UPSTREAM mm: memcg: enable memcg OOM killer only for user faults
[linux-3.10.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 #include <linux/debugfs.h>
47
48 #include <asm/tlbflush.h>
49 #include <asm/div64.h>
50
51 #include <linux/swapops.h>
52
53 #include "internal.h"
54
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/vmscan.h>
57
58 struct scan_control {
59         /* Incremented by the number of inactive pages that were scanned */
60         unsigned long nr_scanned;
61
62         /* Number of pages freed so far during a call to shrink_zones() */
63         unsigned long nr_reclaimed;
64
65         /* How many pages shrink_list() should reclaim */
66         unsigned long nr_to_reclaim;
67
68         unsigned long hibernation_mode;
69
70         /* This context's GFP mask */
71         gfp_t gfp_mask;
72
73         int may_writepage;
74
75         /* Can mapped pages be reclaimed? */
76         int may_unmap;
77
78         /* Can pages be swapped as part of reclaim? */
79         int may_swap;
80
81         int order;
82
83         /* Scan (total_size >> priority) pages at once */
84         int priority;
85
86         /*
87          * The memory cgroup that hit its limit and as a result is the
88          * primary target of this reclaim invocation.
89          */
90         struct mem_cgroup *target_mem_cgroup;
91
92         /*
93          * Nodemask of nodes allowed by the caller. If NULL, all nodes
94          * are scanned.
95          */
96         nodemask_t      *nodemask;
97 };
98
99 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100
101 #ifdef ARCH_HAS_PREFETCH
102 #define prefetch_prev_lru_page(_page, _base, _field)                    \
103         do {                                                            \
104                 if ((_page)->lru.prev != _base) {                       \
105                         struct page *prev;                              \
106                                                                         \
107                         prev = lru_to_page(&(_page->lru));              \
108                         prefetch(&prev->_field);                        \
109                 }                                                       \
110         } while (0)
111 #else
112 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #endif
114
115 #ifdef ARCH_HAS_PREFETCHW
116 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
117         do {                                                            \
118                 if ((_page)->lru.prev != _base) {                       \
119                         struct page *prev;                              \
120                                                                         \
121                         prev = lru_to_page(&(_page->lru));              \
122                         prefetchw(&prev->_field);                       \
123                 }                                                       \
124         } while (0)
125 #else
126 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127 #endif
128
129 /*
130  * From 0 .. 100.  Higher means more swappy.
131  */
132 int vm_swappiness = 60;
133 unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
134
135 static LIST_HEAD(shrinker_list);
136 static DECLARE_RWSEM(shrinker_rwsem);
137
138 #ifdef CONFIG_MEMCG
139 static bool global_reclaim(struct scan_control *sc)
140 {
141         return !sc->target_mem_cgroup;
142 }
143 #else
144 static bool global_reclaim(struct scan_control *sc)
145 {
146         return true;
147 }
148 #endif
149
150 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
151 {
152         if (!mem_cgroup_disabled())
153                 return mem_cgroup_get_lru_size(lruvec, lru);
154
155         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
156 }
157
158 struct dentry *debug_file;
159
160 static int debug_shrinker_show(struct seq_file *s, void *unused)
161 {
162         struct shrinker *shrinker;
163         struct shrink_control sc;
164
165         sc.gfp_mask = -1;
166         sc.nr_to_scan = 0;
167
168         down_read(&shrinker_rwsem);
169         list_for_each_entry(shrinker, &shrinker_list, list) {
170                 char name[64];
171                 int num_objs;
172
173                 num_objs = shrinker->shrink(shrinker, &sc);
174                 seq_printf(s, "%pf %d\n", shrinker->shrink, num_objs);
175         }
176         up_read(&shrinker_rwsem);
177         return 0;
178 }
179
180 static int debug_shrinker_open(struct inode *inode, struct file *file)
181 {
182         return single_open(file, debug_shrinker_show, inode->i_private);
183 }
184
185 static const struct file_operations debug_shrinker_fops = {
186         .open = debug_shrinker_open,
187         .read = seq_read,
188         .llseek = seq_lseek,
189         .release = single_release,
190 };
191
192 /*
193  * Add a shrinker callback to be called from the vm
194  */
195 void register_shrinker(struct shrinker *shrinker)
196 {
197         atomic_long_set(&shrinker->nr_in_batch, 0);
198         down_write(&shrinker_rwsem);
199         list_add_tail(&shrinker->list, &shrinker_list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(register_shrinker);
203
204 static int __init add_shrinker_debug(void)
205 {
206         debugfs_create_file("shrinker", 0644, NULL, NULL,
207                             &debug_shrinker_fops);
208         return 0;
209 }
210
211 late_initcall(add_shrinker_debug);
212
213 /*
214  * Remove one
215  */
216 void unregister_shrinker(struct shrinker *shrinker)
217 {
218         down_write(&shrinker_rwsem);
219         list_del(&shrinker->list);
220         up_write(&shrinker_rwsem);
221 }
222 EXPORT_SYMBOL(unregister_shrinker);
223
224 static inline int do_shrinker_shrink(struct shrinker *shrinker,
225                                      struct shrink_control *sc,
226                                      unsigned long nr_to_scan)
227 {
228         sc->nr_to_scan = nr_to_scan;
229         return (*shrinker->shrink)(shrinker, sc);
230 }
231
232 #define SHRINK_BATCH 128
233 /*
234  * Call the shrink functions to age shrinkable caches
235  *
236  * Here we assume it costs one seek to replace a lru page and that it also
237  * takes a seek to recreate a cache object.  With this in mind we age equal
238  * percentages of the lru and ageable caches.  This should balance the seeks
239  * generated by these structures.
240  *
241  * If the vm encountered mapped pages on the LRU it increase the pressure on
242  * slab to avoid swapping.
243  *
244  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
245  *
246  * `lru_pages' represents the number of on-LRU pages in all the zones which
247  * are eligible for the caller's allocation attempt.  It is used for balancing
248  * slab reclaim versus page reclaim.
249  *
250  * Returns the number of slab objects which we shrunk.
251  */
252 unsigned long shrink_slab(struct shrink_control *shrink,
253                           unsigned long nr_pages_scanned,
254                           unsigned long lru_pages)
255 {
256         struct shrinker *shrinker;
257         unsigned long ret = 0;
258
259         if (nr_pages_scanned == 0)
260                 nr_pages_scanned = SWAP_CLUSTER_MAX;
261
262         if (!down_read_trylock(&shrinker_rwsem)) {
263                 /* Assume we'll be able to shrink next time */
264                 ret = 1;
265                 goto out;
266         }
267
268         list_for_each_entry(shrinker, &shrinker_list, list) {
269                 unsigned long long delta;
270                 long total_scan;
271                 long max_pass;
272                 int shrink_ret = 0;
273                 long nr;
274                 long new_nr;
275                 long batch_size = shrinker->batch ? shrinker->batch
276                                                   : SHRINK_BATCH;
277
278                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
279                 if (max_pass <= 0)
280                         continue;
281
282                 /*
283                  * copy the current shrinker scan count into a local variable
284                  * and zero it so that other concurrent shrinker invocations
285                  * don't also do this scanning work.
286                  */
287                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
288
289                 total_scan = nr;
290                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
291                 delta *= max_pass;
292                 do_div(delta, lru_pages + 1);
293                 total_scan += delta;
294                 if (total_scan < 0) {
295                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
296                                "delete nr=%ld\n",
297                                shrinker->shrink, total_scan);
298                         total_scan = max_pass;
299                 }
300
301                 /*
302                  * We need to avoid excessive windup on filesystem shrinkers
303                  * due to large numbers of GFP_NOFS allocations causing the
304                  * shrinkers to return -1 all the time. This results in a large
305                  * nr being built up so when a shrink that can do some work
306                  * comes along it empties the entire cache due to nr >>>
307                  * max_pass.  This is bad for sustaining a working set in
308                  * memory.
309                  *
310                  * Hence only allow the shrinker to scan the entire cache when
311                  * a large delta change is calculated directly.
312                  */
313                 if (delta < max_pass / 4)
314                         total_scan = min(total_scan, max_pass / 2);
315
316                 /*
317                  * Avoid risking looping forever due to too large nr value:
318                  * never try to free more than twice the estimate number of
319                  * freeable entries.
320                  */
321                 if (total_scan > max_pass * 2)
322                         total_scan = max_pass * 2;
323
324                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
325                                         nr_pages_scanned, lru_pages,
326                                         max_pass, delta, total_scan);
327
328                 while (total_scan >= batch_size) {
329                         int nr_before;
330
331                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
332                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
333                                                         batch_size);
334                         if (shrink_ret == -1)
335                                 break;
336                         if (shrink_ret < nr_before)
337                                 ret += nr_before - shrink_ret;
338                         count_vm_events(SLABS_SCANNED, batch_size);
339                         total_scan -= batch_size;
340
341                         cond_resched();
342                 }
343
344                 /*
345                  * move the unused scan count back into the shrinker in a
346                  * manner that handles concurrent updates. If we exhausted the
347                  * scan, there is no need to do an update.
348                  */
349                 if (total_scan > 0)
350                         new_nr = atomic_long_add_return(total_scan,
351                                         &shrinker->nr_in_batch);
352                 else
353                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
354
355                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
356         }
357         up_read(&shrinker_rwsem);
358 out:
359         cond_resched();
360         return ret;
361 }
362
363 static inline int is_page_cache_freeable(struct page *page)
364 {
365         /*
366          * A freeable page cache page is referenced only by the caller
367          * that isolated the page, the page cache radix tree and
368          * optional buffer heads at page->private.
369          */
370         return page_count(page) - page_has_private(page) == 2;
371 }
372
373 static int may_write_to_queue(struct backing_dev_info *bdi,
374                               struct scan_control *sc)
375 {
376         if (current->flags & PF_SWAPWRITE)
377                 return 1;
378         if (!bdi_write_congested(bdi))
379                 return 1;
380         if (bdi == current->backing_dev_info)
381                 return 1;
382         return 0;
383 }
384
385 /*
386  * We detected a synchronous write error writing a page out.  Probably
387  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
388  * fsync(), msync() or close().
389  *
390  * The tricky part is that after writepage we cannot touch the mapping: nothing
391  * prevents it from being freed up.  But we have a ref on the page and once
392  * that page is locked, the mapping is pinned.
393  *
394  * We're allowed to run sleeping lock_page() here because we know the caller has
395  * __GFP_FS.
396  */
397 static void handle_write_error(struct address_space *mapping,
398                                 struct page *page, int error)
399 {
400         lock_page(page);
401         if (page_mapping(page) == mapping)
402                 mapping_set_error(mapping, error);
403         unlock_page(page);
404 }
405
406 /* possible outcome of pageout() */
407 typedef enum {
408         /* failed to write page out, page is locked */
409         PAGE_KEEP,
410         /* move page to the active list, page is locked */
411         PAGE_ACTIVATE,
412         /* page has been sent to the disk successfully, page is unlocked */
413         PAGE_SUCCESS,
414         /* page is clean and locked */
415         PAGE_CLEAN,
416 } pageout_t;
417
418 /*
419  * pageout is called by shrink_page_list() for each dirty page.
420  * Calls ->writepage().
421  */
422 static pageout_t pageout(struct page *page, struct address_space *mapping,
423                          struct scan_control *sc)
424 {
425         /*
426          * If the page is dirty, only perform writeback if that write
427          * will be non-blocking.  To prevent this allocation from being
428          * stalled by pagecache activity.  But note that there may be
429          * stalls if we need to run get_block().  We could test
430          * PagePrivate for that.
431          *
432          * If this process is currently in __generic_file_aio_write() against
433          * this page's queue, we can perform writeback even if that
434          * will block.
435          *
436          * If the page is swapcache, write it back even if that would
437          * block, for some throttling. This happens by accident, because
438          * swap_backing_dev_info is bust: it doesn't reflect the
439          * congestion state of the swapdevs.  Easy to fix, if needed.
440          */
441         if (!is_page_cache_freeable(page))
442                 return PAGE_KEEP;
443         if (!mapping) {
444                 /*
445                  * Some data journaling orphaned pages can have
446                  * page->mapping == NULL while being dirty with clean buffers.
447                  */
448                 if (page_has_private(page)) {
449                         if (try_to_free_buffers(page)) {
450                                 ClearPageDirty(page);
451                                 printk("%s: orphaned page\n", __func__);
452                                 return PAGE_CLEAN;
453                         }
454                 }
455                 return PAGE_KEEP;
456         }
457         if (mapping->a_ops->writepage == NULL)
458                 return PAGE_ACTIVATE;
459         if (!may_write_to_queue(mapping->backing_dev_info, sc))
460                 return PAGE_KEEP;
461
462         if (clear_page_dirty_for_io(page)) {
463                 int res;
464                 struct writeback_control wbc = {
465                         .sync_mode = WB_SYNC_NONE,
466                         .nr_to_write = SWAP_CLUSTER_MAX,
467                         .range_start = 0,
468                         .range_end = LLONG_MAX,
469                         .for_reclaim = 1,
470                 };
471
472                 SetPageReclaim(page);
473                 res = mapping->a_ops->writepage(page, &wbc);
474                 if (res < 0)
475                         handle_write_error(mapping, page, res);
476                 if (res == AOP_WRITEPAGE_ACTIVATE) {
477                         ClearPageReclaim(page);
478                         return PAGE_ACTIVATE;
479                 }
480
481                 if (!PageWriteback(page)) {
482                         /* synchronous write or broken a_ops? */
483                         ClearPageReclaim(page);
484                 }
485                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
486                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
487                 return PAGE_SUCCESS;
488         }
489
490         return PAGE_CLEAN;
491 }
492
493 /*
494  * Same as remove_mapping, but if the page is removed from the mapping, it
495  * gets returned with a refcount of 0.
496  */
497 static int __remove_mapping(struct address_space *mapping, struct page *page)
498 {
499         BUG_ON(!PageLocked(page));
500         BUG_ON(mapping != page_mapping(page));
501
502         spin_lock_irq(&mapping->tree_lock);
503         /*
504          * The non racy check for a busy page.
505          *
506          * Must be careful with the order of the tests. When someone has
507          * a ref to the page, it may be possible that they dirty it then
508          * drop the reference. So if PageDirty is tested before page_count
509          * here, then the following race may occur:
510          *
511          * get_user_pages(&page);
512          * [user mapping goes away]
513          * write_to(page);
514          *                              !PageDirty(page)    [good]
515          * SetPageDirty(page);
516          * put_page(page);
517          *                              !page_count(page)   [good, discard it]
518          *
519          * [oops, our write_to data is lost]
520          *
521          * Reversing the order of the tests ensures such a situation cannot
522          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
523          * load is not satisfied before that of page->_count.
524          *
525          * Note that if SetPageDirty is always performed via set_page_dirty,
526          * and thus under tree_lock, then this ordering is not required.
527          */
528         if (!page_freeze_refs(page, 2))
529                 goto cannot_free;
530         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
531         if (unlikely(PageDirty(page))) {
532                 page_unfreeze_refs(page, 2);
533                 goto cannot_free;
534         }
535
536         if (PageSwapCache(page)) {
537                 swp_entry_t swap = { .val = page_private(page) };
538                 __delete_from_swap_cache(page);
539                 spin_unlock_irq(&mapping->tree_lock);
540                 swapcache_free(swap, page);
541         } else {
542                 void (*freepage)(struct page *);
543
544                 freepage = mapping->a_ops->freepage;
545
546                 __delete_from_page_cache(page);
547                 spin_unlock_irq(&mapping->tree_lock);
548                 mem_cgroup_uncharge_cache_page(page);
549
550                 if (freepage != NULL)
551                         freepage(page);
552         }
553
554         return 1;
555
556 cannot_free:
557         spin_unlock_irq(&mapping->tree_lock);
558         return 0;
559 }
560
561 /*
562  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
563  * someone else has a ref on the page, abort and return 0.  If it was
564  * successfully detached, return 1.  Assumes the caller has a single ref on
565  * this page.
566  */
567 int remove_mapping(struct address_space *mapping, struct page *page)
568 {
569         if (__remove_mapping(mapping, page)) {
570                 /*
571                  * Unfreezing the refcount with 1 rather than 2 effectively
572                  * drops the pagecache ref for us without requiring another
573                  * atomic operation.
574                  */
575                 page_unfreeze_refs(page, 1);
576                 return 1;
577         }
578         return 0;
579 }
580
581 /**
582  * putback_lru_page - put previously isolated page onto appropriate LRU list
583  * @page: page to be put back to appropriate lru list
584  *
585  * Add previously isolated @page to appropriate LRU list.
586  * Page may still be unevictable for other reasons.
587  *
588  * lru_lock must not be held, interrupts must be enabled.
589  */
590 void putback_lru_page(struct page *page)
591 {
592         int lru;
593         int active = !!TestClearPageActive(page);
594         int was_unevictable = PageUnevictable(page);
595
596         VM_BUG_ON(PageLRU(page));
597
598 redo:
599         ClearPageUnevictable(page);
600
601         if (page_evictable(page)) {
602                 /*
603                  * For evictable pages, we can use the cache.
604                  * In event of a race, worst case is we end up with an
605                  * unevictable page on [in]active list.
606                  * We know how to handle that.
607                  */
608                 lru = active + page_lru_base_type(page);
609                 lru_cache_add_lru(page, lru);
610         } else {
611                 /*
612                  * Put unevictable pages directly on zone's unevictable
613                  * list.
614                  */
615                 lru = LRU_UNEVICTABLE;
616                 add_page_to_unevictable_list(page);
617                 /*
618                  * When racing with an mlock or AS_UNEVICTABLE clearing
619                  * (page is unlocked) make sure that if the other thread
620                  * does not observe our setting of PG_lru and fails
621                  * isolation/check_move_unevictable_pages,
622                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
623                  * the page back to the evictable list.
624                  *
625                  * The other side is TestClearPageMlocked() or shmem_lock().
626                  */
627                 smp_mb();
628         }
629
630         /*
631          * page's status can change while we move it among lru. If an evictable
632          * page is on unevictable list, it never be freed. To avoid that,
633          * check after we added it to the list, again.
634          */
635         if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
636                 if (!isolate_lru_page(page)) {
637                         put_page(page);
638                         goto redo;
639                 }
640                 /* This means someone else dropped this page from LRU
641                  * So, it will be freed or putback to LRU again. There is
642                  * nothing to do here.
643                  */
644         }
645
646         if (was_unevictable && lru != LRU_UNEVICTABLE)
647                 count_vm_event(UNEVICTABLE_PGRESCUED);
648         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
649                 count_vm_event(UNEVICTABLE_PGCULLED);
650
651         put_page(page);         /* drop ref from isolate */
652 }
653
654 enum page_references {
655         PAGEREF_RECLAIM,
656         PAGEREF_RECLAIM_CLEAN,
657         PAGEREF_KEEP,
658         PAGEREF_ACTIVATE,
659 };
660
661 static enum page_references page_check_references(struct page *page,
662                                                   struct scan_control *sc)
663 {
664         int referenced_ptes, referenced_page;
665         unsigned long vm_flags;
666
667         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
668                                           &vm_flags);
669         referenced_page = TestClearPageReferenced(page);
670
671         /*
672          * Mlock lost the isolation race with us.  Let try_to_unmap()
673          * move the page to the unevictable list.
674          */
675         if (vm_flags & VM_LOCKED)
676                 return PAGEREF_RECLAIM;
677
678         if (referenced_ptes) {
679                 if (PageSwapBacked(page))
680                         return PAGEREF_ACTIVATE;
681                 /*
682                  * All mapped pages start out with page table
683                  * references from the instantiating fault, so we need
684                  * to look twice if a mapped file page is used more
685                  * than once.
686                  *
687                  * Mark it and spare it for another trip around the
688                  * inactive list.  Another page table reference will
689                  * lead to its activation.
690                  *
691                  * Note: the mark is set for activated pages as well
692                  * so that recently deactivated but used pages are
693                  * quickly recovered.
694                  */
695                 SetPageReferenced(page);
696
697                 if (referenced_page || referenced_ptes > 1)
698                         return PAGEREF_ACTIVATE;
699
700                 /*
701                  * Activate file-backed executable pages after first usage.
702                  */
703                 if (vm_flags & VM_EXEC)
704                         return PAGEREF_ACTIVATE;
705
706                 return PAGEREF_KEEP;
707         }
708
709         /* Reclaim if clean, defer dirty pages to writeback */
710         if (referenced_page && !PageSwapBacked(page))
711                 return PAGEREF_RECLAIM_CLEAN;
712
713         return PAGEREF_RECLAIM;
714 }
715
716 /*
717  * shrink_page_list() returns the number of reclaimed pages
718  */
719 static unsigned long shrink_page_list(struct list_head *page_list,
720                                       struct zone *zone,
721                                       struct scan_control *sc,
722                                       enum ttu_flags ttu_flags,
723                                       unsigned long *ret_nr_dirty,
724                                       unsigned long *ret_nr_writeback,
725                                       bool force_reclaim)
726 {
727         LIST_HEAD(ret_pages);
728         LIST_HEAD(free_pages);
729         int pgactivate = 0;
730         unsigned long nr_dirty = 0;
731         unsigned long nr_congested = 0;
732         unsigned long nr_reclaimed = 0;
733         unsigned long nr_writeback = 0;
734
735         cond_resched();
736
737         mem_cgroup_uncharge_start();
738         while (!list_empty(page_list)) {
739                 struct address_space *mapping;
740                 struct page *page;
741                 int may_enter_fs;
742                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
743
744                 cond_resched();
745
746                 page = lru_to_page(page_list);
747                 list_del(&page->lru);
748
749                 if (!trylock_page(page))
750                         goto keep;
751
752                 VM_BUG_ON(PageActive(page));
753                 VM_BUG_ON(page_zone(page) != zone);
754
755                 sc->nr_scanned++;
756
757                 if (unlikely(!page_evictable(page)))
758                         goto cull_mlocked;
759
760                 if (!sc->may_unmap && page_mapped(page))
761                         goto keep_locked;
762
763                 /* Double the slab pressure for mapped and swapcache pages */
764                 if (page_mapped(page) || PageSwapCache(page))
765                         sc->nr_scanned++;
766
767                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
768                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
769
770                 if (PageWriteback(page)) {
771                         /*
772                          * memcg doesn't have any dirty pages throttling so we
773                          * could easily OOM just because too many pages are in
774                          * writeback and there is nothing else to reclaim.
775                          *
776                          * Check __GFP_IO, certainly because a loop driver
777                          * thread might enter reclaim, and deadlock if it waits
778                          * on a page for which it is needed to do the write
779                          * (loop masks off __GFP_IO|__GFP_FS for this reason);
780                          * but more thought would probably show more reasons.
781                          *
782                          * Don't require __GFP_FS, since we're not going into
783                          * the FS, just waiting on its writeback completion.
784                          * Worryingly, ext4 gfs2 and xfs allocate pages with
785                          * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
786                          * testing may_enter_fs here is liable to OOM on them.
787                          */
788                         if (global_reclaim(sc) ||
789                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
790                                 /*
791                                  * This is slightly racy - end_page_writeback()
792                                  * might have just cleared PageReclaim, then
793                                  * setting PageReclaim here end up interpreted
794                                  * as PageReadahead - but that does not matter
795                                  * enough to care.  What we do want is for this
796                                  * page to have PageReclaim set next time memcg
797                                  * reclaim reaches the tests above, so it will
798                                  * then wait_on_page_writeback() to avoid OOM;
799                                  * and it's also appropriate in global reclaim.
800                                  */
801                                 SetPageReclaim(page);
802                                 nr_writeback++;
803                                 goto keep_locked;
804                         }
805                         wait_on_page_writeback(page);
806                 }
807
808                 if (!force_reclaim)
809                         references = page_check_references(page, sc);
810
811                 switch (references) {
812                 case PAGEREF_ACTIVATE:
813                         goto activate_locked;
814                 case PAGEREF_KEEP:
815                         goto keep_locked;
816                 case PAGEREF_RECLAIM:
817                 case PAGEREF_RECLAIM_CLEAN:
818                         ; /* try to reclaim the page below */
819                 }
820
821                 /*
822                  * Anonymous process memory has backing store?
823                  * Try to allocate it some swap space here.
824                  */
825                 if (PageAnon(page) && !PageSwapCache(page)) {
826                         if (!(sc->gfp_mask & __GFP_IO))
827                                 goto keep_locked;
828                         if (!add_to_swap(page, page_list))
829                                 goto activate_locked;
830                         may_enter_fs = 1;
831                 }
832
833                 mapping = page_mapping(page);
834
835                 /*
836                  * The page is mapped into the page tables of one or more
837                  * processes. Try to unmap it here.
838                  */
839                 if (page_mapped(page) && mapping) {
840                         switch (try_to_unmap(page, ttu_flags)) {
841                         case SWAP_FAIL:
842                                 goto activate_locked;
843                         case SWAP_AGAIN:
844                                 goto keep_locked;
845                         case SWAP_MLOCK:
846                                 goto cull_mlocked;
847                         case SWAP_SUCCESS:
848                                 ; /* try to free the page below */
849                         }
850                 }
851
852                 if (PageDirty(page)) {
853                         nr_dirty++;
854
855                         /*
856                          * Only kswapd can writeback filesystem pages to
857                          * avoid risk of stack overflow but do not writeback
858                          * unless under significant pressure.
859                          */
860                         if (page_is_file_cache(page) &&
861                                         (!current_is_kswapd() ||
862                                          sc->priority >= DEF_PRIORITY - 2)) {
863                                 /*
864                                  * Immediately reclaim when written back.
865                                  * Similar in principal to deactivate_page()
866                                  * except we already have the page isolated
867                                  * and know it's dirty
868                                  */
869                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
870                                 SetPageReclaim(page);
871
872                                 goto keep_locked;
873                         }
874
875                         if (references == PAGEREF_RECLAIM_CLEAN)
876                                 goto keep_locked;
877                         if (!may_enter_fs)
878                                 goto keep_locked;
879                         if (!sc->may_writepage)
880                                 goto keep_locked;
881
882                         /* Page is dirty, try to write it out here */
883                         switch (pageout(page, mapping, sc)) {
884                         case PAGE_KEEP:
885                                 nr_congested++;
886                                 goto keep_locked;
887                         case PAGE_ACTIVATE:
888                                 goto activate_locked;
889                         case PAGE_SUCCESS:
890                                 if (PageWriteback(page))
891                                         goto keep;
892                                 if (PageDirty(page))
893                                         goto keep;
894
895                                 /*
896                                  * A synchronous write - probably a ramdisk.  Go
897                                  * ahead and try to reclaim the page.
898                                  */
899                                 if (!trylock_page(page))
900                                         goto keep;
901                                 if (PageDirty(page) || PageWriteback(page))
902                                         goto keep_locked;
903                                 mapping = page_mapping(page);
904                         case PAGE_CLEAN:
905                                 ; /* try to free the page below */
906                         }
907                 }
908
909                 /*
910                  * If the page has buffers, try to free the buffer mappings
911                  * associated with this page. If we succeed we try to free
912                  * the page as well.
913                  *
914                  * We do this even if the page is PageDirty().
915                  * try_to_release_page() does not perform I/O, but it is
916                  * possible for a page to have PageDirty set, but it is actually
917                  * clean (all its buffers are clean).  This happens if the
918                  * buffers were written out directly, with submit_bh(). ext3
919                  * will do this, as well as the blockdev mapping.
920                  * try_to_release_page() will discover that cleanness and will
921                  * drop the buffers and mark the page clean - it can be freed.
922                  *
923                  * Rarely, pages can have buffers and no ->mapping.  These are
924                  * the pages which were not successfully invalidated in
925                  * truncate_complete_page().  We try to drop those buffers here
926                  * and if that worked, and the page is no longer mapped into
927                  * process address space (page_count == 1) it can be freed.
928                  * Otherwise, leave the page on the LRU so it is swappable.
929                  */
930                 if (page_has_private(page)) {
931                         if (!try_to_release_page(page, sc->gfp_mask))
932                                 goto activate_locked;
933                         if (!mapping && page_count(page) == 1) {
934                                 unlock_page(page);
935                                 if (put_page_testzero(page))
936                                         goto free_it;
937                                 else {
938                                         /*
939                                          * rare race with speculative reference.
940                                          * the speculative reference will free
941                                          * this page shortly, so we may
942                                          * increment nr_reclaimed here (and
943                                          * leave it off the LRU).
944                                          */
945                                         nr_reclaimed++;
946                                         continue;
947                                 }
948                         }
949                 }
950
951                 if (!mapping || !__remove_mapping(mapping, page))
952                         goto keep_locked;
953
954                 /*
955                  * At this point, we have no other references and there is
956                  * no way to pick any more up (removed from LRU, removed
957                  * from pagecache). Can use non-atomic bitops now (and
958                  * we obviously don't have to worry about waking up a process
959                  * waiting on the page lock, because there are no references.
960                  */
961                 __clear_page_locked(page);
962 free_it:
963                 nr_reclaimed++;
964
965                 /*
966                  * Is there need to periodically free_page_list? It would
967                  * appear not as the counts should be low
968                  */
969                 list_add(&page->lru, &free_pages);
970                 continue;
971
972 cull_mlocked:
973                 if (PageSwapCache(page))
974                         try_to_free_swap(page);
975                 unlock_page(page);
976                 putback_lru_page(page);
977                 continue;
978
979 activate_locked:
980                 /* Not a candidate for swapping, so reclaim swap space. */
981                 if (PageSwapCache(page) && vm_swap_full())
982                         try_to_free_swap(page);
983                 VM_BUG_ON(PageActive(page));
984                 SetPageActive(page);
985                 pgactivate++;
986 keep_locked:
987                 unlock_page(page);
988 keep:
989                 list_add(&page->lru, &ret_pages);
990                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
991         }
992
993         /*
994          * Tag a zone as congested if all the dirty pages encountered were
995          * backed by a congested BDI. In this case, reclaimers should just
996          * back off and wait for congestion to clear because further reclaim
997          * will encounter the same problem
998          */
999         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1000                 zone_set_flag(zone, ZONE_CONGESTED);
1001
1002         free_hot_cold_page_list(&free_pages, 1);
1003
1004         list_splice(&ret_pages, page_list);
1005         count_vm_events(PGACTIVATE, pgactivate);
1006         mem_cgroup_uncharge_end();
1007         *ret_nr_dirty += nr_dirty;
1008         *ret_nr_writeback += nr_writeback;
1009         return nr_reclaimed;
1010 }
1011
1012 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1013                                             struct list_head *page_list)
1014 {
1015         struct scan_control sc = {
1016                 .gfp_mask = GFP_KERNEL,
1017                 .priority = DEF_PRIORITY,
1018                 .may_unmap = 1,
1019         };
1020         unsigned long ret, dummy1, dummy2;
1021         struct page *page, *next;
1022         LIST_HEAD(clean_pages);
1023
1024         list_for_each_entry_safe(page, next, page_list, lru) {
1025                 if (page_is_file_cache(page) && !PageDirty(page)) {
1026                         ClearPageActive(page);
1027                         list_move(&page->lru, &clean_pages);
1028                 }
1029         }
1030
1031         ret = shrink_page_list(&clean_pages, zone, &sc,
1032                                 TTU_UNMAP|TTU_IGNORE_ACCESS,
1033                                 &dummy1, &dummy2, true);
1034         list_splice(&clean_pages, page_list);
1035         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1036         return ret;
1037 }
1038
1039 /*
1040  * Attempt to remove the specified page from its LRU.  Only take this page
1041  * if it is of the appropriate PageActive status.  Pages which are being
1042  * freed elsewhere are also ignored.
1043  *
1044  * page:        page to consider
1045  * mode:        one of the LRU isolation modes defined above
1046  *
1047  * returns 0 on success, -ve errno on failure.
1048  */
1049 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1050 {
1051         int ret = -EINVAL;
1052
1053         /* Only take pages on the LRU. */
1054         if (!PageLRU(page))
1055                 return ret;
1056
1057         /* Compaction should not handle unevictable pages but CMA can do so */
1058         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1059                 return ret;
1060
1061         ret = -EBUSY;
1062
1063         /*
1064          * To minimise LRU disruption, the caller can indicate that it only
1065          * wants to isolate pages it will be able to operate on without
1066          * blocking - clean pages for the most part.
1067          *
1068          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1069          * is used by reclaim when it is cannot write to backing storage
1070          *
1071          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1072          * that it is possible to migrate without blocking
1073          */
1074         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1075                 /* All the caller can do on PageWriteback is block */
1076                 if (PageWriteback(page))
1077                         return ret;
1078
1079                 if (PageDirty(page)) {
1080                         struct address_space *mapping;
1081
1082                         /* ISOLATE_CLEAN means only clean pages */
1083                         if (mode & ISOLATE_CLEAN)
1084                                 return ret;
1085
1086                         /*
1087                          * Only pages without mappings or that have a
1088                          * ->migratepage callback are possible to migrate
1089                          * without blocking
1090                          */
1091                         mapping = page_mapping(page);
1092                         if (mapping && !mapping->a_ops->migratepage)
1093                                 return ret;
1094                 }
1095         }
1096
1097         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1098                 return ret;
1099
1100         if (likely(get_page_unless_zero(page))) {
1101                 /*
1102                  * Be careful not to clear PageLRU until after we're
1103                  * sure the page is not being freed elsewhere -- the
1104                  * page release code relies on it.
1105                  */
1106                 ClearPageLRU(page);
1107                 ret = 0;
1108         }
1109
1110         return ret;
1111 }
1112
1113 /*
1114  * zone->lru_lock is heavily contended.  Some of the functions that
1115  * shrink the lists perform better by taking out a batch of pages
1116  * and working on them outside the LRU lock.
1117  *
1118  * For pagecache intensive workloads, this function is the hottest
1119  * spot in the kernel (apart from copy_*_user functions).
1120  *
1121  * Appropriate locks must be held before calling this function.
1122  *
1123  * @nr_to_scan: The number of pages to look through on the list.
1124  * @lruvec:     The LRU vector to pull pages from.
1125  * @dst:        The temp list to put pages on to.
1126  * @nr_scanned: The number of pages that were scanned.
1127  * @sc:         The scan_control struct for this reclaim session
1128  * @mode:       One of the LRU isolation modes
1129  * @lru:        LRU list id for isolating
1130  *
1131  * returns how many pages were moved onto *@dst.
1132  */
1133 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1134                 struct lruvec *lruvec, struct list_head *dst,
1135                 unsigned long *nr_scanned, struct scan_control *sc,
1136                 isolate_mode_t mode, enum lru_list lru)
1137 {
1138         struct list_head *src = &lruvec->lists[lru];
1139         unsigned long nr_taken = 0;
1140         unsigned long scan;
1141
1142         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1143                 struct page *page;
1144                 int nr_pages;
1145
1146                 page = lru_to_page(src);
1147                 prefetchw_prev_lru_page(page, src, flags);
1148
1149                 VM_BUG_ON(!PageLRU(page));
1150
1151                 switch (__isolate_lru_page(page, mode)) {
1152                 case 0:
1153                         nr_pages = hpage_nr_pages(page);
1154                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1155                         list_move(&page->lru, dst);
1156                         nr_taken += nr_pages;
1157                         break;
1158
1159                 case -EBUSY:
1160                         /* else it is being freed elsewhere */
1161                         list_move(&page->lru, src);
1162                         continue;
1163
1164                 default:
1165                         BUG();
1166                 }
1167         }
1168
1169         *nr_scanned = scan;
1170         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1171                                     nr_taken, mode, is_file_lru(lru));
1172         return nr_taken;
1173 }
1174
1175 /**
1176  * isolate_lru_page - tries to isolate a page from its LRU list
1177  * @page: page to isolate from its LRU list
1178  *
1179  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1180  * vmstat statistic corresponding to whatever LRU list the page was on.
1181  *
1182  * Returns 0 if the page was removed from an LRU list.
1183  * Returns -EBUSY if the page was not on an LRU list.
1184  *
1185  * The returned page will have PageLRU() cleared.  If it was found on
1186  * the active list, it will have PageActive set.  If it was found on
1187  * the unevictable list, it will have the PageUnevictable bit set. That flag
1188  * may need to be cleared by the caller before letting the page go.
1189  *
1190  * The vmstat statistic corresponding to the list on which the page was
1191  * found will be decremented.
1192  *
1193  * Restrictions:
1194  * (1) Must be called with an elevated refcount on the page. This is a
1195  *     fundamentnal difference from isolate_lru_pages (which is called
1196  *     without a stable reference).
1197  * (2) the lru_lock must not be held.
1198  * (3) interrupts must be enabled.
1199  */
1200 int isolate_lru_page(struct page *page)
1201 {
1202         int ret = -EBUSY;
1203
1204         VM_BUG_ON(!page_count(page));
1205
1206         if (PageLRU(page)) {
1207                 struct zone *zone = page_zone(page);
1208                 struct lruvec *lruvec;
1209
1210                 spin_lock_irq(&zone->lru_lock);
1211                 lruvec = mem_cgroup_page_lruvec(page, zone);
1212                 if (PageLRU(page)) {
1213                         int lru = page_lru(page);
1214                         get_page(page);
1215                         ClearPageLRU(page);
1216                         del_page_from_lru_list(page, lruvec, lru);
1217                         ret = 0;
1218                 }
1219                 spin_unlock_irq(&zone->lru_lock);
1220         }
1221         return ret;
1222 }
1223
1224 /*
1225  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1226  * then get resheduled. When there are massive number of tasks doing page
1227  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1228  * the LRU list will go small and be scanned faster than necessary, leading to
1229  * unnecessary swapping, thrashing and OOM.
1230  */
1231 static int too_many_isolated(struct zone *zone, int file,
1232                 struct scan_control *sc)
1233 {
1234         unsigned long inactive, isolated;
1235
1236         if (current_is_kswapd())
1237                 return 0;
1238
1239         if (!global_reclaim(sc))
1240                 return 0;
1241
1242         if (file) {
1243                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1244                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1245         } else {
1246                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1247                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1248         }
1249
1250         /*
1251          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1252          * won't get blocked by normal direct-reclaimers, forming a circular
1253          * deadlock.
1254          */
1255         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1256                 inactive >>= 3;
1257
1258         return isolated > inactive;
1259 }
1260
1261 static noinline_for_stack void
1262 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1263 {
1264         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1265         struct zone *zone = lruvec_zone(lruvec);
1266         LIST_HEAD(pages_to_free);
1267
1268         /*
1269          * Put back any unfreeable pages.
1270          */
1271         while (!list_empty(page_list)) {
1272                 struct page *page = lru_to_page(page_list);
1273                 int lru;
1274
1275                 VM_BUG_ON(PageLRU(page));
1276                 list_del(&page->lru);
1277                 if (unlikely(!page_evictable(page))) {
1278                         spin_unlock_irq(&zone->lru_lock);
1279                         putback_lru_page(page);
1280                         spin_lock_irq(&zone->lru_lock);
1281                         continue;
1282                 }
1283
1284                 lruvec = mem_cgroup_page_lruvec(page, zone);
1285
1286                 SetPageLRU(page);
1287                 lru = page_lru(page);
1288                 add_page_to_lru_list(page, lruvec, lru);
1289
1290                 if (is_active_lru(lru)) {
1291                         int file = is_file_lru(lru);
1292                         int numpages = hpage_nr_pages(page);
1293                         reclaim_stat->recent_rotated[file] += numpages;
1294                 }
1295                 if (put_page_testzero(page)) {
1296                         __ClearPageLRU(page);
1297                         __ClearPageActive(page);
1298                         del_page_from_lru_list(page, lruvec, lru);
1299
1300                         if (unlikely(PageCompound(page))) {
1301                                 spin_unlock_irq(&zone->lru_lock);
1302                                 (*get_compound_page_dtor(page))(page);
1303                                 spin_lock_irq(&zone->lru_lock);
1304                         } else
1305                                 list_add(&page->lru, &pages_to_free);
1306                 }
1307         }
1308
1309         /*
1310          * To save our caller's stack, now use input list for pages to free.
1311          */
1312         list_splice(&pages_to_free, page_list);
1313 }
1314
1315 /*
1316  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1317  * of reclaimed pages
1318  */
1319 static noinline_for_stack unsigned long
1320 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1321                      struct scan_control *sc, enum lru_list lru)
1322 {
1323         LIST_HEAD(page_list);
1324         unsigned long nr_scanned;
1325         unsigned long nr_reclaimed = 0;
1326         unsigned long nr_taken;
1327         unsigned long nr_dirty = 0;
1328         unsigned long nr_writeback = 0;
1329         isolate_mode_t isolate_mode = 0;
1330         int file = is_file_lru(lru);
1331         struct zone *zone = lruvec_zone(lruvec);
1332         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1333
1334         while (unlikely(too_many_isolated(zone, file, sc))) {
1335                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1336
1337                 /* We are about to die and free our memory. Return now. */
1338                 if (fatal_signal_pending(current))
1339                         return SWAP_CLUSTER_MAX;
1340         }
1341
1342         lru_add_drain();
1343
1344         if (!sc->may_unmap)
1345                 isolate_mode |= ISOLATE_UNMAPPED;
1346         if (!sc->may_writepage)
1347                 isolate_mode |= ISOLATE_CLEAN;
1348
1349         spin_lock_irq(&zone->lru_lock);
1350
1351         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1352                                      &nr_scanned, sc, isolate_mode, lru);
1353
1354         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1355         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1356
1357         if (global_reclaim(sc)) {
1358                 zone->pages_scanned += nr_scanned;
1359                 if (current_is_kswapd())
1360                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1361                 else
1362                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1363         }
1364         spin_unlock_irq(&zone->lru_lock);
1365
1366         if (nr_taken == 0)
1367                 return 0;
1368
1369         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1370                                         &nr_dirty, &nr_writeback, false);
1371
1372         spin_lock_irq(&zone->lru_lock);
1373
1374         reclaim_stat->recent_scanned[file] += nr_taken;
1375
1376         if (global_reclaim(sc)) {
1377                 if (current_is_kswapd())
1378                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1379                                                nr_reclaimed);
1380                 else
1381                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1382                                                nr_reclaimed);
1383         }
1384
1385         putback_inactive_pages(lruvec, &page_list);
1386
1387         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1388
1389         spin_unlock_irq(&zone->lru_lock);
1390
1391         free_hot_cold_page_list(&page_list, 1);
1392
1393         /*
1394          * If reclaim is isolating dirty pages under writeback, it implies
1395          * that the long-lived page allocation rate is exceeding the page
1396          * laundering rate. Either the global limits are not being effective
1397          * at throttling processes due to the page distribution throughout
1398          * zones or there is heavy usage of a slow backing device. The
1399          * only option is to throttle from reclaim context which is not ideal
1400          * as there is no guarantee the dirtying process is throttled in the
1401          * same way balance_dirty_pages() manages.
1402          *
1403          * This scales the number of dirty pages that must be under writeback
1404          * before throttling depending on priority. It is a simple backoff
1405          * function that has the most effect in the range DEF_PRIORITY to
1406          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1407          * in trouble and reclaim is considered to be in trouble.
1408          *
1409          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1410          * DEF_PRIORITY-1  50% must be PageWriteback
1411          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1412          * ...
1413          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1414          *                     isolated page is PageWriteback
1415          */
1416         if (nr_writeback && nr_writeback >=
1417                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1418                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1419
1420         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1421                 zone_idx(zone),
1422                 nr_scanned, nr_reclaimed,
1423                 sc->priority,
1424                 trace_shrink_flags(file));
1425         return nr_reclaimed;
1426 }
1427
1428 /*
1429  * This moves pages from the active list to the inactive list.
1430  *
1431  * We move them the other way if the page is referenced by one or more
1432  * processes, from rmap.
1433  *
1434  * If the pages are mostly unmapped, the processing is fast and it is
1435  * appropriate to hold zone->lru_lock across the whole operation.  But if
1436  * the pages are mapped, the processing is slow (page_referenced()) so we
1437  * should drop zone->lru_lock around each page.  It's impossible to balance
1438  * this, so instead we remove the pages from the LRU while processing them.
1439  * It is safe to rely on PG_active against the non-LRU pages in here because
1440  * nobody will play with that bit on a non-LRU page.
1441  *
1442  * The downside is that we have to touch page->_count against each page.
1443  * But we had to alter page->flags anyway.
1444  */
1445
1446 static void move_active_pages_to_lru(struct lruvec *lruvec,
1447                                      struct list_head *list,
1448                                      struct list_head *pages_to_free,
1449                                      enum lru_list lru)
1450 {
1451         struct zone *zone = lruvec_zone(lruvec);
1452         unsigned long pgmoved = 0;
1453         struct page *page;
1454         int nr_pages;
1455
1456         while (!list_empty(list)) {
1457                 page = lru_to_page(list);
1458                 lruvec = mem_cgroup_page_lruvec(page, zone);
1459
1460                 VM_BUG_ON(PageLRU(page));
1461                 SetPageLRU(page);
1462
1463                 nr_pages = hpage_nr_pages(page);
1464                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1465                 list_move(&page->lru, &lruvec->lists[lru]);
1466                 pgmoved += nr_pages;
1467
1468                 if (put_page_testzero(page)) {
1469                         __ClearPageLRU(page);
1470                         __ClearPageActive(page);
1471                         del_page_from_lru_list(page, lruvec, lru);
1472
1473                         if (unlikely(PageCompound(page))) {
1474                                 spin_unlock_irq(&zone->lru_lock);
1475                                 (*get_compound_page_dtor(page))(page);
1476                                 spin_lock_irq(&zone->lru_lock);
1477                         } else
1478                                 list_add(&page->lru, pages_to_free);
1479                 }
1480         }
1481         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1482         if (!is_active_lru(lru))
1483                 __count_vm_events(PGDEACTIVATE, pgmoved);
1484 }
1485
1486 static void shrink_active_list(unsigned long nr_to_scan,
1487                                struct lruvec *lruvec,
1488                                struct scan_control *sc,
1489                                enum lru_list lru)
1490 {
1491         unsigned long nr_taken;
1492         unsigned long nr_scanned;
1493         unsigned long vm_flags;
1494         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1495         LIST_HEAD(l_active);
1496         LIST_HEAD(l_inactive);
1497         struct page *page;
1498         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1499         unsigned long nr_rotated = 0;
1500         isolate_mode_t isolate_mode = 0;
1501         int file = is_file_lru(lru);
1502         struct zone *zone = lruvec_zone(lruvec);
1503
1504         lru_add_drain();
1505
1506         if (!sc->may_unmap)
1507                 isolate_mode |= ISOLATE_UNMAPPED;
1508         if (!sc->may_writepage)
1509                 isolate_mode |= ISOLATE_CLEAN;
1510
1511         spin_lock_irq(&zone->lru_lock);
1512
1513         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1514                                      &nr_scanned, sc, isolate_mode, lru);
1515         if (global_reclaim(sc))
1516                 zone->pages_scanned += nr_scanned;
1517
1518         reclaim_stat->recent_scanned[file] += nr_taken;
1519
1520         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1521         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1522         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1523         spin_unlock_irq(&zone->lru_lock);
1524
1525         while (!list_empty(&l_hold)) {
1526                 cond_resched();
1527                 page = lru_to_page(&l_hold);
1528                 list_del(&page->lru);
1529
1530                 if (unlikely(!page_evictable(page))) {
1531                         putback_lru_page(page);
1532                         continue;
1533                 }
1534
1535                 if (unlikely(buffer_heads_over_limit)) {
1536                         if (page_has_private(page) && trylock_page(page)) {
1537                                 if (page_has_private(page))
1538                                         try_to_release_page(page, 0);
1539                                 unlock_page(page);
1540                         }
1541                 }
1542
1543                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1544                                     &vm_flags)) {
1545                         nr_rotated += hpage_nr_pages(page);
1546                         /*
1547                          * Identify referenced, file-backed active pages and
1548                          * give them one more trip around the active list. So
1549                          * that executable code get better chances to stay in
1550                          * memory under moderate memory pressure.  Anon pages
1551                          * are not likely to be evicted by use-once streaming
1552                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1553                          * so we ignore them here.
1554                          */
1555                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1556                                 list_add(&page->lru, &l_active);
1557                                 continue;
1558                         }
1559                 }
1560
1561                 ClearPageActive(page);  /* we are de-activating */
1562                 list_add(&page->lru, &l_inactive);
1563         }
1564
1565         /*
1566          * Move pages back to the lru list.
1567          */
1568         spin_lock_irq(&zone->lru_lock);
1569         /*
1570          * Count referenced pages from currently used mappings as rotated,
1571          * even though only some of them are actually re-activated.  This
1572          * helps balance scan pressure between file and anonymous pages in
1573          * get_scan_ratio.
1574          */
1575         reclaim_stat->recent_rotated[file] += nr_rotated;
1576
1577         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1578         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1579         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1580         spin_unlock_irq(&zone->lru_lock);
1581
1582         free_hot_cold_page_list(&l_hold, 1);
1583 }
1584
1585 #ifdef CONFIG_SWAP
1586 static int inactive_anon_is_low_global(struct zone *zone)
1587 {
1588         unsigned long active, inactive;
1589
1590         active = zone_page_state(zone, NR_ACTIVE_ANON);
1591         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1592
1593         if (inactive * zone->inactive_ratio < active)
1594                 return 1;
1595
1596         return 0;
1597 }
1598
1599 /**
1600  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1601  * @lruvec: LRU vector to check
1602  *
1603  * Returns true if the zone does not have enough inactive anon pages,
1604  * meaning some active anon pages need to be deactivated.
1605  */
1606 static int inactive_anon_is_low(struct lruvec *lruvec)
1607 {
1608         /*
1609          * If we don't have swap space, anonymous page deactivation
1610          * is pointless.
1611          */
1612         if (!total_swap_pages)
1613                 return 0;
1614
1615         if (!mem_cgroup_disabled())
1616                 return mem_cgroup_inactive_anon_is_low(lruvec);
1617
1618         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1619 }
1620 #else
1621 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1622 {
1623         return 0;
1624 }
1625 #endif
1626
1627 /**
1628  * inactive_file_is_low - check if file pages need to be deactivated
1629  * @lruvec: LRU vector to check
1630  *
1631  * When the system is doing streaming IO, memory pressure here
1632  * ensures that active file pages get deactivated, until more
1633  * than half of the file pages are on the inactive list.
1634  *
1635  * Once we get to that situation, protect the system's working
1636  * set from being evicted by disabling active file page aging.
1637  *
1638  * This uses a different ratio than the anonymous pages, because
1639  * the page cache uses a use-once replacement algorithm.
1640  */
1641 static int inactive_file_is_low(struct lruvec *lruvec)
1642 {
1643         unsigned long inactive;
1644         unsigned long active;
1645
1646         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1647         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1648
1649         return active > inactive;
1650 }
1651
1652 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1653 {
1654         if (is_file_lru(lru))
1655                 return inactive_file_is_low(lruvec);
1656         else
1657                 return inactive_anon_is_low(lruvec);
1658 }
1659
1660 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1661                                  struct lruvec *lruvec, struct scan_control *sc)
1662 {
1663         if (is_active_lru(lru)) {
1664                 if (inactive_list_is_low(lruvec, lru))
1665                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1666                 return 0;
1667         }
1668
1669         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1670 }
1671
1672 static int vmscan_swappiness(struct scan_control *sc)
1673 {
1674         if (global_reclaim(sc))
1675                 return vm_swappiness;
1676         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1677 }
1678
1679 enum scan_balance {
1680         SCAN_EQUAL,
1681         SCAN_FRACT,
1682         SCAN_ANON,
1683         SCAN_FILE,
1684 };
1685
1686 /*
1687  * Determine how aggressively the anon and file LRU lists should be
1688  * scanned.  The relative value of each set of LRU lists is determined
1689  * by looking at the fraction of the pages scanned we did rotate back
1690  * onto the active list instead of evict.
1691  *
1692  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1693  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1694  */
1695 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1696                            unsigned long *nr)
1697 {
1698         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1699         u64 fraction[2];
1700         u64 denominator = 0;    /* gcc */
1701         struct zone *zone = lruvec_zone(lruvec);
1702         unsigned long anon_prio, file_prio;
1703         enum scan_balance scan_balance;
1704         unsigned long anon, file, free;
1705         bool force_scan = false;
1706         unsigned long ap, fp;
1707         enum lru_list lru;
1708
1709         /*
1710          * If the zone or memcg is small, nr[l] can be 0.  This
1711          * results in no scanning on this priority and a potential
1712          * priority drop.  Global direct reclaim can go to the next
1713          * zone and tends to have no problems. Global kswapd is for
1714          * zone balancing and it needs to scan a minimum amount. When
1715          * reclaiming for a memcg, a priority drop can cause high
1716          * latencies, so it's better to scan a minimum amount there as
1717          * well.
1718          */
1719         if (current_is_kswapd() && zone->all_unreclaimable)
1720                 force_scan = true;
1721         if (!global_reclaim(sc))
1722                 force_scan = true;
1723
1724         /* If we have no swap space, do not bother scanning anon pages. */
1725         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1726                 scan_balance = SCAN_FILE;
1727                 goto out;
1728         }
1729
1730         /*
1731          * Global reclaim will swap to prevent OOM even with no
1732          * swappiness, but memcg users want to use this knob to
1733          * disable swapping for individual groups completely when
1734          * using the memory controller's swap limit feature would be
1735          * too expensive.
1736          */
1737         if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1738                 scan_balance = SCAN_FILE;
1739                 goto out;
1740         }
1741
1742         /*
1743          * Do not apply any pressure balancing cleverness when the
1744          * system is close to OOM, scan both anon and file equally
1745          * (unless the swappiness setting disagrees with swapping).
1746          */
1747         if (!sc->priority && vmscan_swappiness(sc)) {
1748                 scan_balance = SCAN_EQUAL;
1749                 goto out;
1750         }
1751
1752         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1753                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1754         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1755                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1756
1757         /*
1758          * If it's foreseeable that reclaiming the file cache won't be
1759          * enough to get the zone back into a desirable shape, we have
1760          * to swap.  Better start now and leave the - probably heavily
1761          * thrashing - remaining file pages alone.
1762          */
1763         if (global_reclaim(sc)) {
1764                 free = zone_page_state(zone, NR_FREE_PAGES);
1765                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1766                         scan_balance = SCAN_ANON;
1767                         goto out;
1768                 }
1769         }
1770
1771         /*
1772          * There is enough inactive page cache, do not reclaim
1773          * anything from the anonymous working set right now.
1774          */
1775         if (!inactive_file_is_low(lruvec)) {
1776                 scan_balance = SCAN_FILE;
1777                 goto out;
1778         }
1779
1780         scan_balance = SCAN_FRACT;
1781
1782         /*
1783          * With swappiness at 100, anonymous and file have the same priority.
1784          * This scanning priority is essentially the inverse of IO cost.
1785          */
1786         anon_prio = vmscan_swappiness(sc);
1787         file_prio = 200 - anon_prio;
1788
1789         /*
1790          * OK, so we have swap space and a fair amount of page cache
1791          * pages.  We use the recently rotated / recently scanned
1792          * ratios to determine how valuable each cache is.
1793          *
1794          * Because workloads change over time (and to avoid overflow)
1795          * we keep these statistics as a floating average, which ends
1796          * up weighing recent references more than old ones.
1797          *
1798          * anon in [0], file in [1]
1799          */
1800         spin_lock_irq(&zone->lru_lock);
1801         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1802                 reclaim_stat->recent_scanned[0] /= 2;
1803                 reclaim_stat->recent_rotated[0] /= 2;
1804         }
1805
1806         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1807                 reclaim_stat->recent_scanned[1] /= 2;
1808                 reclaim_stat->recent_rotated[1] /= 2;
1809         }
1810
1811         /*
1812          * The amount of pressure on anon vs file pages is inversely
1813          * proportional to the fraction of recently scanned pages on
1814          * each list that were recently referenced and in active use.
1815          */
1816         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1817         ap /= reclaim_stat->recent_rotated[0] + 1;
1818
1819         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1820         fp /= reclaim_stat->recent_rotated[1] + 1;
1821         spin_unlock_irq(&zone->lru_lock);
1822
1823         fraction[0] = ap;
1824         fraction[1] = fp;
1825         denominator = ap + fp + 1;
1826 out:
1827         for_each_evictable_lru(lru) {
1828                 int file = is_file_lru(lru);
1829                 unsigned long size;
1830                 unsigned long scan;
1831
1832                 size = get_lru_size(lruvec, lru);
1833                 scan = size >> sc->priority;
1834
1835                 if (!scan && force_scan)
1836                         scan = min(size, SWAP_CLUSTER_MAX);
1837
1838                 switch (scan_balance) {
1839                 case SCAN_EQUAL:
1840                         /* Scan lists relative to size */
1841                         break;
1842                 case SCAN_FRACT:
1843                         /*
1844                          * Scan types proportional to swappiness and
1845                          * their relative recent reclaim efficiency.
1846                          */
1847                         scan = div64_u64(scan * fraction[file], denominator);
1848                         break;
1849                 case SCAN_FILE:
1850                 case SCAN_ANON:
1851                         /* Scan one type exclusively */
1852                         if ((scan_balance == SCAN_FILE) != file)
1853                                 scan = 0;
1854                         break;
1855                 default:
1856                         /* Look ma, no brain */
1857                         BUG();
1858                 }
1859                 nr[lru] = scan;
1860         }
1861 }
1862
1863 /*
1864  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1865  */
1866 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1867 {
1868         unsigned long nr[NR_LRU_LISTS];
1869         unsigned long nr_to_scan;
1870         enum lru_list lru;
1871         unsigned long nr_reclaimed = 0;
1872         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1873         struct blk_plug plug;
1874
1875         get_scan_count(lruvec, sc, nr);
1876
1877         blk_start_plug(&plug);
1878         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1879                                         nr[LRU_INACTIVE_FILE]) {
1880                 for_each_evictable_lru(lru) {
1881                         if (nr[lru]) {
1882                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1883                                 nr[lru] -= nr_to_scan;
1884
1885                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1886                                                             lruvec, sc);
1887                         }
1888                 }
1889                 /*
1890                  * On large memory systems, scan >> priority can become
1891                  * really large. This is fine for the starting priority;
1892                  * we want to put equal scanning pressure on each zone.
1893                  * However, if the VM has a harder time of freeing pages,
1894                  * with multiple processes reclaiming pages, the total
1895                  * freeing target can get unreasonably large.
1896                  */
1897                 if (nr_reclaimed >= nr_to_reclaim &&
1898                     sc->priority < DEF_PRIORITY)
1899                         break;
1900         }
1901         blk_finish_plug(&plug);
1902         sc->nr_reclaimed += nr_reclaimed;
1903
1904         /*
1905          * Even if we did not try to evict anon pages at all, we want to
1906          * rebalance the anon lru active/inactive ratio.
1907          */
1908         if (inactive_anon_is_low(lruvec))
1909                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1910                                    sc, LRU_ACTIVE_ANON);
1911
1912         throttle_vm_writeout(sc->gfp_mask);
1913 }
1914
1915 /* Use reclaim/compaction for costly allocs or under memory pressure */
1916 static bool in_reclaim_compaction(struct scan_control *sc)
1917 {
1918         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1919                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1920                          sc->priority < DEF_PRIORITY - 2))
1921                 return true;
1922
1923         return false;
1924 }
1925
1926 /*
1927  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1928  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1929  * true if more pages should be reclaimed such that when the page allocator
1930  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1931  * It will give up earlier than that if there is difficulty reclaiming pages.
1932  */
1933 static inline bool should_continue_reclaim(struct zone *zone,
1934                                         unsigned long nr_reclaimed,
1935                                         unsigned long nr_scanned,
1936                                         struct scan_control *sc)
1937 {
1938         unsigned long pages_for_compaction;
1939         unsigned long inactive_lru_pages;
1940
1941         /* If not in reclaim/compaction mode, stop */
1942         if (!in_reclaim_compaction(sc))
1943                 return false;
1944
1945         /* Consider stopping depending on scan and reclaim activity */
1946         if (sc->gfp_mask & __GFP_REPEAT) {
1947                 /*
1948                  * For __GFP_REPEAT allocations, stop reclaiming if the
1949                  * full LRU list has been scanned and we are still failing
1950                  * to reclaim pages. This full LRU scan is potentially
1951                  * expensive but a __GFP_REPEAT caller really wants to succeed
1952                  */
1953                 if (!nr_reclaimed && !nr_scanned)
1954                         return false;
1955         } else {
1956                 /*
1957                  * For non-__GFP_REPEAT allocations which can presumably
1958                  * fail without consequence, stop if we failed to reclaim
1959                  * any pages from the last SWAP_CLUSTER_MAX number of
1960                  * pages that were scanned. This will return to the
1961                  * caller faster at the risk reclaim/compaction and
1962                  * the resulting allocation attempt fails
1963                  */
1964                 if (!nr_reclaimed)
1965                         return false;
1966         }
1967
1968         /*
1969          * If we have not reclaimed enough pages for compaction and the
1970          * inactive lists are large enough, continue reclaiming
1971          */
1972         pages_for_compaction = (2UL << sc->order);
1973         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
1974         if (get_nr_swap_pages() > 0)
1975                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
1976         if (sc->nr_reclaimed < pages_for_compaction &&
1977                         inactive_lru_pages > pages_for_compaction)
1978                 return true;
1979
1980         /* If compaction would go ahead or the allocation would succeed, stop */
1981         switch (compaction_suitable(zone, sc->order)) {
1982         case COMPACT_PARTIAL:
1983         case COMPACT_CONTINUE:
1984                 return false;
1985         default:
1986                 return true;
1987         }
1988 }
1989
1990 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1991 {
1992         unsigned long nr_reclaimed, nr_scanned;
1993
1994         do {
1995                 struct mem_cgroup *root = sc->target_mem_cgroup;
1996                 struct mem_cgroup_reclaim_cookie reclaim = {
1997                         .zone = zone,
1998                         .priority = sc->priority,
1999                 };
2000                 struct mem_cgroup *memcg;
2001
2002                 nr_reclaimed = sc->nr_reclaimed;
2003                 nr_scanned = sc->nr_scanned;
2004
2005                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2006                 do {
2007                         struct lruvec *lruvec;
2008
2009                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2010
2011                         shrink_lruvec(lruvec, sc);
2012
2013                         /*
2014                          * Direct reclaim and kswapd have to scan all memory
2015                          * cgroups to fulfill the overall scan target for the
2016                          * zone.
2017                          *
2018                          * Limit reclaim, on the other hand, only cares about
2019                          * nr_to_reclaim pages to be reclaimed and it will
2020                          * retry with decreasing priority if one round over the
2021                          * whole hierarchy is not sufficient.
2022                          */
2023                         if (!global_reclaim(sc) &&
2024                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2025                                 mem_cgroup_iter_break(root, memcg);
2026                                 break;
2027                         }
2028                         memcg = mem_cgroup_iter(root, memcg, &reclaim);
2029                 } while (memcg);
2030
2031                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2032                            sc->nr_scanned - nr_scanned,
2033                            sc->nr_reclaimed - nr_reclaimed);
2034
2035         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2036                                          sc->nr_scanned - nr_scanned, sc));
2037 }
2038
2039 /* Returns true if compaction should go ahead for a high-order request */
2040 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2041 {
2042         unsigned long balance_gap, watermark;
2043         bool watermark_ok;
2044
2045         /* Do not consider compaction for orders reclaim is meant to satisfy */
2046         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2047                 return false;
2048
2049         /*
2050          * Compaction takes time to run and there are potentially other
2051          * callers using the pages just freed. Continue reclaiming until
2052          * there is a buffer of free pages available to give compaction
2053          * a reasonable chance of completing and allocating the page
2054          */
2055         balance_gap = min(low_wmark_pages(zone),
2056                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2057                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2058         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2059         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2060
2061         /*
2062          * If compaction is deferred, reclaim up to a point where
2063          * compaction will have a chance of success when re-enabled
2064          */
2065         if (compaction_deferred(zone, sc->order))
2066                 return watermark_ok;
2067
2068         /* If compaction is not ready to start, keep reclaiming */
2069         if (!compaction_suitable(zone, sc->order))
2070                 return false;
2071
2072         return watermark_ok;
2073 }
2074
2075 /*
2076  * This is the direct reclaim path, for page-allocating processes.  We only
2077  * try to reclaim pages from zones which will satisfy the caller's allocation
2078  * request.
2079  *
2080  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2081  * Because:
2082  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2083  *    allocation or
2084  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2085  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2086  *    zone defense algorithm.
2087  *
2088  * If a zone is deemed to be full of pinned pages then just give it a light
2089  * scan then give up on it.
2090  *
2091  * This function returns true if a zone is being reclaimed for a costly
2092  * high-order allocation and compaction is ready to begin. This indicates to
2093  * the caller that it should consider retrying the allocation instead of
2094  * further reclaim.
2095  */
2096 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2097 {
2098         struct zoneref *z;
2099         struct zone *zone;
2100         unsigned long nr_soft_reclaimed;
2101         unsigned long nr_soft_scanned;
2102         bool aborted_reclaim = false;
2103
2104         /*
2105          * If the number of buffer_heads in the machine exceeds the maximum
2106          * allowed level, force direct reclaim to scan the highmem zone as
2107          * highmem pages could be pinning lowmem pages storing buffer_heads
2108          */
2109         if (buffer_heads_over_limit)
2110                 sc->gfp_mask |= __GFP_HIGHMEM;
2111
2112         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2113                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2114                 if (!populated_zone(zone))
2115                         continue;
2116                 /*
2117                  * Take care memory controller reclaiming has small influence
2118                  * to global LRU.
2119                  */
2120                 if (global_reclaim(sc)) {
2121                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2122                                 continue;
2123                         if (zone->all_unreclaimable &&
2124                                         sc->priority != DEF_PRIORITY)
2125                                 continue;       /* Let kswapd poll it */
2126                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2127                                 /*
2128                                  * If we already have plenty of memory free for
2129                                  * compaction in this zone, don't free any more.
2130                                  * Even though compaction is invoked for any
2131                                  * non-zero order, only frequent costly order
2132                                  * reclamation is disruptive enough to become a
2133                                  * noticeable problem, like transparent huge
2134                                  * page allocations.
2135                                  */
2136                                 if (compaction_ready(zone, sc)) {
2137                                         aborted_reclaim = true;
2138                                         continue;
2139                                 }
2140                         }
2141                         /*
2142                          * This steals pages from memory cgroups over softlimit
2143                          * and returns the number of reclaimed pages and
2144                          * scanned pages. This works for global memory pressure
2145                          * and balancing, not for a memcg's limit.
2146                          */
2147                         nr_soft_scanned = 0;
2148                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2149                                                 sc->order, sc->gfp_mask,
2150                                                 &nr_soft_scanned);
2151                         sc->nr_reclaimed += nr_soft_reclaimed;
2152                         sc->nr_scanned += nr_soft_scanned;
2153                         /* need some check for avoid more shrink_zone() */
2154                 }
2155
2156                 shrink_zone(zone, sc);
2157         }
2158
2159         return aborted_reclaim;
2160 }
2161
2162 static bool zone_reclaimable(struct zone *zone)
2163 {
2164         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2165 }
2166
2167 /* All zones in zonelist are unreclaimable? */
2168 static bool all_unreclaimable(struct zonelist *zonelist,
2169                 struct scan_control *sc)
2170 {
2171         struct zoneref *z;
2172         struct zone *zone;
2173
2174         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2175                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2176                 if (!populated_zone(zone))
2177                         continue;
2178                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2179                         continue;
2180                 if (!zone->all_unreclaimable)
2181                         return false;
2182         }
2183
2184         return true;
2185 }
2186
2187 /*
2188  * This is the main entry point to direct page reclaim.
2189  *
2190  * If a full scan of the inactive list fails to free enough memory then we
2191  * are "out of memory" and something needs to be killed.
2192  *
2193  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2194  * high - the zone may be full of dirty or under-writeback pages, which this
2195  * caller can't do much about.  We kick the writeback threads and take explicit
2196  * naps in the hope that some of these pages can be written.  But if the
2197  * allocating task holds filesystem locks which prevent writeout this might not
2198  * work, and the allocation attempt will fail.
2199  *
2200  * returns:     0, if no pages reclaimed
2201  *              else, the number of pages reclaimed
2202  */
2203 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2204                                         struct scan_control *sc,
2205                                         struct shrink_control *shrink)
2206 {
2207         unsigned long total_scanned = 0;
2208         struct reclaim_state *reclaim_state = current->reclaim_state;
2209         struct zoneref *z;
2210         struct zone *zone;
2211         unsigned long writeback_threshold;
2212         bool aborted_reclaim;
2213
2214         delayacct_freepages_start();
2215
2216         if (global_reclaim(sc))
2217                 count_vm_event(ALLOCSTALL);
2218
2219         do {
2220                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2221                                 sc->priority);
2222                 sc->nr_scanned = 0;
2223                 aborted_reclaim = shrink_zones(zonelist, sc);
2224
2225                 /*
2226                  * Don't shrink slabs when reclaiming memory from
2227                  * over limit cgroups
2228                  */
2229                 if (global_reclaim(sc)) {
2230                         unsigned long lru_pages = 0;
2231                         for_each_zone_zonelist(zone, z, zonelist,
2232                                         gfp_zone(sc->gfp_mask)) {
2233                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2234                                         continue;
2235
2236                                 lru_pages += zone_reclaimable_pages(zone);
2237                         }
2238
2239                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2240                         if (reclaim_state) {
2241                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2242                                 reclaim_state->reclaimed_slab = 0;
2243                         }
2244                 }
2245                 total_scanned += sc->nr_scanned;
2246                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2247                         goto out;
2248
2249                 /*
2250                  * If we're getting trouble reclaiming, start doing
2251                  * writepage even in laptop mode.
2252                  */
2253                 if (sc->priority < DEF_PRIORITY - 2)
2254                         sc->may_writepage = 1;
2255
2256                 /*
2257                  * Try to write back as many pages as we just scanned.  This
2258                  * tends to cause slow streaming writers to write data to the
2259                  * disk smoothly, at the dirtying rate, which is nice.   But
2260                  * that's undesirable in laptop mode, where we *want* lumpy
2261                  * writeout.  So in laptop mode, write out the whole world.
2262                  */
2263                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2264                 if (total_scanned > writeback_threshold) {
2265                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2266                                                 WB_REASON_TRY_TO_FREE_PAGES);
2267                         sc->may_writepage = 1;
2268                 }
2269
2270                 /* Take a nap, wait for some writeback to complete */
2271                 if (!sc->hibernation_mode && sc->nr_scanned &&
2272                     sc->priority < DEF_PRIORITY - 2) {
2273                         struct zone *preferred_zone;
2274
2275                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2276                                                 &cpuset_current_mems_allowed,
2277                                                 &preferred_zone);
2278                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2279                 }
2280         } while (--sc->priority >= 0);
2281
2282 out:
2283         delayacct_freepages_end();
2284
2285         if (sc->nr_reclaimed)
2286                 return sc->nr_reclaimed;
2287
2288         /*
2289          * As hibernation is going on, kswapd is freezed so that it can't mark
2290          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2291          * check.
2292          */
2293         if (oom_killer_disabled)
2294                 return 0;
2295
2296         /* Aborted reclaim to try compaction? don't OOM, then */
2297         if (aborted_reclaim)
2298                 return 1;
2299
2300         /* top priority shrink_zones still had more to do? don't OOM, then */
2301         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2302                 return 1;
2303
2304         return 0;
2305 }
2306
2307 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2308 {
2309         struct zone *zone;
2310         unsigned long pfmemalloc_reserve = 0;
2311         unsigned long free_pages = 0;
2312         int i;
2313         bool wmark_ok;
2314
2315         for (i = 0; i <= ZONE_NORMAL; i++) {
2316                 zone = &pgdat->node_zones[i];
2317                 pfmemalloc_reserve += min_wmark_pages(zone);
2318                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2319         }
2320
2321         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2322
2323         /* kswapd must be awake if processes are being throttled */
2324         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2325                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2326                                                 (enum zone_type)ZONE_NORMAL);
2327                 wake_up_interruptible(&pgdat->kswapd_wait);
2328         }
2329
2330         return wmark_ok;
2331 }
2332
2333 /*
2334  * Throttle direct reclaimers if backing storage is backed by the network
2335  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2336  * depleted. kswapd will continue to make progress and wake the processes
2337  * when the low watermark is reached.
2338  *
2339  * Returns true if a fatal signal was delivered during throttling. If this
2340  * happens, the page allocator should not consider triggering the OOM killer.
2341  */
2342 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2343                                         nodemask_t *nodemask)
2344 {
2345         struct zone *zone;
2346         int high_zoneidx = gfp_zone(gfp_mask);
2347         pg_data_t *pgdat;
2348
2349         /*
2350          * Kernel threads should not be throttled as they may be indirectly
2351          * responsible for cleaning pages necessary for reclaim to make forward
2352          * progress. kjournald for example may enter direct reclaim while
2353          * committing a transaction where throttling it could forcing other
2354          * processes to block on log_wait_commit().
2355          */
2356         if (current->flags & PF_KTHREAD)
2357                 goto out;
2358
2359         /*
2360          * If a fatal signal is pending, this process should not throttle.
2361          * It should return quickly so it can exit and free its memory
2362          */
2363         if (fatal_signal_pending(current))
2364                 goto out;
2365
2366         /* Check if the pfmemalloc reserves are ok */
2367         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2368         pgdat = zone->zone_pgdat;
2369         if (pfmemalloc_watermark_ok(pgdat))
2370                 goto out;
2371
2372         /* Account for the throttling */
2373         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2374
2375         /*
2376          * If the caller cannot enter the filesystem, it's possible that it
2377          * is due to the caller holding an FS lock or performing a journal
2378          * transaction in the case of a filesystem like ext[3|4]. In this case,
2379          * it is not safe to block on pfmemalloc_wait as kswapd could be
2380          * blocked waiting on the same lock. Instead, throttle for up to a
2381          * second before continuing.
2382          */
2383         if (!(gfp_mask & __GFP_FS)) {
2384                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2385                         pfmemalloc_watermark_ok(pgdat), HZ);
2386
2387                 goto check_pending;
2388         }
2389
2390         /* Throttle until kswapd wakes the process */
2391         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2392                 pfmemalloc_watermark_ok(pgdat));
2393
2394 check_pending:
2395         if (fatal_signal_pending(current))
2396                 return true;
2397
2398 out:
2399         return false;
2400 }
2401
2402 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2403                                 gfp_t gfp_mask, nodemask_t *nodemask)
2404 {
2405         unsigned long nr_reclaimed;
2406         struct scan_control sc = {
2407                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2408                 .may_writepage = !laptop_mode,
2409                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2410                 .may_unmap = 1,
2411                 .may_swap = 1,
2412                 .order = order,
2413                 .priority = DEF_PRIORITY,
2414                 .target_mem_cgroup = NULL,
2415                 .nodemask = nodemask,
2416         };
2417         struct shrink_control shrink = {
2418                 .gfp_mask = sc.gfp_mask,
2419         };
2420
2421         /*
2422          * Do not enter reclaim if fatal signal was delivered while throttled.
2423          * 1 is returned so that the page allocator does not OOM kill at this
2424          * point.
2425          */
2426         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2427                 return 1;
2428
2429         trace_mm_vmscan_direct_reclaim_begin(order,
2430                                 sc.may_writepage,
2431                                 gfp_mask);
2432
2433         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2434
2435         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2436
2437         return nr_reclaimed;
2438 }
2439
2440 #ifdef CONFIG_MEMCG
2441
2442 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2443                                                 gfp_t gfp_mask, bool noswap,
2444                                                 struct zone *zone,
2445                                                 unsigned long *nr_scanned)
2446 {
2447         struct scan_control sc = {
2448                 .nr_scanned = 0,
2449                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2450                 .may_writepage = !laptop_mode,
2451                 .may_unmap = 1,
2452                 .may_swap = !noswap,
2453                 .order = 0,
2454                 .priority = 0,
2455                 .target_mem_cgroup = memcg,
2456         };
2457         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2458
2459         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2460                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2461
2462         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2463                                                       sc.may_writepage,
2464                                                       sc.gfp_mask);
2465
2466         /*
2467          * NOTE: Although we can get the priority field, using it
2468          * here is not a good idea, since it limits the pages we can scan.
2469          * if we don't reclaim here, the shrink_zone from balance_pgdat
2470          * will pick up pages from other mem cgroup's as well. We hack
2471          * the priority and make it zero.
2472          */
2473         shrink_lruvec(lruvec, &sc);
2474
2475         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2476
2477         *nr_scanned = sc.nr_scanned;
2478         return sc.nr_reclaimed;
2479 }
2480
2481 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2482                                            gfp_t gfp_mask,
2483                                            bool noswap)
2484 {
2485         struct zonelist *zonelist;
2486         unsigned long nr_reclaimed;
2487         int nid;
2488         struct scan_control sc = {
2489                 .may_writepage = !laptop_mode,
2490                 .may_unmap = 1,
2491                 .may_swap = !noswap,
2492                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2493                 .order = 0,
2494                 .priority = DEF_PRIORITY,
2495                 .target_mem_cgroup = memcg,
2496                 .nodemask = NULL, /* we don't care the placement */
2497                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2498                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2499         };
2500         struct shrink_control shrink = {
2501                 .gfp_mask = sc.gfp_mask,
2502         };
2503
2504         /*
2505          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2506          * take care of from where we get pages. So the node where we start the
2507          * scan does not need to be the current node.
2508          */
2509         nid = mem_cgroup_select_victim_node(memcg);
2510
2511         zonelist = NODE_DATA(nid)->node_zonelists;
2512
2513         trace_mm_vmscan_memcg_reclaim_begin(0,
2514                                             sc.may_writepage,
2515                                             sc.gfp_mask);
2516
2517         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2518
2519         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2520
2521         return nr_reclaimed;
2522 }
2523 #endif
2524
2525 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2526 {
2527         struct mem_cgroup *memcg;
2528
2529         if (!total_swap_pages)
2530                 return;
2531
2532         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2533         do {
2534                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2535
2536                 if (inactive_anon_is_low(lruvec))
2537                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2538                                            sc, LRU_ACTIVE_ANON);
2539
2540                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2541         } while (memcg);
2542 }
2543
2544 static bool zone_balanced(struct zone *zone, int order,
2545                           unsigned long balance_gap, int classzone_idx)
2546 {
2547         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2548                                     balance_gap, classzone_idx, 0))
2549                 return false;
2550
2551         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2552             !compaction_suitable(zone, order))
2553                 return false;
2554
2555         return true;
2556 }
2557
2558 /*
2559  * pgdat_balanced() is used when checking if a node is balanced.
2560  *
2561  * For order-0, all zones must be balanced!
2562  *
2563  * For high-order allocations only zones that meet watermarks and are in a
2564  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2565  * total of balanced pages must be at least 25% of the zones allowed by
2566  * classzone_idx for the node to be considered balanced. Forcing all zones to
2567  * be balanced for high orders can cause excessive reclaim when there are
2568  * imbalanced zones.
2569  * The choice of 25% is due to
2570  *   o a 16M DMA zone that is balanced will not balance a zone on any
2571  *     reasonable sized machine
2572  *   o On all other machines, the top zone must be at least a reasonable
2573  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2574  *     would need to be at least 256M for it to be balance a whole node.
2575  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2576  *     to balance a node on its own. These seemed like reasonable ratios.
2577  */
2578 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2579 {
2580         unsigned long managed_pages = 0;
2581         unsigned long balanced_pages = 0;
2582         int i;
2583
2584         /* Check the watermark levels */
2585         for (i = 0; i <= classzone_idx; i++) {
2586                 struct zone *zone = pgdat->node_zones + i;
2587
2588                 if (!populated_zone(zone))
2589                         continue;
2590
2591                 managed_pages += zone->managed_pages;
2592
2593                 /*
2594                  * A special case here:
2595                  *
2596                  * balance_pgdat() skips over all_unreclaimable after
2597                  * DEF_PRIORITY. Effectively, it considers them balanced so
2598                  * they must be considered balanced here as well!
2599                  */
2600                 if (zone->all_unreclaimable) {
2601                         balanced_pages += zone->managed_pages;
2602                         continue;
2603                 }
2604
2605                 if (zone_balanced(zone, order, 0, i))
2606                         balanced_pages += zone->managed_pages;
2607                 else if (!order)
2608                         return false;
2609         }
2610
2611         if (order)
2612                 return balanced_pages >= (managed_pages >> 2);
2613         else
2614                 return true;
2615 }
2616
2617 /*
2618  * Prepare kswapd for sleeping. This verifies that there are no processes
2619  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2620  *
2621  * Returns true if kswapd is ready to sleep
2622  */
2623 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2624                                         int classzone_idx)
2625 {
2626         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2627         if (remaining)
2628                 return false;
2629
2630         /*
2631          * There is a potential race between when kswapd checks its watermarks
2632          * and a process gets throttled. There is also a potential race if
2633          * processes get throttled, kswapd wakes, a large process exits therby
2634          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2635          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2636          * so wake them now if necessary. If necessary, processes will wake
2637          * kswapd and get throttled again
2638          */
2639         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2640                 wake_up(&pgdat->pfmemalloc_wait);
2641                 return false;
2642         }
2643
2644         return pgdat_balanced(pgdat, order, classzone_idx);
2645 }
2646
2647 /*
2648  * For kswapd, balance_pgdat() will work across all this node's zones until
2649  * they are all at high_wmark_pages(zone).
2650  *
2651  * Returns the final order kswapd was reclaiming at
2652  *
2653  * There is special handling here for zones which are full of pinned pages.
2654  * This can happen if the pages are all mlocked, or if they are all used by
2655  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2656  * What we do is to detect the case where all pages in the zone have been
2657  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2658  * dead and from now on, only perform a short scan.  Basically we're polling
2659  * the zone for when the problem goes away.
2660  *
2661  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2662  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2663  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2664  * lower zones regardless of the number of free pages in the lower zones. This
2665  * interoperates with the page allocator fallback scheme to ensure that aging
2666  * of pages is balanced across the zones.
2667  */
2668 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2669                                                         int *classzone_idx)
2670 {
2671         bool pgdat_is_balanced = false;
2672         int i;
2673         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2674         struct reclaim_state *reclaim_state = current->reclaim_state;
2675         unsigned long nr_soft_reclaimed;
2676         unsigned long nr_soft_scanned;
2677         struct scan_control sc = {
2678                 .gfp_mask = GFP_KERNEL,
2679                 .may_unmap = 1,
2680                 .may_swap = 1,
2681                 /*
2682                  * kswapd doesn't want to be bailed out while reclaim. because
2683                  * we want to put equal scanning pressure on each zone.
2684                  */
2685                 .nr_to_reclaim = ULONG_MAX,
2686                 .order = order,
2687                 .target_mem_cgroup = NULL,
2688         };
2689         struct shrink_control shrink = {
2690                 .gfp_mask = sc.gfp_mask,
2691         };
2692 loop_again:
2693         sc.priority = DEF_PRIORITY;
2694         sc.nr_reclaimed = 0;
2695         sc.may_writepage = !laptop_mode;
2696         count_vm_event(PAGEOUTRUN);
2697
2698         do {
2699                 unsigned long lru_pages = 0;
2700
2701                 /*
2702                  * Scan in the highmem->dma direction for the highest
2703                  * zone which needs scanning
2704                  */
2705                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2706                         struct zone *zone = pgdat->node_zones + i;
2707
2708                         if (!populated_zone(zone))
2709                                 continue;
2710
2711                         if (zone->all_unreclaimable &&
2712                             sc.priority != DEF_PRIORITY)
2713                                 continue;
2714
2715                         /*
2716                          * Do some background aging of the anon list, to give
2717                          * pages a chance to be referenced before reclaiming.
2718                          */
2719                         age_active_anon(zone, &sc);
2720
2721                         /*
2722                          * If the number of buffer_heads in the machine
2723                          * exceeds the maximum allowed level and this node
2724                          * has a highmem zone, force kswapd to reclaim from
2725                          * it to relieve lowmem pressure.
2726                          */
2727                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2728                                 end_zone = i;
2729                                 break;
2730                         }
2731
2732                         if (!zone_balanced(zone, order, 0, 0)) {
2733                                 end_zone = i;
2734                                 break;
2735                         } else {
2736                                 /* If balanced, clear the congested flag */
2737                                 zone_clear_flag(zone, ZONE_CONGESTED);
2738                         }
2739                 }
2740
2741                 if (i < 0) {
2742                         pgdat_is_balanced = true;
2743                         goto out;
2744                 }
2745
2746                 for (i = 0; i <= end_zone; i++) {
2747                         struct zone *zone = pgdat->node_zones + i;
2748
2749                         lru_pages += zone_reclaimable_pages(zone);
2750                 }
2751
2752                 /*
2753                  * Now scan the zone in the dma->highmem direction, stopping
2754                  * at the last zone which needs scanning.
2755                  *
2756                  * We do this because the page allocator works in the opposite
2757                  * direction.  This prevents the page allocator from allocating
2758                  * pages behind kswapd's direction of progress, which would
2759                  * cause too much scanning of the lower zones.
2760                  */
2761                 for (i = 0; i <= end_zone; i++) {
2762                         struct zone *zone = pgdat->node_zones + i;
2763                         int nr_slab, testorder;
2764                         unsigned long balance_gap;
2765
2766                         if (!populated_zone(zone))
2767                                 continue;
2768
2769                         if (zone->all_unreclaimable &&
2770                             sc.priority != DEF_PRIORITY)
2771                                 continue;
2772
2773                         sc.nr_scanned = 0;
2774
2775                         nr_soft_scanned = 0;
2776                         /*
2777                          * Call soft limit reclaim before calling shrink_zone.
2778                          */
2779                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2780                                                         order, sc.gfp_mask,
2781                                                         &nr_soft_scanned);
2782                         sc.nr_reclaimed += nr_soft_reclaimed;
2783
2784                         /*
2785                          * We put equal pressure on every zone, unless
2786                          * one zone has way too many pages free
2787                          * already. The "too many pages" is defined
2788                          * as the high wmark plus a "gap" where the
2789                          * gap is either the low watermark or 1%
2790                          * of the zone, whichever is smaller.
2791                          */
2792                         balance_gap = min(low_wmark_pages(zone),
2793                                 (zone->managed_pages +
2794                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2795                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2796                         /*
2797                          * Kswapd reclaims only single pages with compaction
2798                          * enabled. Trying too hard to reclaim until contiguous
2799                          * free pages have become available can hurt performance
2800                          * by evicting too much useful data from memory.
2801                          * Do not reclaim more than needed for compaction.
2802                          */
2803                         testorder = order;
2804                         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2805                                         compaction_suitable(zone, order) !=
2806                                                 COMPACT_SKIPPED)
2807                                 testorder = 0;
2808
2809                         if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2810                             !zone_balanced(zone, testorder,
2811                                            balance_gap, end_zone)) {
2812                                 shrink_zone(zone, &sc);
2813
2814                                 reclaim_state->reclaimed_slab = 0;
2815                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2816                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2817
2818                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2819                                         zone->all_unreclaimable = 1;
2820                         }
2821
2822                         /*
2823                          * If we're getting trouble reclaiming, start doing
2824                          * writepage even in laptop mode.
2825                          */
2826                         if (sc.priority < DEF_PRIORITY - 2)
2827                                 sc.may_writepage = 1;
2828
2829                         if (zone->all_unreclaimable) {
2830                                 if (end_zone && end_zone == i)
2831                                         end_zone--;
2832                                 continue;
2833                         }
2834
2835                         if (zone_balanced(zone, testorder, 0, end_zone))
2836                                 /*
2837                                  * If a zone reaches its high watermark,
2838                                  * consider it to be no longer congested. It's
2839                                  * possible there are dirty pages backed by
2840                                  * congested BDIs but as pressure is relieved,
2841                                  * speculatively avoid congestion waits
2842                                  */
2843                                 zone_clear_flag(zone, ZONE_CONGESTED);
2844                 }
2845
2846                 /*
2847                  * If the low watermark is met there is no need for processes
2848                  * to be throttled on pfmemalloc_wait as they should not be
2849                  * able to safely make forward progress. Wake them
2850                  */
2851                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2852                                 pfmemalloc_watermark_ok(pgdat))
2853                         wake_up(&pgdat->pfmemalloc_wait);
2854
2855                 if (pgdat_balanced(pgdat, order, *classzone_idx)) {
2856                         pgdat_is_balanced = true;
2857                         break;          /* kswapd: all done */
2858                 }
2859
2860                 /*
2861                  * We do this so kswapd doesn't build up large priorities for
2862                  * example when it is freeing in parallel with allocators. It
2863                  * matches the direct reclaim path behaviour in terms of impact
2864                  * on zone->*_priority.
2865                  */
2866                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2867                         break;
2868         } while (--sc.priority >= 0);
2869
2870 out:
2871         if (!pgdat_is_balanced) {
2872                 cond_resched();
2873
2874                 try_to_freeze();
2875
2876                 /*
2877                  * Fragmentation may mean that the system cannot be
2878                  * rebalanced for high-order allocations in all zones.
2879                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2880                  * it means the zones have been fully scanned and are still
2881                  * not balanced. For high-order allocations, there is
2882                  * little point trying all over again as kswapd may
2883                  * infinite loop.
2884                  *
2885                  * Instead, recheck all watermarks at order-0 as they
2886                  * are the most important. If watermarks are ok, kswapd will go
2887                  * back to sleep. High-order users can still perform direct
2888                  * reclaim if they wish.
2889                  */
2890                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2891                         order = sc.order = 0;
2892
2893                 goto loop_again;
2894         }
2895
2896         /*
2897          * If kswapd was reclaiming at a higher order, it has the option of
2898          * sleeping without all zones being balanced. Before it does, it must
2899          * ensure that the watermarks for order-0 on *all* zones are met and
2900          * that the congestion flags are cleared. The congestion flag must
2901          * be cleared as kswapd is the only mechanism that clears the flag
2902          * and it is potentially going to sleep here.
2903          */
2904         if (order) {
2905                 int zones_need_compaction = 1;
2906
2907                 for (i = 0; i <= end_zone; i++) {
2908                         struct zone *zone = pgdat->node_zones + i;
2909
2910                         if (!populated_zone(zone))
2911                                 continue;
2912
2913                         /* Check if the memory needs to be defragmented. */
2914                         if (zone_watermark_ok(zone, order,
2915                                     low_wmark_pages(zone), *classzone_idx, 0))
2916                                 zones_need_compaction = 0;
2917                 }
2918
2919                 if (zones_need_compaction)
2920                         compact_pgdat(pgdat, order);
2921         }
2922
2923         /*
2924          * Return the order we were reclaiming at so prepare_kswapd_sleep()
2925          * makes a decision on the order we were last reclaiming at. However,
2926          * if another caller entered the allocator slow path while kswapd
2927          * was awake, order will remain at the higher level
2928          */
2929         *classzone_idx = end_zone;
2930         return order;
2931 }
2932
2933 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2934 {
2935         long remaining = 0;
2936         DEFINE_WAIT(wait);
2937
2938         if (freezing(current) || kthread_should_stop())
2939                 return;
2940
2941         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2942
2943         /* Try to sleep for a short interval */
2944         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2945                 remaining = schedule_timeout(HZ/10);
2946                 finish_wait(&pgdat->kswapd_wait, &wait);
2947                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2948         }
2949
2950         /*
2951          * After a short sleep, check if it was a premature sleep. If not, then
2952          * go fully to sleep until explicitly woken up.
2953          */
2954         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2955                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2956
2957                 /*
2958                  * vmstat counters are not perfectly accurate and the estimated
2959                  * value for counters such as NR_FREE_PAGES can deviate from the
2960                  * true value by nr_online_cpus * threshold. To avoid the zone
2961                  * watermarks being breached while under pressure, we reduce the
2962                  * per-cpu vmstat threshold while kswapd is awake and restore
2963                  * them before going back to sleep.
2964                  */
2965                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2966
2967                 /*
2968                  * Compaction records what page blocks it recently failed to
2969                  * isolate pages from and skips them in the future scanning.
2970                  * When kswapd is going to sleep, it is reasonable to assume
2971                  * that pages and compaction may succeed so reset the cache.
2972                  */
2973                 reset_isolation_suitable(pgdat);
2974
2975                 if (!kthread_should_stop())
2976                         schedule();
2977
2978                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2979         } else {
2980                 if (remaining)
2981                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2982                 else
2983                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2984         }
2985         finish_wait(&pgdat->kswapd_wait, &wait);
2986 }
2987
2988 /*
2989  * The background pageout daemon, started as a kernel thread
2990  * from the init process.
2991  *
2992  * This basically trickles out pages so that we have _some_
2993  * free memory available even if there is no other activity
2994  * that frees anything up. This is needed for things like routing
2995  * etc, where we otherwise might have all activity going on in
2996  * asynchronous contexts that cannot page things out.
2997  *
2998  * If there are applications that are active memory-allocators
2999  * (most normal use), this basically shouldn't matter.
3000  */
3001 static int kswapd(void *p)
3002 {
3003         unsigned long order, new_order;
3004         unsigned balanced_order;
3005         int classzone_idx, new_classzone_idx;
3006         int balanced_classzone_idx;
3007         pg_data_t *pgdat = (pg_data_t*)p;
3008         struct task_struct *tsk = current;
3009
3010         struct reclaim_state reclaim_state = {
3011                 .reclaimed_slab = 0,
3012         };
3013         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3014
3015         lockdep_set_current_reclaim_state(GFP_KERNEL);
3016
3017         if (!cpumask_empty(cpumask))
3018                 set_cpus_allowed_ptr(tsk, cpumask);
3019         current->reclaim_state = &reclaim_state;
3020
3021         /*
3022          * Tell the memory management that we're a "memory allocator",
3023          * and that if we need more memory we should get access to it
3024          * regardless (see "__alloc_pages()"). "kswapd" should
3025          * never get caught in the normal page freeing logic.
3026          *
3027          * (Kswapd normally doesn't need memory anyway, but sometimes
3028          * you need a small amount of memory in order to be able to
3029          * page out something else, and this flag essentially protects
3030          * us from recursively trying to free more memory as we're
3031          * trying to free the first piece of memory in the first place).
3032          */
3033         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3034         set_freezable();
3035
3036         order = new_order = 0;
3037         balanced_order = 0;
3038         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3039         balanced_classzone_idx = classzone_idx;
3040         for ( ; ; ) {
3041                 bool ret;
3042
3043                 /*
3044                  * If the last balance_pgdat was unsuccessful it's unlikely a
3045                  * new request of a similar or harder type will succeed soon
3046                  * so consider going to sleep on the basis we reclaimed at
3047                  */
3048                 if (balanced_classzone_idx >= new_classzone_idx &&
3049                                         balanced_order == new_order) {
3050                         new_order = pgdat->kswapd_max_order;
3051                         new_classzone_idx = pgdat->classzone_idx;
3052                         pgdat->kswapd_max_order =  0;
3053                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3054                 }
3055
3056                 if (order < new_order || classzone_idx > new_classzone_idx) {
3057                         /*
3058                          * Don't sleep if someone wants a larger 'order'
3059                          * allocation or has tigher zone constraints
3060                          */
3061                         order = new_order;
3062                         classzone_idx = new_classzone_idx;
3063                 } else {
3064                         kswapd_try_to_sleep(pgdat, balanced_order,
3065                                                 balanced_classzone_idx);
3066                         order = pgdat->kswapd_max_order;
3067                         classzone_idx = pgdat->classzone_idx;
3068                         new_order = order;
3069                         new_classzone_idx = classzone_idx;
3070                         pgdat->kswapd_max_order = 0;
3071                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3072                 }
3073
3074                 ret = try_to_freeze();
3075                 if (kthread_should_stop())
3076                         break;
3077
3078                 /*
3079                  * We can speed up thawing tasks if we don't call balance_pgdat
3080                  * after returning from the refrigerator
3081                  */
3082                 if (!ret) {
3083                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3084                         balanced_classzone_idx = classzone_idx;
3085                         balanced_order = balance_pgdat(pgdat, order,
3086                                                 &balanced_classzone_idx);
3087                 }
3088         }
3089
3090         current->reclaim_state = NULL;
3091         return 0;
3092 }
3093
3094 /*
3095  * A zone is low on free memory, so wake its kswapd task to service it.
3096  */
3097 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3098 {
3099         pg_data_t *pgdat;
3100
3101         if (!populated_zone(zone))
3102                 return;
3103
3104         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3105                 return;
3106         pgdat = zone->zone_pgdat;
3107         if (pgdat->kswapd_max_order < order) {
3108                 pgdat->kswapd_max_order = order;
3109                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3110         }
3111         if (!waitqueue_active(&pgdat->kswapd_wait))
3112                 return;
3113         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3114                 return;
3115
3116         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3117         wake_up_interruptible(&pgdat->kswapd_wait);
3118 }
3119
3120 /*
3121  * The reclaimable count would be mostly accurate.
3122  * The less reclaimable pages may be
3123  * - mlocked pages, which will be moved to unevictable list when encountered
3124  * - mapped pages, which may require several travels to be reclaimed
3125  * - dirty pages, which is not "instantly" reclaimable
3126  */
3127 unsigned long global_reclaimable_pages(void)
3128 {
3129         int nr;
3130
3131         nr = global_page_state(NR_ACTIVE_FILE) +
3132              global_page_state(NR_INACTIVE_FILE);
3133
3134         if (get_nr_swap_pages() > 0)
3135                 nr += global_page_state(NR_ACTIVE_ANON) +
3136                       global_page_state(NR_INACTIVE_ANON);
3137
3138         return nr;
3139 }
3140
3141 unsigned long zone_reclaimable_pages(struct zone *zone)
3142 {
3143         int nr;
3144
3145         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3146              zone_page_state(zone, NR_INACTIVE_FILE);
3147
3148         if (get_nr_swap_pages() > 0)
3149                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3150                       zone_page_state(zone, NR_INACTIVE_ANON);
3151
3152         return nr;
3153 }
3154
3155 #ifdef CONFIG_HIBERNATION
3156 /*
3157  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3158  * freed pages.
3159  *
3160  * Rather than trying to age LRUs the aim is to preserve the overall
3161  * LRU order by reclaiming preferentially
3162  * inactive > active > active referenced > active mapped
3163  */
3164 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3165 {
3166         struct reclaim_state reclaim_state;
3167         struct scan_control sc = {
3168                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3169                 .may_swap = 1,
3170                 .may_unmap = 1,
3171                 .may_writepage = 1,
3172                 .nr_to_reclaim = nr_to_reclaim,
3173                 .hibernation_mode = 1,
3174                 .order = 0,
3175                 .priority = DEF_PRIORITY,
3176         };
3177         struct shrink_control shrink = {
3178                 .gfp_mask = sc.gfp_mask,
3179         };
3180         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3181         struct task_struct *p = current;
3182         unsigned long nr_reclaimed;
3183
3184         p->flags |= PF_MEMALLOC;
3185         lockdep_set_current_reclaim_state(sc.gfp_mask);
3186         reclaim_state.reclaimed_slab = 0;
3187         p->reclaim_state = &reclaim_state;
3188
3189         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3190
3191         p->reclaim_state = NULL;
3192         lockdep_clear_current_reclaim_state();
3193         p->flags &= ~PF_MEMALLOC;
3194
3195         return nr_reclaimed;
3196 }
3197 #endif /* CONFIG_HIBERNATION */
3198
3199 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3200    not required for correctness.  So if the last cpu in a node goes
3201    away, we get changed to run anywhere: as the first one comes back,
3202    restore their cpu bindings. */
3203 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3204                         void *hcpu)
3205 {
3206         int nid;
3207
3208         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3209                 for_each_node_state(nid, N_MEMORY) {
3210                         pg_data_t *pgdat = NODE_DATA(nid);
3211                         const struct cpumask *mask;
3212
3213                         mask = cpumask_of_node(pgdat->node_id);
3214
3215                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3216                                 /* One of our CPUs online: restore mask */
3217                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3218                 }
3219         }
3220         return NOTIFY_OK;
3221 }
3222
3223 /*
3224  * This kswapd start function will be called by init and node-hot-add.
3225  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3226  */
3227 int kswapd_run(int nid)
3228 {
3229         pg_data_t *pgdat = NODE_DATA(nid);
3230         int ret = 0;
3231
3232         if (pgdat->kswapd)
3233                 return 0;
3234
3235         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3236         if (IS_ERR(pgdat->kswapd)) {
3237                 /* failure at boot is fatal */
3238                 BUG_ON(system_state == SYSTEM_BOOTING);
3239                 pr_err("Failed to start kswapd on node %d\n", nid);
3240                 ret = PTR_ERR(pgdat->kswapd);
3241                 pgdat->kswapd = NULL;
3242         }
3243         return ret;
3244 }
3245
3246 /*
3247  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3248  * hold lock_memory_hotplug().
3249  */
3250 void kswapd_stop(int nid)
3251 {
3252         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3253
3254         if (kswapd) {
3255                 kthread_stop(kswapd);
3256                 NODE_DATA(nid)->kswapd = NULL;
3257         }
3258 }
3259
3260 static int __init kswapd_init(void)
3261 {
3262         int nid;
3263
3264         swap_setup();
3265         for_each_node_state(nid, N_MEMORY)
3266                 kswapd_run(nid);
3267         hotcpu_notifier(cpu_callback, 0);
3268         return 0;
3269 }
3270
3271 module_init(kswapd_init)
3272
3273 #ifdef CONFIG_NUMA
3274 /*
3275  * Zone reclaim mode
3276  *
3277  * If non-zero call zone_reclaim when the number of free pages falls below
3278  * the watermarks.
3279  */
3280 int zone_reclaim_mode __read_mostly;
3281
3282 #define RECLAIM_OFF 0
3283 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3284 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3285 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3286
3287 /*
3288  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3289  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3290  * a zone.
3291  */
3292 #define ZONE_RECLAIM_PRIORITY 4
3293
3294 /*
3295  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3296  * occur.
3297  */
3298 int sysctl_min_unmapped_ratio = 1;
3299
3300 /*
3301  * If the number of slab pages in a zone grows beyond this percentage then
3302  * slab reclaim needs to occur.
3303  */
3304 int sysctl_min_slab_ratio = 5;
3305
3306 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3307 {
3308         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3309         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3310                 zone_page_state(zone, NR_ACTIVE_FILE);
3311
3312         /*
3313          * It's possible for there to be more file mapped pages than
3314          * accounted for by the pages on the file LRU lists because
3315          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3316          */
3317         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3318 }
3319
3320 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3321 static long zone_pagecache_reclaimable(struct zone *zone)
3322 {
3323         long nr_pagecache_reclaimable;
3324         long delta = 0;
3325
3326         /*
3327          * If RECLAIM_SWAP is set, then all file pages are considered
3328          * potentially reclaimable. Otherwise, we have to worry about
3329          * pages like swapcache and zone_unmapped_file_pages() provides
3330          * a better estimate
3331          */
3332         if (zone_reclaim_mode & RECLAIM_SWAP)
3333                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3334         else
3335                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3336
3337         /* If we can't clean pages, remove dirty pages from consideration */
3338         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3339                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3340
3341         /* Watch for any possible underflows due to delta */
3342         if (unlikely(delta > nr_pagecache_reclaimable))
3343                 delta = nr_pagecache_reclaimable;
3344
3345         return nr_pagecache_reclaimable - delta;
3346 }
3347
3348 /*
3349  * Try to free up some pages from this zone through reclaim.
3350  */
3351 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3352 {
3353         /* Minimum pages needed in order to stay on node */
3354         const unsigned long nr_pages = 1 << order;
3355         struct task_struct *p = current;
3356         struct reclaim_state reclaim_state;
3357         struct scan_control sc = {
3358                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3359                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3360                 .may_swap = 1,
3361                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3362                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3363                 .order = order,
3364                 .priority = ZONE_RECLAIM_PRIORITY,
3365         };
3366         struct shrink_control shrink = {
3367                 .gfp_mask = sc.gfp_mask,
3368         };
3369         unsigned long nr_slab_pages0, nr_slab_pages1;
3370
3371         cond_resched();
3372         /*
3373          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3374          * and we also need to be able to write out pages for RECLAIM_WRITE
3375          * and RECLAIM_SWAP.
3376          */
3377         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3378         lockdep_set_current_reclaim_state(gfp_mask);
3379         reclaim_state.reclaimed_slab = 0;
3380         p->reclaim_state = &reclaim_state;
3381
3382         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3383                 /*
3384                  * Free memory by calling shrink zone with increasing
3385                  * priorities until we have enough memory freed.
3386                  */
3387                 do {
3388                         shrink_zone(zone, &sc);
3389                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3390         }
3391
3392         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3393         if (nr_slab_pages0 > zone->min_slab_pages) {
3394                 /*
3395                  * shrink_slab() does not currently allow us to determine how
3396                  * many pages were freed in this zone. So we take the current
3397                  * number of slab pages and shake the slab until it is reduced
3398                  * by the same nr_pages that we used for reclaiming unmapped
3399                  * pages.
3400                  *
3401                  * Note that shrink_slab will free memory on all zones and may
3402                  * take a long time.
3403                  */
3404                 for (;;) {
3405                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3406
3407                         /* No reclaimable slab or very low memory pressure */
3408                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3409                                 break;
3410
3411                         /* Freed enough memory */
3412                         nr_slab_pages1 = zone_page_state(zone,
3413                                                         NR_SLAB_RECLAIMABLE);
3414                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3415                                 break;
3416                 }
3417
3418                 /*
3419                  * Update nr_reclaimed by the number of slab pages we
3420                  * reclaimed from this zone.
3421                  */
3422                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3423                 if (nr_slab_pages1 < nr_slab_pages0)
3424                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3425         }
3426
3427         p->reclaim_state = NULL;
3428         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3429         lockdep_clear_current_reclaim_state();
3430         return sc.nr_reclaimed >= nr_pages;
3431 }
3432
3433 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3434 {
3435         int node_id;
3436         int ret;
3437
3438         /*
3439          * Zone reclaim reclaims unmapped file backed pages and
3440          * slab pages if we are over the defined limits.
3441          *
3442          * A small portion of unmapped file backed pages is needed for
3443          * file I/O otherwise pages read by file I/O will be immediately
3444          * thrown out if the zone is overallocated. So we do not reclaim
3445          * if less than a specified percentage of the zone is used by
3446          * unmapped file backed pages.
3447          */
3448         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3449             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3450                 return ZONE_RECLAIM_FULL;
3451
3452         if (zone->all_unreclaimable)
3453                 return ZONE_RECLAIM_FULL;
3454
3455         /*
3456          * Do not scan if the allocation should not be delayed.
3457          */
3458         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3459                 return ZONE_RECLAIM_NOSCAN;
3460
3461         /*
3462          * Only run zone reclaim on the local zone or on zones that do not
3463          * have associated processors. This will favor the local processor
3464          * over remote processors and spread off node memory allocations
3465          * as wide as possible.
3466          */
3467         node_id = zone_to_nid(zone);
3468         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3469                 return ZONE_RECLAIM_NOSCAN;
3470
3471         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3472                 return ZONE_RECLAIM_NOSCAN;
3473
3474         ret = __zone_reclaim(zone, gfp_mask, order);
3475         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3476
3477         if (!ret)
3478                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3479
3480         return ret;
3481 }
3482 #endif
3483
3484 /*
3485  * page_evictable - test whether a page is evictable
3486  * @page: the page to test
3487  *
3488  * Test whether page is evictable--i.e., should be placed on active/inactive
3489  * lists vs unevictable list.
3490  *
3491  * Reasons page might not be evictable:
3492  * (1) page's mapping marked unevictable
3493  * (2) page is part of an mlocked VMA
3494  *
3495  */
3496 int page_evictable(struct page *page)
3497 {
3498         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3499 }
3500
3501 #ifdef CONFIG_SHMEM
3502 /**
3503  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3504  * @pages:      array of pages to check
3505  * @nr_pages:   number of pages to check
3506  *
3507  * Checks pages for evictability and moves them to the appropriate lru list.
3508  *
3509  * This function is only used for SysV IPC SHM_UNLOCK.
3510  */
3511 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3512 {
3513         struct lruvec *lruvec;
3514         struct zone *zone = NULL;
3515         int pgscanned = 0;
3516         int pgrescued = 0;
3517         int i;
3518
3519         for (i = 0; i < nr_pages; i++) {
3520                 struct page *page = pages[i];
3521                 struct zone *pagezone;
3522
3523                 pgscanned++;
3524                 pagezone = page_zone(page);
3525                 if (pagezone != zone) {
3526                         if (zone)
3527                                 spin_unlock_irq(&zone->lru_lock);
3528                         zone = pagezone;
3529                         spin_lock_irq(&zone->lru_lock);
3530                 }
3531                 lruvec = mem_cgroup_page_lruvec(page, zone);
3532
3533                 if (!PageLRU(page) || !PageUnevictable(page))
3534                         continue;
3535
3536                 if (page_evictable(page)) {
3537                         enum lru_list lru = page_lru_base_type(page);
3538
3539                         VM_BUG_ON(PageActive(page));
3540                         ClearPageUnevictable(page);
3541                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3542                         add_page_to_lru_list(page, lruvec, lru);
3543                         pgrescued++;
3544                 }
3545         }
3546
3547         if (zone) {
3548                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3549                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3550                 spin_unlock_irq(&zone->lru_lock);
3551         }
3552 }
3553 #endif /* CONFIG_SHMEM */
3554
3555 static void warn_scan_unevictable_pages(void)
3556 {
3557         printk_once(KERN_WARNING
3558                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3559                     "disabled for lack of a legitimate use case.  If you have "
3560                     "one, please send an email to linux-mm@kvack.org.\n",
3561                     current->comm);
3562 }
3563
3564 /*
3565  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3566  * all nodes' unevictable lists for evictable pages
3567  */
3568 unsigned long scan_unevictable_pages;
3569
3570 int scan_unevictable_handler(struct ctl_table *table, int write,
3571                            void __user *buffer,
3572                            size_t *length, loff_t *ppos)
3573 {
3574         warn_scan_unevictable_pages();
3575         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3576         scan_unevictable_pages = 0;
3577         return 0;
3578 }
3579
3580 #ifdef CONFIG_NUMA
3581 /*
3582  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3583  * a specified node's per zone unevictable lists for evictable pages.
3584  */
3585
3586 static ssize_t read_scan_unevictable_node(struct device *dev,
3587                                           struct device_attribute *attr,
3588                                           char *buf)
3589 {
3590         warn_scan_unevictable_pages();
3591         return sprintf(buf, "0\n");     /* always zero; should fit... */
3592 }
3593
3594 static ssize_t write_scan_unevictable_node(struct device *dev,
3595                                            struct device_attribute *attr,
3596                                         const char *buf, size_t count)
3597 {
3598         warn_scan_unevictable_pages();
3599         return 1;
3600 }
3601
3602
3603 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3604                         read_scan_unevictable_node,
3605                         write_scan_unevictable_node);
3606
3607 int scan_unevictable_register_node(struct node *node)
3608 {
3609         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3610 }
3611
3612 void scan_unevictable_unregister_node(struct node *node)
3613 {
3614         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3615 }
3616 #endif