Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0
2 : /*
3 : * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4 : *
5 : * Swap reorganised 29.12.95, Stephen Tweedie.
6 : * kswapd added: 7.1.96 sct
7 : * Removed kswapd_ctl limits, and swap out as many pages as needed
8 : * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 : * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 : * Multiqueue VM started 5.8.00, Rik van Riel.
11 : */
12 :
13 : #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 :
15 : #include <linux/mm.h>
16 : #include <linux/sched/mm.h>
17 : #include <linux/module.h>
18 : #include <linux/gfp.h>
19 : #include <linux/kernel_stat.h>
20 : #include <linux/swap.h>
21 : #include <linux/pagemap.h>
22 : #include <linux/init.h>
23 : #include <linux/highmem.h>
24 : #include <linux/vmpressure.h>
25 : #include <linux/vmstat.h>
26 : #include <linux/file.h>
27 : #include <linux/writeback.h>
28 : #include <linux/blkdev.h>
29 : #include <linux/buffer_head.h> /* for try_to_release_page(),
30 : buffer_heads_over_limit */
31 : #include <linux/mm_inline.h>
32 : #include <linux/backing-dev.h>
33 : #include <linux/rmap.h>
34 : #include <linux/topology.h>
35 : #include <linux/cpu.h>
36 : #include <linux/cpuset.h>
37 : #include <linux/compaction.h>
38 : #include <linux/notifier.h>
39 : #include <linux/rwsem.h>
40 : #include <linux/delay.h>
41 : #include <linux/kthread.h>
42 : #include <linux/freezer.h>
43 : #include <linux/memcontrol.h>
44 : #include <linux/delayacct.h>
45 : #include <linux/sysctl.h>
46 : #include <linux/oom.h>
47 : #include <linux/pagevec.h>
48 : #include <linux/prefetch.h>
49 : #include <linux/printk.h>
50 : #include <linux/dax.h>
51 : #include <linux/psi.h>
52 :
53 : #include <asm/tlbflush.h>
54 : #include <asm/div64.h>
55 :
56 : #include <linux/swapops.h>
57 : #include <linux/balloon_compaction.h>
58 :
59 : #include "internal.h"
60 :
61 : #define CREATE_TRACE_POINTS
62 : #include <trace/events/vmscan.h>
63 :
64 : struct scan_control {
65 : /* How many pages shrink_list() should reclaim */
66 : unsigned long nr_to_reclaim;
67 :
68 : /*
69 : * Nodemask of nodes allowed by the caller. If NULL, all nodes
70 : * are scanned.
71 : */
72 : nodemask_t *nodemask;
73 :
74 : /*
75 : * The memory cgroup that hit its limit and as a result is the
76 : * primary target of this reclaim invocation.
77 : */
78 : struct mem_cgroup *target_mem_cgroup;
79 :
80 : /*
81 : * Scan pressure balancing between anon and file LRUs
82 : */
83 : unsigned long anon_cost;
84 : unsigned long file_cost;
85 :
86 : /* Can active pages be deactivated as part of reclaim? */
87 : #define DEACTIVATE_ANON 1
88 : #define DEACTIVATE_FILE 2
89 : unsigned int may_deactivate:2;
90 : unsigned int force_deactivate:1;
91 : unsigned int skipped_deactivate:1;
92 :
93 : /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 : unsigned int may_writepage:1;
95 :
96 : /* Can mapped pages be reclaimed? */
97 : unsigned int may_unmap:1;
98 :
99 : /* Can pages be swapped as part of reclaim? */
100 : unsigned int may_swap:1;
101 :
102 : /*
103 : * Cgroups are not reclaimed below their configured memory.low,
104 : * unless we threaten to OOM. If any cgroups are skipped due to
105 : * memory.low and nothing was reclaimed, go back for memory.low.
106 : */
107 : unsigned int memcg_low_reclaim:1;
108 : unsigned int memcg_low_skipped:1;
109 :
110 : unsigned int hibernation_mode:1;
111 :
112 : /* One of the zones is ready for compaction */
113 : unsigned int compaction_ready:1;
114 :
115 : /* There is easily reclaimable cold cache in the current node */
116 : unsigned int cache_trim_mode:1;
117 :
118 : /* The file pages on the current node are dangerously low */
119 : unsigned int file_is_tiny:1;
120 :
121 : /* Allocation order */
122 : s8 order;
123 :
124 : /* Scan (total_size >> priority) pages at once */
125 : s8 priority;
126 :
127 : /* The highest zone to isolate pages for reclaim from */
128 : s8 reclaim_idx;
129 :
130 : /* This context's GFP mask */
131 : gfp_t gfp_mask;
132 :
133 : /* Incremented by the number of inactive pages that were scanned */
134 : unsigned long nr_scanned;
135 :
136 : /* Number of pages freed so far during a call to shrink_zones() */
137 : unsigned long nr_reclaimed;
138 :
139 : struct {
140 : unsigned int dirty;
141 : unsigned int unqueued_dirty;
142 : unsigned int congested;
143 : unsigned int writeback;
144 : unsigned int immediate;
145 : unsigned int file_taken;
146 : unsigned int taken;
147 : } nr;
148 :
149 : /* for recording the reclaimed slab by now */
150 : struct reclaim_state reclaim_state;
151 : };
152 :
153 : #ifdef ARCH_HAS_PREFETCHW
154 : #define prefetchw_prev_lru_page(_page, _base, _field) \
155 : do { \
156 : if ((_page)->lru.prev != _base) { \
157 : struct page *prev; \
158 : \
159 : prev = lru_to_page(&(_page->lru)); \
160 : prefetchw(&prev->_field); \
161 : } \
162 : } while (0)
163 : #else
164 : #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
165 : #endif
166 :
167 : /*
168 : * From 0 .. 200. Higher means more swappy.
169 : */
170 : int vm_swappiness = 60;
171 :
172 0 : static void set_task_reclaim_state(struct task_struct *task,
173 : struct reclaim_state *rs)
174 : {
175 : /* Check for an overwrite */
176 0 : WARN_ON_ONCE(rs && task->reclaim_state);
177 :
178 : /* Check for the nulling of an already-nulled member */
179 0 : WARN_ON_ONCE(!rs && !task->reclaim_state);
180 :
181 0 : task->reclaim_state = rs;
182 0 : }
183 :
184 : static LIST_HEAD(shrinker_list);
185 : static DECLARE_RWSEM(shrinker_rwsem);
186 :
187 : #ifdef CONFIG_MEMCG
188 : /*
189 : * We allow subsystems to populate their shrinker-related
190 : * LRU lists before register_shrinker_prepared() is called
191 : * for the shrinker, since we don't want to impose
192 : * restrictions on their internal registration order.
193 : * In this case shrink_slab_memcg() may find corresponding
194 : * bit is set in the shrinkers map.
195 : *
196 : * This value is used by the function to detect registering
197 : * shrinkers and to skip do_shrink_slab() calls for them.
198 : */
199 : #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
200 :
201 : static DEFINE_IDR(shrinker_idr);
202 : static int shrinker_nr_max;
203 :
204 : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
205 : {
206 : int id, ret = -ENOMEM;
207 :
208 : down_write(&shrinker_rwsem);
209 : /* This may call shrinker, so it must use down_read_trylock() */
210 : id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
211 : if (id < 0)
212 : goto unlock;
213 :
214 : if (id >= shrinker_nr_max) {
215 : if (memcg_expand_shrinker_maps(id)) {
216 : idr_remove(&shrinker_idr, id);
217 : goto unlock;
218 : }
219 :
220 : shrinker_nr_max = id + 1;
221 : }
222 : shrinker->id = id;
223 : ret = 0;
224 : unlock:
225 : up_write(&shrinker_rwsem);
226 : return ret;
227 : }
228 :
229 : static void unregister_memcg_shrinker(struct shrinker *shrinker)
230 : {
231 : int id = shrinker->id;
232 :
233 : BUG_ON(id < 0);
234 :
235 : down_write(&shrinker_rwsem);
236 : idr_remove(&shrinker_idr, id);
237 : up_write(&shrinker_rwsem);
238 : }
239 :
240 : static bool cgroup_reclaim(struct scan_control *sc)
241 : {
242 : return sc->target_mem_cgroup;
243 : }
244 :
245 : /**
246 : * writeback_throttling_sane - is the usual dirty throttling mechanism available?
247 : * @sc: scan_control in question
248 : *
249 : * The normal page dirty throttling mechanism in balance_dirty_pages() is
250 : * completely broken with the legacy memcg and direct stalling in
251 : * shrink_page_list() is used for throttling instead, which lacks all the
252 : * niceties such as fairness, adaptive pausing, bandwidth proportional
253 : * allocation and configurability.
254 : *
255 : * This function tests whether the vmscan currently in progress can assume
256 : * that the normal dirty throttling mechanism is operational.
257 : */
258 : static bool writeback_throttling_sane(struct scan_control *sc)
259 : {
260 : if (!cgroup_reclaim(sc))
261 : return true;
262 : #ifdef CONFIG_CGROUP_WRITEBACK
263 : if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
264 : return true;
265 : #endif
266 : return false;
267 : }
268 : #else
269 : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
270 : {
271 : return 0;
272 : }
273 :
274 : static void unregister_memcg_shrinker(struct shrinker *shrinker)
275 : {
276 : }
277 :
278 0 : static bool cgroup_reclaim(struct scan_control *sc)
279 : {
280 0 : return false;
281 : }
282 :
283 0 : static bool writeback_throttling_sane(struct scan_control *sc)
284 : {
285 0 : return true;
286 : }
287 : #endif
288 :
289 : /*
290 : * This misses isolated pages which are not accounted for to save counters.
291 : * As the data only determines if reclaim or compaction continues, it is
292 : * not expected that isolated pages will be a dominating factor.
293 : */
294 0 : unsigned long zone_reclaimable_pages(struct zone *zone)
295 : {
296 0 : unsigned long nr;
297 :
298 0 : nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
299 0 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
300 0 : if (get_nr_swap_pages() > 0)
301 : nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
302 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
303 :
304 0 : return nr;
305 : }
306 :
307 : /**
308 : * lruvec_lru_size - Returns the number of pages on the given LRU list.
309 : * @lruvec: lru vector
310 : * @lru: lru to use
311 : * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
312 : */
313 0 : static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
314 : int zone_idx)
315 : {
316 0 : unsigned long size = 0;
317 0 : int zid;
318 :
319 0 : for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
320 0 : struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
321 :
322 0 : if (!managed_zone(zone))
323 0 : continue;
324 :
325 0 : if (!mem_cgroup_disabled())
326 : size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
327 : else
328 0 : size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
329 : }
330 0 : return size;
331 : }
332 :
333 : /*
334 : * Add a shrinker callback to be called from the vm.
335 : */
336 129 : int prealloc_shrinker(struct shrinker *shrinker)
337 : {
338 129 : unsigned int size = sizeof(*shrinker->nr_deferred);
339 :
340 129 : if (shrinker->flags & SHRINKER_NUMA_AWARE)
341 125 : size *= nr_node_ids;
342 :
343 129 : shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
344 129 : if (!shrinker->nr_deferred)
345 0 : return -ENOMEM;
346 :
347 : if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
348 129 : if (prealloc_memcg_shrinker(shrinker))
349 : goto free_deferred;
350 : }
351 :
352 : return 0;
353 :
354 : free_deferred:
355 : kfree(shrinker->nr_deferred);
356 : shrinker->nr_deferred = NULL;
357 : return -ENOMEM;
358 : }
359 :
360 0 : void free_prealloced_shrinker(struct shrinker *shrinker)
361 : {
362 0 : if (!shrinker->nr_deferred)
363 : return;
364 :
365 0 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
366 0 : unregister_memcg_shrinker(shrinker);
367 :
368 0 : kfree(shrinker->nr_deferred);
369 0 : shrinker->nr_deferred = NULL;
370 : }
371 :
372 129 : void register_shrinker_prepared(struct shrinker *shrinker)
373 : {
374 129 : down_write(&shrinker_rwsem);
375 129 : list_add_tail(&shrinker->list, &shrinker_list);
376 : #ifdef CONFIG_MEMCG
377 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
378 : idr_replace(&shrinker_idr, shrinker, shrinker->id);
379 : #endif
380 129 : up_write(&shrinker_rwsem);
381 129 : }
382 :
383 5 : int register_shrinker(struct shrinker *shrinker)
384 : {
385 5 : int err = prealloc_shrinker(shrinker);
386 :
387 5 : if (err)
388 : return err;
389 5 : register_shrinker_prepared(shrinker);
390 5 : return 0;
391 : }
392 : EXPORT_SYMBOL(register_shrinker);
393 :
394 : /*
395 : * Remove one
396 : */
397 99 : void unregister_shrinker(struct shrinker *shrinker)
398 : {
399 99 : if (!shrinker->nr_deferred)
400 : return;
401 99 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
402 99 : unregister_memcg_shrinker(shrinker);
403 99 : down_write(&shrinker_rwsem);
404 99 : list_del(&shrinker->list);
405 99 : up_write(&shrinker_rwsem);
406 99 : kfree(shrinker->nr_deferred);
407 99 : shrinker->nr_deferred = NULL;
408 : }
409 : EXPORT_SYMBOL(unregister_shrinker);
410 :
411 : #define SHRINK_BATCH 128
412 :
413 0 : static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
414 : struct shrinker *shrinker, int priority)
415 : {
416 0 : unsigned long freed = 0;
417 0 : unsigned long long delta;
418 0 : long total_scan;
419 0 : long freeable;
420 0 : long nr;
421 0 : long new_nr;
422 0 : int nid = shrinkctl->nid;
423 0 : long batch_size = shrinker->batch ? shrinker->batch
424 0 : : SHRINK_BATCH;
425 0 : long scanned = 0, next_deferred;
426 :
427 0 : if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
428 0 : nid = 0;
429 :
430 0 : freeable = shrinker->count_objects(shrinker, shrinkctl);
431 0 : if (freeable == 0 || freeable == SHRINK_EMPTY)
432 : return freeable;
433 :
434 : /*
435 : * copy the current shrinker scan count into a local variable
436 : * and zero it so that other concurrent shrinker invocations
437 : * don't also do this scanning work.
438 : */
439 0 : nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
440 :
441 0 : total_scan = nr;
442 0 : if (shrinker->seeks) {
443 0 : delta = freeable >> priority;
444 0 : delta *= 4;
445 0 : do_div(delta, shrinker->seeks);
446 : } else {
447 : /*
448 : * These objects don't require any IO to create. Trim
449 : * them aggressively under memory pressure to keep
450 : * them from causing refetches in the IO caches.
451 : */
452 0 : delta = freeable / 2;
453 : }
454 :
455 0 : total_scan += delta;
456 0 : if (total_scan < 0) {
457 0 : pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
458 : shrinker->scan_objects, total_scan);
459 0 : total_scan = freeable;
460 0 : next_deferred = nr;
461 : } else
462 : next_deferred = total_scan;
463 :
464 : /*
465 : * We need to avoid excessive windup on filesystem shrinkers
466 : * due to large numbers of GFP_NOFS allocations causing the
467 : * shrinkers to return -1 all the time. This results in a large
468 : * nr being built up so when a shrink that can do some work
469 : * comes along it empties the entire cache due to nr >>>
470 : * freeable. This is bad for sustaining a working set in
471 : * memory.
472 : *
473 : * Hence only allow the shrinker to scan the entire cache when
474 : * a large delta change is calculated directly.
475 : */
476 0 : if (delta < freeable / 4)
477 0 : total_scan = min(total_scan, freeable / 2);
478 :
479 : /*
480 : * Avoid risking looping forever due to too large nr value:
481 : * never try to free more than twice the estimate number of
482 : * freeable entries.
483 : */
484 0 : if (total_scan > freeable * 2)
485 : total_scan = freeable * 2;
486 :
487 0 : trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
488 : freeable, delta, total_scan, priority);
489 :
490 : /*
491 : * Normally, we should not scan less than batch_size objects in one
492 : * pass to avoid too frequent shrinker calls, but if the slab has less
493 : * than batch_size objects in total and we are really tight on memory,
494 : * we will try to reclaim all available objects, otherwise we can end
495 : * up failing allocations although there are plenty of reclaimable
496 : * objects spread over several slabs with usage less than the
497 : * batch_size.
498 : *
499 : * We detect the "tight on memory" situations by looking at the total
500 : * number of objects we want to scan (total_scan). If it is greater
501 : * than the total number of objects on slab (freeable), we must be
502 : * scanning at high prio and therefore should try to reclaim as much as
503 : * possible.
504 : */
505 0 : while (total_scan >= batch_size ||
506 : total_scan >= freeable) {
507 0 : unsigned long ret;
508 0 : unsigned long nr_to_scan = min(batch_size, total_scan);
509 :
510 0 : shrinkctl->nr_to_scan = nr_to_scan;
511 0 : shrinkctl->nr_scanned = nr_to_scan;
512 0 : ret = shrinker->scan_objects(shrinker, shrinkctl);
513 0 : if (ret == SHRINK_STOP)
514 : break;
515 0 : freed += ret;
516 :
517 0 : count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
518 0 : total_scan -= shrinkctl->nr_scanned;
519 0 : scanned += shrinkctl->nr_scanned;
520 :
521 0 : cond_resched();
522 : }
523 :
524 0 : if (next_deferred >= scanned)
525 0 : next_deferred -= scanned;
526 : else
527 : next_deferred = 0;
528 : /*
529 : * move the unused scan count back into the shrinker in a
530 : * manner that handles concurrent updates. If we exhausted the
531 : * scan, there is no need to do an update.
532 : */
533 0 : if (next_deferred > 0)
534 0 : new_nr = atomic_long_add_return(next_deferred,
535 0 : &shrinker->nr_deferred[nid]);
536 : else
537 0 : new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
538 :
539 0 : trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
540 0 : return freed;
541 : }
542 :
543 : #ifdef CONFIG_MEMCG
544 : static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
545 : struct mem_cgroup *memcg, int priority)
546 : {
547 : struct memcg_shrinker_map *map;
548 : unsigned long ret, freed = 0;
549 : int i;
550 :
551 : if (!mem_cgroup_online(memcg))
552 : return 0;
553 :
554 : if (!down_read_trylock(&shrinker_rwsem))
555 : return 0;
556 :
557 : map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
558 : true);
559 : if (unlikely(!map))
560 : goto unlock;
561 :
562 : for_each_set_bit(i, map->map, shrinker_nr_max) {
563 : struct shrink_control sc = {
564 : .gfp_mask = gfp_mask,
565 : .nid = nid,
566 : .memcg = memcg,
567 : };
568 : struct shrinker *shrinker;
569 :
570 : shrinker = idr_find(&shrinker_idr, i);
571 : if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
572 : if (!shrinker)
573 : clear_bit(i, map->map);
574 : continue;
575 : }
576 :
577 : /* Call non-slab shrinkers even though kmem is disabled */
578 : if (!memcg_kmem_enabled() &&
579 : !(shrinker->flags & SHRINKER_NONSLAB))
580 : continue;
581 :
582 : ret = do_shrink_slab(&sc, shrinker, priority);
583 : if (ret == SHRINK_EMPTY) {
584 : clear_bit(i, map->map);
585 : /*
586 : * After the shrinker reported that it had no objects to
587 : * free, but before we cleared the corresponding bit in
588 : * the memcg shrinker map, a new object might have been
589 : * added. To make sure, we have the bit set in this
590 : * case, we invoke the shrinker one more time and reset
591 : * the bit if it reports that it is not empty anymore.
592 : * The memory barrier here pairs with the barrier in
593 : * memcg_set_shrinker_bit():
594 : *
595 : * list_lru_add() shrink_slab_memcg()
596 : * list_add_tail() clear_bit()
597 : * <MB> <MB>
598 : * set_bit() do_shrink_slab()
599 : */
600 : smp_mb__after_atomic();
601 : ret = do_shrink_slab(&sc, shrinker, priority);
602 : if (ret == SHRINK_EMPTY)
603 : ret = 0;
604 : else
605 : memcg_set_shrinker_bit(memcg, nid, i);
606 : }
607 : freed += ret;
608 :
609 : if (rwsem_is_contended(&shrinker_rwsem)) {
610 : freed = freed ? : 1;
611 : break;
612 : }
613 : }
614 : unlock:
615 : up_read(&shrinker_rwsem);
616 : return freed;
617 : }
618 : #else /* CONFIG_MEMCG */
619 : static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
620 : struct mem_cgroup *memcg, int priority)
621 : {
622 : return 0;
623 : }
624 : #endif /* CONFIG_MEMCG */
625 :
626 : /**
627 : * shrink_slab - shrink slab caches
628 : * @gfp_mask: allocation context
629 : * @nid: node whose slab caches to target
630 : * @memcg: memory cgroup whose slab caches to target
631 : * @priority: the reclaim priority
632 : *
633 : * Call the shrink functions to age shrinkable caches.
634 : *
635 : * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
636 : * unaware shrinkers will receive a node id of 0 instead.
637 : *
638 : * @memcg specifies the memory cgroup to target. Unaware shrinkers
639 : * are called only if it is the root cgroup.
640 : *
641 : * @priority is sc->priority, we take the number of objects and >> by priority
642 : * in order to get the scan target.
643 : *
644 : * Returns the number of reclaimed slab objects.
645 : */
646 0 : static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
647 : struct mem_cgroup *memcg,
648 : int priority)
649 : {
650 0 : unsigned long ret, freed = 0;
651 0 : struct shrinker *shrinker;
652 :
653 : /*
654 : * The root memcg might be allocated even though memcg is disabled
655 : * via "cgroup_disable=memory" boot parameter. This could make
656 : * mem_cgroup_is_root() return false, then just run memcg slab
657 : * shrink, but skip global shrink. This may result in premature
658 : * oom.
659 : */
660 0 : if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
661 : return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
662 :
663 0 : if (!down_read_trylock(&shrinker_rwsem))
664 0 : goto out;
665 :
666 0 : list_for_each_entry(shrinker, &shrinker_list, list) {
667 0 : struct shrink_control sc = {
668 : .gfp_mask = gfp_mask,
669 : .nid = nid,
670 : .memcg = memcg,
671 : };
672 :
673 0 : ret = do_shrink_slab(&sc, shrinker, priority);
674 0 : if (ret == SHRINK_EMPTY)
675 0 : ret = 0;
676 0 : freed += ret;
677 : /*
678 : * Bail out if someone want to register a new shrinker to
679 : * prevent the registration from being stalled for long periods
680 : * by parallel ongoing shrinking.
681 : */
682 0 : if (rwsem_is_contended(&shrinker_rwsem)) {
683 0 : freed = freed ? : 1;
684 0 : break;
685 : }
686 : }
687 :
688 0 : up_read(&shrinker_rwsem);
689 0 : out:
690 0 : cond_resched();
691 0 : return freed;
692 : }
693 :
694 0 : void drop_slab_node(int nid)
695 : {
696 0 : unsigned long freed;
697 :
698 0 : do {
699 0 : struct mem_cgroup *memcg = NULL;
700 :
701 0 : if (fatal_signal_pending(current))
702 : return;
703 :
704 0 : freed = 0;
705 0 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
706 0 : do {
707 0 : freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
708 0 : } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
709 0 : } while (freed > 10);
710 : }
711 :
712 0 : void drop_slab(void)
713 : {
714 0 : int nid;
715 :
716 0 : for_each_online_node(nid)
717 0 : drop_slab_node(nid);
718 0 : }
719 :
720 0 : static inline int is_page_cache_freeable(struct page *page)
721 : {
722 : /*
723 : * A freeable page cache page is referenced only by the caller
724 : * that isolated the page, the page cache and optional buffer
725 : * heads at page->private.
726 : */
727 0 : int page_cache_pins = thp_nr_pages(page);
728 0 : return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
729 : }
730 :
731 0 : static int may_write_to_inode(struct inode *inode)
732 : {
733 0 : if (current->flags & PF_SWAPWRITE)
734 : return 1;
735 0 : if (!inode_write_congested(inode))
736 : return 1;
737 0 : if (inode_to_bdi(inode) == current->backing_dev_info)
738 0 : return 1;
739 : return 0;
740 : }
741 :
742 : /*
743 : * We detected a synchronous write error writing a page out. Probably
744 : * -ENOSPC. We need to propagate that into the address_space for a subsequent
745 : * fsync(), msync() or close().
746 : *
747 : * The tricky part is that after writepage we cannot touch the mapping: nothing
748 : * prevents it from being freed up. But we have a ref on the page and once
749 : * that page is locked, the mapping is pinned.
750 : *
751 : * We're allowed to run sleeping lock_page() here because we know the caller has
752 : * __GFP_FS.
753 : */
754 0 : static void handle_write_error(struct address_space *mapping,
755 : struct page *page, int error)
756 : {
757 0 : lock_page(page);
758 0 : if (page_mapping(page) == mapping)
759 0 : mapping_set_error(mapping, error);
760 0 : unlock_page(page);
761 0 : }
762 :
763 : /* possible outcome of pageout() */
764 : typedef enum {
765 : /* failed to write page out, page is locked */
766 : PAGE_KEEP,
767 : /* move page to the active list, page is locked */
768 : PAGE_ACTIVATE,
769 : /* page has been sent to the disk successfully, page is unlocked */
770 : PAGE_SUCCESS,
771 : /* page is clean and locked */
772 : PAGE_CLEAN,
773 : } pageout_t;
774 :
775 : /*
776 : * pageout is called by shrink_page_list() for each dirty page.
777 : * Calls ->writepage().
778 : */
779 0 : static pageout_t pageout(struct page *page, struct address_space *mapping)
780 : {
781 : /*
782 : * If the page is dirty, only perform writeback if that write
783 : * will be non-blocking. To prevent this allocation from being
784 : * stalled by pagecache activity. But note that there may be
785 : * stalls if we need to run get_block(). We could test
786 : * PagePrivate for that.
787 : *
788 : * If this process is currently in __generic_file_write_iter() against
789 : * this page's queue, we can perform writeback even if that
790 : * will block.
791 : *
792 : * If the page is swapcache, write it back even if that would
793 : * block, for some throttling. This happens by accident, because
794 : * swap_backing_dev_info is bust: it doesn't reflect the
795 : * congestion state of the swapdevs. Easy to fix, if needed.
796 : */
797 0 : if (!is_page_cache_freeable(page))
798 : return PAGE_KEEP;
799 0 : if (!mapping) {
800 : /*
801 : * Some data journaling orphaned pages can have
802 : * page->mapping == NULL while being dirty with clean buffers.
803 : */
804 0 : if (page_has_private(page)) {
805 0 : if (try_to_free_buffers(page)) {
806 0 : ClearPageDirty(page);
807 0 : pr_info("%s: orphaned page\n", __func__);
808 0 : return PAGE_CLEAN;
809 : }
810 : }
811 0 : return PAGE_KEEP;
812 : }
813 0 : if (mapping->a_ops->writepage == NULL)
814 : return PAGE_ACTIVATE;
815 0 : if (!may_write_to_inode(mapping->host))
816 : return PAGE_KEEP;
817 :
818 0 : if (clear_page_dirty_for_io(page)) {
819 0 : int res;
820 0 : struct writeback_control wbc = {
821 : .sync_mode = WB_SYNC_NONE,
822 : .nr_to_write = SWAP_CLUSTER_MAX,
823 : .range_start = 0,
824 : .range_end = LLONG_MAX,
825 : .for_reclaim = 1,
826 : };
827 :
828 0 : SetPageReclaim(page);
829 0 : res = mapping->a_ops->writepage(page, &wbc);
830 0 : if (res < 0)
831 0 : handle_write_error(mapping, page, res);
832 0 : if (res == AOP_WRITEPAGE_ACTIVATE) {
833 0 : ClearPageReclaim(page);
834 0 : return PAGE_ACTIVATE;
835 : }
836 :
837 0 : if (!PageWriteback(page)) {
838 : /* synchronous write or broken a_ops? */
839 0 : ClearPageReclaim(page);
840 : }
841 0 : trace_mm_vmscan_writepage(page);
842 0 : inc_node_page_state(page, NR_VMSCAN_WRITE);
843 0 : return PAGE_SUCCESS;
844 : }
845 :
846 : return PAGE_CLEAN;
847 : }
848 :
849 : /*
850 : * Same as remove_mapping, but if the page is removed from the mapping, it
851 : * gets returned with a refcount of 0.
852 : */
853 0 : static int __remove_mapping(struct address_space *mapping, struct page *page,
854 : bool reclaimed, struct mem_cgroup *target_memcg)
855 : {
856 0 : unsigned long flags;
857 0 : int refcount;
858 0 : void *shadow = NULL;
859 :
860 0 : BUG_ON(!PageLocked(page));
861 0 : BUG_ON(mapping != page_mapping(page));
862 :
863 0 : xa_lock_irqsave(&mapping->i_pages, flags);
864 : /*
865 : * The non racy check for a busy page.
866 : *
867 : * Must be careful with the order of the tests. When someone has
868 : * a ref to the page, it may be possible that they dirty it then
869 : * drop the reference. So if PageDirty is tested before page_count
870 : * here, then the following race may occur:
871 : *
872 : * get_user_pages(&page);
873 : * [user mapping goes away]
874 : * write_to(page);
875 : * !PageDirty(page) [good]
876 : * SetPageDirty(page);
877 : * put_page(page);
878 : * !page_count(page) [good, discard it]
879 : *
880 : * [oops, our write_to data is lost]
881 : *
882 : * Reversing the order of the tests ensures such a situation cannot
883 : * escape unnoticed. The smp_rmb is needed to ensure the page->flags
884 : * load is not satisfied before that of page->_refcount.
885 : *
886 : * Note that if SetPageDirty is always performed via set_page_dirty,
887 : * and thus under the i_pages lock, then this ordering is not required.
888 : */
889 0 : refcount = 1 + compound_nr(page);
890 0 : if (!page_ref_freeze(page, refcount))
891 0 : goto cannot_free;
892 : /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
893 0 : if (unlikely(PageDirty(page))) {
894 0 : page_ref_unfreeze(page, refcount);
895 0 : goto cannot_free;
896 : }
897 :
898 0 : if (PageSwapCache(page)) {
899 : swp_entry_t swap = { .val = page_private(page) };
900 : mem_cgroup_swapout(page, swap);
901 : if (reclaimed && !mapping_exiting(mapping))
902 : shadow = workingset_eviction(page, target_memcg);
903 : __delete_from_swap_cache(page, swap, shadow);
904 : xa_unlock_irqrestore(&mapping->i_pages, flags);
905 : put_swap_page(page, swap);
906 : } else {
907 0 : void (*freepage)(struct page *);
908 :
909 0 : freepage = mapping->a_ops->freepage;
910 : /*
911 : * Remember a shadow entry for reclaimed file cache in
912 : * order to detect refaults, thus thrashing, later on.
913 : *
914 : * But don't store shadows in an address space that is
915 : * already exiting. This is not just an optimization,
916 : * inode reclaim needs to empty out the radix tree or
917 : * the nodes are lost. Don't plant shadows behind its
918 : * back.
919 : *
920 : * We also don't store shadows for DAX mappings because the
921 : * only page cache pages found in these are zero pages
922 : * covering holes, and because we don't want to mix DAX
923 : * exceptional entries and shadow exceptional entries in the
924 : * same address_space.
925 : */
926 0 : if (reclaimed && page_is_file_lru(page) &&
927 0 : !mapping_exiting(mapping) && !dax_mapping(mapping))
928 0 : shadow = workingset_eviction(page, target_memcg);
929 0 : __delete_from_page_cache(page, shadow);
930 0 : xa_unlock_irqrestore(&mapping->i_pages, flags);
931 :
932 0 : if (freepage != NULL)
933 0 : freepage(page);
934 : }
935 :
936 : return 1;
937 :
938 0 : cannot_free:
939 0 : xa_unlock_irqrestore(&mapping->i_pages, flags);
940 0 : return 0;
941 : }
942 :
943 : /*
944 : * Attempt to detach a locked page from its ->mapping. If it is dirty or if
945 : * someone else has a ref on the page, abort and return 0. If it was
946 : * successfully detached, return 1. Assumes the caller has a single ref on
947 : * this page.
948 : */
949 0 : int remove_mapping(struct address_space *mapping, struct page *page)
950 : {
951 0 : if (__remove_mapping(mapping, page, false, NULL)) {
952 : /*
953 : * Unfreezing the refcount with 1 rather than 2 effectively
954 : * drops the pagecache ref for us without requiring another
955 : * atomic operation.
956 : */
957 0 : page_ref_unfreeze(page, 1);
958 0 : return 1;
959 : }
960 : return 0;
961 : }
962 :
963 : /**
964 : * putback_lru_page - put previously isolated page onto appropriate LRU list
965 : * @page: page to be put back to appropriate lru list
966 : *
967 : * Add previously isolated @page to appropriate LRU list.
968 : * Page may still be unevictable for other reasons.
969 : *
970 : * lru_lock must not be held, interrupts must be enabled.
971 : */
972 11 : void putback_lru_page(struct page *page)
973 : {
974 11 : lru_cache_add(page);
975 11 : put_page(page); /* drop ref from isolate */
976 0 : }
977 :
978 : enum page_references {
979 : PAGEREF_RECLAIM,
980 : PAGEREF_RECLAIM_CLEAN,
981 : PAGEREF_KEEP,
982 : PAGEREF_ACTIVATE,
983 : };
984 :
985 0 : static enum page_references page_check_references(struct page *page,
986 : struct scan_control *sc)
987 : {
988 0 : int referenced_ptes, referenced_page;
989 0 : unsigned long vm_flags;
990 :
991 0 : referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
992 : &vm_flags);
993 0 : referenced_page = TestClearPageReferenced(page);
994 :
995 : /*
996 : * Mlock lost the isolation race with us. Let try_to_unmap()
997 : * move the page to the unevictable list.
998 : */
999 0 : if (vm_flags & VM_LOCKED)
1000 : return PAGEREF_RECLAIM;
1001 :
1002 0 : if (referenced_ptes) {
1003 : /*
1004 : * All mapped pages start out with page table
1005 : * references from the instantiating fault, so we need
1006 : * to look twice if a mapped file page is used more
1007 : * than once.
1008 : *
1009 : * Mark it and spare it for another trip around the
1010 : * inactive list. Another page table reference will
1011 : * lead to its activation.
1012 : *
1013 : * Note: the mark is set for activated pages as well
1014 : * so that recently deactivated but used pages are
1015 : * quickly recovered.
1016 : */
1017 0 : SetPageReferenced(page);
1018 :
1019 0 : if (referenced_page || referenced_ptes > 1)
1020 : return PAGEREF_ACTIVATE;
1021 :
1022 : /*
1023 : * Activate file-backed executable pages after first usage.
1024 : */
1025 0 : if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1026 : return PAGEREF_ACTIVATE;
1027 :
1028 0 : return PAGEREF_KEEP;
1029 : }
1030 :
1031 : /* Reclaim if clean, defer dirty pages to writeback */
1032 0 : if (referenced_page && !PageSwapBacked(page))
1033 0 : return PAGEREF_RECLAIM_CLEAN;
1034 :
1035 : return PAGEREF_RECLAIM;
1036 : }
1037 :
1038 : /* Check if a page is dirty or under writeback */
1039 0 : static void page_check_dirty_writeback(struct page *page,
1040 : bool *dirty, bool *writeback)
1041 : {
1042 0 : struct address_space *mapping;
1043 :
1044 : /*
1045 : * Anonymous pages are not handled by flushers and must be written
1046 : * from reclaim context. Do not stall reclaim based on them
1047 : */
1048 0 : if (!page_is_file_lru(page) ||
1049 0 : (PageAnon(page) && !PageSwapBacked(page))) {
1050 0 : *dirty = false;
1051 0 : *writeback = false;
1052 0 : return;
1053 : }
1054 :
1055 : /* By default assume that the page flags are accurate */
1056 0 : *dirty = PageDirty(page);
1057 0 : *writeback = PageWriteback(page);
1058 :
1059 : /* Verify dirty/writeback state if the filesystem supports it */
1060 0 : if (!page_has_private(page))
1061 : return;
1062 :
1063 0 : mapping = page_mapping(page);
1064 0 : if (mapping && mapping->a_ops->is_dirty_writeback)
1065 0 : mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1066 : }
1067 :
1068 : /*
1069 : * shrink_page_list() returns the number of reclaimed pages
1070 : */
1071 0 : static unsigned int shrink_page_list(struct list_head *page_list,
1072 : struct pglist_data *pgdat,
1073 : struct scan_control *sc,
1074 : struct reclaim_stat *stat,
1075 : bool ignore_references)
1076 : {
1077 0 : LIST_HEAD(ret_pages);
1078 0 : LIST_HEAD(free_pages);
1079 0 : unsigned int nr_reclaimed = 0;
1080 0 : unsigned int pgactivate = 0;
1081 :
1082 0 : memset(stat, 0, sizeof(*stat));
1083 0 : cond_resched();
1084 :
1085 0 : while (!list_empty(page_list)) {
1086 0 : struct address_space *mapping;
1087 0 : struct page *page;
1088 0 : enum page_references references = PAGEREF_RECLAIM;
1089 0 : bool dirty, writeback, may_enter_fs;
1090 0 : unsigned int nr_pages;
1091 :
1092 0 : cond_resched();
1093 :
1094 0 : page = lru_to_page(page_list);
1095 0 : list_del(&page->lru);
1096 :
1097 0 : if (!trylock_page(page))
1098 0 : goto keep;
1099 :
1100 0 : VM_BUG_ON_PAGE(PageActive(page), page);
1101 :
1102 0 : nr_pages = compound_nr(page);
1103 :
1104 : /* Account the number of base pages even though THP */
1105 0 : sc->nr_scanned += nr_pages;
1106 :
1107 0 : if (unlikely(!page_evictable(page)))
1108 0 : goto activate_locked;
1109 :
1110 0 : if (!sc->may_unmap && page_mapped(page))
1111 0 : goto keep_locked;
1112 :
1113 0 : may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1114 0 : (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1115 :
1116 : /*
1117 : * The number of dirty pages determines if a node is marked
1118 : * reclaim_congested which affects wait_iff_congested. kswapd
1119 : * will stall and start writing pages if the tail of the LRU
1120 : * is all dirty unqueued pages.
1121 : */
1122 0 : page_check_dirty_writeback(page, &dirty, &writeback);
1123 0 : if (dirty || writeback)
1124 0 : stat->nr_dirty++;
1125 :
1126 0 : if (dirty && !writeback)
1127 0 : stat->nr_unqueued_dirty++;
1128 :
1129 : /*
1130 : * Treat this page as congested if the underlying BDI is or if
1131 : * pages are cycling through the LRU so quickly that the
1132 : * pages marked for immediate reclaim are making it to the
1133 : * end of the LRU a second time.
1134 : */
1135 0 : mapping = page_mapping(page);
1136 0 : if (((dirty || writeback) && mapping &&
1137 0 : inode_write_congested(mapping->host)) ||
1138 0 : (writeback && PageReclaim(page)))
1139 0 : stat->nr_congested++;
1140 :
1141 : /*
1142 : * If a page at the tail of the LRU is under writeback, there
1143 : * are three cases to consider.
1144 : *
1145 : * 1) If reclaim is encountering an excessive number of pages
1146 : * under writeback and this page is both under writeback and
1147 : * PageReclaim then it indicates that pages are being queued
1148 : * for IO but are being recycled through the LRU before the
1149 : * IO can complete. Waiting on the page itself risks an
1150 : * indefinite stall if it is impossible to writeback the
1151 : * page due to IO error or disconnected storage so instead
1152 : * note that the LRU is being scanned too quickly and the
1153 : * caller can stall after page list has been processed.
1154 : *
1155 : * 2) Global or new memcg reclaim encounters a page that is
1156 : * not marked for immediate reclaim, or the caller does not
1157 : * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1158 : * not to fs). In this case mark the page for immediate
1159 : * reclaim and continue scanning.
1160 : *
1161 : * Require may_enter_fs because we would wait on fs, which
1162 : * may not have submitted IO yet. And the loop driver might
1163 : * enter reclaim, and deadlock if it waits on a page for
1164 : * which it is needed to do the write (loop masks off
1165 : * __GFP_IO|__GFP_FS for this reason); but more thought
1166 : * would probably show more reasons.
1167 : *
1168 : * 3) Legacy memcg encounters a page that is already marked
1169 : * PageReclaim. memcg does not have any dirty pages
1170 : * throttling so we could easily OOM just because too many
1171 : * pages are in writeback and there is nothing else to
1172 : * reclaim. Wait for the writeback to complete.
1173 : *
1174 : * In cases 1) and 2) we activate the pages to get them out of
1175 : * the way while we continue scanning for clean pages on the
1176 : * inactive list and refilling from the active list. The
1177 : * observation here is that waiting for disk writes is more
1178 : * expensive than potentially causing reloads down the line.
1179 : * Since they're marked for immediate reclaim, they won't put
1180 : * memory pressure on the cache working set any longer than it
1181 : * takes to write them to disk.
1182 : */
1183 0 : if (PageWriteback(page)) {
1184 : /* Case 1 above */
1185 0 : if (current_is_kswapd() &&
1186 0 : PageReclaim(page) &&
1187 0 : test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1188 0 : stat->nr_immediate++;
1189 0 : goto activate_locked;
1190 :
1191 : /* Case 2 above */
1192 0 : } else if (writeback_throttling_sane(sc) ||
1193 : !PageReclaim(page) || !may_enter_fs) {
1194 : /*
1195 : * This is slightly racy - end_page_writeback()
1196 : * might have just cleared PageReclaim, then
1197 : * setting PageReclaim here end up interpreted
1198 : * as PageReadahead - but that does not matter
1199 : * enough to care. What we do want is for this
1200 : * page to have PageReclaim set next time memcg
1201 : * reclaim reaches the tests above, so it will
1202 : * then wait_on_page_writeback() to avoid OOM;
1203 : * and it's also appropriate in global reclaim.
1204 : */
1205 0 : SetPageReclaim(page);
1206 0 : stat->nr_writeback++;
1207 0 : goto activate_locked;
1208 :
1209 : /* Case 3 above */
1210 : } else {
1211 : unlock_page(page);
1212 : wait_on_page_writeback(page);
1213 : /* then go back and try same page again */
1214 : list_add_tail(&page->lru, page_list);
1215 0 : continue;
1216 : }
1217 : }
1218 :
1219 0 : if (!ignore_references)
1220 0 : references = page_check_references(page, sc);
1221 :
1222 0 : switch (references) {
1223 0 : case PAGEREF_ACTIVATE:
1224 0 : goto activate_locked;
1225 0 : case PAGEREF_KEEP:
1226 0 : stat->nr_ref_keep += nr_pages;
1227 0 : goto keep_locked;
1228 : case PAGEREF_RECLAIM:
1229 : case PAGEREF_RECLAIM_CLEAN:
1230 0 : ; /* try to reclaim the page below */
1231 : }
1232 :
1233 : /*
1234 : * Anonymous process memory has backing store?
1235 : * Try to allocate it some swap space here.
1236 : * Lazyfree page could be freed directly
1237 : */
1238 0 : if (PageAnon(page) && PageSwapBacked(page)) {
1239 0 : if (!PageSwapCache(page)) {
1240 0 : if (!(sc->gfp_mask & __GFP_IO))
1241 0 : goto keep_locked;
1242 0 : if (page_maybe_dma_pinned(page))
1243 0 : goto keep_locked;
1244 0 : if (PageTransHuge(page)) {
1245 : /* cannot split THP, skip it */
1246 0 : if (!can_split_huge_page(page, NULL))
1247 0 : goto activate_locked;
1248 : /*
1249 : * Split pages without a PMD map right
1250 : * away. Chances are some or all of the
1251 : * tail pages can be freed without IO.
1252 : */
1253 0 : if (!compound_mapcount(page) &&
1254 0 : split_huge_page_to_list(page,
1255 : page_list))
1256 0 : goto activate_locked;
1257 : }
1258 0 : if (!add_to_swap(page)) {
1259 0 : if (!PageTransHuge(page))
1260 0 : goto activate_locked_split;
1261 : /* Fallback to swap normal pages */
1262 0 : if (split_huge_page_to_list(page,
1263 : page_list))
1264 0 : goto activate_locked;
1265 : #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1266 0 : count_vm_event(THP_SWPOUT_FALLBACK);
1267 : #endif
1268 0 : if (!add_to_swap(page))
1269 0 : goto activate_locked_split;
1270 : }
1271 :
1272 : may_enter_fs = true;
1273 :
1274 : /* Adding to swap updated mapping */
1275 : mapping = page_mapping(page);
1276 : }
1277 0 : } else if (unlikely(PageTransHuge(page))) {
1278 : /* Split file THP */
1279 0 : if (split_huge_page_to_list(page, page_list))
1280 0 : goto keep_locked;
1281 : }
1282 :
1283 : /*
1284 : * THP may get split above, need minus tail pages and update
1285 : * nr_pages to avoid accounting tail pages twice.
1286 : *
1287 : * The tail pages that are added into swap cache successfully
1288 : * reach here.
1289 : */
1290 0 : if ((nr_pages > 1) && !PageTransHuge(page)) {
1291 0 : sc->nr_scanned -= (nr_pages - 1);
1292 0 : nr_pages = 1;
1293 : }
1294 :
1295 : /*
1296 : * The page is mapped into the page tables of one or more
1297 : * processes. Try to unmap it here.
1298 : */
1299 0 : if (page_mapped(page)) {
1300 0 : enum ttu_flags flags = TTU_BATCH_FLUSH;
1301 0 : bool was_swapbacked = PageSwapBacked(page);
1302 :
1303 0 : if (unlikely(PageTransHuge(page)))
1304 0 : flags |= TTU_SPLIT_HUGE_PMD;
1305 :
1306 0 : if (!try_to_unmap(page, flags)) {
1307 0 : stat->nr_unmap_fail += nr_pages;
1308 0 : if (!was_swapbacked && PageSwapBacked(page))
1309 0 : stat->nr_lazyfree_fail += nr_pages;
1310 0 : goto activate_locked;
1311 : }
1312 : }
1313 :
1314 0 : if (PageDirty(page)) {
1315 : /*
1316 : * Only kswapd can writeback filesystem pages
1317 : * to avoid risk of stack overflow. But avoid
1318 : * injecting inefficient single-page IO into
1319 : * flusher writeback as much as possible: only
1320 : * write pages when we've encountered many
1321 : * dirty pages, and when we've already scanned
1322 : * the rest of the LRU for clean pages and see
1323 : * the same dirty pages again (PageReclaim).
1324 : */
1325 0 : if (page_is_file_lru(page) &&
1326 0 : (!current_is_kswapd() || !PageReclaim(page) ||
1327 0 : !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1328 : /*
1329 : * Immediately reclaim when written back.
1330 : * Similar in principal to deactivate_page()
1331 : * except we already have the page isolated
1332 : * and know it's dirty
1333 : */
1334 0 : inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1335 0 : SetPageReclaim(page);
1336 :
1337 0 : goto activate_locked;
1338 : }
1339 :
1340 0 : if (references == PAGEREF_RECLAIM_CLEAN)
1341 0 : goto keep_locked;
1342 0 : if (!may_enter_fs)
1343 0 : goto keep_locked;
1344 0 : if (!sc->may_writepage)
1345 0 : goto keep_locked;
1346 :
1347 : /*
1348 : * Page is dirty. Flush the TLB if a writable entry
1349 : * potentially exists to avoid CPU writes after IO
1350 : * starts and then write it out here.
1351 : */
1352 0 : try_to_unmap_flush_dirty();
1353 0 : switch (pageout(page, mapping)) {
1354 0 : case PAGE_KEEP:
1355 0 : goto keep_locked;
1356 0 : case PAGE_ACTIVATE:
1357 0 : goto activate_locked;
1358 : case PAGE_SUCCESS:
1359 0 : stat->nr_pageout += thp_nr_pages(page);
1360 :
1361 0 : if (PageWriteback(page))
1362 0 : goto keep;
1363 0 : if (PageDirty(page))
1364 0 : goto keep;
1365 :
1366 : /*
1367 : * A synchronous write - probably a ramdisk. Go
1368 : * ahead and try to reclaim the page.
1369 : */
1370 0 : if (!trylock_page(page))
1371 0 : goto keep;
1372 0 : if (PageDirty(page) || PageWriteback(page))
1373 0 : goto keep_locked;
1374 0 : mapping = page_mapping(page);
1375 0 : fallthrough;
1376 0 : case PAGE_CLEAN:
1377 0 : ; /* try to free the page below */
1378 : }
1379 : }
1380 :
1381 : /*
1382 : * If the page has buffers, try to free the buffer mappings
1383 : * associated with this page. If we succeed we try to free
1384 : * the page as well.
1385 : *
1386 : * We do this even if the page is PageDirty().
1387 : * try_to_release_page() does not perform I/O, but it is
1388 : * possible for a page to have PageDirty set, but it is actually
1389 : * clean (all its buffers are clean). This happens if the
1390 : * buffers were written out directly, with submit_bh(). ext3
1391 : * will do this, as well as the blockdev mapping.
1392 : * try_to_release_page() will discover that cleanness and will
1393 : * drop the buffers and mark the page clean - it can be freed.
1394 : *
1395 : * Rarely, pages can have buffers and no ->mapping. These are
1396 : * the pages which were not successfully invalidated in
1397 : * truncate_cleanup_page(). We try to drop those buffers here
1398 : * and if that worked, and the page is no longer mapped into
1399 : * process address space (page_count == 1) it can be freed.
1400 : * Otherwise, leave the page on the LRU so it is swappable.
1401 : */
1402 0 : if (page_has_private(page)) {
1403 0 : if (!try_to_release_page(page, sc->gfp_mask))
1404 0 : goto activate_locked;
1405 0 : if (!mapping && page_count(page) == 1) {
1406 0 : unlock_page(page);
1407 0 : if (put_page_testzero(page))
1408 0 : goto free_it;
1409 : else {
1410 : /*
1411 : * rare race with speculative reference.
1412 : * the speculative reference will free
1413 : * this page shortly, so we may
1414 : * increment nr_reclaimed here (and
1415 : * leave it off the LRU).
1416 : */
1417 0 : nr_reclaimed++;
1418 0 : continue;
1419 : }
1420 : }
1421 : }
1422 :
1423 0 : if (PageAnon(page) && !PageSwapBacked(page)) {
1424 : /* follow __remove_mapping for reference */
1425 0 : if (!page_ref_freeze(page, 1))
1426 0 : goto keep_locked;
1427 0 : if (PageDirty(page)) {
1428 0 : page_ref_unfreeze(page, 1);
1429 0 : goto keep_locked;
1430 : }
1431 :
1432 0 : count_vm_event(PGLAZYFREED);
1433 0 : count_memcg_page_event(page, PGLAZYFREED);
1434 0 : } else if (!mapping || !__remove_mapping(mapping, page, true,
1435 : sc->target_mem_cgroup))
1436 0 : goto keep_locked;
1437 :
1438 0 : unlock_page(page);
1439 0 : free_it:
1440 : /*
1441 : * THP may get swapped out in a whole, need account
1442 : * all base pages.
1443 : */
1444 0 : nr_reclaimed += nr_pages;
1445 :
1446 : /*
1447 : * Is there need to periodically free_page_list? It would
1448 : * appear not as the counts should be low
1449 : */
1450 0 : if (unlikely(PageTransHuge(page)))
1451 0 : destroy_compound_page(page);
1452 : else
1453 0 : list_add(&page->lru, &free_pages);
1454 0 : continue;
1455 :
1456 0 : activate_locked_split:
1457 : /*
1458 : * The tail pages that are failed to add into swap cache
1459 : * reach here. Fixup nr_scanned and nr_pages.
1460 : */
1461 0 : if (nr_pages > 1) {
1462 0 : sc->nr_scanned -= (nr_pages - 1);
1463 0 : nr_pages = 1;
1464 : }
1465 0 : activate_locked:
1466 : /* Not a candidate for swapping, so reclaim swap space. */
1467 0 : if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1468 : PageMlocked(page)))
1469 0 : try_to_free_swap(page);
1470 0 : VM_BUG_ON_PAGE(PageActive(page), page);
1471 0 : if (!PageMlocked(page)) {
1472 0 : int type = page_is_file_lru(page);
1473 0 : SetPageActive(page);
1474 0 : stat->nr_activate[type] += nr_pages;
1475 0 : count_memcg_page_event(page, PGACTIVATE);
1476 : }
1477 0 : keep_locked:
1478 0 : unlock_page(page);
1479 0 : keep:
1480 0 : list_add(&page->lru, &ret_pages);
1481 0 : VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1482 : }
1483 :
1484 0 : pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1485 :
1486 0 : mem_cgroup_uncharge_list(&free_pages);
1487 0 : try_to_unmap_flush();
1488 0 : free_unref_page_list(&free_pages);
1489 :
1490 0 : list_splice(&ret_pages, page_list);
1491 0 : count_vm_events(PGACTIVATE, pgactivate);
1492 :
1493 0 : return nr_reclaimed;
1494 : }
1495 :
1496 0 : unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1497 : struct list_head *page_list)
1498 : {
1499 0 : struct scan_control sc = {
1500 : .gfp_mask = GFP_KERNEL,
1501 : .priority = DEF_PRIORITY,
1502 : .may_unmap = 1,
1503 : };
1504 0 : struct reclaim_stat stat;
1505 0 : unsigned int nr_reclaimed;
1506 0 : struct page *page, *next;
1507 0 : LIST_HEAD(clean_pages);
1508 :
1509 0 : list_for_each_entry_safe(page, next, page_list, lru) {
1510 0 : if (page_is_file_lru(page) && !PageDirty(page) &&
1511 0 : !__PageMovable(page) && !PageUnevictable(page)) {
1512 0 : ClearPageActive(page);
1513 0 : list_move(&page->lru, &clean_pages);
1514 : }
1515 : }
1516 :
1517 0 : nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1518 : &stat, true);
1519 0 : list_splice(&clean_pages, page_list);
1520 0 : mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1521 0 : -(long)nr_reclaimed);
1522 : /*
1523 : * Since lazyfree pages are isolated from file LRU from the beginning,
1524 : * they will rotate back to anonymous LRU in the end if it failed to
1525 : * discard so isolated count will be mismatched.
1526 : * Compensate the isolated count for both LRU lists.
1527 : */
1528 0 : mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1529 0 : stat.nr_lazyfree_fail);
1530 0 : mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1531 0 : -(long)stat.nr_lazyfree_fail);
1532 0 : return nr_reclaimed;
1533 : }
1534 :
1535 : /*
1536 : * Attempt to remove the specified page from its LRU. Only take this page
1537 : * if it is of the appropriate PageActive status. Pages which are being
1538 : * freed elsewhere are also ignored.
1539 : *
1540 : * page: page to consider
1541 : * mode: one of the LRU isolation modes defined above
1542 : *
1543 : * returns true on success, false on failure.
1544 : */
1545 0 : bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1546 : {
1547 : /* Only take pages on the LRU. */
1548 0 : if (!PageLRU(page))
1549 : return false;
1550 :
1551 : /* Compaction should not handle unevictable pages but CMA can do so */
1552 0 : if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1553 : return false;
1554 :
1555 : /*
1556 : * To minimise LRU disruption, the caller can indicate that it only
1557 : * wants to isolate pages it will be able to operate on without
1558 : * blocking - clean pages for the most part.
1559 : *
1560 : * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1561 : * that it is possible to migrate without blocking
1562 : */
1563 0 : if (mode & ISOLATE_ASYNC_MIGRATE) {
1564 : /* All the caller can do on PageWriteback is block */
1565 0 : if (PageWriteback(page))
1566 : return false;
1567 :
1568 0 : if (PageDirty(page)) {
1569 0 : struct address_space *mapping;
1570 0 : bool migrate_dirty;
1571 :
1572 : /*
1573 : * Only pages without mappings or that have a
1574 : * ->migratepage callback are possible to migrate
1575 : * without blocking. However, we can be racing with
1576 : * truncation so it's necessary to lock the page
1577 : * to stabilise the mapping as truncation holds
1578 : * the page lock until after the page is removed
1579 : * from the page cache.
1580 : */
1581 0 : if (!trylock_page(page))
1582 : return false;
1583 :
1584 0 : mapping = page_mapping(page);
1585 0 : migrate_dirty = !mapping || mapping->a_ops->migratepage;
1586 0 : unlock_page(page);
1587 0 : if (!migrate_dirty)
1588 : return false;
1589 : }
1590 : }
1591 :
1592 0 : if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1593 0 : return false;
1594 :
1595 : return true;
1596 : }
1597 :
1598 : /*
1599 : * Update LRU sizes after isolating pages. The LRU size updates must
1600 : * be complete before mem_cgroup_update_lru_size due to a sanity check.
1601 : */
1602 0 : static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1603 : enum lru_list lru, unsigned long *nr_zone_taken)
1604 : {
1605 0 : int zid;
1606 :
1607 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1608 0 : if (!nr_zone_taken[zid])
1609 0 : continue;
1610 :
1611 0 : update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1612 : }
1613 :
1614 : }
1615 :
1616 : /**
1617 : * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1618 : *
1619 : * lruvec->lru_lock is heavily contended. Some of the functions that
1620 : * shrink the lists perform better by taking out a batch of pages
1621 : * and working on them outside the LRU lock.
1622 : *
1623 : * For pagecache intensive workloads, this function is the hottest
1624 : * spot in the kernel (apart from copy_*_user functions).
1625 : *
1626 : * Lru_lock must be held before calling this function.
1627 : *
1628 : * @nr_to_scan: The number of eligible pages to look through on the list.
1629 : * @lruvec: The LRU vector to pull pages from.
1630 : * @dst: The temp list to put pages on to.
1631 : * @nr_scanned: The number of pages that were scanned.
1632 : * @sc: The scan_control struct for this reclaim session
1633 : * @lru: LRU list id for isolating
1634 : *
1635 : * returns how many pages were moved onto *@dst.
1636 : */
1637 0 : static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1638 : struct lruvec *lruvec, struct list_head *dst,
1639 : unsigned long *nr_scanned, struct scan_control *sc,
1640 : enum lru_list lru)
1641 : {
1642 0 : struct list_head *src = &lruvec->lists[lru];
1643 0 : unsigned long nr_taken = 0;
1644 0 : unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1645 0 : unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1646 0 : unsigned long skipped = 0;
1647 0 : unsigned long scan, total_scan, nr_pages;
1648 0 : LIST_HEAD(pages_skipped);
1649 0 : isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1650 :
1651 0 : total_scan = 0;
1652 0 : scan = 0;
1653 0 : while (scan < nr_to_scan && !list_empty(src)) {
1654 0 : struct page *page;
1655 :
1656 0 : page = lru_to_page(src);
1657 0 : prefetchw_prev_lru_page(page, src, flags);
1658 :
1659 0 : nr_pages = compound_nr(page);
1660 0 : total_scan += nr_pages;
1661 :
1662 0 : if (page_zonenum(page) > sc->reclaim_idx) {
1663 0 : list_move(&page->lru, &pages_skipped);
1664 0 : nr_skipped[page_zonenum(page)] += nr_pages;
1665 0 : continue;
1666 : }
1667 :
1668 : /*
1669 : * Do not count skipped pages because that makes the function
1670 : * return with no isolated pages if the LRU mostly contains
1671 : * ineligible pages. This causes the VM to not reclaim any
1672 : * pages, triggering a premature OOM.
1673 : *
1674 : * Account all tail pages of THP. This would not cause
1675 : * premature OOM since __isolate_lru_page() returns -EBUSY
1676 : * only when the page is being freed somewhere else.
1677 : */
1678 0 : scan += nr_pages;
1679 0 : if (!__isolate_lru_page_prepare(page, mode)) {
1680 : /* It is being freed elsewhere */
1681 0 : list_move(&page->lru, src);
1682 0 : continue;
1683 : }
1684 : /*
1685 : * Be careful not to clear PageLRU until after we're
1686 : * sure the page is not being freed elsewhere -- the
1687 : * page release code relies on it.
1688 : */
1689 0 : if (unlikely(!get_page_unless_zero(page))) {
1690 0 : list_move(&page->lru, src);
1691 0 : continue;
1692 : }
1693 :
1694 0 : if (!TestClearPageLRU(page)) {
1695 : /* Another thread is already isolating this page */
1696 0 : put_page(page);
1697 0 : list_move(&page->lru, src);
1698 0 : continue;
1699 : }
1700 :
1701 0 : nr_taken += nr_pages;
1702 0 : nr_zone_taken[page_zonenum(page)] += nr_pages;
1703 0 : list_move(&page->lru, dst);
1704 : }
1705 :
1706 : /*
1707 : * Splice any skipped pages to the start of the LRU list. Note that
1708 : * this disrupts the LRU order when reclaiming for lower zones but
1709 : * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1710 : * scanning would soon rescan the same pages to skip and put the
1711 : * system at risk of premature OOM.
1712 : */
1713 0 : if (!list_empty(&pages_skipped)) {
1714 0 : int zid;
1715 :
1716 0 : list_splice(&pages_skipped, src);
1717 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1718 0 : if (!nr_skipped[zid])
1719 0 : continue;
1720 :
1721 0 : __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1722 0 : skipped += nr_skipped[zid];
1723 : }
1724 : }
1725 0 : *nr_scanned = total_scan;
1726 0 : trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1727 : total_scan, skipped, nr_taken, mode, lru);
1728 0 : update_lru_sizes(lruvec, lru, nr_zone_taken);
1729 0 : return nr_taken;
1730 : }
1731 :
1732 : /**
1733 : * isolate_lru_page - tries to isolate a page from its LRU list
1734 : * @page: page to isolate from its LRU list
1735 : *
1736 : * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1737 : * vmstat statistic corresponding to whatever LRU list the page was on.
1738 : *
1739 : * Returns 0 if the page was removed from an LRU list.
1740 : * Returns -EBUSY if the page was not on an LRU list.
1741 : *
1742 : * The returned page will have PageLRU() cleared. If it was found on
1743 : * the active list, it will have PageActive set. If it was found on
1744 : * the unevictable list, it will have the PageUnevictable bit set. That flag
1745 : * may need to be cleared by the caller before letting the page go.
1746 : *
1747 : * The vmstat statistic corresponding to the list on which the page was
1748 : * found will be decremented.
1749 : *
1750 : * Restrictions:
1751 : *
1752 : * (1) Must be called with an elevated refcount on the page. This is a
1753 : * fundamental difference from isolate_lru_pages (which is called
1754 : * without a stable reference).
1755 : * (2) the lru_lock must not be held.
1756 : * (3) interrupts must be enabled.
1757 : */
1758 11 : int isolate_lru_page(struct page *page)
1759 : {
1760 11 : int ret = -EBUSY;
1761 :
1762 11 : VM_BUG_ON_PAGE(!page_count(page), page);
1763 11 : WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1764 :
1765 22 : if (TestClearPageLRU(page)) {
1766 11 : struct lruvec *lruvec;
1767 :
1768 11 : get_page(page);
1769 11 : lruvec = lock_page_lruvec_irq(page);
1770 11 : del_page_from_lru_list(page, lruvec);
1771 11 : unlock_page_lruvec_irq(lruvec);
1772 11 : ret = 0;
1773 : }
1774 :
1775 11 : return ret;
1776 : }
1777 :
1778 : /*
1779 : * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1780 : * then get rescheduled. When there are massive number of tasks doing page
1781 : * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1782 : * the LRU list will go small and be scanned faster than necessary, leading to
1783 : * unnecessary swapping, thrashing and OOM.
1784 : */
1785 0 : static int too_many_isolated(struct pglist_data *pgdat, int file,
1786 : struct scan_control *sc)
1787 : {
1788 0 : unsigned long inactive, isolated;
1789 :
1790 0 : if (current_is_kswapd())
1791 : return 0;
1792 :
1793 0 : if (!writeback_throttling_sane(sc))
1794 : return 0;
1795 :
1796 0 : if (file) {
1797 0 : inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1798 0 : isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1799 : } else {
1800 0 : inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1801 0 : isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1802 : }
1803 :
1804 : /*
1805 : * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1806 : * won't get blocked by normal direct-reclaimers, forming a circular
1807 : * deadlock.
1808 : */
1809 0 : if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1810 0 : inactive >>= 3;
1811 :
1812 0 : return isolated > inactive;
1813 : }
1814 :
1815 : /*
1816 : * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
1817 : * On return, @list is reused as a list of pages to be freed by the caller.
1818 : *
1819 : * Returns the number of pages moved to the given lruvec.
1820 : */
1821 0 : static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1822 : struct list_head *list)
1823 : {
1824 0 : int nr_pages, nr_moved = 0;
1825 0 : LIST_HEAD(pages_to_free);
1826 0 : struct page *page;
1827 :
1828 0 : while (!list_empty(list)) {
1829 0 : page = lru_to_page(list);
1830 0 : VM_BUG_ON_PAGE(PageLRU(page), page);
1831 0 : list_del(&page->lru);
1832 0 : if (unlikely(!page_evictable(page))) {
1833 0 : spin_unlock_irq(&lruvec->lru_lock);
1834 0 : putback_lru_page(page);
1835 0 : spin_lock_irq(&lruvec->lru_lock);
1836 0 : continue;
1837 : }
1838 :
1839 : /*
1840 : * The SetPageLRU needs to be kept here for list integrity.
1841 : * Otherwise:
1842 : * #0 move_pages_to_lru #1 release_pages
1843 : * if !put_page_testzero
1844 : * if (put_page_testzero())
1845 : * !PageLRU //skip lru_lock
1846 : * SetPageLRU()
1847 : * list_add(&page->lru,)
1848 : * list_add(&page->lru,)
1849 : */
1850 0 : SetPageLRU(page);
1851 :
1852 0 : if (unlikely(put_page_testzero(page))) {
1853 0 : __clear_page_lru_flags(page);
1854 :
1855 0 : if (unlikely(PageCompound(page))) {
1856 0 : spin_unlock_irq(&lruvec->lru_lock);
1857 0 : destroy_compound_page(page);
1858 0 : spin_lock_irq(&lruvec->lru_lock);
1859 : } else
1860 0 : list_add(&page->lru, &pages_to_free);
1861 :
1862 0 : continue;
1863 : }
1864 :
1865 : /*
1866 : * All pages were isolated from the same lruvec (and isolation
1867 : * inhibits memcg migration).
1868 : */
1869 0 : VM_BUG_ON_PAGE(!lruvec_holds_page_lru_lock(page, lruvec), page);
1870 0 : add_page_to_lru_list(page, lruvec);
1871 0 : nr_pages = thp_nr_pages(page);
1872 0 : nr_moved += nr_pages;
1873 0 : if (PageActive(page))
1874 0 : workingset_age_nonresident(lruvec, nr_pages);
1875 : }
1876 :
1877 : /*
1878 : * To save our caller's stack, now use input list for pages to free.
1879 : */
1880 0 : list_splice(&pages_to_free, list);
1881 :
1882 0 : return nr_moved;
1883 : }
1884 :
1885 : /*
1886 : * If a kernel thread (such as nfsd for loop-back mounts) services
1887 : * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1888 : * In that case we should only throttle if the backing device it is
1889 : * writing to is congested. In other cases it is safe to throttle.
1890 : */
1891 0 : static int current_may_throttle(void)
1892 : {
1893 0 : return !(current->flags & PF_LOCAL_THROTTLE) ||
1894 0 : current->backing_dev_info == NULL ||
1895 0 : bdi_write_congested(current->backing_dev_info);
1896 : }
1897 :
1898 : /*
1899 : * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1900 : * of reclaimed pages
1901 : */
1902 : static noinline_for_stack unsigned long
1903 0 : shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1904 : struct scan_control *sc, enum lru_list lru)
1905 : {
1906 0 : LIST_HEAD(page_list);
1907 0 : unsigned long nr_scanned;
1908 0 : unsigned int nr_reclaimed = 0;
1909 0 : unsigned long nr_taken;
1910 0 : struct reclaim_stat stat;
1911 0 : bool file = is_file_lru(lru);
1912 0 : enum vm_event_item item;
1913 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1914 0 : bool stalled = false;
1915 :
1916 0 : while (unlikely(too_many_isolated(pgdat, file, sc))) {
1917 0 : if (stalled)
1918 : return 0;
1919 :
1920 : /* wait a bit for the reclaimer. */
1921 0 : msleep(100);
1922 0 : stalled = true;
1923 :
1924 : /* We are about to die and free our memory. Return now. */
1925 0 : if (fatal_signal_pending(current))
1926 : return SWAP_CLUSTER_MAX;
1927 : }
1928 :
1929 0 : lru_add_drain();
1930 :
1931 0 : spin_lock_irq(&lruvec->lru_lock);
1932 :
1933 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1934 : &nr_scanned, sc, lru);
1935 :
1936 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1937 0 : item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1938 0 : if (!cgroup_reclaim(sc))
1939 0 : __count_vm_events(item, nr_scanned);
1940 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1941 0 : __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1942 :
1943 0 : spin_unlock_irq(&lruvec->lru_lock);
1944 :
1945 0 : if (nr_taken == 0)
1946 : return 0;
1947 :
1948 0 : nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
1949 :
1950 0 : spin_lock_irq(&lruvec->lru_lock);
1951 0 : move_pages_to_lru(lruvec, &page_list);
1952 :
1953 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1954 0 : item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1955 0 : if (!cgroup_reclaim(sc))
1956 0 : __count_vm_events(item, nr_reclaimed);
1957 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1958 0 : __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1959 0 : spin_unlock_irq(&lruvec->lru_lock);
1960 :
1961 0 : lru_note_cost(lruvec, file, stat.nr_pageout);
1962 0 : mem_cgroup_uncharge_list(&page_list);
1963 0 : free_unref_page_list(&page_list);
1964 :
1965 : /*
1966 : * If dirty pages are scanned that are not queued for IO, it
1967 : * implies that flushers are not doing their job. This can
1968 : * happen when memory pressure pushes dirty pages to the end of
1969 : * the LRU before the dirty limits are breached and the dirty
1970 : * data has expired. It can also happen when the proportion of
1971 : * dirty pages grows not through writes but through memory
1972 : * pressure reclaiming all the clean cache. And in some cases,
1973 : * the flushers simply cannot keep up with the allocation
1974 : * rate. Nudge the flusher threads in case they are asleep.
1975 : */
1976 0 : if (stat.nr_unqueued_dirty == nr_taken)
1977 0 : wakeup_flusher_threads(WB_REASON_VMSCAN);
1978 :
1979 0 : sc->nr.dirty += stat.nr_dirty;
1980 0 : sc->nr.congested += stat.nr_congested;
1981 0 : sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1982 0 : sc->nr.writeback += stat.nr_writeback;
1983 0 : sc->nr.immediate += stat.nr_immediate;
1984 0 : sc->nr.taken += nr_taken;
1985 0 : if (file)
1986 0 : sc->nr.file_taken += nr_taken;
1987 :
1988 0 : trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1989 0 : nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1990 0 : return nr_reclaimed;
1991 : }
1992 :
1993 : /*
1994 : * shrink_active_list() moves pages from the active LRU to the inactive LRU.
1995 : *
1996 : * We move them the other way if the page is referenced by one or more
1997 : * processes.
1998 : *
1999 : * If the pages are mostly unmapped, the processing is fast and it is
2000 : * appropriate to hold lru_lock across the whole operation. But if
2001 : * the pages are mapped, the processing is slow (page_referenced()), so
2002 : * we should drop lru_lock around each page. It's impossible to balance
2003 : * this, so instead we remove the pages from the LRU while processing them.
2004 : * It is safe to rely on PG_active against the non-LRU pages in here because
2005 : * nobody will play with that bit on a non-LRU page.
2006 : *
2007 : * The downside is that we have to touch page->_refcount against each page.
2008 : * But we had to alter page->flags anyway.
2009 : */
2010 0 : static void shrink_active_list(unsigned long nr_to_scan,
2011 : struct lruvec *lruvec,
2012 : struct scan_control *sc,
2013 : enum lru_list lru)
2014 : {
2015 0 : unsigned long nr_taken;
2016 0 : unsigned long nr_scanned;
2017 0 : unsigned long vm_flags;
2018 0 : LIST_HEAD(l_hold); /* The pages which were snipped off */
2019 0 : LIST_HEAD(l_active);
2020 0 : LIST_HEAD(l_inactive);
2021 0 : struct page *page;
2022 0 : unsigned nr_deactivate, nr_activate;
2023 0 : unsigned nr_rotated = 0;
2024 0 : int file = is_file_lru(lru);
2025 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2026 :
2027 0 : lru_add_drain();
2028 :
2029 0 : spin_lock_irq(&lruvec->lru_lock);
2030 :
2031 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2032 : &nr_scanned, sc, lru);
2033 :
2034 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2035 :
2036 0 : if (!cgroup_reclaim(sc))
2037 0 : __count_vm_events(PGREFILL, nr_scanned);
2038 0 : __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2039 :
2040 0 : spin_unlock_irq(&lruvec->lru_lock);
2041 :
2042 0 : while (!list_empty(&l_hold)) {
2043 0 : cond_resched();
2044 0 : page = lru_to_page(&l_hold);
2045 0 : list_del(&page->lru);
2046 :
2047 0 : if (unlikely(!page_evictable(page))) {
2048 0 : putback_lru_page(page);
2049 0 : continue;
2050 : }
2051 :
2052 0 : if (unlikely(buffer_heads_over_limit)) {
2053 0 : if (page_has_private(page) && trylock_page(page)) {
2054 0 : if (page_has_private(page))
2055 0 : try_to_release_page(page, 0);
2056 0 : unlock_page(page);
2057 : }
2058 : }
2059 :
2060 0 : if (page_referenced(page, 0, sc->target_mem_cgroup,
2061 : &vm_flags)) {
2062 : /*
2063 : * Identify referenced, file-backed active pages and
2064 : * give them one more trip around the active list. So
2065 : * that executable code get better chances to stay in
2066 : * memory under moderate memory pressure. Anon pages
2067 : * are not likely to be evicted by use-once streaming
2068 : * IO, plus JVM can create lots of anon VM_EXEC pages,
2069 : * so we ignore them here.
2070 : */
2071 0 : if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2072 0 : nr_rotated += thp_nr_pages(page);
2073 0 : list_add(&page->lru, &l_active);
2074 0 : continue;
2075 : }
2076 : }
2077 :
2078 0 : ClearPageActive(page); /* we are de-activating */
2079 0 : SetPageWorkingset(page);
2080 0 : list_add(&page->lru, &l_inactive);
2081 : }
2082 :
2083 : /*
2084 : * Move pages back to the lru list.
2085 : */
2086 0 : spin_lock_irq(&lruvec->lru_lock);
2087 :
2088 0 : nr_activate = move_pages_to_lru(lruvec, &l_active);
2089 0 : nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2090 : /* Keep all free pages in l_active list */
2091 0 : list_splice(&l_inactive, &l_active);
2092 :
2093 0 : __count_vm_events(PGDEACTIVATE, nr_deactivate);
2094 0 : __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2095 :
2096 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2097 0 : spin_unlock_irq(&lruvec->lru_lock);
2098 :
2099 0 : mem_cgroup_uncharge_list(&l_active);
2100 0 : free_unref_page_list(&l_active);
2101 0 : trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2102 0 : nr_deactivate, nr_rotated, sc->priority, file);
2103 0 : }
2104 :
2105 0 : unsigned long reclaim_pages(struct list_head *page_list)
2106 : {
2107 0 : int nid = NUMA_NO_NODE;
2108 0 : unsigned int nr_reclaimed = 0;
2109 0 : LIST_HEAD(node_page_list);
2110 0 : struct reclaim_stat dummy_stat;
2111 0 : struct page *page;
2112 0 : struct scan_control sc = {
2113 : .gfp_mask = GFP_KERNEL,
2114 : .priority = DEF_PRIORITY,
2115 : .may_writepage = 1,
2116 : .may_unmap = 1,
2117 : .may_swap = 1,
2118 : };
2119 :
2120 0 : while (!list_empty(page_list)) {
2121 0 : page = lru_to_page(page_list);
2122 0 : if (nid == NUMA_NO_NODE) {
2123 0 : nid = page_to_nid(page);
2124 0 : INIT_LIST_HEAD(&node_page_list);
2125 : }
2126 :
2127 0 : if (nid == page_to_nid(page)) {
2128 0 : ClearPageActive(page);
2129 0 : list_move(&page->lru, &node_page_list);
2130 0 : continue;
2131 : }
2132 :
2133 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2134 : NODE_DATA(nid),
2135 : &sc, &dummy_stat, false);
2136 0 : while (!list_empty(&node_page_list)) {
2137 0 : page = lru_to_page(&node_page_list);
2138 0 : list_del(&page->lru);
2139 0 : putback_lru_page(page);
2140 : }
2141 :
2142 : nid = NUMA_NO_NODE;
2143 : }
2144 :
2145 0 : if (!list_empty(&node_page_list)) {
2146 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2147 : NODE_DATA(nid),
2148 : &sc, &dummy_stat, false);
2149 0 : while (!list_empty(&node_page_list)) {
2150 0 : page = lru_to_page(&node_page_list);
2151 0 : list_del(&page->lru);
2152 0 : putback_lru_page(page);
2153 : }
2154 : }
2155 :
2156 0 : return nr_reclaimed;
2157 : }
2158 :
2159 0 : static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2160 : struct lruvec *lruvec, struct scan_control *sc)
2161 : {
2162 0 : if (is_active_lru(lru)) {
2163 0 : if (sc->may_deactivate & (1 << is_file_lru(lru)))
2164 0 : shrink_active_list(nr_to_scan, lruvec, sc, lru);
2165 : else
2166 0 : sc->skipped_deactivate = 1;
2167 0 : return 0;
2168 : }
2169 :
2170 0 : return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2171 : }
2172 :
2173 : /*
2174 : * The inactive anon list should be small enough that the VM never has
2175 : * to do too much work.
2176 : *
2177 : * The inactive file list should be small enough to leave most memory
2178 : * to the established workingset on the scan-resistant active list,
2179 : * but large enough to avoid thrashing the aggregate readahead window.
2180 : *
2181 : * Both inactive lists should also be large enough that each inactive
2182 : * page has a chance to be referenced again before it is reclaimed.
2183 : *
2184 : * If that fails and refaulting is observed, the inactive list grows.
2185 : *
2186 : * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2187 : * on this LRU, maintained by the pageout code. An inactive_ratio
2188 : * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2189 : *
2190 : * total target max
2191 : * memory ratio inactive
2192 : * -------------------------------------
2193 : * 10MB 1 5MB
2194 : * 100MB 1 50MB
2195 : * 1GB 3 250MB
2196 : * 10GB 10 0.9GB
2197 : * 100GB 31 3GB
2198 : * 1TB 101 10GB
2199 : * 10TB 320 32GB
2200 : */
2201 0 : static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2202 : {
2203 0 : enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2204 0 : unsigned long inactive, active;
2205 0 : unsigned long inactive_ratio;
2206 0 : unsigned long gb;
2207 :
2208 0 : inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2209 0 : active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2210 :
2211 0 : gb = (inactive + active) >> (30 - PAGE_SHIFT);
2212 0 : if (gb)
2213 0 : inactive_ratio = int_sqrt(10 * gb);
2214 : else
2215 : inactive_ratio = 1;
2216 :
2217 0 : return inactive * inactive_ratio < active;
2218 : }
2219 :
2220 : enum scan_balance {
2221 : SCAN_EQUAL,
2222 : SCAN_FRACT,
2223 : SCAN_ANON,
2224 : SCAN_FILE,
2225 : };
2226 :
2227 : /*
2228 : * Determine how aggressively the anon and file LRU lists should be
2229 : * scanned. The relative value of each set of LRU lists is determined
2230 : * by looking at the fraction of the pages scanned we did rotate back
2231 : * onto the active list instead of evict.
2232 : *
2233 : * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2234 : * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2235 : */
2236 0 : static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2237 : unsigned long *nr)
2238 : {
2239 0 : struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2240 0 : unsigned long anon_cost, file_cost, total_cost;
2241 0 : int swappiness = mem_cgroup_swappiness(memcg);
2242 0 : u64 fraction[ANON_AND_FILE];
2243 0 : u64 denominator = 0; /* gcc */
2244 0 : enum scan_balance scan_balance;
2245 0 : unsigned long ap, fp;
2246 0 : enum lru_list lru;
2247 :
2248 : /* If we have no swap space, do not bother scanning anon pages. */
2249 0 : if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2250 0 : scan_balance = SCAN_FILE;
2251 0 : goto out;
2252 : }
2253 :
2254 : /*
2255 : * Global reclaim will swap to prevent OOM even with no
2256 : * swappiness, but memcg users want to use this knob to
2257 : * disable swapping for individual groups completely when
2258 : * using the memory controller's swap limit feature would be
2259 : * too expensive.
2260 : */
2261 : if (cgroup_reclaim(sc) && !swappiness) {
2262 : scan_balance = SCAN_FILE;
2263 : goto out;
2264 : }
2265 :
2266 : /*
2267 : * Do not apply any pressure balancing cleverness when the
2268 : * system is close to OOM, scan both anon and file equally
2269 : * (unless the swappiness setting disagrees with swapping).
2270 : */
2271 : if (!sc->priority && swappiness) {
2272 : scan_balance = SCAN_EQUAL;
2273 : goto out;
2274 : }
2275 :
2276 : /*
2277 : * If the system is almost out of file pages, force-scan anon.
2278 : */
2279 : if (sc->file_is_tiny) {
2280 : scan_balance = SCAN_ANON;
2281 : goto out;
2282 : }
2283 :
2284 : /*
2285 : * If there is enough inactive page cache, we do not reclaim
2286 : * anything from the anonymous working right now.
2287 : */
2288 : if (sc->cache_trim_mode) {
2289 : scan_balance = SCAN_FILE;
2290 : goto out;
2291 : }
2292 :
2293 : scan_balance = SCAN_FRACT;
2294 : /*
2295 : * Calculate the pressure balance between anon and file pages.
2296 : *
2297 : * The amount of pressure we put on each LRU is inversely
2298 : * proportional to the cost of reclaiming each list, as
2299 : * determined by the share of pages that are refaulting, times
2300 : * the relative IO cost of bringing back a swapped out
2301 : * anonymous page vs reloading a filesystem page (swappiness).
2302 : *
2303 : * Although we limit that influence to ensure no list gets
2304 : * left behind completely: at least a third of the pressure is
2305 : * applied, before swappiness.
2306 : *
2307 : * With swappiness at 100, anon and file have equal IO cost.
2308 : */
2309 : total_cost = sc->anon_cost + sc->file_cost;
2310 : anon_cost = total_cost + sc->anon_cost;
2311 : file_cost = total_cost + sc->file_cost;
2312 : total_cost = anon_cost + file_cost;
2313 :
2314 : ap = swappiness * (total_cost + 1);
2315 : ap /= anon_cost + 1;
2316 :
2317 : fp = (200 - swappiness) * (total_cost + 1);
2318 : fp /= file_cost + 1;
2319 :
2320 : fraction[0] = ap;
2321 : fraction[1] = fp;
2322 : denominator = ap + fp;
2323 0 : out:
2324 0 : for_each_evictable_lru(lru) {
2325 0 : int file = is_file_lru(lru);
2326 0 : unsigned long lruvec_size;
2327 0 : unsigned long scan;
2328 0 : unsigned long protection;
2329 :
2330 0 : lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2331 0 : protection = mem_cgroup_protection(sc->target_mem_cgroup,
2332 : memcg,
2333 0 : sc->memcg_low_reclaim);
2334 :
2335 0 : if (protection) {
2336 : /*
2337 : * Scale a cgroup's reclaim pressure by proportioning
2338 : * its current usage to its memory.low or memory.min
2339 : * setting.
2340 : *
2341 : * This is important, as otherwise scanning aggression
2342 : * becomes extremely binary -- from nothing as we
2343 : * approach the memory protection threshold, to totally
2344 : * nominal as we exceed it. This results in requiring
2345 : * setting extremely liberal protection thresholds. It
2346 : * also means we simply get no protection at all if we
2347 : * set it too low, which is not ideal.
2348 : *
2349 : * If there is any protection in place, we reduce scan
2350 : * pressure by how much of the total memory used is
2351 : * within protection thresholds.
2352 : *
2353 : * There is one special case: in the first reclaim pass,
2354 : * we skip over all groups that are within their low
2355 : * protection. If that fails to reclaim enough pages to
2356 : * satisfy the reclaim goal, we come back and override
2357 : * the best-effort low protection. However, we still
2358 : * ideally want to honor how well-behaved groups are in
2359 : * that case instead of simply punishing them all
2360 : * equally. As such, we reclaim them based on how much
2361 : * memory they are using, reducing the scan pressure
2362 : * again by how much of the total memory used is under
2363 : * hard protection.
2364 : */
2365 : unsigned long cgroup_size = mem_cgroup_size(memcg);
2366 :
2367 : /* Avoid TOCTOU with earlier protection check */
2368 : cgroup_size = max(cgroup_size, protection);
2369 :
2370 : scan = lruvec_size - lruvec_size * protection /
2371 : cgroup_size;
2372 :
2373 : /*
2374 : * Minimally target SWAP_CLUSTER_MAX pages to keep
2375 : * reclaim moving forwards, avoiding decrementing
2376 : * sc->priority further than desirable.
2377 : */
2378 : scan = max(scan, SWAP_CLUSTER_MAX);
2379 : } else {
2380 0 : scan = lruvec_size;
2381 : }
2382 :
2383 0 : scan >>= sc->priority;
2384 :
2385 : /*
2386 : * If the cgroup's already been deleted, make sure to
2387 : * scrape out the remaining cache.
2388 : */
2389 0 : if (!scan && !mem_cgroup_online(memcg))
2390 : scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2391 :
2392 0 : switch (scan_balance) {
2393 : case SCAN_EQUAL:
2394 : /* Scan lists relative to size */
2395 : break;
2396 : case SCAN_FRACT:
2397 : /*
2398 : * Scan types proportional to swappiness and
2399 : * their relative recent reclaim efficiency.
2400 : * Make sure we don't miss the last page on
2401 : * the offlined memory cgroups because of a
2402 : * round-off error.
2403 : */
2404 : scan = mem_cgroup_online(memcg) ?
2405 : div64_u64(scan * fraction[file], denominator) :
2406 : DIV64_U64_ROUND_UP(scan * fraction[file],
2407 : denominator);
2408 : break;
2409 : case SCAN_FILE:
2410 : case SCAN_ANON:
2411 : /* Scan one type exclusively */
2412 0 : if ((scan_balance == SCAN_FILE) != file)
2413 0 : scan = 0;
2414 : break;
2415 : default:
2416 : /* Look ma, no brain */
2417 0 : BUG();
2418 : }
2419 :
2420 0 : nr[lru] = scan;
2421 : }
2422 0 : }
2423 :
2424 0 : static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2425 : {
2426 0 : unsigned long nr[NR_LRU_LISTS];
2427 0 : unsigned long targets[NR_LRU_LISTS];
2428 0 : unsigned long nr_to_scan;
2429 0 : enum lru_list lru;
2430 0 : unsigned long nr_reclaimed = 0;
2431 0 : unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2432 0 : struct blk_plug plug;
2433 0 : bool scan_adjusted;
2434 :
2435 0 : get_scan_count(lruvec, sc, nr);
2436 :
2437 : /* Record the original scan target for proportional adjustments later */
2438 0 : memcpy(targets, nr, sizeof(nr));
2439 :
2440 : /*
2441 : * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2442 : * event that can occur when there is little memory pressure e.g.
2443 : * multiple streaming readers/writers. Hence, we do not abort scanning
2444 : * when the requested number of pages are reclaimed when scanning at
2445 : * DEF_PRIORITY on the assumption that the fact we are direct
2446 : * reclaiming implies that kswapd is not keeping up and it is best to
2447 : * do a batch of work at once. For memcg reclaim one check is made to
2448 : * abort proportional reclaim if either the file or anon lru has already
2449 : * dropped to zero at the first pass.
2450 : */
2451 0 : scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2452 0 : sc->priority == DEF_PRIORITY);
2453 :
2454 0 : blk_start_plug(&plug);
2455 0 : while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2456 0 : nr[LRU_INACTIVE_FILE]) {
2457 : unsigned long nr_anon, nr_file, percentage;
2458 : unsigned long nr_scanned;
2459 :
2460 0 : for_each_evictable_lru(lru) {
2461 0 : if (nr[lru]) {
2462 0 : nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2463 0 : nr[lru] -= nr_to_scan;
2464 :
2465 0 : nr_reclaimed += shrink_list(lru, nr_to_scan,
2466 : lruvec, sc);
2467 : }
2468 : }
2469 :
2470 0 : cond_resched();
2471 :
2472 0 : if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2473 0 : continue;
2474 :
2475 : /*
2476 : * For kswapd and memcg, reclaim at least the number of pages
2477 : * requested. Ensure that the anon and file LRUs are scanned
2478 : * proportionally what was requested by get_scan_count(). We
2479 : * stop reclaiming one LRU and reduce the amount scanning
2480 : * proportional to the original scan target.
2481 : */
2482 0 : nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2483 0 : nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2484 :
2485 : /*
2486 : * It's just vindictive to attack the larger once the smaller
2487 : * has gone to zero. And given the way we stop scanning the
2488 : * smaller below, this makes sure that we only make one nudge
2489 : * towards proportionality once we've got nr_to_reclaim.
2490 : */
2491 0 : if (!nr_file || !nr_anon)
2492 : break;
2493 :
2494 0 : if (nr_file > nr_anon) {
2495 0 : unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2496 0 : targets[LRU_ACTIVE_ANON] + 1;
2497 0 : lru = LRU_BASE;
2498 0 : percentage = nr_anon * 100 / scan_target;
2499 : } else {
2500 0 : unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2501 0 : targets[LRU_ACTIVE_FILE] + 1;
2502 0 : lru = LRU_FILE;
2503 0 : percentage = nr_file * 100 / scan_target;
2504 : }
2505 :
2506 : /* Stop scanning the smaller of the LRU */
2507 0 : nr[lru] = 0;
2508 0 : nr[lru + LRU_ACTIVE] = 0;
2509 :
2510 : /*
2511 : * Recalculate the other LRU scan count based on its original
2512 : * scan target and the percentage scanning already complete
2513 : */
2514 0 : lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2515 0 : nr_scanned = targets[lru] - nr[lru];
2516 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2517 0 : nr[lru] -= min(nr[lru], nr_scanned);
2518 :
2519 0 : lru += LRU_ACTIVE;
2520 0 : nr_scanned = targets[lru] - nr[lru];
2521 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2522 0 : nr[lru] -= min(nr[lru], nr_scanned);
2523 :
2524 0 : scan_adjusted = true;
2525 : }
2526 0 : blk_finish_plug(&plug);
2527 0 : sc->nr_reclaimed += nr_reclaimed;
2528 :
2529 : /*
2530 : * Even if we did not try to evict anon pages at all, we want to
2531 : * rebalance the anon lru active/inactive ratio.
2532 : */
2533 0 : if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2534 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2535 : sc, LRU_ACTIVE_ANON);
2536 0 : }
2537 :
2538 : /* Use reclaim/compaction for costly allocs or under memory pressure */
2539 0 : static bool in_reclaim_compaction(struct scan_control *sc)
2540 : {
2541 0 : if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2542 0 : (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2543 0 : sc->priority < DEF_PRIORITY - 2))
2544 0 : return true;
2545 :
2546 : return false;
2547 : }
2548 :
2549 : /*
2550 : * Reclaim/compaction is used for high-order allocation requests. It reclaims
2551 : * order-0 pages before compacting the zone. should_continue_reclaim() returns
2552 : * true if more pages should be reclaimed such that when the page allocator
2553 : * calls try_to_compact_pages() that it will have enough free pages to succeed.
2554 : * It will give up earlier than that if there is difficulty reclaiming pages.
2555 : */
2556 0 : static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2557 : unsigned long nr_reclaimed,
2558 : struct scan_control *sc)
2559 : {
2560 0 : unsigned long pages_for_compaction;
2561 0 : unsigned long inactive_lru_pages;
2562 0 : int z;
2563 :
2564 : /* If not in reclaim/compaction mode, stop */
2565 0 : if (!in_reclaim_compaction(sc))
2566 : return false;
2567 :
2568 : /*
2569 : * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2570 : * number of pages that were scanned. This will return to the caller
2571 : * with the risk reclaim/compaction and the resulting allocation attempt
2572 : * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2573 : * allocations through requiring that the full LRU list has been scanned
2574 : * first, by assuming that zero delta of sc->nr_scanned means full LRU
2575 : * scan, but that approximation was wrong, and there were corner cases
2576 : * where always a non-zero amount of pages were scanned.
2577 : */
2578 0 : if (!nr_reclaimed)
2579 : return false;
2580 :
2581 : /* If compaction would go ahead or the allocation would succeed, stop */
2582 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
2583 0 : struct zone *zone = &pgdat->node_zones[z];
2584 0 : if (!managed_zone(zone))
2585 0 : continue;
2586 :
2587 0 : switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2588 : case COMPACT_SUCCESS:
2589 : case COMPACT_CONTINUE:
2590 : return false;
2591 0 : default:
2592 : /* check next zone */
2593 0 : ;
2594 : }
2595 : }
2596 :
2597 : /*
2598 : * If we have not reclaimed enough pages for compaction and the
2599 : * inactive lists are large enough, continue reclaiming
2600 : */
2601 0 : pages_for_compaction = compact_gap(sc->order);
2602 0 : inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2603 0 : if (get_nr_swap_pages() > 0)
2604 : inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2605 :
2606 0 : return inactive_lru_pages > pages_for_compaction;
2607 : }
2608 :
2609 0 : static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2610 : {
2611 0 : struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2612 0 : struct mem_cgroup *memcg;
2613 :
2614 0 : memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2615 0 : do {
2616 0 : struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2617 0 : unsigned long reclaimed;
2618 0 : unsigned long scanned;
2619 :
2620 : /*
2621 : * This loop can become CPU-bound when target memcgs
2622 : * aren't eligible for reclaim - either because they
2623 : * don't have any reclaimable pages, or because their
2624 : * memory is explicitly protected. Avoid soft lockups.
2625 : */
2626 0 : cond_resched();
2627 :
2628 0 : mem_cgroup_calculate_protection(target_memcg, memcg);
2629 :
2630 0 : if (mem_cgroup_below_min(memcg)) {
2631 : /*
2632 : * Hard protection.
2633 : * If there is no reclaimable memory, OOM.
2634 : */
2635 : continue;
2636 0 : } else if (mem_cgroup_below_low(memcg)) {
2637 : /*
2638 : * Soft protection.
2639 : * Respect the protection only as long as
2640 : * there is an unprotected supply
2641 : * of reclaimable memory from other cgroups.
2642 : */
2643 : if (!sc->memcg_low_reclaim) {
2644 : sc->memcg_low_skipped = 1;
2645 : continue;
2646 : }
2647 0 : memcg_memory_event(memcg, MEMCG_LOW);
2648 : }
2649 :
2650 0 : reclaimed = sc->nr_reclaimed;
2651 0 : scanned = sc->nr_scanned;
2652 :
2653 0 : shrink_lruvec(lruvec, sc);
2654 :
2655 0 : shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2656 0 : sc->priority);
2657 :
2658 : /* Record the group's reclaim efficiency */
2659 0 : vmpressure(sc->gfp_mask, memcg, false,
2660 0 : sc->nr_scanned - scanned,
2661 0 : sc->nr_reclaimed - reclaimed);
2662 :
2663 0 : } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2664 0 : }
2665 :
2666 0 : static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2667 : {
2668 0 : struct reclaim_state *reclaim_state = current->reclaim_state;
2669 0 : unsigned long nr_reclaimed, nr_scanned;
2670 0 : struct lruvec *target_lruvec;
2671 0 : bool reclaimable = false;
2672 0 : unsigned long file;
2673 :
2674 0 : target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2675 :
2676 0 : again:
2677 0 : memset(&sc->nr, 0, sizeof(sc->nr));
2678 :
2679 0 : nr_reclaimed = sc->nr_reclaimed;
2680 0 : nr_scanned = sc->nr_scanned;
2681 :
2682 : /*
2683 : * Determine the scan balance between anon and file LRUs.
2684 : */
2685 0 : spin_lock_irq(&target_lruvec->lru_lock);
2686 0 : sc->anon_cost = target_lruvec->anon_cost;
2687 0 : sc->file_cost = target_lruvec->file_cost;
2688 0 : spin_unlock_irq(&target_lruvec->lru_lock);
2689 :
2690 : /*
2691 : * Target desirable inactive:active list ratios for the anon
2692 : * and file LRU lists.
2693 : */
2694 0 : if (!sc->force_deactivate) {
2695 0 : unsigned long refaults;
2696 :
2697 0 : refaults = lruvec_page_state(target_lruvec,
2698 : WORKINGSET_ACTIVATE_ANON);
2699 0 : if (refaults != target_lruvec->refaults[0] ||
2700 0 : inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2701 0 : sc->may_deactivate |= DEACTIVATE_ANON;
2702 : else
2703 0 : sc->may_deactivate &= ~DEACTIVATE_ANON;
2704 :
2705 : /*
2706 : * When refaults are being observed, it means a new
2707 : * workingset is being established. Deactivate to get
2708 : * rid of any stale active pages quickly.
2709 : */
2710 0 : refaults = lruvec_page_state(target_lruvec,
2711 : WORKINGSET_ACTIVATE_FILE);
2712 0 : if (refaults != target_lruvec->refaults[1] ||
2713 0 : inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2714 0 : sc->may_deactivate |= DEACTIVATE_FILE;
2715 : else
2716 0 : sc->may_deactivate &= ~DEACTIVATE_FILE;
2717 : } else
2718 0 : sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2719 :
2720 : /*
2721 : * If we have plenty of inactive file pages that aren't
2722 : * thrashing, try to reclaim those first before touching
2723 : * anonymous pages.
2724 : */
2725 0 : file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2726 0 : if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2727 0 : sc->cache_trim_mode = 1;
2728 : else
2729 0 : sc->cache_trim_mode = 0;
2730 :
2731 : /*
2732 : * Prevent the reclaimer from falling into the cache trap: as
2733 : * cache pages start out inactive, every cache fault will tip
2734 : * the scan balance towards the file LRU. And as the file LRU
2735 : * shrinks, so does the window for rotation from references.
2736 : * This means we have a runaway feedback loop where a tiny
2737 : * thrashing file LRU becomes infinitely more attractive than
2738 : * anon pages. Try to detect this based on file LRU size.
2739 : */
2740 0 : if (!cgroup_reclaim(sc)) {
2741 0 : unsigned long total_high_wmark = 0;
2742 0 : unsigned long free, anon;
2743 0 : int z;
2744 :
2745 0 : free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2746 0 : file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2747 0 : node_page_state(pgdat, NR_INACTIVE_FILE);
2748 :
2749 0 : for (z = 0; z < MAX_NR_ZONES; z++) {
2750 0 : struct zone *zone = &pgdat->node_zones[z];
2751 0 : if (!managed_zone(zone))
2752 0 : continue;
2753 :
2754 0 : total_high_wmark += high_wmark_pages(zone);
2755 : }
2756 :
2757 : /*
2758 : * Consider anon: if that's low too, this isn't a
2759 : * runaway file reclaim problem, but rather just
2760 : * extreme pressure. Reclaim as per usual then.
2761 : */
2762 0 : anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2763 :
2764 0 : sc->file_is_tiny =
2765 0 : file + free <= total_high_wmark &&
2766 0 : !(sc->may_deactivate & DEACTIVATE_ANON) &&
2767 0 : anon >> sc->priority;
2768 : }
2769 :
2770 0 : shrink_node_memcgs(pgdat, sc);
2771 :
2772 0 : if (reclaim_state) {
2773 0 : sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2774 0 : reclaim_state->reclaimed_slab = 0;
2775 : }
2776 :
2777 : /* Record the subtree's reclaim efficiency */
2778 0 : vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2779 0 : sc->nr_scanned - nr_scanned,
2780 0 : sc->nr_reclaimed - nr_reclaimed);
2781 :
2782 0 : if (sc->nr_reclaimed - nr_reclaimed)
2783 0 : reclaimable = true;
2784 :
2785 0 : if (current_is_kswapd()) {
2786 : /*
2787 : * If reclaim is isolating dirty pages under writeback,
2788 : * it implies that the long-lived page allocation rate
2789 : * is exceeding the page laundering rate. Either the
2790 : * global limits are not being effective at throttling
2791 : * processes due to the page distribution throughout
2792 : * zones or there is heavy usage of a slow backing
2793 : * device. The only option is to throttle from reclaim
2794 : * context which is not ideal as there is no guarantee
2795 : * the dirtying process is throttled in the same way
2796 : * balance_dirty_pages() manages.
2797 : *
2798 : * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2799 : * count the number of pages under pages flagged for
2800 : * immediate reclaim and stall if any are encountered
2801 : * in the nr_immediate check below.
2802 : */
2803 0 : if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2804 0 : set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2805 :
2806 : /* Allow kswapd to start writing pages during reclaim.*/
2807 0 : if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2808 0 : set_bit(PGDAT_DIRTY, &pgdat->flags);
2809 :
2810 : /*
2811 : * If kswapd scans pages marked for immediate
2812 : * reclaim and under writeback (nr_immediate), it
2813 : * implies that pages are cycling through the LRU
2814 : * faster than they are written so also forcibly stall.
2815 : */
2816 0 : if (sc->nr.immediate)
2817 0 : congestion_wait(BLK_RW_ASYNC, HZ/10);
2818 : }
2819 :
2820 : /*
2821 : * Tag a node/memcg as congested if all the dirty pages
2822 : * scanned were backed by a congested BDI and
2823 : * wait_iff_congested will stall.
2824 : *
2825 : * Legacy memcg will stall in page writeback so avoid forcibly
2826 : * stalling in wait_iff_congested().
2827 : */
2828 0 : if ((current_is_kswapd() ||
2829 0 : (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2830 0 : sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2831 0 : set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2832 :
2833 : /*
2834 : * Stall direct reclaim for IO completions if underlying BDIs
2835 : * and node is congested. Allow kswapd to continue until it
2836 : * starts encountering unqueued dirty pages or cycling through
2837 : * the LRU too quickly.
2838 : */
2839 0 : if (!current_is_kswapd() && current_may_throttle() &&
2840 0 : !sc->hibernation_mode &&
2841 0 : test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2842 0 : wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2843 :
2844 0 : if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2845 : sc))
2846 0 : goto again;
2847 :
2848 : /*
2849 : * Kswapd gives up on balancing particular nodes after too
2850 : * many failures to reclaim anything from them and goes to
2851 : * sleep. On reclaim progress, reset the failure counter. A
2852 : * successful direct reclaim run will revive a dormant kswapd.
2853 : */
2854 0 : if (reclaimable)
2855 0 : pgdat->kswapd_failures = 0;
2856 0 : }
2857 :
2858 : /*
2859 : * Returns true if compaction should go ahead for a costly-order request, or
2860 : * the allocation would already succeed without compaction. Return false if we
2861 : * should reclaim first.
2862 : */
2863 0 : static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2864 : {
2865 0 : unsigned long watermark;
2866 0 : enum compact_result suitable;
2867 :
2868 0 : suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2869 0 : if (suitable == COMPACT_SUCCESS)
2870 : /* Allocation should succeed already. Don't reclaim. */
2871 : return true;
2872 0 : if (suitable == COMPACT_SKIPPED)
2873 : /* Compaction cannot yet proceed. Do reclaim. */
2874 : return false;
2875 :
2876 : /*
2877 : * Compaction is already possible, but it takes time to run and there
2878 : * are potentially other callers using the pages just freed. So proceed
2879 : * with reclaim to make a buffer of free pages available to give
2880 : * compaction a reasonable chance of completing and allocating the page.
2881 : * Note that we won't actually reclaim the whole buffer in one attempt
2882 : * as the target watermark in should_continue_reclaim() is lower. But if
2883 : * we are already above the high+gap watermark, don't reclaim at all.
2884 : */
2885 0 : watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2886 :
2887 0 : return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2888 : }
2889 :
2890 : /*
2891 : * This is the direct reclaim path, for page-allocating processes. We only
2892 : * try to reclaim pages from zones which will satisfy the caller's allocation
2893 : * request.
2894 : *
2895 : * If a zone is deemed to be full of pinned pages then just give it a light
2896 : * scan then give up on it.
2897 : */
2898 0 : static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2899 : {
2900 0 : struct zoneref *z;
2901 0 : struct zone *zone;
2902 0 : unsigned long nr_soft_reclaimed;
2903 0 : unsigned long nr_soft_scanned;
2904 0 : gfp_t orig_mask;
2905 0 : pg_data_t *last_pgdat = NULL;
2906 :
2907 : /*
2908 : * If the number of buffer_heads in the machine exceeds the maximum
2909 : * allowed level, force direct reclaim to scan the highmem zone as
2910 : * highmem pages could be pinning lowmem pages storing buffer_heads
2911 : */
2912 0 : orig_mask = sc->gfp_mask;
2913 0 : if (buffer_heads_over_limit) {
2914 0 : sc->gfp_mask |= __GFP_HIGHMEM;
2915 0 : sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2916 : }
2917 :
2918 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
2919 : sc->reclaim_idx, sc->nodemask) {
2920 : /*
2921 : * Take care memory controller reclaiming has small influence
2922 : * to global LRU.
2923 : */
2924 0 : if (!cgroup_reclaim(sc)) {
2925 0 : if (!cpuset_zone_allowed(zone,
2926 : GFP_KERNEL | __GFP_HARDWALL))
2927 : continue;
2928 :
2929 : /*
2930 : * If we already have plenty of memory free for
2931 : * compaction in this zone, don't free any more.
2932 : * Even though compaction is invoked for any
2933 : * non-zero order, only frequent costly order
2934 : * reclamation is disruptive enough to become a
2935 : * noticeable problem, like transparent huge
2936 : * page allocations.
2937 : */
2938 0 : if (IS_ENABLED(CONFIG_COMPACTION) &&
2939 0 : sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2940 0 : compaction_ready(zone, sc)) {
2941 0 : sc->compaction_ready = true;
2942 0 : continue;
2943 : }
2944 :
2945 : /*
2946 : * Shrink each node in the zonelist once. If the
2947 : * zonelist is ordered by zone (not the default) then a
2948 : * node may be shrunk multiple times but in that case
2949 : * the user prefers lower zones being preserved.
2950 : */
2951 0 : if (zone->zone_pgdat == last_pgdat)
2952 0 : continue;
2953 :
2954 : /*
2955 : * This steals pages from memory cgroups over softlimit
2956 : * and returns the number of reclaimed pages and
2957 : * scanned pages. This works for global memory pressure
2958 : * and balancing, not for a memcg's limit.
2959 : */
2960 0 : nr_soft_scanned = 0;
2961 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2962 0 : sc->order, sc->gfp_mask,
2963 : &nr_soft_scanned);
2964 0 : sc->nr_reclaimed += nr_soft_reclaimed;
2965 0 : sc->nr_scanned += nr_soft_scanned;
2966 : /* need some check for avoid more shrink_zone() */
2967 : }
2968 :
2969 : /* See comment about same check for global reclaim above */
2970 0 : if (zone->zone_pgdat == last_pgdat)
2971 : continue;
2972 0 : last_pgdat = zone->zone_pgdat;
2973 0 : shrink_node(zone->zone_pgdat, sc);
2974 : }
2975 :
2976 : /*
2977 : * Restore to original mask to avoid the impact on the caller if we
2978 : * promoted it to __GFP_HIGHMEM.
2979 : */
2980 0 : sc->gfp_mask = orig_mask;
2981 0 : }
2982 :
2983 0 : static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2984 : {
2985 0 : struct lruvec *target_lruvec;
2986 0 : unsigned long refaults;
2987 :
2988 0 : target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2989 0 : refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2990 0 : target_lruvec->refaults[0] = refaults;
2991 0 : refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
2992 0 : target_lruvec->refaults[1] = refaults;
2993 0 : }
2994 :
2995 : /*
2996 : * This is the main entry point to direct page reclaim.
2997 : *
2998 : * If a full scan of the inactive list fails to free enough memory then we
2999 : * are "out of memory" and something needs to be killed.
3000 : *
3001 : * If the caller is !__GFP_FS then the probability of a failure is reasonably
3002 : * high - the zone may be full of dirty or under-writeback pages, which this
3003 : * caller can't do much about. We kick the writeback threads and take explicit
3004 : * naps in the hope that some of these pages can be written. But if the
3005 : * allocating task holds filesystem locks which prevent writeout this might not
3006 : * work, and the allocation attempt will fail.
3007 : *
3008 : * returns: 0, if no pages reclaimed
3009 : * else, the number of pages reclaimed
3010 : */
3011 0 : static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3012 : struct scan_control *sc)
3013 : {
3014 0 : int initial_priority = sc->priority;
3015 0 : pg_data_t *last_pgdat;
3016 0 : struct zoneref *z;
3017 0 : struct zone *zone;
3018 : retry:
3019 0 : delayacct_freepages_start();
3020 :
3021 0 : if (!cgroup_reclaim(sc))
3022 0 : __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3023 :
3024 0 : do {
3025 0 : vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3026 0 : sc->priority);
3027 0 : sc->nr_scanned = 0;
3028 0 : shrink_zones(zonelist, sc);
3029 :
3030 0 : if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3031 : break;
3032 :
3033 0 : if (sc->compaction_ready)
3034 : break;
3035 :
3036 : /*
3037 : * If we're getting trouble reclaiming, start doing
3038 : * writepage even in laptop mode.
3039 : */
3040 0 : if (sc->priority < DEF_PRIORITY - 2)
3041 0 : sc->may_writepage = 1;
3042 0 : } while (--sc->priority >= 0);
3043 :
3044 0 : last_pgdat = NULL;
3045 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3046 : sc->nodemask) {
3047 0 : if (zone->zone_pgdat == last_pgdat)
3048 0 : continue;
3049 0 : last_pgdat = zone->zone_pgdat;
3050 :
3051 0 : snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3052 :
3053 0 : if (cgroup_reclaim(sc)) {
3054 : struct lruvec *lruvec;
3055 :
3056 : lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3057 : zone->zone_pgdat);
3058 : clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3059 : }
3060 : }
3061 :
3062 0 : delayacct_freepages_end();
3063 :
3064 0 : if (sc->nr_reclaimed)
3065 0 : return sc->nr_reclaimed;
3066 :
3067 : /* Aborted reclaim to try compaction? don't OOM, then */
3068 0 : if (sc->compaction_ready)
3069 : return 1;
3070 :
3071 : /*
3072 : * We make inactive:active ratio decisions based on the node's
3073 : * composition of memory, but a restrictive reclaim_idx or a
3074 : * memory.low cgroup setting can exempt large amounts of
3075 : * memory from reclaim. Neither of which are very common, so
3076 : * instead of doing costly eligibility calculations of the
3077 : * entire cgroup subtree up front, we assume the estimates are
3078 : * good, and retry with forcible deactivation if that fails.
3079 : */
3080 0 : if (sc->skipped_deactivate) {
3081 0 : sc->priority = initial_priority;
3082 0 : sc->force_deactivate = 1;
3083 0 : sc->skipped_deactivate = 0;
3084 0 : goto retry;
3085 : }
3086 :
3087 : /* Untapped cgroup reserves? Don't OOM, retry. */
3088 0 : if (sc->memcg_low_skipped) {
3089 0 : sc->priority = initial_priority;
3090 0 : sc->force_deactivate = 0;
3091 0 : sc->memcg_low_reclaim = 1;
3092 0 : sc->memcg_low_skipped = 0;
3093 0 : goto retry;
3094 : }
3095 :
3096 : return 0;
3097 : }
3098 :
3099 0 : static bool allow_direct_reclaim(pg_data_t *pgdat)
3100 : {
3101 0 : struct zone *zone;
3102 0 : unsigned long pfmemalloc_reserve = 0;
3103 0 : unsigned long free_pages = 0;
3104 0 : int i;
3105 0 : bool wmark_ok;
3106 :
3107 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3108 : return true;
3109 :
3110 0 : for (i = 0; i <= ZONE_NORMAL; i++) {
3111 0 : zone = &pgdat->node_zones[i];
3112 0 : if (!managed_zone(zone))
3113 0 : continue;
3114 :
3115 0 : if (!zone_reclaimable_pages(zone))
3116 0 : continue;
3117 :
3118 0 : pfmemalloc_reserve += min_wmark_pages(zone);
3119 0 : free_pages += zone_page_state(zone, NR_FREE_PAGES);
3120 : }
3121 :
3122 : /* If there are no reserves (unexpected config) then do not throttle */
3123 0 : if (!pfmemalloc_reserve)
3124 : return true;
3125 :
3126 0 : wmark_ok = free_pages > pfmemalloc_reserve / 2;
3127 :
3128 : /* kswapd must be awake if processes are being throttled */
3129 0 : if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3130 0 : if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3131 0 : WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3132 :
3133 0 : wake_up_interruptible(&pgdat->kswapd_wait);
3134 : }
3135 :
3136 : return wmark_ok;
3137 : }
3138 :
3139 : /*
3140 : * Throttle direct reclaimers if backing storage is backed by the network
3141 : * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3142 : * depleted. kswapd will continue to make progress and wake the processes
3143 : * when the low watermark is reached.
3144 : *
3145 : * Returns true if a fatal signal was delivered during throttling. If this
3146 : * happens, the page allocator should not consider triggering the OOM killer.
3147 : */
3148 0 : static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3149 : nodemask_t *nodemask)
3150 : {
3151 0 : struct zoneref *z;
3152 0 : struct zone *zone;
3153 0 : pg_data_t *pgdat = NULL;
3154 :
3155 : /*
3156 : * Kernel threads should not be throttled as they may be indirectly
3157 : * responsible for cleaning pages necessary for reclaim to make forward
3158 : * progress. kjournald for example may enter direct reclaim while
3159 : * committing a transaction where throttling it could forcing other
3160 : * processes to block on log_wait_commit().
3161 : */
3162 0 : if (current->flags & PF_KTHREAD)
3163 0 : goto out;
3164 :
3165 : /*
3166 : * If a fatal signal is pending, this process should not throttle.
3167 : * It should return quickly so it can exit and free its memory
3168 : */
3169 0 : if (fatal_signal_pending(current))
3170 0 : goto out;
3171 :
3172 : /*
3173 : * Check if the pfmemalloc reserves are ok by finding the first node
3174 : * with a usable ZONE_NORMAL or lower zone. The expectation is that
3175 : * GFP_KERNEL will be required for allocating network buffers when
3176 : * swapping over the network so ZONE_HIGHMEM is unusable.
3177 : *
3178 : * Throttling is based on the first usable node and throttled processes
3179 : * wait on a queue until kswapd makes progress and wakes them. There
3180 : * is an affinity then between processes waking up and where reclaim
3181 : * progress has been made assuming the process wakes on the same node.
3182 : * More importantly, processes running on remote nodes will not compete
3183 : * for remote pfmemalloc reserves and processes on different nodes
3184 : * should make reasonable progress.
3185 : */
3186 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
3187 : gfp_zone(gfp_mask), nodemask) {
3188 0 : if (zone_idx(zone) > ZONE_NORMAL)
3189 0 : continue;
3190 :
3191 : /* Throttle based on the first usable node */
3192 0 : pgdat = zone->zone_pgdat;
3193 0 : if (allow_direct_reclaim(pgdat))
3194 0 : goto out;
3195 : break;
3196 : }
3197 :
3198 : /* If no zone was usable by the allocation flags then do not throttle */
3199 0 : if (!pgdat)
3200 0 : goto out;
3201 :
3202 : /* Account for the throttling */
3203 0 : count_vm_event(PGSCAN_DIRECT_THROTTLE);
3204 :
3205 : /*
3206 : * If the caller cannot enter the filesystem, it's possible that it
3207 : * is due to the caller holding an FS lock or performing a journal
3208 : * transaction in the case of a filesystem like ext[3|4]. In this case,
3209 : * it is not safe to block on pfmemalloc_wait as kswapd could be
3210 : * blocked waiting on the same lock. Instead, throttle for up to a
3211 : * second before continuing.
3212 : */
3213 0 : if (!(gfp_mask & __GFP_FS)) {
3214 0 : wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3215 : allow_direct_reclaim(pgdat), HZ);
3216 :
3217 0 : goto check_pending;
3218 : }
3219 :
3220 : /* Throttle until kswapd wakes the process */
3221 0 : wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3222 : allow_direct_reclaim(pgdat));
3223 :
3224 0 : check_pending:
3225 0 : if (fatal_signal_pending(current))
3226 0 : return true;
3227 :
3228 0 : out:
3229 : return false;
3230 : }
3231 :
3232 0 : unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3233 : gfp_t gfp_mask, nodemask_t *nodemask)
3234 : {
3235 0 : unsigned long nr_reclaimed;
3236 0 : struct scan_control sc = {
3237 : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3238 0 : .gfp_mask = current_gfp_context(gfp_mask),
3239 0 : .reclaim_idx = gfp_zone(gfp_mask),
3240 : .order = order,
3241 : .nodemask = nodemask,
3242 : .priority = DEF_PRIORITY,
3243 0 : .may_writepage = !laptop_mode,
3244 : .may_unmap = 1,
3245 : .may_swap = 1,
3246 : };
3247 :
3248 : /*
3249 : * scan_control uses s8 fields for order, priority, and reclaim_idx.
3250 : * Confirm they are large enough for max values.
3251 : */
3252 0 : BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3253 0 : BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3254 0 : BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3255 :
3256 : /*
3257 : * Do not enter reclaim if fatal signal was delivered while throttled.
3258 : * 1 is returned so that the page allocator does not OOM kill at this
3259 : * point.
3260 : */
3261 0 : if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3262 : return 1;
3263 :
3264 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3265 0 : trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3266 :
3267 0 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3268 :
3269 0 : trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3270 0 : set_task_reclaim_state(current, NULL);
3271 :
3272 0 : return nr_reclaimed;
3273 : }
3274 :
3275 : #ifdef CONFIG_MEMCG
3276 :
3277 : /* Only used by soft limit reclaim. Do not reuse for anything else. */
3278 : unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3279 : gfp_t gfp_mask, bool noswap,
3280 : pg_data_t *pgdat,
3281 : unsigned long *nr_scanned)
3282 : {
3283 : struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3284 : struct scan_control sc = {
3285 : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3286 : .target_mem_cgroup = memcg,
3287 : .may_writepage = !laptop_mode,
3288 : .may_unmap = 1,
3289 : .reclaim_idx = MAX_NR_ZONES - 1,
3290 : .may_swap = !noswap,
3291 : };
3292 :
3293 : WARN_ON_ONCE(!current->reclaim_state);
3294 :
3295 : sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3296 : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3297 :
3298 : trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3299 : sc.gfp_mask);
3300 :
3301 : /*
3302 : * NOTE: Although we can get the priority field, using it
3303 : * here is not a good idea, since it limits the pages we can scan.
3304 : * if we don't reclaim here, the shrink_node from balance_pgdat
3305 : * will pick up pages from other mem cgroup's as well. We hack
3306 : * the priority and make it zero.
3307 : */
3308 : shrink_lruvec(lruvec, &sc);
3309 :
3310 : trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3311 :
3312 : *nr_scanned = sc.nr_scanned;
3313 :
3314 : return sc.nr_reclaimed;
3315 : }
3316 :
3317 : unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3318 : unsigned long nr_pages,
3319 : gfp_t gfp_mask,
3320 : bool may_swap)
3321 : {
3322 : unsigned long nr_reclaimed;
3323 : unsigned int noreclaim_flag;
3324 : struct scan_control sc = {
3325 : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3326 : .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3327 : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3328 : .reclaim_idx = MAX_NR_ZONES - 1,
3329 : .target_mem_cgroup = memcg,
3330 : .priority = DEF_PRIORITY,
3331 : .may_writepage = !laptop_mode,
3332 : .may_unmap = 1,
3333 : .may_swap = may_swap,
3334 : };
3335 : /*
3336 : * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3337 : * equal pressure on all the nodes. This is based on the assumption that
3338 : * the reclaim does not bail out early.
3339 : */
3340 : struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3341 :
3342 : set_task_reclaim_state(current, &sc.reclaim_state);
3343 : trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3344 : noreclaim_flag = memalloc_noreclaim_save();
3345 :
3346 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3347 :
3348 : memalloc_noreclaim_restore(noreclaim_flag);
3349 : trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3350 : set_task_reclaim_state(current, NULL);
3351 :
3352 : return nr_reclaimed;
3353 : }
3354 : #endif
3355 :
3356 0 : static void age_active_anon(struct pglist_data *pgdat,
3357 : struct scan_control *sc)
3358 : {
3359 0 : struct mem_cgroup *memcg;
3360 0 : struct lruvec *lruvec;
3361 :
3362 0 : if (!total_swap_pages)
3363 0 : return;
3364 :
3365 : lruvec = mem_cgroup_lruvec(NULL, pgdat);
3366 : if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3367 : return;
3368 :
3369 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
3370 : do {
3371 : lruvec = mem_cgroup_lruvec(memcg, pgdat);
3372 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3373 : sc, LRU_ACTIVE_ANON);
3374 : memcg = mem_cgroup_iter(NULL, memcg, NULL);
3375 : } while (memcg);
3376 : }
3377 :
3378 0 : static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3379 : {
3380 0 : int i;
3381 0 : struct zone *zone;
3382 :
3383 : /*
3384 : * Check for watermark boosts top-down as the higher zones
3385 : * are more likely to be boosted. Both watermarks and boosts
3386 : * should not be checked at the same time as reclaim would
3387 : * start prematurely when there is no boosting and a lower
3388 : * zone is balanced.
3389 : */
3390 0 : for (i = highest_zoneidx; i >= 0; i--) {
3391 0 : zone = pgdat->node_zones + i;
3392 0 : if (!managed_zone(zone))
3393 0 : continue;
3394 :
3395 0 : if (zone->watermark_boost)
3396 : return true;
3397 : }
3398 :
3399 : return false;
3400 : }
3401 :
3402 : /*
3403 : * Returns true if there is an eligible zone balanced for the request order
3404 : * and highest_zoneidx
3405 : */
3406 2 : static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3407 : {
3408 2 : int i;
3409 2 : unsigned long mark = -1;
3410 2 : struct zone *zone;
3411 :
3412 : /*
3413 : * Check watermarks bottom-up as lower zones are more likely to
3414 : * meet watermarks.
3415 : */
3416 2 : for (i = 0; i <= highest_zoneidx; i++) {
3417 2 : zone = pgdat->node_zones + i;
3418 :
3419 2 : if (!managed_zone(zone))
3420 0 : continue;
3421 :
3422 2 : mark = high_wmark_pages(zone);
3423 2 : if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3424 : return true;
3425 : }
3426 :
3427 : /*
3428 : * If a node has no populated zone within highest_zoneidx, it does not
3429 : * need balancing by definition. This can happen if a zone-restricted
3430 : * allocation tries to wake a remote kswapd.
3431 : */
3432 0 : if (mark == -1)
3433 0 : return true;
3434 :
3435 : return false;
3436 : }
3437 :
3438 : /* Clear pgdat state for congested, dirty or under writeback. */
3439 2 : static void clear_pgdat_congested(pg_data_t *pgdat)
3440 : {
3441 2 : struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3442 :
3443 2 : clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3444 2 : clear_bit(PGDAT_DIRTY, &pgdat->flags);
3445 2 : clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3446 2 : }
3447 :
3448 : /*
3449 : * Prepare kswapd for sleeping. This verifies that there are no processes
3450 : * waiting in throttle_direct_reclaim() and that watermarks have been met.
3451 : *
3452 : * Returns true if kswapd is ready to sleep
3453 : */
3454 2 : static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3455 : int highest_zoneidx)
3456 : {
3457 : /*
3458 : * The throttled processes are normally woken up in balance_pgdat() as
3459 : * soon as allow_direct_reclaim() is true. But there is a potential
3460 : * race between when kswapd checks the watermarks and a process gets
3461 : * throttled. There is also a potential race if processes get
3462 : * throttled, kswapd wakes, a large process exits thereby balancing the
3463 : * zones, which causes kswapd to exit balance_pgdat() before reaching
3464 : * the wake up checks. If kswapd is going to sleep, no process should
3465 : * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3466 : * the wake up is premature, processes will wake kswapd and get
3467 : * throttled again. The difference from wake ups in balance_pgdat() is
3468 : * that here we are under prepare_to_wait().
3469 : */
3470 2 : if (waitqueue_active(&pgdat->pfmemalloc_wait))
3471 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3472 :
3473 : /* Hopeless node, leave it to direct reclaim */
3474 2 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3475 : return true;
3476 :
3477 2 : if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3478 2 : clear_pgdat_congested(pgdat);
3479 2 : return true;
3480 : }
3481 :
3482 : return false;
3483 : }
3484 :
3485 : /*
3486 : * kswapd shrinks a node of pages that are at or below the highest usable
3487 : * zone that is currently unbalanced.
3488 : *
3489 : * Returns true if kswapd scanned at least the requested number of pages to
3490 : * reclaim or if the lack of progress was due to pages under writeback.
3491 : * This is used to determine if the scanning priority needs to be raised.
3492 : */
3493 0 : static bool kswapd_shrink_node(pg_data_t *pgdat,
3494 : struct scan_control *sc)
3495 : {
3496 0 : struct zone *zone;
3497 0 : int z;
3498 :
3499 : /* Reclaim a number of pages proportional to the number of zones */
3500 0 : sc->nr_to_reclaim = 0;
3501 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
3502 0 : zone = pgdat->node_zones + z;
3503 0 : if (!managed_zone(zone))
3504 0 : continue;
3505 :
3506 0 : sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3507 : }
3508 :
3509 : /*
3510 : * Historically care was taken to put equal pressure on all zones but
3511 : * now pressure is applied based on node LRU order.
3512 : */
3513 0 : shrink_node(pgdat, sc);
3514 :
3515 : /*
3516 : * Fragmentation may mean that the system cannot be rebalanced for
3517 : * high-order allocations. If twice the allocation size has been
3518 : * reclaimed then recheck watermarks only at order-0 to prevent
3519 : * excessive reclaim. Assume that a process requested a high-order
3520 : * can direct reclaim/compact.
3521 : */
3522 0 : if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3523 0 : sc->order = 0;
3524 :
3525 0 : return sc->nr_scanned >= sc->nr_to_reclaim;
3526 : }
3527 :
3528 : /*
3529 : * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3530 : * that are eligible for use by the caller until at least one zone is
3531 : * balanced.
3532 : *
3533 : * Returns the order kswapd finished reclaiming at.
3534 : *
3535 : * kswapd scans the zones in the highmem->normal->dma direction. It skips
3536 : * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3537 : * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3538 : * or lower is eligible for reclaim until at least one usable zone is
3539 : * balanced.
3540 : */
3541 0 : static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3542 : {
3543 0 : int i;
3544 0 : unsigned long nr_soft_reclaimed;
3545 0 : unsigned long nr_soft_scanned;
3546 0 : unsigned long pflags;
3547 0 : unsigned long nr_boost_reclaim;
3548 0 : unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3549 0 : bool boosted;
3550 0 : struct zone *zone;
3551 0 : struct scan_control sc = {
3552 : .gfp_mask = GFP_KERNEL,
3553 : .order = order,
3554 : .may_unmap = 1,
3555 : };
3556 :
3557 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3558 0 : psi_memstall_enter(&pflags);
3559 0 : __fs_reclaim_acquire();
3560 :
3561 0 : count_vm_event(PAGEOUTRUN);
3562 :
3563 : /*
3564 : * Account for the reclaim boost. Note that the zone boost is left in
3565 : * place so that parallel allocations that are near the watermark will
3566 : * stall or direct reclaim until kswapd is finished.
3567 : */
3568 0 : nr_boost_reclaim = 0;
3569 0 : for (i = 0; i <= highest_zoneidx; i++) {
3570 0 : zone = pgdat->node_zones + i;
3571 0 : if (!managed_zone(zone))
3572 0 : continue;
3573 :
3574 0 : nr_boost_reclaim += zone->watermark_boost;
3575 0 : zone_boosts[i] = zone->watermark_boost;
3576 : }
3577 0 : boosted = nr_boost_reclaim;
3578 :
3579 0 : restart:
3580 0 : sc.priority = DEF_PRIORITY;
3581 0 : do {
3582 0 : unsigned long nr_reclaimed = sc.nr_reclaimed;
3583 0 : bool raise_priority = true;
3584 0 : bool balanced;
3585 0 : bool ret;
3586 :
3587 0 : sc.reclaim_idx = highest_zoneidx;
3588 :
3589 : /*
3590 : * If the number of buffer_heads exceeds the maximum allowed
3591 : * then consider reclaiming from all zones. This has a dual
3592 : * purpose -- on 64-bit systems it is expected that
3593 : * buffer_heads are stripped during active rotation. On 32-bit
3594 : * systems, highmem pages can pin lowmem memory and shrinking
3595 : * buffers can relieve lowmem pressure. Reclaim may still not
3596 : * go ahead if all eligible zones for the original allocation
3597 : * request are balanced to avoid excessive reclaim from kswapd.
3598 : */
3599 0 : if (buffer_heads_over_limit) {
3600 0 : for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3601 0 : zone = pgdat->node_zones + i;
3602 0 : if (!managed_zone(zone))
3603 0 : continue;
3604 :
3605 0 : sc.reclaim_idx = i;
3606 0 : break;
3607 : }
3608 : }
3609 :
3610 : /*
3611 : * If the pgdat is imbalanced then ignore boosting and preserve
3612 : * the watermarks for a later time and restart. Note that the
3613 : * zone watermarks will be still reset at the end of balancing
3614 : * on the grounds that the normal reclaim should be enough to
3615 : * re-evaluate if boosting is required when kswapd next wakes.
3616 : */
3617 0 : balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3618 0 : if (!balanced && nr_boost_reclaim) {
3619 0 : nr_boost_reclaim = 0;
3620 0 : goto restart;
3621 : }
3622 :
3623 : /*
3624 : * If boosting is not active then only reclaim if there are no
3625 : * eligible zones. Note that sc.reclaim_idx is not used as
3626 : * buffer_heads_over_limit may have adjusted it.
3627 : */
3628 0 : if (!nr_boost_reclaim && balanced)
3629 0 : goto out;
3630 :
3631 : /* Limit the priority of boosting to avoid reclaim writeback */
3632 0 : if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3633 0 : raise_priority = false;
3634 :
3635 : /*
3636 : * Do not writeback or swap pages for boosted reclaim. The
3637 : * intent is to relieve pressure not issue sub-optimal IO
3638 : * from reclaim context. If no pages are reclaimed, the
3639 : * reclaim will be aborted.
3640 : */
3641 0 : sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3642 0 : sc.may_swap = !nr_boost_reclaim;
3643 :
3644 : /*
3645 : * Do some background aging of the anon list, to give
3646 : * pages a chance to be referenced before reclaiming. All
3647 : * pages are rotated regardless of classzone as this is
3648 : * about consistent aging.
3649 : */
3650 0 : age_active_anon(pgdat, &sc);
3651 :
3652 : /*
3653 : * If we're getting trouble reclaiming, start doing writepage
3654 : * even in laptop mode.
3655 : */
3656 0 : if (sc.priority < DEF_PRIORITY - 2)
3657 0 : sc.may_writepage = 1;
3658 :
3659 : /* Call soft limit reclaim before calling shrink_node. */
3660 0 : sc.nr_scanned = 0;
3661 0 : nr_soft_scanned = 0;
3662 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3663 : sc.gfp_mask, &nr_soft_scanned);
3664 0 : sc.nr_reclaimed += nr_soft_reclaimed;
3665 :
3666 : /*
3667 : * There should be no need to raise the scanning priority if
3668 : * enough pages are already being scanned that that high
3669 : * watermark would be met at 100% efficiency.
3670 : */
3671 0 : if (kswapd_shrink_node(pgdat, &sc))
3672 0 : raise_priority = false;
3673 :
3674 : /*
3675 : * If the low watermark is met there is no need for processes
3676 : * to be throttled on pfmemalloc_wait as they should not be
3677 : * able to safely make forward progress. Wake them
3678 : */
3679 0 : if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3680 0 : allow_direct_reclaim(pgdat))
3681 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3682 :
3683 : /* Check if kswapd should be suspending */
3684 0 : __fs_reclaim_release();
3685 0 : ret = try_to_freeze();
3686 0 : __fs_reclaim_acquire();
3687 0 : if (ret || kthread_should_stop())
3688 : break;
3689 :
3690 : /*
3691 : * Raise priority if scanning rate is too low or there was no
3692 : * progress in reclaiming pages
3693 : */
3694 0 : nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3695 0 : nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3696 :
3697 : /*
3698 : * If reclaim made no progress for a boost, stop reclaim as
3699 : * IO cannot be queued and it could be an infinite loop in
3700 : * extreme circumstances.
3701 : */
3702 0 : if (nr_boost_reclaim && !nr_reclaimed)
3703 : break;
3704 :
3705 0 : if (raise_priority || !nr_reclaimed)
3706 0 : sc.priority--;
3707 0 : } while (sc.priority >= 1);
3708 :
3709 0 : if (!sc.nr_reclaimed)
3710 0 : pgdat->kswapd_failures++;
3711 :
3712 0 : out:
3713 : /* If reclaim was boosted, account for the reclaim done in this pass */
3714 0 : if (boosted) {
3715 : unsigned long flags;
3716 :
3717 0 : for (i = 0; i <= highest_zoneidx; i++) {
3718 0 : if (!zone_boosts[i])
3719 0 : continue;
3720 :
3721 : /* Increments are under the zone lock */
3722 0 : zone = pgdat->node_zones + i;
3723 0 : spin_lock_irqsave(&zone->lock, flags);
3724 0 : zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3725 0 : spin_unlock_irqrestore(&zone->lock, flags);
3726 : }
3727 :
3728 : /*
3729 : * As there is now likely space, wakeup kcompact to defragment
3730 : * pageblocks.
3731 : */
3732 0 : wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3733 : }
3734 :
3735 0 : snapshot_refaults(NULL, pgdat);
3736 0 : __fs_reclaim_release();
3737 0 : psi_memstall_leave(&pflags);
3738 0 : set_task_reclaim_state(current, NULL);
3739 :
3740 : /*
3741 : * Return the order kswapd stopped reclaiming at as
3742 : * prepare_kswapd_sleep() takes it into account. If another caller
3743 : * entered the allocator slow path while kswapd was awake, order will
3744 : * remain at the higher level.
3745 : */
3746 0 : return sc.order;
3747 : }
3748 :
3749 : /*
3750 : * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3751 : * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3752 : * not a valid index then either kswapd runs for first time or kswapd couldn't
3753 : * sleep after previous reclaim attempt (node is still unbalanced). In that
3754 : * case return the zone index of the previous kswapd reclaim cycle.
3755 : */
3756 1 : static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3757 : enum zone_type prev_highest_zoneidx)
3758 : {
3759 1 : enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3760 :
3761 1 : return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3762 : }
3763 :
3764 1 : static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3765 : unsigned int highest_zoneidx)
3766 : {
3767 1 : long remaining = 0;
3768 1 : DEFINE_WAIT(wait);
3769 :
3770 1 : if (freezing(current) || kthread_should_stop())
3771 0 : return;
3772 :
3773 1 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3774 :
3775 : /*
3776 : * Try to sleep for a short interval. Note that kcompactd will only be
3777 : * woken if it is possible to sleep for a short interval. This is
3778 : * deliberate on the assumption that if reclaim cannot keep an
3779 : * eligible zone balanced that it's also unlikely that compaction will
3780 : * succeed.
3781 : */
3782 1 : if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3783 : /*
3784 : * Compaction records what page blocks it recently failed to
3785 : * isolate pages from and skips them in the future scanning.
3786 : * When kswapd is going to sleep, it is reasonable to assume
3787 : * that pages and compaction may succeed so reset the cache.
3788 : */
3789 1 : reset_isolation_suitable(pgdat);
3790 :
3791 : /*
3792 : * We have freed the memory, now we should compact it to make
3793 : * allocation of the requested order possible.
3794 : */
3795 1 : wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3796 :
3797 1 : remaining = schedule_timeout(HZ/10);
3798 :
3799 : /*
3800 : * If woken prematurely then reset kswapd_highest_zoneidx and
3801 : * order. The values will either be from a wakeup request or
3802 : * the previous request that slept prematurely.
3803 : */
3804 1 : if (remaining) {
3805 0 : WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3806 : kswapd_highest_zoneidx(pgdat,
3807 : highest_zoneidx));
3808 :
3809 0 : if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3810 0 : WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3811 : }
3812 :
3813 1 : finish_wait(&pgdat->kswapd_wait, &wait);
3814 1 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3815 : }
3816 :
3817 : /*
3818 : * After a short sleep, check if it was a premature sleep. If not, then
3819 : * go fully to sleep until explicitly woken up.
3820 : */
3821 2 : if (!remaining &&
3822 1 : prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3823 1 : trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3824 :
3825 : /*
3826 : * vmstat counters are not perfectly accurate and the estimated
3827 : * value for counters such as NR_FREE_PAGES can deviate from the
3828 : * true value by nr_online_cpus * threshold. To avoid the zone
3829 : * watermarks being breached while under pressure, we reduce the
3830 : * per-cpu vmstat threshold while kswapd is awake and restore
3831 : * them before going back to sleep.
3832 : */
3833 1 : set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3834 :
3835 1 : if (!kthread_should_stop())
3836 1 : schedule();
3837 :
3838 0 : set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3839 : } else {
3840 0 : if (remaining)
3841 0 : count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3842 : else
3843 0 : count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3844 : }
3845 0 : finish_wait(&pgdat->kswapd_wait, &wait);
3846 : }
3847 :
3848 : /*
3849 : * The background pageout daemon, started as a kernel thread
3850 : * from the init process.
3851 : *
3852 : * This basically trickles out pages so that we have _some_
3853 : * free memory available even if there is no other activity
3854 : * that frees anything up. This is needed for things like routing
3855 : * etc, where we otherwise might have all activity going on in
3856 : * asynchronous contexts that cannot page things out.
3857 : *
3858 : * If there are applications that are active memory-allocators
3859 : * (most normal use), this basically shouldn't matter.
3860 : */
3861 1 : static int kswapd(void *p)
3862 : {
3863 1 : unsigned int alloc_order, reclaim_order;
3864 1 : unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3865 1 : pg_data_t *pgdat = (pg_data_t*)p;
3866 1 : struct task_struct *tsk = current;
3867 1 : const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3868 :
3869 1 : if (!cpumask_empty(cpumask))
3870 1 : set_cpus_allowed_ptr(tsk, cpumask);
3871 :
3872 : /*
3873 : * Tell the memory management that we're a "memory allocator",
3874 : * and that if we need more memory we should get access to it
3875 : * regardless (see "__alloc_pages()"). "kswapd" should
3876 : * never get caught in the normal page freeing logic.
3877 : *
3878 : * (Kswapd normally doesn't need memory anyway, but sometimes
3879 : * you need a small amount of memory in order to be able to
3880 : * page out something else, and this flag essentially protects
3881 : * us from recursively trying to free more memory as we're
3882 : * trying to free the first piece of memory in the first place).
3883 : */
3884 1 : tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3885 1 : set_freezable();
3886 :
3887 1 : WRITE_ONCE(pgdat->kswapd_order, 0);
3888 1 : WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3889 1 : for ( ; ; ) {
3890 1 : bool ret;
3891 :
3892 1 : alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3893 1 : highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3894 : highest_zoneidx);
3895 :
3896 1 : kswapd_try_sleep:
3897 1 : kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3898 : highest_zoneidx);
3899 :
3900 : /* Read the new order and highest_zoneidx */
3901 0 : alloc_order = READ_ONCE(pgdat->kswapd_order);
3902 0 : highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3903 : highest_zoneidx);
3904 0 : WRITE_ONCE(pgdat->kswapd_order, 0);
3905 0 : WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3906 :
3907 0 : ret = try_to_freeze();
3908 0 : if (kthread_should_stop())
3909 : break;
3910 :
3911 : /*
3912 : * We can speed up thawing tasks if we don't call balance_pgdat
3913 : * after returning from the refrigerator
3914 : */
3915 0 : if (ret)
3916 : continue;
3917 :
3918 : /*
3919 : * Reclaim begins at the requested order but if a high-order
3920 : * reclaim fails then kswapd falls back to reclaiming for
3921 : * order-0. If that happens, kswapd will consider sleeping
3922 : * for the order it finished reclaiming at (reclaim_order)
3923 : * but kcompactd is woken to compact for the original
3924 : * request (alloc_order).
3925 : */
3926 0 : trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3927 : alloc_order);
3928 0 : reclaim_order = balance_pgdat(pgdat, alloc_order,
3929 : highest_zoneidx);
3930 0 : if (reclaim_order < alloc_order)
3931 0 : goto kswapd_try_sleep;
3932 : }
3933 :
3934 0 : tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3935 :
3936 0 : return 0;
3937 : }
3938 :
3939 : /*
3940 : * A zone is low on free memory or too fragmented for high-order memory. If
3941 : * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3942 : * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3943 : * has failed or is not needed, still wake up kcompactd if only compaction is
3944 : * needed.
3945 : */
3946 0 : void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3947 : enum zone_type highest_zoneidx)
3948 : {
3949 0 : pg_data_t *pgdat;
3950 0 : enum zone_type curr_idx;
3951 :
3952 0 : if (!managed_zone(zone))
3953 : return;
3954 :
3955 0 : if (!cpuset_zone_allowed(zone, gfp_flags))
3956 : return;
3957 :
3958 0 : pgdat = zone->zone_pgdat;
3959 0 : curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3960 :
3961 0 : if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3962 0 : WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3963 :
3964 0 : if (READ_ONCE(pgdat->kswapd_order) < order)
3965 0 : WRITE_ONCE(pgdat->kswapd_order, order);
3966 :
3967 0 : if (!waitqueue_active(&pgdat->kswapd_wait))
3968 : return;
3969 :
3970 : /* Hopeless node, leave it to direct reclaim if possible */
3971 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3972 0 : (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3973 0 : !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3974 : /*
3975 : * There may be plenty of free memory available, but it's too
3976 : * fragmented for high-order allocations. Wake up kcompactd
3977 : * and rely on compaction_suitable() to determine if it's
3978 : * needed. If it fails, it will defer subsequent attempts to
3979 : * ratelimit its work.
3980 : */
3981 0 : if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3982 0 : wakeup_kcompactd(pgdat, order, highest_zoneidx);
3983 0 : return;
3984 : }
3985 :
3986 0 : trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3987 : gfp_flags);
3988 0 : wake_up_interruptible(&pgdat->kswapd_wait);
3989 : }
3990 :
3991 : #ifdef CONFIG_HIBERNATION
3992 : /*
3993 : * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3994 : * freed pages.
3995 : *
3996 : * Rather than trying to age LRUs the aim is to preserve the overall
3997 : * LRU order by reclaiming preferentially
3998 : * inactive > active > active referenced > active mapped
3999 : */
4000 : unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4001 : {
4002 : struct scan_control sc = {
4003 : .nr_to_reclaim = nr_to_reclaim,
4004 : .gfp_mask = GFP_HIGHUSER_MOVABLE,
4005 : .reclaim_idx = MAX_NR_ZONES - 1,
4006 : .priority = DEF_PRIORITY,
4007 : .may_writepage = 1,
4008 : .may_unmap = 1,
4009 : .may_swap = 1,
4010 : .hibernation_mode = 1,
4011 : };
4012 : struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4013 : unsigned long nr_reclaimed;
4014 : unsigned int noreclaim_flag;
4015 :
4016 : fs_reclaim_acquire(sc.gfp_mask);
4017 : noreclaim_flag = memalloc_noreclaim_save();
4018 : set_task_reclaim_state(current, &sc.reclaim_state);
4019 :
4020 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4021 :
4022 : set_task_reclaim_state(current, NULL);
4023 : memalloc_noreclaim_restore(noreclaim_flag);
4024 : fs_reclaim_release(sc.gfp_mask);
4025 :
4026 : return nr_reclaimed;
4027 : }
4028 : #endif /* CONFIG_HIBERNATION */
4029 :
4030 : /*
4031 : * This kswapd start function will be called by init and node-hot-add.
4032 : * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4033 : */
4034 1 : int kswapd_run(int nid)
4035 : {
4036 1 : pg_data_t *pgdat = NODE_DATA(nid);
4037 1 : int ret = 0;
4038 :
4039 1 : if (pgdat->kswapd)
4040 : return 0;
4041 :
4042 1 : pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4043 1 : if (IS_ERR(pgdat->kswapd)) {
4044 : /* failure at boot is fatal */
4045 0 : BUG_ON(system_state < SYSTEM_RUNNING);
4046 0 : pr_err("Failed to start kswapd on node %d\n", nid);
4047 0 : ret = PTR_ERR(pgdat->kswapd);
4048 0 : pgdat->kswapd = NULL;
4049 : }
4050 : return ret;
4051 : }
4052 :
4053 : /*
4054 : * Called by memory hotplug when all memory in a node is offlined. Caller must
4055 : * hold mem_hotplug_begin/end().
4056 : */
4057 0 : void kswapd_stop(int nid)
4058 : {
4059 0 : struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4060 :
4061 0 : if (kswapd) {
4062 0 : kthread_stop(kswapd);
4063 0 : NODE_DATA(nid)->kswapd = NULL;
4064 : }
4065 0 : }
4066 :
4067 1 : static int __init kswapd_init(void)
4068 : {
4069 1 : int nid;
4070 :
4071 1 : swap_setup();
4072 2 : for_each_node_state(nid, N_MEMORY)
4073 1 : kswapd_run(nid);
4074 1 : return 0;
4075 : }
4076 :
4077 : module_init(kswapd_init)
4078 :
4079 : #ifdef CONFIG_NUMA
4080 : /*
4081 : * Node reclaim mode
4082 : *
4083 : * If non-zero call node_reclaim when the number of free pages falls below
4084 : * the watermarks.
4085 : */
4086 : int node_reclaim_mode __read_mostly;
4087 :
4088 : /*
4089 : * These bit locations are exposed in the vm.zone_reclaim_mode sysctl
4090 : * ABI. New bits are OK, but existing bits can never change.
4091 : */
4092 : #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4093 : #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4094 : #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4095 :
4096 : /*
4097 : * Priority for NODE_RECLAIM. This determines the fraction of pages
4098 : * of a node considered for each zone_reclaim. 4 scans 1/16th of
4099 : * a zone.
4100 : */
4101 : #define NODE_RECLAIM_PRIORITY 4
4102 :
4103 : /*
4104 : * Percentage of pages in a zone that must be unmapped for node_reclaim to
4105 : * occur.
4106 : */
4107 : int sysctl_min_unmapped_ratio = 1;
4108 :
4109 : /*
4110 : * If the number of slab pages in a zone grows beyond this percentage then
4111 : * slab reclaim needs to occur.
4112 : */
4113 : int sysctl_min_slab_ratio = 5;
4114 :
4115 0 : static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4116 : {
4117 0 : unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4118 0 : unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4119 0 : node_page_state(pgdat, NR_ACTIVE_FILE);
4120 :
4121 : /*
4122 : * It's possible for there to be more file mapped pages than
4123 : * accounted for by the pages on the file LRU lists because
4124 : * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4125 : */
4126 0 : return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4127 : }
4128 :
4129 : /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4130 0 : static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4131 : {
4132 0 : unsigned long nr_pagecache_reclaimable;
4133 0 : unsigned long delta = 0;
4134 :
4135 : /*
4136 : * If RECLAIM_UNMAP is set, then all file pages are considered
4137 : * potentially reclaimable. Otherwise, we have to worry about
4138 : * pages like swapcache and node_unmapped_file_pages() provides
4139 : * a better estimate
4140 : */
4141 0 : if (node_reclaim_mode & RECLAIM_UNMAP)
4142 0 : nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4143 : else
4144 0 : nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4145 :
4146 : /* If we can't clean pages, remove dirty pages from consideration */
4147 0 : if (!(node_reclaim_mode & RECLAIM_WRITE))
4148 0 : delta += node_page_state(pgdat, NR_FILE_DIRTY);
4149 :
4150 : /* Watch for any possible underflows due to delta */
4151 0 : if (unlikely(delta > nr_pagecache_reclaimable))
4152 0 : delta = nr_pagecache_reclaimable;
4153 :
4154 0 : return nr_pagecache_reclaimable - delta;
4155 : }
4156 :
4157 : /*
4158 : * Try to free up some pages from this node through reclaim.
4159 : */
4160 0 : static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4161 : {
4162 : /* Minimum pages needed in order to stay on node */
4163 0 : const unsigned long nr_pages = 1 << order;
4164 0 : struct task_struct *p = current;
4165 0 : unsigned int noreclaim_flag;
4166 0 : struct scan_control sc = {
4167 0 : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4168 0 : .gfp_mask = current_gfp_context(gfp_mask),
4169 : .order = order,
4170 : .priority = NODE_RECLAIM_PRIORITY,
4171 0 : .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4172 0 : .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4173 : .may_swap = 1,
4174 0 : .reclaim_idx = gfp_zone(gfp_mask),
4175 : };
4176 :
4177 0 : trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4178 : sc.gfp_mask);
4179 :
4180 0 : cond_resched();
4181 0 : fs_reclaim_acquire(sc.gfp_mask);
4182 : /*
4183 : * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4184 : * and we also need to be able to write out pages for RECLAIM_WRITE
4185 : * and RECLAIM_UNMAP.
4186 : */
4187 0 : noreclaim_flag = memalloc_noreclaim_save();
4188 0 : p->flags |= PF_SWAPWRITE;
4189 0 : set_task_reclaim_state(p, &sc.reclaim_state);
4190 :
4191 0 : if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4192 : /*
4193 : * Free memory by calling shrink node with increasing
4194 : * priorities until we have enough memory freed.
4195 : */
4196 0 : do {
4197 0 : shrink_node(pgdat, &sc);
4198 0 : } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4199 : }
4200 :
4201 0 : set_task_reclaim_state(p, NULL);
4202 0 : current->flags &= ~PF_SWAPWRITE;
4203 0 : memalloc_noreclaim_restore(noreclaim_flag);
4204 0 : fs_reclaim_release(sc.gfp_mask);
4205 :
4206 0 : trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4207 :
4208 0 : return sc.nr_reclaimed >= nr_pages;
4209 : }
4210 :
4211 0 : int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4212 : {
4213 0 : int ret;
4214 :
4215 : /*
4216 : * Node reclaim reclaims unmapped file backed pages and
4217 : * slab pages if we are over the defined limits.
4218 : *
4219 : * A small portion of unmapped file backed pages is needed for
4220 : * file I/O otherwise pages read by file I/O will be immediately
4221 : * thrown out if the node is overallocated. So we do not reclaim
4222 : * if less than a specified percentage of the node is used by
4223 : * unmapped file backed pages.
4224 : */
4225 0 : if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4226 0 : node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4227 0 : pgdat->min_slab_pages)
4228 : return NODE_RECLAIM_FULL;
4229 :
4230 : /*
4231 : * Do not scan if the allocation should not be delayed.
4232 : */
4233 0 : if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4234 : return NODE_RECLAIM_NOSCAN;
4235 :
4236 : /*
4237 : * Only run node reclaim on the local node or on nodes that do not
4238 : * have associated processors. This will favor the local processor
4239 : * over remote processors and spread off node memory allocations
4240 : * as wide as possible.
4241 : */
4242 0 : if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4243 : return NODE_RECLAIM_NOSCAN;
4244 :
4245 0 : if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4246 : return NODE_RECLAIM_NOSCAN;
4247 :
4248 0 : ret = __node_reclaim(pgdat, gfp_mask, order);
4249 0 : clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4250 :
4251 0 : if (!ret)
4252 0 : count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4253 :
4254 : return ret;
4255 : }
4256 : #endif
4257 :
4258 : /**
4259 : * check_move_unevictable_pages - check pages for evictability and move to
4260 : * appropriate zone lru list
4261 : * @pvec: pagevec with lru pages to check
4262 : *
4263 : * Checks pages for evictability, if an evictable page is in the unevictable
4264 : * lru list, moves it to the appropriate evictable lru list. This function
4265 : * should be only used for lru pages.
4266 : */
4267 0 : void check_move_unevictable_pages(struct pagevec *pvec)
4268 : {
4269 0 : struct lruvec *lruvec = NULL;
4270 0 : int pgscanned = 0;
4271 0 : int pgrescued = 0;
4272 0 : int i;
4273 :
4274 0 : for (i = 0; i < pvec->nr; i++) {
4275 0 : struct page *page = pvec->pages[i];
4276 0 : int nr_pages;
4277 :
4278 0 : if (PageTransTail(page))
4279 0 : continue;
4280 :
4281 0 : nr_pages = thp_nr_pages(page);
4282 0 : pgscanned += nr_pages;
4283 :
4284 : /* block memcg migration during page moving between lru */
4285 0 : if (!TestClearPageLRU(page))
4286 0 : continue;
4287 :
4288 0 : lruvec = relock_page_lruvec_irq(page, lruvec);
4289 0 : if (page_evictable(page) && PageUnevictable(page)) {
4290 0 : del_page_from_lru_list(page, lruvec);
4291 0 : ClearPageUnevictable(page);
4292 0 : add_page_to_lru_list(page, lruvec);
4293 0 : pgrescued += nr_pages;
4294 : }
4295 0 : SetPageLRU(page);
4296 : }
4297 :
4298 0 : if (lruvec) {
4299 0 : __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4300 0 : __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4301 0 : unlock_page_lruvec_irq(lruvec);
4302 0 : } else if (pgscanned) {
4303 0 : count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4304 : }
4305 0 : }
4306 : EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
|
