~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/vmscan.c

Version: ~ [ linux-5.11-rc3 ] ~ [ linux-5.10.7 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.89 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.167 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.215 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.251 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.251 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  *  linux/mm/vmscan.c
  3  *
  4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  5  *
  6  *  Swap reorganised 29.12.95, Stephen Tweedie.
  7  *  kswapd added: 7.1.96  sct
  8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
  9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
 10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
 11  *  Multiqueue VM started 5.8.00, Rik van Riel.
 12  */
 13 
 14 #include <linux/mm.h>
 15 #include <linux/module.h>
 16 #include <linux/gfp.h>
 17 #include <linux/kernel_stat.h>
 18 #include <linux/swap.h>
 19 #include <linux/pagemap.h>
 20 #include <linux/init.h>
 21 #include <linux/highmem.h>
 22 #include <linux/vmpressure.h>
 23 #include <linux/vmstat.h>
 24 #include <linux/file.h>
 25 #include <linux/writeback.h>
 26 #include <linux/blkdev.h>
 27 #include <linux/buffer_head.h>  /* for try_to_release_page(),
 28                                         buffer_heads_over_limit */
 29 #include <linux/mm_inline.h>
 30 #include <linux/backing-dev.h>
 31 #include <linux/rmap.h>
 32 #include <linux/topology.h>
 33 #include <linux/cpu.h>
 34 #include <linux/cpuset.h>
 35 #include <linux/compaction.h>
 36 #include <linux/notifier.h>
 37 #include <linux/rwsem.h>
 38 #include <linux/delay.h>
 39 #include <linux/kthread.h>
 40 #include <linux/freezer.h>
 41 #include <linux/memcontrol.h>
 42 #include <linux/delayacct.h>
 43 #include <linux/sysctl.h>
 44 #include <linux/oom.h>
 45 #include <linux/prefetch.h>
 46 
 47 #include <asm/tlbflush.h>
 48 #include <asm/div64.h>
 49 
 50 #include <linux/swapops.h>
 51 #include <linux/balloon_compaction.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 /*
159  * Add a shrinker callback to be called from the vm
160  */
161 void register_shrinker(struct shrinker *shrinker)
162 {
163         atomic_long_set(&shrinker->nr_in_batch, 0);
164         down_write(&shrinker_rwsem);
165         list_add_tail(&shrinker->list, &shrinker_list);
166         up_write(&shrinker_rwsem);
167 }
168 EXPORT_SYMBOL(register_shrinker);
169 
170 /*
171  * Remove one
172  */
173 void unregister_shrinker(struct shrinker *shrinker)
174 {
175         down_write(&shrinker_rwsem);
176         list_del(&shrinker->list);
177         up_write(&shrinker_rwsem);
178 }
179 EXPORT_SYMBOL(unregister_shrinker);
180 
181 static inline int do_shrinker_shrink(struct shrinker *shrinker,
182                                      struct shrink_control *sc,
183                                      unsigned long nr_to_scan)
184 {
185         sc->nr_to_scan = nr_to_scan;
186         return (*shrinker->shrink)(shrinker, sc);
187 }
188 
189 #define SHRINK_BATCH 128
190 /*
191  * Call the shrink functions to age shrinkable caches
192  *
193  * Here we assume it costs one seek to replace a lru page and that it also
194  * takes a seek to recreate a cache object.  With this in mind we age equal
195  * percentages of the lru and ageable caches.  This should balance the seeks
196  * generated by these structures.
197  *
198  * If the vm encountered mapped pages on the LRU it increase the pressure on
199  * slab to avoid swapping.
200  *
201  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
202  *
203  * `lru_pages' represents the number of on-LRU pages in all the zones which
204  * are eligible for the caller's allocation attempt.  It is used for balancing
205  * slab reclaim versus page reclaim.
206  *
207  * Returns the number of slab objects which we shrunk.
208  */
209 unsigned long shrink_slab(struct shrink_control *shrink,
210                           unsigned long nr_pages_scanned,
211                           unsigned long lru_pages)
212 {
213         struct shrinker *shrinker;
214         unsigned long ret = 0;
215 
216         if (nr_pages_scanned == 0)
217                 nr_pages_scanned = SWAP_CLUSTER_MAX;
218 
219         if (!down_read_trylock(&shrinker_rwsem)) {
220                 /* Assume we'll be able to shrink next time */
221                 ret = 1;
222                 goto out;
223         }
224 
225         list_for_each_entry(shrinker, &shrinker_list, list) {
226                 unsigned long long delta;
227                 long total_scan;
228                 long max_pass;
229                 int shrink_ret = 0;
230                 long nr;
231                 long new_nr;
232                 long batch_size = shrinker->batch ? shrinker->batch
233                                                   : SHRINK_BATCH;
234 
235                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
236                 if (max_pass <= 0)
237                         continue;
238 
239                 /*
240                  * copy the current shrinker scan count into a local variable
241                  * and zero it so that other concurrent shrinker invocations
242                  * don't also do this scanning work.
243                  */
244                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
245 
246                 total_scan = nr;
247                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
248                 delta *= max_pass;
249                 do_div(delta, lru_pages + 1);
250                 total_scan += delta;
251                 if (total_scan < 0) {
252                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
253                                "delete nr=%ld\n",
254                                shrinker->shrink, total_scan);
255                         total_scan = max_pass;
256                 }
257 
258                 /*
259                  * We need to avoid excessive windup on filesystem shrinkers
260                  * due to large numbers of GFP_NOFS allocations causing the
261                  * shrinkers to return -1 all the time. This results in a large
262                  * nr being built up so when a shrink that can do some work
263                  * comes along it empties the entire cache due to nr >>>
264                  * max_pass.  This is bad for sustaining a working set in
265                  * memory.
266                  *
267                  * Hence only allow the shrinker to scan the entire cache when
268                  * a large delta change is calculated directly.
269                  */
270                 if (delta < max_pass / 4)
271                         total_scan = min(total_scan, max_pass / 2);
272 
273                 /*
274                  * Avoid risking looping forever due to too large nr value:
275                  * never try to free more than twice the estimate number of
276                  * freeable entries.
277                  */
278                 if (total_scan > max_pass * 2)
279                         total_scan = max_pass * 2;
280 
281                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
282                                         nr_pages_scanned, lru_pages,
283                                         max_pass, delta, total_scan);
284 
285                 while (total_scan >= batch_size) {
286                         int nr_before;
287 
288                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
289                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
290                                                         batch_size);
291                         if (shrink_ret == -1)
292                                 break;
293                         if (shrink_ret < nr_before)
294                                 ret += nr_before - shrink_ret;
295                         count_vm_events(SLABS_SCANNED, batch_size);
296                         total_scan -= batch_size;
297 
298                         cond_resched();
299                 }
300 
301                 /*
302                  * move the unused scan count back into the shrinker in a
303                  * manner that handles concurrent updates. If we exhausted the
304                  * scan, there is no need to do an update.
305                  */
306                 if (total_scan > 0)
307                         new_nr = atomic_long_add_return(total_scan,
308                                         &shrinker->nr_in_batch);
309                 else
310                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
311 
312                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
313         }
314         up_read(&shrinker_rwsem);
315 out:
316         cond_resched();
317         return ret;
318 }
319 
320 static inline int is_page_cache_freeable(struct page *page)
321 {
322         /*
323          * A freeable page cache page is referenced only by the caller
324          * that isolated the page, the page cache radix tree and
325          * optional buffer heads at page->private.
326          */
327         return page_count(page) - page_has_private(page) == 2;
328 }
329 
330 static int may_write_to_queue(struct backing_dev_info *bdi,
331                               struct scan_control *sc)
332 {
333         if (current->flags & PF_SWAPWRITE)
334                 return 1;
335         if (!bdi_write_congested(bdi))
336                 return 1;
337         if (bdi == current->backing_dev_info)
338                 return 1;
339         return 0;
340 }
341 
342 /*
343  * We detected a synchronous write error writing a page out.  Probably
344  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
345  * fsync(), msync() or close().
346  *
347  * The tricky part is that after writepage we cannot touch the mapping: nothing
348  * prevents it from being freed up.  But we have a ref on the page and once
349  * that page is locked, the mapping is pinned.
350  *
351  * We're allowed to run sleeping lock_page() here because we know the caller has
352  * __GFP_FS.
353  */
354 static void handle_write_error(struct address_space *mapping,
355                                 struct page *page, int error)
356 {
357         lock_page(page);
358         if (page_mapping(page) == mapping)
359                 mapping_set_error(mapping, error);
360         unlock_page(page);
361 }
362 
363 /* possible outcome of pageout() */
364 typedef enum {
365         /* failed to write page out, page is locked */
366         PAGE_KEEP,
367         /* move page to the active list, page is locked */
368         PAGE_ACTIVATE,
369         /* page has been sent to the disk successfully, page is unlocked */
370         PAGE_SUCCESS,
371         /* page is clean and locked */
372         PAGE_CLEAN,
373 } pageout_t;
374 
375 /*
376  * pageout is called by shrink_page_list() for each dirty page.
377  * Calls ->writepage().
378  */
379 static pageout_t pageout(struct page *page, struct address_space *mapping,
380                          struct scan_control *sc)
381 {
382         /*
383          * If the page is dirty, only perform writeback if that write
384          * will be non-blocking.  To prevent this allocation from being
385          * stalled by pagecache activity.  But note that there may be
386          * stalls if we need to run get_block().  We could test
387          * PagePrivate for that.
388          *
389          * If this process is currently in __generic_file_aio_write() against
390          * this page's queue, we can perform writeback even if that
391          * will block.
392          *
393          * If the page is swapcache, write it back even if that would
394          * block, for some throttling. This happens by accident, because
395          * swap_backing_dev_info is bust: it doesn't reflect the
396          * congestion state of the swapdevs.  Easy to fix, if needed.
397          */
398         if (!is_page_cache_freeable(page))
399                 return PAGE_KEEP;
400         if (!mapping) {
401                 /*
402                  * Some data journaling orphaned pages can have
403                  * page->mapping == NULL while being dirty with clean buffers.
404                  */
405                 if (page_has_private(page)) {
406                         if (try_to_free_buffers(page)) {
407                                 ClearPageDirty(page);
408                                 printk("%s: orphaned page\n", __func__);
409                                 return PAGE_CLEAN;
410                         }
411                 }
412                 return PAGE_KEEP;
413         }
414         if (mapping->a_ops->writepage == NULL)
415                 return PAGE_ACTIVATE;
416         if (!may_write_to_queue(mapping->backing_dev_info, sc))
417                 return PAGE_KEEP;
418 
419         if (clear_page_dirty_for_io(page)) {
420                 int res;
421                 struct writeback_control wbc = {
422                         .sync_mode = WB_SYNC_NONE,
423                         .nr_to_write = SWAP_CLUSTER_MAX,
424                         .range_start = 0,
425                         .range_end = LLONG_MAX,
426                         .for_reclaim = 1,
427                 };
428 
429                 SetPageReclaim(page);
430                 res = mapping->a_ops->writepage(page, &wbc);
431                 if (res < 0)
432                         handle_write_error(mapping, page, res);
433                 if (res == AOP_WRITEPAGE_ACTIVATE) {
434                         ClearPageReclaim(page);
435                         return PAGE_ACTIVATE;
436                 }
437 
438                 if (!PageWriteback(page)) {
439                         /* synchronous write or broken a_ops? */
440                         ClearPageReclaim(page);
441                 }
442                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
443                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
444                 return PAGE_SUCCESS;
445         }
446 
447         return PAGE_CLEAN;
448 }
449 
450 /*
451  * Same as remove_mapping, but if the page is removed from the mapping, it
452  * gets returned with a refcount of 0.
453  */
454 static int __remove_mapping(struct address_space *mapping, struct page *page)
455 {
456         BUG_ON(!PageLocked(page));
457         BUG_ON(mapping != page_mapping(page));
458 
459         spin_lock_irq(&mapping->tree_lock);
460         /*
461          * The non racy check for a busy page.
462          *
463          * Must be careful with the order of the tests. When someone has
464          * a ref to the page, it may be possible that they dirty it then
465          * drop the reference. So if PageDirty is tested before page_count
466          * here, then the following race may occur:
467          *
468          * get_user_pages(&page);
469          * [user mapping goes away]
470          * write_to(page);
471          *                              !PageDirty(page)    [good]
472          * SetPageDirty(page);
473          * put_page(page);
474          *                              !page_count(page)   [good, discard it]
475          *
476          * [oops, our write_to data is lost]
477          *
478          * Reversing the order of the tests ensures such a situation cannot
479          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
480          * load is not satisfied before that of page->_count.
481          *
482          * Note that if SetPageDirty is always performed via set_page_dirty,
483          * and thus under tree_lock, then this ordering is not required.
484          */
485         if (!page_freeze_refs(page, 2))
486                 goto cannot_free;
487         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
488         if (unlikely(PageDirty(page))) {
489                 page_unfreeze_refs(page, 2);
490                 goto cannot_free;
491         }
492 
493         if (PageSwapCache(page)) {
494                 swp_entry_t swap = { .val = page_private(page) };
495                 __delete_from_swap_cache(page);
496                 spin_unlock_irq(&mapping->tree_lock);
497                 swapcache_free(swap, page);
498         } else {
499                 void (*freepage)(struct page *);
500 
501                 freepage = mapping->a_ops->freepage;
502 
503                 __delete_from_page_cache(page);
504                 spin_unlock_irq(&mapping->tree_lock);
505                 mem_cgroup_uncharge_cache_page(page);
506 
507                 if (freepage != NULL)
508                         freepage(page);
509         }
510 
511         return 1;
512 
513 cannot_free:
514         spin_unlock_irq(&mapping->tree_lock);
515         return 0;
516 }
517 
518 /*
519  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
520  * someone else has a ref on the page, abort and return 0.  If it was
521  * successfully detached, return 1.  Assumes the caller has a single ref on
522  * this page.
523  */
524 int remove_mapping(struct address_space *mapping, struct page *page)
525 {
526         if (__remove_mapping(mapping, page)) {
527                 /*
528                  * Unfreezing the refcount with 1 rather than 2 effectively
529                  * drops the pagecache ref for us without requiring another
530                  * atomic operation.
531                  */
532                 page_unfreeze_refs(page, 1);
533                 return 1;
534         }
535         return 0;
536 }
537 
538 /**
539  * putback_lru_page - put previously isolated page onto appropriate LRU list
540  * @page: page to be put back to appropriate lru list
541  *
542  * Add previously isolated @page to appropriate LRU list.
543  * Page may still be unevictable for other reasons.
544  *
545  * lru_lock must not be held, interrupts must be enabled.
546  */
547 void putback_lru_page(struct page *page)
548 {
549         int lru;
550         int active = !!TestClearPageActive(page);
551         int was_unevictable = PageUnevictable(page);
552 
553         VM_BUG_ON(PageLRU(page));
554 
555 redo:
556         ClearPageUnevictable(page);
557 
558         if (page_evictable(page)) {
559                 /*
560                  * For evictable pages, we can use the cache.
561                  * In event of a race, worst case is we end up with an
562                  * unevictable page on [in]active list.
563                  * We know how to handle that.
564                  */
565                 lru = active + page_lru_base_type(page);
566                 lru_cache_add_lru(page, lru);
567         } else {
568                 /*
569                  * Put unevictable pages directly on zone's unevictable
570                  * list.
571                  */
572                 lru = LRU_UNEVICTABLE;
573                 add_page_to_unevictable_list(page);
574                 /*
575                  * When racing with an mlock or AS_UNEVICTABLE clearing
576                  * (page is unlocked) make sure that if the other thread
577                  * does not observe our setting of PG_lru and fails
578                  * isolation/check_move_unevictable_pages,
579                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
580                  * the page back to the evictable list.
581                  *
582                  * The other side is TestClearPageMlocked() or shmem_lock().
583                  */
584                 smp_mb();
585         }
586 
587         /*
588          * page's status can change while we move it among lru. If an evictable
589          * page is on unevictable list, it never be freed. To avoid that,
590          * check after we added it to the list, again.
591          */
592         if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
593                 if (!isolate_lru_page(page)) {
594                         put_page(page);
595                         goto redo;
596                 }
597                 /* This means someone else dropped this page from LRU
598                  * So, it will be freed or putback to LRU again. There is
599                  * nothing to do here.
600                  */
601         }
602 
603         if (was_unevictable && lru != LRU_UNEVICTABLE)
604                 count_vm_event(UNEVICTABLE_PGRESCUED);
605         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
606                 count_vm_event(UNEVICTABLE_PGCULLED);
607 
608         put_page(page);         /* drop ref from isolate */
609 }
610 
611 enum page_references {
612         PAGEREF_RECLAIM,
613         PAGEREF_RECLAIM_CLEAN,
614         PAGEREF_KEEP,
615         PAGEREF_ACTIVATE,
616 };
617 
618 static enum page_references page_check_references(struct page *page,
619                                                   struct scan_control *sc)
620 {
621         int referenced_ptes, referenced_page;
622         unsigned long vm_flags;
623 
624         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
625                                           &vm_flags);
626         referenced_page = TestClearPageReferenced(page);
627 
628         /*
629          * Mlock lost the isolation race with us.  Let try_to_unmap()
630          * move the page to the unevictable list.
631          */
632         if (vm_flags & VM_LOCKED)
633                 return PAGEREF_RECLAIM;
634 
635         if (referenced_ptes) {
636                 if (PageSwapBacked(page))
637                         return PAGEREF_ACTIVATE;
638                 /*
639                  * All mapped pages start out with page table
640                  * references from the instantiating fault, so we need
641                  * to look twice if a mapped file page is used more
642                  * than once.
643                  *
644                  * Mark it and spare it for another trip around the
645                  * inactive list.  Another page table reference will
646                  * lead to its activation.
647                  *
648                  * Note: the mark is set for activated pages as well
649                  * so that recently deactivated but used pages are
650                  * quickly recovered.
651                  */
652                 SetPageReferenced(page);
653 
654                 if (referenced_page || referenced_ptes > 1)
655                         return PAGEREF_ACTIVATE;
656 
657                 /*
658                  * Activate file-backed executable pages after first usage.
659                  */
660                 if (vm_flags & VM_EXEC)
661                         return PAGEREF_ACTIVATE;
662 
663                 return PAGEREF_KEEP;
664         }
665 
666         /* Reclaim if clean, defer dirty pages to writeback */
667         if (referenced_page && !PageSwapBacked(page))
668                 return PAGEREF_RECLAIM_CLEAN;
669 
670         return PAGEREF_RECLAIM;
671 }
672 
673 /*
674  * shrink_page_list() returns the number of reclaimed pages
675  */
676 static unsigned long shrink_page_list(struct list_head *page_list,
677                                       struct zone *zone,
678                                       struct scan_control *sc,
679                                       enum ttu_flags ttu_flags,
680                                       unsigned long *ret_nr_dirty,
681                                       unsigned long *ret_nr_writeback,
682                                       bool force_reclaim)
683 {
684         LIST_HEAD(ret_pages);
685         LIST_HEAD(free_pages);
686         int pgactivate = 0;
687         unsigned long nr_dirty = 0;
688         unsigned long nr_congested = 0;
689         unsigned long nr_reclaimed = 0;
690         unsigned long nr_writeback = 0;
691 
692         cond_resched();
693 
694         mem_cgroup_uncharge_start();
695         while (!list_empty(page_list)) {
696                 struct address_space *mapping;
697                 struct page *page;
698                 int may_enter_fs;
699                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
700 
701                 cond_resched();
702 
703                 page = lru_to_page(page_list);
704                 list_del(&page->lru);
705 
706                 if (!trylock_page(page))
707                         goto keep;
708 
709                 VM_BUG_ON(PageActive(page));
710                 VM_BUG_ON(page_zone(page) != zone);
711 
712                 sc->nr_scanned++;
713 
714                 if (unlikely(!page_evictable(page)))
715                         goto cull_mlocked;
716 
717                 if (!sc->may_unmap && page_mapped(page))
718                         goto keep_locked;
719 
720                 /* Double the slab pressure for mapped and swapcache pages */
721                 if (page_mapped(page) || PageSwapCache(page))
722                         sc->nr_scanned++;
723 
724                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
725                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
726 
727                 if (PageWriteback(page)) {
728                         /*
729                          * memcg doesn't have any dirty pages throttling so we
730                          * could easily OOM just because too many pages are in
731                          * writeback and there is nothing else to reclaim.
732                          *
733                          * Require may_enter_fs to wait on writeback, because
734                          * fs may not have submitted IO yet. And a loop driver
735                          * thread might enter reclaim, and deadlock if it waits
736                          * on a page for which it is needed to do the write
737                          * (loop masks off __GFP_IO|__GFP_FS for this reason);
738                          * but more thought would probably show more reasons.
739                          */
740                         if (global_reclaim(sc) ||
741                             !PageReclaim(page) || !may_enter_fs) {
742                                 /*
743                                  * This is slightly racy - end_page_writeback()
744                                  * might have just cleared PageReclaim, then
745                                  * setting PageReclaim here end up interpreted
746                                  * as PageReadahead - but that does not matter
747                                  * enough to care.  What we do want is for this
748                                  * page to have PageReclaim set next time memcg
749                                  * reclaim reaches the tests above, so it will
750                                  * then wait_on_page_writeback() to avoid OOM;
751                                  * and it's also appropriate in global reclaim.
752                                  */
753                                 SetPageReclaim(page);
754                                 nr_writeback++;
755                                 goto keep_locked;
756                         }
757                         wait_on_page_writeback(page);
758                 }
759 
760                 if (!force_reclaim)
761                         references = page_check_references(page, sc);
762 
763                 switch (references) {
764                 case PAGEREF_ACTIVATE:
765                         goto activate_locked;
766                 case PAGEREF_KEEP:
767                         goto keep_locked;
768                 case PAGEREF_RECLAIM:
769                 case PAGEREF_RECLAIM_CLEAN:
770                         ; /* try to reclaim the page below */
771                 }
772 
773                 /*
774                  * Anonymous process memory has backing store?
775                  * Try to allocate it some swap space here.
776                  */
777                 if (PageAnon(page) && !PageSwapCache(page)) {
778                         if (!(sc->gfp_mask & __GFP_IO))
779                                 goto keep_locked;
780                         if (!add_to_swap(page, page_list))
781                                 goto activate_locked;
782                         may_enter_fs = 1;
783                 }
784 
785                 mapping = page_mapping(page);
786 
787                 /*
788                  * The page is mapped into the page tables of one or more
789                  * processes. Try to unmap it here.
790                  */
791                 if (page_mapped(page) && mapping) {
792                         switch (try_to_unmap(page, ttu_flags)) {
793                         case SWAP_FAIL:
794                                 goto activate_locked;
795                         case SWAP_AGAIN:
796                                 goto keep_locked;
797                         case SWAP_MLOCK:
798                                 goto cull_mlocked;
799                         case SWAP_SUCCESS:
800                                 ; /* try to free the page below */
801                         }
802                 }
803 
804                 if (PageDirty(page)) {
805                         nr_dirty++;
806 
807                         /*
808                          * Only kswapd can writeback filesystem pages to
809                          * avoid risk of stack overflow but do not writeback
810                          * unless under significant pressure.
811                          */
812                         if (page_is_file_cache(page) &&
813                                         (!current_is_kswapd() ||
814                                          sc->priority >= DEF_PRIORITY - 2)) {
815                                 /*
816                                  * Immediately reclaim when written back.
817                                  * Similar in principal to deactivate_page()
818                                  * except we already have the page isolated
819                                  * and know it's dirty
820                                  */
821                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
822                                 SetPageReclaim(page);
823 
824                                 goto keep_locked;
825                         }
826 
827                         if (references == PAGEREF_RECLAIM_CLEAN)
828                                 goto keep_locked;
829                         if (!may_enter_fs)
830                                 goto keep_locked;
831                         if (!sc->may_writepage)
832                                 goto keep_locked;
833 
834                         /* Page is dirty, try to write it out here */
835                         switch (pageout(page, mapping, sc)) {
836                         case PAGE_KEEP:
837                                 nr_congested++;
838                                 goto keep_locked;
839                         case PAGE_ACTIVATE:
840                                 goto activate_locked;
841                         case PAGE_SUCCESS:
842                                 if (PageWriteback(page))
843                                         goto keep;
844                                 if (PageDirty(page))
845                                         goto keep;
846 
847                                 /*
848                                  * A synchronous write - probably a ramdisk.  Go
849                                  * ahead and try to reclaim the page.
850                                  */
851                                 if (!trylock_page(page))
852                                         goto keep;
853                                 if (PageDirty(page) || PageWriteback(page))
854                                         goto keep_locked;
855                                 mapping = page_mapping(page);
856                         case PAGE_CLEAN:
857                                 ; /* try to free the page below */
858                         }
859                 }
860 
861                 /*
862                  * If the page has buffers, try to free the buffer mappings
863                  * associated with this page. If we succeed we try to free
864                  * the page as well.
865                  *
866                  * We do this even if the page is PageDirty().
867                  * try_to_release_page() does not perform I/O, but it is
868                  * possible for a page to have PageDirty set, but it is actually
869                  * clean (all its buffers are clean).  This happens if the
870                  * buffers were written out directly, with submit_bh(). ext3
871                  * will do this, as well as the blockdev mapping.
872                  * try_to_release_page() will discover that cleanness and will
873                  * drop the buffers and mark the page clean - it can be freed.
874                  *
875                  * Rarely, pages can have buffers and no ->mapping.  These are
876                  * the pages which were not successfully invalidated in
877                  * truncate_complete_page().  We try to drop those buffers here
878                  * and if that worked, and the page is no longer mapped into
879                  * process address space (page_count == 1) it can be freed.
880                  * Otherwise, leave the page on the LRU so it is swappable.
881                  */
882                 if (page_has_private(page)) {
883                         if (!try_to_release_page(page, sc->gfp_mask))
884                                 goto activate_locked;
885                         if (!mapping && page_count(page) == 1) {
886                                 unlock_page(page);
887                                 if (put_page_testzero(page))
888                                         goto free_it;
889                                 else {
890                                         /*
891                                          * rare race with speculative reference.
892                                          * the speculative reference will free
893                                          * this page shortly, so we may
894                                          * increment nr_reclaimed here (and
895                                          * leave it off the LRU).
896                                          */
897                                         nr_reclaimed++;
898                                         continue;
899                                 }
900                         }
901                 }
902 
903                 if (!mapping || !__remove_mapping(mapping, page))
904                         goto keep_locked;
905 
906                 /*
907                  * At this point, we have no other references and there is
908                  * no way to pick any more up (removed from LRU, removed
909                  * from pagecache). Can use non-atomic bitops now (and
910                  * we obviously don't have to worry about waking up a process
911                  * waiting on the page lock, because there are no references.
912                  */
913                 __clear_page_locked(page);
914 free_it:
915                 nr_reclaimed++;
916 
917                 /*
918                  * Is there need to periodically free_page_list? It would
919                  * appear not as the counts should be low
920                  */
921                 list_add(&page->lru, &free_pages);
922                 continue;
923 
924 cull_mlocked:
925                 if (PageSwapCache(page))
926                         try_to_free_swap(page);
927                 unlock_page(page);
928                 list_add(&page->lru, &ret_pages);
929                 continue;
930 
931 activate_locked:
932                 /* Not a candidate for swapping, so reclaim swap space. */
933                 if (PageSwapCache(page) && vm_swap_full())
934                         try_to_free_swap(page);
935                 VM_BUG_ON(PageActive(page));
936                 SetPageActive(page);
937                 pgactivate++;
938 keep_locked:
939                 unlock_page(page);
940 keep:
941                 list_add(&page->lru, &ret_pages);
942                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
943         }
944 
945         /*
946          * Tag a zone as congested if all the dirty pages encountered were
947          * backed by a congested BDI. In this case, reclaimers should just
948          * back off and wait for congestion to clear because further reclaim
949          * will encounter the same problem
950          */
951         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
952                 zone_set_flag(zone, ZONE_CONGESTED);
953 
954         free_hot_cold_page_list(&free_pages, 1);
955 
956         list_splice(&ret_pages, page_list);
957         count_vm_events(PGACTIVATE, pgactivate);
958         mem_cgroup_uncharge_end();
959         *ret_nr_dirty += nr_dirty;
960         *ret_nr_writeback += nr_writeback;
961         return nr_reclaimed;
962 }
963 
964 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
965                                             struct list_head *page_list)
966 {
967         struct scan_control sc = {
968                 .gfp_mask = GFP_KERNEL,
969                 .priority = DEF_PRIORITY,
970                 .may_unmap = 1,
971         };
972         unsigned long ret, dummy1, dummy2;
973         struct page *page, *next;
974         LIST_HEAD(clean_pages);
975 
976         list_for_each_entry_safe(page, next, page_list, lru) {
977                 if (page_is_file_cache(page) && !PageDirty(page) &&
978                     !isolated_balloon_page(page)) {
979                         ClearPageActive(page);
980                         list_move(&page->lru, &clean_pages);
981                 }
982         }
983 
984         ret = shrink_page_list(&clean_pages, zone, &sc,
985                                 TTU_UNMAP|TTU_IGNORE_ACCESS,
986                                 &dummy1, &dummy2, true);
987         list_splice(&clean_pages, page_list);
988         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
989         return ret;
990 }
991 
992 /*
993  * Attempt to remove the specified page from its LRU.  Only take this page
994  * if it is of the appropriate PageActive status.  Pages which are being
995  * freed elsewhere are also ignored.
996  *
997  * page:        page to consider
998  * mode:        one of the LRU isolation modes defined above
999  *
1000  * returns 0 on success, -ve errno on failure.
1001  */
1002 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1003 {
1004         int ret = -EINVAL;
1005 
1006         /* Only take pages on the LRU. */
1007         if (!PageLRU(page))
1008                 return ret;
1009 
1010         /* Compaction should not handle unevictable pages but CMA can do so */
1011         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1012                 return ret;
1013 
1014         ret = -EBUSY;
1015 
1016         /*
1017          * To minimise LRU disruption, the caller can indicate that it only
1018          * wants to isolate pages it will be able to operate on without
1019          * blocking - clean pages for the most part.
1020          *
1021          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1022          * is used by reclaim when it is cannot write to backing storage
1023          *
1024          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1025          * that it is possible to migrate without blocking
1026          */
1027         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1028                 /* All the caller can do on PageWriteback is block */
1029                 if (PageWriteback(page))
1030                         return ret;
1031 
1032                 if (PageDirty(page)) {
1033                         struct address_space *mapping;
1034 
1035                         /* ISOLATE_CLEAN means only clean pages */
1036                         if (mode & ISOLATE_CLEAN)
1037                                 return ret;
1038 
1039                         /*
1040                          * Only pages without mappings or that have a
1041                          * ->migratepage callback are possible to migrate
1042                          * without blocking
1043                          */
1044                         mapping = page_mapping(page);
1045                         if (mapping && !mapping->a_ops->migratepage)
1046                                 return ret;
1047                 }
1048         }
1049 
1050         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1051                 return ret;
1052 
1053         if (likely(get_page_unless_zero(page))) {
1054                 /*
1055                  * Be careful not to clear PageLRU until after we're
1056                  * sure the page is not being freed elsewhere -- the
1057                  * page release code relies on it.
1058                  */
1059                 ClearPageLRU(page);
1060                 ret = 0;
1061         }
1062 
1063         return ret;
1064 }
1065 
1066 /*
1067  * zone->lru_lock is heavily contended.  Some of the functions that
1068  * shrink the lists perform better by taking out a batch of pages
1069  * and working on them outside the LRU lock.
1070  *
1071  * For pagecache intensive workloads, this function is the hottest
1072  * spot in the kernel (apart from copy_*_user functions).
1073  *
1074  * Appropriate locks must be held before calling this function.
1075  *
1076  * @nr_to_scan: The number of pages to look through on the list.
1077  * @lruvec:     The LRU vector to pull pages from.
1078  * @dst:        The temp list to put pages on to.
1079  * @nr_scanned: The number of pages that were scanned.
1080  * @sc:         The scan_control struct for this reclaim session
1081  * @mode:       One of the LRU isolation modes
1082  * @lru:        LRU list id for isolating
1083  *
1084  * returns how many pages were moved onto *@dst.
1085  */
1086 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1087                 struct lruvec *lruvec, struct list_head *dst,
1088                 unsigned long *nr_scanned, struct scan_control *sc,
1089                 isolate_mode_t mode, enum lru_list lru)
1090 {
1091         struct list_head *src = &lruvec->lists[lru];
1092         unsigned long nr_taken = 0;
1093         unsigned long scan;
1094 
1095         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1096                 struct page *page;
1097                 int nr_pages;
1098 
1099                 page = lru_to_page(src);
1100                 prefetchw_prev_lru_page(page, src, flags);
1101 
1102                 VM_BUG_ON(!PageLRU(page));
1103 
1104                 switch (__isolate_lru_page(page, mode)) {
1105                 case 0:
1106                         nr_pages = hpage_nr_pages(page);
1107                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1108                         list_move(&page->lru, dst);
1109                         nr_taken += nr_pages;
1110                         break;
1111 
1112                 case -EBUSY:
1113                         /* else it is being freed elsewhere */
1114                         list_move(&page->lru, src);
1115                         continue;
1116 
1117                 default:
1118                         BUG();
1119                 }
1120         }
1121 
1122         *nr_scanned = scan;
1123         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1124                                     nr_taken, mode, is_file_lru(lru));
1125         return nr_taken;
1126 }
1127 
1128 /**
1129  * isolate_lru_page - tries to isolate a page from its LRU list
1130  * @page: page to isolate from its LRU list
1131  *
1132  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1133  * vmstat statistic corresponding to whatever LRU list the page was on.
1134  *
1135  * Returns 0 if the page was removed from an LRU list.
1136  * Returns -EBUSY if the page was not on an LRU list.
1137  *
1138  * The returned page will have PageLRU() cleared.  If it was found on
1139  * the active list, it will have PageActive set.  If it was found on
1140  * the unevictable list, it will have the PageUnevictable bit set. That flag
1141  * may need to be cleared by the caller before letting the page go.
1142  *
1143  * The vmstat statistic corresponding to the list on which the page was
1144  * found will be decremented.
1145  *
1146  * Restrictions:
1147  * (1) Must be called with an elevated refcount on the page. This is a
1148  *     fundamentnal difference from isolate_lru_pages (which is called
1149  *     without a stable reference).
1150  * (2) the lru_lock must not be held.
1151  * (3) interrupts must be enabled.
1152  */
1153 int isolate_lru_page(struct page *page)
1154 {
1155         int ret = -EBUSY;
1156 
1157         VM_BUG_ON(!page_count(page));
1158 
1159         if (PageLRU(page)) {
1160                 struct zone *zone = page_zone(page);
1161                 struct lruvec *lruvec;
1162 
1163                 spin_lock_irq(&zone->lru_lock);
1164                 lruvec = mem_cgroup_page_lruvec(page, zone);
1165                 if (PageLRU(page)) {
1166                         int lru = page_lru(page);
1167                         get_page(page);
1168                         ClearPageLRU(page);
1169                         del_page_from_lru_list(page, lruvec, lru);
1170                         ret = 0;
1171                 }
1172                 spin_unlock_irq(&zone->lru_lock);
1173         }
1174         return ret;
1175 }
1176 
1177 /*
1178  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1179  * then get resheduled. When there are massive number of tasks doing page
1180  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1181  * the LRU list will go small and be scanned faster than necessary, leading to
1182  * unnecessary swapping, thrashing and OOM.
1183  */
1184 static int too_many_isolated(struct zone *zone, int file,
1185                 struct scan_control *sc)
1186 {
1187         unsigned long inactive, isolated;
1188 
1189         if (current_is_kswapd())
1190                 return 0;
1191 
1192         if (!global_reclaim(sc))
1193                 return 0;
1194 
1195         if (file) {
1196                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1197                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1198         } else {
1199                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1200                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1201         }
1202 
1203         /*
1204          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1205          * won't get blocked by normal direct-reclaimers, forming a circular
1206          * deadlock.
1207          */
1208         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1209                 inactive >>= 3;
1210 
1211         return isolated > inactive;
1212 }
1213 
1214 static noinline_for_stack void
1215 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1216 {
1217         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1218         struct zone *zone = lruvec_zone(lruvec);
1219         LIST_HEAD(pages_to_free);
1220 
1221         /*
1222          * Put back any unfreeable pages.
1223          */
1224         while (!list_empty(page_list)) {
1225                 struct page *page = lru_to_page(page_list);
1226                 int lru;
1227 
1228                 VM_BUG_ON(PageLRU(page));
1229                 list_del(&page->lru);
1230                 if (unlikely(!page_evictable(page))) {
1231                         spin_unlock_irq(&zone->lru_lock);
1232                         putback_lru_page(page);
1233                         spin_lock_irq(&zone->lru_lock);
1234                         continue;
1235                 }
1236 
1237                 lruvec = mem_cgroup_page_lruvec(page, zone);
1238 
1239                 SetPageLRU(page);
1240                 lru = page_lru(page);
1241                 add_page_to_lru_list(page, lruvec, lru);
1242 
1243                 if (is_active_lru(lru)) {
1244                         int file = is_file_lru(lru);
1245                         int numpages = hpage_nr_pages(page);
1246                         reclaim_stat->recent_rotated[file] += numpages;
1247                 }
1248                 if (put_page_testzero(page)) {
1249                         __ClearPageLRU(page);
1250                         __ClearPageActive(page);
1251                         del_page_from_lru_list(page, lruvec, lru);
1252 
1253                         if (unlikely(PageCompound(page))) {
1254                                 spin_unlock_irq(&zone->lru_lock);
1255                                 (*get_compound_page_dtor(page))(page);
1256                                 spin_lock_irq(&zone->lru_lock);
1257                         } else
1258                                 list_add(&page->lru, &pages_to_free);
1259                 }
1260         }
1261 
1262         /*
1263          * To save our caller's stack, now use input list for pages to free.
1264          */
1265         list_splice(&pages_to_free, page_list);
1266 }
1267 
1268 /*
1269  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1270  * of reclaimed pages
1271  */
1272 static noinline_for_stack unsigned long
1273 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1274                      struct scan_control *sc, enum lru_list lru)
1275 {
1276         LIST_HEAD(page_list);
1277         unsigned long nr_scanned;
1278         unsigned long nr_reclaimed = 0;
1279         unsigned long nr_taken;
1280         unsigned long nr_dirty = 0;
1281         unsigned long nr_writeback = 0;
1282         isolate_mode_t isolate_mode = 0;
1283         int file = is_file_lru(lru);
1284         struct zone *zone = lruvec_zone(lruvec);
1285         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1286 
1287         while (unlikely(too_many_isolated(zone, file, sc))) {
1288                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1289 
1290                 /* We are about to die and free our memory. Return now. */
1291                 if (fatal_signal_pending(current))
1292                         return SWAP_CLUSTER_MAX;
1293         }
1294 
1295         lru_add_drain();
1296 
1297         if (!sc->may_unmap)
1298                 isolate_mode |= ISOLATE_UNMAPPED;
1299         if (!sc->may_writepage)
1300                 isolate_mode |= ISOLATE_CLEAN;
1301 
1302         spin_lock_irq(&zone->lru_lock);
1303 
1304         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1305                                      &nr_scanned, sc, isolate_mode, lru);
1306 
1307         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1308         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1309 
1310         if (global_reclaim(sc)) {
1311                 zone->pages_scanned += nr_scanned;
1312                 if (current_is_kswapd())
1313                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1314                 else
1315                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1316         }
1317         spin_unlock_irq(&zone->lru_lock);
1318 
1319         if (nr_taken == 0)
1320                 return 0;
1321 
1322         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1323                                         &nr_dirty, &nr_writeback, false);
1324 
1325         spin_lock_irq(&zone->lru_lock);
1326 
1327         reclaim_stat->recent_scanned[file] += nr_taken;
1328 
1329         if (global_reclaim(sc)) {
1330                 if (current_is_kswapd())
1331                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1332                                                nr_reclaimed);
1333                 else
1334                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1335                                                nr_reclaimed);
1336         }
1337 
1338         putback_inactive_pages(lruvec, &page_list);
1339 
1340         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1341 
1342         spin_unlock_irq(&zone->lru_lock);
1343 
1344         free_hot_cold_page_list(&page_list, 1);
1345 
1346         /*
1347          * If reclaim is isolating dirty pages under writeback, it implies
1348          * that the long-lived page allocation rate is exceeding the page
1349          * laundering rate. Either the global limits are not being effective
1350          * at throttling processes due to the page distribution throughout
1351          * zones or there is heavy usage of a slow backing device. The
1352          * only option is to throttle from reclaim context which is not ideal
1353          * as there is no guarantee the dirtying process is throttled in the
1354          * same way balance_dirty_pages() manages.
1355          *
1356          * This scales the number of dirty pages that must be under writeback
1357          * before throttling depending on priority. It is a simple backoff
1358          * function that has the most effect in the range DEF_PRIORITY to
1359          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1360          * in trouble and reclaim is considered to be in trouble.
1361          *
1362          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1363          * DEF_PRIORITY-1  50% must be PageWriteback
1364          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1365          * ...
1366          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1367          *                     isolated page is PageWriteback
1368          */
1369         if (nr_writeback && nr_writeback >=
1370                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1371                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1372 
1373         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1374                 zone_idx(zone),
1375                 nr_scanned, nr_reclaimed,
1376                 sc->priority,
1377                 trace_shrink_flags(file));
1378         return nr_reclaimed;
1379 }
1380 
1381 /*
1382  * This moves pages from the active list to the inactive list.
1383  *
1384  * We move them the other way if the page is referenced by one or more
1385  * processes, from rmap.
1386  *
1387  * If the pages are mostly unmapped, the processing is fast and it is
1388  * appropriate to hold zone->lru_lock across the whole operation.  But if
1389  * the pages are mapped, the processing is slow (page_referenced()) so we
1390  * should drop zone->lru_lock around each page.  It's impossible to balance
1391  * this, so instead we remove the pages from the LRU while processing them.
1392  * It is safe to rely on PG_active against the non-LRU pages in here because
1393  * nobody will play with that bit on a non-LRU page.
1394  *
1395  * The downside is that we have to touch page->_count against each page.
1396  * But we had to alter page->flags anyway.
1397  */
1398 
1399 static void move_active_pages_to_lru(struct lruvec *lruvec,
1400                                      struct list_head *list,
1401                                      struct list_head *pages_to_free,
1402                                      enum lru_list lru)
1403 {
1404         struct zone *zone = lruvec_zone(lruvec);
1405         unsigned long pgmoved = 0;
1406         struct page *page;
1407         int nr_pages;
1408 
1409         while (!list_empty(list)) {
1410                 page = lru_to_page(list);
1411                 lruvec = mem_cgroup_page_lruvec(page, zone);
1412 
1413                 VM_BUG_ON(PageLRU(page));
1414                 SetPageLRU(page);
1415 
1416                 nr_pages = hpage_nr_pages(page);
1417                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1418                 list_move(&page->lru, &lruvec->lists[lru]);
1419                 pgmoved += nr_pages;
1420 
1421                 if (put_page_testzero(page)) {
1422                         __ClearPageLRU(page);
1423                         __ClearPageActive(page);
1424                         del_page_from_lru_list(page, lruvec, lru);
1425 
1426                         if (unlikely(PageCompound(page))) {
1427                                 spin_unlock_irq(&zone->lru_lock);
1428                                 (*get_compound_page_dtor(page))(page);
1429                                 spin_lock_irq(&zone->lru_lock);
1430                         } else
1431                                 list_add(&page->lru, pages_to_free);
1432                 }
1433         }
1434         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1435         if (!is_active_lru(lru))
1436                 __count_vm_events(PGDEACTIVATE, pgmoved);
1437 }
1438 
1439 static void shrink_active_list(unsigned long nr_to_scan,
1440                                struct lruvec *lruvec,
1441                                struct scan_control *sc,
1442                                enum lru_list lru)
1443 {
1444         unsigned long nr_taken;
1445         unsigned long nr_scanned;
1446         unsigned long vm_flags;
1447         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1448         LIST_HEAD(l_active);
1449         LIST_HEAD(l_inactive);
1450         struct page *page;
1451         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1452         unsigned long nr_rotated = 0;
1453         isolate_mode_t isolate_mode = 0;
1454         int file = is_file_lru(lru);
1455         struct zone *zone = lruvec_zone(lruvec);
1456 
1457         lru_add_drain();
1458 
1459         if (!sc->may_unmap)
1460                 isolate_mode |= ISOLATE_UNMAPPED;
1461         if (!sc->may_writepage)
1462                 isolate_mode |= ISOLATE_CLEAN;
1463 
1464         spin_lock_irq(&zone->lru_lock);
1465 
1466         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1467                                      &nr_scanned, sc, isolate_mode, lru);
1468         if (global_reclaim(sc))
1469                 zone->pages_scanned += nr_scanned;
1470 
1471         reclaim_stat->recent_scanned[file] += nr_taken;
1472 
1473         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1474         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1475         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1476         spin_unlock_irq(&zone->lru_lock);
1477 
1478         while (!list_empty(&l_hold)) {
1479                 cond_resched();
1480                 page = lru_to_page(&l_hold);
1481                 list_del(&page->lru);
1482 
1483                 if (unlikely(!page_evictable(page))) {
1484                         putback_lru_page(page);
1485                         continue;
1486                 }
1487 
1488                 if (unlikely(buffer_heads_over_limit)) {
1489                         if (page_has_private(page) && trylock_page(page)) {
1490                                 if (page_has_private(page))
1491                                         try_to_release_page(page, 0);
1492                                 unlock_page(page);
1493                         }
1494                 }
1495 
1496                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1497                                     &vm_flags)) {
1498                         nr_rotated += hpage_nr_pages(page);
1499                         /*
1500                          * Identify referenced, file-backed active pages and
1501                          * give them one more trip around the active list. So
1502                          * that executable code get better chances to stay in
1503                          * memory under moderate memory pressure.  Anon pages
1504                          * are not likely to be evicted by use-once streaming
1505                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1506                          * so we ignore them here.
1507                          */
1508                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1509                                 list_add(&page->lru, &l_active);
1510                                 continue;
1511                         }
1512                 }
1513 
1514                 ClearPageActive(page);  /* we are de-activating */
1515                 list_add(&page->lru, &l_inactive);
1516         }
1517 
1518         /*
1519          * Move pages back to the lru list.
1520          */
1521         spin_lock_irq(&zone->lru_lock);
1522         /*
1523          * Count referenced pages from currently used mappings as rotated,
1524          * even though only some of them are actually re-activated.  This
1525          * helps balance scan pressure between file and anonymous pages in
1526          * get_scan_ratio.
1527          */
1528         reclaim_stat->recent_rotated[file] += nr_rotated;
1529 
1530         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1531         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1532         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1533         spin_unlock_irq(&zone->lru_lock);
1534 
1535         free_hot_cold_page_list(&l_hold, 1);
1536 }
1537 
1538 #ifdef CONFIG_SWAP
1539 static int inactive_anon_is_low_global(struct zone *zone)
1540 {
1541         unsigned long active, inactive;
1542 
1543         active = zone_page_state(zone, NR_ACTIVE_ANON);
1544         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1545 
1546         if (inactive * zone->inactive_ratio < active)
1547                 return 1;
1548 
1549         return 0;
1550 }
1551 
1552 /**
1553  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1554  * @lruvec: LRU vector to check
1555  *
1556  * Returns true if the zone does not have enough inactive anon pages,
1557  * meaning some active anon pages need to be deactivated.
1558  */
1559 static int inactive_anon_is_low(struct lruvec *lruvec)
1560 {
1561         /*
1562          * If we don't have swap space, anonymous page deactivation
1563          * is pointless.
1564          */
1565         if (!total_swap_pages)
1566                 return 0;
1567 
1568         if (!mem_cgroup_disabled())
1569                 return mem_cgroup_inactive_anon_is_low(lruvec);
1570 
1571         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1572 }
1573 #else
1574 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1575 {
1576         return 0;
1577 }
1578 #endif
1579 
1580 /**
1581  * inactive_file_is_low - check if file pages need to be deactivated
1582  * @lruvec: LRU vector to check
1583  *
1584  * When the system is doing streaming IO, memory pressure here
1585  * ensures that active file pages get deactivated, until more
1586  * than half of the file pages are on the inactive list.
1587  *
1588  * Once we get to that situation, protect the system's working
1589  * set from being evicted by disabling active file page aging.
1590  *
1591  * This uses a different ratio than the anonymous pages, because
1592  * the page cache uses a use-once replacement algorithm.
1593  */
1594 static int inactive_file_is_low(struct lruvec *lruvec)
1595 {
1596         unsigned long inactive;
1597         unsigned long active;
1598 
1599         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1600         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1601 
1602         return active > inactive;
1603 }
1604 
1605 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1606 {
1607         if (is_file_lru(lru))
1608                 return inactive_file_is_low(lruvec);
1609         else
1610                 return inactive_anon_is_low(lruvec);
1611 }
1612 
1613 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1614                                  struct lruvec *lruvec, struct scan_control *sc)
1615 {
1616         if (is_active_lru(lru)) {
1617                 if (inactive_list_is_low(lruvec, lru))
1618                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1619                 return 0;
1620         }
1621 
1622         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1623 }
1624 
1625 static int vmscan_swappiness(struct scan_control *sc)
1626 {
1627         if (global_reclaim(sc))
1628                 return vm_swappiness;
1629         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1630 }
1631 
1632 enum scan_balance {
1633         SCAN_EQUAL,
1634         SCAN_FRACT,
1635         SCAN_ANON,
1636         SCAN_FILE,
1637 };
1638 
1639 /*
1640  * Determine how aggressively the anon and file LRU lists should be
1641  * scanned.  The relative value of each set of LRU lists is determined
1642  * by looking at the fraction of the pages scanned we did rotate back
1643  * onto the active list instead of evict.
1644  *
1645  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1646  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1647  */
1648 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1649                            unsigned long *nr)
1650 {
1651         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1652         u64 fraction[2];
1653         u64 denominator = 0;    /* gcc */
1654         struct zone *zone = lruvec_zone(lruvec);
1655         unsigned long anon_prio, file_prio;
1656         enum scan_balance scan_balance;
1657         unsigned long anon, file, free;
1658         bool force_scan = false;
1659         unsigned long ap, fp;
1660         enum lru_list lru;
1661 
1662         /*
1663          * If the zone or memcg is small, nr[l] can be 0.  This
1664          * results in no scanning on this priority and a potential
1665          * priority drop.  Global direct reclaim can go to the next
1666          * zone and tends to have no problems. Global kswapd is for
1667          * zone balancing and it needs to scan a minimum amount. When
1668          * reclaiming for a memcg, a priority drop can cause high
1669          * latencies, so it's better to scan a minimum amount there as
1670          * well.
1671          */
1672         if (current_is_kswapd() && zone->all_unreclaimable)
1673                 force_scan = true;
1674         if (!global_reclaim(sc))
1675                 force_scan = true;
1676 
1677         /* If we have no swap space, do not bother scanning anon pages. */
1678         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1679                 scan_balance = SCAN_FILE;
1680                 goto out;
1681         }
1682 
1683         /*
1684          * Global reclaim will swap to prevent OOM even with no
1685          * swappiness, but memcg users want to use this knob to
1686          * disable swapping for individual groups completely when
1687          * using the memory controller's swap limit feature would be
1688          * too expensive.
1689          */
1690         if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1691                 scan_balance = SCAN_FILE;
1692                 goto out;
1693         }
1694 
1695         /*
1696          * Do not apply any pressure balancing cleverness when the
1697          * system is close to OOM, scan both anon and file equally
1698          * (unless the swappiness setting disagrees with swapping).
1699          */
1700         if (!sc->priority && vmscan_swappiness(sc)) {
1701                 scan_balance = SCAN_EQUAL;
1702                 goto out;
1703         }
1704 
1705         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1706                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1707         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1708                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1709 
1710         /*
1711          * If it's foreseeable that reclaiming the file cache won't be
1712          * enough to get the zone back into a desirable shape, we have
1713          * to swap.  Better start now and leave the - probably heavily
1714          * thrashing - remaining file pages alone.
1715          */
1716         if (global_reclaim(sc)) {
1717                 free = zone_page_state(zone, NR_FREE_PAGES);
1718                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1719                         scan_balance = SCAN_ANON;
1720                         goto out;
1721                 }
1722         }
1723 
1724         /*
1725          * There is enough inactive page cache, do not reclaim
1726          * anything from the anonymous working set right now.
1727          */
1728         if (!inactive_file_is_low(lruvec)) {
1729                 scan_balance = SCAN_FILE;
1730                 goto out;
1731         }
1732 
1733         scan_balance = SCAN_FRACT;
1734 
1735         /*
1736          * With swappiness at 100, anonymous and file have the same priority.
1737          * This scanning priority is essentially the inverse of IO cost.
1738          */
1739         anon_prio = vmscan_swappiness(sc);
1740         file_prio = 200 - anon_prio;
1741 
1742         /*
1743          * OK, so we have swap space and a fair amount of page cache
1744          * pages.  We use the recently rotated / recently scanned
1745          * ratios to determine how valuable each cache is.
1746          *
1747          * Because workloads change over time (and to avoid overflow)
1748          * we keep these statistics as a floating average, which ends
1749          * up weighing recent references more than old ones.
1750          *
1751          * anon in [0], file in [1]
1752          */
1753         spin_lock_irq(&zone->lru_lock);
1754         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1755                 reclaim_stat->recent_scanned[0] /= 2;
1756                 reclaim_stat->recent_rotated[0] /= 2;
1757         }
1758 
1759         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1760                 reclaim_stat->recent_scanned[1] /= 2;
1761                 reclaim_stat->recent_rotated[1] /= 2;
1762         }
1763 
1764         /*
1765          * The amount of pressure on anon vs file pages is inversely
1766          * proportional to the fraction of recently scanned pages on
1767          * each list that were recently referenced and in active use.
1768          */
1769         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1770         ap /= reclaim_stat->recent_rotated[0] + 1;
1771 
1772         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1773         fp /= reclaim_stat->recent_rotated[1] + 1;
1774         spin_unlock_irq(&zone->lru_lock);
1775 
1776         fraction[0] = ap;
1777         fraction[1] = fp;
1778         denominator = ap + fp + 1;
1779 out:
1780         for_each_evictable_lru(lru) {
1781                 int file = is_file_lru(lru);
1782                 unsigned long size;
1783                 unsigned long scan;
1784 
1785                 size = get_lru_size(lruvec, lru);
1786                 scan = size >> sc->priority;
1787 
1788                 if (!scan && force_scan)
1789                         scan = min(size, SWAP_CLUSTER_MAX);
1790 
1791                 switch (scan_balance) {
1792                 case SCAN_EQUAL:
1793                         /* Scan lists relative to size */
1794                         break;
1795                 case SCAN_FRACT:
1796                         /*
1797                          * Scan types proportional to swappiness and
1798                          * their relative recent reclaim efficiency.
1799                          */
1800                         scan = div64_u64(scan * fraction[file], denominator);
1801                         break;
1802                 case SCAN_FILE:
1803                 case SCAN_ANON:
1804                         /* Scan one type exclusively */
1805                         if ((scan_balance == SCAN_FILE) != file)
1806                                 scan = 0;
1807                         break;
1808                 default:
1809                         /* Look ma, no brain */
1810                         BUG();
1811                 }
1812                 nr[lru] = scan;
1813         }
1814 }
1815 
1816 /*
1817  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1818  */
1819 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1820 {
1821         unsigned long nr[NR_LRU_LISTS];
1822         unsigned long nr_to_scan;
1823         enum lru_list lru;
1824         unsigned long nr_reclaimed = 0;
1825         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1826         struct blk_plug plug;
1827 
1828         get_scan_count(lruvec, sc, nr);
1829 
1830         blk_start_plug(&plug);
1831         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1832                                         nr[LRU_INACTIVE_FILE]) {
1833                 for_each_evictable_lru(lru) {
1834                         if (nr[lru]) {
1835                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1836                                 nr[lru] -= nr_to_scan;
1837 
1838                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1839                                                             lruvec, sc);
1840                         }
1841                 }
1842                 /*
1843                  * On large memory systems, scan >> priority can become
1844                  * really large. This is fine for the starting priority;
1845                  * we want to put equal scanning pressure on each zone.
1846                  * However, if the VM has a harder time of freeing pages,
1847                  * with multiple processes reclaiming pages, the total
1848                  * freeing target can get unreasonably large.
1849                  */
1850                 if (nr_reclaimed >= nr_to_reclaim &&
1851                     sc->priority < DEF_PRIORITY)
1852                         break;
1853         }
1854         blk_finish_plug(&plug);
1855         sc->nr_reclaimed += nr_reclaimed;
1856 
1857         /*
1858          * Even if we did not try to evict anon pages at all, we want to
1859          * rebalance the anon lru active/inactive ratio.
1860          */
1861         if (inactive_anon_is_low(lruvec))
1862                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1863                                    sc, LRU_ACTIVE_ANON);
1864 
1865         throttle_vm_writeout(sc->gfp_mask);
1866 }
1867 
1868 /* Use reclaim/compaction for costly allocs or under memory pressure */
1869 static bool in_reclaim_compaction(struct scan_control *sc)
1870 {
1871         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1872                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1873                          sc->priority < DEF_PRIORITY - 2))
1874                 return true;
1875 
1876         return false;
1877 }
1878 
1879 /*
1880  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1881  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1882  * true if more pages should be reclaimed such that when the page allocator
1883  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1884  * It will give up earlier than that if there is difficulty reclaiming pages.
1885  */
1886 static inline bool should_continue_reclaim(struct zone *zone,
1887                                         unsigned long nr_reclaimed,
1888                                         unsigned long nr_scanned,
1889                                         struct scan_control *sc)
1890 {
1891         unsigned long pages_for_compaction;
1892         unsigned long inactive_lru_pages;
1893 
1894         /* If not in reclaim/compaction mode, stop */
1895         if (!in_reclaim_compaction(sc))
1896                 return false;
1897 
1898         /* Consider stopping depending on scan and reclaim activity */
1899         if (sc->gfp_mask & __GFP_REPEAT) {
1900                 /*
1901                  * For __GFP_REPEAT allocations, stop reclaiming if the
1902                  * full LRU list has been scanned and we are still failing
1903                  * to reclaim pages. This full LRU scan is potentially
1904                  * expensive but a __GFP_REPEAT caller really wants to succeed
1905                  */
1906                 if (!nr_reclaimed && !nr_scanned)
1907                         return false;
1908         } else {
1909                 /*
1910                  * For non-__GFP_REPEAT allocations which can presumably
1911                  * fail without consequence, stop if we failed to reclaim
1912                  * any pages from the last SWAP_CLUSTER_MAX number of
1913                  * pages that were scanned. This will return to the
1914                  * caller faster at the risk reclaim/compaction and
1915                  * the resulting allocation attempt fails
1916                  */
1917                 if (!nr_reclaimed)
1918                         return false;
1919         }
1920 
1921         /*
1922          * If we have not reclaimed enough pages for compaction and the
1923          * inactive lists are large enough, continue reclaiming
1924          */
1925         pages_for_compaction = (2UL << sc->order);
1926         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
1927         if (get_nr_swap_pages() > 0)
1928                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
1929         if (sc->nr_reclaimed < pages_for_compaction &&
1930                         inactive_lru_pages > pages_for_compaction)
1931                 return true;
1932 
1933         /* If compaction would go ahead or the allocation would succeed, stop */
1934         switch (compaction_suitable(zone, sc->order)) {
1935         case COMPACT_PARTIAL:
1936         case COMPACT_CONTINUE:
1937                 return false;
1938         default:
1939                 return true;
1940         }
1941 }
1942 
1943 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1944 {
1945         unsigned long nr_reclaimed, nr_scanned;
1946 
1947         do {
1948                 struct mem_cgroup *root = sc->target_mem_cgroup;
1949                 struct mem_cgroup_reclaim_cookie reclaim = {
1950                         .zone = zone,
1951                         .priority = sc->priority,
1952                 };
1953                 struct mem_cgroup *memcg;
1954 
1955                 nr_reclaimed = sc->nr_reclaimed;
1956                 nr_scanned = sc->nr_scanned;
1957 
1958                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1959                 do {
1960                         struct lruvec *lruvec;
1961 
1962                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1963 
1964                         shrink_lruvec(lruvec, sc);
1965 
1966                         /*
1967                          * Direct reclaim and kswapd have to scan all memory
1968                          * cgroups to fulfill the overall scan target for the
1969                          * zone.
1970                          *
1971                          * Limit reclaim, on the other hand, only cares about
1972                          * nr_to_reclaim pages to be reclaimed and it will
1973                          * retry with decreasing priority if one round over the
1974                          * whole hierarchy is not sufficient.
1975                          */
1976                         if (!global_reclaim(sc) &&
1977                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
1978                                 mem_cgroup_iter_break(root, memcg);
1979                                 break;
1980                         }
1981                         memcg = mem_cgroup_iter(root, memcg, &reclaim);
1982                 } while (memcg);
1983 
1984                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
1985                            sc->nr_scanned - nr_scanned,
1986                            sc->nr_reclaimed - nr_reclaimed);
1987 
1988         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
1989                                          sc->nr_scanned - nr_scanned, sc));
1990 }
1991 
1992 /* Returns true if compaction should go ahead for a high-order request */
1993 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1994 {
1995         unsigned long balance_gap, watermark;
1996         bool watermark_ok;
1997 
1998         /* Do not consider compaction for orders reclaim is meant to satisfy */
1999         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2000                 return false;
2001 
2002         /*
2003          * Compaction takes time to run and there are potentially other
2004          * callers using the pages just freed. Continue reclaiming until
2005          * there is a buffer of free pages available to give compaction
2006          * a reasonable chance of completing and allocating the page
2007          */
2008         balance_gap = min(low_wmark_pages(zone),
2009                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2010                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2011         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2012         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2013 
2014         /*
2015          * If compaction is deferred, reclaim up to a point where
2016          * compaction will have a chance of success when re-enabled
2017          */
2018         if (compaction_deferred(zone, sc->order))
2019                 return watermark_ok;
2020 
2021         /* If compaction is not ready to start, keep reclaiming */
2022         if (!compaction_suitable(zone, sc->order))
2023                 return false;
2024 
2025         return watermark_ok;
2026 }
2027 
2028 /*
2029  * This is the direct reclaim path, for page-allocating processes.  We only
2030  * try to reclaim pages from zones which will satisfy the caller's allocation
2031  * request.
2032  *
2033  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2034  * Because:
2035  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2036  *    allocation or
2037  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2038  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2039  *    zone defense algorithm.
2040  *
2041  * If a zone is deemed to be full of pinned pages then just give it a light
2042  * scan then give up on it.
2043  *
2044  * This function returns true if a zone is being reclaimed for a costly
2045  * high-order allocation and compaction is ready to begin. This indicates to
2046  * the caller that it should consider retrying the allocation instead of
2047  * further reclaim.
2048  */
2049 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2050 {
2051         struct zoneref *z;
2052         struct zone *zone;
2053         unsigned long nr_soft_reclaimed;
2054         unsigned long nr_soft_scanned;
2055         bool aborted_reclaim = false;
2056 
2057         /*
2058          * If the number of buffer_heads in the machine exceeds the maximum
2059          * allowed level, force direct reclaim to scan the highmem zone as
2060          * highmem pages could be pinning lowmem pages storing buffer_heads
2061          */
2062         if (buffer_heads_over_limit)
2063                 sc->gfp_mask |= __GFP_HIGHMEM;
2064 
2065         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2066                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2067                 if (!populated_zone(zone))
2068                         continue;
2069                 /*
2070                  * Take care memory controller reclaiming has small influence
2071                  * to global LRU.
2072                  */
2073                 if (global_reclaim(sc)) {
2074                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2075                                 continue;
2076                         if (zone->all_unreclaimable &&
2077                                         sc->priority != DEF_PRIORITY)
2078                                 continue;       /* Let kswapd poll it */
2079                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2080                                 /*
2081                                  * If we already have plenty of memory free for
2082                                  * compaction in this zone, don't free any more.
2083                                  * Even though compaction is invoked for any
2084                                  * non-zero order, only frequent costly order
2085                                  * reclamation is disruptive enough to become a
2086                                  * noticeable problem, like transparent huge
2087                                  * page allocations.
2088                                  */
2089                                 if (compaction_ready(zone, sc)) {
2090                                         aborted_reclaim = true;
2091                                         continue;
2092                                 }
2093                         }
2094                         /*
2095                          * This steals pages from memory cgroups over softlimit
2096                          * and returns the number of reclaimed pages and
2097                          * scanned pages. This works for global memory pressure
2098                          * and balancing, not for a memcg's limit.
2099                          */
2100                         nr_soft_scanned = 0;
2101                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2102                                                 sc->order, sc->gfp_mask,
2103                                                 &nr_soft_scanned);
2104                         sc->nr_reclaimed += nr_soft_reclaimed;
2105                         sc->nr_scanned += nr_soft_scanned;
2106                         /* need some check for avoid more shrink_zone() */
2107                 }
2108 
2109                 shrink_zone(zone, sc);
2110         }
2111 
2112         return aborted_reclaim;
2113 }
2114 
2115 static unsigned long zone_reclaimable_pages(struct zone *zone)
2116 {
2117         int nr;
2118 
2119         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2120              zone_page_state(zone, NR_INACTIVE_FILE);
2121 
2122         if (get_nr_swap_pages() > 0)
2123                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2124                       zone_page_state(zone, NR_INACTIVE_ANON);
2125 
2126         return nr;
2127 }
2128 
2129 static bool zone_reclaimable(struct zone *zone)
2130 {
2131         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2132 }
2133 
2134 /* All zones in zonelist are unreclaimable? */
2135 static bool all_unreclaimable(struct zonelist *zonelist,
2136                 struct scan_control *sc)
2137 {
2138         struct zoneref *z;
2139         struct zone *zone;
2140 
2141         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2142                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2143                 if (!populated_zone(zone))
2144                         continue;
2145                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2146                         continue;
2147                 if (!zone->all_unreclaimable)
2148                         return false;
2149         }
2150 
2151         return true;
2152 }
2153 
2154 /*
2155  * This is the main entry point to direct page reclaim.
2156  *
2157  * If a full scan of the inactive list fails to free enough memory then we
2158  * are "out of memory" and something needs to be killed.
2159  *
2160  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2161  * high - the zone may be full of dirty or under-writeback pages, which this
2162  * caller can't do much about.  We kick the writeback threads and take explicit
2163  * naps in the hope that some of these pages can be written.  But if the
2164  * allocating task holds filesystem locks which prevent writeout this might not
2165  * work, and the allocation attempt will fail.
2166  *
2167  * returns:     0, if no pages reclaimed
2168  *              else, the number of pages reclaimed
2169  */
2170 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2171                                         struct scan_control *sc,
2172                                         struct shrink_control *shrink)
2173 {
2174         unsigned long total_scanned = 0;
2175         struct reclaim_state *reclaim_state = current->reclaim_state;
2176         struct zoneref *z;
2177         struct zone *zone;
2178         unsigned long writeback_threshold;
2179         bool aborted_reclaim;
2180 
2181         delayacct_freepages_start();
2182 
2183         if (global_reclaim(sc))
2184                 count_vm_event(ALLOCSTALL);
2185 
2186         do {
2187                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2188                                 sc->priority);
2189                 sc->nr_scanned = 0;
2190                 aborted_reclaim = shrink_zones(zonelist, sc);
2191 
2192                 /*
2193                  * Don't shrink slabs when reclaiming memory from
2194                  * over limit cgroups
2195                  */
2196                 if (global_reclaim(sc)) {
2197                         unsigned long lru_pages = 0;
2198                         for_each_zone_zonelist(zone, z, zonelist,
2199                                         gfp_zone(sc->gfp_mask)) {
2200                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2201                                         continue;
2202 
2203                                 lru_pages += zone_reclaimable_pages(zone);
2204                         }
2205 
2206                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2207                         if (reclaim_state) {
2208                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2209                                 reclaim_state->reclaimed_slab = 0;
2210                         }
2211                 }
2212                 total_scanned += sc->nr_scanned;
2213                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2214                         goto out;
2215 
2216                 /*
2217                  * If we're getting trouble reclaiming, start doing
2218                  * writepage even in laptop mode.
2219                  */
2220                 if (sc->priority < DEF_PRIORITY - 2)
2221                         sc->may_writepage = 1;
2222 
2223                 /*
2224                  * Try to write back as many pages as we just scanned.  This
2225                  * tends to cause slow streaming writers to write data to the
2226                  * disk smoothly, at the dirtying rate, which is nice.   But
2227                  * that's undesirable in laptop mode, where we *want* lumpy
2228                  * writeout.  So in laptop mode, write out the whole world.
2229                  */
2230                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2231                 if (total_scanned > writeback_threshold) {
2232                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2233                                                 WB_REASON_TRY_TO_FREE_PAGES);
2234                         sc->may_writepage = 1;
2235                 }
2236 
2237                 /* Take a nap, wait for some writeback to complete */
2238                 if (!sc->hibernation_mode && sc->nr_scanned &&
2239                     sc->priority < DEF_PRIORITY - 2) {
2240                         struct zone *preferred_zone;
2241 
2242                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2243                                                 &cpuset_current_mems_allowed,
2244                                                 &preferred_zone);
2245                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2246                 }
2247         } while (--sc->priority >= 0);
2248 
2249 out:
2250         delayacct_freepages_end();
2251 
2252         if (sc->nr_reclaimed)
2253                 return sc->nr_reclaimed;
2254 
2255         /*
2256          * As hibernation is going on, kswapd is freezed so that it can't mark
2257          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2258          * check.
2259          */
2260         if (oom_killer_disabled)
2261                 return 0;
2262 
2263         /* Aborted reclaim to try compaction? don't OOM, then */
2264         if (aborted_reclaim)
2265                 return 1;
2266 
2267         /* top priority shrink_zones still had more to do? don't OOM, then */
2268         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2269                 return 1;
2270 
2271         return 0;
2272 }
2273 
2274 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2275 {
2276         struct zone *zone;
2277         unsigned long pfmemalloc_reserve = 0;
2278         unsigned long free_pages = 0;
2279         int i;
2280         bool wmark_ok;
2281 
2282         for (i = 0; i <= ZONE_NORMAL; i++) {
2283                 zone = &pgdat->node_zones[i];
2284                 if (!populated_zone(zone))
2285                         continue;
2286 
2287                 pfmemalloc_reserve += min_wmark_pages(zone);
2288                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2289         }
2290 
2291         /* If there are no reserves (unexpected config) then do not throttle */
2292         if (!pfmemalloc_reserve)
2293                 return true;
2294 
2295         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2296 
2297         /* kswapd must be awake if processes are being throttled */
2298         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2299                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2300                                                 (enum zone_type)ZONE_NORMAL);
2301                 wake_up_interruptible(&pgdat->kswapd_wait);
2302         }
2303 
2304         return wmark_ok;
2305 }
2306 
2307 /*
2308  * Throttle direct reclaimers if backing storage is backed by the network
2309  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2310  * depleted. kswapd will continue to make progress and wake the processes
2311  * when the low watermark is reached.
2312  *
2313  * Returns true if a fatal signal was delivered during throttling. If this
2314  * happens, the page allocator should not consider triggering the OOM killer.
2315  */
2316 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2317                                         nodemask_t *nodemask)
2318 {
2319         struct zoneref *z;
2320         struct zone *zone;
2321         pg_data_t *pgdat = NULL;
2322 
2323         /*
2324          * Kernel threads should not be throttled as they may be indirectly
2325          * responsible for cleaning pages necessary for reclaim to make forward
2326          * progress. kjournald for example may enter direct reclaim while
2327          * committing a transaction where throttling it could forcing other
2328          * processes to block on log_wait_commit().
2329          */
2330         if (current->flags & PF_KTHREAD)
2331                 goto out;
2332 
2333         /*
2334          * If a fatal signal is pending, this process should not throttle.
2335          * It should return quickly so it can exit and free its memory
2336          */
2337         if (fatal_signal_pending(current))
2338                 goto out;
2339 
2340         /*
2341          * Check if the pfmemalloc reserves are ok by finding the first node
2342          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2343          * GFP_KERNEL will be required for allocating network buffers when
2344          * swapping over the network so ZONE_HIGHMEM is unusable.
2345          *
2346          * Throttling is based on the first usable node and throttled processes
2347          * wait on a queue until kswapd makes progress and wakes them. There
2348          * is an affinity then between processes waking up and where reclaim
2349          * progress has been made assuming the process wakes on the same node.
2350          * More importantly, processes running on remote nodes will not compete
2351          * for remote pfmemalloc reserves and processes on different nodes
2352          * should make reasonable progress.
2353          */
2354         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2355                                         gfp_mask, nodemask) {
2356                 if (zone_idx(zone) > ZONE_NORMAL)
2357                         continue;
2358 
2359                 /* Throttle based on the first usable node */
2360                 pgdat = zone->zone_pgdat;
2361                 if (pfmemalloc_watermark_ok(pgdat))
2362                         goto out;
2363                 break;
2364         }
2365 
2366         /* If no zone was usable by the allocation flags then do not throttle */
2367         if (!pgdat)
2368                 goto out;
2369 
2370         /* Account for the throttling */
2371         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2372 
2373         /*
2374          * If the caller cannot enter the filesystem, it's possible that it
2375          * is due to the caller holding an FS lock or performing a journal
2376          * transaction in the case of a filesystem like ext[3|4]. In this case,
2377          * it is not safe to block on pfmemalloc_wait as kswapd could be
2378          * blocked waiting on the same lock. Instead, throttle for up to a
2379          * second before continuing.
2380          */
2381         if (!(gfp_mask & __GFP_FS)) {
2382                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2383                         pfmemalloc_watermark_ok(pgdat), HZ);
2384 
2385                 goto check_pending;
2386         }
2387 
2388         /* Throttle until kswapd wakes the process */
2389         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2390                 pfmemalloc_watermark_ok(pgdat));
2391 
2392 check_pending:
2393         if (fatal_signal_pending(current))
2394                 return true;
2395 
2396 out:
2397         return false;
2398 }
2399 
2400 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2401                                 gfp_t gfp_mask, nodemask_t *nodemask)
2402 {
2403         unsigned long nr_reclaimed;
2404         struct scan_control sc = {
2405                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2406                 .may_writepage = !laptop_mode,
2407                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2408                 .may_unmap = 1,
2409                 .may_swap = 1,
2410                 .order = order,
2411                 .priority = DEF_PRIORITY,
2412                 .target_mem_cgroup = NULL,
2413                 .nodemask = nodemask,
2414         };
2415         struct shrink_control shrink = {
2416                 .gfp_mask = sc.gfp_mask,
2417         };
2418 
2419         /*
2420          * Do not enter reclaim if fatal signal was delivered while throttled.
2421          * 1 is returned so that the page allocator does not OOM kill at this
2422          * point.
2423          */
2424         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2425                 return 1;
2426 
2427         trace_mm_vmscan_direct_reclaim_begin(order,
2428                                 sc.may_writepage,
2429                                 gfp_mask);
2430 
2431         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2432 
2433         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2434 
2435         return nr_reclaimed;
2436 }
2437 
2438 #ifdef CONFIG_MEMCG
2439 
2440 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2441                                                 gfp_t gfp_mask, bool noswap,
2442                                                 struct zone *zone,
2443                                                 unsigned long *nr_scanned)
2444 {
2445         struct scan_control sc = {
2446                 .nr_scanned = 0,
2447                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2448                 .may_writepage = !laptop_mode,
2449                 .may_unmap = 1,
2450                 .may_swap = !noswap,
2451                 .order = 0,
2452                 .priority = 0,
2453                 .target_mem_cgroup = memcg,
2454         };
2455         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2456 
2457         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2458                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2459 
2460         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2461                                                       sc.may_writepage,
2462                                                       sc.gfp_mask);
2463 
2464         /*
2465          * NOTE: Although we can get the priority field, using it
2466          * here is not a good idea, since it limits the pages we can scan.
2467          * if we don't reclaim here, the shrink_zone from balance_pgdat
2468          * will pick up pages from other mem cgroup's as well. We hack
2469          * the priority and make it zero.
2470          */
2471         shrink_lruvec(lruvec, &sc);
2472 
2473         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2474 
2475         *nr_scanned = sc.nr_scanned;
2476         return sc.nr_reclaimed;
2477 }
2478 
2479 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2480                                            gfp_t gfp_mask,
2481                                            bool noswap)
2482 {
2483         struct zonelist *zonelist;
2484         unsigned long nr_reclaimed;
2485         int nid;
2486         struct scan_control sc = {
2487                 .may_writepage = !laptop_mode,
2488                 .may_unmap = 1,
2489                 .may_swap = !noswap,
2490                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2491                 .order = 0,
2492                 .priority = DEF_PRIORITY,
2493                 .target_mem_cgroup = memcg,
2494                 .nodemask = NULL, /* we don't care the placement */
2495                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2496                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2497         };
2498         struct shrink_control shrink = {
2499                 .gfp_mask = sc.gfp_mask,
2500         };
2501 
2502         /*
2503          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2504          * take care of from where we get pages. So the node where we start the
2505          * scan does not need to be the current node.
2506          */
2507         nid = mem_cgroup_select_victim_node(memcg);
2508 
2509         zonelist = NODE_DATA(nid)->node_zonelists;
2510 
2511         trace_mm_vmscan_memcg_reclaim_begin(0,
2512                                             sc.may_writepage,
2513                                             sc.gfp_mask);
2514 
2515         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2516 
2517         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2518 
2519         return nr_reclaimed;
2520 }
2521 #endif
2522 
2523 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2524 {
2525         struct mem_cgroup *memcg;
2526 
2527         if (!total_swap_pages)
2528                 return;
2529 
2530         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2531         do {
2532                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2533 
2534                 if (inactive_anon_is_low(lruvec))
2535                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2536                                            sc, LRU_ACTIVE_ANON);
2537 
2538                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2539         } while (memcg);
2540 }
2541 
2542 static bool zone_balanced(struct zone *zone, int order,
2543                           unsigned long balance_gap, int classzone_idx)
2544 {
2545         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2546                                     balance_gap, classzone_idx, 0))
2547                 return false;
2548 
2549         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2550             !compaction_suitable(zone, order))
2551                 return false;
2552 
2553         return true;
2554 }
2555 
2556 /*
2557  * pgdat_balanced() is used when checking if a node is balanced.
2558  *
2559  * For order-0, all zones must be balanced!
2560  *
2561  * For high-order allocations only zones that meet watermarks and are in a
2562  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2563  * total of balanced pages must be at least 25% of the zones allowed by
2564  * classzone_idx for the node to be considered balanced. Forcing all zones to
2565  * be balanced for high orders can cause excessive reclaim when there are
2566  * imbalanced zones.
2567  * The choice of 25% is due to
2568  *   o a 16M DMA zone that is balanced will not balance a zone on any
2569  *     reasonable sized machine
2570  *   o On all other machines, the top zone must be at least a reasonable
2571  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2572  *     would need to be at least 256M for it to be balance a whole node.
2573  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2574  *     to balance a node on its own. These seemed like reasonable ratios.
2575  */
2576 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2577 {
2578         unsigned long managed_pages = 0;
2579         unsigned long balanced_pages = 0;
2580         int i;
2581 
2582         /* Check the watermark levels */
2583         for (i = 0; i <= classzone_idx; i++) {
2584                 struct zone *zone = pgdat->node_zones + i;
2585 
2586                 if (!populated_zone(zone))
2587                         continue;
2588 
2589                 managed_pages += zone->managed_pages;
2590 
2591                 /*
2592                  * A special case here:
2593                  *
2594                  * balance_pgdat() skips over all_unreclaimable after
2595                  * DEF_PRIORITY. Effectively, it considers them balanced so
2596                  * they must be considered balanced here as well!
2597                  */
2598                 if (zone->all_unreclaimable) {
2599                         balanced_pages += zone->managed_pages;
2600                         continue;
2601                 }
2602 
2603                 if (zone_balanced(zone, order, 0, i))
2604                         balanced_pages += zone->managed_pages;
2605                 else if (!order)
2606                         return false;
2607         }
2608 
2609         if (order)
2610                 return balanced_pages >= (managed_pages >> 2);
2611         else
2612                 return true;
2613 }
2614 
2615 /*
2616  * Prepare kswapd for sleeping. This verifies that there are no processes
2617  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2618  *
2619  * Returns true if kswapd is ready to sleep
2620  */
2621 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2622                                         int classzone_idx)
2623 {
2624         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2625         if (remaining)
2626                 return false;
2627 
2628         /*
2629          * The throttled processes are normally woken up in balance_pgdat() as
2630          * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2631          * race between when kswapd checks the watermarks and a process gets
2632          * throttled. There is also a potential race if processes get
2633          * throttled, kswapd wakes, a large process exits thereby balancing the
2634          * zones, which causes kswapd to exit balance_pgdat() before reaching
2635          * the wake up checks. If kswapd is going to sleep, no process should
2636          * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2637          * the wake up is premature, processes will wake kswapd and get
2638          * throttled again. The difference from wake ups in balance_pgdat() is
2639          * that here we are under prepare_to_wait().
2640          */
2641         if (waitqueue_active(&pgdat->pfmemalloc_wait))
2642                 wake_up_all(&pgdat->pfmemalloc_wait);
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         tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3091         current->reclaim_state = NULL;
3092         lockdep_clear_current_reclaim_state();
3093 
3094         return 0;
3095 }
3096 
3097 /*
3098  * A zone is low on free memory, so wake its kswapd task to service it.
3099  */
3100 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3101 {
3102         pg_data_t *pgdat;
3103 
3104         if (!populated_zone(zone))
3105                 return;
3106 
3107         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3108                 return;
3109         pgdat = zone->zone_pgdat;
3110         if (pgdat->kswapd_max_order < order) {
3111                 pgdat->kswapd_max_order = order;
3112                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3113         }
3114         if (!waitqueue_active(&pgdat->kswapd_wait))
3115                 return;
3116         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3117                 return;
3118 
3119         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3120         wake_up_interruptible(&pgdat->kswapd_wait);
3121 }
3122 
3123 #ifdef CONFIG_HIBERNATION
3124 /*
3125  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3126  * freed pages.
3127  *
3128  * Rather than trying to age LRUs the aim is to preserve the overall
3129  * LRU order by reclaiming preferentially
3130  * inactive > active > active referenced > active mapped
3131  */
3132 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3133 {
3134         struct reclaim_state reclaim_state;
3135         struct scan_control sc = {
3136                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3137                 .may_swap = 1,
3138                 .may_unmap = 1,
3139                 .may_writepage = 1,
3140                 .nr_to_reclaim = nr_to_reclaim,
3141                 .hibernation_mode = 1,
3142                 .order = 0,
3143                 .priority = DEF_PRIORITY,
3144         };
3145         struct shrink_control shrink = {
3146                 .gfp_mask = sc.gfp_mask,
3147         };
3148         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3149         struct task_struct *p = current;
3150         unsigned long nr_reclaimed;
3151 
3152         p->flags |= PF_MEMALLOC;
3153         lockdep_set_current_reclaim_state(sc.gfp_mask);
3154         reclaim_state.reclaimed_slab = 0;
3155         p->reclaim_state = &reclaim_state;
3156 
3157         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3158 
3159         p->reclaim_state = NULL;
3160         lockdep_clear_current_reclaim_state();
3161         p->flags &= ~PF_MEMALLOC;
3162 
3163         return nr_reclaimed;
3164 }
3165 #endif /* CONFIG_HIBERNATION */
3166 
3167 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3168    not required for correctness.  So if the last cpu in a node goes
3169    away, we get changed to run anywhere: as the first one comes back,
3170    restore their cpu bindings. */
3171 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3172                         void *hcpu)
3173 {
3174         int nid;
3175 
3176         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3177                 for_each_node_state(nid, N_MEMORY) {
3178                         pg_data_t *pgdat = NODE_DATA(nid);
3179                         const struct cpumask *mask;
3180 
3181                         mask = cpumask_of_node(pgdat->node_id);
3182 
3183                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3184                                 /* One of our CPUs online: restore mask */
3185                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3186                 }
3187         }
3188         return NOTIFY_OK;
3189 }
3190 
3191 /*
3192  * This kswapd start function will be called by init and node-hot-add.
3193  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3194  */
3195 int kswapd_run(int nid)
3196 {
3197         pg_data_t *pgdat = NODE_DATA(nid);
3198         int ret = 0;
3199 
3200         if (pgdat->kswapd)
3201                 return 0;
3202 
3203         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3204         if (IS_ERR(pgdat->kswapd)) {
3205                 /* failure at boot is fatal */
3206                 BUG_ON(system_state == SYSTEM_BOOTING);
3207                 pr_err("Failed to start kswapd on node %d\n", nid);
3208                 ret = PTR_ERR(pgdat->kswapd);
3209                 pgdat->kswapd = NULL;
3210         }
3211         return ret;
3212 }
3213 
3214 /*
3215  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3216  * hold lock_memory_hotplug().
3217  */
3218 void kswapd_stop(int nid)
3219 {
3220         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3221 
3222         if (kswapd) {
3223                 kthread_stop(kswapd);
3224                 NODE_DATA(nid)->kswapd = NULL;
3225         }
3226 }
3227 
3228 static int __init kswapd_init(void)
3229 {
3230         int nid;
3231 
3232         swap_setup();
3233         for_each_node_state(nid, N_MEMORY)
3234                 kswapd_run(nid);
3235         hotcpu_notifier(cpu_callback, 0);
3236         return 0;
3237 }
3238 
3239 module_init(kswapd_init)
3240 
3241 #ifdef CONFIG_NUMA
3242 /*
3243  * Zone reclaim mode
3244  *
3245  * If non-zero call zone_reclaim when the number of free pages falls below
3246  * the watermarks.
3247  */
3248 int zone_reclaim_mode __read_mostly;
3249 
3250 #define RECLAIM_OFF 0
3251 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3252 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3253 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3254 
3255 /*
3256  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3257  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3258  * a zone.
3259  */
3260 #define ZONE_RECLAIM_PRIORITY 4
3261 
3262 /*
3263  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3264  * occur.
3265  */
3266 int sysctl_min_unmapped_ratio = 1;
3267 
3268 /*
3269  * If the number of slab pages in a zone grows beyond this percentage then
3270  * slab reclaim needs to occur.
3271  */
3272 int sysctl_min_slab_ratio = 5;
3273 
3274 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3275 {
3276         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3277         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3278                 zone_page_state(zone, NR_ACTIVE_FILE);
3279 
3280         /*
3281          * It's possible for there to be more file mapped pages than
3282          * accounted for by the pages on the file LRU lists because
3283          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3284          */
3285         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3286 }
3287 
3288 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3289 static long zone_pagecache_reclaimable(struct zone *zone)
3290 {
3291         long nr_pagecache_reclaimable;
3292         long delta = 0;
3293 
3294         /*
3295          * If RECLAIM_SWAP is set, then all file pages are considered
3296          * potentially reclaimable. Otherwise, we have to worry about
3297          * pages like swapcache and zone_unmapped_file_pages() provides
3298          * a better estimate
3299          */
3300         if (zone_reclaim_mode & RECLAIM_SWAP)
3301                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3302         else
3303                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3304 
3305         /* If we can't clean pages, remove dirty pages from consideration */
3306         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3307                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3308 
3309         /* Watch for any possible underflows due to delta */
3310         if (unlikely(delta > nr_pagecache_reclaimable))
3311                 delta = nr_pagecache_reclaimable;
3312 
3313         return nr_pagecache_reclaimable - delta;
3314 }
3315 
3316 /*
3317  * Try to free up some pages from this zone through reclaim.
3318  */
3319 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3320 {
3321         /* Minimum pages needed in order to stay on node */
3322         const unsigned long nr_pages = 1 << order;
3323         struct task_struct *p = current;
3324         struct reclaim_state reclaim_state;
3325         struct scan_control sc = {
3326                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3327                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3328                 .may_swap = 1,
3329                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3330                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3331                 .order = order,
3332                 .priority = ZONE_RECLAIM_PRIORITY,
3333         };
3334         struct shrink_control shrink = {
3335                 .gfp_mask = sc.gfp_mask,
3336         };
3337         unsigned long nr_slab_pages0, nr_slab_pages1;
3338 
3339         cond_resched();
3340         /*
3341          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3342          * and we also need to be able to write out pages for RECLAIM_WRITE
3343          * and RECLAIM_SWAP.
3344          */
3345         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3346         lockdep_set_current_reclaim_state(gfp_mask);
3347         reclaim_state.reclaimed_slab = 0;
3348         p->reclaim_state = &reclaim_state;
3349 
3350         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3351                 /*
3352                  * Free memory by calling shrink zone with increasing
3353                  * priorities until we have enough memory freed.
3354                  */
3355                 do {
3356                         shrink_zone(zone, &sc);
3357                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3358         }
3359 
3360         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3361         if (nr_slab_pages0 > zone->min_slab_pages) {
3362                 /*
3363                  * shrink_slab() does not currently allow us to determine how
3364                  * many pages were freed in this zone. So we take the current
3365                  * number of slab pages and shake the slab until it is reduced
3366                  * by the same nr_pages that we used for reclaiming unmapped
3367                  * pages.
3368                  *
3369                  * Note that shrink_slab will free memory on all zones and may
3370                  * take a long time.
3371                  */
3372                 for (;;) {
3373                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3374 
3375                         /* No reclaimable slab or very low memory pressure */
3376                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3377                                 break;
3378 
3379                         /* Freed enough memory */
3380                         nr_slab_pages1 = zone_page_state(zone,
3381                                                         NR_SLAB_RECLAIMABLE);
3382                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3383                                 break;
3384                 }
3385 
3386                 /*
3387                  * Update nr_reclaimed by the number of slab pages we
3388                  * reclaimed from this zone.
3389                  */
3390                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3391                 if (nr_slab_pages1 < nr_slab_pages0)
3392                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3393         }
3394 
3395         p->reclaim_state = NULL;
3396         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3397         lockdep_clear_current_reclaim_state();
3398         return sc.nr_reclaimed >= nr_pages;
3399 }
3400 
3401 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3402 {
3403         int node_id;
3404         int ret;
3405 
3406         /*
3407          * Zone reclaim reclaims unmapped file backed pages and
3408          * slab pages if we are over the defined limits.
3409          *
3410          * A small portion of unmapped file backed pages is needed for
3411          * file I/O otherwise pages read by file I/O will be immediately
3412          * thrown out if the zone is overallocated. So we do not reclaim
3413          * if less than a specified percentage of the zone is used by
3414          * unmapped file backed pages.
3415          */
3416         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3417             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3418                 return ZONE_RECLAIM_FULL;
3419 
3420         if (zone->all_unreclaimable)
3421                 return ZONE_RECLAIM_FULL;
3422 
3423         /*
3424          * Do not scan if the allocation should not be delayed.
3425          */
3426         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3427                 return ZONE_RECLAIM_NOSCAN;
3428 
3429         /*
3430          * Only run zone reclaim on the local zone or on zones that do not
3431          * have associated processors. This will favor the local processor
3432          * over remote processors and spread off node memory allocations
3433          * as wide as possible.
3434          */
3435         node_id = zone_to_nid(zone);
3436         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3437                 return ZONE_RECLAIM_NOSCAN;
3438 
3439         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3440                 return ZONE_RECLAIM_NOSCAN;
3441 
3442         ret = __zone_reclaim(zone, gfp_mask, order);
3443         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3444 
3445         if (!ret)
3446                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3447 
3448         return ret;
3449 }
3450 #endif
3451 
3452 /*
3453  * page_evictable - test whether a page is evictable
3454  * @page: the page to test
3455  *
3456  * Test whether page is evictable--i.e., should be placed on active/inactive
3457  * lists vs unevictable list.
3458  *
3459  * Reasons page might not be evictable:
3460  * (1) page's mapping marked unevictable
3461  * (2) page is part of an mlocked VMA
3462  *
3463  */
3464 int page_evictable(struct page *page)
3465 {
3466         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3467 }
3468 
3469 #ifdef CONFIG_SHMEM
3470 /**
3471  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3472  * @pages:      array of pages to check
3473  * @nr_pages:   number of pages to check
3474  *
3475  * Checks pages for evictability and moves them to the appropriate lru list.
3476  *
3477  * This function is only used for SysV IPC SHM_UNLOCK.
3478  */
3479 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3480 {
3481         struct lruvec *lruvec;
3482         struct zone *zone = NULL;
3483         int pgscanned = 0;
3484         int pgrescued = 0;
3485         int i;
3486 
3487         for (i = 0; i < nr_pages; i++) {
3488                 struct page *page = pages[i];
3489                 struct zone *pagezone;
3490 
3491                 pgscanned++;
3492                 pagezone = page_zone(page);
3493                 if (pagezone != zone) {
3494                         if (zone)
3495                                 spin_unlock_irq(&zone->lru_lock);
3496                         zone = pagezone;
3497                         spin_lock_irq(&zone->lru_lock);
3498                 }
3499                 lruvec = mem_cgroup_page_lruvec(page, zone);
3500 
3501                 if (!PageLRU(page) || !PageUnevictable(page))
3502                         continue;
3503 
3504                 if (page_evictable(page)) {
3505                         enum lru_list lru = page_lru_base_type(page);
3506 
3507                         VM_BUG_ON(PageActive(page));
3508                         ClearPageUnevictable(page);
3509                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3510                         add_page_to_lru_list(page, lruvec, lru);
3511                         pgrescued++;
3512                 }
3513         }
3514 
3515         if (zone) {
3516                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3517                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3518                 spin_unlock_irq(&zone->lru_lock);
3519         }
3520 }
3521 #endif /* CONFIG_SHMEM */
3522 
3523 static void warn_scan_unevictable_pages(void)
3524 {
3525         printk_once(KERN_WARNING
3526                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3527                     "disabled for lack of a legitimate use case.  If you have "
3528                     "one, please send an email to linux-mm@kvack.org.\n",
3529                     current->comm);
3530 }
3531 
3532 /*
3533  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3534  * all nodes' unevictable lists for evictable pages
3535  */
3536 unsigned long scan_unevictable_pages;
3537 
3538 int scan_unevictable_handler(struct ctl_table *table, int write,
3539                            void __user *buffer,
3540                            size_t *length, loff_t *ppos)
3541 {
3542         warn_scan_unevictable_pages();
3543         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3544         scan_unevictable_pages = 0;
3545         return 0;
3546 }
3547 
3548 #ifdef CONFIG_NUMA
3549 /*
3550  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3551  * a specified node's per zone unevictable lists for evictable pages.
3552  */
3553 
3554 static ssize_t read_scan_unevictable_node(struct device *dev,
3555                                           struct device_attribute *attr,
3556                                           char *buf)
3557 {
3558         warn_scan_unevictable_pages();
3559         return sprintf(buf, "\n");     /* always zero; should fit... */
3560 }
3561 
3562 static ssize_t write_scan_unevictable_node(struct device *dev,
3563                                            struct device_attribute *attr,
3564                                         const char *buf, size_t count)
3565 {
3566         warn_scan_unevictable_pages();
3567         return 1;
3568 }
3569 
3570 
3571 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3572                         read_scan_unevictable_node,
3573                         write_scan_unevictable_node);
3574 
3575 int scan_unevictable_register_node(struct node *node)
3576 {
3577         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3578 }
3579 
3580 void scan_unevictable_unregister_node(struct node *node)
3581 {
3582         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3583 }
3584 #endif
3585 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp