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

TOMOYO Linux Cross Reference
Linux/mm/vmscan.c

Version: ~ [ linux-5.13-rc7 ] ~ [ linux-5.12.12 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.45 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.127 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.195 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.237 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.273 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.273 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ 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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 15 
 16 #include <linux/mm.h>
 17 #include <linux/module.h>
 18 #include <linux/gfp.h>
 19 #include <linux/kernel_stat.h>
 20 #include <linux/swap.h>
 21 #include <linux/pagemap.h>
 22 #include <linux/init.h>
 23 #include <linux/highmem.h>
 24 #include <linux/vmpressure.h>
 25 #include <linux/vmstat.h>
 26 #include <linux/file.h>
 27 #include <linux/writeback.h>
 28 #include <linux/blkdev.h>
 29 #include <linux/buffer_head.h>  /* for try_to_release_page(),
 30                                         buffer_heads_over_limit */
 31 #include <linux/mm_inline.h>
 32 #include <linux/backing-dev.h>
 33 #include <linux/rmap.h>
 34 #include <linux/topology.h>
 35 #include <linux/cpu.h>
 36 #include <linux/cpuset.h>
 37 #include <linux/compaction.h>
 38 #include <linux/notifier.h>
 39 #include <linux/rwsem.h>
 40 #include <linux/delay.h>
 41 #include <linux/kthread.h>
 42 #include <linux/freezer.h>
 43 #include <linux/memcontrol.h>
 44 #include <linux/delayacct.h>
 45 #include <linux/sysctl.h>
 46 #include <linux/oom.h>
 47 #include <linux/prefetch.h>
 48 #include <linux/printk.h>
 49 
 50 #include <asm/tlbflush.h>
 51 #include <asm/div64.h>
 52 
 53 #include <linux/swapops.h>
 54 #include <linux/balloon_compaction.h>
 55 
 56 #include "internal.h"
 57 
 58 #define CREATE_TRACE_POINTS
 59 #include <trace/events/vmscan.h>
 60 
 61 struct scan_control {
 62         /* How many pages shrink_list() should reclaim */
 63         unsigned long nr_to_reclaim;
 64 
 65         /* This context's GFP mask */
 66         gfp_t gfp_mask;
 67 
 68         /* Allocation order */
 69         int order;
 70 
 71         /*
 72          * Nodemask of nodes allowed by the caller. If NULL, all nodes
 73          * are scanned.
 74          */
 75         nodemask_t      *nodemask;
 76 
 77         /*
 78          * The memory cgroup that hit its limit and as a result is the
 79          * primary target of this reclaim invocation.
 80          */
 81         struct mem_cgroup *target_mem_cgroup;
 82 
 83         /* Scan (total_size >> priority) pages at once */
 84         int priority;
 85 
 86         unsigned int may_writepage:1;
 87 
 88         /* Can mapped pages be reclaimed? */
 89         unsigned int may_unmap:1;
 90 
 91         /* Can pages be swapped as part of reclaim? */
 92         unsigned int may_swap:1;
 93 
 94         /* Can cgroups be reclaimed below their normal consumption range? */
 95         unsigned int may_thrash:1;
 96 
 97         unsigned int hibernation_mode:1;
 98 
 99         /* One of the zones is ready for compaction */
100         unsigned int compaction_ready:1;
101 
102         /* Incremented by the number of inactive pages that were scanned */
103         unsigned long nr_scanned;
104 
105         /* Number of pages freed so far during a call to shrink_zones() */
106         unsigned long nr_reclaimed;
107 };
108 
109 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
110 
111 #ifdef ARCH_HAS_PREFETCH
112 #define prefetch_prev_lru_page(_page, _base, _field)                    \
113         do {                                                            \
114                 if ((_page)->lru.prev != _base) {                       \
115                         struct page *prev;                              \
116                                                                         \
117                         prev = lru_to_page(&(_page->lru));              \
118                         prefetch(&prev->_field);                        \
119                 }                                                       \
120         } while (0)
121 #else
122 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
123 #endif
124 
125 #ifdef ARCH_HAS_PREFETCHW
126 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
127         do {                                                            \
128                 if ((_page)->lru.prev != _base) {                       \
129                         struct page *prev;                              \
130                                                                         \
131                         prev = lru_to_page(&(_page->lru));              \
132                         prefetchw(&prev->_field);                       \
133                 }                                                       \
134         } while (0)
135 #else
136 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137 #endif
138 
139 /*
140  * From 0 .. 100.  Higher means more swappy.
141  */
142 int vm_swappiness = 60;
143 /*
144  * The total number of pages which are beyond the high watermark within all
145  * zones.
146  */
147 unsigned long vm_total_pages;
148 
149 static LIST_HEAD(shrinker_list);
150 static DECLARE_RWSEM(shrinker_rwsem);
151 
152 #ifdef CONFIG_MEMCG
153 static bool global_reclaim(struct scan_control *sc)
154 {
155         return !sc->target_mem_cgroup;
156 }
157 
158 /**
159  * sane_reclaim - is the usual dirty throttling mechanism operational?
160  * @sc: scan_control in question
161  *
162  * The normal page dirty throttling mechanism in balance_dirty_pages() is
163  * completely broken with the legacy memcg and direct stalling in
164  * shrink_page_list() is used for throttling instead, which lacks all the
165  * niceties such as fairness, adaptive pausing, bandwidth proportional
166  * allocation and configurability.
167  *
168  * This function tests whether the vmscan currently in progress can assume
169  * that the normal dirty throttling mechanism is operational.
170  */
171 static bool sane_reclaim(struct scan_control *sc)
172 {
173         struct mem_cgroup *memcg = sc->target_mem_cgroup;
174 
175         if (!memcg)
176                 return true;
177 #ifdef CONFIG_CGROUP_WRITEBACK
178         if (cgroup_on_dfl(memcg->css.cgroup))
179                 return true;
180 #endif
181         return false;
182 }
183 #else
184 static bool global_reclaim(struct scan_control *sc)
185 {
186         return true;
187 }
188 
189 static bool sane_reclaim(struct scan_control *sc)
190 {
191         return true;
192 }
193 #endif
194 
195 static unsigned long zone_reclaimable_pages(struct zone *zone)
196 {
197         int nr;
198 
199         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
200              zone_page_state(zone, NR_INACTIVE_FILE);
201 
202         if (get_nr_swap_pages() > 0)
203                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
204                       zone_page_state(zone, NR_INACTIVE_ANON);
205 
206         return nr;
207 }
208 
209 bool zone_reclaimable(struct zone *zone)
210 {
211         return zone_page_state(zone, NR_PAGES_SCANNED) <
212                 zone_reclaimable_pages(zone) * 6;
213 }
214 
215 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
216 {
217         if (!mem_cgroup_disabled())
218                 return mem_cgroup_get_lru_size(lruvec, lru);
219 
220         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
221 }
222 
223 /*
224  * Add a shrinker callback to be called from the vm.
225  */
226 int register_shrinker(struct shrinker *shrinker)
227 {
228         size_t size = sizeof(*shrinker->nr_deferred);
229 
230         /*
231          * If we only have one possible node in the system anyway, save
232          * ourselves the trouble and disable NUMA aware behavior. This way we
233          * will save memory and some small loop time later.
234          */
235         if (nr_node_ids == 1)
236                 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
237 
238         if (shrinker->flags & SHRINKER_NUMA_AWARE)
239                 size *= nr_node_ids;
240 
241         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
242         if (!shrinker->nr_deferred)
243                 return -ENOMEM;
244 
245         down_write(&shrinker_rwsem);
246         list_add_tail(&shrinker->list, &shrinker_list);
247         up_write(&shrinker_rwsem);
248         return 0;
249 }
250 EXPORT_SYMBOL(register_shrinker);
251 
252 /*
253  * Remove one
254  */
255 void unregister_shrinker(struct shrinker *shrinker)
256 {
257         down_write(&shrinker_rwsem);
258         list_del(&shrinker->list);
259         up_write(&shrinker_rwsem);
260         kfree(shrinker->nr_deferred);
261 }
262 EXPORT_SYMBOL(unregister_shrinker);
263 
264 #define SHRINK_BATCH 128
265 
266 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
267                                     struct shrinker *shrinker,
268                                     unsigned long nr_scanned,
269                                     unsigned long nr_eligible)
270 {
271         unsigned long freed = 0;
272         unsigned long long delta;
273         long total_scan;
274         long freeable;
275         long nr;
276         long new_nr;
277         int nid = shrinkctl->nid;
278         long batch_size = shrinker->batch ? shrinker->batch
279                                           : SHRINK_BATCH;
280 
281         freeable = shrinker->count_objects(shrinker, shrinkctl);
282         if (freeable == 0)
283                 return 0;
284 
285         /*
286          * copy the current shrinker scan count into a local variable
287          * and zero it so that other concurrent shrinker invocations
288          * don't also do this scanning work.
289          */
290         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
291 
292         total_scan = nr;
293         delta = (4 * nr_scanned) / shrinker->seeks;
294         delta *= freeable;
295         do_div(delta, nr_eligible + 1);
296         total_scan += delta;
297         if (total_scan < 0) {
298                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
299                        shrinker->scan_objects, total_scan);
300                 total_scan = freeable;
301         }
302 
303         /*
304          * We need to avoid excessive windup on filesystem shrinkers
305          * due to large numbers of GFP_NOFS allocations causing the
306          * shrinkers to return -1 all the time. This results in a large
307          * nr being built up so when a shrink that can do some work
308          * comes along it empties the entire cache due to nr >>>
309          * freeable. This is bad for sustaining a working set in
310          * memory.
311          *
312          * Hence only allow the shrinker to scan the entire cache when
313          * a large delta change is calculated directly.
314          */
315         if (delta < freeable / 4)
316                 total_scan = min(total_scan, freeable / 2);
317 
318         /*
319          * Avoid risking looping forever due to too large nr value:
320          * never try to free more than twice the estimate number of
321          * freeable entries.
322          */
323         if (total_scan > freeable * 2)
324                 total_scan = freeable * 2;
325 
326         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
327                                    nr_scanned, nr_eligible,
328                                    freeable, delta, total_scan);
329 
330         /*
331          * Normally, we should not scan less than batch_size objects in one
332          * pass to avoid too frequent shrinker calls, but if the slab has less
333          * than batch_size objects in total and we are really tight on memory,
334          * we will try to reclaim all available objects, otherwise we can end
335          * up failing allocations although there are plenty of reclaimable
336          * objects spread over several slabs with usage less than the
337          * batch_size.
338          *
339          * We detect the "tight on memory" situations by looking at the total
340          * number of objects we want to scan (total_scan). If it is greater
341          * than the total number of objects on slab (freeable), we must be
342          * scanning at high prio and therefore should try to reclaim as much as
343          * possible.
344          */
345         while (total_scan >= batch_size ||
346                total_scan >= freeable) {
347                 unsigned long ret;
348                 unsigned long nr_to_scan = min(batch_size, total_scan);
349 
350                 shrinkctl->nr_to_scan = nr_to_scan;
351                 ret = shrinker->scan_objects(shrinker, shrinkctl);
352                 if (ret == SHRINK_STOP)
353                         break;
354                 freed += ret;
355 
356                 count_vm_events(SLABS_SCANNED, nr_to_scan);
357                 total_scan -= nr_to_scan;
358 
359                 cond_resched();
360         }
361 
362         /*
363          * move the unused scan count back into the shrinker in a
364          * manner that handles concurrent updates. If we exhausted the
365          * scan, there is no need to do an update.
366          */
367         if (total_scan > 0)
368                 new_nr = atomic_long_add_return(total_scan,
369                                                 &shrinker->nr_deferred[nid]);
370         else
371                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
372 
373         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
374         return freed;
375 }
376 
377 /**
378  * shrink_slab - shrink slab caches
379  * @gfp_mask: allocation context
380  * @nid: node whose slab caches to target
381  * @memcg: memory cgroup whose slab caches to target
382  * @nr_scanned: pressure numerator
383  * @nr_eligible: pressure denominator
384  *
385  * Call the shrink functions to age shrinkable caches.
386  *
387  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
388  * unaware shrinkers will receive a node id of 0 instead.
389  *
390  * @memcg specifies the memory cgroup to target. If it is not NULL,
391  * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
392  * objects from the memory cgroup specified. Otherwise all shrinkers
393  * are called, and memcg aware shrinkers are supposed to scan the
394  * global list then.
395  *
396  * @nr_scanned and @nr_eligible form a ratio that indicate how much of
397  * the available objects should be scanned.  Page reclaim for example
398  * passes the number of pages scanned and the number of pages on the
399  * LRU lists that it considered on @nid, plus a bias in @nr_scanned
400  * when it encountered mapped pages.  The ratio is further biased by
401  * the ->seeks setting of the shrink function, which indicates the
402  * cost to recreate an object relative to that of an LRU page.
403  *
404  * Returns the number of reclaimed slab objects.
405  */
406 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
407                                  struct mem_cgroup *memcg,
408                                  unsigned long nr_scanned,
409                                  unsigned long nr_eligible)
410 {
411         struct shrinker *shrinker;
412         unsigned long freed = 0;
413 
414         if (memcg && !memcg_kmem_is_active(memcg))
415                 return 0;
416 
417         if (nr_scanned == 0)
418                 nr_scanned = SWAP_CLUSTER_MAX;
419 
420         if (!down_read_trylock(&shrinker_rwsem)) {
421                 /*
422                  * If we would return 0, our callers would understand that we
423                  * have nothing else to shrink and give up trying. By returning
424                  * 1 we keep it going and assume we'll be able to shrink next
425                  * time.
426                  */
427                 freed = 1;
428                 goto out;
429         }
430 
431         list_for_each_entry(shrinker, &shrinker_list, list) {
432                 struct shrink_control sc = {
433                         .gfp_mask = gfp_mask,
434                         .nid = nid,
435                         .memcg = memcg,
436                 };
437 
438                 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
439                         continue;
440 
441                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
442                         sc.nid = 0;
443 
444                 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
445         }
446 
447         up_read(&shrinker_rwsem);
448 out:
449         cond_resched();
450         return freed;
451 }
452 
453 void drop_slab_node(int nid)
454 {
455         unsigned long freed;
456 
457         do {
458                 struct mem_cgroup *memcg = NULL;
459 
460                 freed = 0;
461                 do {
462                         freed += shrink_slab(GFP_KERNEL, nid, memcg,
463                                              1000, 1000);
464                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
465         } while (freed > 10);
466 }
467 
468 void drop_slab(void)
469 {
470         int nid;
471 
472         for_each_online_node(nid)
473                 drop_slab_node(nid);
474 }
475 
476 static inline int is_page_cache_freeable(struct page *page)
477 {
478         /*
479          * A freeable page cache page is referenced only by the caller
480          * that isolated the page, the page cache radix tree and
481          * optional buffer heads at page->private.
482          */
483         return page_count(page) - page_has_private(page) == 2;
484 }
485 
486 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
487 {
488         if (current->flags & PF_SWAPWRITE)
489                 return 1;
490         if (!inode_write_congested(inode))
491                 return 1;
492         if (inode_to_bdi(inode) == current->backing_dev_info)
493                 return 1;
494         return 0;
495 }
496 
497 /*
498  * We detected a synchronous write error writing a page out.  Probably
499  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
500  * fsync(), msync() or close().
501  *
502  * The tricky part is that after writepage we cannot touch the mapping: nothing
503  * prevents it from being freed up.  But we have a ref on the page and once
504  * that page is locked, the mapping is pinned.
505  *
506  * We're allowed to run sleeping lock_page() here because we know the caller has
507  * __GFP_FS.
508  */
509 static void handle_write_error(struct address_space *mapping,
510                                 struct page *page, int error)
511 {
512         lock_page(page);
513         if (page_mapping(page) == mapping)
514                 mapping_set_error(mapping, error);
515         unlock_page(page);
516 }
517 
518 /* possible outcome of pageout() */
519 typedef enum {
520         /* failed to write page out, page is locked */
521         PAGE_KEEP,
522         /* move page to the active list, page is locked */
523         PAGE_ACTIVATE,
524         /* page has been sent to the disk successfully, page is unlocked */
525         PAGE_SUCCESS,
526         /* page is clean and locked */
527         PAGE_CLEAN,
528 } pageout_t;
529 
530 /*
531  * pageout is called by shrink_page_list() for each dirty page.
532  * Calls ->writepage().
533  */
534 static pageout_t pageout(struct page *page, struct address_space *mapping,
535                          struct scan_control *sc)
536 {
537         /*
538          * If the page is dirty, only perform writeback if that write
539          * will be non-blocking.  To prevent this allocation from being
540          * stalled by pagecache activity.  But note that there may be
541          * stalls if we need to run get_block().  We could test
542          * PagePrivate for that.
543          *
544          * If this process is currently in __generic_file_write_iter() against
545          * this page's queue, we can perform writeback even if that
546          * will block.
547          *
548          * If the page is swapcache, write it back even if that would
549          * block, for some throttling. This happens by accident, because
550          * swap_backing_dev_info is bust: it doesn't reflect the
551          * congestion state of the swapdevs.  Easy to fix, if needed.
552          */
553         if (!is_page_cache_freeable(page))
554                 return PAGE_KEEP;
555         if (!mapping) {
556                 /*
557                  * Some data journaling orphaned pages can have
558                  * page->mapping == NULL while being dirty with clean buffers.
559                  */
560                 if (page_has_private(page)) {
561                         if (try_to_free_buffers(page)) {
562                                 ClearPageDirty(page);
563                                 pr_info("%s: orphaned page\n", __func__);
564                                 return PAGE_CLEAN;
565                         }
566                 }
567                 return PAGE_KEEP;
568         }
569         if (mapping->a_ops->writepage == NULL)
570                 return PAGE_ACTIVATE;
571         if (!may_write_to_inode(mapping->host, sc))
572                 return PAGE_KEEP;
573 
574         if (clear_page_dirty_for_io(page)) {
575                 int res;
576                 struct writeback_control wbc = {
577                         .sync_mode = WB_SYNC_NONE,
578                         .nr_to_write = SWAP_CLUSTER_MAX,
579                         .range_start = 0,
580                         .range_end = LLONG_MAX,
581                         .for_reclaim = 1,
582                 };
583 
584                 SetPageReclaim(page);
585                 res = mapping->a_ops->writepage(page, &wbc);
586                 if (res < 0)
587                         handle_write_error(mapping, page, res);
588                 if (res == AOP_WRITEPAGE_ACTIVATE) {
589                         ClearPageReclaim(page);
590                         return PAGE_ACTIVATE;
591                 }
592 
593                 if (!PageWriteback(page)) {
594                         /* synchronous write or broken a_ops? */
595                         ClearPageReclaim(page);
596                 }
597                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
598                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
599                 return PAGE_SUCCESS;
600         }
601 
602         return PAGE_CLEAN;
603 }
604 
605 /*
606  * Same as remove_mapping, but if the page is removed from the mapping, it
607  * gets returned with a refcount of 0.
608  */
609 static int __remove_mapping(struct address_space *mapping, struct page *page,
610                             bool reclaimed)
611 {
612         unsigned long flags;
613         struct mem_cgroup *memcg;
614 
615         BUG_ON(!PageLocked(page));
616         BUG_ON(mapping != page_mapping(page));
617 
618         memcg = mem_cgroup_begin_page_stat(page);
619         spin_lock_irqsave(&mapping->tree_lock, flags);
620         /*
621          * The non racy check for a busy page.
622          *
623          * Must be careful with the order of the tests. When someone has
624          * a ref to the page, it may be possible that they dirty it then
625          * drop the reference. So if PageDirty is tested before page_count
626          * here, then the following race may occur:
627          *
628          * get_user_pages(&page);
629          * [user mapping goes away]
630          * write_to(page);
631          *                              !PageDirty(page)    [good]
632          * SetPageDirty(page);
633          * put_page(page);
634          *                              !page_count(page)   [good, discard it]
635          *
636          * [oops, our write_to data is lost]
637          *
638          * Reversing the order of the tests ensures such a situation cannot
639          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
640          * load is not satisfied before that of page->_count.
641          *
642          * Note that if SetPageDirty is always performed via set_page_dirty,
643          * and thus under tree_lock, then this ordering is not required.
644          */
645         if (!page_freeze_refs(page, 2))
646                 goto cannot_free;
647         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
648         if (unlikely(PageDirty(page))) {
649                 page_unfreeze_refs(page, 2);
650                 goto cannot_free;
651         }
652 
653         if (PageSwapCache(page)) {
654                 swp_entry_t swap = { .val = page_private(page) };
655                 mem_cgroup_swapout(page, swap);
656                 __delete_from_swap_cache(page);
657                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
658                 mem_cgroup_end_page_stat(memcg);
659                 swapcache_free(swap);
660         } else {
661                 void (*freepage)(struct page *);
662                 void *shadow = NULL;
663 
664                 freepage = mapping->a_ops->freepage;
665                 /*
666                  * Remember a shadow entry for reclaimed file cache in
667                  * order to detect refaults, thus thrashing, later on.
668                  *
669                  * But don't store shadows in an address space that is
670                  * already exiting.  This is not just an optizimation,
671                  * inode reclaim needs to empty out the radix tree or
672                  * the nodes are lost.  Don't plant shadows behind its
673                  * back.
674                  */
675                 if (reclaimed && page_is_file_cache(page) &&
676                     !mapping_exiting(mapping))
677                         shadow = workingset_eviction(mapping, page);
678                 __delete_from_page_cache(page, shadow, memcg);
679                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
680                 mem_cgroup_end_page_stat(memcg);
681 
682                 if (freepage != NULL)
683                         freepage(page);
684         }
685 
686         return 1;
687 
688 cannot_free:
689         spin_unlock_irqrestore(&mapping->tree_lock, flags);
690         mem_cgroup_end_page_stat(memcg);
691         return 0;
692 }
693 
694 /*
695  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
696  * someone else has a ref on the page, abort and return 0.  If it was
697  * successfully detached, return 1.  Assumes the caller has a single ref on
698  * this page.
699  */
700 int remove_mapping(struct address_space *mapping, struct page *page)
701 {
702         if (__remove_mapping(mapping, page, false)) {
703                 /*
704                  * Unfreezing the refcount with 1 rather than 2 effectively
705                  * drops the pagecache ref for us without requiring another
706                  * atomic operation.
707                  */
708                 page_unfreeze_refs(page, 1);
709                 return 1;
710         }
711         return 0;
712 }
713 
714 /**
715  * putback_lru_page - put previously isolated page onto appropriate LRU list
716  * @page: page to be put back to appropriate lru list
717  *
718  * Add previously isolated @page to appropriate LRU list.
719  * Page may still be unevictable for other reasons.
720  *
721  * lru_lock must not be held, interrupts must be enabled.
722  */
723 void putback_lru_page(struct page *page)
724 {
725         bool is_unevictable;
726         int was_unevictable = PageUnevictable(page);
727 
728         VM_BUG_ON_PAGE(PageLRU(page), page);
729 
730 redo:
731         ClearPageUnevictable(page);
732 
733         if (page_evictable(page)) {
734                 /*
735                  * For evictable pages, we can use the cache.
736                  * In event of a race, worst case is we end up with an
737                  * unevictable page on [in]active list.
738                  * We know how to handle that.
739                  */
740                 is_unevictable = false;
741                 lru_cache_add(page);
742         } else {
743                 /*
744                  * Put unevictable pages directly on zone's unevictable
745                  * list.
746                  */
747                 is_unevictable = true;
748                 add_page_to_unevictable_list(page);
749                 /*
750                  * When racing with an mlock or AS_UNEVICTABLE clearing
751                  * (page is unlocked) make sure that if the other thread
752                  * does not observe our setting of PG_lru and fails
753                  * isolation/check_move_unevictable_pages,
754                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
755                  * the page back to the evictable list.
756                  *
757                  * The other side is TestClearPageMlocked() or shmem_lock().
758                  */
759                 smp_mb();
760         }
761 
762         /*
763          * page's status can change while we move it among lru. If an evictable
764          * page is on unevictable list, it never be freed. To avoid that,
765          * check after we added it to the list, again.
766          */
767         if (is_unevictable && page_evictable(page)) {
768                 if (!isolate_lru_page(page)) {
769                         put_page(page);
770                         goto redo;
771                 }
772                 /* This means someone else dropped this page from LRU
773                  * So, it will be freed or putback to LRU again. There is
774                  * nothing to do here.
775                  */
776         }
777 
778         if (was_unevictable && !is_unevictable)
779                 count_vm_event(UNEVICTABLE_PGRESCUED);
780         else if (!was_unevictable && is_unevictable)
781                 count_vm_event(UNEVICTABLE_PGCULLED);
782 
783         put_page(page);         /* drop ref from isolate */
784 }
785 
786 enum page_references {
787         PAGEREF_RECLAIM,
788         PAGEREF_RECLAIM_CLEAN,
789         PAGEREF_KEEP,
790         PAGEREF_ACTIVATE,
791 };
792 
793 static enum page_references page_check_references(struct page *page,
794                                                   struct scan_control *sc)
795 {
796         int referenced_ptes, referenced_page;
797         unsigned long vm_flags;
798 
799         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
800                                           &vm_flags);
801         referenced_page = TestClearPageReferenced(page);
802 
803         /*
804          * Mlock lost the isolation race with us.  Let try_to_unmap()
805          * move the page to the unevictable list.
806          */
807         if (vm_flags & VM_LOCKED)
808                 return PAGEREF_RECLAIM;
809 
810         if (referenced_ptes) {
811                 if (PageSwapBacked(page))
812                         return PAGEREF_ACTIVATE;
813                 /*
814                  * All mapped pages start out with page table
815                  * references from the instantiating fault, so we need
816                  * to look twice if a mapped file page is used more
817                  * than once.
818                  *
819                  * Mark it and spare it for another trip around the
820                  * inactive list.  Another page table reference will
821                  * lead to its activation.
822                  *
823                  * Note: the mark is set for activated pages as well
824                  * so that recently deactivated but used pages are
825                  * quickly recovered.
826                  */
827                 SetPageReferenced(page);
828 
829                 if (referenced_page || referenced_ptes > 1)
830                         return PAGEREF_ACTIVATE;
831 
832                 /*
833                  * Activate file-backed executable pages after first usage.
834                  */
835                 if (vm_flags & VM_EXEC)
836                         return PAGEREF_ACTIVATE;
837 
838                 return PAGEREF_KEEP;
839         }
840 
841         /* Reclaim if clean, defer dirty pages to writeback */
842         if (referenced_page && !PageSwapBacked(page))
843                 return PAGEREF_RECLAIM_CLEAN;
844 
845         return PAGEREF_RECLAIM;
846 }
847 
848 /* Check if a page is dirty or under writeback */
849 static void page_check_dirty_writeback(struct page *page,
850                                        bool *dirty, bool *writeback)
851 {
852         struct address_space *mapping;
853 
854         /*
855          * Anonymous pages are not handled by flushers and must be written
856          * from reclaim context. Do not stall reclaim based on them
857          */
858         if (!page_is_file_cache(page)) {
859                 *dirty = false;
860                 *writeback = false;
861                 return;
862         }
863 
864         /* By default assume that the page flags are accurate */
865         *dirty = PageDirty(page);
866         *writeback = PageWriteback(page);
867 
868         /* Verify dirty/writeback state if the filesystem supports it */
869         if (!page_has_private(page))
870                 return;
871 
872         mapping = page_mapping(page);
873         if (mapping && mapping->a_ops->is_dirty_writeback)
874                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
875 }
876 
877 /*
878  * shrink_page_list() returns the number of reclaimed pages
879  */
880 static unsigned long shrink_page_list(struct list_head *page_list,
881                                       struct zone *zone,
882                                       struct scan_control *sc,
883                                       enum ttu_flags ttu_flags,
884                                       unsigned long *ret_nr_dirty,
885                                       unsigned long *ret_nr_unqueued_dirty,
886                                       unsigned long *ret_nr_congested,
887                                       unsigned long *ret_nr_writeback,
888                                       unsigned long *ret_nr_immediate,
889                                       bool force_reclaim)
890 {
891         LIST_HEAD(ret_pages);
892         LIST_HEAD(free_pages);
893         int pgactivate = 0;
894         unsigned long nr_unqueued_dirty = 0;
895         unsigned long nr_dirty = 0;
896         unsigned long nr_congested = 0;
897         unsigned long nr_reclaimed = 0;
898         unsigned long nr_writeback = 0;
899         unsigned long nr_immediate = 0;
900 
901         cond_resched();
902 
903         while (!list_empty(page_list)) {
904                 struct address_space *mapping;
905                 struct page *page;
906                 int may_enter_fs;
907                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
908                 bool dirty, writeback;
909 
910                 cond_resched();
911 
912                 page = lru_to_page(page_list);
913                 list_del(&page->lru);
914 
915                 if (!trylock_page(page))
916                         goto keep;
917 
918                 VM_BUG_ON_PAGE(PageActive(page), page);
919                 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
920 
921                 sc->nr_scanned++;
922 
923                 if (unlikely(!page_evictable(page)))
924                         goto cull_mlocked;
925 
926                 if (!sc->may_unmap && page_mapped(page))
927                         goto keep_locked;
928 
929                 /* Double the slab pressure for mapped and swapcache pages */
930                 if (page_mapped(page) || PageSwapCache(page))
931                         sc->nr_scanned++;
932 
933                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
934                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
935 
936                 /*
937                  * The number of dirty pages determines if a zone is marked
938                  * reclaim_congested which affects wait_iff_congested. kswapd
939                  * will stall and start writing pages if the tail of the LRU
940                  * is all dirty unqueued pages.
941                  */
942                 page_check_dirty_writeback(page, &dirty, &writeback);
943                 if (dirty || writeback)
944                         nr_dirty++;
945 
946                 if (dirty && !writeback)
947                         nr_unqueued_dirty++;
948 
949                 /*
950                  * Treat this page as congested if the underlying BDI is or if
951                  * pages are cycling through the LRU so quickly that the
952                  * pages marked for immediate reclaim are making it to the
953                  * end of the LRU a second time.
954                  */
955                 mapping = page_mapping(page);
956                 if (((dirty || writeback) && mapping &&
957                      inode_write_congested(mapping->host)) ||
958                     (writeback && PageReclaim(page)))
959                         nr_congested++;
960 
961                 /*
962                  * If a page at the tail of the LRU is under writeback, there
963                  * are three cases to consider.
964                  *
965                  * 1) If reclaim is encountering an excessive number of pages
966                  *    under writeback and this page is both under writeback and
967                  *    PageReclaim then it indicates that pages are being queued
968                  *    for IO but are being recycled through the LRU before the
969                  *    IO can complete. Waiting on the page itself risks an
970                  *    indefinite stall if it is impossible to writeback the
971                  *    page due to IO error or disconnected storage so instead
972                  *    note that the LRU is being scanned too quickly and the
973                  *    caller can stall after page list has been processed.
974                  *
975                  * 2) Global or new memcg reclaim encounters a page that is
976                  *    not marked for immediate reclaim, or the caller does not
977                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
978                  *    not to fs). In this case mark the page for immediate
979                  *    reclaim and continue scanning.
980                  *
981                  *    Require may_enter_fs because we would wait on fs, which
982                  *    may not have submitted IO yet. And the loop driver might
983                  *    enter reclaim, and deadlock if it waits on a page for
984                  *    which it is needed to do the write (loop masks off
985                  *    __GFP_IO|__GFP_FS for this reason); but more thought
986                  *    would probably show more reasons.
987                  *
988                  * 3) Legacy memcg encounters a page that is already marked
989                  *    PageReclaim. memcg does not have any dirty pages
990                  *    throttling so we could easily OOM just because too many
991                  *    pages are in writeback and there is nothing else to
992                  *    reclaim. Wait for the writeback to complete.
993                  */
994                 if (PageWriteback(page)) {
995                         /* Case 1 above */
996                         if (current_is_kswapd() &&
997                             PageReclaim(page) &&
998                             test_bit(ZONE_WRITEBACK, &zone->flags)) {
999                                 nr_immediate++;
1000                                 goto keep_locked;
1001 
1002                         /* Case 2 above */
1003                         } else if (sane_reclaim(sc) ||
1004                             !PageReclaim(page) || !may_enter_fs) {
1005                                 /*
1006                                  * This is slightly racy - end_page_writeback()
1007                                  * might have just cleared PageReclaim, then
1008                                  * setting PageReclaim here end up interpreted
1009                                  * as PageReadahead - but that does not matter
1010                                  * enough to care.  What we do want is for this
1011                                  * page to have PageReclaim set next time memcg
1012                                  * reclaim reaches the tests above, so it will
1013                                  * then wait_on_page_writeback() to avoid OOM;
1014                                  * and it's also appropriate in global reclaim.
1015                                  */
1016                                 SetPageReclaim(page);
1017                                 nr_writeback++;
1018                                 goto keep_locked;
1019 
1020                         /* Case 3 above */
1021                         } else {
1022                                 unlock_page(page);
1023                                 wait_on_page_writeback(page);
1024                                 /* then go back and try same page again */
1025                                 list_add_tail(&page->lru, page_list);
1026                                 continue;
1027                         }
1028                 }
1029 
1030                 if (!force_reclaim)
1031                         references = page_check_references(page, sc);
1032 
1033                 switch (references) {
1034                 case PAGEREF_ACTIVATE:
1035                         goto activate_locked;
1036                 case PAGEREF_KEEP:
1037                         goto keep_locked;
1038                 case PAGEREF_RECLAIM:
1039                 case PAGEREF_RECLAIM_CLEAN:
1040                         ; /* try to reclaim the page below */
1041                 }
1042 
1043                 /*
1044                  * Anonymous process memory has backing store?
1045                  * Try to allocate it some swap space here.
1046                  */
1047                 if (PageAnon(page) && !PageSwapCache(page)) {
1048                         if (!(sc->gfp_mask & __GFP_IO))
1049                                 goto keep_locked;
1050                         if (!add_to_swap(page, page_list))
1051                                 goto activate_locked;
1052                         may_enter_fs = 1;
1053 
1054                         /* Adding to swap updated mapping */
1055                         mapping = page_mapping(page);
1056                 }
1057 
1058                 /*
1059                  * The page is mapped into the page tables of one or more
1060                  * processes. Try to unmap it here.
1061                  */
1062                 if (page_mapped(page) && mapping) {
1063                         switch (try_to_unmap(page,
1064                                         ttu_flags|TTU_BATCH_FLUSH)) {
1065                         case SWAP_FAIL:
1066                                 goto activate_locked;
1067                         case SWAP_AGAIN:
1068                                 goto keep_locked;
1069                         case SWAP_MLOCK:
1070                                 goto cull_mlocked;
1071                         case SWAP_SUCCESS:
1072                                 ; /* try to free the page below */
1073                         }
1074                 }
1075 
1076                 if (PageDirty(page)) {
1077                         /*
1078                          * Only kswapd can writeback filesystem pages to
1079                          * avoid risk of stack overflow but only writeback
1080                          * if many dirty pages have been encountered.
1081                          */
1082                         if (page_is_file_cache(page) &&
1083                                         (!current_is_kswapd() ||
1084                                          !test_bit(ZONE_DIRTY, &zone->flags))) {
1085                                 /*
1086                                  * Immediately reclaim when written back.
1087                                  * Similar in principal to deactivate_page()
1088                                  * except we already have the page isolated
1089                                  * and know it's dirty
1090                                  */
1091                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1092                                 SetPageReclaim(page);
1093 
1094                                 goto keep_locked;
1095                         }
1096 
1097                         if (references == PAGEREF_RECLAIM_CLEAN)
1098                                 goto keep_locked;
1099                         if (!may_enter_fs)
1100                                 goto keep_locked;
1101                         if (!sc->may_writepage)
1102                                 goto keep_locked;
1103 
1104                         /*
1105                          * Page is dirty. Flush the TLB if a writable entry
1106                          * potentially exists to avoid CPU writes after IO
1107                          * starts and then write it out here.
1108                          */
1109                         try_to_unmap_flush_dirty();
1110                         switch (pageout(page, mapping, sc)) {
1111                         case PAGE_KEEP:
1112                                 goto keep_locked;
1113                         case PAGE_ACTIVATE:
1114                                 goto activate_locked;
1115                         case PAGE_SUCCESS:
1116                                 if (PageWriteback(page))
1117                                         goto keep;
1118                                 if (PageDirty(page))
1119                                         goto keep;
1120 
1121                                 /*
1122                                  * A synchronous write - probably a ramdisk.  Go
1123                                  * ahead and try to reclaim the page.
1124                                  */
1125                                 if (!trylock_page(page))
1126                                         goto keep;
1127                                 if (PageDirty(page) || PageWriteback(page))
1128                                         goto keep_locked;
1129                                 mapping = page_mapping(page);
1130                         case PAGE_CLEAN:
1131                                 ; /* try to free the page below */
1132                         }
1133                 }
1134 
1135                 /*
1136                  * If the page has buffers, try to free the buffer mappings
1137                  * associated with this page. If we succeed we try to free
1138                  * the page as well.
1139                  *
1140                  * We do this even if the page is PageDirty().
1141                  * try_to_release_page() does not perform I/O, but it is
1142                  * possible for a page to have PageDirty set, but it is actually
1143                  * clean (all its buffers are clean).  This happens if the
1144                  * buffers were written out directly, with submit_bh(). ext3
1145                  * will do this, as well as the blockdev mapping.
1146                  * try_to_release_page() will discover that cleanness and will
1147                  * drop the buffers and mark the page clean - it can be freed.
1148                  *
1149                  * Rarely, pages can have buffers and no ->mapping.  These are
1150                  * the pages which were not successfully invalidated in
1151                  * truncate_complete_page().  We try to drop those buffers here
1152                  * and if that worked, and the page is no longer mapped into
1153                  * process address space (page_count == 1) it can be freed.
1154                  * Otherwise, leave the page on the LRU so it is swappable.
1155                  */
1156                 if (page_has_private(page)) {
1157                         if (!try_to_release_page(page, sc->gfp_mask))
1158                                 goto activate_locked;
1159                         if (!mapping && page_count(page) == 1) {
1160                                 unlock_page(page);
1161                                 if (put_page_testzero(page))
1162                                         goto free_it;
1163                                 else {
1164                                         /*
1165                                          * rare race with speculative reference.
1166                                          * the speculative reference will free
1167                                          * this page shortly, so we may
1168                                          * increment nr_reclaimed here (and
1169                                          * leave it off the LRU).
1170                                          */
1171                                         nr_reclaimed++;
1172                                         continue;
1173                                 }
1174                         }
1175                 }
1176 
1177                 if (!mapping || !__remove_mapping(mapping, page, true))
1178                         goto keep_locked;
1179 
1180                 /*
1181                  * At this point, we have no other references and there is
1182                  * no way to pick any more up (removed from LRU, removed
1183                  * from pagecache). Can use non-atomic bitops now (and
1184                  * we obviously don't have to worry about waking up a process
1185                  * waiting on the page lock, because there are no references.
1186                  */
1187                 __clear_page_locked(page);
1188 free_it:
1189                 nr_reclaimed++;
1190 
1191                 /*
1192                  * Is there need to periodically free_page_list? It would
1193                  * appear not as the counts should be low
1194                  */
1195                 list_add(&page->lru, &free_pages);
1196                 continue;
1197 
1198 cull_mlocked:
1199                 if (PageSwapCache(page))
1200                         try_to_free_swap(page);
1201                 unlock_page(page);
1202                 list_add(&page->lru, &ret_pages);
1203                 continue;
1204 
1205 activate_locked:
1206                 /* Not a candidate for swapping, so reclaim swap space. */
1207                 if (PageSwapCache(page) && vm_swap_full())
1208                         try_to_free_swap(page);
1209                 VM_BUG_ON_PAGE(PageActive(page), page);
1210                 SetPageActive(page);
1211                 pgactivate++;
1212 keep_locked:
1213                 unlock_page(page);
1214 keep:
1215                 list_add(&page->lru, &ret_pages);
1216                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1217         }
1218 
1219         mem_cgroup_uncharge_list(&free_pages);
1220         try_to_unmap_flush();
1221         free_hot_cold_page_list(&free_pages, true);
1222 
1223         list_splice(&ret_pages, page_list);
1224         count_vm_events(PGACTIVATE, pgactivate);
1225 
1226         *ret_nr_dirty += nr_dirty;
1227         *ret_nr_congested += nr_congested;
1228         *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1229         *ret_nr_writeback += nr_writeback;
1230         *ret_nr_immediate += nr_immediate;
1231         return nr_reclaimed;
1232 }
1233 
1234 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1235                                             struct list_head *page_list)
1236 {
1237         struct scan_control sc = {
1238                 .gfp_mask = GFP_KERNEL,
1239                 .priority = DEF_PRIORITY,
1240                 .may_unmap = 1,
1241         };
1242         unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1243         struct page *page, *next;
1244         LIST_HEAD(clean_pages);
1245 
1246         list_for_each_entry_safe(page, next, page_list, lru) {
1247                 if (page_is_file_cache(page) && !PageDirty(page) &&
1248                     !isolated_balloon_page(page)) {
1249                         ClearPageActive(page);
1250                         list_move(&page->lru, &clean_pages);
1251                 }
1252         }
1253 
1254         ret = shrink_page_list(&clean_pages, zone, &sc,
1255                         TTU_UNMAP|TTU_IGNORE_ACCESS,
1256                         &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1257         list_splice(&clean_pages, page_list);
1258         mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1259         return ret;
1260 }
1261 
1262 /*
1263  * Attempt to remove the specified page from its LRU.  Only take this page
1264  * if it is of the appropriate PageActive status.  Pages which are being
1265  * freed elsewhere are also ignored.
1266  *
1267  * page:        page to consider
1268  * mode:        one of the LRU isolation modes defined above
1269  *
1270  * returns 0 on success, -ve errno on failure.
1271  */
1272 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1273 {
1274         int ret = -EINVAL;
1275 
1276         /* Only take pages on the LRU. */
1277         if (!PageLRU(page))
1278                 return ret;
1279 
1280         /* Compaction should not handle unevictable pages but CMA can do so */
1281         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1282                 return ret;
1283 
1284         ret = -EBUSY;
1285 
1286         /*
1287          * To minimise LRU disruption, the caller can indicate that it only
1288          * wants to isolate pages it will be able to operate on without
1289          * blocking - clean pages for the most part.
1290          *
1291          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1292          * is used by reclaim when it is cannot write to backing storage
1293          *
1294          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1295          * that it is possible to migrate without blocking
1296          */
1297         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1298                 /* All the caller can do on PageWriteback is block */
1299                 if (PageWriteback(page))
1300                         return ret;
1301 
1302                 if (PageDirty(page)) {
1303                         struct address_space *mapping;
1304 
1305                         /* ISOLATE_CLEAN means only clean pages */
1306                         if (mode & ISOLATE_CLEAN)
1307                                 return ret;
1308 
1309                         /*
1310                          * Only pages without mappings or that have a
1311                          * ->migratepage callback are possible to migrate
1312                          * without blocking
1313                          */
1314                         mapping = page_mapping(page);
1315                         if (mapping && !mapping->a_ops->migratepage)
1316                                 return ret;
1317                 }
1318         }
1319 
1320         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1321                 return ret;
1322 
1323         if (likely(get_page_unless_zero(page))) {
1324                 /*
1325                  * Be careful not to clear PageLRU until after we're
1326                  * sure the page is not being freed elsewhere -- the
1327                  * page release code relies on it.
1328                  */
1329                 ClearPageLRU(page);
1330                 ret = 0;
1331         }
1332 
1333         return ret;
1334 }
1335 
1336 /*
1337  * zone->lru_lock is heavily contended.  Some of the functions that
1338  * shrink the lists perform better by taking out a batch of pages
1339  * and working on them outside the LRU lock.
1340  *
1341  * For pagecache intensive workloads, this function is the hottest
1342  * spot in the kernel (apart from copy_*_user functions).
1343  *
1344  * Appropriate locks must be held before calling this function.
1345  *
1346  * @nr_to_scan: The number of pages to look through on the list.
1347  * @lruvec:     The LRU vector to pull pages from.
1348  * @dst:        The temp list to put pages on to.
1349  * @nr_scanned: The number of pages that were scanned.
1350  * @sc:         The scan_control struct for this reclaim session
1351  * @mode:       One of the LRU isolation modes
1352  * @lru:        LRU list id for isolating
1353  *
1354  * returns how many pages were moved onto *@dst.
1355  */
1356 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1357                 struct lruvec *lruvec, struct list_head *dst,
1358                 unsigned long *nr_scanned, struct scan_control *sc,
1359                 isolate_mode_t mode, enum lru_list lru)
1360 {
1361         struct list_head *src = &lruvec->lists[lru];
1362         unsigned long nr_taken = 0;
1363         unsigned long scan;
1364 
1365         for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1366                                         !list_empty(src); scan++) {
1367                 struct page *page;
1368                 int nr_pages;
1369 
1370                 page = lru_to_page(src);
1371                 prefetchw_prev_lru_page(page, src, flags);
1372 
1373                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1374 
1375                 switch (__isolate_lru_page(page, mode)) {
1376                 case 0:
1377                         nr_pages = hpage_nr_pages(page);
1378                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1379                         list_move(&page->lru, dst);
1380                         nr_taken += nr_pages;
1381                         break;
1382 
1383                 case -EBUSY:
1384                         /* else it is being freed elsewhere */
1385                         list_move(&page->lru, src);
1386                         continue;
1387 
1388                 default:
1389                         BUG();
1390                 }
1391         }
1392 
1393         *nr_scanned = scan;
1394         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1395                                     nr_taken, mode, is_file_lru(lru));
1396         return nr_taken;
1397 }
1398 
1399 /**
1400  * isolate_lru_page - tries to isolate a page from its LRU list
1401  * @page: page to isolate from its LRU list
1402  *
1403  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1404  * vmstat statistic corresponding to whatever LRU list the page was on.
1405  *
1406  * Returns 0 if the page was removed from an LRU list.
1407  * Returns -EBUSY if the page was not on an LRU list.
1408  *
1409  * The returned page will have PageLRU() cleared.  If it was found on
1410  * the active list, it will have PageActive set.  If it was found on
1411  * the unevictable list, it will have the PageUnevictable bit set. That flag
1412  * may need to be cleared by the caller before letting the page go.
1413  *
1414  * The vmstat statistic corresponding to the list on which the page was
1415  * found will be decremented.
1416  *
1417  * Restrictions:
1418  * (1) Must be called with an elevated refcount on the page. This is a
1419  *     fundamentnal difference from isolate_lru_pages (which is called
1420  *     without a stable reference).
1421  * (2) the lru_lock must not be held.
1422  * (3) interrupts must be enabled.
1423  */
1424 int isolate_lru_page(struct page *page)
1425 {
1426         int ret = -EBUSY;
1427 
1428         VM_BUG_ON_PAGE(!page_count(page), page);
1429 
1430         if (PageLRU(page)) {
1431                 struct zone *zone = page_zone(page);
1432                 struct lruvec *lruvec;
1433 
1434                 spin_lock_irq(&zone->lru_lock);
1435                 lruvec = mem_cgroup_page_lruvec(page, zone);
1436                 if (PageLRU(page)) {
1437                         int lru = page_lru(page);
1438                         get_page(page);
1439                         ClearPageLRU(page);
1440                         del_page_from_lru_list(page, lruvec, lru);
1441                         ret = 0;
1442                 }
1443                 spin_unlock_irq(&zone->lru_lock);
1444         }
1445         return ret;
1446 }
1447 
1448 /*
1449  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1450  * then get resheduled. When there are massive number of tasks doing page
1451  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1452  * the LRU list will go small and be scanned faster than necessary, leading to
1453  * unnecessary swapping, thrashing and OOM.
1454  */
1455 static int too_many_isolated(struct zone *zone, int file,
1456                 struct scan_control *sc)
1457 {
1458         unsigned long inactive, isolated;
1459 
1460         if (current_is_kswapd())
1461                 return 0;
1462 
1463         if (!sane_reclaim(sc))
1464                 return 0;
1465 
1466         if (file) {
1467                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1468                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1469         } else {
1470                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1471                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1472         }
1473 
1474         /*
1475          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1476          * won't get blocked by normal direct-reclaimers, forming a circular
1477          * deadlock.
1478          */
1479         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1480                 inactive >>= 3;
1481 
1482         return isolated > inactive;
1483 }
1484 
1485 static noinline_for_stack void
1486 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1487 {
1488         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1489         struct zone *zone = lruvec_zone(lruvec);
1490         LIST_HEAD(pages_to_free);
1491 
1492         /*
1493          * Put back any unfreeable pages.
1494          */
1495         while (!list_empty(page_list)) {
1496                 struct page *page = lru_to_page(page_list);
1497                 int lru;
1498 
1499                 VM_BUG_ON_PAGE(PageLRU(page), page);
1500                 list_del(&page->lru);
1501                 if (unlikely(!page_evictable(page))) {
1502                         spin_unlock_irq(&zone->lru_lock);
1503                         putback_lru_page(page);
1504                         spin_lock_irq(&zone->lru_lock);
1505                         continue;
1506                 }
1507 
1508                 lruvec = mem_cgroup_page_lruvec(page, zone);
1509 
1510                 SetPageLRU(page);
1511                 lru = page_lru(page);
1512                 add_page_to_lru_list(page, lruvec, lru);
1513 
1514                 if (is_active_lru(lru)) {
1515                         int file = is_file_lru(lru);
1516                         int numpages = hpage_nr_pages(page);
1517                         reclaim_stat->recent_rotated[file] += numpages;
1518                 }
1519                 if (put_page_testzero(page)) {
1520                         __ClearPageLRU(page);
1521                         __ClearPageActive(page);
1522                         del_page_from_lru_list(page, lruvec, lru);
1523 
1524                         if (unlikely(PageCompound(page))) {
1525                                 spin_unlock_irq(&zone->lru_lock);
1526                                 mem_cgroup_uncharge(page);
1527                                 (*get_compound_page_dtor(page))(page);
1528                                 spin_lock_irq(&zone->lru_lock);
1529                         } else
1530                                 list_add(&page->lru, &pages_to_free);
1531                 }
1532         }
1533 
1534         /*
1535          * To save our caller's stack, now use input list for pages to free.
1536          */
1537         list_splice(&pages_to_free, page_list);
1538 }
1539 
1540 /*
1541  * If a kernel thread (such as nfsd for loop-back mounts) services
1542  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1543  * In that case we should only throttle if the backing device it is
1544  * writing to is congested.  In other cases it is safe to throttle.
1545  */
1546 static int current_may_throttle(void)
1547 {
1548         return !(current->flags & PF_LESS_THROTTLE) ||
1549                 current->backing_dev_info == NULL ||
1550                 bdi_write_congested(current->backing_dev_info);
1551 }
1552 
1553 /*
1554  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1555  * of reclaimed pages
1556  */
1557 static noinline_for_stack unsigned long
1558 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1559                      struct scan_control *sc, enum lru_list lru)
1560 {
1561         LIST_HEAD(page_list);
1562         unsigned long nr_scanned;
1563         unsigned long nr_reclaimed = 0;
1564         unsigned long nr_taken;
1565         unsigned long nr_dirty = 0;
1566         unsigned long nr_congested = 0;
1567         unsigned long nr_unqueued_dirty = 0;
1568         unsigned long nr_writeback = 0;
1569         unsigned long nr_immediate = 0;
1570         isolate_mode_t isolate_mode = 0;
1571         int file = is_file_lru(lru);
1572         struct zone *zone = lruvec_zone(lruvec);
1573         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1574 
1575         while (unlikely(too_many_isolated(zone, file, sc))) {
1576                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1577 
1578                 /* We are about to die and free our memory. Return now. */
1579                 if (fatal_signal_pending(current))
1580                         return SWAP_CLUSTER_MAX;
1581         }
1582 
1583         lru_add_drain();
1584 
1585         if (!sc->may_unmap)
1586                 isolate_mode |= ISOLATE_UNMAPPED;
1587         if (!sc->may_writepage)
1588                 isolate_mode |= ISOLATE_CLEAN;
1589 
1590         spin_lock_irq(&zone->lru_lock);
1591 
1592         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1593                                      &nr_scanned, sc, isolate_mode, lru);
1594 
1595         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1596         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1597 
1598         if (global_reclaim(sc)) {
1599                 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1600                 if (current_is_kswapd())
1601                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1602                 else
1603                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1604         }
1605         spin_unlock_irq(&zone->lru_lock);
1606 
1607         if (nr_taken == 0)
1608                 return 0;
1609 
1610         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1611                                 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1612                                 &nr_writeback, &nr_immediate,
1613                                 false);
1614 
1615         spin_lock_irq(&zone->lru_lock);
1616 
1617         reclaim_stat->recent_scanned[file] += nr_taken;
1618 
1619         if (global_reclaim(sc)) {
1620                 if (current_is_kswapd())
1621                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1622                                                nr_reclaimed);
1623                 else
1624                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1625                                                nr_reclaimed);
1626         }
1627 
1628         putback_inactive_pages(lruvec, &page_list);
1629 
1630         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1631 
1632         spin_unlock_irq(&zone->lru_lock);
1633 
1634         mem_cgroup_uncharge_list(&page_list);
1635         free_hot_cold_page_list(&page_list, true);
1636 
1637         /*
1638          * If reclaim is isolating dirty pages under writeback, it implies
1639          * that the long-lived page allocation rate is exceeding the page
1640          * laundering rate. Either the global limits are not being effective
1641          * at throttling processes due to the page distribution throughout
1642          * zones or there is heavy usage of a slow backing device. The
1643          * only option is to throttle from reclaim context which is not ideal
1644          * as there is no guarantee the dirtying process is throttled in the
1645          * same way balance_dirty_pages() manages.
1646          *
1647          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1648          * of pages under pages flagged for immediate reclaim and stall if any
1649          * are encountered in the nr_immediate check below.
1650          */
1651         if (nr_writeback && nr_writeback == nr_taken)
1652                 set_bit(ZONE_WRITEBACK, &zone->flags);
1653 
1654         /*
1655          * Legacy memcg will stall in page writeback so avoid forcibly
1656          * stalling here.
1657          */
1658         if (sane_reclaim(sc)) {
1659                 /*
1660                  * Tag a zone as congested if all the dirty pages scanned were
1661                  * backed by a congested BDI and wait_iff_congested will stall.
1662                  */
1663                 if (nr_dirty && nr_dirty == nr_congested)
1664                         set_bit(ZONE_CONGESTED, &zone->flags);
1665 
1666                 /*
1667                  * If dirty pages are scanned that are not queued for IO, it
1668                  * implies that flushers are not keeping up. In this case, flag
1669                  * the zone ZONE_DIRTY and kswapd will start writing pages from
1670                  * reclaim context.
1671                  */
1672                 if (nr_unqueued_dirty == nr_taken)
1673                         set_bit(ZONE_DIRTY, &zone->flags);
1674 
1675                 /*
1676                  * If kswapd scans pages marked marked for immediate
1677                  * reclaim and under writeback (nr_immediate), it implies
1678                  * that pages are cycling through the LRU faster than
1679                  * they are written so also forcibly stall.
1680                  */
1681                 if (nr_immediate && current_may_throttle())
1682                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1683         }
1684 
1685         /*
1686          * Stall direct reclaim for IO completions if underlying BDIs or zone
1687          * is congested. Allow kswapd to continue until it starts encountering
1688          * unqueued dirty pages or cycling through the LRU too quickly.
1689          */
1690         if (!sc->hibernation_mode && !current_is_kswapd() &&
1691             current_may_throttle())
1692                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1693 
1694         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1695                 zone_idx(zone),
1696                 nr_scanned, nr_reclaimed,
1697                 sc->priority,
1698                 trace_shrink_flags(file));
1699         return nr_reclaimed;
1700 }
1701 
1702 /*
1703  * This moves pages from the active list to the inactive list.
1704  *
1705  * We move them the other way if the page is referenced by one or more
1706  * processes, from rmap.
1707  *
1708  * If the pages are mostly unmapped, the processing is fast and it is
1709  * appropriate to hold zone->lru_lock across the whole operation.  But if
1710  * the pages are mapped, the processing is slow (page_referenced()) so we
1711  * should drop zone->lru_lock around each page.  It's impossible to balance
1712  * this, so instead we remove the pages from the LRU while processing them.
1713  * It is safe to rely on PG_active against the non-LRU pages in here because
1714  * nobody will play with that bit on a non-LRU page.
1715  *
1716  * The downside is that we have to touch page->_count against each page.
1717  * But we had to alter page->flags anyway.
1718  */
1719 
1720 static void move_active_pages_to_lru(struct lruvec *lruvec,
1721                                      struct list_head *list,
1722                                      struct list_head *pages_to_free,
1723                                      enum lru_list lru)
1724 {
1725         struct zone *zone = lruvec_zone(lruvec);
1726         unsigned long pgmoved = 0;
1727         struct page *page;
1728         int nr_pages;
1729 
1730         while (!list_empty(list)) {
1731                 page = lru_to_page(list);
1732                 lruvec = mem_cgroup_page_lruvec(page, zone);
1733 
1734                 VM_BUG_ON_PAGE(PageLRU(page), page);
1735                 SetPageLRU(page);
1736 
1737                 nr_pages = hpage_nr_pages(page);
1738                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1739                 list_move(&page->lru, &lruvec->lists[lru]);
1740                 pgmoved += nr_pages;
1741 
1742                 if (put_page_testzero(page)) {
1743                         __ClearPageLRU(page);
1744                         __ClearPageActive(page);
1745                         del_page_from_lru_list(page, lruvec, lru);
1746 
1747                         if (unlikely(PageCompound(page))) {
1748                                 spin_unlock_irq(&zone->lru_lock);
1749                                 mem_cgroup_uncharge(page);
1750                                 (*get_compound_page_dtor(page))(page);
1751                                 spin_lock_irq(&zone->lru_lock);
1752                         } else
1753                                 list_add(&page->lru, pages_to_free);
1754                 }
1755         }
1756         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1757         if (!is_active_lru(lru))
1758                 __count_vm_events(PGDEACTIVATE, pgmoved);
1759 }
1760 
1761 static void shrink_active_list(unsigned long nr_to_scan,
1762                                struct lruvec *lruvec,
1763                                struct scan_control *sc,
1764                                enum lru_list lru)
1765 {
1766         unsigned long nr_taken;
1767         unsigned long nr_scanned;
1768         unsigned long vm_flags;
1769         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1770         LIST_HEAD(l_active);
1771         LIST_HEAD(l_inactive);
1772         struct page *page;
1773         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1774         unsigned long nr_rotated = 0;
1775         isolate_mode_t isolate_mode = 0;
1776         int file = is_file_lru(lru);
1777         struct zone *zone = lruvec_zone(lruvec);
1778 
1779         lru_add_drain();
1780 
1781         if (!sc->may_unmap)
1782                 isolate_mode |= ISOLATE_UNMAPPED;
1783         if (!sc->may_writepage)
1784                 isolate_mode |= ISOLATE_CLEAN;
1785 
1786         spin_lock_irq(&zone->lru_lock);
1787 
1788         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1789                                      &nr_scanned, sc, isolate_mode, lru);
1790         if (global_reclaim(sc))
1791                 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1792 
1793         reclaim_stat->recent_scanned[file] += nr_taken;
1794 
1795         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1796         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1797         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1798         spin_unlock_irq(&zone->lru_lock);
1799 
1800         while (!list_empty(&l_hold)) {
1801                 cond_resched();
1802                 page = lru_to_page(&l_hold);
1803                 list_del(&page->lru);
1804 
1805                 if (unlikely(!page_evictable(page))) {
1806                         putback_lru_page(page);
1807                         continue;
1808                 }
1809 
1810                 if (unlikely(buffer_heads_over_limit)) {
1811                         if (page_has_private(page) && trylock_page(page)) {
1812                                 if (page_has_private(page))
1813                                         try_to_release_page(page, 0);
1814                                 unlock_page(page);
1815                         }
1816                 }
1817 
1818                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1819                                     &vm_flags)) {
1820                         nr_rotated += hpage_nr_pages(page);
1821                         /*
1822                          * Identify referenced, file-backed active pages and
1823                          * give them one more trip around the active list. So
1824                          * that executable code get better chances to stay in
1825                          * memory under moderate memory pressure.  Anon pages
1826                          * are not likely to be evicted by use-once streaming
1827                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1828                          * so we ignore them here.
1829                          */
1830                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1831                                 list_add(&page->lru, &l_active);
1832                                 continue;
1833                         }
1834                 }
1835 
1836                 ClearPageActive(page);  /* we are de-activating */
1837                 list_add(&page->lru, &l_inactive);
1838         }
1839 
1840         /*
1841          * Move pages back to the lru list.
1842          */
1843         spin_lock_irq(&zone->lru_lock);
1844         /*
1845          * Count referenced pages from currently used mappings as rotated,
1846          * even though only some of them are actually re-activated.  This
1847          * helps balance scan pressure between file and anonymous pages in
1848          * get_scan_count.
1849          */
1850         reclaim_stat->recent_rotated[file] += nr_rotated;
1851 
1852         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1853         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1854         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1855         spin_unlock_irq(&zone->lru_lock);
1856 
1857         mem_cgroup_uncharge_list(&l_hold);
1858         free_hot_cold_page_list(&l_hold, true);
1859 }
1860 
1861 #ifdef CONFIG_SWAP
1862 static int inactive_anon_is_low_global(struct zone *zone)
1863 {
1864         unsigned long active, inactive;
1865 
1866         active = zone_page_state(zone, NR_ACTIVE_ANON);
1867         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1868 
1869         if (inactive * zone->inactive_ratio < active)
1870                 return 1;
1871 
1872         return 0;
1873 }
1874 
1875 /**
1876  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1877  * @lruvec: LRU vector to check
1878  *
1879  * Returns true if the zone does not have enough inactive anon pages,
1880  * meaning some active anon pages need to be deactivated.
1881  */
1882 static int inactive_anon_is_low(struct lruvec *lruvec)
1883 {
1884         /*
1885          * If we don't have swap space, anonymous page deactivation
1886          * is pointless.
1887          */
1888         if (!total_swap_pages)
1889                 return 0;
1890 
1891         if (!mem_cgroup_disabled())
1892                 return mem_cgroup_inactive_anon_is_low(lruvec);
1893 
1894         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1895 }
1896 #else
1897 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1898 {
1899         return 0;
1900 }
1901 #endif
1902 
1903 /**
1904  * inactive_file_is_low - check if file pages need to be deactivated
1905  * @lruvec: LRU vector to check
1906  *
1907  * When the system is doing streaming IO, memory pressure here
1908  * ensures that active file pages get deactivated, until more
1909  * than half of the file pages are on the inactive list.
1910  *
1911  * Once we get to that situation, protect the system's working
1912  * set from being evicted by disabling active file page aging.
1913  *
1914  * This uses a different ratio than the anonymous pages, because
1915  * the page cache uses a use-once replacement algorithm.
1916  */
1917 static int inactive_file_is_low(struct lruvec *lruvec)
1918 {
1919         unsigned long inactive;
1920         unsigned long active;
1921 
1922         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1923         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1924 
1925         return active > inactive;
1926 }
1927 
1928 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1929 {
1930         if (is_file_lru(lru))
1931                 return inactive_file_is_low(lruvec);
1932         else
1933                 return inactive_anon_is_low(lruvec);
1934 }
1935 
1936 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1937                                  struct lruvec *lruvec, struct scan_control *sc)
1938 {
1939         if (is_active_lru(lru)) {
1940                 if (inactive_list_is_low(lruvec, lru))
1941                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1942                 return 0;
1943         }
1944 
1945         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1946 }
1947 
1948 enum scan_balance {
1949         SCAN_EQUAL,
1950         SCAN_FRACT,
1951         SCAN_ANON,
1952         SCAN_FILE,
1953 };
1954 
1955 /*
1956  * Determine how aggressively the anon and file LRU lists should be
1957  * scanned.  The relative value of each set of LRU lists is determined
1958  * by looking at the fraction of the pages scanned we did rotate back
1959  * onto the active list instead of evict.
1960  *
1961  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1962  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1963  */
1964 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1965                            struct scan_control *sc, unsigned long *nr,
1966                            unsigned long *lru_pages)
1967 {
1968         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1969         u64 fraction[2];
1970         u64 denominator = 0;    /* gcc */
1971         struct zone *zone = lruvec_zone(lruvec);
1972         unsigned long anon_prio, file_prio;
1973         enum scan_balance scan_balance;
1974         unsigned long anon, file;
1975         bool force_scan = false;
1976         unsigned long ap, fp;
1977         enum lru_list lru;
1978         bool some_scanned;
1979         int pass;
1980 
1981         /*
1982          * If the zone or memcg is small, nr[l] can be 0.  This
1983          * results in no scanning on this priority and a potential
1984          * priority drop.  Global direct reclaim can go to the next
1985          * zone and tends to have no problems. Global kswapd is for
1986          * zone balancing and it needs to scan a minimum amount. When
1987          * reclaiming for a memcg, a priority drop can cause high
1988          * latencies, so it's better to scan a minimum amount there as
1989          * well.
1990          */
1991         if (current_is_kswapd()) {
1992                 if (!zone_reclaimable(zone))
1993                         force_scan = true;
1994                 if (!mem_cgroup_lruvec_online(lruvec))
1995                         force_scan = true;
1996         }
1997         if (!global_reclaim(sc))
1998                 force_scan = true;
1999 
2000         /* If we have no swap space, do not bother scanning anon pages. */
2001         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
2002                 scan_balance = SCAN_FILE;
2003                 goto out;
2004         }
2005 
2006         /*
2007          * Global reclaim will swap to prevent OOM even with no
2008          * swappiness, but memcg users want to use this knob to
2009          * disable swapping for individual groups completely when
2010          * using the memory controller's swap limit feature would be
2011          * too expensive.
2012          */
2013         if (!global_reclaim(sc) && !swappiness) {
2014                 scan_balance = SCAN_FILE;
2015                 goto out;
2016         }
2017 
2018         /*
2019          * Do not apply any pressure balancing cleverness when the
2020          * system is close to OOM, scan both anon and file equally
2021          * (unless the swappiness setting disagrees with swapping).
2022          */
2023         if (!sc->priority && swappiness) {
2024                 scan_balance = SCAN_EQUAL;
2025                 goto out;
2026         }
2027 
2028         /*
2029          * Prevent the reclaimer from falling into the cache trap: as
2030          * cache pages start out inactive, every cache fault will tip
2031          * the scan balance towards the file LRU.  And as the file LRU
2032          * shrinks, so does the window for rotation from references.
2033          * This means we have a runaway feedback loop where a tiny
2034          * thrashing file LRU becomes infinitely more attractive than
2035          * anon pages.  Try to detect this based on file LRU size.
2036          */
2037         if (global_reclaim(sc)) {
2038                 unsigned long zonefile;
2039                 unsigned long zonefree;
2040 
2041                 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2042                 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2043                            zone_page_state(zone, NR_INACTIVE_FILE);
2044 
2045                 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2046                         scan_balance = SCAN_ANON;
2047                         goto out;
2048                 }
2049         }
2050 
2051         /*
2052          * There is enough inactive page cache, do not reclaim
2053          * anything from the anonymous working set right now.
2054          */
2055         if (!inactive_file_is_low(lruvec)) {
2056                 scan_balance = SCAN_FILE;
2057                 goto out;
2058         }
2059 
2060         scan_balance = SCAN_FRACT;
2061 
2062         /*
2063          * With swappiness at 100, anonymous and file have the same priority.
2064          * This scanning priority is essentially the inverse of IO cost.
2065          */
2066         anon_prio = swappiness;
2067         file_prio = 200 - anon_prio;
2068 
2069         /*
2070          * OK, so we have swap space and a fair amount of page cache
2071          * pages.  We use the recently rotated / recently scanned
2072          * ratios to determine how valuable each cache is.
2073          *
2074          * Because workloads change over time (and to avoid overflow)
2075          * we keep these statistics as a floating average, which ends
2076          * up weighing recent references more than old ones.
2077          *
2078          * anon in [0], file in [1]
2079          */
2080 
2081         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2082                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2083         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2084                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2085 
2086         spin_lock_irq(&zone->lru_lock);
2087         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2088                 reclaim_stat->recent_scanned[0] /= 2;
2089                 reclaim_stat->recent_rotated[0] /= 2;
2090         }
2091 
2092         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2093                 reclaim_stat->recent_scanned[1] /= 2;
2094                 reclaim_stat->recent_rotated[1] /= 2;
2095         }
2096 
2097         /*
2098          * The amount of pressure on anon vs file pages is inversely
2099          * proportional to the fraction of recently scanned pages on
2100          * each list that were recently referenced and in active use.
2101          */
2102         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2103         ap /= reclaim_stat->recent_rotated[0] + 1;
2104 
2105         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2106         fp /= reclaim_stat->recent_rotated[1] + 1;
2107         spin_unlock_irq(&zone->lru_lock);
2108 
2109         fraction[0] = ap;
2110         fraction[1] = fp;
2111         denominator = ap + fp + 1;
2112 out:
2113         some_scanned = false;
2114         /* Only use force_scan on second pass. */
2115         for (pass = 0; !some_scanned && pass < 2; pass++) {
2116                 *lru_pages = 0;
2117                 for_each_evictable_lru(lru) {
2118                         int file = is_file_lru(lru);
2119                         unsigned long size;
2120                         unsigned long scan;
2121 
2122                         size = get_lru_size(lruvec, lru);
2123                         scan = size >> sc->priority;
2124 
2125                         if (!scan && pass && force_scan)
2126                                 scan = min(size, SWAP_CLUSTER_MAX);
2127 
2128                         switch (scan_balance) {
2129                         case SCAN_EQUAL:
2130                                 /* Scan lists relative to size */
2131                                 break;
2132                         case SCAN_FRACT:
2133                                 /*
2134                                  * Scan types proportional to swappiness and
2135                                  * their relative recent reclaim efficiency.
2136                                  */
2137                                 scan = div64_u64(scan * fraction[file],
2138                                                         denominator);
2139                                 break;
2140                         case SCAN_FILE:
2141                         case SCAN_ANON:
2142                                 /* Scan one type exclusively */
2143                                 if ((scan_balance == SCAN_FILE) != file) {
2144                                         size = 0;
2145                                         scan = 0;
2146                                 }
2147                                 break;
2148                         default:
2149                                 /* Look ma, no brain */
2150                                 BUG();
2151                         }
2152 
2153                         *lru_pages += size;
2154                         nr[lru] = scan;
2155 
2156                         /*
2157                          * Skip the second pass and don't force_scan,
2158                          * if we found something to scan.
2159                          */
2160                         some_scanned |= !!scan;
2161                 }
2162         }
2163 }
2164 
2165 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2166 static void init_tlb_ubc(void)
2167 {
2168         /*
2169          * This deliberately does not clear the cpumask as it's expensive
2170          * and unnecessary. If there happens to be data in there then the
2171          * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2172          * then will be cleared.
2173          */
2174         current->tlb_ubc.flush_required = false;
2175 }
2176 #else
2177 static inline void init_tlb_ubc(void)
2178 {
2179 }
2180 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2181 
2182 /*
2183  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2184  */
2185 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2186                           struct scan_control *sc, unsigned long *lru_pages)
2187 {
2188         unsigned long nr[NR_LRU_LISTS];
2189         unsigned long targets[NR_LRU_LISTS];
2190         unsigned long nr_to_scan;
2191         enum lru_list lru;
2192         unsigned long nr_reclaimed = 0;
2193         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2194         struct blk_plug plug;
2195         bool scan_adjusted;
2196 
2197         get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2198 
2199         /* Record the original scan target for proportional adjustments later */
2200         memcpy(targets, nr, sizeof(nr));
2201 
2202         /*
2203          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2204          * event that can occur when there is little memory pressure e.g.
2205          * multiple streaming readers/writers. Hence, we do not abort scanning
2206          * when the requested number of pages are reclaimed when scanning at
2207          * DEF_PRIORITY on the assumption that the fact we are direct
2208          * reclaiming implies that kswapd is not keeping up and it is best to
2209          * do a batch of work at once. For memcg reclaim one check is made to
2210          * abort proportional reclaim if either the file or anon lru has already
2211          * dropped to zero at the first pass.
2212          */
2213         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2214                          sc->priority == DEF_PRIORITY);
2215 
2216         init_tlb_ubc();
2217 
2218         blk_start_plug(&plug);
2219         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2220                                         nr[LRU_INACTIVE_FILE]) {
2221                 unsigned long nr_anon, nr_file, percentage;
2222                 unsigned long nr_scanned;
2223 
2224                 for_each_evictable_lru(lru) {
2225                         if (nr[lru]) {
2226                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2227                                 nr[lru] -= nr_to_scan;
2228 
2229                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2230                                                             lruvec, sc);
2231                         }
2232                 }
2233 
2234                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2235                         continue;
2236 
2237                 /*
2238                  * For kswapd and memcg, reclaim at least the number of pages
2239                  * requested. Ensure that the anon and file LRUs are scanned
2240                  * proportionally what was requested by get_scan_count(). We
2241                  * stop reclaiming one LRU and reduce the amount scanning
2242                  * proportional to the original scan target.
2243                  */
2244                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2245                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2246 
2247                 /*
2248                  * It's just vindictive to attack the larger once the smaller
2249                  * has gone to zero.  And given the way we stop scanning the
2250                  * smaller below, this makes sure that we only make one nudge
2251                  * towards proportionality once we've got nr_to_reclaim.
2252                  */
2253                 if (!nr_file || !nr_anon)
2254                         break;
2255 
2256                 if (nr_file > nr_anon) {
2257                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2258                                                 targets[LRU_ACTIVE_ANON] + 1;
2259                         lru = LRU_BASE;
2260                         percentage = nr_anon * 100 / scan_target;
2261                 } else {
2262                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2263                                                 targets[LRU_ACTIVE_FILE] + 1;
2264                         lru = LRU_FILE;
2265                         percentage = nr_file * 100 / scan_target;
2266                 }
2267 
2268                 /* Stop scanning the smaller of the LRU */
2269                 nr[lru] = 0;
2270                 nr[lru + LRU_ACTIVE] = 0;
2271 
2272                 /*
2273                  * Recalculate the other LRU scan count based on its original
2274                  * scan target and the percentage scanning already complete
2275                  */
2276                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2277                 nr_scanned = targets[lru] - nr[lru];
2278                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2279                 nr[lru] -= min(nr[lru], nr_scanned);
2280 
2281                 lru += LRU_ACTIVE;
2282                 nr_scanned = targets[lru] - nr[lru];
2283                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2284                 nr[lru] -= min(nr[lru], nr_scanned);
2285 
2286                 scan_adjusted = true;
2287         }
2288         blk_finish_plug(&plug);
2289         sc->nr_reclaimed += nr_reclaimed;
2290 
2291         /*
2292          * Even if we did not try to evict anon pages at all, we want to
2293          * rebalance the anon lru active/inactive ratio.
2294          */
2295         if (inactive_anon_is_low(lruvec))
2296                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2297                                    sc, LRU_ACTIVE_ANON);
2298 
2299         throttle_vm_writeout(sc->gfp_mask);
2300 }
2301 
2302 /* Use reclaim/compaction for costly allocs or under memory pressure */
2303 static bool in_reclaim_compaction(struct scan_control *sc)
2304 {
2305         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2306                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2307                          sc->priority < DEF_PRIORITY - 2))
2308                 return true;
2309 
2310         return false;
2311 }
2312 
2313 /*
2314  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2315  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2316  * true if more pages should be reclaimed such that when the page allocator
2317  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2318  * It will give up earlier than that if there is difficulty reclaiming pages.
2319  */
2320 static inline bool should_continue_reclaim(struct zone *zone,
2321                                         unsigned long nr_reclaimed,
2322                                         unsigned long nr_scanned,
2323                                         struct scan_control *sc)
2324 {
2325         unsigned long pages_for_compaction;
2326         unsigned long inactive_lru_pages;
2327 
2328         /* If not in reclaim/compaction mode, stop */
2329         if (!in_reclaim_compaction(sc))
2330                 return false;
2331 
2332         /* Consider stopping depending on scan and reclaim activity */
2333         if (sc->gfp_mask & __GFP_REPEAT) {
2334                 /*
2335                  * For __GFP_REPEAT allocations, stop reclaiming if the
2336                  * full LRU list has been scanned and we are still failing
2337                  * to reclaim pages. This full LRU scan is potentially
2338                  * expensive but a __GFP_REPEAT caller really wants to succeed
2339                  */
2340                 if (!nr_reclaimed && !nr_scanned)
2341                         return false;
2342         } else {
2343                 /*
2344                  * For non-__GFP_REPEAT allocations which can presumably
2345                  * fail without consequence, stop if we failed to reclaim
2346                  * any pages from the last SWAP_CLUSTER_MAX number of
2347                  * pages that were scanned. This will return to the
2348                  * caller faster at the risk reclaim/compaction and
2349                  * the resulting allocation attempt fails
2350                  */
2351                 if (!nr_reclaimed)
2352                         return false;
2353         }
2354 
2355         /*
2356          * If we have not reclaimed enough pages for compaction and the
2357          * inactive lists are large enough, continue reclaiming
2358          */
2359         pages_for_compaction = (2UL << sc->order);
2360         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2361         if (get_nr_swap_pages() > 0)
2362                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2363         if (sc->nr_reclaimed < pages_for_compaction &&
2364                         inactive_lru_pages > pages_for_compaction)
2365                 return true;
2366 
2367         /* If compaction would go ahead or the allocation would succeed, stop */
2368         switch (compaction_suitable(zone, sc->order, 0, 0)) {
2369         case COMPACT_PARTIAL:
2370         case COMPACT_CONTINUE:
2371                 return false;
2372         default:
2373                 return true;
2374         }
2375 }
2376 
2377 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2378                         bool is_classzone)
2379 {
2380         struct reclaim_state *reclaim_state = current->reclaim_state;
2381         unsigned long nr_reclaimed, nr_scanned;
2382         bool reclaimable = false;
2383 
2384         do {
2385                 struct mem_cgroup *root = sc->target_mem_cgroup;
2386                 struct mem_cgroup_reclaim_cookie reclaim = {
2387                         .zone = zone,
2388                         .priority = sc->priority,
2389                 };
2390                 unsigned long zone_lru_pages = 0;
2391                 struct mem_cgroup *memcg;
2392 
2393                 nr_reclaimed = sc->nr_reclaimed;
2394                 nr_scanned = sc->nr_scanned;
2395 
2396                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2397                 do {
2398                         unsigned long lru_pages;
2399                         unsigned long scanned;
2400                         struct lruvec *lruvec;
2401                         int swappiness;
2402 
2403                         if (mem_cgroup_low(root, memcg)) {
2404                                 if (!sc->may_thrash)
2405                                         continue;
2406                                 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2407                         }
2408 
2409                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2410                         swappiness = mem_cgroup_swappiness(memcg);
2411                         scanned = sc->nr_scanned;
2412 
2413                         shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2414                         zone_lru_pages += lru_pages;
2415 
2416                         if (memcg && is_classzone)
2417                                 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2418                                             memcg, sc->nr_scanned - scanned,
2419                                             lru_pages);
2420 
2421                         /*
2422                          * Direct reclaim and kswapd have to scan all memory
2423                          * cgroups to fulfill the overall scan target for the
2424                          * zone.
2425                          *
2426                          * Limit reclaim, on the other hand, only cares about
2427                          * nr_to_reclaim pages to be reclaimed and it will
2428                          * retry with decreasing priority if one round over the
2429                          * whole hierarchy is not sufficient.
2430                          */
2431                         if (!global_reclaim(sc) &&
2432                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2433                                 mem_cgroup_iter_break(root, memcg);
2434                                 break;
2435                         }
2436                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2437 
2438                 /*
2439                  * Shrink the slab caches in the same proportion that
2440                  * the eligible LRU pages were scanned.
2441                  */
2442                 if (global_reclaim(sc) && is_classzone)
2443                         shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2444                                     sc->nr_scanned - nr_scanned,
2445                                     zone_lru_pages);
2446 
2447                 if (reclaim_state) {
2448                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2449                         reclaim_state->reclaimed_slab = 0;
2450                 }
2451 
2452                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2453                            sc->nr_scanned - nr_scanned,
2454                            sc->nr_reclaimed - nr_reclaimed);
2455 
2456                 if (sc->nr_reclaimed - nr_reclaimed)
2457                         reclaimable = true;
2458 
2459         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2460                                          sc->nr_scanned - nr_scanned, sc));
2461 
2462         return reclaimable;
2463 }
2464 
2465 /*
2466  * Returns true if compaction should go ahead for a high-order request, or
2467  * the high-order allocation would succeed without compaction.
2468  */
2469 static inline bool compaction_ready(struct zone *zone, int order)
2470 {
2471         unsigned long balance_gap, watermark;
2472         bool watermark_ok;
2473 
2474         /*
2475          * Compaction takes time to run and there are potentially other
2476          * callers using the pages just freed. Continue reclaiming until
2477          * there is a buffer of free pages available to give compaction
2478          * a reasonable chance of completing and allocating the page
2479          */
2480         balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2481                         zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2482         watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2483         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2484 
2485         /*
2486          * If compaction is deferred, reclaim up to a point where
2487          * compaction will have a chance of success when re-enabled
2488          */
2489         if (compaction_deferred(zone, order))
2490                 return watermark_ok;
2491 
2492         /*
2493          * If compaction is not ready to start and allocation is not likely
2494          * to succeed without it, then keep reclaiming.
2495          */
2496         if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2497                 return false;
2498 
2499         return watermark_ok;
2500 }
2501 
2502 /*
2503  * This is the direct reclaim path, for page-allocating processes.  We only
2504  * try to reclaim pages from zones which will satisfy the caller's allocation
2505  * request.
2506  *
2507  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2508  * Because:
2509  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2510  *    allocation or
2511  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2512  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2513  *    zone defense algorithm.
2514  *
2515  * If a zone is deemed to be full of pinned pages then just give it a light
2516  * scan then give up on it.
2517  *
2518  * Returns true if a zone was reclaimable.
2519  */
2520 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2521 {
2522         struct zoneref *z;
2523         struct zone *zone;
2524         unsigned long nr_soft_reclaimed;
2525         unsigned long nr_soft_scanned;
2526         gfp_t orig_mask;
2527         enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2528         bool reclaimable = false;
2529 
2530         /*
2531          * If the number of buffer_heads in the machine exceeds the maximum
2532          * allowed level, force direct reclaim to scan the highmem zone as
2533          * highmem pages could be pinning lowmem pages storing buffer_heads
2534          */
2535         orig_mask = sc->gfp_mask;
2536         if (buffer_heads_over_limit)
2537                 sc->gfp_mask |= __GFP_HIGHMEM;
2538 
2539         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2540                                         requested_highidx, sc->nodemask) {
2541                 enum zone_type classzone_idx;
2542 
2543                 if (!populated_zone(zone))
2544                         continue;
2545 
2546                 classzone_idx = requested_highidx;
2547                 while (!populated_zone(zone->zone_pgdat->node_zones +
2548                                                         classzone_idx))
2549                         classzone_idx--;
2550 
2551                 /*
2552                  * Take care memory controller reclaiming has small influence
2553                  * to global LRU.
2554                  */
2555                 if (global_reclaim(sc)) {
2556                         if (!cpuset_zone_allowed(zone,
2557                                                  GFP_KERNEL | __GFP_HARDWALL))
2558                                 continue;
2559 
2560                         if (sc->priority != DEF_PRIORITY &&
2561                             !zone_reclaimable(zone))
2562                                 continue;       /* Let kswapd poll it */
2563 
2564                         /*
2565                          * If we already have plenty of memory free for
2566                          * compaction in this zone, don't free any more.
2567                          * Even though compaction is invoked for any
2568                          * non-zero order, only frequent costly order
2569                          * reclamation is disruptive enough to become a
2570                          * noticeable problem, like transparent huge
2571                          * page allocations.
2572                          */
2573                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2574                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2575                             zonelist_zone_idx(z) <= requested_highidx &&
2576                             compaction_ready(zone, sc->order)) {
2577                                 sc->compaction_ready = true;
2578                                 continue;
2579                         }
2580 
2581                         /*
2582                          * This steals pages from memory cgroups over softlimit
2583                          * and returns the number of reclaimed pages and
2584                          * scanned pages. This works for global memory pressure
2585                          * and balancing, not for a memcg's limit.
2586                          */
2587                         nr_soft_scanned = 0;
2588                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2589                                                 sc->order, sc->gfp_mask,
2590                                                 &nr_soft_scanned);
2591                         sc->nr_reclaimed += nr_soft_reclaimed;
2592                         sc->nr_scanned += nr_soft_scanned;
2593                         if (nr_soft_reclaimed)
2594                                 reclaimable = true;
2595                         /* need some check for avoid more shrink_zone() */
2596                 }
2597 
2598                 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2599                         reclaimable = true;
2600 
2601                 if (global_reclaim(sc) &&
2602                     !reclaimable && zone_reclaimable(zone))
2603                         reclaimable = true;
2604         }
2605 
2606         /*
2607          * Restore to original mask to avoid the impact on the caller if we
2608          * promoted it to __GFP_HIGHMEM.
2609          */
2610         sc->gfp_mask = orig_mask;
2611 
2612         return reclaimable;
2613 }
2614 
2615 /*
2616  * This is the main entry point to direct page reclaim.
2617  *
2618  * If a full scan of the inactive list fails to free enough memory then we
2619  * are "out of memory" and something needs to be killed.
2620  *
2621  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2622  * high - the zone may be full of dirty or under-writeback pages, which this
2623  * caller can't do much about.  We kick the writeback threads and take explicit
2624  * naps in the hope that some of these pages can be written.  But if the
2625  * allocating task holds filesystem locks which prevent writeout this might not
2626  * work, and the allocation attempt will fail.
2627  *
2628  * returns:     0, if no pages reclaimed
2629  *              else, the number of pages reclaimed
2630  */
2631 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2632                                           struct scan_control *sc)
2633 {
2634         int initial_priority = sc->priority;
2635         unsigned long total_scanned = 0;
2636         unsigned long writeback_threshold;
2637         bool zones_reclaimable;
2638 retry:
2639         delayacct_freepages_start();
2640 
2641         if (global_reclaim(sc))
2642                 count_vm_event(ALLOCSTALL);
2643 
2644         do {
2645                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2646                                 sc->priority);
2647                 sc->nr_scanned = 0;
2648                 zones_reclaimable = shrink_zones(zonelist, sc);
2649 
2650                 total_scanned += sc->nr_scanned;
2651                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2652                         break;
2653 
2654                 if (sc->compaction_ready)
2655                         break;
2656 
2657                 /*
2658                  * If we're getting trouble reclaiming, start doing
2659                  * writepage even in laptop mode.
2660                  */
2661                 if (sc->priority < DEF_PRIORITY - 2)
2662                         sc->may_writepage = 1;
2663 
2664                 /*
2665                  * Try to write back as many pages as we just scanned.  This
2666                  * tends to cause slow streaming writers to write data to the
2667                  * disk smoothly, at the dirtying rate, which is nice.   But
2668                  * that's undesirable in laptop mode, where we *want* lumpy
2669                  * writeout.  So in laptop mode, write out the whole world.
2670                  */
2671                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2672                 if (total_scanned > writeback_threshold) {
2673                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2674                                                 WB_REASON_TRY_TO_FREE_PAGES);
2675                         sc->may_writepage = 1;
2676                 }
2677         } while (--sc->priority >= 0);
2678 
2679         delayacct_freepages_end();
2680 
2681         if (sc->nr_reclaimed)
2682                 return sc->nr_reclaimed;
2683 
2684         /* Aborted reclaim to try compaction? don't OOM, then */
2685         if (sc->compaction_ready)
2686                 return 1;
2687 
2688         /* Untapped cgroup reserves?  Don't OOM, retry. */
2689         if (!sc->may_thrash) {
2690                 sc->priority = initial_priority;
2691                 sc->may_thrash = 1;
2692                 goto retry;
2693         }
2694 
2695         /* Any of the zones still reclaimable?  Don't OOM. */
2696         if (zones_reclaimable)
2697                 return 1;
2698 
2699         return 0;
2700 }
2701 
2702 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2703 {
2704         struct zone *zone;
2705         unsigned long pfmemalloc_reserve = 0;
2706         unsigned long free_pages = 0;
2707         int i;
2708         bool wmark_ok;
2709 
2710         for (i = 0; i <= ZONE_NORMAL; i++) {
2711                 zone = &pgdat->node_zones[i];
2712                 if (!populated_zone(zone) ||
2713                     zone_reclaimable_pages(zone) == 0)
2714                         continue;
2715 
2716                 pfmemalloc_reserve += min_wmark_pages(zone);
2717                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2718         }
2719 
2720         /* If there are no reserves (unexpected config) then do not throttle */
2721         if (!pfmemalloc_reserve)
2722                 return true;
2723 
2724         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2725 
2726         /* kswapd must be awake if processes are being throttled */
2727         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2728                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2729                                                 (enum zone_type)ZONE_NORMAL);
2730                 wake_up_interruptible(&pgdat->kswapd_wait);
2731         }
2732 
2733         return wmark_ok;
2734 }
2735 
2736 /*
2737  * Throttle direct reclaimers if backing storage is backed by the network
2738  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2739  * depleted. kswapd will continue to make progress and wake the processes
2740  * when the low watermark is reached.
2741  *
2742  * Returns true if a fatal signal was delivered during throttling. If this
2743  * happens, the page allocator should not consider triggering the OOM killer.
2744  */
2745 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2746                                         nodemask_t *nodemask)
2747 {
2748         struct zoneref *z;
2749         struct zone *zone;
2750         pg_data_t *pgdat = NULL;
2751 
2752         /*
2753          * Kernel threads should not be throttled as they may be indirectly
2754          * responsible for cleaning pages necessary for reclaim to make forward
2755          * progress. kjournald for example may enter direct reclaim while
2756          * committing a transaction where throttling it could forcing other
2757          * processes to block on log_wait_commit().
2758          */
2759         if (current->flags & PF_KTHREAD)
2760                 goto out;
2761 
2762         /*
2763          * If a fatal signal is pending, this process should not throttle.
2764          * It should return quickly so it can exit and free its memory
2765          */
2766         if (fatal_signal_pending(current))
2767                 goto out;
2768 
2769         /*
2770          * Check if the pfmemalloc reserves are ok by finding the first node
2771          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2772          * GFP_KERNEL will be required for allocating network buffers when
2773          * swapping over the network so ZONE_HIGHMEM is unusable.
2774          *
2775          * Throttling is based on the first usable node and throttled processes
2776          * wait on a queue until kswapd makes progress and wakes them. There
2777          * is an affinity then between processes waking up and where reclaim
2778          * progress has been made assuming the process wakes on the same node.
2779          * More importantly, processes running on remote nodes will not compete
2780          * for remote pfmemalloc reserves and processes on different nodes
2781          * should make reasonable progress.
2782          */
2783         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2784                                         gfp_zone(gfp_mask), nodemask) {
2785                 if (zone_idx(zone) > ZONE_NORMAL)
2786                         continue;
2787 
2788                 /* Throttle based on the first usable node */
2789                 pgdat = zone->zone_pgdat;
2790                 if (pfmemalloc_watermark_ok(pgdat))
2791                         goto out;
2792                 break;
2793         }
2794 
2795         /* If no zone was usable by the allocation flags then do not throttle */
2796         if (!pgdat)
2797                 goto out;
2798 
2799         /* Account for the throttling */
2800         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2801 
2802         /*
2803          * If the caller cannot enter the filesystem, it's possible that it
2804          * is due to the caller holding an FS lock or performing a journal
2805          * transaction in the case of a filesystem like ext[3|4]. In this case,
2806          * it is not safe to block on pfmemalloc_wait as kswapd could be
2807          * blocked waiting on the same lock. Instead, throttle for up to a
2808          * second before continuing.
2809          */
2810         if (!(gfp_mask & __GFP_FS)) {
2811                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2812                         pfmemalloc_watermark_ok(pgdat), HZ);
2813 
2814                 goto check_pending;
2815         }
2816 
2817         /* Throttle until kswapd wakes the process */
2818         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2819                 pfmemalloc_watermark_ok(pgdat));
2820 
2821 check_pending:
2822         if (fatal_signal_pending(current))
2823                 return true;
2824 
2825 out:
2826         return false;
2827 }
2828 
2829 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2830                                 gfp_t gfp_mask, nodemask_t *nodemask)
2831 {
2832         unsigned long nr_reclaimed;
2833         struct scan_control sc = {
2834                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2835                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2836                 .order = order,
2837                 .nodemask = nodemask,
2838                 .priority = DEF_PRIORITY,
2839                 .may_writepage = !laptop_mode,
2840                 .may_unmap = 1,
2841                 .may_swap = 1,
2842         };
2843 
2844         /*
2845          * Do not enter reclaim if fatal signal was delivered while throttled.
2846          * 1 is returned so that the page allocator does not OOM kill at this
2847          * point.
2848          */
2849         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2850                 return 1;
2851 
2852         trace_mm_vmscan_direct_reclaim_begin(order,
2853                                 sc.may_writepage,
2854                                 gfp_mask);
2855 
2856         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2857 
2858         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2859 
2860         return nr_reclaimed;
2861 }
2862 
2863 #ifdef CONFIG_MEMCG
2864 
2865 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2866                                                 gfp_t gfp_mask, bool noswap,
2867                                                 struct zone *zone,
2868                                                 unsigned long *nr_scanned)
2869 {
2870         struct scan_control sc = {
2871                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2872                 .target_mem_cgroup = memcg,
2873                 .may_writepage = !laptop_mode,
2874                 .may_unmap = 1,
2875                 .may_swap = !noswap,
2876         };
2877         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2878         int swappiness = mem_cgroup_swappiness(memcg);
2879         unsigned long lru_pages;
2880 
2881         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2882                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2883 
2884         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2885                                                       sc.may_writepage,
2886                                                       sc.gfp_mask);
2887 
2888         /*
2889          * NOTE: Although we can get the priority field, using it
2890          * here is not a good idea, since it limits the pages we can scan.
2891          * if we don't reclaim here, the shrink_zone from balance_pgdat
2892          * will pick up pages from other mem cgroup's as well. We hack
2893          * the priority and make it zero.
2894          */
2895         shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2896 
2897         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2898 
2899         *nr_scanned = sc.nr_scanned;
2900         return sc.nr_reclaimed;
2901 }
2902 
2903 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2904                                            unsigned long nr_pages,
2905                                            gfp_t gfp_mask,
2906                                            bool may_swap)
2907 {
2908         struct zonelist *zonelist;
2909         unsigned long nr_reclaimed;
2910         int nid;
2911         struct scan_control sc = {
2912                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2913                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2914                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2915                 .target_mem_cgroup = memcg,
2916                 .priority = DEF_PRIORITY,
2917                 .may_writepage = !laptop_mode,
2918                 .may_unmap = 1,
2919                 .may_swap = may_swap,
2920         };
2921 
2922         /*
2923          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2924          * take care of from where we get pages. So the node where we start the
2925          * scan does not need to be the current node.
2926          */
2927         nid = mem_cgroup_select_victim_node(memcg);
2928 
2929         zonelist = NODE_DATA(nid)->node_zonelists;
2930 
2931         trace_mm_vmscan_memcg_reclaim_begin(0,
2932                                             sc.may_writepage,
2933                                             sc.gfp_mask);
2934 
2935         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2936 
2937         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2938 
2939         return nr_reclaimed;
2940 }
2941 #endif
2942 
2943 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2944 {
2945         struct mem_cgroup *memcg;
2946 
2947         if (!total_swap_pages)
2948                 return;
2949 
2950         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2951         do {
2952                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2953 
2954                 if (inactive_anon_is_low(lruvec))
2955                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2956                                            sc, LRU_ACTIVE_ANON);
2957 
2958                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2959         } while (memcg);
2960 }
2961 
2962 static bool zone_balanced(struct zone *zone, int order,
2963                           unsigned long balance_gap, int classzone_idx)
2964 {
2965         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2966                                     balance_gap, classzone_idx, 0))
2967                 return false;
2968 
2969         if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2970                                 order, 0, classzone_idx) == COMPACT_SKIPPED)
2971                 return false;
2972 
2973         return true;
2974 }
2975 
2976 /*
2977  * pgdat_balanced() is used when checking if a node is balanced.
2978  *
2979  * For order-0, all zones must be balanced!
2980  *
2981  * For high-order allocations only zones that meet watermarks and are in a
2982  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2983  * total of balanced pages must be at least 25% of the zones allowed by
2984  * classzone_idx for the node to be considered balanced. Forcing all zones to
2985  * be balanced for high orders can cause excessive reclaim when there are
2986  * imbalanced zones.
2987  * The choice of 25% is due to
2988  *   o a 16M DMA zone that is balanced will not balance a zone on any
2989  *     reasonable sized machine
2990  *   o On all other machines, the top zone must be at least a reasonable
2991  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2992  *     would need to be at least 256M for it to be balance a whole node.
2993  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2994  *     to balance a node on its own. These seemed like reasonable ratios.
2995  */
2996 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2997 {
2998         unsigned long managed_pages = 0;
2999         unsigned long balanced_pages = 0;
3000         int i;
3001 
3002         /* Check the watermark levels */
3003         for (i = 0; i <= classzone_idx; i++) {
3004                 struct zone *zone = pgdat->node_zones + i;
3005 
3006                 if (!populated_zone(zone))
3007                         continue;
3008 
3009                 managed_pages += zone->managed_pages;
3010 
3011                 /*
3012                  * A special case here:
3013                  *
3014                  * balance_pgdat() skips over all_unreclaimable after
3015                  * DEF_PRIORITY. Effectively, it considers them balanced so
3016                  * they must be considered balanced here as well!
3017                  */
3018                 if (!zone_reclaimable(zone)) {
3019                         balanced_pages += zone->managed_pages;
3020                         continue;
3021                 }
3022 
3023                 if (zone_balanced(zone, order, 0, i))
3024                         balanced_pages += zone->managed_pages;
3025                 else if (!order)
3026                         return false;
3027         }
3028 
3029         if (order)
3030                 return balanced_pages >= (managed_pages >> 2);
3031         else
3032                 return true;
3033 }
3034 
3035 /*
3036  * Prepare kswapd for sleeping. This verifies that there are no processes
3037  * waiting in throttle_direct_reclaim() and that watermarks have been met.
3038  *
3039  * Returns true if kswapd is ready to sleep
3040  */
3041 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3042                                         int classzone_idx)
3043 {
3044         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3045         if (remaining)
3046                 return false;
3047 
3048         /*
3049          * The throttled processes are normally woken up in balance_pgdat() as
3050          * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3051          * race between when kswapd checks the watermarks and a process gets
3052          * throttled. There is also a potential race if processes get
3053          * throttled, kswapd wakes, a large process exits thereby balancing the
3054          * zones, which causes kswapd to exit balance_pgdat() before reaching
3055          * the wake up checks. If kswapd is going to sleep, no process should
3056          * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3057          * the wake up is premature, processes will wake kswapd and get
3058          * throttled again. The difference from wake ups in balance_pgdat() is
3059          * that here we are under prepare_to_wait().
3060          */
3061         if (waitqueue_active(&pgdat->pfmemalloc_wait))
3062                 wake_up_all(&pgdat->pfmemalloc_wait);
3063 
3064         return pgdat_balanced(pgdat, order, classzone_idx);
3065 }
3066 
3067 /*
3068  * kswapd shrinks the zone by the number of pages required to reach
3069  * the high watermark.
3070  *
3071  * Returns true if kswapd scanned at least the requested number of pages to
3072  * reclaim or if the lack of progress was due to pages under writeback.
3073  * This is used to determine if the scanning priority needs to be raised.
3074  */
3075 static bool kswapd_shrink_zone(struct zone *zone,
3076                                int classzone_idx,
3077                                struct scan_control *sc,
3078                                unsigned long *nr_attempted)
3079 {
3080         int testorder = sc->order;
3081         unsigned long balance_gap;
3082         bool lowmem_pressure;
3083 
3084         /* Reclaim above the high watermark. */
3085         sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3086 
3087         /*
3088          * Kswapd reclaims only single pages with compaction enabled. Trying
3089          * too hard to reclaim until contiguous free pages have become
3090          * available can hurt performance by evicting too much useful data
3091          * from memory. Do not reclaim more than needed for compaction.
3092          */
3093         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3094                         compaction_suitable(zone, sc->order, 0, classzone_idx)
3095                                                         != COMPACT_SKIPPED)
3096                 testorder = 0;
3097 
3098         /*
3099          * We put equal pressure on every zone, unless one zone has way too
3100          * many pages free already. The "too many pages" is defined as the
3101          * high wmark plus a "gap" where the gap is either the low
3102          * watermark or 1% of the zone, whichever is smaller.
3103          */
3104         balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3105                         zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3106 
3107         /*
3108          * If there is no low memory pressure or the zone is balanced then no
3109          * reclaim is necessary
3110          */
3111         lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3112         if (!lowmem_pressure && zone_balanced(zone, testorder,
3113                                                 balance_gap, classzone_idx))
3114                 return true;
3115 
3116         shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3117 
3118         /* Account for the number of pages attempted to reclaim */
3119         *nr_attempted += sc->nr_to_reclaim;
3120 
3121         clear_bit(ZONE_WRITEBACK, &zone->flags);
3122 
3123         /*
3124          * If a zone reaches its high watermark, consider it to be no longer
3125          * congested. It's possible there are dirty pages backed by congested
3126          * BDIs but as pressure is relieved, speculatively avoid congestion
3127          * waits.
3128          */
3129         if (zone_reclaimable(zone) &&
3130             zone_balanced(zone, testorder, 0, classzone_idx)) {
3131                 clear_bit(ZONE_CONGESTED, &zone->flags);
3132                 clear_bit(ZONE_DIRTY, &zone->flags);
3133         }
3134 
3135         return sc->nr_scanned >= sc->nr_to_reclaim;
3136 }
3137 
3138 /*
3139  * For kswapd, balance_pgdat() will work across all this node's zones until
3140  * they are all at high_wmark_pages(zone).
3141  *
3142  * Returns the final order kswapd was reclaiming at
3143  *
3144  * There is special handling here for zones which are full of pinned pages.
3145  * This can happen if the pages are all mlocked, or if they are all used by
3146  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
3147  * What we do is to detect the case where all pages in the zone have been
3148  * scanned twice and there has been zero successful reclaim.  Mark the zone as
3149  * dead and from now on, only perform a short scan.  Basically we're polling
3150  * the zone for when the problem goes away.
3151  *
3152  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3153  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3154  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3155  * lower zones regardless of the number of free pages in the lower zones. This
3156  * interoperates with the page allocator fallback scheme to ensure that aging
3157  * of pages is balanced across the zones.
3158  */
3159 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3160                                                         int *classzone_idx)
3161 {
3162         int i;
3163         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
3164         unsigned long nr_soft_reclaimed;
3165         unsigned long nr_soft_scanned;
3166         struct scan_control sc = {
3167                 .gfp_mask = GFP_KERNEL,
3168                 .order = order,
3169                 .priority = DEF_PRIORITY,
3170                 .may_writepage = !laptop_mode,
3171                 .may_unmap = 1,
3172                 .may_swap = 1,
3173         };
3174         count_vm_event(PAGEOUTRUN);
3175 
3176         do {
3177                 unsigned long nr_attempted = 0;
3178                 bool raise_priority = true;
3179                 bool pgdat_needs_compaction = (order > 0);
3180 
3181                 sc.nr_reclaimed = 0;
3182 
3183                 /*
3184                  * Scan in the highmem->dma direction for the highest
3185                  * zone which needs scanning
3186                  */
3187                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188                         struct zone *zone = pgdat->node_zones + i;
3189 
3190                         if (!populated_zone(zone))
3191                                 continue;
3192 
3193                         if (sc.priority != DEF_PRIORITY &&
3194                             !zone_reclaimable(zone))
3195                                 continue;
3196 
3197                         /*
3198                          * Do some background aging of the anon list, to give
3199                          * pages a chance to be referenced before reclaiming.
3200                          */
3201                         age_active_anon(zone, &sc);
3202 
3203                         /*
3204                          * If the number of buffer_heads in the machine
3205                          * exceeds the maximum allowed level and this node
3206                          * has a highmem zone, force kswapd to reclaim from
3207                          * it to relieve lowmem pressure.
3208                          */
3209                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210                                 end_zone = i;
3211                                 break;
3212                         }
3213 
3214                         if (!zone_balanced(zone, order, 0, 0)) {
3215                                 end_zone = i;
3216                                 break;
3217                         } else {
3218                                 /*
3219                                  * If balanced, clear the dirty and congested
3220                                  * flags
3221                                  */
3222                                 clear_bit(ZONE_CONGESTED, &zone->flags);
3223                                 clear_bit(ZONE_DIRTY, &zone->flags);
3224                         }
3225                 }
3226 
3227                 if (i < 0)
3228                         goto out;
3229 
3230                 for (i = 0; i <= end_zone; i++) {
3231                         struct zone *zone = pgdat->node_zones + i;
3232 
3233                         if (!populated_zone(zone))
3234                                 continue;
3235 
3236                         /*
3237                          * If any zone is currently balanced then kswapd will
3238                          * not call compaction as it is expected that the
3239                          * necessary pages are already available.
3240                          */
3241                         if (pgdat_needs_compaction &&
3242                                         zone_watermark_ok(zone, order,
3243                                                 low_wmark_pages(zone),
3244                                                 *classzone_idx, 0))
3245                                 pgdat_needs_compaction = false;
3246                 }
3247 
3248                 /*
3249                  * If we're getting trouble reclaiming, start doing writepage
3250                  * even in laptop mode.
3251                  */
3252                 if (sc.priority < DEF_PRIORITY - 2)
3253                         sc.may_writepage = 1;
3254 
3255                 /*
3256                  * Now scan the zone in the dma->highmem direction, stopping
3257                  * at the last zone which needs scanning.
3258                  *
3259                  * We do this because the page allocator works in the opposite
3260                  * direction.  This prevents the page allocator from allocating
3261                  * pages behind kswapd's direction of progress, which would
3262                  * cause too much scanning of the lower zones.
3263                  */
3264                 for (i = 0; i <= end_zone; i++) {
3265                         struct zone *zone = pgdat->node_zones + i;
3266 
3267                         if (!populated_zone(zone))
3268                                 continue;
3269 
3270                         if (sc.priority != DEF_PRIORITY &&
3271                             !zone_reclaimable(zone))
3272                                 continue;
3273 
3274                         sc.nr_scanned = 0;
3275 
3276                         nr_soft_scanned = 0;
3277                         /*
3278                          * Call soft limit reclaim before calling shrink_zone.
3279                          */
3280                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3281                                                         order, sc.gfp_mask,
3282                                                         &nr_soft_scanned);
3283                         sc.nr_reclaimed += nr_soft_reclaimed;
3284 
3285                         /*
3286                          * There should be no need to raise the scanning
3287                          * priority if enough pages are already being scanned
3288                          * that that high watermark would be met at 100%
3289                          * efficiency.
3290                          */
3291                         if (kswapd_shrink_zone(zone, end_zone,
3292                                                &sc, &nr_attempted))
3293                                 raise_priority = false;
3294                 }
3295 
3296                 /*
3297                  * If the low watermark is met there is no need for processes
3298                  * to be throttled on pfmemalloc_wait as they should not be
3299                  * able to safely make forward progress. Wake them
3300                  */
3301                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3302                                 pfmemalloc_watermark_ok(pgdat))
3303                         wake_up_all(&pgdat->pfmemalloc_wait);
3304 
3305                 /*
3306                  * Fragmentation may mean that the system cannot be rebalanced
3307                  * for high-order allocations in all zones. If twice the
3308                  * allocation size has been reclaimed and the zones are still
3309                  * not balanced then recheck the watermarks at order-0 to
3310                  * prevent kswapd reclaiming excessively. Assume that a
3311                  * process requested a high-order can direct reclaim/compact.
3312                  */
3313                 if (order && sc.nr_reclaimed >= 2UL << order)
3314                         order = sc.order = 0;
3315 
3316                 /* Check if kswapd should be suspending */
3317                 if (try_to_freeze() || kthread_should_stop())
3318                         break;
3319 
3320                 /*
3321                  * Compact if necessary and kswapd is reclaiming at least the
3322                  * high watermark number of pages as requsted
3323                  */
3324                 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3325                         compact_pgdat(pgdat, order);
3326 
3327                 /*
3328                  * Raise priority if scanning rate is too low or there was no
3329                  * progress in reclaiming pages
3330                  */
3331                 if (raise_priority || !sc.nr_reclaimed)
3332                         sc.priority--;
3333         } while (sc.priority >= 1 &&
3334                  !pgdat_balanced(pgdat, order, *classzone_idx));
3335 
3336 out:
3337         /*
3338          * Return the order we were reclaiming at so prepare_kswapd_sleep()
3339          * makes a decision on the order we were last reclaiming at. However,
3340          * if another caller entered the allocator slow path while kswapd
3341          * was awake, order will remain at the higher level
3342          */
3343         *classzone_idx = end_zone;
3344         return order;
3345 }
3346 
3347 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3348 {
3349         long remaining = 0;
3350         DEFINE_WAIT(wait);
3351 
3352         if (freezing(current) || kthread_should_stop())
3353                 return;
3354 
3355         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3356 
3357         /* Try to sleep for a short interval */
3358         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3359                 remaining = schedule_timeout(HZ/10);
3360                 finish_wait(&pgdat->kswapd_wait, &wait);
3361                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3362         }
3363 
3364         /*
3365          * After a short sleep, check if it was a premature sleep. If not, then
3366          * go fully to sleep until explicitly woken up.
3367          */
3368         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3369                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3370 
3371                 /*
3372                  * vmstat counters are not perfectly accurate and the estimated
3373                  * value for counters such as NR_FREE_PAGES can deviate from the
3374                  * true value by nr_online_cpus * threshold. To avoid the zone
3375                  * watermarks being breached while under pressure, we reduce the
3376                  * per-cpu vmstat threshold while kswapd is awake and restore
3377                  * them before going back to sleep.
3378                  */
3379                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3380 
3381                 /*
3382                  * Compaction records what page blocks it recently failed to
3383                  * isolate pages from and skips them in the future scanning.
3384                  * When kswapd is going to sleep, it is reasonable to assume
3385                  * that pages and compaction may succeed so reset the cache.
3386                  */
3387                 reset_isolation_suitable(pgdat);
3388 
3389                 if (!kthread_should_stop())
3390                         schedule();
3391 
3392                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3393         } else {
3394                 if (remaining)
3395                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3396                 else
3397                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3398         }
3399         finish_wait(&pgdat->kswapd_wait, &wait);
3400 }
3401 
3402 /*
3403  * The background pageout daemon, started as a kernel thread
3404  * from the init process.
3405  *
3406  * This basically trickles out pages so that we have _some_
3407  * free memory available even if there is no other activity
3408  * that frees anything up. This is needed for things like routing
3409  * etc, where we otherwise might have all activity going on in
3410  * asynchronous contexts that cannot page things out.
3411  *
3412  * If there are applications that are active memory-allocators
3413  * (most normal use), this basically shouldn't matter.
3414  */
3415 static int kswapd(void *p)
3416 {
3417         unsigned long order, new_order;
3418         unsigned balanced_order;
3419         int classzone_idx, new_classzone_idx;
3420         int balanced_classzone_idx;
3421         pg_data_t *pgdat = (pg_data_t*)p;
3422         struct task_struct *tsk = current;
3423 
3424         struct reclaim_state reclaim_state = {
3425                 .reclaimed_slab = 0,
3426         };
3427         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3428 
3429         lockdep_set_current_reclaim_state(GFP_KERNEL);
3430 
3431         if (!cpumask_empty(cpumask))
3432                 set_cpus_allowed_ptr(tsk, cpumask);
3433         current->reclaim_state = &reclaim_state;
3434 
3435         /*
3436          * Tell the memory management that we're a "memory allocator",
3437          * and that if we need more memory we should get access to it
3438          * regardless (see "__alloc_pages()"). "kswapd" should
3439          * never get caught in the normal page freeing logic.
3440          *
3441          * (Kswapd normally doesn't need memory anyway, but sometimes
3442          * you need a small amount of memory in order to be able to
3443          * page out something else, and this flag essentially protects
3444          * us from recursively trying to free more memory as we're
3445          * trying to free the first piece of memory in the first place).
3446          */
3447         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3448         set_freezable();
3449 
3450         order = new_order = 0;
3451         balanced_order = 0;
3452         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3453         balanced_classzone_idx = classzone_idx;
3454         for ( ; ; ) {
3455                 bool ret;
3456 
3457                 /*
3458                  * If the last balance_pgdat was unsuccessful it's unlikely a
3459                  * new request of a similar or harder type will succeed soon
3460                  * so consider going to sleep on the basis we reclaimed at
3461                  */
3462                 if (balanced_classzone_idx >= new_classzone_idx &&
3463                                         balanced_order == new_order) {
3464                         new_order = pgdat->kswapd_max_order;
3465                         new_classzone_idx = pgdat->classzone_idx;
3466                         pgdat->kswapd_max_order =  0;
3467                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3468                 }
3469 
3470                 if (order < new_order || classzone_idx > new_classzone_idx) {
3471                         /*
3472                          * Don't sleep if someone wants a larger 'order'
3473                          * allocation or has tigher zone constraints
3474                          */
3475                         order = new_order;
3476                         classzone_idx = new_classzone_idx;
3477                 } else {
3478                         kswapd_try_to_sleep(pgdat, balanced_order,
3479                                                 balanced_classzone_idx);
3480                         order = pgdat->kswapd_max_order;
3481                         classzone_idx = pgdat->classzone_idx;
3482                         new_order = order;
3483                         new_classzone_idx = classzone_idx;
3484                         pgdat->kswapd_max_order = 0;
3485                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3486                 }
3487 
3488                 ret = try_to_freeze();
3489                 if (kthread_should_stop())
3490                         break;
3491 
3492                 /*
3493                  * We can speed up thawing tasks if we don't call balance_pgdat
3494                  * after returning from the refrigerator
3495                  */
3496                 if (!ret) {
3497                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3498                         balanced_classzone_idx = classzone_idx;
3499                         balanced_order = balance_pgdat(pgdat, order,
3500                                                 &balanced_classzone_idx);
3501                 }
3502         }
3503 
3504         tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3505         current->reclaim_state = NULL;
3506         lockdep_clear_current_reclaim_state();
3507 
3508         return 0;
3509 }
3510 
3511 /*
3512  * A zone is low on free memory, so wake its kswapd task to service it.
3513  */
3514 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3515 {
3516         pg_data_t *pgdat;
3517 
3518         if (!populated_zone(zone))
3519                 return;
3520 
3521         if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3522                 return;
3523         pgdat = zone->zone_pgdat;
3524         if (pgdat->kswapd_max_order < order) {
3525                 pgdat->kswapd_max_order = order;
3526                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3527         }
3528         if (!waitqueue_active(&pgdat->kswapd_wait))
3529                 return;
3530         if (zone_balanced(zone, order, 0, 0))
3531                 return;
3532 
3533         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3534         wake_up_interruptible(&pgdat->kswapd_wait);
3535 }
3536 
3537 #ifdef CONFIG_HIBERNATION
3538 /*
3539  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3540  * freed pages.
3541  *
3542  * Rather than trying to age LRUs the aim is to preserve the overall
3543  * LRU order by reclaiming preferentially
3544  * inactive > active > active referenced > active mapped
3545  */
3546 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3547 {
3548         struct reclaim_state reclaim_state;
3549         struct scan_control sc = {
3550                 .nr_to_reclaim = nr_to_reclaim,
3551                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3552                 .priority = DEF_PRIORITY,
3553                 .may_writepage = 1,
3554                 .may_unmap = 1,
3555                 .may_swap = 1,
3556                 .hibernation_mode = 1,
3557         };
3558         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3559         struct task_struct *p = current;
3560         unsigned long nr_reclaimed;
3561 
3562         p->flags |= PF_MEMALLOC;
3563         lockdep_set_current_reclaim_state(sc.gfp_mask);
3564         reclaim_state.reclaimed_slab = 0;
3565         p->reclaim_state = &reclaim_state;
3566 
3567         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3568 
3569         p->reclaim_state = NULL;
3570         lockdep_clear_current_reclaim_state();
3571         p->flags &= ~PF_MEMALLOC;
3572 
3573         return nr_reclaimed;
3574 }
3575 #endif /* CONFIG_HIBERNATION */
3576 
3577 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3578    not required for correctness.  So if the last cpu in a node goes
3579    away, we get changed to run anywhere: as the first one comes back,
3580    restore their cpu bindings. */
3581 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3582                         void *hcpu)
3583 {
3584         int nid;
3585 
3586         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3587                 for_each_node_state(nid, N_MEMORY) {
3588                         pg_data_t *pgdat = NODE_DATA(nid);
3589                         const struct cpumask *mask;
3590 
3591                         mask = cpumask_of_node(pgdat->node_id);
3592 
3593                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3594                                 /* One of our CPUs online: restore mask */
3595                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3596                 }
3597         }
3598         return NOTIFY_OK;
3599 }
3600 
3601 /*
3602  * This kswapd start function will be called by init and node-hot-add.
3603  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3604  */
3605 int kswapd_run(int nid)
3606 {
3607         pg_data_t *pgdat = NODE_DATA(nid);
3608         int ret = 0;
3609 
3610         if (pgdat->kswapd)
3611                 return 0;
3612 
3613         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3614         if (IS_ERR(pgdat->kswapd)) {
3615                 /* failure at boot is fatal */
3616                 BUG_ON(system_state == SYSTEM_BOOTING);
3617                 pr_err("Failed to start kswapd on node %d\n", nid);
3618                 ret = PTR_ERR(pgdat->kswapd);
3619                 pgdat->kswapd = NULL;
3620         }
3621         return ret;
3622 }
3623 
3624 /*
3625  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3626  * hold mem_hotplug_begin/end().
3627  */
3628 void kswapd_stop(int nid)
3629 {
3630         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3631 
3632         if (kswapd) {
3633                 kthread_stop(kswapd);
3634                 NODE_DATA(nid)->kswapd = NULL;
3635         }
3636 }
3637 
3638 static int __init kswapd_init(void)
3639 {
3640         int nid;
3641 
3642         swap_setup();
3643         for_each_node_state(nid, N_MEMORY)
3644                 kswapd_run(nid);
3645         hotcpu_notifier(cpu_callback, 0);
3646         return 0;
3647 }
3648 
3649 module_init(kswapd_init)
3650 
3651 #ifdef CONFIG_NUMA
3652 /*
3653  * Zone reclaim mode
3654  *
3655  * If non-zero call zone_reclaim when the number of free pages falls below
3656  * the watermarks.
3657  */
3658 int zone_reclaim_mode __read_mostly;
3659 
3660 #define RECLAIM_OFF 0
3661 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3662 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3663 #define RECLAIM_UNMAP (1<<2)    /* Unmap pages during reclaim */
3664 
3665 /*
3666  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3667  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3668  * a zone.
3669  */
3670 #define ZONE_RECLAIM_PRIORITY 4
3671 
3672 /*
3673  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3674  * occur.
3675  */
3676 int sysctl_min_unmapped_ratio = 1;
3677 
3678 /*
3679  * If the number of slab pages in a zone grows beyond this percentage then
3680  * slab reclaim needs to occur.
3681  */
3682 int sysctl_min_slab_ratio = 5;
3683 
3684 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3685 {
3686         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3687         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3688                 zone_page_state(zone, NR_ACTIVE_FILE);
3689 
3690         /*
3691          * It's possible for there to be more file mapped pages than
3692          * accounted for by the pages on the file LRU lists because
3693          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3694          */
3695         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3696 }
3697 
3698 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3699 static long zone_pagecache_reclaimable(struct zone *zone)
3700 {
3701         long nr_pagecache_reclaimable;
3702         long delta = 0;
3703 
3704         /*
3705          * If RECLAIM_UNMAP is set, then all file pages are considered
3706          * potentially reclaimable. Otherwise, we have to worry about
3707          * pages like swapcache and zone_unmapped_file_pages() provides
3708          * a better estimate
3709          */
3710         if (zone_reclaim_mode & RECLAIM_UNMAP)
3711                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3712         else
3713                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3714 
3715         /* If we can't clean pages, remove dirty pages from consideration */
3716         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3717                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3718 
3719         /* Watch for any possible underflows due to delta */
3720         if (unlikely(delta > nr_pagecache_reclaimable))
3721                 delta = nr_pagecache_reclaimable;
3722 
3723         return nr_pagecache_reclaimable - delta;
3724 }
3725 
3726 /*
3727  * Try to free up some pages from this zone through reclaim.
3728  */
3729 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3730 {
3731         /* Minimum pages needed in order to stay on node */
3732         const unsigned long nr_pages = 1 << order;
3733         struct task_struct *p = current;
3734         struct reclaim_state reclaim_state;
3735         struct scan_control sc = {
3736                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3737                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3738                 .order = order,
3739                 .priority = ZONE_RECLAIM_PRIORITY,
3740                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3741                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3742                 .may_swap = 1,
3743         };
3744 
3745         cond_resched();
3746         /*
3747          * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3748          * and we also need to be able to write out pages for RECLAIM_WRITE
3749          * and RECLAIM_UNMAP.
3750          */
3751         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3752         lockdep_set_current_reclaim_state(gfp_mask);
3753         reclaim_state.reclaimed_slab = 0;
3754         p->reclaim_state = &reclaim_state;
3755 
3756         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3757                 /*
3758                  * Free memory by calling shrink zone with increasing
3759                  * priorities until we have enough memory freed.
3760                  */
3761                 do {
3762                         shrink_zone(zone, &sc, true);
3763                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3764         }
3765 
3766         p->reclaim_state = NULL;
3767         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3768         lockdep_clear_current_reclaim_state();
3769         return sc.nr_reclaimed >= nr_pages;
3770 }
3771 
3772 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3773 {
3774         int node_id;
3775         int ret;
3776 
3777         /*
3778          * Zone reclaim reclaims unmapped file backed pages and
3779          * slab pages if we are over the defined limits.
3780          *
3781          * A small portion of unmapped file backed pages is needed for
3782          * file I/O otherwise pages read by file I/O will be immediately
3783          * thrown out if the zone is overallocated. So we do not reclaim
3784          * if less than a specified percentage of the zone is used by
3785          * unmapped file backed pages.
3786          */
3787         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3788             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3789                 return ZONE_RECLAIM_FULL;
3790 
3791         if (!zone_reclaimable(zone))
3792                 return ZONE_RECLAIM_FULL;
3793 
3794         /*
3795          * Do not scan if the allocation should not be delayed.
3796          */
3797         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3798                 return ZONE_RECLAIM_NOSCAN;
3799 
3800         /*
3801          * Only run zone reclaim on the local zone or on zones that do not
3802          * have associated processors. This will favor the local processor
3803          * over remote processors and spread off node memory allocations
3804          * as wide as possible.
3805          */
3806         node_id = zone_to_nid(zone);
3807         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3808                 return ZONE_RECLAIM_NOSCAN;
3809 
3810         if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3811                 return ZONE_RECLAIM_NOSCAN;
3812 
3813         ret = __zone_reclaim(zone, gfp_mask, order);
3814         clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3815 
3816         if (!ret)
3817                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3818 
3819         return ret;
3820 }
3821 #endif
3822 
3823 /*
3824  * page_evictable - test whether a page is evictable
3825  * @page: the page to test
3826  *
3827  * Test whether page is evictable--i.e., should be placed on active/inactive
3828  * lists vs unevictable list.
3829  *
3830  * Reasons page might not be evictable:
3831  * (1) page's mapping marked unevictable
3832  * (2) page is part of an mlocked VMA
3833  *
3834  */
3835 int page_evictable(struct page *page)
3836 {
3837         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3838 }
3839 
3840 #ifdef CONFIG_SHMEM
3841 /**
3842  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3843  * @pages:      array of pages to check
3844  * @nr_pages:   number of pages to check
3845  *
3846  * Checks pages for evictability and moves them to the appropriate lru list.
3847  *
3848  * This function is only used for SysV IPC SHM_UNLOCK.
3849  */
3850 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3851 {
3852         struct lruvec *lruvec;
3853         struct zone *zone = NULL;
3854         int pgscanned = 0;
3855         int pgrescued = 0;
3856         int i;
3857 
3858         for (i = 0; i < nr_pages; i++) {
3859                 struct page *page = pages[i];
3860                 struct zone *pagezone;
3861 
3862                 pgscanned++;
3863                 pagezone = page_zone(page);
3864                 if (pagezone != zone) {
3865                         if (zone)
3866                                 spin_unlock_irq(&zone->lru_lock);
3867                         zone = pagezone;
3868                         spin_lock_irq(&zone->lru_lock);
3869                 }
3870                 lruvec = mem_cgroup_page_lruvec(page, zone);
3871 
3872                 if (!PageLRU(page) || !PageUnevictable(page))
3873                         continue;
3874 
3875                 if (page_evictable(page)) {
3876                         enum lru_list lru = page_lru_base_type(page);
3877 
3878                         VM_BUG_ON_PAGE(PageActive(page), page);
3879                         ClearPageUnevictable(page);
3880                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3881                         add_page_to_lru_list(page, lruvec, lru);
3882                         pgrescued++;
3883                 }
3884         }
3885 
3886         if (zone) {
3887                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3888                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3889                 spin_unlock_irq(&zone->lru_lock);
3890         }
3891 }
3892 #endif /* CONFIG_SHMEM */
3893 

~ [ 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