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

TOMOYO Linux Cross Reference
Linux/mm/vmscan.c

Version: ~ [ linux-5.14-rc3 ] ~ [ linux-5.13.5 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.53 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.135 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.198 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.240 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.276 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.276 ] ~ [ 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/sched/mm.h>
 18 #include <linux/module.h>
 19 #include <linux/gfp.h>
 20 #include <linux/kernel_stat.h>
 21 #include <linux/swap.h>
 22 #include <linux/pagemap.h>
 23 #include <linux/init.h>
 24 #include <linux/highmem.h>
 25 #include <linux/vmpressure.h>
 26 #include <linux/vmstat.h>
 27 #include <linux/file.h>
 28 #include <linux/writeback.h>
 29 #include <linux/blkdev.h>
 30 #include <linux/buffer_head.h>  /* for try_to_release_page(),
 31                                         buffer_heads_over_limit */
 32 #include <linux/mm_inline.h>
 33 #include <linux/backing-dev.h>
 34 #include <linux/rmap.h>
 35 #include <linux/topology.h>
 36 #include <linux/cpu.h>
 37 #include <linux/cpuset.h>
 38 #include <linux/compaction.h>
 39 #include <linux/notifier.h>
 40 #include <linux/rwsem.h>
 41 #include <linux/delay.h>
 42 #include <linux/kthread.h>
 43 #include <linux/freezer.h>
 44 #include <linux/memcontrol.h>
 45 #include <linux/delayacct.h>
 46 #include <linux/sysctl.h>
 47 #include <linux/oom.h>
 48 #include <linux/prefetch.h>
 49 #include <linux/printk.h>
 50 #include <linux/dax.h>
 51 
 52 #include <asm/tlbflush.h>
 53 #include <asm/div64.h>
 54 
 55 #include <linux/swapops.h>
 56 #include <linux/balloon_compaction.h>
 57 
 58 #include "internal.h"
 59 
 60 #define CREATE_TRACE_POINTS
 61 #include <trace/events/vmscan.h>
 62 
 63 struct scan_control {
 64         /* How many pages shrink_list() should reclaim */
 65         unsigned long nr_to_reclaim;
 66 
 67         /* This context's GFP mask */
 68         gfp_t gfp_mask;
 69 
 70         /* Allocation order */
 71         int order;
 72 
 73         /*
 74          * Nodemask of nodes allowed by the caller. If NULL, all nodes
 75          * are scanned.
 76          */
 77         nodemask_t      *nodemask;
 78 
 79         /*
 80          * The memory cgroup that hit its limit and as a result is the
 81          * primary target of this reclaim invocation.
 82          */
 83         struct mem_cgroup *target_mem_cgroup;
 84 
 85         /* Scan (total_size >> priority) pages at once */
 86         int priority;
 87 
 88         /* The highest zone to isolate pages for reclaim from */
 89         enum zone_type reclaim_idx;
 90 
 91         /* Writepage batching in laptop mode; RECLAIM_WRITE */
 92         unsigned int may_writepage:1;
 93 
 94         /* Can mapped pages be reclaimed? */
 95         unsigned int may_unmap:1;
 96 
 97         /* Can pages be swapped as part of reclaim? */
 98         unsigned int may_swap:1;
 99 
100         /* Can cgroups be reclaimed below their normal consumption range? */
101         unsigned int may_thrash:1;
102 
103         unsigned int hibernation_mode:1;
104 
105         /* One of the zones is ready for compaction */
106         unsigned int compaction_ready:1;
107 
108         /* Incremented by the number of inactive pages that were scanned */
109         unsigned long nr_scanned;
110 
111         /* Number of pages freed so far during a call to shrink_zones() */
112         unsigned long nr_reclaimed;
113 };
114 
115 #ifdef ARCH_HAS_PREFETCH
116 #define prefetch_prev_lru_page(_page, _base, _field)                    \
117         do {                                                            \
118                 if ((_page)->lru.prev != _base) {                       \
119                         struct page *prev;                              \
120                                                                         \
121                         prev = lru_to_page(&(_page->lru));              \
122                         prefetch(&prev->_field);                        \
123                 }                                                       \
124         } while (0)
125 #else
126 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
127 #endif
128 
129 #ifdef ARCH_HAS_PREFETCHW
130 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
131         do {                                                            \
132                 if ((_page)->lru.prev != _base) {                       \
133                         struct page *prev;                              \
134                                                                         \
135                         prev = lru_to_page(&(_page->lru));              \
136                         prefetchw(&prev->_field);                       \
137                 }                                                       \
138         } while (0)
139 #else
140 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
141 #endif
142 
143 /*
144  * From 0 .. 100.  Higher means more swappy.
145  */
146 int vm_swappiness = 60;
147 /*
148  * The total number of pages which are beyond the high watermark within all
149  * zones.
150  */
151 unsigned long vm_total_pages;
152 
153 static LIST_HEAD(shrinker_list);
154 static DECLARE_RWSEM(shrinker_rwsem);
155 
156 #ifdef CONFIG_MEMCG
157 static bool global_reclaim(struct scan_control *sc)
158 {
159         return !sc->target_mem_cgroup;
160 }
161 
162 /**
163  * sane_reclaim - is the usual dirty throttling mechanism operational?
164  * @sc: scan_control in question
165  *
166  * The normal page dirty throttling mechanism in balance_dirty_pages() is
167  * completely broken with the legacy memcg and direct stalling in
168  * shrink_page_list() is used for throttling instead, which lacks all the
169  * niceties such as fairness, adaptive pausing, bandwidth proportional
170  * allocation and configurability.
171  *
172  * This function tests whether the vmscan currently in progress can assume
173  * that the normal dirty throttling mechanism is operational.
174  */
175 static bool sane_reclaim(struct scan_control *sc)
176 {
177         struct mem_cgroup *memcg = sc->target_mem_cgroup;
178 
179         if (!memcg)
180                 return true;
181 #ifdef CONFIG_CGROUP_WRITEBACK
182         if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
183                 return true;
184 #endif
185         return false;
186 }
187 #else
188 static bool global_reclaim(struct scan_control *sc)
189 {
190         return true;
191 }
192 
193 static bool sane_reclaim(struct scan_control *sc)
194 {
195         return true;
196 }
197 #endif
198 
199 /*
200  * This misses isolated pages which are not accounted for to save counters.
201  * As the data only determines if reclaim or compaction continues, it is
202  * not expected that isolated pages will be a dominating factor.
203  */
204 unsigned long zone_reclaimable_pages(struct zone *zone)
205 {
206         unsigned long nr;
207 
208         nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
209                 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
210         if (get_nr_swap_pages() > 0)
211                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
212                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
213 
214         return nr;
215 }
216 
217 unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
218 {
219         unsigned long nr;
220 
221         nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
222              node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
223              node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
224 
225         if (get_nr_swap_pages() > 0)
226                 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
227                       node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
228                       node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
229 
230         return nr;
231 }
232 
233 bool pgdat_reclaimable(struct pglist_data *pgdat)
234 {
235         return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
236                 pgdat_reclaimable_pages(pgdat) * 6;
237 }
238 
239 /**
240  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
241  * @lruvec: lru vector
242  * @lru: lru to use
243  * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
244  */
245 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
246 {
247         unsigned long lru_size;
248         int zid;
249 
250         if (!mem_cgroup_disabled())
251                 lru_size = mem_cgroup_get_lru_size(lruvec, lru);
252         else
253                 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
254 
255         for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
256                 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
257                 unsigned long size;
258 
259                 if (!managed_zone(zone))
260                         continue;
261 
262                 if (!mem_cgroup_disabled())
263                         size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
264                 else
265                         size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
266                                        NR_ZONE_LRU_BASE + lru);
267                 lru_size -= min(size, lru_size);
268         }
269 
270         return lru_size;
271 
272 }
273 
274 /*
275  * Add a shrinker callback to be called from the vm.
276  */
277 int register_shrinker(struct shrinker *shrinker)
278 {
279         size_t size = sizeof(*shrinker->nr_deferred);
280 
281         if (shrinker->flags & SHRINKER_NUMA_AWARE)
282                 size *= nr_node_ids;
283 
284         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
285         if (!shrinker->nr_deferred)
286                 return -ENOMEM;
287 
288         down_write(&shrinker_rwsem);
289         list_add_tail(&shrinker->list, &shrinker_list);
290         up_write(&shrinker_rwsem);
291         return 0;
292 }
293 EXPORT_SYMBOL(register_shrinker);
294 
295 /*
296  * Remove one
297  */
298 void unregister_shrinker(struct shrinker *shrinker)
299 {
300         down_write(&shrinker_rwsem);
301         list_del(&shrinker->list);
302         up_write(&shrinker_rwsem);
303         kfree(shrinker->nr_deferred);
304 }
305 EXPORT_SYMBOL(unregister_shrinker);
306 
307 #define SHRINK_BATCH 128
308 
309 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
310                                     struct shrinker *shrinker,
311                                     unsigned long nr_scanned,
312                                     unsigned long nr_eligible)
313 {
314         unsigned long freed = 0;
315         unsigned long long delta;
316         long total_scan;
317         long freeable;
318         long nr;
319         long new_nr;
320         int nid = shrinkctl->nid;
321         long batch_size = shrinker->batch ? shrinker->batch
322                                           : SHRINK_BATCH;
323         long scanned = 0, next_deferred;
324 
325         freeable = shrinker->count_objects(shrinker, shrinkctl);
326         if (freeable == 0)
327                 return 0;
328 
329         /*
330          * copy the current shrinker scan count into a local variable
331          * and zero it so that other concurrent shrinker invocations
332          * don't also do this scanning work.
333          */
334         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
335 
336         total_scan = nr;
337         delta = (4 * nr_scanned) / shrinker->seeks;
338         delta *= freeable;
339         do_div(delta, nr_eligible + 1);
340         total_scan += delta;
341         if (total_scan < 0) {
342                 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
343                        shrinker->scan_objects, total_scan);
344                 total_scan = freeable;
345                 next_deferred = nr;
346         } else
347                 next_deferred = total_scan;
348 
349         /*
350          * We need to avoid excessive windup on filesystem shrinkers
351          * due to large numbers of GFP_NOFS allocations causing the
352          * shrinkers to return -1 all the time. This results in a large
353          * nr being built up so when a shrink that can do some work
354          * comes along it empties the entire cache due to nr >>>
355          * freeable. This is bad for sustaining a working set in
356          * memory.
357          *
358          * Hence only allow the shrinker to scan the entire cache when
359          * a large delta change is calculated directly.
360          */
361         if (delta < freeable / 4)
362                 total_scan = min(total_scan, freeable / 2);
363 
364         /*
365          * Avoid risking looping forever due to too large nr value:
366          * never try to free more than twice the estimate number of
367          * freeable entries.
368          */
369         if (total_scan > freeable * 2)
370                 total_scan = freeable * 2;
371 
372         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
373                                    nr_scanned, nr_eligible,
374                                    freeable, delta, total_scan);
375 
376         /*
377          * Normally, we should not scan less than batch_size objects in one
378          * pass to avoid too frequent shrinker calls, but if the slab has less
379          * than batch_size objects in total and we are really tight on memory,
380          * we will try to reclaim all available objects, otherwise we can end
381          * up failing allocations although there are plenty of reclaimable
382          * objects spread over several slabs with usage less than the
383          * batch_size.
384          *
385          * We detect the "tight on memory" situations by looking at the total
386          * number of objects we want to scan (total_scan). If it is greater
387          * than the total number of objects on slab (freeable), we must be
388          * scanning at high prio and therefore should try to reclaim as much as
389          * possible.
390          */
391         while (total_scan >= batch_size ||
392                total_scan >= freeable) {
393                 unsigned long ret;
394                 unsigned long nr_to_scan = min(batch_size, total_scan);
395 
396                 shrinkctl->nr_to_scan = nr_to_scan;
397                 ret = shrinker->scan_objects(shrinker, shrinkctl);
398                 if (ret == SHRINK_STOP)
399                         break;
400                 freed += ret;
401 
402                 count_vm_events(SLABS_SCANNED, nr_to_scan);
403                 total_scan -= nr_to_scan;
404                 scanned += nr_to_scan;
405 
406                 cond_resched();
407         }
408 
409         if (next_deferred >= scanned)
410                 next_deferred -= scanned;
411         else
412                 next_deferred = 0;
413         /*
414          * move the unused scan count back into the shrinker in a
415          * manner that handles concurrent updates. If we exhausted the
416          * scan, there is no need to do an update.
417          */
418         if (next_deferred > 0)
419                 new_nr = atomic_long_add_return(next_deferred,
420                                                 &shrinker->nr_deferred[nid]);
421         else
422                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
423 
424         trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
425         return freed;
426 }
427 
428 /**
429  * shrink_slab - shrink slab caches
430  * @gfp_mask: allocation context
431  * @nid: node whose slab caches to target
432  * @memcg: memory cgroup whose slab caches to target
433  * @nr_scanned: pressure numerator
434  * @nr_eligible: pressure denominator
435  *
436  * Call the shrink functions to age shrinkable caches.
437  *
438  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
439  * unaware shrinkers will receive a node id of 0 instead.
440  *
441  * @memcg specifies the memory cgroup to target. If it is not NULL,
442  * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
443  * objects from the memory cgroup specified. Otherwise, only unaware
444  * shrinkers are called.
445  *
446  * @nr_scanned and @nr_eligible form a ratio that indicate how much of
447  * the available objects should be scanned.  Page reclaim for example
448  * passes the number of pages scanned and the number of pages on the
449  * LRU lists that it considered on @nid, plus a bias in @nr_scanned
450  * when it encountered mapped pages.  The ratio is further biased by
451  * the ->seeks setting of the shrink function, which indicates the
452  * cost to recreate an object relative to that of an LRU page.
453  *
454  * Returns the number of reclaimed slab objects.
455  */
456 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
457                                  struct mem_cgroup *memcg,
458                                  unsigned long nr_scanned,
459                                  unsigned long nr_eligible)
460 {
461         struct shrinker *shrinker;
462         unsigned long freed = 0;
463 
464         if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
465                 return 0;
466 
467         if (nr_scanned == 0)
468                 nr_scanned = SWAP_CLUSTER_MAX;
469 
470         if (!down_read_trylock(&shrinker_rwsem)) {
471                 /*
472                  * If we would return 0, our callers would understand that we
473                  * have nothing else to shrink and give up trying. By returning
474                  * 1 we keep it going and assume we'll be able to shrink next
475                  * time.
476                  */
477                 freed = 1;
478                 goto out;
479         }
480 
481         list_for_each_entry(shrinker, &shrinker_list, list) {
482                 struct shrink_control sc = {
483                         .gfp_mask = gfp_mask,
484                         .nid = nid,
485                         .memcg = memcg,
486                 };
487 
488                 /*
489                  * If kernel memory accounting is disabled, we ignore
490                  * SHRINKER_MEMCG_AWARE flag and call all shrinkers
491                  * passing NULL for memcg.
492                  */
493                 if (memcg_kmem_enabled() &&
494                     !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
495                         continue;
496 
497                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
498                         sc.nid = 0;
499 
500                 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
501         }
502 
503         up_read(&shrinker_rwsem);
504 out:
505         cond_resched();
506         return freed;
507 }
508 
509 void drop_slab_node(int nid)
510 {
511         unsigned long freed;
512 
513         do {
514                 struct mem_cgroup *memcg = NULL;
515 
516                 freed = 0;
517                 do {
518                         freed += shrink_slab(GFP_KERNEL, nid, memcg,
519                                              1000, 1000);
520                 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
521         } while (freed > 10);
522 }
523 
524 void drop_slab(void)
525 {
526         int nid;
527 
528         for_each_online_node(nid)
529                 drop_slab_node(nid);
530 }
531 
532 static inline int is_page_cache_freeable(struct page *page)
533 {
534         /*
535          * A freeable page cache page is referenced only by the caller
536          * that isolated the page, the page cache radix tree and
537          * optional buffer heads at page->private.
538          */
539         return page_count(page) - page_has_private(page) == 2;
540 }
541 
542 static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
543 {
544         if (current->flags & PF_SWAPWRITE)
545                 return 1;
546         if (!inode_write_congested(inode))
547                 return 1;
548         if (inode_to_bdi(inode) == current->backing_dev_info)
549                 return 1;
550         return 0;
551 }
552 
553 /*
554  * We detected a synchronous write error writing a page out.  Probably
555  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
556  * fsync(), msync() or close().
557  *
558  * The tricky part is that after writepage we cannot touch the mapping: nothing
559  * prevents it from being freed up.  But we have a ref on the page and once
560  * that page is locked, the mapping is pinned.
561  *
562  * We're allowed to run sleeping lock_page() here because we know the caller has
563  * __GFP_FS.
564  */
565 static void handle_write_error(struct address_space *mapping,
566                                 struct page *page, int error)
567 {
568         lock_page(page);
569         if (page_mapping(page) == mapping)
570                 mapping_set_error(mapping, error);
571         unlock_page(page);
572 }
573 
574 /* possible outcome of pageout() */
575 typedef enum {
576         /* failed to write page out, page is locked */
577         PAGE_KEEP,
578         /* move page to the active list, page is locked */
579         PAGE_ACTIVATE,
580         /* page has been sent to the disk successfully, page is unlocked */
581         PAGE_SUCCESS,
582         /* page is clean and locked */
583         PAGE_CLEAN,
584 } pageout_t;
585 
586 /*
587  * pageout is called by shrink_page_list() for each dirty page.
588  * Calls ->writepage().
589  */
590 static pageout_t pageout(struct page *page, struct address_space *mapping,
591                          struct scan_control *sc)
592 {
593         /*
594          * If the page is dirty, only perform writeback if that write
595          * will be non-blocking.  To prevent this allocation from being
596          * stalled by pagecache activity.  But note that there may be
597          * stalls if we need to run get_block().  We could test
598          * PagePrivate for that.
599          *
600          * If this process is currently in __generic_file_write_iter() against
601          * this page's queue, we can perform writeback even if that
602          * will block.
603          *
604          * If the page is swapcache, write it back even if that would
605          * block, for some throttling. This happens by accident, because
606          * swap_backing_dev_info is bust: it doesn't reflect the
607          * congestion state of the swapdevs.  Easy to fix, if needed.
608          */
609         if (!is_page_cache_freeable(page))
610                 return PAGE_KEEP;
611         if (!mapping) {
612                 /*
613                  * Some data journaling orphaned pages can have
614                  * page->mapping == NULL while being dirty with clean buffers.
615                  */
616                 if (page_has_private(page)) {
617                         if (try_to_free_buffers(page)) {
618                                 ClearPageDirty(page);
619                                 pr_info("%s: orphaned page\n", __func__);
620                                 return PAGE_CLEAN;
621                         }
622                 }
623                 return PAGE_KEEP;
624         }
625         if (mapping->a_ops->writepage == NULL)
626                 return PAGE_ACTIVATE;
627         if (!may_write_to_inode(mapping->host, sc))
628                 return PAGE_KEEP;
629 
630         if (clear_page_dirty_for_io(page)) {
631                 int res;
632                 struct writeback_control wbc = {
633                         .sync_mode = WB_SYNC_NONE,
634                         .nr_to_write = SWAP_CLUSTER_MAX,
635                         .range_start = 0,
636                         .range_end = LLONG_MAX,
637                         .for_reclaim = 1,
638                 };
639 
640                 SetPageReclaim(page);
641                 res = mapping->a_ops->writepage(page, &wbc);
642                 if (res < 0)
643                         handle_write_error(mapping, page, res);
644                 if (res == AOP_WRITEPAGE_ACTIVATE) {
645                         ClearPageReclaim(page);
646                         return PAGE_ACTIVATE;
647                 }
648 
649                 if (!PageWriteback(page)) {
650                         /* synchronous write or broken a_ops? */
651                         ClearPageReclaim(page);
652                 }
653                 trace_mm_vmscan_writepage(page);
654                 inc_node_page_state(page, NR_VMSCAN_WRITE);
655                 return PAGE_SUCCESS;
656         }
657 
658         return PAGE_CLEAN;
659 }
660 
661 /*
662  * Same as remove_mapping, but if the page is removed from the mapping, it
663  * gets returned with a refcount of 0.
664  */
665 static int __remove_mapping(struct address_space *mapping, struct page *page,
666                             bool reclaimed)
667 {
668         unsigned long flags;
669 
670         BUG_ON(!PageLocked(page));
671         BUG_ON(mapping != page_mapping(page));
672 
673         spin_lock_irqsave(&mapping->tree_lock, flags);
674         /*
675          * The non racy check for a busy page.
676          *
677          * Must be careful with the order of the tests. When someone has
678          * a ref to the page, it may be possible that they dirty it then
679          * drop the reference. So if PageDirty is tested before page_count
680          * here, then the following race may occur:
681          *
682          * get_user_pages(&page);
683          * [user mapping goes away]
684          * write_to(page);
685          *                              !PageDirty(page)    [good]
686          * SetPageDirty(page);
687          * put_page(page);
688          *                              !page_count(page)   [good, discard it]
689          *
690          * [oops, our write_to data is lost]
691          *
692          * Reversing the order of the tests ensures such a situation cannot
693          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
694          * load is not satisfied before that of page->_refcount.
695          *
696          * Note that if SetPageDirty is always performed via set_page_dirty,
697          * and thus under tree_lock, then this ordering is not required.
698          */
699         if (!page_ref_freeze(page, 2))
700                 goto cannot_free;
701         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
702         if (unlikely(PageDirty(page))) {
703                 page_ref_unfreeze(page, 2);
704                 goto cannot_free;
705         }
706 
707         if (PageSwapCache(page)) {
708                 swp_entry_t swap = { .val = page_private(page) };
709                 mem_cgroup_swapout(page, swap);
710                 __delete_from_swap_cache(page);
711                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
712                 swapcache_free(swap);
713         } else {
714                 void (*freepage)(struct page *);
715                 void *shadow = NULL;
716 
717                 freepage = mapping->a_ops->freepage;
718                 /*
719                  * Remember a shadow entry for reclaimed file cache in
720                  * order to detect refaults, thus thrashing, later on.
721                  *
722                  * But don't store shadows in an address space that is
723                  * already exiting.  This is not just an optizimation,
724                  * inode reclaim needs to empty out the radix tree or
725                  * the nodes are lost.  Don't plant shadows behind its
726                  * back.
727                  *
728                  * We also don't store shadows for DAX mappings because the
729                  * only page cache pages found in these are zero pages
730                  * covering holes, and because we don't want to mix DAX
731                  * exceptional entries and shadow exceptional entries in the
732                  * same page_tree.
733                  */
734                 if (reclaimed && page_is_file_cache(page) &&
735                     !mapping_exiting(mapping) && !dax_mapping(mapping))
736                         shadow = workingset_eviction(mapping, page);
737                 __delete_from_page_cache(page, shadow);
738                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
739 
740                 if (freepage != NULL)
741                         freepage(page);
742         }
743 
744         return 1;
745 
746 cannot_free:
747         spin_unlock_irqrestore(&mapping->tree_lock, flags);
748         return 0;
749 }
750 
751 /*
752  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
753  * someone else has a ref on the page, abort and return 0.  If it was
754  * successfully detached, return 1.  Assumes the caller has a single ref on
755  * this page.
756  */
757 int remove_mapping(struct address_space *mapping, struct page *page)
758 {
759         if (__remove_mapping(mapping, page, false)) {
760                 /*
761                  * Unfreezing the refcount with 1 rather than 2 effectively
762                  * drops the pagecache ref for us without requiring another
763                  * atomic operation.
764                  */
765                 page_ref_unfreeze(page, 1);
766                 return 1;
767         }
768         return 0;
769 }
770 
771 /**
772  * putback_lru_page - put previously isolated page onto appropriate LRU list
773  * @page: page to be put back to appropriate lru list
774  *
775  * Add previously isolated @page to appropriate LRU list.
776  * Page may still be unevictable for other reasons.
777  *
778  * lru_lock must not be held, interrupts must be enabled.
779  */
780 void putback_lru_page(struct page *page)
781 {
782         bool is_unevictable;
783         int was_unevictable = PageUnevictable(page);
784 
785         VM_BUG_ON_PAGE(PageLRU(page), page);
786 
787 redo:
788         ClearPageUnevictable(page);
789 
790         if (page_evictable(page)) {
791                 /*
792                  * For evictable pages, we can use the cache.
793                  * In event of a race, worst case is we end up with an
794                  * unevictable page on [in]active list.
795                  * We know how to handle that.
796                  */
797                 is_unevictable = false;
798                 lru_cache_add(page);
799         } else {
800                 /*
801                  * Put unevictable pages directly on zone's unevictable
802                  * list.
803                  */
804                 is_unevictable = true;
805                 add_page_to_unevictable_list(page);
806                 /*
807                  * When racing with an mlock or AS_UNEVICTABLE clearing
808                  * (page is unlocked) make sure that if the other thread
809                  * does not observe our setting of PG_lru and fails
810                  * isolation/check_move_unevictable_pages,
811                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
812                  * the page back to the evictable list.
813                  *
814                  * The other side is TestClearPageMlocked() or shmem_lock().
815                  */
816                 smp_mb();
817         }
818 
819         /*
820          * page's status can change while we move it among lru. If an evictable
821          * page is on unevictable list, it never be freed. To avoid that,
822          * check after we added it to the list, again.
823          */
824         if (is_unevictable && page_evictable(page)) {
825                 if (!isolate_lru_page(page)) {
826                         put_page(page);
827                         goto redo;
828                 }
829                 /* This means someone else dropped this page from LRU
830                  * So, it will be freed or putback to LRU again. There is
831                  * nothing to do here.
832                  */
833         }
834 
835         if (was_unevictable && !is_unevictable)
836                 count_vm_event(UNEVICTABLE_PGRESCUED);
837         else if (!was_unevictable && is_unevictable)
838                 count_vm_event(UNEVICTABLE_PGCULLED);
839 
840         put_page(page);         /* drop ref from isolate */
841 }
842 
843 enum page_references {
844         PAGEREF_RECLAIM,
845         PAGEREF_RECLAIM_CLEAN,
846         PAGEREF_KEEP,
847         PAGEREF_ACTIVATE,
848 };
849 
850 static enum page_references page_check_references(struct page *page,
851                                                   struct scan_control *sc)
852 {
853         int referenced_ptes, referenced_page;
854         unsigned long vm_flags;
855 
856         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
857                                           &vm_flags);
858         referenced_page = TestClearPageReferenced(page);
859 
860         /*
861          * Mlock lost the isolation race with us.  Let try_to_unmap()
862          * move the page to the unevictable list.
863          */
864         if (vm_flags & VM_LOCKED)
865                 return PAGEREF_RECLAIM;
866 
867         if (referenced_ptes) {
868                 if (PageSwapBacked(page))
869                         return PAGEREF_ACTIVATE;
870                 /*
871                  * All mapped pages start out with page table
872                  * references from the instantiating fault, so we need
873                  * to look twice if a mapped file page is used more
874                  * than once.
875                  *
876                  * Mark it and spare it for another trip around the
877                  * inactive list.  Another page table reference will
878                  * lead to its activation.
879                  *
880                  * Note: the mark is set for activated pages as well
881                  * so that recently deactivated but used pages are
882                  * quickly recovered.
883                  */
884                 SetPageReferenced(page);
885 
886                 if (referenced_page || referenced_ptes > 1)
887                         return PAGEREF_ACTIVATE;
888 
889                 /*
890                  * Activate file-backed executable pages after first usage.
891                  */
892                 if (vm_flags & VM_EXEC)
893                         return PAGEREF_ACTIVATE;
894 
895                 return PAGEREF_KEEP;
896         }
897 
898         /* Reclaim if clean, defer dirty pages to writeback */
899         if (referenced_page && !PageSwapBacked(page))
900                 return PAGEREF_RECLAIM_CLEAN;
901 
902         return PAGEREF_RECLAIM;
903 }
904 
905 /* Check if a page is dirty or under writeback */
906 static void page_check_dirty_writeback(struct page *page,
907                                        bool *dirty, bool *writeback)
908 {
909         struct address_space *mapping;
910 
911         /*
912          * Anonymous pages are not handled by flushers and must be written
913          * from reclaim context. Do not stall reclaim based on them
914          */
915         if (!page_is_file_cache(page)) {
916                 *dirty = false;
917                 *writeback = false;
918                 return;
919         }
920 
921         /* By default assume that the page flags are accurate */
922         *dirty = PageDirty(page);
923         *writeback = PageWriteback(page);
924 
925         /* Verify dirty/writeback state if the filesystem supports it */
926         if (!page_has_private(page))
927                 return;
928 
929         mapping = page_mapping(page);
930         if (mapping && mapping->a_ops->is_dirty_writeback)
931                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
932 }
933 
934 struct reclaim_stat {
935         unsigned nr_dirty;
936         unsigned nr_unqueued_dirty;
937         unsigned nr_congested;
938         unsigned nr_writeback;
939         unsigned nr_immediate;
940         unsigned nr_activate;
941         unsigned nr_ref_keep;
942         unsigned nr_unmap_fail;
943 };
944 
945 /*
946  * shrink_page_list() returns the number of reclaimed pages
947  */
948 static unsigned long shrink_page_list(struct list_head *page_list,
949                                       struct pglist_data *pgdat,
950                                       struct scan_control *sc,
951                                       enum ttu_flags ttu_flags,
952                                       struct reclaim_stat *stat,
953                                       bool force_reclaim)
954 {
955         LIST_HEAD(ret_pages);
956         LIST_HEAD(free_pages);
957         int pgactivate = 0;
958         unsigned nr_unqueued_dirty = 0;
959         unsigned nr_dirty = 0;
960         unsigned nr_congested = 0;
961         unsigned nr_reclaimed = 0;
962         unsigned nr_writeback = 0;
963         unsigned nr_immediate = 0;
964         unsigned nr_ref_keep = 0;
965         unsigned nr_unmap_fail = 0;
966 
967         cond_resched();
968 
969         while (!list_empty(page_list)) {
970                 struct address_space *mapping;
971                 struct page *page;
972                 int may_enter_fs;
973                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
974                 bool dirty, writeback;
975                 bool lazyfree = false;
976                 int ret = SWAP_SUCCESS;
977 
978                 cond_resched();
979 
980                 page = lru_to_page(page_list);
981                 list_del(&page->lru);
982 
983                 if (!trylock_page(page))
984                         goto keep;
985 
986                 VM_BUG_ON_PAGE(PageActive(page), page);
987 
988                 sc->nr_scanned++;
989 
990                 if (unlikely(!page_evictable(page)))
991                         goto cull_mlocked;
992 
993                 if (!sc->may_unmap && page_mapped(page))
994                         goto keep_locked;
995 
996                 /* Double the slab pressure for mapped and swapcache pages */
997                 if (page_mapped(page) || PageSwapCache(page))
998                         sc->nr_scanned++;
999 
1000                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1001                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1002 
1003                 /*
1004                  * The number of dirty pages determines if a zone is marked
1005                  * reclaim_congested which affects wait_iff_congested. kswapd
1006                  * will stall and start writing pages if the tail of the LRU
1007                  * is all dirty unqueued pages.
1008                  */
1009                 page_check_dirty_writeback(page, &dirty, &writeback);
1010                 if (dirty || writeback)
1011                         nr_dirty++;
1012 
1013                 if (dirty && !writeback)
1014                         nr_unqueued_dirty++;
1015 
1016                 /*
1017                  * Treat this page as congested if the underlying BDI is or if
1018                  * pages are cycling through the LRU so quickly that the
1019                  * pages marked for immediate reclaim are making it to the
1020                  * end of the LRU a second time.
1021                  */
1022                 mapping = page_mapping(page);
1023                 if (((dirty || writeback) && mapping &&
1024                      inode_write_congested(mapping->host)) ||
1025                     (writeback && PageReclaim(page)))
1026                         nr_congested++;
1027 
1028                 /*
1029                  * If a page at the tail of the LRU is under writeback, there
1030                  * are three cases to consider.
1031                  *
1032                  * 1) If reclaim is encountering an excessive number of pages
1033                  *    under writeback and this page is both under writeback and
1034                  *    PageReclaim then it indicates that pages are being queued
1035                  *    for IO but are being recycled through the LRU before the
1036                  *    IO can complete. Waiting on the page itself risks an
1037                  *    indefinite stall if it is impossible to writeback the
1038                  *    page due to IO error or disconnected storage so instead
1039                  *    note that the LRU is being scanned too quickly and the
1040                  *    caller can stall after page list has been processed.
1041                  *
1042                  * 2) Global or new memcg reclaim encounters a page that is
1043                  *    not marked for immediate reclaim, or the caller does not
1044                  *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1045                  *    not to fs). In this case mark the page for immediate
1046                  *    reclaim and continue scanning.
1047                  *
1048                  *    Require may_enter_fs because we would wait on fs, which
1049                  *    may not have submitted IO yet. And the loop driver might
1050                  *    enter reclaim, and deadlock if it waits on a page for
1051                  *    which it is needed to do the write (loop masks off
1052                  *    __GFP_IO|__GFP_FS for this reason); but more thought
1053                  *    would probably show more reasons.
1054                  *
1055                  * 3) Legacy memcg encounters a page that is already marked
1056                  *    PageReclaim. memcg does not have any dirty pages
1057                  *    throttling so we could easily OOM just because too many
1058                  *    pages are in writeback and there is nothing else to
1059                  *    reclaim. Wait for the writeback to complete.
1060                  *
1061                  * In cases 1) and 2) we activate the pages to get them out of
1062                  * the way while we continue scanning for clean pages on the
1063                  * inactive list and refilling from the active list. The
1064                  * observation here is that waiting for disk writes is more
1065                  * expensive than potentially causing reloads down the line.
1066                  * Since they're marked for immediate reclaim, they won't put
1067                  * memory pressure on the cache working set any longer than it
1068                  * takes to write them to disk.
1069                  */
1070                 if (PageWriteback(page)) {
1071                         /* Case 1 above */
1072                         if (current_is_kswapd() &&
1073                             PageReclaim(page) &&
1074                             test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1075                                 nr_immediate++;
1076                                 goto activate_locked;
1077 
1078                         /* Case 2 above */
1079                         } else if (sane_reclaim(sc) ||
1080                             !PageReclaim(page) || !may_enter_fs) {
1081                                 /*
1082                                  * This is slightly racy - end_page_writeback()
1083                                  * might have just cleared PageReclaim, then
1084                                  * setting PageReclaim here end up interpreted
1085                                  * as PageReadahead - but that does not matter
1086                                  * enough to care.  What we do want is for this
1087                                  * page to have PageReclaim set next time memcg
1088                                  * reclaim reaches the tests above, so it will
1089                                  * then wait_on_page_writeback() to avoid OOM;
1090                                  * and it's also appropriate in global reclaim.
1091                                  */
1092                                 SetPageReclaim(page);
1093                                 nr_writeback++;
1094                                 goto activate_locked;
1095 
1096                         /* Case 3 above */
1097                         } else {
1098                                 unlock_page(page);
1099                                 wait_on_page_writeback(page);
1100                                 /* then go back and try same page again */
1101                                 list_add_tail(&page->lru, page_list);
1102                                 continue;
1103                         }
1104                 }
1105 
1106                 if (!force_reclaim)
1107                         references = page_check_references(page, sc);
1108 
1109                 switch (references) {
1110                 case PAGEREF_ACTIVATE:
1111                         goto activate_locked;
1112                 case PAGEREF_KEEP:
1113                         nr_ref_keep++;
1114                         goto keep_locked;
1115                 case PAGEREF_RECLAIM:
1116                 case PAGEREF_RECLAIM_CLEAN:
1117                         ; /* try to reclaim the page below */
1118                 }
1119 
1120                 /*
1121                  * Anonymous process memory has backing store?
1122                  * Try to allocate it some swap space here.
1123                  */
1124                 if (PageAnon(page) && !PageSwapCache(page)) {
1125                         if (!(sc->gfp_mask & __GFP_IO))
1126                                 goto keep_locked;
1127                         if (!add_to_swap(page, page_list))
1128                                 goto activate_locked;
1129                         lazyfree = true;
1130                         may_enter_fs = 1;
1131 
1132                         /* Adding to swap updated mapping */
1133                         mapping = page_mapping(page);
1134                 } else if (unlikely(PageTransHuge(page))) {
1135                         /* Split file THP */
1136                         if (split_huge_page_to_list(page, page_list))
1137                                 goto keep_locked;
1138                 }
1139 
1140                 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1141 
1142                 /*
1143                  * The page is mapped into the page tables of one or more
1144                  * processes. Try to unmap it here.
1145                  */
1146                 if (page_mapped(page) && mapping) {
1147                         switch (ret = try_to_unmap(page, lazyfree ?
1148                                 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1149                                 (ttu_flags | TTU_BATCH_FLUSH))) {
1150                         case SWAP_FAIL:
1151                                 nr_unmap_fail++;
1152                                 goto activate_locked;
1153                         case SWAP_AGAIN:
1154                                 goto keep_locked;
1155                         case SWAP_MLOCK:
1156                                 goto cull_mlocked;
1157                         case SWAP_LZFREE:
1158                                 goto lazyfree;
1159                         case SWAP_SUCCESS:
1160                                 ; /* try to free the page below */
1161                         }
1162                 }
1163 
1164                 if (PageDirty(page)) {
1165                         /*
1166                          * Only kswapd can writeback filesystem pages
1167                          * to avoid risk of stack overflow. But avoid
1168                          * injecting inefficient single-page IO into
1169                          * flusher writeback as much as possible: only
1170                          * write pages when we've encountered many
1171                          * dirty pages, and when we've already scanned
1172                          * the rest of the LRU for clean pages and see
1173                          * the same dirty pages again (PageReclaim).
1174                          */
1175                         if (page_is_file_cache(page) &&
1176                             (!current_is_kswapd() || !PageReclaim(page) ||
1177                              !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1178                                 /*
1179                                  * Immediately reclaim when written back.
1180                                  * Similar in principal to deactivate_page()
1181                                  * except we already have the page isolated
1182                                  * and know it's dirty
1183                                  */
1184                                 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1185                                 SetPageReclaim(page);
1186 
1187                                 goto activate_locked;
1188                         }
1189 
1190                         if (references == PAGEREF_RECLAIM_CLEAN)
1191                                 goto keep_locked;
1192                         if (!may_enter_fs)
1193                                 goto keep_locked;
1194                         if (!sc->may_writepage)
1195                                 goto keep_locked;
1196 
1197                         /*
1198                          * Page is dirty. Flush the TLB if a writable entry
1199                          * potentially exists to avoid CPU writes after IO
1200                          * starts and then write it out here.
1201                          */
1202                         try_to_unmap_flush_dirty();
1203                         switch (pageout(page, mapping, sc)) {
1204                         case PAGE_KEEP:
1205                                 goto keep_locked;
1206                         case PAGE_ACTIVATE:
1207                                 goto activate_locked;
1208                         case PAGE_SUCCESS:
1209                                 if (PageWriteback(page))
1210                                         goto keep;
1211                                 if (PageDirty(page))
1212                                         goto keep;
1213 
1214                                 /*
1215                                  * A synchronous write - probably a ramdisk.  Go
1216                                  * ahead and try to reclaim the page.
1217                                  */
1218                                 if (!trylock_page(page))
1219                                         goto keep;
1220                                 if (PageDirty(page) || PageWriteback(page))
1221                                         goto keep_locked;
1222                                 mapping = page_mapping(page);
1223                         case PAGE_CLEAN:
1224                                 ; /* try to free the page below */
1225                         }
1226                 }
1227 
1228                 /*
1229                  * If the page has buffers, try to free the buffer mappings
1230                  * associated with this page. If we succeed we try to free
1231                  * the page as well.
1232                  *
1233                  * We do this even if the page is PageDirty().
1234                  * try_to_release_page() does not perform I/O, but it is
1235                  * possible for a page to have PageDirty set, but it is actually
1236                  * clean (all its buffers are clean).  This happens if the
1237                  * buffers were written out directly, with submit_bh(). ext3
1238                  * will do this, as well as the blockdev mapping.
1239                  * try_to_release_page() will discover that cleanness and will
1240                  * drop the buffers and mark the page clean - it can be freed.
1241                  *
1242                  * Rarely, pages can have buffers and no ->mapping.  These are
1243                  * the pages which were not successfully invalidated in
1244                  * truncate_complete_page().  We try to drop those buffers here
1245                  * and if that worked, and the page is no longer mapped into
1246                  * process address space (page_count == 1) it can be freed.
1247                  * Otherwise, leave the page on the LRU so it is swappable.
1248                  */
1249                 if (page_has_private(page)) {
1250                         if (!try_to_release_page(page, sc->gfp_mask))
1251                                 goto activate_locked;
1252                         if (!mapping && page_count(page) == 1) {
1253                                 unlock_page(page);
1254                                 if (put_page_testzero(page))
1255                                         goto free_it;
1256                                 else {
1257                                         /*
1258                                          * rare race with speculative reference.
1259                                          * the speculative reference will free
1260                                          * this page shortly, so we may
1261                                          * increment nr_reclaimed here (and
1262                                          * leave it off the LRU).
1263                                          */
1264                                         nr_reclaimed++;
1265                                         continue;
1266                                 }
1267                         }
1268                 }
1269 
1270 lazyfree:
1271                 if (!mapping || !__remove_mapping(mapping, page, true))
1272                         goto keep_locked;
1273 
1274                 /*
1275                  * At this point, we have no other references and there is
1276                  * no way to pick any more up (removed from LRU, removed
1277                  * from pagecache). Can use non-atomic bitops now (and
1278                  * we obviously don't have to worry about waking up a process
1279                  * waiting on the page lock, because there are no references.
1280                  */
1281                 __ClearPageLocked(page);
1282 free_it:
1283                 if (ret == SWAP_LZFREE)
1284                         count_vm_event(PGLAZYFREED);
1285 
1286                 nr_reclaimed++;
1287 
1288                 /*
1289                  * Is there need to periodically free_page_list? It would
1290                  * appear not as the counts should be low
1291                  */
1292                 list_add(&page->lru, &free_pages);
1293                 continue;
1294 
1295 cull_mlocked:
1296                 if (PageSwapCache(page))
1297                         try_to_free_swap(page);
1298                 unlock_page(page);
1299                 list_add(&page->lru, &ret_pages);
1300                 continue;
1301 
1302 activate_locked:
1303                 /* Not a candidate for swapping, so reclaim swap space. */
1304                 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1305                         try_to_free_swap(page);
1306                 VM_BUG_ON_PAGE(PageActive(page), page);
1307                 SetPageActive(page);
1308                 pgactivate++;
1309 keep_locked:
1310                 unlock_page(page);
1311 keep:
1312                 list_add(&page->lru, &ret_pages);
1313                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1314         }
1315 
1316         mem_cgroup_uncharge_list(&free_pages);
1317         try_to_unmap_flush();
1318         free_hot_cold_page_list(&free_pages, true);
1319 
1320         list_splice(&ret_pages, page_list);
1321         count_vm_events(PGACTIVATE, pgactivate);
1322 
1323         if (stat) {
1324                 stat->nr_dirty = nr_dirty;
1325                 stat->nr_congested = nr_congested;
1326                 stat->nr_unqueued_dirty = nr_unqueued_dirty;
1327                 stat->nr_writeback = nr_writeback;
1328                 stat->nr_immediate = nr_immediate;
1329                 stat->nr_activate = pgactivate;
1330                 stat->nr_ref_keep = nr_ref_keep;
1331                 stat->nr_unmap_fail = nr_unmap_fail;
1332         }
1333         return nr_reclaimed;
1334 }
1335 
1336 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1337                                             struct list_head *page_list)
1338 {
1339         struct scan_control sc = {
1340                 .gfp_mask = GFP_KERNEL,
1341                 .priority = DEF_PRIORITY,
1342                 .may_unmap = 1,
1343         };
1344         unsigned long ret;
1345         struct page *page, *next;
1346         LIST_HEAD(clean_pages);
1347 
1348         list_for_each_entry_safe(page, next, page_list, lru) {
1349                 if (page_is_file_cache(page) && !PageDirty(page) &&
1350                     !__PageMovable(page)) {
1351                         ClearPageActive(page);
1352                         list_move(&page->lru, &clean_pages);
1353                 }
1354         }
1355 
1356         ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1357                         TTU_UNMAP|TTU_IGNORE_ACCESS, NULL, true);
1358         list_splice(&clean_pages, page_list);
1359         mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1360         return ret;
1361 }
1362 
1363 /*
1364  * Attempt to remove the specified page from its LRU.  Only take this page
1365  * if it is of the appropriate PageActive status.  Pages which are being
1366  * freed elsewhere are also ignored.
1367  *
1368  * page:        page to consider
1369  * mode:        one of the LRU isolation modes defined above
1370  *
1371  * returns 0 on success, -ve errno on failure.
1372  */
1373 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1374 {
1375         int ret = -EINVAL;
1376 
1377         /* Only take pages on the LRU. */
1378         if (!PageLRU(page))
1379                 return ret;
1380 
1381         /* Compaction should not handle unevictable pages but CMA can do so */
1382         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1383                 return ret;
1384 
1385         ret = -EBUSY;
1386 
1387         /*
1388          * To minimise LRU disruption, the caller can indicate that it only
1389          * wants to isolate pages it will be able to operate on without
1390          * blocking - clean pages for the most part.
1391          *
1392          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1393          * that it is possible to migrate without blocking
1394          */
1395         if (mode & ISOLATE_ASYNC_MIGRATE) {
1396                 /* All the caller can do on PageWriteback is block */
1397                 if (PageWriteback(page))
1398                         return ret;
1399 
1400                 if (PageDirty(page)) {
1401                         struct address_space *mapping;
1402 
1403                         /*
1404                          * Only pages without mappings or that have a
1405                          * ->migratepage callback are possible to migrate
1406                          * without blocking
1407                          */
1408                         mapping = page_mapping(page);
1409                         if (mapping && !mapping->a_ops->migratepage)
1410                                 return ret;
1411                 }
1412         }
1413 
1414         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1415                 return ret;
1416 
1417         if (likely(get_page_unless_zero(page))) {
1418                 /*
1419                  * Be careful not to clear PageLRU until after we're
1420                  * sure the page is not being freed elsewhere -- the
1421                  * page release code relies on it.
1422                  */
1423                 ClearPageLRU(page);
1424                 ret = 0;
1425         }
1426 
1427         return ret;
1428 }
1429 
1430 
1431 /*
1432  * Update LRU sizes after isolating pages. The LRU size updates must
1433  * be complete before mem_cgroup_update_lru_size due to a santity check.
1434  */
1435 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1436                         enum lru_list lru, unsigned long *nr_zone_taken)
1437 {
1438         int zid;
1439 
1440         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1441                 if (!nr_zone_taken[zid])
1442                         continue;
1443 
1444                 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1445 #ifdef CONFIG_MEMCG
1446                 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1447 #endif
1448         }
1449 
1450 }
1451 
1452 /*
1453  * zone_lru_lock is heavily contended.  Some of the functions that
1454  * shrink the lists perform better by taking out a batch of pages
1455  * and working on them outside the LRU lock.
1456  *
1457  * For pagecache intensive workloads, this function is the hottest
1458  * spot in the kernel (apart from copy_*_user functions).
1459  *
1460  * Appropriate locks must be held before calling this function.
1461  *
1462  * @nr_to_scan: The number of pages to look through on the list.
1463  * @lruvec:     The LRU vector to pull pages from.
1464  * @dst:        The temp list to put pages on to.
1465  * @nr_scanned: The number of pages that were scanned.
1466  * @sc:         The scan_control struct for this reclaim session
1467  * @mode:       One of the LRU isolation modes
1468  * @lru:        LRU list id for isolating
1469  *
1470  * returns how many pages were moved onto *@dst.
1471  */
1472 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1473                 struct lruvec *lruvec, struct list_head *dst,
1474                 unsigned long *nr_scanned, struct scan_control *sc,
1475                 isolate_mode_t mode, enum lru_list lru)
1476 {
1477         struct list_head *src = &lruvec->lists[lru];
1478         unsigned long nr_taken = 0;
1479         unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1480         unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1481         unsigned long skipped = 0, total_skipped = 0;
1482         unsigned long scan, nr_pages;
1483         LIST_HEAD(pages_skipped);
1484 
1485         for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1486                                         !list_empty(src);) {
1487                 struct page *page;
1488 
1489                 page = lru_to_page(src);
1490                 prefetchw_prev_lru_page(page, src, flags);
1491 
1492                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1493 
1494                 if (page_zonenum(page) > sc->reclaim_idx) {
1495                         list_move(&page->lru, &pages_skipped);
1496                         nr_skipped[page_zonenum(page)]++;
1497                         continue;
1498                 }
1499 
1500                 /*
1501                  * Account for scanned and skipped separetly to avoid the pgdat
1502                  * being prematurely marked unreclaimable by pgdat_reclaimable.
1503                  */
1504                 scan++;
1505 
1506                 switch (__isolate_lru_page(page, mode)) {
1507                 case 0:
1508                         nr_pages = hpage_nr_pages(page);
1509                         nr_taken += nr_pages;
1510                         nr_zone_taken[page_zonenum(page)] += nr_pages;
1511                         list_move(&page->lru, dst);
1512                         break;
1513 
1514                 case -EBUSY:
1515                         /* else it is being freed elsewhere */
1516                         list_move(&page->lru, src);
1517                         continue;
1518 
1519                 default:
1520                         BUG();
1521                 }
1522         }
1523 
1524         /*
1525          * Splice any skipped pages to the start of the LRU list. Note that
1526          * this disrupts the LRU order when reclaiming for lower zones but
1527          * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1528          * scanning would soon rescan the same pages to skip and put the
1529          * system at risk of premature OOM.
1530          */
1531         if (!list_empty(&pages_skipped)) {
1532                 int zid;
1533 
1534                 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1535                         if (!nr_skipped[zid])
1536                                 continue;
1537 
1538                         __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1539                         skipped += nr_skipped[zid];
1540                 }
1541 
1542                 /*
1543                  * Account skipped pages as a partial scan as the pgdat may be
1544                  * close to unreclaimable. If the LRU list is empty, account
1545                  * skipped pages as a full scan.
1546                  */
1547                 total_skipped = list_empty(src) ? skipped : skipped >> 2;
1548 
1549                 list_splice(&pages_skipped, src);
1550         }
1551         *nr_scanned = scan + total_skipped;
1552         trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1553                                     scan, skipped, nr_taken, mode, lru);
1554         update_lru_sizes(lruvec, lru, nr_zone_taken);
1555         return nr_taken;
1556 }
1557 
1558 /**
1559  * isolate_lru_page - tries to isolate a page from its LRU list
1560  * @page: page to isolate from its LRU list
1561  *
1562  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1563  * vmstat statistic corresponding to whatever LRU list the page was on.
1564  *
1565  * Returns 0 if the page was removed from an LRU list.
1566  * Returns -EBUSY if the page was not on an LRU list.
1567  *
1568  * The returned page will have PageLRU() cleared.  If it was found on
1569  * the active list, it will have PageActive set.  If it was found on
1570  * the unevictable list, it will have the PageUnevictable bit set. That flag
1571  * may need to be cleared by the caller before letting the page go.
1572  *
1573  * The vmstat statistic corresponding to the list on which the page was
1574  * found will be decremented.
1575  *
1576  * Restrictions:
1577  * (1) Must be called with an elevated refcount on the page. This is a
1578  *     fundamentnal difference from isolate_lru_pages (which is called
1579  *     without a stable reference).
1580  * (2) the lru_lock must not be held.
1581  * (3) interrupts must be enabled.
1582  */
1583 int isolate_lru_page(struct page *page)
1584 {
1585         int ret = -EBUSY;
1586 
1587         VM_BUG_ON_PAGE(!page_count(page), page);
1588         WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1589 
1590         if (PageLRU(page)) {
1591                 struct zone *zone = page_zone(page);
1592                 struct lruvec *lruvec;
1593 
1594                 spin_lock_irq(zone_lru_lock(zone));
1595                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1596                 if (PageLRU(page)) {
1597                         int lru = page_lru(page);
1598                         get_page(page);
1599                         ClearPageLRU(page);
1600                         del_page_from_lru_list(page, lruvec, lru);
1601                         ret = 0;
1602                 }
1603                 spin_unlock_irq(zone_lru_lock(zone));
1604         }
1605         return ret;
1606 }
1607 
1608 /*
1609  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1610  * then get resheduled. When there are massive number of tasks doing page
1611  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1612  * the LRU list will go small and be scanned faster than necessary, leading to
1613  * unnecessary swapping, thrashing and OOM.
1614  */
1615 static int too_many_isolated(struct pglist_data *pgdat, int file,
1616                 struct scan_control *sc)
1617 {
1618         unsigned long inactive, isolated;
1619 
1620         if (current_is_kswapd())
1621                 return 0;
1622 
1623         if (!sane_reclaim(sc))
1624                 return 0;
1625 
1626         if (file) {
1627                 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1628                 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1629         } else {
1630                 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1631                 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1632         }
1633 
1634         /*
1635          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1636          * won't get blocked by normal direct-reclaimers, forming a circular
1637          * deadlock.
1638          */
1639         if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1640                 inactive >>= 3;
1641 
1642         return isolated > inactive;
1643 }
1644 
1645 static noinline_for_stack void
1646 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1647 {
1648         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1649         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1650         LIST_HEAD(pages_to_free);
1651 
1652         /*
1653          * Put back any unfreeable pages.
1654          */
1655         while (!list_empty(page_list)) {
1656                 struct page *page = lru_to_page(page_list);
1657                 int lru;
1658 
1659                 VM_BUG_ON_PAGE(PageLRU(page), page);
1660                 list_del(&page->lru);
1661                 if (unlikely(!page_evictable(page))) {
1662                         spin_unlock_irq(&pgdat->lru_lock);
1663                         putback_lru_page(page);
1664                         spin_lock_irq(&pgdat->lru_lock);
1665                         continue;
1666                 }
1667 
1668                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1669 
1670                 SetPageLRU(page);
1671                 lru = page_lru(page);
1672                 add_page_to_lru_list(page, lruvec, lru);
1673 
1674                 if (is_active_lru(lru)) {
1675                         int file = is_file_lru(lru);
1676                         int numpages = hpage_nr_pages(page);
1677                         reclaim_stat->recent_rotated[file] += numpages;
1678                 }
1679                 if (put_page_testzero(page)) {
1680                         __ClearPageLRU(page);
1681                         __ClearPageActive(page);
1682                         del_page_from_lru_list(page, lruvec, lru);
1683 
1684                         if (unlikely(PageCompound(page))) {
1685                                 spin_unlock_irq(&pgdat->lru_lock);
1686                                 mem_cgroup_uncharge(page);
1687                                 (*get_compound_page_dtor(page))(page);
1688                                 spin_lock_irq(&pgdat->lru_lock);
1689                         } else
1690                                 list_add(&page->lru, &pages_to_free);
1691                 }
1692         }
1693 
1694         /*
1695          * To save our caller's stack, now use input list for pages to free.
1696          */
1697         list_splice(&pages_to_free, page_list);
1698 }
1699 
1700 /*
1701  * If a kernel thread (such as nfsd for loop-back mounts) services
1702  * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1703  * In that case we should only throttle if the backing device it is
1704  * writing to is congested.  In other cases it is safe to throttle.
1705  */
1706 static int current_may_throttle(void)
1707 {
1708         return !(current->flags & PF_LESS_THROTTLE) ||
1709                 current->backing_dev_info == NULL ||
1710                 bdi_write_congested(current->backing_dev_info);
1711 }
1712 
1713 /*
1714  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1715  * of reclaimed pages
1716  */
1717 static noinline_for_stack unsigned long
1718 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1719                      struct scan_control *sc, enum lru_list lru)
1720 {
1721         LIST_HEAD(page_list);
1722         unsigned long nr_scanned;
1723         unsigned long nr_reclaimed = 0;
1724         unsigned long nr_taken;
1725         struct reclaim_stat stat = {};
1726         isolate_mode_t isolate_mode = 0;
1727         int file = is_file_lru(lru);
1728         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1729         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1730 
1731         while (unlikely(too_many_isolated(pgdat, file, sc))) {
1732                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1733 
1734                 /* We are about to die and free our memory. Return now. */
1735                 if (fatal_signal_pending(current))
1736                         return SWAP_CLUSTER_MAX;
1737         }
1738 
1739         lru_add_drain();
1740 
1741         if (!sc->may_unmap)
1742                 isolate_mode |= ISOLATE_UNMAPPED;
1743 
1744         spin_lock_irq(&pgdat->lru_lock);
1745 
1746         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1747                                      &nr_scanned, sc, isolate_mode, lru);
1748 
1749         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1750         reclaim_stat->recent_scanned[file] += nr_taken;
1751 
1752         if (global_reclaim(sc)) {
1753                 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1754                 if (current_is_kswapd())
1755                         __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1756                 else
1757                         __count_vm_events(PGSCAN_DIRECT, nr_scanned);
1758         }
1759         spin_unlock_irq(&pgdat->lru_lock);
1760 
1761         if (nr_taken == 0)
1762                 return 0;
1763 
1764         nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
1765                                 &stat, false);
1766 
1767         spin_lock_irq(&pgdat->lru_lock);
1768 
1769         if (global_reclaim(sc)) {
1770                 if (current_is_kswapd())
1771                         __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1772                 else
1773                         __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1774         }
1775 
1776         putback_inactive_pages(lruvec, &page_list);
1777 
1778         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1779 
1780         spin_unlock_irq(&pgdat->lru_lock);
1781 
1782         mem_cgroup_uncharge_list(&page_list);
1783         free_hot_cold_page_list(&page_list, true);
1784 
1785         /*
1786          * If reclaim is isolating dirty pages under writeback, it implies
1787          * that the long-lived page allocation rate is exceeding the page
1788          * laundering rate. Either the global limits are not being effective
1789          * at throttling processes due to the page distribution throughout
1790          * zones or there is heavy usage of a slow backing device. The
1791          * only option is to throttle from reclaim context which is not ideal
1792          * as there is no guarantee the dirtying process is throttled in the
1793          * same way balance_dirty_pages() manages.
1794          *
1795          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1796          * of pages under pages flagged for immediate reclaim and stall if any
1797          * are encountered in the nr_immediate check below.
1798          */
1799         if (stat.nr_writeback && stat.nr_writeback == nr_taken)
1800                 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
1801 
1802         /*
1803          * Legacy memcg will stall in page writeback so avoid forcibly
1804          * stalling here.
1805          */
1806         if (sane_reclaim(sc)) {
1807                 /*
1808                  * Tag a zone as congested if all the dirty pages scanned were
1809                  * backed by a congested BDI and wait_iff_congested will stall.
1810                  */
1811                 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested)
1812                         set_bit(PGDAT_CONGESTED, &pgdat->flags);
1813 
1814                 /*
1815                  * If dirty pages are scanned that are not queued for IO, it
1816                  * implies that flushers are not doing their job. This can
1817                  * happen when memory pressure pushes dirty pages to the end of
1818                  * the LRU before the dirty limits are breached and the dirty
1819                  * data has expired. It can also happen when the proportion of
1820                  * dirty pages grows not through writes but through memory
1821                  * pressure reclaiming all the clean cache. And in some cases,
1822                  * the flushers simply cannot keep up with the allocation
1823                  * rate. Nudge the flusher threads in case they are asleep, but
1824                  * also allow kswapd to start writing pages during reclaim.
1825                  */
1826                 if (stat.nr_unqueued_dirty == nr_taken) {
1827                         wakeup_flusher_threads(0, WB_REASON_VMSCAN);
1828                         set_bit(PGDAT_DIRTY, &pgdat->flags);
1829                 }
1830 
1831                 /*
1832                  * If kswapd scans pages marked marked for immediate
1833                  * reclaim and under writeback (nr_immediate), it implies
1834                  * that pages are cycling through the LRU faster than
1835                  * they are written so also forcibly stall.
1836                  */
1837                 if (stat.nr_immediate && current_may_throttle())
1838                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1839         }
1840 
1841         /*
1842          * Stall direct reclaim for IO completions if underlying BDIs or zone
1843          * is congested. Allow kswapd to continue until it starts encountering
1844          * unqueued dirty pages or cycling through the LRU too quickly.
1845          */
1846         if (!sc->hibernation_mode && !current_is_kswapd() &&
1847             current_may_throttle())
1848                 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
1849 
1850         trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1851                         nr_scanned, nr_reclaimed,
1852                         stat.nr_dirty,  stat.nr_writeback,
1853                         stat.nr_congested, stat.nr_immediate,
1854                         stat.nr_activate, stat.nr_ref_keep,
1855                         stat.nr_unmap_fail,
1856                         sc->priority, file);
1857         return nr_reclaimed;
1858 }
1859 
1860 /*
1861  * This moves pages from the active list to the inactive list.
1862  *
1863  * We move them the other way if the page is referenced by one or more
1864  * processes, from rmap.
1865  *
1866  * If the pages are mostly unmapped, the processing is fast and it is
1867  * appropriate to hold zone_lru_lock across the whole operation.  But if
1868  * the pages are mapped, the processing is slow (page_referenced()) so we
1869  * should drop zone_lru_lock around each page.  It's impossible to balance
1870  * this, so instead we remove the pages from the LRU while processing them.
1871  * It is safe to rely on PG_active against the non-LRU pages in here because
1872  * nobody will play with that bit on a non-LRU page.
1873  *
1874  * The downside is that we have to touch page->_refcount against each page.
1875  * But we had to alter page->flags anyway.
1876  *
1877  * Returns the number of pages moved to the given lru.
1878  */
1879 
1880 static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1881                                      struct list_head *list,
1882                                      struct list_head *pages_to_free,
1883                                      enum lru_list lru)
1884 {
1885         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1886         struct page *page;
1887         int nr_pages;
1888         int nr_moved = 0;
1889 
1890         while (!list_empty(list)) {
1891                 page = lru_to_page(list);
1892                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1893 
1894                 VM_BUG_ON_PAGE(PageLRU(page), page);
1895                 SetPageLRU(page);
1896 
1897                 nr_pages = hpage_nr_pages(page);
1898                 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1899                 list_move(&page->lru, &lruvec->lists[lru]);
1900 
1901                 if (put_page_testzero(page)) {
1902                         __ClearPageLRU(page);
1903                         __ClearPageActive(page);
1904                         del_page_from_lru_list(page, lruvec, lru);
1905 
1906                         if (unlikely(PageCompound(page))) {
1907                                 spin_unlock_irq(&pgdat->lru_lock);
1908                                 mem_cgroup_uncharge(page);
1909                                 (*get_compound_page_dtor(page))(page);
1910                                 spin_lock_irq(&pgdat->lru_lock);
1911                         } else
1912                                 list_add(&page->lru, pages_to_free);
1913                 } else {
1914                         nr_moved += nr_pages;
1915                 }
1916         }
1917 
1918         if (!is_active_lru(lru))
1919                 __count_vm_events(PGDEACTIVATE, nr_moved);
1920 
1921         return nr_moved;
1922 }
1923 
1924 static void shrink_active_list(unsigned long nr_to_scan,
1925                                struct lruvec *lruvec,
1926                                struct scan_control *sc,
1927                                enum lru_list lru)
1928 {
1929         unsigned long nr_taken;
1930         unsigned long nr_scanned;
1931         unsigned long vm_flags;
1932         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1933         LIST_HEAD(l_active);
1934         LIST_HEAD(l_inactive);
1935         struct page *page;
1936         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1937         unsigned nr_deactivate, nr_activate;
1938         unsigned nr_rotated = 0;
1939         isolate_mode_t isolate_mode = 0;
1940         int file = is_file_lru(lru);
1941         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1942 
1943         lru_add_drain();
1944 
1945         if (!sc->may_unmap)
1946                 isolate_mode |= ISOLATE_UNMAPPED;
1947 
1948         spin_lock_irq(&pgdat->lru_lock);
1949 
1950         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1951                                      &nr_scanned, sc, isolate_mode, lru);
1952 
1953         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1954         reclaim_stat->recent_scanned[file] += nr_taken;
1955 
1956         if (global_reclaim(sc))
1957                 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1958         __count_vm_events(PGREFILL, nr_scanned);
1959 
1960         spin_unlock_irq(&pgdat->lru_lock);
1961 
1962         while (!list_empty(&l_hold)) {
1963                 cond_resched();
1964                 page = lru_to_page(&l_hold);
1965                 list_del(&page->lru);
1966 
1967                 if (unlikely(!page_evictable(page))) {
1968                         putback_lru_page(page);
1969                         continue;
1970                 }
1971 
1972                 if (unlikely(buffer_heads_over_limit)) {
1973                         if (page_has_private(page) && trylock_page(page)) {
1974                                 if (page_has_private(page))
1975                                         try_to_release_page(page, 0);
1976                                 unlock_page(page);
1977                         }
1978                 }
1979 
1980                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1981                                     &vm_flags)) {
1982                         nr_rotated += hpage_nr_pages(page);
1983                         /*
1984                          * Identify referenced, file-backed active pages and
1985                          * give them one more trip around the active list. So
1986                          * that executable code get better chances to stay in
1987                          * memory under moderate memory pressure.  Anon pages
1988                          * are not likely to be evicted by use-once streaming
1989                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1990                          * so we ignore them here.
1991                          */
1992                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1993                                 list_add(&page->lru, &l_active);
1994                                 continue;
1995                         }
1996                 }
1997 
1998                 ClearPageActive(page);  /* we are de-activating */
1999                 list_add(&page->lru, &l_inactive);
2000         }
2001 
2002         /*
2003          * Move pages back to the lru list.
2004          */
2005         spin_lock_irq(&pgdat->lru_lock);
2006         /*
2007          * Count referenced pages from currently used mappings as rotated,
2008          * even though only some of them are actually re-activated.  This
2009          * helps balance scan pressure between file and anonymous pages in
2010          * get_scan_count.
2011          */
2012         reclaim_stat->recent_rotated[file] += nr_rotated;
2013 
2014         nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
2015         nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
2016         __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2017         spin_unlock_irq(&pgdat->lru_lock);
2018 
2019         mem_cgroup_uncharge_list(&l_hold);
2020         free_hot_cold_page_list(&l_hold, true);
2021         trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2022                         nr_deactivate, nr_rotated, sc->priority, file);
2023 }
2024 
2025 /*
2026  * The inactive anon list should be small enough that the VM never has
2027  * to do too much work.
2028  *
2029  * The inactive file list should be small enough to leave most memory
2030  * to the established workingset on the scan-resistant active list,
2031  * but large enough to avoid thrashing the aggregate readahead window.
2032  *
2033  * Both inactive lists should also be large enough that each inactive
2034  * page has a chance to be referenced again before it is reclaimed.
2035  *
2036  * If that fails and refaulting is observed, the inactive list grows.
2037  *
2038  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2039  * on this LRU, maintained by the pageout code. A zone->inactive_ratio
2040  * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2041  *
2042  * total     target    max
2043  * memory    ratio     inactive
2044  * -------------------------------------
2045  *   10MB       1         5MB
2046  *  100MB       1        50MB
2047  *    1GB       3       250MB
2048  *   10GB      10       0.9GB
2049  *  100GB      31         3GB
2050  *    1TB     101        10GB
2051  *   10TB     320        32GB
2052  */
2053 static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2054                                  struct mem_cgroup *memcg,
2055                                  struct scan_control *sc, bool actual_reclaim)
2056 {
2057         enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2058         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2059         enum lru_list inactive_lru = file * LRU_FILE;
2060         unsigned long inactive, active;
2061         unsigned long inactive_ratio;
2062         unsigned long refaults;
2063         unsigned long gb;
2064 
2065         /*
2066          * If we don't have swap space, anonymous page deactivation
2067          * is pointless.
2068          */
2069         if (!file && !total_swap_pages)
2070                 return false;
2071 
2072         inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2073         active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2074 
2075         if (memcg)
2076                 refaults = mem_cgroup_read_stat(memcg,
2077                                                 MEMCG_WORKINGSET_ACTIVATE);
2078         else
2079                 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2080 
2081         /*
2082          * When refaults are being observed, it means a new workingset
2083          * is being established. Disable active list protection to get
2084          * rid of the stale workingset quickly.
2085          */
2086         if (file && actual_reclaim && lruvec->refaults != refaults) {
2087                 inactive_ratio = 0;
2088         } else {
2089                 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2090                 if (gb)
2091                         inactive_ratio = int_sqrt(10 * gb);
2092                 else
2093                         inactive_ratio = 1;
2094         }
2095 
2096         if (actual_reclaim)
2097                 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2098                         lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2099                         lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2100                         inactive_ratio, file);
2101 
2102         return inactive * inactive_ratio < active;
2103 }
2104 
2105 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2106                                  struct lruvec *lruvec, struct mem_cgroup *memcg,
2107                                  struct scan_control *sc)
2108 {
2109         if (is_active_lru(lru)) {
2110                 if (inactive_list_is_low(lruvec, is_file_lru(lru),
2111                                          memcg, sc, true))
2112                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
2113                 return 0;
2114         }
2115 
2116         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2117 }
2118 
2119 enum scan_balance {
2120         SCAN_EQUAL,
2121         SCAN_FRACT,
2122         SCAN_ANON,
2123         SCAN_FILE,
2124 };
2125 
2126 /*
2127  * Determine how aggressively the anon and file LRU lists should be
2128  * scanned.  The relative value of each set of LRU lists is determined
2129  * by looking at the fraction of the pages scanned we did rotate back
2130  * onto the active list instead of evict.
2131  *
2132  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2133  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2134  */
2135 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2136                            struct scan_control *sc, unsigned long *nr,
2137                            unsigned long *lru_pages)
2138 {
2139         int swappiness = mem_cgroup_swappiness(memcg);
2140         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2141         u64 fraction[2];
2142         u64 denominator = 0;    /* gcc */
2143         struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2144         unsigned long anon_prio, file_prio;
2145         enum scan_balance scan_balance;
2146         unsigned long anon, file;
2147         bool force_scan = false;
2148         unsigned long ap, fp;
2149         enum lru_list lru;
2150         bool some_scanned;
2151         int pass;
2152 
2153         /*
2154          * If the zone or memcg is small, nr[l] can be 0.  This
2155          * results in no scanning on this priority and a potential
2156          * priority drop.  Global direct reclaim can go to the next
2157          * zone and tends to have no problems. Global kswapd is for
2158          * zone balancing and it needs to scan a minimum amount. When
2159          * reclaiming for a memcg, a priority drop can cause high
2160          * latencies, so it's better to scan a minimum amount there as
2161          * well.
2162          */
2163         if (current_is_kswapd()) {
2164                 if (!pgdat_reclaimable(pgdat))
2165                         force_scan = true;
2166                 if (!mem_cgroup_online(memcg))
2167                         force_scan = true;
2168         }
2169         if (!global_reclaim(sc))
2170                 force_scan = true;
2171 
2172         /* If we have no swap space, do not bother scanning anon pages. */
2173         if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2174                 scan_balance = SCAN_FILE;
2175                 goto out;
2176         }
2177 
2178         /*
2179          * Global reclaim will swap to prevent OOM even with no
2180          * swappiness, but memcg users want to use this knob to
2181          * disable swapping for individual groups completely when
2182          * using the memory controller's swap limit feature would be
2183          * too expensive.
2184          */
2185         if (!global_reclaim(sc) && !swappiness) {
2186                 scan_balance = SCAN_FILE;
2187                 goto out;
2188         }
2189 
2190         /*
2191          * Do not apply any pressure balancing cleverness when the
2192          * system is close to OOM, scan both anon and file equally
2193          * (unless the swappiness setting disagrees with swapping).
2194          */
2195         if (!sc->priority && swappiness) {
2196                 scan_balance = SCAN_EQUAL;
2197                 goto out;
2198         }
2199 
2200         /*
2201          * Prevent the reclaimer from falling into the cache trap: as
2202          * cache pages start out inactive, every cache fault will tip
2203          * the scan balance towards the file LRU.  And as the file LRU
2204          * shrinks, so does the window for rotation from references.
2205          * This means we have a runaway feedback loop where a tiny
2206          * thrashing file LRU becomes infinitely more attractive than
2207          * anon pages.  Try to detect this based on file LRU size.
2208          */
2209         if (global_reclaim(sc)) {
2210                 unsigned long pgdatfile;
2211                 unsigned long pgdatfree;
2212                 int z;
2213                 unsigned long total_high_wmark = 0;
2214 
2215                 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2216                 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2217                            node_page_state(pgdat, NR_INACTIVE_FILE);
2218 
2219                 for (z = 0; z < MAX_NR_ZONES; z++) {
2220                         struct zone *zone = &pgdat->node_zones[z];
2221                         if (!managed_zone(zone))
2222                                 continue;
2223 
2224                         total_high_wmark += high_wmark_pages(zone);
2225                 }
2226 
2227                 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2228                         scan_balance = SCAN_ANON;
2229                         goto out;
2230                 }
2231         }
2232 
2233         /*
2234          * If there is enough inactive page cache, i.e. if the size of the
2235          * inactive list is greater than that of the active list *and* the
2236          * inactive list actually has some pages to scan on this priority, we
2237          * do not reclaim anything from the anonymous working set right now.
2238          * Without the second condition we could end up never scanning an
2239          * lruvec even if it has plenty of old anonymous pages unless the
2240          * system is under heavy pressure.
2241          */
2242         if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2243             lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2244                 scan_balance = SCAN_FILE;
2245                 goto out;
2246         }
2247 
2248         scan_balance = SCAN_FRACT;
2249 
2250         /*
2251          * With swappiness at 100, anonymous and file have the same priority.
2252          * This scanning priority is essentially the inverse of IO cost.
2253          */
2254         anon_prio = swappiness;
2255         file_prio = 200 - anon_prio;
2256 
2257         /*
2258          * OK, so we have swap space and a fair amount of page cache
2259          * pages.  We use the recently rotated / recently scanned
2260          * ratios to determine how valuable each cache is.
2261          *
2262          * Because workloads change over time (and to avoid overflow)
2263          * we keep these statistics as a floating average, which ends
2264          * up weighing recent references more than old ones.
2265          *
2266          * anon in [0], file in [1]
2267          */
2268 
2269         anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2270                 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2271         file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2272                 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2273 
2274         spin_lock_irq(&pgdat->lru_lock);
2275         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2276                 reclaim_stat->recent_scanned[0] /= 2;
2277                 reclaim_stat->recent_rotated[0] /= 2;
2278         }
2279 
2280         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2281                 reclaim_stat->recent_scanned[1] /= 2;
2282                 reclaim_stat->recent_rotated[1] /= 2;
2283         }
2284 
2285         /*
2286          * The amount of pressure on anon vs file pages is inversely
2287          * proportional to the fraction of recently scanned pages on
2288          * each list that were recently referenced and in active use.
2289          */
2290         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2291         ap /= reclaim_stat->recent_rotated[0] + 1;
2292 
2293         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2294         fp /= reclaim_stat->recent_rotated[1] + 1;
2295         spin_unlock_irq(&pgdat->lru_lock);
2296 
2297         fraction[0] = ap;
2298         fraction[1] = fp;
2299         denominator = ap + fp + 1;
2300 out:
2301         some_scanned = false;
2302         /* Only use force_scan on second pass. */
2303         for (pass = 0; !some_scanned && pass < 2; pass++) {
2304                 *lru_pages = 0;
2305                 for_each_evictable_lru(lru) {
2306                         int file = is_file_lru(lru);
2307                         unsigned long size;
2308                         unsigned long scan;
2309 
2310                         size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2311                         scan = size >> sc->priority;
2312 
2313                         if (!scan && pass && force_scan)
2314                                 scan = min(size, SWAP_CLUSTER_MAX);
2315 
2316                         switch (scan_balance) {
2317                         case SCAN_EQUAL:
2318                                 /* Scan lists relative to size */
2319                                 break;
2320                         case SCAN_FRACT:
2321                                 /*
2322                                  * Scan types proportional to swappiness and
2323                                  * their relative recent reclaim efficiency.
2324                                  */
2325                                 scan = div64_u64(scan * fraction[file],
2326                                                         denominator);
2327                                 break;
2328                         case SCAN_FILE:
2329                         case SCAN_ANON:
2330                                 /* Scan one type exclusively */
2331                                 if ((scan_balance == SCAN_FILE) != file) {
2332                                         size = 0;
2333                                         scan = 0;
2334                                 }
2335                                 break;
2336                         default:
2337                                 /* Look ma, no brain */
2338                                 BUG();
2339                         }
2340 
2341                         *lru_pages += size;
2342                         nr[lru] = scan;
2343 
2344                         /*
2345                          * Skip the second pass and don't force_scan,
2346                          * if we found something to scan.
2347                          */
2348                         some_scanned |= !!scan;
2349                 }
2350         }
2351 }
2352 
2353 /*
2354  * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2355  */
2356 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2357                               struct scan_control *sc, unsigned long *lru_pages)
2358 {
2359         struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2360         unsigned long nr[NR_LRU_LISTS];
2361         unsigned long targets[NR_LRU_LISTS];
2362         unsigned long nr_to_scan;
2363         enum lru_list lru;
2364         unsigned long nr_reclaimed = 0;
2365         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2366         struct blk_plug plug;
2367         bool scan_adjusted;
2368 
2369         get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2370 
2371         /* Record the original scan target for proportional adjustments later */
2372         memcpy(targets, nr, sizeof(nr));
2373 
2374         /*
2375          * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2376          * event that can occur when there is little memory pressure e.g.
2377          * multiple streaming readers/writers. Hence, we do not abort scanning
2378          * when the requested number of pages are reclaimed when scanning at
2379          * DEF_PRIORITY on the assumption that the fact we are direct
2380          * reclaiming implies that kswapd is not keeping up and it is best to
2381          * do a batch of work at once. For memcg reclaim one check is made to
2382          * abort proportional reclaim if either the file or anon lru has already
2383          * dropped to zero at the first pass.
2384          */
2385         scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2386                          sc->priority == DEF_PRIORITY);
2387 
2388         blk_start_plug(&plug);
2389         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2390                                         nr[LRU_INACTIVE_FILE]) {
2391                 unsigned long nr_anon, nr_file, percentage;
2392                 unsigned long nr_scanned;
2393 
2394                 for_each_evictable_lru(lru) {
2395                         if (nr[lru]) {
2396                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2397                                 nr[lru] -= nr_to_scan;
2398 
2399                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2400                                                             lruvec, memcg, sc);
2401                         }
2402                 }
2403 
2404                 cond_resched();
2405 
2406                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2407                         continue;
2408 
2409                 /*
2410                  * For kswapd and memcg, reclaim at least the number of pages
2411                  * requested. Ensure that the anon and file LRUs are scanned
2412                  * proportionally what was requested by get_scan_count(). We
2413                  * stop reclaiming one LRU and reduce the amount scanning
2414                  * proportional to the original scan target.
2415                  */
2416                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2417                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2418 
2419                 /*
2420                  * It's just vindictive to attack the larger once the smaller
2421                  * has gone to zero.  And given the way we stop scanning the
2422                  * smaller below, this makes sure that we only make one nudge
2423                  * towards proportionality once we've got nr_to_reclaim.
2424                  */
2425                 if (!nr_file || !nr_anon)
2426                         break;
2427 
2428                 if (nr_file > nr_anon) {
2429                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2430                                                 targets[LRU_ACTIVE_ANON] + 1;
2431                         lru = LRU_BASE;
2432                         percentage = nr_anon * 100 / scan_target;
2433                 } else {
2434                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2435                                                 targets[LRU_ACTIVE_FILE] + 1;
2436                         lru = LRU_FILE;
2437                         percentage = nr_file * 100 / scan_target;
2438                 }
2439 
2440                 /* Stop scanning the smaller of the LRU */
2441                 nr[lru] = 0;
2442                 nr[lru + LRU_ACTIVE] = 0;
2443 
2444                 /*
2445                  * Recalculate the other LRU scan count based on its original
2446                  * scan target and the percentage scanning already complete
2447                  */
2448                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2449                 nr_scanned = targets[lru] - nr[lru];
2450                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2451                 nr[lru] -= min(nr[lru], nr_scanned);
2452 
2453                 lru += LRU_ACTIVE;
2454                 nr_scanned = targets[lru] - nr[lru];
2455                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2456                 nr[lru] -= min(nr[lru], nr_scanned);
2457 
2458                 scan_adjusted = true;
2459         }
2460         blk_finish_plug(&plug);
2461         sc->nr_reclaimed += nr_reclaimed;
2462 
2463         /*
2464          * Even if we did not try to evict anon pages at all, we want to
2465          * rebalance the anon lru active/inactive ratio.
2466          */
2467         if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2468                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2469                                    sc, LRU_ACTIVE_ANON);
2470 }
2471 
2472 /* Use reclaim/compaction for costly allocs or under memory pressure */
2473 static bool in_reclaim_compaction(struct scan_control *sc)
2474 {
2475         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2476                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2477                          sc->priority < DEF_PRIORITY - 2))
2478                 return true;
2479 
2480         return false;
2481 }
2482 
2483 /*
2484  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2485  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2486  * true if more pages should be reclaimed such that when the page allocator
2487  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2488  * It will give up earlier than that if there is difficulty reclaiming pages.
2489  */
2490 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2491                                         unsigned long nr_reclaimed,
2492                                         unsigned long nr_scanned,
2493                                         struct scan_control *sc)
2494 {
2495         unsigned long pages_for_compaction;
2496         unsigned long inactive_lru_pages;
2497         int z;
2498 
2499         /* If not in reclaim/compaction mode, stop */
2500         if (!in_reclaim_compaction(sc))
2501                 return false;
2502 
2503         /* Consider stopping depending on scan and reclaim activity */
2504         if (sc->gfp_mask & __GFP_REPEAT) {
2505                 /*
2506                  * For __GFP_REPEAT allocations, stop reclaiming if the
2507                  * full LRU list has been scanned and we are still failing
2508                  * to reclaim pages. This full LRU scan is potentially
2509                  * expensive but a __GFP_REPEAT caller really wants to succeed
2510                  */
2511                 if (!nr_reclaimed && !nr_scanned)
2512                         return false;
2513         } else {
2514                 /*
2515                  * For non-__GFP_REPEAT allocations which can presumably
2516                  * fail without consequence, stop if we failed to reclaim
2517                  * any pages from the last SWAP_CLUSTER_MAX number of
2518                  * pages that were scanned. This will return to the
2519                  * caller faster at the risk reclaim/compaction and
2520                  * the resulting allocation attempt fails
2521                  */
2522                 if (!nr_reclaimed)
2523                         return false;
2524         }
2525 
2526         /*
2527          * If we have not reclaimed enough pages for compaction and the
2528          * inactive lists are large enough, continue reclaiming
2529          */
2530         pages_for_compaction = compact_gap(sc->order);
2531         inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2532         if (get_nr_swap_pages() > 0)
2533                 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2534         if (sc->nr_reclaimed < pages_for_compaction &&
2535                         inactive_lru_pages > pages_for_compaction)
2536                 return true;
2537 
2538         /* If compaction would go ahead or the allocation would succeed, stop */
2539         for (z = 0; z <= sc->reclaim_idx; z++) {
2540                 struct zone *zone = &pgdat->node_zones[z];
2541                 if (!managed_zone(zone))
2542                         continue;
2543 
2544                 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2545                 case COMPACT_SUCCESS:
2546                 case COMPACT_CONTINUE:
2547                         return false;
2548                 default:
2549                         /* check next zone */
2550                         ;
2551                 }
2552         }
2553         return true;
2554 }
2555 
2556 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2557 {
2558         struct reclaim_state *reclaim_state = current->reclaim_state;
2559         unsigned long nr_reclaimed, nr_scanned;
2560         bool reclaimable = false;
2561 
2562         do {
2563                 struct mem_cgroup *root = sc->target_mem_cgroup;
2564                 struct mem_cgroup_reclaim_cookie reclaim = {
2565                         .pgdat = pgdat,
2566                         .priority = sc->priority,
2567                 };
2568                 unsigned long node_lru_pages = 0;
2569                 struct mem_cgroup *memcg;
2570 
2571                 nr_reclaimed = sc->nr_reclaimed;
2572                 nr_scanned = sc->nr_scanned;
2573 
2574                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2575                 do {
2576                         unsigned long lru_pages;
2577                         unsigned long reclaimed;
2578                         unsigned long scanned;
2579 
2580                         if (mem_cgroup_low(root, memcg)) {
2581                                 if (!sc->may_thrash)
2582                                         continue;
2583                                 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2584                         }
2585 
2586                         reclaimed = sc->nr_reclaimed;
2587                         scanned = sc->nr_scanned;
2588 
2589                         shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2590                         node_lru_pages += lru_pages;
2591 
2592                         if (memcg)
2593                                 shrink_slab(sc->gfp_mask, pgdat->node_id,
2594                                             memcg, sc->nr_scanned - scanned,
2595                                             lru_pages);
2596 
2597                         /* Record the group's reclaim efficiency */
2598                         vmpressure(sc->gfp_mask, memcg, false,
2599                                    sc->nr_scanned - scanned,
2600                                    sc->nr_reclaimed - reclaimed);
2601 
2602                         /*
2603                          * Direct reclaim and kswapd have to scan all memory
2604                          * cgroups to fulfill the overall scan target for the
2605                          * node.
2606                          *
2607                          * Limit reclaim, on the other hand, only cares about
2608                          * nr_to_reclaim pages to be reclaimed and it will
2609                          * retry with decreasing priority if one round over the
2610                          * whole hierarchy is not sufficient.
2611                          */
2612                         if (!global_reclaim(sc) &&
2613                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2614                                 mem_cgroup_iter_break(root, memcg);
2615                                 break;
2616                         }
2617                 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2618 
2619                 /*
2620                  * Shrink the slab caches in the same proportion that
2621                  * the eligible LRU pages were scanned.
2622                  */
2623                 if (global_reclaim(sc))
2624                         shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2625                                     sc->nr_scanned - nr_scanned,
2626                                     node_lru_pages);
2627 
2628                 if (reclaim_state) {
2629                         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2630                         reclaim_state->reclaimed_slab = 0;
2631                 }
2632 
2633                 /* Record the subtree's reclaim efficiency */
2634                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2635                            sc->nr_scanned - nr_scanned,
2636                            sc->nr_reclaimed - nr_reclaimed);
2637 
2638                 if (sc->nr_reclaimed - nr_reclaimed)
2639                         reclaimable = true;
2640 
2641         } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2642                                          sc->nr_scanned - nr_scanned, sc));
2643 
2644         return reclaimable;
2645 }
2646 
2647 /*
2648  * Returns true if compaction should go ahead for a costly-order request, or
2649  * the allocation would already succeed without compaction. Return false if we
2650  * should reclaim first.
2651  */
2652 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2653 {
2654         unsigned long watermark;
2655         enum compact_result suitable;
2656 
2657         suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2658         if (suitable == COMPACT_SUCCESS)
2659                 /* Allocation should succeed already. Don't reclaim. */
2660                 return true;
2661         if (suitable == COMPACT_SKIPPED)
2662                 /* Compaction cannot yet proceed. Do reclaim. */
2663                 return false;
2664 
2665         /*
2666          * Compaction is already possible, but it takes time to run and there
2667          * are potentially other callers using the pages just freed. So proceed
2668          * with reclaim to make a buffer of free pages available to give
2669          * compaction a reasonable chance of completing and allocating the page.
2670          * Note that we won't actually reclaim the whole buffer in one attempt
2671          * as the target watermark in should_continue_reclaim() is lower. But if
2672          * we are already above the high+gap watermark, don't reclaim at all.
2673          */
2674         watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2675 
2676         return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2677 }
2678 
2679 /*
2680  * This is the direct reclaim path, for page-allocating processes.  We only
2681  * try to reclaim pages from zones which will satisfy the caller's allocation
2682  * request.
2683  *
2684  * If a zone is deemed to be full of pinned pages then just give it a light
2685  * scan then give up on it.
2686  */
2687 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2688 {
2689         struct zoneref *z;
2690         struct zone *zone;
2691         unsigned long nr_soft_reclaimed;
2692         unsigned long nr_soft_scanned;
2693         gfp_t orig_mask;
2694         pg_data_t *last_pgdat = NULL;
2695 
2696         /*
2697          * If the number of buffer_heads in the machine exceeds the maximum
2698          * allowed level, force direct reclaim to scan the highmem zone as
2699          * highmem pages could be pinning lowmem pages storing buffer_heads
2700          */
2701         orig_mask = sc->gfp_mask;
2702         if (buffer_heads_over_limit) {
2703                 sc->gfp_mask |= __GFP_HIGHMEM;
2704                 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2705         }
2706 
2707         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2708                                         sc->reclaim_idx, sc->nodemask) {
2709                 /*
2710                  * Take care memory controller reclaiming has small influence
2711                  * to global LRU.
2712                  */
2713                 if (global_reclaim(sc)) {
2714                         if (!cpuset_zone_allowed(zone,
2715                                                  GFP_KERNEL | __GFP_HARDWALL))
2716                                 continue;
2717 
2718                         if (sc->priority != DEF_PRIORITY &&
2719                             !pgdat_reclaimable(zone->zone_pgdat))
2720                                 continue;       /* Let kswapd poll it */
2721 
2722                         /*
2723                          * If we already have plenty of memory free for
2724                          * compaction in this zone, don't free any more.
2725                          * Even though compaction is invoked for any
2726                          * non-zero order, only frequent costly order
2727                          * reclamation is disruptive enough to become a
2728                          * noticeable problem, like transparent huge
2729                          * page allocations.
2730                          */
2731                         if (IS_ENABLED(CONFIG_COMPACTION) &&
2732                             sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2733                             compaction_ready(zone, sc)) {
2734                                 sc->compaction_ready = true;
2735                                 continue;
2736                         }
2737 
2738                         /*
2739                          * Shrink each node in the zonelist once. If the
2740                          * zonelist is ordered by zone (not the default) then a
2741                          * node may be shrunk multiple times but in that case
2742                          * the user prefers lower zones being preserved.
2743                          */
2744                         if (zone->zone_pgdat == last_pgdat)
2745                                 continue;
2746 
2747                         /*
2748                          * This steals pages from memory cgroups over softlimit
2749                          * and returns the number of reclaimed pages and
2750                          * scanned pages. This works for global memory pressure
2751                          * and balancing, not for a memcg's limit.
2752                          */
2753                         nr_soft_scanned = 0;
2754                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2755                                                 sc->order, sc->gfp_mask,
2756                                                 &nr_soft_scanned);
2757                         sc->nr_reclaimed += nr_soft_reclaimed;
2758                         sc->nr_scanned += nr_soft_scanned;
2759                         /* need some check for avoid more shrink_zone() */
2760                 }
2761 
2762                 /* See comment about same check for global reclaim above */
2763                 if (zone->zone_pgdat == last_pgdat)
2764                         continue;
2765                 last_pgdat = zone->zone_pgdat;
2766                 shrink_node(zone->zone_pgdat, sc);
2767         }
2768 
2769         /*
2770          * Restore to original mask to avoid the impact on the caller if we
2771          * promoted it to __GFP_HIGHMEM.
2772          */
2773         sc->gfp_mask = orig_mask;
2774 }
2775 
2776 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2777 {
2778         struct mem_cgroup *memcg;
2779 
2780         memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2781         do {
2782                 unsigned long refaults;
2783                 struct lruvec *lruvec;
2784 
2785                 if (memcg)
2786                         refaults = mem_cgroup_read_stat(memcg,
2787                                                 MEMCG_WORKINGSET_ACTIVATE);
2788                 else
2789                         refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2790 
2791                 lruvec = mem_cgroup_lruvec(pgdat, memcg);
2792                 lruvec->refaults = refaults;
2793         } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2794 }
2795 
2796 /*
2797  * This is the main entry point to direct page reclaim.
2798  *
2799  * If a full scan of the inactive list fails to free enough memory then we
2800  * are "out of memory" and something needs to be killed.
2801  *
2802  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2803  * high - the zone may be full of dirty or under-writeback pages, which this
2804  * caller can't do much about.  We kick the writeback threads and take explicit
2805  * naps in the hope that some of these pages can be written.  But if the
2806  * allocating task holds filesystem locks which prevent writeout this might not
2807  * work, and the allocation attempt will fail.
2808  *
2809  * returns:     0, if no pages reclaimed
2810  *              else, the number of pages reclaimed
2811  */
2812 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2813                                           struct scan_control *sc)
2814 {
2815         int initial_priority = sc->priority;
2816         pg_data_t *last_pgdat;
2817         struct zoneref *z;
2818         struct zone *zone;
2819 retry:
2820         delayacct_freepages_start();
2821 
2822         if (global_reclaim(sc))
2823                 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2824 
2825         do {
2826                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2827                                 sc->priority);
2828                 sc->nr_scanned = 0;
2829                 shrink_zones(zonelist, sc);
2830 
2831                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2832                         break;
2833 
2834                 if (sc->compaction_ready)
2835                         break;
2836 
2837                 /*
2838                  * If we're getting trouble reclaiming, start doing
2839                  * writepage even in laptop mode.
2840                  */
2841                 if (sc->priority < DEF_PRIORITY - 2)
2842                         sc->may_writepage = 1;
2843         } while (--sc->priority >= 0);
2844 
2845         last_pgdat = NULL;
2846         for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2847                                         sc->nodemask) {
2848                 if (zone->zone_pgdat == last_pgdat)
2849                         continue;
2850                 last_pgdat = zone->zone_pgdat;
2851                 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2852         }
2853 
2854         delayacct_freepages_end();
2855 
2856         if (sc->nr_reclaimed)
2857                 return sc->nr_reclaimed;
2858 
2859         /* Aborted reclaim to try compaction? don't OOM, then */
2860         if (sc->compaction_ready)
2861                 return 1;
2862 
2863         /* Untapped cgroup reserves?  Don't OOM, retry. */
2864         if (!sc->may_thrash) {
2865                 sc->priority = initial_priority;
2866                 sc->may_thrash = 1;
2867                 goto retry;
2868         }
2869 
2870         return 0;
2871 }
2872 
2873 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2874 {
2875         struct zone *zone;
2876         unsigned long pfmemalloc_reserve = 0;
2877         unsigned long free_pages = 0;
2878         int i;
2879         bool wmark_ok;
2880 
2881         for (i = 0; i <= ZONE_NORMAL; i++) {
2882                 zone = &pgdat->node_zones[i];
2883                 if (!managed_zone(zone) ||
2884                     pgdat_reclaimable_pages(pgdat) == 0)
2885                         continue;
2886 
2887                 pfmemalloc_reserve += min_wmark_pages(zone);
2888                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2889         }
2890 
2891         /* If there are no reserves (unexpected config) then do not throttle */
2892         if (!pfmemalloc_reserve)
2893                 return true;
2894 
2895         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2896 
2897         /* kswapd must be awake if processes are being throttled */
2898         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2899                 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2900                                                 (enum zone_type)ZONE_NORMAL);
2901                 wake_up_interruptible(&pgdat->kswapd_wait);
2902         }
2903 
2904         return wmark_ok;
2905 }
2906 
2907 /*
2908  * Throttle direct reclaimers if backing storage is backed by the network
2909  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2910  * depleted. kswapd will continue to make progress and wake the processes
2911  * when the low watermark is reached.
2912  *
2913  * Returns true if a fatal signal was delivered during throttling. If this
2914  * happens, the page allocator should not consider triggering the OOM killer.
2915  */
2916 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2917                                         nodemask_t *nodemask)
2918 {
2919         struct zoneref *z;
2920         struct zone *zone;
2921         pg_data_t *pgdat = NULL;
2922 
2923         /*
2924          * Kernel threads should not be throttled as they may be indirectly
2925          * responsible for cleaning pages necessary for reclaim to make forward
2926          * progress. kjournald for example may enter direct reclaim while
2927          * committing a transaction where throttling it could forcing other
2928          * processes to block on log_wait_commit().
2929          */
2930         if (current->flags & PF_KTHREAD)
2931                 goto out;
2932 
2933         /*
2934          * If a fatal signal is pending, this process should not throttle.
2935          * It should return quickly so it can exit and free its memory
2936          */
2937         if (fatal_signal_pending(current))
2938                 goto out;
2939 
2940         /*
2941          * Check if the pfmemalloc reserves are ok by finding the first node
2942          * with a usable ZONE_NORMAL or lower zone. The expectation is that
2943          * GFP_KERNEL will be required for allocating network buffers when
2944          * swapping over the network so ZONE_HIGHMEM is unusable.
2945          *
2946          * Throttling is based on the first usable node and throttled processes
2947          * wait on a queue until kswapd makes progress and wakes them. There
2948          * is an affinity then between processes waking up and where reclaim
2949          * progress has been made assuming the process wakes on the same node.
2950          * More importantly, processes running on remote nodes will not compete
2951          * for remote pfmemalloc reserves and processes on different nodes
2952          * should make reasonable progress.
2953          */
2954         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2955                                         gfp_zone(gfp_mask), nodemask) {
2956                 if (zone_idx(zone) > ZONE_NORMAL)
2957                         continue;
2958 
2959                 /* Throttle based on the first usable node */
2960                 pgdat = zone->zone_pgdat;
2961                 if (pfmemalloc_watermark_ok(pgdat))
2962                         goto out;
2963                 break;
2964         }
2965 
2966         /* If no zone was usable by the allocation flags then do not throttle */
2967         if (!pgdat)
2968                 goto out;
2969 
2970         /* Account for the throttling */
2971         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2972 
2973         /*
2974          * If the caller cannot enter the filesystem, it's possible that it
2975          * is due to the caller holding an FS lock or performing a journal
2976          * transaction in the case of a filesystem like ext[3|4]. In this case,
2977          * it is not safe to block on pfmemalloc_wait as kswapd could be
2978          * blocked waiting on the same lock. Instead, throttle for up to a
2979          * second before continuing.
2980          */
2981         if (!(gfp_mask & __GFP_FS)) {
2982                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2983                         pfmemalloc_watermark_ok(pgdat), HZ);
2984 
2985                 goto check_pending;
2986         }
2987 
2988         /* Throttle until kswapd wakes the process */
2989         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2990                 pfmemalloc_watermark_ok(pgdat));
2991 
2992 check_pending:
2993         if (fatal_signal_pending(current))
2994                 return true;
2995 
2996 out:
2997         return false;
2998 }
2999 
3000 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3001                                 gfp_t gfp_mask, nodemask_t *nodemask)
3002 {
3003         unsigned long nr_reclaimed;
3004         struct scan_control sc = {
3005                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3006                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3007                 .reclaim_idx = gfp_zone(gfp_mask),
3008                 .order = order,
3009                 .nodemask = nodemask,
3010                 .priority = DEF_PRIORITY,
3011                 .may_writepage = !laptop_mode,
3012                 .may_unmap = 1,
3013                 .may_swap = 1,
3014         };
3015 
3016         /*
3017          * Do not enter reclaim if fatal signal was delivered while throttled.
3018          * 1 is returned so that the page allocator does not OOM kill at this
3019          * point.
3020          */
3021         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
3022                 return 1;
3023 
3024         trace_mm_vmscan_direct_reclaim_begin(order,
3025                                 sc.may_writepage,
3026                                 gfp_mask,
3027                                 sc.reclaim_idx);
3028 
3029         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3030 
3031         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3032 
3033         return nr_reclaimed;
3034 }
3035 
3036 #ifdef CONFIG_MEMCG
3037 
3038 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3039                                                 gfp_t gfp_mask, bool noswap,
3040                                                 pg_data_t *pgdat,
3041                                                 unsigned long *nr_scanned)
3042 {
3043         struct scan_control sc = {
3044                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3045                 .target_mem_cgroup = memcg,
3046                 .may_writepage = !laptop_mode,
3047                 .may_unmap = 1,
3048                 .reclaim_idx = MAX_NR_ZONES - 1,
3049                 .may_swap = !noswap,
3050         };
3051         unsigned long lru_pages;
3052 
3053         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3054                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3055 
3056         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3057                                                       sc.may_writepage,
3058                                                       sc.gfp_mask,
3059                                                       sc.reclaim_idx);
3060 
3061         /*
3062          * NOTE: Although we can get the priority field, using it
3063          * here is not a good idea, since it limits the pages we can scan.
3064          * if we don't reclaim here, the shrink_node from balance_pgdat
3065          * will pick up pages from other mem cgroup's as well. We hack
3066          * the priority and make it zero.
3067          */
3068         shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3069 
3070         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3071 
3072         *nr_scanned = sc.nr_scanned;
3073         return sc.nr_reclaimed;
3074 }
3075 
3076 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3077                                            unsigned long nr_pages,
3078                                            gfp_t gfp_mask,
3079                                            bool may_swap)
3080 {
3081         struct zonelist *zonelist;
3082         unsigned long nr_reclaimed;
3083         int nid;
3084         struct scan_control sc = {
3085                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3086                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3087                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3088                 .reclaim_idx = MAX_NR_ZONES - 1,
3089                 .target_mem_cgroup = memcg,
3090                 .priority = DEF_PRIORITY,
3091                 .may_writepage = !laptop_mode,
3092                 .may_unmap = 1,
3093                 .may_swap = may_swap,
3094         };
3095 
3096         /*
3097          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3098          * take care of from where we get pages. So the node where we start the
3099          * scan does not need to be the current node.
3100          */
3101         nid = mem_cgroup_select_victim_node(memcg);
3102 
3103         zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3104 
3105         trace_mm_vmscan_memcg_reclaim_begin(0,
3106                                             sc.may_writepage,
3107                                             sc.gfp_mask,
3108                                             sc.reclaim_idx);
3109 
3110         current->flags |= PF_MEMALLOC;
3111         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3112         current->flags &= ~PF_MEMALLOC;
3113 
3114         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3115 
3116         return nr_reclaimed;
3117 }
3118 #endif
3119 
3120 static void age_active_anon(struct pglist_data *pgdat,
3121                                 struct scan_control *sc)
3122 {
3123         struct mem_cgroup *memcg;
3124 
3125         if (!total_swap_pages)
3126                 return;
3127 
3128         memcg = mem_cgroup_iter(NULL, NULL, NULL);
3129         do {
3130                 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3131 
3132                 if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3133                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3134                                            sc, LRU_ACTIVE_ANON);
3135 
3136                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3137         } while (memcg);
3138 }
3139 
3140 static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
3141 {
3142         unsigned long mark = high_wmark_pages(zone);
3143 
3144         if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3145                 return false;
3146 
3147         /*
3148          * If any eligible zone is balanced then the node is not considered
3149          * to be congested or dirty
3150          */
3151         clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3152         clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3153         clear_bit(PGDAT_WRITEBACK, &zone->zone_pgdat->flags);
3154 
3155         return true;
3156 }
3157 
3158 /*
3159  * Prepare kswapd for sleeping. This verifies that there are no processes
3160  * waiting in throttle_direct_reclaim() and that watermarks have been met.
3161  *
3162  * Returns true if kswapd is ready to sleep
3163  */
3164 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3165 {
3166         int i;
3167 
3168         /*
3169          * The throttled processes are normally woken up in balance_pgdat() as
3170          * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3171          * race between when kswapd checks the watermarks and a process gets
3172          * throttled. There is also a potential race if processes get
3173          * throttled, kswapd wakes, a large process exits thereby balancing the
3174          * zones, which causes kswapd to exit balance_pgdat() before reaching
3175          * the wake up checks. If kswapd is going to sleep, no process should
3176          * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3177          * the wake up is premature, processes will wake kswapd and get
3178          * throttled again. The difference from wake ups in balance_pgdat() is
3179          * that here we are under prepare_to_wait().
3180          */
3181         if (waitqueue_active(&pgdat->pfmemalloc_wait))
3182                 wake_up_all(&pgdat->pfmemalloc_wait);
3183 
3184         for (i = 0; i <= classzone_idx; i++) {
3185                 struct zone *zone = pgdat->node_zones + i;
3186 
3187                 if (!managed_zone(zone))
3188                         continue;
3189 
3190                 if (!zone_balanced(zone, order, classzone_idx))
3191                         return false;
3192         }
3193 
3194         return true;
3195 }
3196 
3197 /*
3198  * kswapd shrinks a node of pages that are at or below the highest usable
3199  * zone that is currently unbalanced.
3200  *
3201  * Returns true if kswapd scanned at least the requested number of pages to
3202  * reclaim or if the lack of progress was due to pages under writeback.
3203  * This is used to determine if the scanning priority needs to be raised.
3204  */
3205 static bool kswapd_shrink_node(pg_data_t *pgdat,
3206                                struct scan_control *sc)
3207 {
3208         struct zone *zone;
3209         int z;
3210 
3211         /* Reclaim a number of pages proportional to the number of zones */
3212         sc->nr_to_reclaim = 0;
3213         for (z = 0; z <= sc->reclaim_idx; z++) {
3214                 zone = pgdat->node_zones + z;
3215                 if (!managed_zone(zone))
3216                         continue;
3217 
3218                 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3219         }
3220 
3221         /*
3222          * Historically care was taken to put equal pressure on all zones but
3223          * now pressure is applied based on node LRU order.
3224          */
3225         shrink_node(pgdat, sc);
3226 
3227         /*
3228          * Fragmentation may mean that the system cannot be rebalanced for
3229          * high-order allocations. If twice the allocation size has been
3230          * reclaimed then recheck watermarks only at order-0 to prevent
3231          * excessive reclaim. Assume that a process requested a high-order
3232          * can direct reclaim/compact.
3233          */
3234         if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3235                 sc->order = 0;
3236 
3237         return sc->nr_scanned >= sc->nr_to_reclaim;
3238 }
3239 
3240 /*
3241  * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3242  * that are eligible for use by the caller until at least one zone is
3243  * balanced.
3244  *
3245  * Returns the order kswapd finished reclaiming at.
3246  *
3247  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3248  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3249  * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3250  * or lower is eligible for reclaim until at least one usable zone is
3251  * balanced.
3252  */
3253 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3254 {
3255         int i;
3256         unsigned long nr_soft_reclaimed;
3257         unsigned long nr_soft_scanned;
3258         struct zone *zone;
3259         struct scan_control sc = {
3260                 .gfp_mask = GFP_KERNEL,
3261                 .order = order,
3262                 .priority = DEF_PRIORITY,
3263                 .may_writepage = !laptop_mode,
3264                 .may_unmap = 1,
3265                 .may_swap = 1,
3266         };
3267         count_vm_event(PAGEOUTRUN);
3268 
3269         do {
3270                 bool raise_priority = true;
3271 
3272                 sc.nr_reclaimed = 0;
3273                 sc.reclaim_idx = classzone_idx;
3274 
3275                 /*
3276                  * If the number of buffer_heads exceeds the maximum allowed
3277                  * then consider reclaiming from all zones. This has a dual
3278                  * purpose -- on 64-bit systems it is expected that
3279                  * buffer_heads are stripped during active rotation. On 32-bit
3280                  * systems, highmem pages can pin lowmem memory and shrinking
3281                  * buffers can relieve lowmem pressure. Reclaim may still not
3282                  * go ahead if all eligible zones for the original allocation
3283                  * request are balanced to avoid excessive reclaim from kswapd.
3284                  */
3285                 if (buffer_heads_over_limit) {
3286                         for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3287                                 zone = pgdat->node_zones + i;
3288                                 if (!managed_zone(zone))
3289                                         continue;
3290 
3291                                 sc.reclaim_idx = i;
3292                                 break;
3293                         }
3294                 }
3295 
3296                 /*
3297                  * Only reclaim if there are no eligible zones. Check from
3298                  * high to low zone as allocations prefer higher zones.
3299                  * Scanning from low to high zone would allow congestion to be
3300                  * cleared during a very small window when a small low
3301                  * zone was balanced even under extreme pressure when the
3302                  * overall node may be congested. Note that sc.reclaim_idx
3303                  * is not used as buffer_heads_over_limit may have adjusted
3304                  * it.
3305                  */
3306                 for (i = classzone_idx; i >= 0; i--) {
3307                         zone = pgdat->node_zones + i;
3308                         if (!managed_zone(zone))
3309                                 continue;
3310 
3311                         if (zone_balanced(zone, sc.order, classzone_idx))
3312                                 goto out;
3313                 }
3314 
3315                 /*
3316                  * Do some background aging of the anon list, to give
3317                  * pages a chance to be referenced before reclaiming. All
3318                  * pages are rotated regardless of classzone as this is
3319                  * about consistent aging.
3320                  */
3321                 age_active_anon(pgdat, &sc);
3322 
3323                 /*
3324                  * If we're getting trouble reclaiming, start doing writepage
3325                  * even in laptop mode.
3326                  */
3327                 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
3328                         sc.may_writepage = 1;
3329 
3330                 /* Call soft limit reclaim before calling shrink_node. */
3331                 sc.nr_scanned = 0;
3332                 nr_soft_scanned = 0;
3333                 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3334                                                 sc.gfp_mask, &nr_soft_scanned);
3335                 sc.nr_reclaimed += nr_soft_reclaimed;
3336 
3337                 /*
3338                  * There should be no need to raise the scanning priority if
3339                  * enough pages are already being scanned that that high
3340                  * watermark would be met at 100% efficiency.
3341                  */
3342                 if (kswapd_shrink_node(pgdat, &sc))
3343                         raise_priority = false;
3344 
3345                 /*
3346                  * If the low watermark is met there is no need for processes
3347                  * to be throttled on pfmemalloc_wait as they should not be
3348                  * able to safely make forward progress. Wake them
3349                  */
3350                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3351                                 pfmemalloc_watermark_ok(pgdat))
3352                         wake_up_all(&pgdat->pfmemalloc_wait);
3353 
3354                 /* Check if kswapd should be suspending */
3355                 if (try_to_freeze() || kthread_should_stop())
3356                         break;
3357 
3358                 /*
3359                  * Raise priority if scanning rate is too low or there was no
3360                  * progress in reclaiming pages
3361                  */
3362                 if (raise_priority || !sc.nr_reclaimed)
3363                         sc.priority--;
3364         } while (sc.priority >= 1);
3365 
3366 out:
3367         snapshot_refaults(NULL, pgdat);
3368         /*
3369          * Return the order kswapd stopped reclaiming at as
3370          * prepare_kswapd_sleep() takes it into account. If another caller
3371          * entered the allocator slow path while kswapd was awake, order will
3372          * remain at the higher level.
3373          */
3374         return sc.order;
3375 }
3376 
3377 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3378                                 unsigned int classzone_idx)
3379 {
3380         long remaining = 0;
3381         DEFINE_WAIT(wait);
3382 
3383         if (freezing(current) || kthread_should_stop())
3384                 return;
3385 
3386         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3387 
3388         /* Try to sleep for a short interval */
3389         if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3390                 /*
3391                  * Compaction records what page blocks it recently failed to
3392                  * isolate pages from and skips them in the future scanning.
3393                  * When kswapd is going to sleep, it is reasonable to assume
3394                  * that pages and compaction may succeed so reset the cache.
3395                  */
3396                 reset_isolation_suitable(pgdat);
3397 
3398                 /*
3399                  * We have freed the memory, now we should compact it to make
3400                  * allocation of the requested order possible.
3401                  */
3402                 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3403 
3404                 remaining = schedule_timeout(HZ/10);
3405 
3406                 /*
3407                  * If woken prematurely then reset kswapd_classzone_idx and
3408                  * order. The values will either be from a wakeup request or
3409                  * the previous request that slept prematurely.
3410                  */
3411                 if (remaining) {
3412                         pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3413                         pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3414                 }
3415 
3416                 finish_wait(&pgdat->kswapd_wait, &wait);
3417                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3418         }
3419 
3420         /*
3421          * After a short sleep, check if it was a premature sleep. If not, then
3422          * go fully to sleep until explicitly woken up.
3423          */
3424         if (!remaining &&
3425             prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3426                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3427 
3428                 /*
3429                  * vmstat counters are not perfectly accurate and the estimated
3430                  * value for counters such as NR_FREE_PAGES can deviate from the
3431                  * true value by nr_online_cpus * threshold. To avoid the zone
3432                  * watermarks being breached while under pressure, we reduce the
3433                  * per-cpu vmstat threshold while kswapd is awake and restore
3434                  * them before going back to sleep.
3435                  */
3436                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3437 
3438                 if (!kthread_should_stop())
3439                         schedule();
3440 
3441                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3442         } else {
3443                 if (remaining)
3444                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3445                 else
3446                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3447         }
3448         finish_wait(&pgdat->kswapd_wait, &wait);
3449 }
3450 
3451 /*
3452  * The background pageout daemon, started as a kernel thread
3453  * from the init process.
3454  *
3455  * This basically trickles out pages so that we have _some_
3456  * free memory available even if there is no other activity
3457  * that frees anything up. This is needed for things like routing
3458  * etc, where we otherwise might have all activity going on in
3459  * asynchronous contexts that cannot page things out.
3460  *
3461  * If there are applications that are active memory-allocators
3462  * (most normal use), this basically shouldn't matter.
3463  */
3464 static int kswapd(void *p)
3465 {
3466         unsigned int alloc_order, reclaim_order, classzone_idx;
3467         pg_data_t *pgdat = (pg_data_t*)p;
3468         struct task_struct *tsk = current;
3469 
3470         struct reclaim_state reclaim_state = {
3471                 .reclaimed_slab = 0,
3472         };
3473         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3474 
3475         lockdep_set_current_reclaim_state(GFP_KERNEL);
3476 
3477         if (!cpumask_empty(cpumask))
3478                 set_cpus_allowed_ptr(tsk, cpumask);
3479         current->reclaim_state = &reclaim_state;
3480 
3481         /*
3482          * Tell the memory management that we're a "memory allocator",
3483          * and that if we need more memory we should get access to it
3484          * regardless (see "__alloc_pages()"). "kswapd" should
3485          * never get caught in the normal page freeing logic.
3486          *
3487          * (Kswapd normally doesn't need memory anyway, but sometimes
3488          * you need a small amount of memory in order to be able to
3489          * page out something else, and this flag essentially protects
3490          * us from recursively trying to free more memory as we're
3491          * trying to free the first piece of memory in the first place).
3492          */
3493         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3494         set_freezable();
3495 
3496         pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3497         pgdat->kswapd_classzone_idx = classzone_idx = 0;
3498         for ( ; ; ) {
3499                 bool ret;
3500 
3501 kswapd_try_sleep:
3502                 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3503                                         classzone_idx);
3504 
3505                 /* Read the new order and classzone_idx */
3506                 alloc_order = reclaim_order = pgdat->kswapd_order;
3507                 classzone_idx = pgdat->kswapd_classzone_idx;
3508                 pgdat->kswapd_order = 0;
3509                 pgdat->kswapd_classzone_idx = 0;
3510 
3511                 ret = try_to_freeze();
3512                 if (kthread_should_stop())
3513                         break;
3514 
3515                 /*
3516                  * We can speed up thawing tasks if we don't call balance_pgdat
3517                  * after returning from the refrigerator
3518                  */
3519                 if (ret)
3520                         continue;
3521 
3522                 /*
3523                  * Reclaim begins at the requested order but if a high-order
3524                  * reclaim fails then kswapd falls back to reclaiming for
3525                  * order-0. If that happens, kswapd will consider sleeping
3526                  * for the order it finished reclaiming at (reclaim_order)
3527                  * but kcompactd is woken to compact for the original
3528                  * request (alloc_order).
3529                  */
3530                 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3531                                                 alloc_order);
3532                 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3533                 if (reclaim_order < alloc_order)
3534                         goto kswapd_try_sleep;
3535 
3536                 alloc_order = reclaim_order = pgdat->kswapd_order;
3537                 classzone_idx = pgdat->kswapd_classzone_idx;
3538         }
3539 
3540         tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3541         current->reclaim_state = NULL;
3542         lockdep_clear_current_reclaim_state();
3543 
3544         return 0;
3545 }
3546 
3547 /*
3548  * A zone is low on free memory, so wake its kswapd task to service it.
3549  */
3550 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3551 {
3552         pg_data_t *pgdat;
3553         int z;
3554 
3555         if (!managed_zone(zone))
3556                 return;
3557 
3558         if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3559                 return;
3560         pgdat = zone->zone_pgdat;
3561         pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3562         pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3563         if (!waitqueue_active(&pgdat->kswapd_wait))
3564                 return;
3565 
3566         /* Only wake kswapd if all zones are unbalanced */
3567         for (z = 0; z <= classzone_idx; z++) {
3568                 zone = pgdat->node_zones + z;
3569                 if (!managed_zone(zone))
3570                         continue;
3571 
3572                 if (zone_balanced(zone, order, classzone_idx))
3573                         return;
3574         }
3575 
3576         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3577         wake_up_interruptible(&pgdat->kswapd_wait);
3578 }
3579 
3580 #ifdef CONFIG_HIBERNATION
3581 /*
3582  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3583  * freed pages.
3584  *
3585  * Rather than trying to age LRUs the aim is to preserve the overall
3586  * LRU order by reclaiming preferentially
3587  * inactive > active > active referenced > active mapped
3588  */
3589 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3590 {
3591         struct reclaim_state reclaim_state;
3592         struct scan_control sc = {
3593                 .nr_to_reclaim = nr_to_reclaim,
3594                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3595                 .reclaim_idx = MAX_NR_ZONES - 1,
3596                 .priority = DEF_PRIORITY,
3597                 .may_writepage = 1,
3598                 .may_unmap = 1,
3599                 .may_swap = 1,
3600                 .hibernation_mode = 1,
3601         };
3602         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3603         struct task_struct *p = current;
3604         unsigned long nr_reclaimed;
3605 
3606         p->flags |= PF_MEMALLOC;
3607         lockdep_set_current_reclaim_state(sc.gfp_mask);
3608         reclaim_state.reclaimed_slab = 0;
3609         p->reclaim_state = &reclaim_state;
3610 
3611         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3612 
3613         p->reclaim_state = NULL;
3614         lockdep_clear_current_reclaim_state();
3615         p->flags &= ~PF_MEMALLOC;
3616 
3617         return nr_reclaimed;
3618 }
3619 #endif /* CONFIG_HIBERNATION */
3620 
3621 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3622    not required for correctness.  So if the last cpu in a node goes
3623    away, we get changed to run anywhere: as the first one comes back,
3624    restore their cpu bindings. */
3625 static int kswapd_cpu_online(unsigned int cpu)
3626 {
3627         int nid;
3628 
3629         for_each_node_state(nid, N_MEMORY) {
3630                 pg_data_t *pgdat = NODE_DATA(nid);
3631                 const struct cpumask *mask;
3632 
3633                 mask = cpumask_of_node(pgdat->node_id);
3634 
3635                 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3636                         /* One of our CPUs online: restore mask */
3637                         set_cpus_allowed_ptr(pgdat->kswapd, mask);
3638         }
3639         return 0;
3640 }
3641 
3642 /*
3643  * This kswapd start function will be called by init and node-hot-add.
3644  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3645  */
3646 int kswapd_run(int nid)
3647 {
3648         pg_data_t *pgdat = NODE_DATA(nid);
3649         int ret = 0;
3650 
3651         if (pgdat->kswapd)
3652                 return 0;
3653 
3654         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3655         if (IS_ERR(pgdat->kswapd)) {
3656                 /* failure at boot is fatal */
3657                 BUG_ON(system_state == SYSTEM_BOOTING);
3658                 pr_err("Failed to start kswapd on node %d\n", nid);
3659                 ret = PTR_ERR(pgdat->kswapd);
3660                 pgdat->kswapd = NULL;
3661         }
3662         return ret;
3663 }
3664 
3665 /*
3666  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3667  * hold mem_hotplug_begin/end().
3668  */
3669 void kswapd_stop(int nid)
3670 {
3671         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3672 
3673         if (kswapd) {
3674                 kthread_stop(kswapd);
3675                 NODE_DATA(nid)->kswapd = NULL;
3676         }
3677 }
3678 
3679 static int __init kswapd_init(void)
3680 {
3681         int nid, ret;
3682 
3683         swap_setup();
3684         for_each_node_state(nid, N_MEMORY)
3685                 kswapd_run(nid);
3686         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3687                                         "mm/vmscan:online", kswapd_cpu_online,
3688                                         NULL);
3689         WARN_ON(ret < 0);
3690         return 0;
3691 }
3692 
3693 module_init(kswapd_init)
3694 
3695 #ifdef CONFIG_NUMA
3696 /*
3697  * Node reclaim mode
3698  *
3699  * If non-zero call node_reclaim when the number of free pages falls below
3700  * the watermarks.
3701  */
3702 int node_reclaim_mode __read_mostly;
3703 
3704 #define RECLAIM_OFF 0
3705 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3706 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3707 #define RECLAIM_UNMAP (1<<2)    /* Unmap pages during reclaim */
3708 
3709 /*
3710  * Priority for NODE_RECLAIM. This determines the fraction of pages
3711  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3712  * a zone.
3713  */
3714 #define NODE_RECLAIM_PRIORITY 4
3715 
3716 /*
3717  * Percentage of pages in a zone that must be unmapped for node_reclaim to
3718  * occur.
3719  */
3720 int sysctl_min_unmapped_ratio = 1;
3721 
3722 /*
3723  * If the number of slab pages in a zone grows beyond this percentage then
3724  * slab reclaim needs to occur.
3725  */
3726 int sysctl_min_slab_ratio = 5;
3727 
3728 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3729 {
3730         unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3731         unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3732                 node_page_state(pgdat, NR_ACTIVE_FILE);
3733 
3734         /*
3735          * It's possible for there to be more file mapped pages than
3736          * accounted for by the pages on the file LRU lists because
3737          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3738          */
3739         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3740 }
3741 
3742 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3743 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3744 {
3745         unsigned long nr_pagecache_reclaimable;
3746         unsigned long delta = 0;
3747 
3748         /*
3749          * If RECLAIM_UNMAP is set, then all file pages are considered
3750          * potentially reclaimable. Otherwise, we have to worry about
3751          * pages like swapcache and node_unmapped_file_pages() provides
3752          * a better estimate
3753          */
3754         if (node_reclaim_mode & RECLAIM_UNMAP)
3755                 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3756         else
3757                 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3758 
3759         /* If we can't clean pages, remove dirty pages from consideration */
3760         if (!(node_reclaim_mode & RECLAIM_WRITE))
3761                 delta += node_page_state(pgdat, NR_FILE_DIRTY);
3762 
3763         /* Watch for any possible underflows due to delta */
3764         if (unlikely(delta > nr_pagecache_reclaimable))
3765                 delta = nr_pagecache_reclaimable;
3766 
3767         return nr_pagecache_reclaimable - delta;
3768 }
3769 
3770 /*
3771  * Try to free up some pages from this node through reclaim.
3772  */
3773 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3774 {
3775         /* Minimum pages needed in order to stay on node */
3776         const unsigned long nr_pages = 1 << order;
3777         struct task_struct *p = current;
3778         struct reclaim_state reclaim_state;
3779         int classzone_idx = gfp_zone(gfp_mask);
3780         struct scan_control sc = {
3781                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3782                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3783                 .order = order,
3784                 .priority = NODE_RECLAIM_PRIORITY,
3785                 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3786                 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3787                 .may_swap = 1,
3788                 .reclaim_idx = classzone_idx,
3789         };
3790 
3791         cond_resched();
3792         /*
3793          * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3794          * and we also need to be able to write out pages for RECLAIM_WRITE
3795          * and RECLAIM_UNMAP.
3796          */
3797         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3798         lockdep_set_current_reclaim_state(gfp_mask);
3799         reclaim_state.reclaimed_slab = 0;
3800         p->reclaim_state = &reclaim_state;
3801 
3802         if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3803                 /*
3804                  * Free memory by calling shrink zone with increasing
3805                  * priorities until we have enough memory freed.
3806                  */
3807                 do {
3808                         shrink_node(pgdat, &sc);
3809                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3810         }
3811 
3812         p->reclaim_state = NULL;
3813         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3814         lockdep_clear_current_reclaim_state();
3815         return sc.nr_reclaimed >= nr_pages;
3816 }
3817 
3818 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3819 {
3820         int ret;
3821 
3822         /*
3823          * Node reclaim reclaims unmapped file backed pages and
3824          * slab pages if we are over the defined limits.
3825          *
3826          * A small portion of unmapped file backed pages is needed for
3827          * file I/O otherwise pages read by file I/O will be immediately
3828          * thrown out if the node is overallocated. So we do not reclaim
3829          * if less than a specified percentage of the node is used by
3830          * unmapped file backed pages.
3831          */
3832         if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3833             sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3834                 return NODE_RECLAIM_FULL;
3835 
3836         if (!pgdat_reclaimable(pgdat))
3837                 return NODE_RECLAIM_FULL;
3838 
3839         /*
3840          * Do not scan if the allocation should not be delayed.
3841          */
3842         if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3843                 return NODE_RECLAIM_NOSCAN;
3844 
3845         /*
3846          * Only run node reclaim on the local node or on nodes that do not
3847          * have associated processors. This will favor the local processor
3848          * over remote processors and spread off node memory allocations
3849          * as wide as possible.
3850          */
3851         if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3852                 return NODE_RECLAIM_NOSCAN;
3853 
3854         if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3855                 return NODE_RECLAIM_NOSCAN;
3856 
3857         ret = __node_reclaim(pgdat, gfp_mask, order);
3858         clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3859 
3860         if (!ret)
3861                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3862 
3863         return ret;
3864 }
3865 #endif
3866 
3867 /*
3868  * page_evictable - test whether a page is evictable
3869  * @page: the page to test
3870  *
3871  * Test whether page is evictable--i.e., should be placed on active/inactive
3872  * lists vs unevictable list.
3873  *
3874  * Reasons page might not be evictable:
3875  * (1) page's mapping marked unevictable
3876  * (2) page is part of an mlocked VMA
3877  *
3878  */
3879 int page_evictable(struct page *page)
3880 {
3881         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3882 }
3883 
3884 #ifdef CONFIG_SHMEM
3885 /**
3886  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3887  * @pages:      array of pages to check
3888  * @nr_pages:   number of pages to check
3889  *
3890  * Checks pages for evictability and moves them to the appropriate lru list.
3891  *
3892  * This function is only used for SysV IPC SHM_UNLOCK.
3893  */
3894 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3895 {
3896         struct lruvec *lruvec;
3897         struct pglist_data *pgdat = NULL;
3898         int pgscanned = 0;
3899         int pgrescued = 0;
3900         int i;
3901 
3902         for (i = 0; i < nr_pages; i++) {
3903                 struct page *page = pages[i];
3904                 struct pglist_data *pagepgdat = page_pgdat(page);
3905 
3906                 pgscanned++;
3907                 if (pagepgdat != pgdat) {
3908                         if (pgdat)
3909                                 spin_unlock_irq(&pgdat->lru_lock);
3910                         pgdat = pagepgdat;
3911                         spin_lock_irq(&pgdat->lru_lock);
3912                 }
3913                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3914 
3915                 if (!PageLRU(page) || !PageUnevictable(page))
3916                         continue;
3917 
3918                 if (page_evictable(page)) {
3919                         enum lru_list lru = page_lru_base_type(page);
3920 
3921                         VM_BUG_ON_PAGE(PageActive(page), page);
3922                         ClearPageUnevictable(page);
3923                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3924                         add_page_to_lru_list(page, lruvec, lru);
3925                         pgrescued++;
3926                 }
3927         }
3928 
3929         if (pgdat) {
3930                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3931                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3932                 spin_unlock_irq(&pgdat->lru_lock);
3933         }
3934 }
3935 #endif /* CONFIG_SHMEM */
3936 

~ [ 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