zone->lru_lock);
1317                         putback_lru_page(page);
1318                         spin_lock_irq(&zone->lru_lock);
1319                         continue;
1320                 }
1321
1322                 lruvec = mem_cgroup_page_lruvec(page, zone);
1323
1324                 SetPageLRU(page);
1325                 lru = page_lru(page);
1326                 add_page_to_lru_list(page, lruvec, lru);
1327
1328                 if (is_active_lru(lru)) {
1329                         int file = is_file_lru(lru);
1330                         int numpages = hpage_nr_pages(page);
1331                         reclaim_stat->recent_rotated[file] += numpages;
1332                 }
1333                 if (put_page_testzero(page)) {
1334                         __ClearPageLRU(page);
1335                         __ClearPageActive(page);
1336                         del_page_from_lru_list(page, lruvec, lru);
1337
1338                         if (unlikely(PageCompound(page))) {
1339                                 spin_unlock_irq(&zone->lru_lock);
1340                                 (*get_compound_page_dtor(page))(page);
1341                                 spin_lock_irq(&zone->lru_lock);
1342                         } else
1343                                 list_add(&page->lru, &pages_to_free);
1344                 }
1345         }
1346
1347         /*
1348          * To save our caller's stack, now use input list for pages to free.
1349          */
1350         list_splice(&pages_to_free, page_list);
1351 }
1352
1353 /*
1354  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1355  * of reclaimed pages
1356  */
1357 static noinline_for_stack unsigned long
1358 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1359                      struct scan_control *sc, enum lru_list lru)
1360 {
1361         LIST_HEAD(page_list);
1362         unsigned long nr_scanned;
1363         unsigned long nr_reclaimed = 0;
1364         unsigned long nr_taken;
1365         unsigned long nr_dirty = 0;
1366         unsigned long nr_congested = 0;
1367         unsigned long nr_unqueued_dirty = 0;
1368         unsigned long nr_writeback = 0;
1369         unsigned long nr_immediate = 0;
1370         isolate_mode_t isolate_mode = 0;
1371         int file = is_file_lru(lru);
1372         struct zone *zone = lruvec_zone(lruvec);
1373         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1374
1375         while (unlikely(too_many_isolated(zone, file, sc))) {
1376                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1377
1378                 /* We are about to die and free our memory. Return now. */
1379                 if (fatal_signal_pending(current))
1380                         return SWAP_CLUSTER_MAX;
1381         }
1382
1383         lru_add_drain();
1384
1385         if (!sc->may_unmap)
1386                 isolate_mode |= ISOLATE_UNMAPPED;
1387         if (!sc->may_writepage)
1388                 isolate_mode |= ISOLATE_CLEAN;
1389
1390         spin_lock_irq(&zone->lru_lock);
1391
1392         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1393                                      &nr_scanned, sc, isolate_mode, lru);
1394
1395         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1396         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1397
1398         if (global_reclaim(sc)) {
1399                 zone->pages_scanned += nr_scanned;
1400                 if (current_is_kswapd())
1401                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1402                 else
1403                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1404         }
1405         spin_unlock_irq(&zone->lru_lock);
1406
1407         if (nr_taken == 0)
1408                 return 0;
1409
1410         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1411                                 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1412                                 &nr_writeback, &nr_immediate,
1413                                 false);
1414
1415         spin_lock_irq(&zone->lru_lock);
1416
1417         reclaim_stat->recent_scanned[file] += nr_taken;
1418
1419         if (global_reclaim(sc)) {
1420                 if (current_is_kswapd())
1421                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1422                                                nr_reclaimed);
1423                 else
1424                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1425                                                nr_reclaimed);
1426         }
1427
1428         putback_inactive_pages(lruvec, &page_list);
1429
1430         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1431
1432         spin_unlock_irq(&zone->lru_lock);
1433
1434         free_hot_cold_page_list(&page_list, 1);
1435
1436         /*
1437          * If reclaim is isolating dirty pages under writeback, it implies
1438          * that the long-lived page allocation rate is exceeding the page
1439          * laundering rate. Either the global limits are not being effective
1440          * at throttling processes due to the page distribution throughout
1441          * zones or there is heavy usage of a slow backing device. The
1442          * only option is to throttle from reclaim context which is not ideal
1443          * as there is no guarantee the dirtying process is throttled in the
1444          * same way balance_dirty_pages() manages.
1445          *
1446          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1447          * of pages under pages flagged for immediate reclaim and stall if any
1448          * are encountered in the nr_immediate check below.
1449          */
1450         if (nr_writeback && nr_writeback == nr_taken)
1451                 zone_set_flag(zone, ZONE_WRITEBACK);
1452
1453         /*
1454          * memcg will stall in page writeback so only consider forcibly
1455          * stalling for global reclaim
1456          */
1457         if (global_reclaim(sc)) {
1458                 /*
1459                  * Tag a zone as congested if all the dirty pages scanned were
1460                  * backed by a congested BDI and wait_iff_congested will stall.
1461                  */
1462                 if (nr_dirty && nr_dirty == nr_congested)
1463                         zone_set_flag(zone, ZONE_CONGESTED);
1464
1465                 /*
1466                  * If dirty pages are scanned that are not queued for IO, it
1467                  * implies that flushers are not keeping up. In this case, flag
1468                  * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1469                  * pages from reclaim context. It will forcibly stall in the
1470                  * next check.
1471                  */
1472                 if (nr_unqueued_dirty == nr_taken)
1473                         zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1474
1475                 /*
1476                  * In addition, if kswapd scans pages marked marked for
1477                  * immediate reclaim and under writeback (nr_immediate), it
1478                  * implies that pages are cycling through the LRU faster than
1479                  * they are written so also forcibly stall.
1480                  */
1481                 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1482                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1483         }
1484
1485         /*
1486          * Stall direct reclaim for IO completions if underlying BDIs or zone
1487          * is congested. Allow kswapd to continue until it starts encountering
1488          * unqueued dirty pages or cycling through the LRU too quickly.
1489          */
1490         if (!sc->hibernation_mode && !current_is_kswapd())
1491                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1492
1493         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1494                 zone_idx(zone),
1495                 nr_scanned, nr_reclaimed,
1496                 sc->priority,
1497                 trace_shrink_flags(file));
1498         return nr_reclaimed;
1499 }
1500
1501 /*
1502  * This moves pages from the active list to the inactive list.
1503  *
1504  * We move them the other way if the page is referenced by one or more
1505  * processes, from rmap.
1506  *
1507  * If the pages are mostly unmapped, the processing is fast and it is
1508  * appropriate to hold zone->lru_lock across the whole operation.  But if
1509  * the pages are mapped, the processing is slow (page_referenced()) so we
1510  * should drop zone->lru_lock around each page.  It's impossible to balance
1511  * this, so instead we remove the pages from the LRU while processing them.
1512  * It is safe to rely on PG_active against the non-LRU pages in here because
1513  * nobody will play with that bit on a non-LRU page.
1514  *
1515  * The downside is that we have to touch page->_count against each page.
1516  * But we had to alter page->flags anyway.
1517  */
1518
1519 static void move_active_pages_to_lru(struct lruvec *lruvec,
1520                                      struct list_head *list,
1521                                      struct list_head *pages_to_free,
1522                                      enum lru_list lru)
1523 {
1524         struct zone *zone = lruvec_zone(lruvec);
1525         unsigned long pgmoved = 0;
1526         struct page *page;
1527         int nr_pages;
1528
1529         while (!list_empty(list)) {
1530                 page = lru_to_page(list);
1531                 lruvec = mem_cgroup_page_lruvec(page, zone);
1532
1533                 VM_BUG_ON(PageLRU(page));
1534                 SetPageLRU(page);
1535
1536                 nr_pages = hpage_nr_pages(page);
1537                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1538                 list_move(&page->lru, &lruvec->lists[lru]);
1539                 pgmoved += nr_pages;
1540
1541                 if (put_page_testzero(page)) {
1542                         __ClearPageLRU(page);
1543                         __ClearPageActive(page);
1544                         del_page_from_lru_list(page, lruvec, lru);
1545
1546                         if (unlikely(PageCompound(page))) {
1547                                 spin_unlock_irq(&zone->lru_lock);
1548                                 (*get_compound_page_dtor(page))(page);
1549                                 spin_lock_irq(&zone->lru_lock);
1550                         } else
1551                                 list_add(&page->lru, pages_to_free);
1552                 }
1553         }
1554         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1555         if (!is_active_lru(lru))
1556                 __count_vm_events(PGDEACTIVATE, pgmoved);
1557 }
1558
1559 static void shrink_active_list(unsigned long nr_to_scan,
1560                                struct lruvec *lruvec,
1561                                struct scan_control *sc,
1562                                enum lru_list lru)
1563 {
1564         unsigned long nr_taken;
1565         unsigned long nr_scanned;
1566         unsigned long vm_flags;
1567         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1568         LIST_HEAD(l_active);
1569         LIST_HEAD(l_inactive);
1570         struct page *page;
1571         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1572         unsigned long nr_rotated = 0;
1573         isolate_mode_t isolate_mode = 0;
1574         int file = is_file_lru(lru);
1575         struct zone *zone = lruvec_zone(lruvec);
1576
1577         lru_add_drain();
1578
1579         if (!sc->may_unmap)
1580                 isolate_mode |= ISOLATE_UNMAPPED;
1581         if (!sc->may_writepage)
1582                 isolate_mode |= ISOLATE_CLEAN;
1583
1584         spin_lock_irq(&zone->lru_lock);
1585
1586         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1587                                      &nr_scanned, sc, isolate_mode, lru);
1588         if (global_reclaim(sc))
1589                 zone->pages_scanned += nr_scanned;
1590
1591         reclaim_stat->recent_scanned[file] += nr_taken;
1592
1593         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1594         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1595         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1596         spin_unlock_irq(&zone->lru_lock);
1597
1598         while (!list_empty(&l_hold)) {
1599                 cond_resched();
1600                 page = lru_to_page(&l_hold);
1601                 list_del(&page->lru);
1602
1603                 if (unlikely(!page_evictable(page))) {
1604                         putback_lru_page(page);
1605                         continue;
1606                 }
1607
1608                 if (unlikely(buffer_heads_over_limit)) {
1609                         if (page_has_private(page) && trylock_page(page)) {
1610                                 if (page_has_private(page))
1611                                         try_to_release_page(page, 0);
1612                                 unlock_page(page);
1613                         }
1614                 }
1615
1616                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1617                                     &vm_flags)) {
1618                         nr_rotated += hpage_nr_pages(page);
1619                         /*
1620                          * Identify referenced, file-backed active pages and
1621                          * give them one more trip around the active list. So
1622                          * that executable code get better chances to stay in
1623                          * memory under moderate memory pressure.  Anon pages
1624                          * are not likely to be evicted by use-once streaming
1625                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1626                          * so we ignore them here.
1627                          */
1628                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1629                                 list_add(&page->lru, &l_active);
1630                                 continue;
1631                         }
1632                 }
1633
1634                 ClearPageActive(page);  /* we are de-activating */
1635                 list_add(&page->lru, &l_inactive);
1636         }
1637
1638         /*
1639          * Move pages back to the lru list.
1640          */
1641         spin_lock_irq(&zone->lru_lock);
1642         /*
1643          * Count referenced pages from currently used mappings as rotated,
1644          * even though only some of them are actually re-activated.  This
1645          * helps balance scan pressure between file and anonymous pages in
1646          * get_scan_ratio.
1647          */
1648         reclaim_stat->recent_rotated[file] += nr_rotated;
1649
1650         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1651         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1652         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1653         spin_unlock_irq(&zone->lru_lock);
1654
1655         free_hot_cold_page_list(&l_hold, 1);
1656 }
1657
1658 #ifdef CONFIG_SWAP
1659 static int inactive_anon_is_low_global(struct zone *zone)
1660 {
1661         unsigned long active, inactive;
1662
1663         active = zone_page_state(zone, NR_ACTIVE_ANON);
1664         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1665
1666         if (inactive * zone->inactive_ratio < active)
1667                 return 1;
1668
1669         return 0;
1670 }
1671
1672 /**
1673  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1674  * @lruvec: LRU vector to check
1675  *
1676  * Returns true if the zone does not have enough inactive anon pages,
1677  * meaning some active anon pages need to be deactivated.
1678  */
1679 static int inactive_anon_is_low(struct lruvec *lruvec)
1680 {
1681         /*
1682          * If we don't have swap space, anonymous page deactivation
1683          * is pointless.
1684          */
1685         if (!total_swap_pages)
1686                 return 0;
1687
1688         if (!mem_cgroup_disabled())
1689                 return mem_cgroup_inactive_anon_is_low(lruvec);
1690
1691         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1692 }
1693 #else
1694 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1695 {
1696         return 0;
1697 }
1698 #endif
1699
1700 /**
1701  * inactive_file_is_low - check if file pages need to be deactivated
1702  * @lruvec: LRU vector to check
1703  *
1704  * When the system is doing streaming IO, memory pressure here
1705  * ensures that active file pages get deactivated, until more
1706  * than half of the file pages are on the inactive list.
1707  *
1708  * Once we get to that situation, protect the system's working
1709  * set from being evicted by disabling active file page aging.
1710  *
1711  * This uses a different ratio than the anonymous pages, because
1712  * the page cache uses a use-once replacement algorithm.
1713  */
1714 static int inactive_file_is_low(struct lruvec *lruvec)
1715 {
1716         unsigned long inactive;
1717         unsigned long active;
1718
1719         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1720         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1721
1722         return active > inactive;
1723 }
1724
1725 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1726 {
1727         if (is_file_lru(lru))
1728                 return inactive_file_is_low(lruvec);
1729         else
1730                 return inactive_anon_is_low(lruvec);
1731 }
1732
1733 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1734                                  struct lruvec *lruvec, struct scan_control *sc)
1735 {
1736         if (is_active_lru(lru)) {
1737                 if (inactive_list_is_low(lruvec, lru))
1738                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1739                 return 0;
1740         }
1741
1742         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1743 }
1744
1745 static int vmscan_swappiness(struct scan_control *sc)
1746 {
1747         if (global_reclaim(sc))
1748                 return vm_swappiness;
1749         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1750 }
1751
1752 enum scan_balance {
1753         SCAN_EQUAL,
1754         SCAN_FRACT,
1755         SCAN_ANON,
1756         SCAN_FILE,
1757 };
1758
1759 /*
1760  * Determine how aggressively the anon and file LRU lists should be
1761  * scanned.  The relative value of each set of LRU lists is determined
1762  * by looking at the fraction of the pages scanned we did rotate back
1763  * onto the active list instead of evict.
1764  *
1765  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1766  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1767  */
1768 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1769                            unsigned long *nr)
1770 {
1771         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1772         u64 fraction[2];
1773         u64 denominator = 0;    /* gcc */
1774         struct zone *zone = lruvec_zone(lruvec);
1775         unsigned long anon_prio, file_prio;
1776         enum scan_balance scan_balance;
1777         unsigned long anon, file, free;
1778         bool force_scan = false;
1779         unsigned long ap, fp;
1780         enum lru_list lru;
1781
1782         /*
1783          * If the zone or memcg is small, nr[l] can be 0.  This
1784          * results in no scanning on this priority and a potential
1785          * priority drop.  Global direct reclaim can go to the next
1786          * zone and tends to have no problems. Global kswapd is for
1787          * zone balancing and it needs to scan a minimum amount. When
1788          * reclaiming for a memcg, a priority drop can cause high
1789          * latencies, so it's better to scan a minimum amount there as
1790          * well.
1791          */
1792         if (current_is_kswapd() && zone->all_unreclaimable)
1793                 force_scan = true;
1794         if (!global_reclaim(sc))
1795                 force_scan = true;
1796
1797         /* If we have no swap space, do not bother scanning anon pages. */
1798         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1799                 scan_balance = SCAN_FILE;
1800                 goto out;
1801         }
1802
1803         /*
1804          * Global reclaim will swap to prevent OOM even with no
1805          * swappiness, but memcg users want to use this knob to
1806          * disable swapping for individual groups completely when
1807          * using the memory controller's swap limit feature would be
1808          * too expensive.
1809          */
1810         if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1811                 scan_balance = SCAN_FILE;
1812                 goto out;
1813         }
1814
1815         /*
1816          * Do not apply any pressure balancing cleverness when the
1817          * system is close to OOM, scan both anon and file equally
1818          * (unless the swappiness setting disagrees with swapping).
1819          */
1820         if (!sc->priority && vmscan_swappiness(sc)) {
1821                 scan_balance = SCAN_EQUAL;
1822                 goto out;
1823         }
1824
1825         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1826                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1827         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1828                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1829
1830         /*
1831          * If it's foreseeable that reclaiming the file cache won't be
1832          * enough to get the zone back into a desirable shape, we have
1833          * to swap.  Better start now and leave the - probably heavily
1834          * thrashing - remaining file pages alone.
1835          */
1836         if (global_reclaim(sc)) {
1837                 free = zone_page_state(zone, NR_FREE_PAGES);
1838                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1839                         scan_balance = SCAN_ANON;
1840                         goto out;
1841                 }
1842         }
1843
1844         /*
1845          * There is enough inactive page cache, do not reclaim
1846          * anything from the anonymous working set right now.
1847          */
1848         if (!inactive_file_is_low(lruvec)) {
1849                 scan_balance = SCAN_FILE;
1850                 goto out;
1851         }
1852
1853         scan_balance = SCAN_FRACT;
1854
1855         /*
1856          * With swappiness at 100, anonymous and file have the same priority.
1857          * This scanning priority is essentially the inverse of IO cost.
1858          */
1859         anon_prio = vmscan_swappiness(sc);
1860         file_prio = 200 - anon_prio;
1861
1862         /*
1863          * OK, so we have swap space and a fair amount of page cache
1864          * pages.  We use the recently rotated / recently scanned
1865          * ratios to determine how valuable each cache is.
1866          *
1867          * Because workloads change over time (and to avoid overflow)
1868          * we keep these statistics as a floating average, which ends
1869          * up weighing recent references more than old ones.
1870          *
1871          * anon in [0], file in [1]
1872          */
1873         spin_lock_irq(&zone->lru_lock);
1874         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1875                 reclaim_stat->recent_scanned[0] /= 2;
1876                 reclaim_stat->recent_rotated[0] /= 2;
1877         }
1878
1879         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1880                 reclaim_stat->recent_scanned[1] /= 2;
1881                 reclaim_stat->recent_rotated[1] /= 2;
1882         }
1883
1884         /*
1885          * The amount of pressure on anon vs file pages is inversely
1886          * proportional to the fraction of recently scanned pages on
1887          * each list that were recently referenced and in active use.
1888          */
1889         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1890         ap /= reclaim_stat->recent_rotated[0] + 1;
1891
1892         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1893         fp /= reclaim_stat->recent_rotated[1] + 1;
1894         spin_unlock_irq(&zone->lru_lock);
1895
1896         fraction[0] = ap;
1897         fraction[1] = fp;
1898         denominator = ap + fp + 1;
1899 out:
1900         for_each_evictable_lru(lru) {
1901                 int file = is_file_lru(lru);
1902                 unsigned long size;
1903                 unsigned long scan;
1904
1905                 size = get_lru_size(lruvec, lru);
1906                 scan = size >> sc->priority;
1907
1908                 if (!scan && force_scan)
1909                         scan = min(size, SWAP_CLUSTER_MAX);
1910
1911                 switch (scan_balance) {
1912                 case SCAN_EQUAL:
1913                         /* Scan lists relative to size */
1914                         break;
1915                 case SCAN_FRACT:
1916                         /*
1917                          * Scan types proportional to swappiness and
1918                          * their relative recent reclaim efficiency.
1919                          */
1920                         scan = div64_u64(scan * fraction[file], denominator);
1921                         break;
1922                 case SCAN_FILE:
1923                 case SCAN_ANON:
1924                         /* Scan one type exclusively */
1925                         if ((scan_balance == SCAN_FILE) != file)
1926                                 scan = 0;
1927                         break;
1928                 default:
1929                         /* Look ma, no brain */
1930                         BUG();
1931                 }
1932                 nr[lru] = scan;
1933         }
1934 }
1935
1936 /*
1937  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1938  */
1939 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1940 {
1941         unsigned long nr[NR_LRU_LISTS];
1942         unsigned long targets[NR_LRU_LISTS];
1943         unsigned long nr_to_scan;
1944         enum lru_list lru;
1945         unsigned long nr_reclaimed = 0;
1946         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1947         struct blk_plug plug;
1948         bool scan_adjusted = false;
1949
1950         get_scan_count(lruvec, sc, nr);
1951
1952         /* Record the original scan target for proportional adjustments later */
1953         memcpy(targets, nr, sizeof(nr));
1954
1955         blk_start_plug(&plug);
1956         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1957                                         nr[LRU_INACTIVE_FILE]) {
1958                 unsigned long nr_anon, nr_file, percentage;
1959                 unsigned long nr_scanned;
1960
1961                 for_each_evictable_lru(lru) {
1962                         if (nr[lru]) {
1963                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1964                                 nr[lru] -= nr_to_scan;
1965
1966                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1967                                                             lruvec, sc);
1968                         }
1969                 }
1970
1971                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
1972                         continue;
1973
1974                 /*
1975                  * For global direct reclaim, reclaim only the number of pages
1976                  * requested. Less care is taken to scan proportionally as it
1977                  * is more important to minimise direct reclaim stall latency
1978                  * than it is to properly age the LRU lists.
1979                  */
1980                 if (global_reclaim(sc) && !current_is_kswapd())
1981                         break;
1982
1983                 /*
1984                  * For kswapd and memcg, reclaim at least the number of pages
1985                  * requested. Ensure that the anon and file LRUs shrink
1986                  * proportionally what was requested by get_scan_count(). We
1987                  * stop reclaiming one LRU and reduce the amount scanning
1988                  * proportional to the original scan target.
1989                  */
1990                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
1991                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
1992
1993                 if (nr_file > nr_anon) {
1994                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
1995                                                 targets[LRU_ACTIVE_ANON] + 1;
1996                         lru = LRU_BASE;
1997                         percentage = nr_anon * 100 / scan_target;
1998                 } else {
1999                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2000                                                 targets[LRU_ACTIVE_FILE] + 1;
2001                         lru = LRU_FILE;
2002                         percentage = nr_file * 100 / scan_target;
2003                 }
2004
2005                 /* Stop scanning the smaller of the LRU */
2006                 nr[lru] = 0;
2007                 nr[lru + LRU_ACTIVE] = 0;
2008
2009                 /*
2010                  * Recalculate the other LRU scan count based on its original
2011                  * scan target and the percentage scanning already complete
2012                  */
2013                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2014                 nr_scanned = targets[lru] - nr[lru];
2015                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2016                 nr[lru] -= min(nr[lru], nr_scanned);
2017
2018                 lru += LRU_ACTIVE;
2019                 nr_scanned = targets[lru] - nr[lru];
2020                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2021                 nr[lru] -= min(nr[lru], nr_scanned);
2022
2023                 scan_adjusted = true;
2024         }
2025         blk_finish_plug(&plug);
2026         sc->nr_reclaimed += nr_reclaimed;
2027
2028         /*
2029          * Even if we did not try to evict anon pages at all, we want to
2030          * rebalance the anon lru active/inactive ratio.
2031          */
2032         if (inactive_anon_is_low(lruvec))
2033                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2034                                    sc, LRU_ACTIVE_ANON);
2035
2036         throttle_vm_writeout(sc->gfp_mask);
2037 }
2038
2039 /* Use reclaim/compaction for costly allocs or under memory pressure */
2040 static bool in_reclaim_compaction(struct scan_control *sc)
2041 {
2042         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2043                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2044                          sc->priority < DEF_PRIORITY - 2))
2045                 return true;
2046
2047         return false;
2048 }
2049
2050 /*
2051  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2052  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2053  * true if more pages should be reclaimed such that when the page allocator
2054  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2055  * It will give up earlier than that if there is difficulty reclaiming pages.
2056  */
2057 static inline bool should_continue_reclaim(struct zone *zone,
2058                                         unsigned long nr_reclaimed,
2059                                         unsigned long nr_scanned,
2060                                         struct scan_control *sc)
2061 {
2062         unsigned long pages_for_compaction;
2063         unsigned long inactive_lru_pages;
2064
2065         /* If not in reclaim/compaction mode, stop */
2066         if (!in_reclaim_compaction(sc))
2067                 return false;
2068
2069         /* Consider stopping depending on scan and reclaim activity */
2070         if (sc->gfp_mask & __GFP_REPEAT) {
2071                 /*
2072                  * For __GFP_REPEAT allocations, stop reclaiming if the
2073                  * full LRU list has been scanned and we are still failing
2074                  * to reclaim pages. This full LRU scan is potentially
2075                  * expensive but a __GFP_REPEAT caller really wants to succeed
2076                  */
2077                 if (!nr_reclaimed && !nr_scanned)
2078                         return false;
2079         } else {
2080                 /*
2081                  * For non-__GFP_REPEAT allocations which can presumably
2082                  * fail without consequence, stop if we failed to reclaim
2083                  * any pages from the last SWAP_CLUSTER_MAX number of
2084                  * pages that were scanned. This will return to the
2085                  * caller faster at the risk reclaim/compaction and
2086                  * the resulting allocation attempt fails
2087                  */
2088                 if (!nr_reclaimed)
2089                         return false;
2090         }
2091
2092         /*
2093          * If we have not reclaimed enough pages for compaction and the
2094          * inactive lists are large enough, continue reclaiming
2095          */
2096         pages_for_compaction = (2UL << sc->order);
2097         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2098         if (get_nr_swap_pages() > 0)
2099                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2100         if (sc->nr_reclaimed < pages_for_compaction &&
2101                         inactive_lru_pages > pages_for_compaction)
2102                 return true;
2103
2104         /* If compaction would go ahead or the allocation would succeed, stop */
2105         switch (compaction_suitable(zone, sc->order)) {
2106         case COMPACT_PARTIAL:
2107         case COMPACT_CONTINUE:
2108                 return false;
2109         default:
2110                 return true;
2111         }
2112 }
2113
2114 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2115 {
2116         unsigned long nr_reclaimed, nr_scanned;
2117
2118         do {
2119                 struct mem_cgroup *root = sc->target_mem_cgroup;
2120                 struct mem_cgroup_reclaim_cookie reclaim = {
2121                         .zone = zone,
2122                         .priority = sc->priority,
2123                 };
2124                 struct mem_cgroup *memcg;
2125
2126                 nr_reclaimed = sc->nr_reclaimed;
2127                 nr_scanned = sc->nr_scanned;
2128
2129                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2130                 do {
2131                         struct lruvec *lruvec;
2132
2133                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2134
2135                         shrink_lruvec(lruvec, sc);
2136
2137                         /*
2138                          * Direct reclaim and kswapd have to scan all memory
2139                          * cgroups to fulfill the overall scan target for the
2140                          * zone.
2141                          *
2142                          * Limit reclaim, on the other hand, only cares about
2143                          * nr_to_reclaim pages to be reclaimed and it will
2144                          * retry with decreasing priority if one round over the
2145                          * whole hierarchy is not sufficient.
2146                          */
2147                         if (!global_reclaim(sc) &&
2148                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2149                                 mem_cgroup_iter_break(root, memcg);
2150                                 break;
2151                         }
2152                         memcg = mem_cgroup_iter(root, memcg, &reclaim);
2153                 } while (memcg);
2154
2155                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2156                            sc->nr_scanned - nr_scanned,
2157                            sc->nr_reclaimed - nr_reclaimed);
2158
2159         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2160                                          sc->nr_scanned - nr_scanned, sc));
2161 }
2162
2163 /* Returns true if compaction should go ahead for a high-order request */
2164 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2165 {
2166         unsigned long balance_gap, watermark;
2167         bool watermark_ok;
2168
2169         /* Do not consider compaction for orders reclaim is meant to satisfy */
2170         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2171                 return false;
2172
2173         /*
2174          * Compaction takes time to run and there are potentially other
2175          * callers using the pages just freed. Continue reclaiming until
2176          * there is a buffer of free pages available to give compaction
2177          * a reasonable chance of completing and allocating the page
2178          */
2179         balance_gap = min(low_wmark_pages(zone),
2180                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2181                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2182         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2183         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2184
2185         /*
2186          * If compaction is deferred, reclaim up to a point where
2187          * compaction will have a chance of success when re-enabled
2188          */
2189         if (compaction_deferred(zone, sc->order))
2190                 return watermark_ok;
2191
2192         /* If compaction is not ready to start, keep reclaiming */
2193         if (!compaction_suitable(zone, sc->order))
2194                 return false;
2195
2196         return watermark_ok;
2197 }
2198
2199 /*
2200  * This is the direct reclaim path, for page-allocating processes.  We only
2201  * try to reclaim pages from zones which will satisfy the caller's allocation
2202  * request.
2203  *
2204  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2205  * Because:
2206  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2207  *    allocation or
2208  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2209  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2210  *    zone defense algorithm.
2211  *
2212  * If a zone is deemed to be full of pinned pages then just give it a light
2213  * scan then give up on it.
2214  *
2215  * This function returns true if a zone is being reclaimed for a costly
2216  * high-order allocation and compaction is ready to begin. This indicates to
2217  * the caller that it should consider retrying the allocation instead of
2218  * further reclaim.
2219  */
2220 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2221 {
2222         struct zoneref *z;
2223         struct zone *zone;
2224         unsigned long nr_soft_reclaimed;
2225         unsigned long nr_soft_scanned;
2226         bool aborted_reclaim = false;
2227
2228         /*
2229          * If the number of buffer_heads in the machine exceeds the maximum
2230          * allowed level, force direct reclaim to scan the highmem zone as
2231          * highmem pages could be pinning lowmem pages storing buffer_heads
2232          */
2233         if (buffer_heads_over_limit)
2234                 sc->gfp_mask |= __GFP_HIGHMEM;
2235
2236         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2237                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2238                 if (!populated_zone(zone))
2239                         continue;
2240                 /*
2241                  * Take care memory controller reclaiming has small influence
2242                  * to global LRU.
2243                  */
2244                 if (global_reclaim(sc)) {
2245                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2246                                 continue;
2247                         if (zone->all_unreclaimable &&
2248                                         sc->priority != DEF_PRIORITY)
2249                                 continue;       /* Let kswapd poll it */
2250                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2251                                 /*
2252                                  * If we already have plenty of memory free for
2253                                  * compaction in this zone, don't free any more.
2254                                  * Even though compaction is invoked for any
2255                                  * non-zero order, only frequent costly order
2256                                  * reclamation is disruptive enough to become a
2257                                  * noticeable problem, like transparent huge
2258                                  * page allocations.
2259                                  */
2260                                 if (compaction_ready(zone, sc)) {
2261                                         aborted_reclaim = true;
2262                                         continue;
2263                                 }
2264                         }
2265                         /*
2266                          * This steals pages from memory cgroups over softlimit
2267                          * and returns the number of reclaimed pages and
2268                          * scanned pages. This works for global memory pressure
2269                          * and balancing, not for a memcg's limit.
2270                          */
2271                         nr_soft_scanned = 0;
2272                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2273                                                 sc->order, sc->gfp_mask,
2274                                                 &nr_soft_scanned);
2275                         sc->nr_reclaimed += nr_soft_reclaimed;
2276                         sc->nr_scanned += nr_soft_scanned;
2277                         /* need some check for avoid more shrink_zone() */
2278                 }
2279
2280                 shrink_zone(zone, sc);
2281         }
2282
2283         return aborted_reclaim;
2284 }
2285
2286 static bool zone_reclaimable(struct zone *zone)
2287 {
2288         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2289 }
2290
2291 /* All zones in zonelist are unreclaimable? */
2292 static bool all_unreclaimable(struct zonelist *zonelist,
2293                 struct scan_control *sc)
2294 {
2295         struct zoneref *z;
2296         struct zone *zone;
2297
2298         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2299                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2300                 if (!populated_zone(zone))
2301                         continue;
2302                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2303                         continue;
2304                 if (!zone->all_unreclaimable)
2305                         return false;
2306         }
2307
2308         return true;
2309 }
2310
2311 /*
2312  * This is the main entry point to direct page reclaim.
2313  *
2314  * If a full scan of the inactive list fails to free enough memory then we
2315  * are "out of memory" and something needs to be killed.
2316  *
2317  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2318  * high - the zone may be full of dirty or under-writeback pages, which this
2319  * caller can't do much about.  We kick the writeback threads and take explicit
2320  * naps in the hope that some of these pages can be written.  But if the
2321  * allocating task holds filesystem locks which prevent writeout this might not
2322  * work, and the allocation attempt will fail.
2323  *
2324  * returns:     0, if no pages reclaimed
2325  *              else, the number of pages reclaimed
2326  */
2327 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2328                                         struct scan_control *sc,
2329                                         struct shrink_control *shrink)
2330 {
2331         unsigned long total_scanned = 0;
2332         struct reclaim_state *reclaim_state = current->reclaim_state;
2333         struct zoneref *z;
2334         struct zone *zone;
2335         unsigned long writeback_threshold;
2336         bool aborted_reclaim;
2337
2338         delayacct_freepages_start();
2339
2340         if (global_reclaim(sc))
2341                 count_vm_event(ALLOCSTALL);
2342
2343         do {
2344                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2345                                 sc->priority);
2346                 sc->nr_scanned = 0;
2347                 aborted_reclaim = shrink_zones(zonelist, sc);
2348
2349                 /*
2350                  * Don't shrink slabs when reclaiming memory from over limit
2351                  * cgroups but do shrink slab at least once when aborting
2352                  * reclaim for compaction to avoid unevenly scanning file/anon
2353                  * LRU pages over slab pages.
2354                  */
2355                 if (global_reclaim(sc)) {
2356                         unsigned long lru_pages = 0;
2357                         for_each_zone_zonelist(zone, z, zonelist,
2358                                         gfp_zone(sc->gfp_mask)) {
2359                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2360                                         continue;
2361
2362                                 lru_pages += zone_reclaimable_pages(zone);
2363                         }
2364
2365                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2366                         if (reclaim_state) {
2367                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2368                                 reclaim_state->reclaimed_slab = 0;
2369                         }
2370                 }
2371                 total_scanned += sc->nr_scanned;
2372                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2373                         goto out;
2374
2375                 /*
2376                  * If we're getting trouble reclaiming, start doing
2377                  * writepage even in laptop mode.
2378                  */
2379                 if (sc->priority < DEF_PRIORITY - 2)
2380                         sc->may_writepage = 1;
2381
2382                 /*
2383                  * Try to write back as many pages as we just scanned.  This
2384                  * tends to cause slow streaming writers to write data to the
2385                  * disk smoothly, at the dirtying rate, which is nice.   But
2386                  * that's undesirable in laptop mode, where we *want* lumpy
2387                  * writeout.  So in laptop mode, write out the whole world.
2388                  */
2389                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2390                 if (total_scanned > writeback_threshold) {
2391                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2392                                                 WB_REASON_TRY_TO_FREE_PAGES);
2393                         sc->may_writepage = 1;
2394                 }
2395         } while (--sc->priority >= 0 && !aborted_reclaim);
2396
2397 out:
2398         delayacct_freepages_end();
2399
2400         if (sc->nr_reclaimed)
2401                 return sc->nr_reclaimed;
2402
2403         /*
2404          * As hibernation is going on, kswapd is freezed so that it can't mark
2405          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2406          * check.
2407          */
2408         if (oom_killer_disabled)
2409                 return 0;
2410
2411         /* Aborted reclaim to try compaction? don't OOM, then */
2412         if (aborted_reclaim)
2413                 return 1;
2414
2415         /* top priority shrink_zones still had more to do? don't OOM, then */
2416         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2417                 return 1;
2418
2419         return 0;
2420 }
2421
2422 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2423 {
2424         struct zone *zone;
2425         unsigned long pfmemalloc_reserve = 0;
2426         unsigned long free_pages = 0;
2427         int i;
2428         bool wmark_ok;
2429
2430         for (i = 0; i <= ZONE_NORMAL; i++) {
2431                 zone = &pgdat->node_zones[i];
2432                 pfmemalloc_reserve += min_wmark_pages(zone);
2433                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2434         }
2435
2436         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2437
2438         /* kswapd must be awake if processes are being throttled */
2439         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2440                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2441                                                 (enum zone_type)ZONE_NORMAL);
2442                 wake_up_interruptible(&pgdat->kswapd_wait);
2443         }
2444
2445         return wmark_ok;
2446 }
2447
2448 /*
2449  * Throttle direct reclaimers if backing storage is backed by the network
2450  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2451  * depleted. kswapd will continue to make progress and wake the processes
2452  * when the low watermark is reached.
2453  *
2454  * Returns true if a fatal signal was delivered during throttling. If this
2455  * happens, the page allocator should not consider triggering the OOM killer.
2456  */
2457 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2458                                         nodemask_t *nodemask)
2459 {
2460         struct zone *zone;
2461         int high_zoneidx = gfp_zone(gfp_mask);
2462         pg_data_t *pgdat;
2463
2464         /*
2465          * Kernel threads should not be throttled as they may be indirectly
2466          * responsible for cleaning pages necessary for reclaim to make forward
2467          * progress. kjournald for example may enter direct reclaim while
2468          * committing a transaction where throttling it could forcing other
2469          * processes to block on log_wait_commit().
2470          */
2471         if (current->flags & PF_KTHREAD)
2472                 goto out;
2473
2474         /*
2475          * If a fatal signal is pending, this process should not throttle.
2476          * It should return quickly so it can exit and free its memory
2477          */
2478         if (fatal_signal_pending(current))
2479                 goto out;
2480
2481         /* Check if the pfmemalloc reserves are ok */
2482         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2483         pgdat = zone->zone_pgdat;
2484         if (pfmemalloc_watermark_ok(pgdat))
2485                 goto out;
2486
2487         /* Account for the throttling */
2488         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2489
2490         /*
2491          * If the caller cannot enter the filesystem, it's possible that it
2492          * is due to the caller holding an FS lock or performing a journal
2493          * transaction in the case of a filesystem like ext[3|4]. In this case,
2494          * it is not safe to block on pfmemalloc_wait as kswapd could be
2495          * blocked waiting on the same lock. Instead, throttle for up to a
2496          * second before continuing.
2497          */
2498         if (!(gfp_mask & __GFP_FS)) {
2499                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2500                         pfmemalloc_watermark_ok(pgdat), HZ);
2501
2502                 goto check_pending;
2503         }
2504
2505         /* Throttle until kswapd wakes the process */
2506         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2507                 pfmemalloc_watermark_ok(pgdat));
2508
2509 check_pending:
2510         if (fatal_signal_pending(current))
2511                 return true;
2512
2513 out:
2514         return false;
2515 }
2516
2517 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2518                                 gfp_t gfp_mask, nodemask_t *nodemask)
2519 {
2520         unsigned long nr_reclaimed;
2521         struct scan_control sc = {
2522                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2523                 .may_writepage = !laptop_mode,
2524                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2525                 .may_unmap = 1,
2526                 .may_swap = 1,
2527                 .order = order,
2528                 .priority = DEF_PRIORITY,
2529                 .target_mem_cgroup = NULL,
2530                 .nodemask = nodemask,
2531         };
2532         struct shrink_control shrink = {
2533                 .gfp_mask = sc.gfp_mask,
2534         };
2535
2536         /*
2537          * Do not enter reclaim if fatal signal was delivered while throttled.
2538          * 1 is returned so that the page allocator does not OOM kill at this
2539          * point.
2540          */
2541         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2542                 return 1;
2543
2544         trace_mm_vmscan_direct_reclaim_begin(order,
2545                                 sc.may_writepage,
2546                                 gfp_mask);
2547
2548         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2549
2550         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2551
2552         return nr_reclaimed;
2553 }
2554
2555 #ifdef CONFIG_MEMCG
2556
2557 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2558                                                 gfp_t gfp_mask, bool noswap,
2559                                                 struct zone *zone,
2560                                                 unsigned long *nr_scanned)
2561 {
2562         struct scan_control sc = {
2563                 .nr_scanned = 0,
2564                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2565                 .may_writepage = !laptop_mode,
2566                 .may_unmap = 1,
2567                 .may_swap = !noswap,
2568                 .order = 0,
2569                 .priority = 0,
2570                 .target_mem_cgroup = memcg,
2571         };
2572         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2573
2574         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2575                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2576
2577         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2578                                                       sc.may_writepage,
2579                                                       sc.gfp_mask);
2580
2581         /*
2582          * NOTE: Although we can get the priority field, using it
2583          * here is not a good idea, since it limits the pages we can scan.
2584          * if we don't reclaim here, the shrink_zone from balance_pgdat
2585          * will pick up pages from other mem cgroup's as well. We hack
2586          * the priority and make it zero.
2587          */
2588         shrink_lruvec(lruvec, &sc);
2589
2590         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2591
2592         *nr_scanned = sc.nr_scanned;
2593         return sc.nr_reclaimed;
2594 }
2595
2596 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2597                                            gfp_t gfp_mask,
2598                                            bool noswap)
2599 {
2600         struct zonelist *zonelist;
2601         unsigned long nr_reclaimed;
2602         int nid;
2603         struct scan_control sc = {
2604                 .may_writepage = !laptop_mode,
2605                 .may_unmap = 1,
2606                 .may_swap = !noswap,
2607                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2608                 .order = 0,
2609                 .priority = DEF_PRIORITY,
2610                 .target_mem_cgroup = memcg,
2611                 .nodemask = NULL, /* we don't care the placement */
2612                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2613                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2614         };
2615         struct shrink_control shrink = {
2616                 .gfp_mask = sc.gfp_mask,
2617         };
2618
2619         /*
2620          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2621          * take care of from where we get pages. So the node where we start the
2622          * scan does not need to be the current node.
2623          */
2624         nid = mem_cgroup_select_victim_node(memcg);
2625
2626         zonelist = NODE_DATA(nid)->node_zonelists;
2627
2628         trace_mm_vmscan_memcg_reclaim_begin(0,
2629                                             sc.may_writepage,
2630                                             sc.gfp_mask);
2631
2632         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2633
2634         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2635
2636         return nr_reclaimed;
2637 }
2638 #endif
2639
2640 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2641 {
2642         struct mem_cgroup *memcg;
2643
2644         if (!total_swap_pages)
2645                 return;
2646
2647         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2648         do {
2649                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2650
2651                 if (inactive_anon_is_low(lruvec))
2652                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2653                                            sc, LRU_ACTIVE_ANON);
2654
2655                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2656         } while (memcg);
2657 }
2658
2659 static bool zone_balanced(struct zone *zone, int order,
2660                           unsigned long balance_gap, int classzone_idx)
2661 {
2662         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2663                                     balance_gap, classzone_idx, 0))
2664                 return false;
2665
2666         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2667             !compaction_suitable(zone, order))
2668                 return false;
2669
2670         return true;
2671 }
2672
2673 /*
2674  * pgdat_balanced() is used when checking if a node is balanced.
2675  *
2676  * For order-0, all zones must be balanced!
2677  *
2678  * For high-order allocations only zones that meet watermarks and are in a
2679  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2680  * total of balanced pages must be at least 25% of the zones allowed by
2681  * classzone_idx for the node to be considered balanced. Forcing all zones to
2682  * be balanced for high orders can cause excessive reclaim when there are
2683  * imbalanced zones.
2684  * The choice of 25% is due to
2685  *   o a 16M DMA zone that is balanced will not balance a zone on any
2686  *     reasonable sized machine
2687  *   o On all other machines, the top zone must be at least a reasonable
2688  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2689  *     would need to be at least 256M for it to be balance a whole node.
2690  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2691  *     to balance a node on its own. These seemed like reasonable ratios.
2692  */
2693 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2694 {
2695         unsigned long managed_pages = 0;
2696         unsigned long balanced_pages = 0;
2697         int i;
2698
2699         /* Check the watermark levels */
2700         for (i = 0; i <= classzone_idx; i++) {
2701                 struct zone *zone = pgdat->node_zones + i;
2702
2703                 if (!populated_zone(zone))
2704                         continue;
2705
2706                 managed_pages += zone->managed_pages;
2707
2708                 /*
2709                  * A special case here:
2710                  *
2711                  * balance_pgdat() skips over all_unreclaimable after
2712                  * DEF_PRIORITY. Effectively, it considers them balanced so
2713                  * they must be considered balanced here as well!
2714                  */
2715                 if (zone->all_unreclaimable) {
2716                         balanced_pages += zone->managed_pages;
2717                         continue;
2718                 }
2719
2720                 if (zone_balanced(zone, order, 0, i))
2721                         balanced_pages += zone->managed_pages;
2722                 else if (!order)
2723                         return false;
2724         }
2725
2726         if (order)
2727                 return balanced_pages >= (managed_pages >> 2);
2728         else
2729                 return true;
2730 }
2731
2732 /*
2733  * Prepare kswapd for sleeping. This verifies that there are no processes
2734  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2735  *
2736  * Returns true if kswapd is ready to sleep
2737  */
2738 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2739                                         int classzone_idx)
2740 {
2741         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2742         if (remaining)
2743                 return false;
2744
2745         /*
2746          * There is a potential race between when kswapd checks its watermarks
2747          * and a process gets throttled. There is also a potential race if
2748          * processes get throttled, kswapd wakes, a large process exits therby
2749          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2750          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2751          * so wake them now if necessary. If necessary, processes will wake
2752          * kswapd and get throttled again
2753          */
2754         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2755                 wake_up(&pgdat->pfmemalloc_wait);
2756                 return false;
2757         }
2758
2759         return pgdat_balanced(pgdat, order, classzone_idx);
2760 }
2761
2762 /*
2763  * kswapd shrinks the zone by the number of pages required to reach
2764  * the high watermark.
2765  *
2766  * Returns true if kswapd scanned at least the requested number of pages to
2767  * reclaim or if the lack of progress was due to pages under writeback.
2768  * This is used to determine if the scanning priority needs to be raised.
2769  */
2770 static bool kswapd_shrink_zone(struct zone *zone,
2771                                int classzone_idx,
2772                                struct scan_control *sc,
2773                                unsigned long lru_pages,
2774                                unsigned long *nr_attempted)
2775 {
2776         unsigned long nr_slab;
2777         int testorder = sc->order;
2778         unsigned long balance_gap;
2779         struct reclaim_state *reclaim_state = current->reclaim_state;
2780         struct shrink_control shrink = {
2781                 .gfp_mask = sc->gfp_mask,
2782         };
2783         bool lowmem_pressure;
2784
2785         /* Reclaim above the high watermark. */
2786         sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2787
2788         /*
2789          * Kswapd reclaims only single pages with compaction enabled. Trying
2790          * too hard to reclaim until contiguous free pages have become
2791          * available can hurt performance by evicting too much useful data
2792          * from memory. Do not reclaim more than needed for compaction.
2793          */
2794         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2795                         compaction_suitable(zone, sc->order) !=
2796                                 COMPACT_SKIPPED)
2797                 testorder = 0;
2798
2799         /*
2800          * We put equal pressure on every zone, unless one zone has way too
2801          * many pages free already. The "too many pages" is defined as the
2802          * high wmark plus a "gap" where the gap is either the low
2803          * watermark or 1% of the zone, whichever is smaller.
2804          */
2805         balance_gap = min(low_wmark_pages(zone),
2806                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2807                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2808
2809         /*
2810          * If there is no low memory pressure or the zone is balanced then no
2811          * reclaim is necessary
2812          */
2813         lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2814         if (!lowmem_pressure && zone_balanced(zone, testorder,
2815                                                 balance_gap, classzone_idx))
2816                 return true;
2817
2818         shrink_zone(zone, sc);
2819
2820         reclaim_state->reclaimed_slab = 0;
2821         nr_slab = shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2822         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823
2824         /* Account for the number of pages attempted to reclaim */
2825         *nr_attempted += sc->nr_to_reclaim;
2826
2827         if (nr_slab == 0 && !zone_reclaimable(zone))
2828                 zone->all_unreclaimable = 1;
2829
2830         zone_clear_flag(zone, ZONE_WRITEBACK);
2831
2832         /*
2833          * If a zone reaches its high watermark, consider it to be no longer
2834          * congested. It's possible there are dirty pages backed by congested
2835          * BDIs but as pressure is relieved, speculatively avoid congestion
2836          * waits.
2837          */
2838         if (!zone->all_unreclaimable &&
2839             zone_balanced(zone, testorder, 0, classzone_idx)) {
2840                 zone_clear_flag(zone, ZONE_CONGESTED);
2841                 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2842         }
2843
2844         return sc->nr_scanned >= sc->nr_to_reclaim;
2845 }
2846
2847 /*
2848  * For kswapd, balance_pgdat() will work across all this node's zones until
2849  * they are all at high_wmark_pages(zone).
2850  *
2851  * Returns the final order kswapd was reclaiming at
2852  *
2853  * There is special handling here for zones which are full of pinned pages.
2854  * This can happen if the pages are all mlocked, or if they are all used by
2855  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2856  * What we do is to detect the case where all pages in the zone have been
2857  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2858  * dead and from now on, only perform a short scan.  Basically we're polling
2859  * the zone for when the problem goes away.
2860  *
2861  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2862  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2863  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2864  * lower zones regardless of the number of free pages in the lower zones. This
2865  * interoperates with the page allocator fallback scheme to ensure that aging
2866  * of pages is balanced across the zones.
2867  */
2868 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2869                                                         int *classzone_idx)
2870 {
2871         int i;
2872         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2873         unsigned long nr_soft_reclaimed;
2874         unsigned long nr_soft_scanned;
2875         struct scan_control sc = {
2876                 .gfp_mask = GFP_KERNEL,
2877                 .priority = DEF_PRIORITY,
2878                 .may_unmap = 1,
2879                 .may_swap = 1,
2880                 .may_writepage = !laptop_mode,
2881                 .order = order,
2882                 .target_mem_cgroup = NULL,
2883         };
2884         count_vm_event(PAGEOUTRUN);
2885
2886         do {
2887                 unsigned long lru_pages = 0;
2888                 unsigned long nr_attempted = 0;
2889                 bool raise_priority = true;
2890                 bool pgdat_needs_compaction = (order > 0);
2891
2892                 sc.nr_reclaimed = 0;
2893
2894                 /*
2895                  * Scan in the highmem->dma direction for the highest
2896                  * zone which needs scanning
2897                  */
2898                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2899                         struct zone *zone = pgdat->node_zones + i;
2900
2901                         if (!populated_zone(zone))
2902                                 continue;
2903
2904                         if (zone->all_unreclaimable &&
2905                             sc.priority != DEF_PRIORITY)
2906                                 continue;
2907
2908                         /*
2909                          * Do some background aging of the anon list, to give
2910                          * pages a chance to be referenced before reclaiming.
2911                          */
2912                         age_active_anon(zone, &sc);
2913
2914                         /*
2915                          * If the number of buffer_heads in the machine
2916                          * exceeds the maximum allowed level and this node
2917                          * has a highmem zone, force kswapd to reclaim from
2918                          * it to relieve lowmem pressure.
2919                          */
2920                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2921                                 end_zone = i;
2922                                 break;
2923                         }
2924
2925                         if (!zone_balanced(zone, order, 0, 0)) {
2926                                 end_zone = i;
2927                                 break;
2928                         } else {
2929                                 /*
2930                                  * If balanced, clear the dirty and congested
2931                                  * flags
2932                                  */
2933                                 zone_clear_flag(zone, ZONE_CONGESTED);
2934                                 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2935                         }
2936                 }
2937
2938                 if (i < 0)
2939                         goto out;
2940
2941                 for (i = 0; i <= end_zone; i++) {
2942                         struct zone *zone = pgdat->node_zones + i;
2943
2944                         if (!populated_zone(zone))
2945                                 continue;
2946
2947                         lru_pages += zone_reclaimable_pages(zone);
2948
2949                         /*
2950                          * If any zone is currently balanced then kswapd will
2951                          * not call compaction as it is expected that the
2952                          * necessary pages are already available.
2953                          */
2954                         if (pgdat_needs_compaction &&
2955                                         zone_watermark_ok(zone, order,
2956                                                 low_wmark_pages(zone),
2957                                                 *classzone_idx, 0))
2958                                 pgdat_needs_compaction = false;
2959                 }
2960
2961                 /*
2962                  * If we're getting trouble reclaiming, start doing writepage
2963                  * even in laptop mode.
2964                  */
2965                 if (sc.priority < DEF_PRIORITY - 2)
2966                         sc.may_writepage = 1;
2967
2968                 /*
2969                  * Now scan the zone in the dma->highmem direction, stopping
2970                  * at the last zone which needs scanning.
2971                  *
2972                  * We do this because the page allocator works in the opposite
2973                  * direction.  This prevents the page allocator from allocating
2974                  * pages behind kswapd's direction of progress, which would
2975                  * cause too much scanning of the lower zones.
2976                  */
2977                 for (i = 0; i <= end_zone; i++) {
2978                         struct zone *zone = pgdat->node_zones + i;
2979
2980                         if (!populated_zone(zone))
2981                                 continue;
2982
2983                         if (zone->all_unreclaimable &&
2984                             sc.priority != DEF_PRIORITY)
2985                                 continue;
2986
2987                         sc.nr_scanned = 0;
2988
2989                         nr_soft_scanned = 0;
2990                         /*
2991                          * Call soft limit reclaim before calling shrink_zone.
2992                          */
2993                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2994                                                         order, sc.gfp_mask,
2995                                                         &nr_soft_scanned);
2996                         sc.nr_reclaimed += nr_soft_reclaimed;
2997
2998                         /*
2999                          * There should be no need to raise the scanning
3000                          * priority if enough pages are already being scanned
3001                          * that that high watermark would be met at 100%
3002                          * efficiency.
3003                          */
3004                         if (kswapd_shrink_zone(zone, end_zone, &sc,
3005                                         lru_pages, &nr_attempted))
3006                                 raise_priority = false;
3007                 }
3008
3009                 /*
3010                  * If the low watermark is met there is no need for processes
3011                  * to be throttled on pfmemalloc_wait as they should not be
3012                  * able to safely make forward progress. Wake them
3013                  */
3014                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3015                                 pfmemalloc_watermark_ok(pgdat))
3016                         wake_up(&pgdat->pfmemalloc_wait);
3017
3018                 /*
3019                  * Fragmentation may mean that the system cannot be rebalanced
3020                  * for high-order allocations in all zones. If twice the
3021                  * allocation size has been reclaimed and the zones are still
3022                  * not balanced then recheck the watermarks at order-0 to
3023                  * prevent kswapd reclaiming excessively. Assume that a
3024                  * process requested a high-order can direct reclaim/compact.
3025                  */
3026                 if (order && sc.nr_reclaimed >= 2UL << order)
3027                         order = sc.order = 0;
3028
3029                 /* Check if kswapd should be suspending */
3030                 if (try_to_freeze() || kthread_should_stop())
3031                         break;
3032
3033                 /*
3034                  * Compact if necessary and kswapd is reclaiming at least the
3035                  * high watermark number of pages as requsted
3036                  */
3037                 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3038                         compact_pgdat(pgdat, order);
3039
3040                 /*
3041                  * Raise priority if scanning rate is too low or there was no
3042                  * progress in reclaiming pages
3043                  */
3044                 if (raise_priority || !sc.nr_reclaimed)
3045                         sc.priority--;
3046         } while (sc.priority >= 1 &&
3047                  !pgdat_balanced(pgdat, order, *classzone_idx));
3048
3049 out:
3050         /*
3051          * Return the order we were reclaiming at so prepare_kswapd_sleep()
3052          * makes a decision on the order we were last reclaiming at. However,
3053          * if another caller entered the allocator slow path while kswapd
3054          * was awake, order will remain at the higher level
3055          */
3056         *classzone_idx = end_zone;
3057         return order;
3058 }
3059
3060 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3061 {
3062         long remaining = 0;
3063         DEFINE_WAIT(wait);
3064
3065         if (freezing(current) || kthread_should_stop())
3066                 return;
3067
3068         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3069
3070         /* Try to sleep for a short interval */
3071         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3072                 remaining = schedule_timeout(HZ/10);
3073                 finish_wait(&pgdat->kswapd_wait, &wait);
3074                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3075         }
3076
3077         /*
3078          * After a short sleep, check if it was a premature sleep. If not, then
3079          * go fully to sleep until explicitly woken up.
3080          */
3081         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3082                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3083
3084                 /*
3085                  * vmstat counters are not perfectly accurate and the estimated
3086                  * value for counters such as NR_FREE_PAGES can deviate from the
3087                  * true value by nr_online_cpus * threshold. To avoid the zone
3088                  * watermarks being breached while under pressure, we reduce the
3089                  * per-cpu vmstat threshold while kswapd is awake and restore
3090                  * them before going back to sleep.
3091                  */
3092                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3093
3094                 /*
3095                  * Compaction records what page blocks it recently failed to
3096                  * isolate pages from and skips them in the future scanning.
3097                  * When kswapd is going to sleep, it is reasonable to assume
3098                  * that pages and compaction may succeed so reset the cache.
3099                  */
3100                 reset_isolation_suitable(pgdat);
3101
3102                 if (!kthread_should_stop())
3103                         schedule();
3104
3105                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3106         } else {
3107                 if (remaining)
3108                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3109                 else
3110                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3111         }
3112         finish_wait(&pgdat->kswapd_wait, &wait);
3113 }
3114
3115 /*
3116  * The background pageout daemon, started as a kernel thread
3117  * from the init process.
3118  *
3119  * This basically trickles out pages so that we have _some_
3120  * free memory available even if there is no other activity
3121  * that frees anything up. This is needed for things like routing
3122  * etc, where we otherwise might have all activity going on in
3123  * asynchronous contexts that cannot page things out.
3124  *
3125  * If there are applications that are active memory-allocators
3126  * (most normal use), this basically shouldn't matter.
3127  */
3128 static int kswapd(void *p)
3129 {
3130         unsigned long order, new_order;
3131         unsigned balanced_order;
3132         int classzone_idx, new_classzone_idx;
3133         int balanced_classzone_idx;
3134         pg_data_t *pgdat = (pg_data_t*)p;
3135         struct task_struct *tsk = current;
3136
3137         struct reclaim_state reclaim_state = {
3138                 .reclaimed_slab = 0,
3139         };
3140         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3141
3142         lockdep_set_current_reclaim_state(GFP_KERNEL);
3143
3144         if (!cpumask_empty(cpumask))
3145                 set_cpus_allowed_ptr(tsk, cpumask);
3146         current->reclaim_state = &reclaim_state;
3147
3148         /*
3149          * Tell the memory management that we're a "memory allocator",
3150          * and that if we need more memory we should get access to it
3151          * regardless (see "__alloc_pages()"). "kswapd" should
3152          * never get caught in the normal page freeing logic.
3153          *
3154          * (Kswapd normally doesn't need memory anyway, but sometimes
3155          * you need a small amount of memory in order to be able to
3156          * page out something else, and this flag essentially protects
3157          * us from recursively trying to free more memory as we're
3158          * trying to free the first piece of memory in the first place).
3159          */
3160         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3161         set_freezable();
3162
3163         order = new_order = 0;
3164         balanced_order = 0;
3165         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3166         balanced_classzone_idx = classzone_idx;
3167         for ( ; ; ) {
3168                 bool ret;
3169
3170                 /*
3171                  * If the last balance_pgdat was unsuccessful it's unlikely a
3172                  * new request of a similar or harder type will succeed soon
3173                  * so consider going to sleep on the basis we reclaimed at
3174                  */
3175                 if (balanced_classzone_idx >= new_classzone_idx &&
3176                                         balanced_order == new_order) {
3177                         new_order = pgdat->kswapd_max_order;
3178                         new_classzone_idx = pgdat->classzone_idx;
3179                         pgdat->kswapd_max_order =  0;
3180                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3181                 }
3182
3183                 if (order < new_order || classzone_idx > new_classzone_idx) {
3184                         /*
3185                          * Don't sleep if someone wants a larger 'order'
3186                          * allocation or has tigher zone constraints
3187                          */
3188                         order = new_order;
3189                         classzone_idx = new_classzone_idx;
3190                 } else {
3191                         kswapd_try_to_sleep(pgdat, balanced_order,
3192                                                 balanced_classzone_idx);
3193                         order = pgdat->kswapd_max_order;
3194                         classzone_idx = pgdat->classzone_idx;
3195                         new_order = order;
3196                         new_classzone_idx = classzone_idx;
3197                         pgdat->kswapd_max_order = 0;
3198                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3199                 }
3200
3201                 ret = try_to_freeze();
3202                 if (kthread_should_stop())
3203                         break;
3204
3205                 /*
3206                  * We can speed up thawing tasks if we don't call balance_pgdat
3207                  * after returning from the refrigerator
3208                  */
3209                 if (!ret) {
3210                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3211                         balanced_classzone_idx = classzone_idx;
3212                         balanced_order = balance_pgdat(pgdat, order,
3213                                                 &balanced_classzone_idx);
3214                 }
3215         }
3216
3217         current->reclaim_state = NULL;
3218         return 0;
3219 }
3220
3221 /*
3222  * A zone is low on free memory, so wake its kswapd task to service it.
3223  */
3224 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3225 {
3226         pg_data_t *pgdat;
3227
3228         if (!populated_zone(zone))
3229                 return;
3230
3231         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3232                 return;
3233         pgdat = zone->zone_pgdat;
3234         if (pgdat->kswapd_max_order < order) {
3235                 pgdat->kswapd_max_order = order;
3236                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3237         }
3238         if (!waitqueue_active(&pgdat->kswapd_wait))
3239                 return;
3240         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3241                 return;
3242
3243         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3244         wake_up_interruptible(&pgdat->kswapd_wait);
3245 }
3246
3247 /*
3248  * The reclaimable count would be mostly accurate.
3249  * The less reclaimable pages may be
3250  * - mlocked pages, which will be moved to unevictable list when encountered
3251  * - mapped pages, which may require several travels to be reclaimed
3252  * - dirty pages, which is not "instantly" reclaimable
3253  */
3254 unsigned long global_reclaimable_pages(void)
3255 {
3256         int nr;
3257
3258         nr = global_page_state(NR_ACTIVE_FILE) +
3259              global_page_state(NR_INACTIVE_FILE);
3260
3261         if (get_nr_swap_pages() > 0)
3262                 nr += global_page_state(NR_ACTIVE_ANON) +
3263                       global_page_state(NR_INACTIVE_ANON);
3264
3265         return nr;
3266 }
3267
3268 unsigned long zone_reclaimable_pages(struct zone *zone)
3269 {
3270         int nr;
3271
3272         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3273              zone_page_state(zone, NR_INACTIVE_FILE);
3274
3275         if (get_nr_swap_pages() > 0)
3276                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3277                       zone_page_state(zone, NR_INACTIVE_ANON);
3278
3279         return nr;
3280 }
3281
3282 #ifdef CONFIG_HIBERNATION
3283 /*
3284  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3285  * freed pages.
3286  *
3287  * Rather than trying to age LRUs the aim is to preserve the overall
3288  * LRU order by reclaiming preferentially
3289  * inactive > active > active referenced > active mapped
3290  */
3291 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3292 {
3293         struct reclaim_state reclaim_state;
3294         struct scan_control sc = {
3295                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3296                 .may_swap = 1,
3297                 .may_unmap = 1,
3298                 .may_writepage = 1,
3299                 .nr_to_reclaim = nr_to_reclaim,
3300                 .hibernation_mode = 1,
3301                 .order = 0,
3302                 .priority = DEF_PRIORITY,
3303         };
3304         struct shrink_control shrink = {
3305                 .gfp_mask = sc.gfp_mask,
3306         };
3307         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3308         struct task_struct *p = current;
3309         unsigned long nr_reclaimed;
3310
3311         p->flags |= PF_MEMALLOC;
3312         lockdep_set_current_reclaim_state(sc.gfp_mask);
3313         reclaim_state.reclaimed_slab = 0;
3314         p->reclaim_state = &reclaim_state;
3315
3316         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3317
3318         p->reclaim_state = NULL;
3319         lockdep_clear_current_reclaim_state();
3320         p->flags &= ~PF_MEMALLOC;
3321
3322         return nr_reclaimed;
3323 }
3324 #endif /* CONFIG_HIBERNATION */
3325
3326 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3327    not required for correctness.  So if the last cpu in a node goes
3328    away, we get changed to run anywhere: as the first one comes back,
3329    restore their cpu bindings. */
3330 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3331                         void *hcpu)
3332 {
3333         int nid;
3334
3335         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3336                 for_each_node_state(nid, N_MEMORY) {
3337                         pg_data_t *pgdat = NODE_DATA(nid);
3338                         const struct cpumask *mask;
3339
3340                         mask = cpumask_of_node(pgdat->node_id);
3341
3342                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3343                                 /* One of our CPUs online: restore mask */
3344                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3345                 }
3346         }
3347         return NOTIFY_OK;
3348 }
3349
3350 /*
3351  * This kswapd start function will be called by init and node-hot-add.
3352  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3353  */
3354 int kswapd_run(int nid)
3355 {
3356         pg_data_t *pgdat = NODE_DATA(nid);
3357         int ret = 0;
3358
3359         if (pgdat->kswapd)
3360                 return 0;
3361
3362         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3363         if (IS_ERR(pgdat->kswapd)) {
3364                 /* failure at boot is fatal */
3365                 BUG_ON(system_state == SYSTEM_BOOTING);
3366                 pr_err("Failed to start kswapd on node %d\n", nid);
3367                 ret = PTR_ERR(pgdat->kswapd);
3368                 pgdat->kswapd = NULL;
3369         }
3370         return ret;
3371 }
3372
3373 /*
3374  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3375  * hold lock_memory_hotplug().
3376  */
3377 void kswapd_stop(int nid)
3378 {
3379         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3380
3381         if (kswapd) {
3382                 kthread_stop(kswapd);
3383                 NODE_DATA(nid)->kswapd = NULL;
3384         }
3385 }
3386
3387 static int __init kswapd_init(void)
3388 {
3389         int nid;
3390
3391         swap_setup();
3392         for_each_node_state(nid, N_MEMORY)
3393                 kswapd_run(nid);
3394         hotcpu_notifier(cpu_callback, 0);
3395         return 0;
3396 }
3397
3398 module_init(kswapd_init)
3399
3400 #ifdef CONFIG_NUMA
3401 /*
3402  * Zone reclaim mode
3403  *
3404  * If non-zero call zone_reclaim when the number of free pages falls below
3405  * the watermarks.
3406  */
3407 int zone_reclaim_mode __read_mostly;
3408
3409 #define RECLAIM_OFF 0
3410 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3411 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3412 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3413
3414 /*
3415  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3416  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3417  * a zone.
3418  */
3419 #define ZONE_RECLAIM_PRIORITY 4
3420
3421 /*
3422  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3423  * occur.
3424  */
3425 int sysctl_min_unmapped_ratio = 1;
3426
3427 /*
3428  * If the number of slab pages in a zone grows beyond this percentage then
3429  * slab reclaim needs to occur.
3430  */
3431 int sysctl_min_slab_ratio = 5;
3432
3433 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3434 {
3435         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3436         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3437                 zone_page_state(zone, NR_ACTIVE_FILE);
3438
3439         /*
3440          * It's possible for there to be more file mapped pages than
3441          * accounted for by the pages on the file LRU lists because
3442          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3443          */
3444         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3445 }
3446
3447 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3448 static long zone_pagecache_reclaimable(struct zone *zone)
3449 {
3450         long nr_pagecache_reclaimable;
3451         long delta = 0;
3452
3453         /*
3454          * If RECLAIM_SWAP is set, then all file pages are considered
3455          * potentially reclaimable. Otherwise, we have to worry about
3456          * pages like swapcache and zone_unmapped_file_pages() provides
3457          * a better estimate
3458          */
3459         if (zone_reclaim_mode & RECLAIM_SWAP)
3460                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3461         else
3462                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3463
3464         /* If we can't clean pages, remove dirty pages from consideration */
3465         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3466                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3467
3468         /* Watch for any possible underflows due to delta */
3469         if (unlikely(delta > nr_pagecache_reclaimable))
3470                 delta = nr_pagecache_reclaimable;
3471
3472         return nr_pagecache_reclaimable - delta;
3473 }
3474
3475 /*
3476  * Try to free up some pages from this zone through reclaim.
3477  */
3478 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3479 {
3480         /* Minimum pages needed in order to stay on node */
3481         const unsigned long nr_pages = 1 << order;
3482         struct task_struct *p = current;
3483         struct reclaim_state reclaim_state;
3484         struct scan_control sc = {
3485                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3486                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3487                 .may_swap = 1,
3488                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3489                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3490                 .order = order,
3491                 .priority = ZONE_RECLAIM_PRIORITY,
3492         };
3493         struct shrink_control shrink = {
3494                 .gfp_mask = sc.gfp_mask,
3495         };
3496         unsigned long nr_slab_pages0, nr_slab_pages1;
3497
3498         cond_resched();
3499         /*
3500          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3501          * and we also need to be able to write out pages for RECLAIM_WRITE
3502          * and RECLAIM_SWAP.
3503          */
3504         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3505         lockdep_set_current_reclaim_state(gfp_mask);
3506         reclaim_state.reclaimed_slab = 0;
3507         p->reclaim_state = &reclaim_state;
3508
3509         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3510                 /*
3511                  * Free memory by calling shrink zone with increasing
3512                  * priorities until we have enough memory freed.
3513                  */
3514                 do {
3515                         shrink_zone(zone, &sc);
3516                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3517         }
3518
3519         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3520         if (nr_slab_pages0 > zone->min_slab_pages) {
3521                 /*
3522                  * shrink_slab() does not currently allow us to determine how
3523                  * many pages were freed in this zone. So we take the current
3524                  * number of slab pages and shake the slab until it is reduced
3525                  * by the same nr_pages that we used for reclaiming unmapped
3526                  * pages.
3527                  *
3528                  * Note that shrink_slab will free memory on all zones and may
3529                  * take a long time.
3530                  */
3531                 for (;;) {
3532                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3533
3534                         /* No reclaimable slab or very low memory pressure */
3535                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3536                                 break;
3537
3538                         /* Freed enough memory */
3539                         nr_slab_pages1 = zone_page_state(zone,
3540                                                         NR_SLAB_RECLAIMABLE);
3541                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3542                                 break;
3543                 }
3544
3545                 /*
3546                  * Update nr_reclaimed by the number of slab pages we
3547                  * reclaimed from this zone.
3548                  */
3549                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3550                 if (nr_slab_pages1 < nr_slab_pages0)
3551                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3552         }
3553
3554         p->reclaim_state = NULL;
3555         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3556         lockdep_clear_current_reclaim_state();
3557         return sc.nr_reclaimed >= nr_pages;
3558 }
3559
3560 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3561 {
3562         int node_id;
3563         int ret;
3564
3565         /*
3566          * Zone reclaim reclaims unmapped file backed pages and
3567          * slab pages if we are over the defined limits.
3568          *
3569          * A small portion of unmapped file backed pages is needed for
3570          * file I/O otherwise pages read by file I/O will be immediately
3571          * thrown out if the zone is overallocated. So we do not reclaim
3572          * if less than a specified percentage of the zone is used by
3573          * unmapped file backed pages.
3574          */
3575         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3576             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3577                 return ZONE_RECLAIM_FULL;
3578
3579         if (zone->all_unreclaimable)
3580                 return ZONE_RECLAIM_FULL;
3581
3582         /*
3583          * Do not scan if the allocation should not be delayed.
3584          */
3585         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3586                 return ZONE_RECLAIM_NOSCAN;
3587
3588         /*
3589          * Only run zone reclaim on the local zone or on zones that do not
3590          * have associated processors. This will favor the local processor
3591          * over remote processors and spread off node memory allocations
3592          * as wide as possible.
3593          */
3594         node_id = zone_to_nid(zone);
3595         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3596                 return ZONE_RECLAIM_NOSCAN;
3597
3598         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3599                 return ZONE_RECLAIM_NOSCAN;
3600
3601         ret = __zone_reclaim(zone, gfp_mask, order);
3602         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3603
3604         if (!ret)
3605                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3606
3607         return ret;
3608 }
3609 #endif
3610
3611 /*
3612  * page_evictable - test whether a page is evictable
3613  * @page: the page to test
3614  *
3615  * Test whether page is evictable--i.e., should be placed on active/inactive
3616  * lists vs unevictable list.
3617  *
3618  * Reasons page might not be evictable:
3619  * (1) page's mapping marked unevictable
3620  * (2) page is part of an mlocked VMA
3621  *
3622  */
3623 int page_evictable(struct page *page)
3624 {
3625         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3626 }
3627
3628 #ifdef CONFIG_SHMEM
3629 /**
3630  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3631  * @pages:      array of pages to check
3632  * @nr_pages:   number of pages to check
3633  *
3634  * Checks pages for evictability and moves them to the appropriate lru list.
3635  *
3636  * This function is only used for SysV IPC SHM_UNLOCK.
3637  */
3638 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3639 {
3640         struct lruvec *lruvec;
3641         struct zone *zone = NULL;
3642         int pgscanned = 0;
3643         int pgrescued = 0;
3644         int i;
3645
3646         for (i = 0; i < nr_pages; i++) {
3647                 struct page *page = pages[i];
3648                 struct zone *pagezone;
3649
3650                 pgscanned++;
3651                 pagezone = page_zone(page);
3652                 if (pagezone != zone) {
3653                         if (zone)
3654                                 spin_unlock_irq(&zone->lru_lock);
3655                         zone = pagezone;
3656                         spin_lock_irq(&zone->lru_lock);
3657                 }
3658                 lruvec = mem_cgroup_page_lruvec(page, zone);
3659
3660                 if (!PageLRU(page) || !PageUnevictable(page))
3661                         continue;
3662
3663                 if (page_evictable(page)) {
3664                         enum lru_list lru = page_lru_base_type(page);
3665
3666                         VM_BUG_ON(PageActive(page));
3667                         ClearPageUnevictable(page);
3668                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3669                         add_page_to_lru_list(page, lruvec, lru);
3670                         pgrescued++;
3671                 }
3672         }
3673
3674         if (zone) {
3675                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3676                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3677                 spin_unlock_irq(&zone->lru_lock);
3678         }
3679 }
3680 #endif /* CONFIG_SHMEM */
3681
3682 static void warn_scan_unevictable_pages(void)
3683 {
3684         printk_once(KERN_WARNING
3685                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3686                     "disabled for lack of a legitimate use case.  If you have "
3687                     "one, please send an email to linux-mm@kvack.org.\n",
3688                     current->comm);
3689 }
3690
3691 /*
3692  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3693  * all nodes' unevictable lists for evictable pages
3694  */
3695 unsigned long scan_unevictable_pages;
3696
3697 int scan_unevictable_handler(struct ctl_table *table, int write,
3698                            void __user *buffer,
3699                            size_t *length, loff_t *ppos)
3700 {
3701         warn_scan_unevictable_pages();
3702         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3703         scan_unevictable_pages = 0;
3704         return 0;
3705 }
3706
3707 #ifdef CONFIG_NUMA
3708 /*
3709  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3710  * a specified node's per zone unevictable lists for evictable pages.
3711  */
3712
3713 static ssize_t read_scan_unevictable_node(struct device *dev,
3714                                           struct device_attribute *attr,
3715                                           char *buf)
3716 {
3717         warn_scan_unevictable_pages();
3718         return sprintf(buf, "0\n");     /* always zero; should fit... */
3719 }
3720
3721 static ssize_t write_scan_unevictable_node(struct device *dev,
3722                                            struct device_attribute *attr,
3723                                         const char *buf, size_t count)
3724 {
3725         warn_scan_unevictable_pages();
3726         return 1;
3727 }
3728
3729
3730 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3731                         read_scan_unevictable_node,
3732                         write_scan_unevictable_node);
3733
3734 int scan_unevictable_register_node(struct node *node)
3735 {
3736         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3737 }
3738
3739 void scan_unevictable_unregister_node(struct node *node)
3740 {
3741         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3742 }
3743 #endif