Branch data Line data Source code
1 : : // SPDX-License-Identifier: GPL-2.0
2 : : /*
3 : : * linux/mm/vmscan.c
4 : : *
5 : : * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 : : *
7 : : * Swap reorganised 29.12.95, Stephen Tweedie.
8 : : * kswapd added: 7.1.96 sct
9 : : * Removed kswapd_ctl limits, and swap out as many pages as needed
10 : : * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 : : * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 : : * Multiqueue VM started 5.8.00, Rik van Riel.
13 : : */
14 : :
15 : : #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 : :
17 : : #include <linux/mm.h>
18 : : #include <linux/sched/mm.h>
19 : : #include <linux/module.h>
20 : : #include <linux/gfp.h>
21 : : #include <linux/kernel_stat.h>
22 : : #include <linux/swap.h>
23 : : #include <linux/pagemap.h>
24 : : #include <linux/init.h>
25 : : #include <linux/highmem.h>
26 : : #include <linux/vmpressure.h>
27 : : #include <linux/vmstat.h>
28 : : #include <linux/file.h>
29 : : #include <linux/writeback.h>
30 : : #include <linux/blkdev.h>
31 : : #include <linux/buffer_head.h> /* for try_to_release_page(),
32 : : buffer_heads_over_limit */
33 : : #include <linux/mm_inline.h>
34 : : #include <linux/backing-dev.h>
35 : : #include <linux/rmap.h>
36 : : #include <linux/topology.h>
37 : : #include <linux/cpu.h>
38 : : #include <linux/cpuset.h>
39 : : #include <linux/compaction.h>
40 : : #include <linux/notifier.h>
41 : : #include <linux/rwsem.h>
42 : : #include <linux/delay.h>
43 : : #include <linux/kthread.h>
44 : : #include <linux/freezer.h>
45 : : #include <linux/memcontrol.h>
46 : : #include <linux/delayacct.h>
47 : : #include <linux/sysctl.h>
48 : : #include <linux/oom.h>
49 : : #include <linux/pagevec.h>
50 : : #include <linux/prefetch.h>
51 : : #include <linux/printk.h>
52 : : #include <linux/dax.h>
53 : : #include <linux/psi.h>
54 : :
55 : : #include <asm/tlbflush.h>
56 : : #include <asm/div64.h>
57 : :
58 : : #include <linux/swapops.h>
59 : : #include <linux/balloon_compaction.h>
60 : :
61 : : #include "internal.h"
62 : :
63 : : #define CREATE_TRACE_POINTS
64 : : #include <trace/events/vmscan.h>
65 : :
66 : : struct scan_control {
67 : : /* How many pages shrink_list() should reclaim */
68 : : unsigned long nr_to_reclaim;
69 : :
70 : : /*
71 : : * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 : : * are scanned.
73 : : */
74 : : nodemask_t *nodemask;
75 : :
76 : : /*
77 : : * The memory cgroup that hit its limit and as a result is the
78 : : * primary target of this reclaim invocation.
79 : : */
80 : : struct mem_cgroup *target_mem_cgroup;
81 : :
82 : : /* Can active pages be deactivated as part of reclaim? */
83 : : #define DEACTIVATE_ANON 1
84 : : #define DEACTIVATE_FILE 2
85 : : unsigned int may_deactivate:2;
86 : : unsigned int force_deactivate:1;
87 : : unsigned int skipped_deactivate:1;
88 : :
89 : : /* Writepage batching in laptop mode; RECLAIM_WRITE */
90 : : unsigned int may_writepage:1;
91 : :
92 : : /* Can mapped pages be reclaimed? */
93 : : unsigned int may_unmap:1;
94 : :
95 : : /* Can pages be swapped as part of reclaim? */
96 : : unsigned int may_swap:1;
97 : :
98 : : /*
99 : : * Cgroups are not reclaimed below their configured memory.low,
100 : : * unless we threaten to OOM. If any cgroups are skipped due to
101 : : * memory.low and nothing was reclaimed, go back for memory.low.
102 : : */
103 : : unsigned int memcg_low_reclaim:1;
104 : : unsigned int memcg_low_skipped:1;
105 : :
106 : : unsigned int hibernation_mode:1;
107 : :
108 : : /* One of the zones is ready for compaction */
109 : : unsigned int compaction_ready:1;
110 : :
111 : : /* There is easily reclaimable cold cache in the current node */
112 : : unsigned int cache_trim_mode:1;
113 : :
114 : : /* The file pages on the current node are dangerously low */
115 : : unsigned int file_is_tiny:1;
116 : :
117 : : /* Allocation order */
118 : : s8 order;
119 : :
120 : : /* Scan (total_size >> priority) pages at once */
121 : : s8 priority;
122 : :
123 : : /* The highest zone to isolate pages for reclaim from */
124 : : s8 reclaim_idx;
125 : :
126 : : /* This context's GFP mask */
127 : : gfp_t gfp_mask;
128 : :
129 : : /* Incremented by the number of inactive pages that were scanned */
130 : : unsigned long nr_scanned;
131 : :
132 : : /* Number of pages freed so far during a call to shrink_zones() */
133 : : unsigned long nr_reclaimed;
134 : :
135 : : struct {
136 : : unsigned int dirty;
137 : : unsigned int unqueued_dirty;
138 : : unsigned int congested;
139 : : unsigned int writeback;
140 : : unsigned int immediate;
141 : : unsigned int file_taken;
142 : : unsigned int taken;
143 : : } nr;
144 : :
145 : : /* for recording the reclaimed slab by now */
146 : : struct reclaim_state reclaim_state;
147 : : };
148 : :
149 : : #ifdef ARCH_HAS_PREFETCHW
150 : : #define prefetchw_prev_lru_page(_page, _base, _field) \
151 : : do { \
152 : : if ((_page)->lru.prev != _base) { \
153 : : struct page *prev; \
154 : : \
155 : : prev = lru_to_page(&(_page->lru)); \
156 : : prefetchw(&prev->_field); \
157 : : } \
158 : : } while (0)
159 : : #else
160 : : #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
161 : : #endif
162 : :
163 : : /*
164 : : * From 0 .. 100. Higher means more swappy.
165 : : */
166 : : int vm_swappiness = 60;
167 : : /*
168 : : * The total number of pages which are beyond the high watermark within all
169 : : * zones.
170 : : */
171 : : unsigned long vm_total_pages;
172 : :
173 : 0 : static void set_task_reclaim_state(struct task_struct *task,
174 : : struct reclaim_state *rs)
175 : : {
176 : : /* Check for an overwrite */
177 [ # # # # : 0 : WARN_ON_ONCE(rs && task->reclaim_state);
# # ]
178 : :
179 : : /* Check for the nulling of an already-nulled member */
180 [ # # # # : 0 : WARN_ON_ONCE(!rs && !task->reclaim_state);
# # ]
181 : :
182 : 0 : task->reclaim_state = rs;
183 : 0 : }
184 : :
185 : : static LIST_HEAD(shrinker_list);
186 : : static DECLARE_RWSEM(shrinker_rwsem);
187 : :
188 : : #ifdef CONFIG_MEMCG
189 : : /*
190 : : * We allow subsystems to populate their shrinker-related
191 : : * LRU lists before register_shrinker_prepared() is called
192 : : * for the shrinker, since we don't want to impose
193 : : * restrictions on their internal registration order.
194 : : * In this case shrink_slab_memcg() may find corresponding
195 : : * bit is set in the shrinkers map.
196 : : *
197 : : * This value is used by the function to detect registering
198 : : * shrinkers and to skip do_shrink_slab() calls for them.
199 : : */
200 : : #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
201 : :
202 : : static DEFINE_IDR(shrinker_idr);
203 : : static int shrinker_nr_max;
204 : :
205 : : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
206 : : {
207 : : int id, ret = -ENOMEM;
208 : :
209 : : down_write(&shrinker_rwsem);
210 : : /* This may call shrinker, so it must use down_read_trylock() */
211 : : id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
212 : : if (id < 0)
213 : : goto unlock;
214 : :
215 : : if (id >= shrinker_nr_max) {
216 : : if (memcg_expand_shrinker_maps(id)) {
217 : : idr_remove(&shrinker_idr, id);
218 : : goto unlock;
219 : : }
220 : :
221 : : shrinker_nr_max = id + 1;
222 : : }
223 : : shrinker->id = id;
224 : : ret = 0;
225 : : unlock:
226 : : up_write(&shrinker_rwsem);
227 : : return ret;
228 : : }
229 : :
230 : : static void unregister_memcg_shrinker(struct shrinker *shrinker)
231 : : {
232 : : int id = shrinker->id;
233 : :
234 : : BUG_ON(id < 0);
235 : :
236 : : down_write(&shrinker_rwsem);
237 : : idr_remove(&shrinker_idr, id);
238 : : up_write(&shrinker_rwsem);
239 : : }
240 : :
241 : : static bool cgroup_reclaim(struct scan_control *sc)
242 : : {
243 : : return sc->target_mem_cgroup;
244 : : }
245 : :
246 : : /**
247 : : * writeback_throttling_sane - is the usual dirty throttling mechanism available?
248 : : * @sc: scan_control in question
249 : : *
250 : : * The normal page dirty throttling mechanism in balance_dirty_pages() is
251 : : * completely broken with the legacy memcg and direct stalling in
252 : : * shrink_page_list() is used for throttling instead, which lacks all the
253 : : * niceties such as fairness, adaptive pausing, bandwidth proportional
254 : : * allocation and configurability.
255 : : *
256 : : * This function tests whether the vmscan currently in progress can assume
257 : : * that the normal dirty throttling mechanism is operational.
258 : : */
259 : : static bool writeback_throttling_sane(struct scan_control *sc)
260 : : {
261 : : if (!cgroup_reclaim(sc))
262 : : return true;
263 : : #ifdef CONFIG_CGROUP_WRITEBACK
264 : : if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
265 : : return true;
266 : : #endif
267 : : return false;
268 : : }
269 : : #else
270 : : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
271 : : {
272 : : return 0;
273 : : }
274 : :
275 : : static void unregister_memcg_shrinker(struct shrinker *shrinker)
276 : : {
277 : : }
278 : :
279 : 0 : static bool cgroup_reclaim(struct scan_control *sc)
280 : : {
281 : 0 : return false;
282 : : }
283 : :
284 : 0 : static bool writeback_throttling_sane(struct scan_control *sc)
285 : : {
286 : 0 : return true;
287 : : }
288 : : #endif
289 : :
290 : : /*
291 : : * This misses isolated pages which are not accounted for to save counters.
292 : : * As the data only determines if reclaim or compaction continues, it is
293 : : * not expected that isolated pages will be a dominating factor.
294 : : */
295 : 0 : unsigned long zone_reclaimable_pages(struct zone *zone)
296 : : {
297 : 0 : unsigned long nr;
298 : :
299 : 0 : nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
300 : 0 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
301 [ # # ]: 0 : if (get_nr_swap_pages() > 0)
302 : 0 : nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
303 : 0 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
304 : :
305 : 0 : return nr;
306 : : }
307 : :
308 : : /**
309 : : * lruvec_lru_size - Returns the number of pages on the given LRU list.
310 : : * @lruvec: lru vector
311 : : * @lru: lru to use
312 : : * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
313 : : */
314 : 0 : unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, 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 : 1260 : int prealloc_shrinker(struct shrinker *shrinker)
337 : : {
338 : 1260 : unsigned int size = sizeof(*shrinker->nr_deferred);
339 : :
340 [ + + ]: 1260 : if (shrinker->flags & SHRINKER_NUMA_AWARE)
341 : 1064 : size *= nr_node_ids;
342 : :
343 : 1260 : shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
344 [ - + ]: 1260 : if (!shrinker->nr_deferred)
345 : 0 : return -ENOMEM;
346 : :
347 : : if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
348 : : 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 : : unregister_memcg_shrinker(shrinker);
367 : :
368 : 0 : kfree(shrinker->nr_deferred);
369 : 0 : shrinker->nr_deferred = NULL;
370 : : }
371 : :
372 : 1260 : void register_shrinker_prepared(struct shrinker *shrinker)
373 : : {
374 : 1260 : down_write(&shrinker_rwsem);
375 : 1260 : 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 : 1260 : up_write(&shrinker_rwsem);
381 : 1260 : }
382 : :
383 : 196 : int register_shrinker(struct shrinker *shrinker)
384 : : {
385 : 196 : int err = prealloc_shrinker(shrinker);
386 : :
387 [ + - ]: 196 : if (err)
388 : : return err;
389 : 196 : register_shrinker_prepared(shrinker);
390 : 196 : return 0;
391 : : }
392 : : EXPORT_SYMBOL(register_shrinker);
393 : :
394 : : /*
395 : : * Remove one
396 : : */
397 : 56 : void unregister_shrinker(struct shrinker *shrinker)
398 : : {
399 [ + - ]: 56 : if (!shrinker->nr_deferred)
400 : : return;
401 : 56 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
402 : : unregister_memcg_shrinker(shrinker);
403 : 56 : down_write(&shrinker_rwsem);
404 : 56 : list_del(&shrinker->list);
405 : 56 : up_write(&shrinker_rwsem);
406 : 56 : kfree(shrinker->nr_deferred);
407 : 56 : 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 regsitration 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 : freed = 0;
702 : 0 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
703 : 0 : do {
704 : 0 : freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
705 [ # # # # ]: 0 : } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
706 [ # # # # ]: 0 : } while (freed > 10);
707 : 0 : }
708 : :
709 : 0 : void drop_slab(void)
710 : : {
711 : 0 : int nid;
712 : :
713 [ # # ]: 0 : for_each_online_node(nid)
714 : : drop_slab_node(nid);
715 : 0 : }
716 : :
717 : 0 : static inline int is_page_cache_freeable(struct page *page)
718 : : {
719 : : /*
720 : : * A freeable page cache page is referenced only by the caller
721 : : * that isolated the page, the page cache and optional buffer
722 : : * heads at page->private.
723 : : */
724 [ # # ]: 0 : int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
725 : : HPAGE_PMD_NR : 1;
726 [ # # ]: 0 : return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
727 : : }
728 : :
729 : 0 : static int may_write_to_inode(struct inode *inode)
730 : : {
731 [ # # ]: 0 : if (current->flags & PF_SWAPWRITE)
732 : : return 1;
733 [ # # ]: 0 : if (!inode_write_congested(inode))
734 : : return 1;
735 [ # # ]: 0 : if (inode_to_bdi(inode) == current->backing_dev_info)
736 : 0 : return 1;
737 : : return 0;
738 : : }
739 : :
740 : : /*
741 : : * We detected a synchronous write error writing a page out. Probably
742 : : * -ENOSPC. We need to propagate that into the address_space for a subsequent
743 : : * fsync(), msync() or close().
744 : : *
745 : : * The tricky part is that after writepage we cannot touch the mapping: nothing
746 : : * prevents it from being freed up. But we have a ref on the page and once
747 : : * that page is locked, the mapping is pinned.
748 : : *
749 : : * We're allowed to run sleeping lock_page() here because we know the caller has
750 : : * __GFP_FS.
751 : : */
752 : 0 : static void handle_write_error(struct address_space *mapping,
753 : : struct page *page, int error)
754 : : {
755 : 0 : lock_page(page);
756 [ # # ]: 0 : if (page_mapping(page) == mapping)
757 : 0 : mapping_set_error(mapping, error);
758 : 0 : unlock_page(page);
759 : 0 : }
760 : :
761 : : /* possible outcome of pageout() */
762 : : typedef enum {
763 : : /* failed to write page out, page is locked */
764 : : PAGE_KEEP,
765 : : /* move page to the active list, page is locked */
766 : : PAGE_ACTIVATE,
767 : : /* page has been sent to the disk successfully, page is unlocked */
768 : : PAGE_SUCCESS,
769 : : /* page is clean and locked */
770 : : PAGE_CLEAN,
771 : : } pageout_t;
772 : :
773 : : /*
774 : : * pageout is called by shrink_page_list() for each dirty page.
775 : : * Calls ->writepage().
776 : : */
777 : 0 : static pageout_t pageout(struct page *page, struct address_space *mapping)
778 : : {
779 : : /*
780 : : * If the page is dirty, only perform writeback if that write
781 : : * will be non-blocking. To prevent this allocation from being
782 : : * stalled by pagecache activity. But note that there may be
783 : : * stalls if we need to run get_block(). We could test
784 : : * PagePrivate for that.
785 : : *
786 : : * If this process is currently in __generic_file_write_iter() against
787 : : * this page's queue, we can perform writeback even if that
788 : : * will block.
789 : : *
790 : : * If the page is swapcache, write it back even if that would
791 : : * block, for some throttling. This happens by accident, because
792 : : * swap_backing_dev_info is bust: it doesn't reflect the
793 : : * congestion state of the swapdevs. Easy to fix, if needed.
794 : : */
795 [ # # ]: 0 : if (!is_page_cache_freeable(page))
796 : : return PAGE_KEEP;
797 [ # # ]: 0 : if (!mapping) {
798 : : /*
799 : : * Some data journaling orphaned pages can have
800 : : * page->mapping == NULL while being dirty with clean buffers.
801 : : */
802 [ # # ]: 0 : if (page_has_private(page)) {
803 [ # # ]: 0 : if (try_to_free_buffers(page)) {
804 [ # # ]: 0 : ClearPageDirty(page);
805 : 0 : pr_info("%s: orphaned page\n", __func__);
806 : 0 : return PAGE_CLEAN;
807 : : }
808 : : }
809 : 0 : return PAGE_KEEP;
810 : : }
811 [ # # ]: 0 : if (mapping->a_ops->writepage == NULL)
812 : : return PAGE_ACTIVATE;
813 [ # # ]: 0 : if (!may_write_to_inode(mapping->host))
814 : : return PAGE_KEEP;
815 : :
816 [ # # ]: 0 : if (clear_page_dirty_for_io(page)) {
817 : 0 : int res;
818 : 0 : struct writeback_control wbc = {
819 : : .sync_mode = WB_SYNC_NONE,
820 : : .nr_to_write = SWAP_CLUSTER_MAX,
821 : : .range_start = 0,
822 : : .range_end = LLONG_MAX,
823 : : .for_reclaim = 1,
824 : : };
825 : :
826 [ # # ]: 0 : SetPageReclaim(page);
827 : 0 : res = mapping->a_ops->writepage(page, &wbc);
828 [ # # ]: 0 : if (res < 0)
829 : 0 : handle_write_error(mapping, page, res);
830 [ # # ]: 0 : if (res == AOP_WRITEPAGE_ACTIVATE) {
831 [ # # ]: 0 : ClearPageReclaim(page);
832 : 0 : return PAGE_ACTIVATE;
833 : : }
834 : :
835 [ # # # # ]: 0 : if (!PageWriteback(page)) {
836 : : /* synchronous write or broken a_ops? */
837 [ # # ]: 0 : ClearPageReclaim(page);
838 : : }
839 : 0 : trace_mm_vmscan_writepage(page);
840 : 0 : inc_node_page_state(page, NR_VMSCAN_WRITE);
841 : 0 : return PAGE_SUCCESS;
842 : : }
843 : :
844 : : return PAGE_CLEAN;
845 : : }
846 : :
847 : : /*
848 : : * Same as remove_mapping, but if the page is removed from the mapping, it
849 : : * gets returned with a refcount of 0.
850 : : */
851 : 0 : static int __remove_mapping(struct address_space *mapping, struct page *page,
852 : : bool reclaimed, struct mem_cgroup *target_memcg)
853 : : {
854 : 0 : unsigned long flags;
855 : 0 : int refcount;
856 : :
857 [ # # # # ]: 0 : BUG_ON(!PageLocked(page));
858 [ # # ]: 0 : BUG_ON(mapping != page_mapping(page));
859 : :
860 : 0 : xa_lock_irqsave(&mapping->i_pages, flags);
861 : : /*
862 : : * The non racy check for a busy page.
863 : : *
864 : : * Must be careful with the order of the tests. When someone has
865 : : * a ref to the page, it may be possible that they dirty it then
866 : : * drop the reference. So if PageDirty is tested before page_count
867 : : * here, then the following race may occur:
868 : : *
869 : : * get_user_pages(&page);
870 : : * [user mapping goes away]
871 : : * write_to(page);
872 : : * !PageDirty(page) [good]
873 : : * SetPageDirty(page);
874 : : * put_page(page);
875 : : * !page_count(page) [good, discard it]
876 : : *
877 : : * [oops, our write_to data is lost]
878 : : *
879 : : * Reversing the order of the tests ensures such a situation cannot
880 : : * escape unnoticed. The smp_rmb is needed to ensure the page->flags
881 : : * load is not satisfied before that of page->_refcount.
882 : : *
883 : : * Note that if SetPageDirty is always performed via set_page_dirty,
884 : : * and thus under the i_pages lock, then this ordering is not required.
885 : : */
886 : 0 : refcount = 1 + compound_nr(page);
887 [ # # ]: 0 : if (!page_ref_freeze(page, refcount))
888 : 0 : goto cannot_free;
889 : : /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
890 [ # # # # ]: 0 : if (unlikely(PageDirty(page))) {
891 : 0 : page_ref_unfreeze(page, refcount);
892 : 0 : goto cannot_free;
893 : : }
894 : :
895 [ # # ]: 0 : if (PageSwapCache(page)) {
896 : 0 : swp_entry_t swap = { .val = page_private(page) };
897 : 0 : mem_cgroup_swapout(page, swap);
898 : 0 : __delete_from_swap_cache(page, swap);
899 : 0 : xa_unlock_irqrestore(&mapping->i_pages, flags);
900 : 0 : put_swap_page(page, swap);
901 : : } else {
902 : 0 : void (*freepage)(struct page *);
903 : 0 : void *shadow = NULL;
904 : :
905 : 0 : freepage = mapping->a_ops->freepage;
906 : : /*
907 : : * Remember a shadow entry for reclaimed file cache in
908 : : * order to detect refaults, thus thrashing, later on.
909 : : *
910 : : * But don't store shadows in an address space that is
911 : : * already exiting. This is not just an optizimation,
912 : : * inode reclaim needs to empty out the radix tree or
913 : : * the nodes are lost. Don't plant shadows behind its
914 : : * back.
915 : : *
916 : : * We also don't store shadows for DAX mappings because the
917 : : * only page cache pages found in these are zero pages
918 : : * covering holes, and because we don't want to mix DAX
919 : : * exceptional entries and shadow exceptional entries in the
920 : : * same address_space.
921 : : */
922 [ # # # # : 0 : if (reclaimed && page_is_file_cache(page) &&
# # ]
923 : : !mapping_exiting(mapping) && !dax_mapping(mapping))
924 : 0 : shadow = workingset_eviction(page, target_memcg);
925 : 0 : __delete_from_page_cache(page, shadow);
926 : 0 : xa_unlock_irqrestore(&mapping->i_pages, flags);
927 : :
928 [ # # ]: 0 : if (freepage != NULL)
929 : 0 : freepage(page);
930 : : }
931 : :
932 : : return 1;
933 : :
934 : 0 : cannot_free:
935 : 0 : xa_unlock_irqrestore(&mapping->i_pages, flags);
936 : 0 : return 0;
937 : : }
938 : :
939 : : /*
940 : : * Attempt to detach a locked page from its ->mapping. If it is dirty or if
941 : : * someone else has a ref on the page, abort and return 0. If it was
942 : : * successfully detached, return 1. Assumes the caller has a single ref on
943 : : * this page.
944 : : */
945 : 0 : int remove_mapping(struct address_space *mapping, struct page *page)
946 : : {
947 [ # # ]: 0 : if (__remove_mapping(mapping, page, false, NULL)) {
948 : : /*
949 : : * Unfreezing the refcount with 1 rather than 2 effectively
950 : : * drops the pagecache ref for us without requiring another
951 : : * atomic operation.
952 : : */
953 : 0 : page_ref_unfreeze(page, 1);
954 : 0 : return 1;
955 : : }
956 : : return 0;
957 : : }
958 : :
959 : : /**
960 : : * putback_lru_page - put previously isolated page onto appropriate LRU list
961 : : * @page: page to be put back to appropriate lru list
962 : : *
963 : : * Add previously isolated @page to appropriate LRU list.
964 : : * Page may still be unevictable for other reasons.
965 : : *
966 : : * lru_lock must not be held, interrupts must be enabled.
967 : : */
968 : 0 : void putback_lru_page(struct page *page)
969 : : {
970 : 0 : lru_cache_add(page);
971 : 0 : put_page(page); /* drop ref from isolate */
972 : 0 : }
973 : :
974 : : enum page_references {
975 : : PAGEREF_RECLAIM,
976 : : PAGEREF_RECLAIM_CLEAN,
977 : : PAGEREF_KEEP,
978 : : PAGEREF_ACTIVATE,
979 : : };
980 : :
981 : : static enum page_references page_check_references(struct page *page,
982 : : struct scan_control *sc)
983 : : {
984 : : int referenced_ptes, referenced_page;
985 : : unsigned long vm_flags;
986 : :
987 : : referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
988 : : &vm_flags);
989 : : referenced_page = TestClearPageReferenced(page);
990 : :
991 : : /*
992 : : * Mlock lost the isolation race with us. Let try_to_unmap()
993 : : * move the page to the unevictable list.
994 : : */
995 : : if (vm_flags & VM_LOCKED)
996 : : return PAGEREF_RECLAIM;
997 : :
998 : : if (referenced_ptes) {
999 : : if (PageSwapBacked(page))
1000 : : return PAGEREF_ACTIVATE;
1001 : : /*
1002 : : * All mapped pages start out with page table
1003 : : * references from the instantiating fault, so we need
1004 : : * to look twice if a mapped file page is used more
1005 : : * than once.
1006 : : *
1007 : : * Mark it and spare it for another trip around the
1008 : : * inactive list. Another page table reference will
1009 : : * lead to its activation.
1010 : : *
1011 : : * Note: the mark is set for activated pages as well
1012 : : * so that recently deactivated but used pages are
1013 : : * quickly recovered.
1014 : : */
1015 : : SetPageReferenced(page);
1016 : :
1017 : : if (referenced_page || referenced_ptes > 1)
1018 : : return PAGEREF_ACTIVATE;
1019 : :
1020 : : /*
1021 : : * Activate file-backed executable pages after first usage.
1022 : : */
1023 : : if (vm_flags & VM_EXEC)
1024 : : return PAGEREF_ACTIVATE;
1025 : :
1026 : : return PAGEREF_KEEP;
1027 : : }
1028 : :
1029 : : /* Reclaim if clean, defer dirty pages to writeback */
1030 : : if (referenced_page && !PageSwapBacked(page))
1031 : : return PAGEREF_RECLAIM_CLEAN;
1032 : :
1033 : : return PAGEREF_RECLAIM;
1034 : : }
1035 : :
1036 : : /* Check if a page is dirty or under writeback */
1037 : 0 : static void page_check_dirty_writeback(struct page *page,
1038 : : bool *dirty, bool *writeback)
1039 : : {
1040 : 0 : struct address_space *mapping;
1041 : :
1042 : : /*
1043 : : * Anonymous pages are not handled by flushers and must be written
1044 : : * from reclaim context. Do not stall reclaim based on them
1045 : : */
1046 [ # # # # ]: 0 : if (!page_is_file_cache(page) ||
1047 [ # # ]: 0 : (PageAnon(page) && !PageSwapBacked(page))) {
1048 : 0 : *dirty = false;
1049 : 0 : *writeback = false;
1050 : 0 : return;
1051 : : }
1052 : :
1053 : : /* By default assume that the page flags are accurate */
1054 [ # # ]: 0 : *dirty = PageDirty(page);
1055 [ # # ]: 0 : *writeback = PageWriteback(page);
1056 : :
1057 : : /* Verify dirty/writeback state if the filesystem supports it */
1058 [ # # ]: 0 : if (!page_has_private(page))
1059 : : return;
1060 : :
1061 : 0 : mapping = page_mapping(page);
1062 [ # # # # ]: 0 : if (mapping && mapping->a_ops->is_dirty_writeback)
1063 : 0 : mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1064 : : }
1065 : :
1066 : : /*
1067 : : * shrink_page_list() returns the number of reclaimed pages
1068 : : */
1069 : 0 : static unsigned long shrink_page_list(struct list_head *page_list,
1070 : : struct pglist_data *pgdat,
1071 : : struct scan_control *sc,
1072 : : enum ttu_flags ttu_flags,
1073 : : struct reclaim_stat *stat,
1074 : : bool ignore_references)
1075 : : {
1076 : 0 : LIST_HEAD(ret_pages);
1077 : 0 : LIST_HEAD(free_pages);
1078 : 0 : unsigned nr_reclaimed = 0;
1079 : 0 : unsigned pgactivate = 0;
1080 : :
1081 : 0 : memset(stat, 0, sizeof(*stat));
1082 : 0 : cond_resched();
1083 : :
1084 [ # # ]: 0 : while (!list_empty(page_list)) {
1085 : 0 : struct address_space *mapping;
1086 : 0 : struct page *page;
1087 : 0 : int may_enter_fs;
1088 : 0 : enum page_references references = PAGEREF_RECLAIM;
1089 : 0 : bool dirty, writeback;
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 (PageTransHuge(page)) {
1243 : : /* cannot split THP, skip it */
1244 : : if (!can_split_huge_page(page, NULL))
1245 : : goto activate_locked;
1246 : : /*
1247 : : * Split pages without a PMD map right
1248 : : * away. Chances are some or all of the
1249 : : * tail pages can be freed without IO.
1250 : : */
1251 : : if (!compound_mapcount(page) &&
1252 : : split_huge_page_to_list(page,
1253 : : page_list))
1254 : : goto activate_locked;
1255 : : }
1256 [ # # ]: 0 : if (!add_to_swap(page)) {
1257 [ # # ]: 0 : if (!PageTransHuge(page))
1258 : 0 : goto activate_locked_split;
1259 : : /* Fallback to swap normal pages */
1260 : : if (split_huge_page_to_list(page,
1261 : : page_list))
1262 : : goto activate_locked;
1263 : : #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1264 : : count_vm_event(THP_SWPOUT_FALLBACK);
1265 : : #endif
1266 : : if (!add_to_swap(page))
1267 : : goto activate_locked_split;
1268 : : }
1269 : :
1270 : 0 : may_enter_fs = 1;
1271 : :
1272 : : /* Adding to swap updated mapping */
1273 : 0 : mapping = page_mapping(page);
1274 : : }
1275 : : } else if (unlikely(PageTransHuge(page))) {
1276 : : /* Split file THP */
1277 : : if (split_huge_page_to_list(page, page_list))
1278 : : goto keep_locked;
1279 : : }
1280 : :
1281 : : /*
1282 : : * THP may get split above, need minus tail pages and update
1283 : : * nr_pages to avoid accounting tail pages twice.
1284 : : *
1285 : : * The tail pages that are added into swap cache successfully
1286 : : * reach here.
1287 : : */
1288 [ # # ]: 0 : if ((nr_pages > 1) && !PageTransHuge(page)) {
1289 : 0 : sc->nr_scanned -= (nr_pages - 1);
1290 : 0 : nr_pages = 1;
1291 : : }
1292 : :
1293 : : /*
1294 : : * The page is mapped into the page tables of one or more
1295 : : * processes. Try to unmap it here.
1296 : : */
1297 [ # # ]: 0 : if (page_mapped(page)) {
1298 : 0 : enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1299 : :
1300 : 0 : if (unlikely(PageTransHuge(page)))
1301 : : flags |= TTU_SPLIT_HUGE_PMD;
1302 [ # # ]: 0 : if (!try_to_unmap(page, flags)) {
1303 : 0 : stat->nr_unmap_fail += nr_pages;
1304 : 0 : goto activate_locked;
1305 : : }
1306 : : }
1307 : :
1308 [ # # # # ]: 0 : if (PageDirty(page)) {
1309 : : /*
1310 : : * Only kswapd can writeback filesystem pages
1311 : : * to avoid risk of stack overflow. But avoid
1312 : : * injecting inefficient single-page IO into
1313 : : * flusher writeback as much as possible: only
1314 : : * write pages when we've encountered many
1315 : : * dirty pages, and when we've already scanned
1316 : : * the rest of the LRU for clean pages and see
1317 : : * the same dirty pages again (PageReclaim).
1318 : : */
1319 [ # # # # ]: 0 : if (page_is_file_cache(page) &&
1320 [ # # # # ]: 0 : (!current_is_kswapd() || !PageReclaim(page) ||
1321 : 0 : !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1322 : : /*
1323 : : * Immediately reclaim when written back.
1324 : : * Similar in principal to deactivate_page()
1325 : : * except we already have the page isolated
1326 : : * and know it's dirty
1327 : : */
1328 : 0 : inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1329 [ # # ]: 0 : SetPageReclaim(page);
1330 : :
1331 : 0 : goto activate_locked;
1332 : : }
1333 : :
1334 [ # # ]: 0 : if (references == PAGEREF_RECLAIM_CLEAN)
1335 : 0 : goto keep_locked;
1336 [ # # ]: 0 : if (!may_enter_fs)
1337 : 0 : goto keep_locked;
1338 [ # # ]: 0 : if (!sc->may_writepage)
1339 : 0 : goto keep_locked;
1340 : :
1341 : : /*
1342 : : * Page is dirty. Flush the TLB if a writable entry
1343 : : * potentially exists to avoid CPU writes after IO
1344 : : * starts and then write it out here.
1345 : : */
1346 : 0 : try_to_unmap_flush_dirty();
1347 [ # # # # ]: 0 : switch (pageout(page, mapping)) {
1348 : 0 : case PAGE_KEEP:
1349 : 0 : goto keep_locked;
1350 : 0 : case PAGE_ACTIVATE:
1351 : 0 : goto activate_locked;
1352 : : case PAGE_SUCCESS:
1353 [ # # # # ]: 0 : if (PageWriteback(page))
1354 : 0 : goto keep;
1355 [ # # # # ]: 0 : if (PageDirty(page))
1356 : 0 : goto keep;
1357 : :
1358 : : /*
1359 : : * A synchronous write - probably a ramdisk. Go
1360 : : * ahead and try to reclaim the page.
1361 : : */
1362 [ # # # # ]: 0 : if (!trylock_page(page))
1363 : 0 : goto keep;
1364 [ # # # # : 0 : if (PageDirty(page) || PageWriteback(page))
# # ]
1365 : 0 : goto keep_locked;
1366 : 0 : mapping = page_mapping(page);
1367 : 0 : case PAGE_CLEAN:
1368 : 0 : ; /* try to free the page below */
1369 : : }
1370 : : }
1371 : :
1372 : : /*
1373 : : * If the page has buffers, try to free the buffer mappings
1374 : : * associated with this page. If we succeed we try to free
1375 : : * the page as well.
1376 : : *
1377 : : * We do this even if the page is PageDirty().
1378 : : * try_to_release_page() does not perform I/O, but it is
1379 : : * possible for a page to have PageDirty set, but it is actually
1380 : : * clean (all its buffers are clean). This happens if the
1381 : : * buffers were written out directly, with submit_bh(). ext3
1382 : : * will do this, as well as the blockdev mapping.
1383 : : * try_to_release_page() will discover that cleanness and will
1384 : : * drop the buffers and mark the page clean - it can be freed.
1385 : : *
1386 : : * Rarely, pages can have buffers and no ->mapping. These are
1387 : : * the pages which were not successfully invalidated in
1388 : : * truncate_complete_page(). We try to drop those buffers here
1389 : : * and if that worked, and the page is no longer mapped into
1390 : : * process address space (page_count == 1) it can be freed.
1391 : : * Otherwise, leave the page on the LRU so it is swappable.
1392 : : */
1393 [ # # ]: 0 : if (page_has_private(page)) {
1394 [ # # ]: 0 : if (!try_to_release_page(page, sc->gfp_mask))
1395 : 0 : goto activate_locked;
1396 [ # # # # ]: 0 : if (!mapping && page_count(page) == 1) {
1397 : 0 : unlock_page(page);
1398 [ # # ]: 0 : if (put_page_testzero(page))
1399 : 0 : goto free_it;
1400 : : else {
1401 : : /*
1402 : : * rare race with speculative reference.
1403 : : * the speculative reference will free
1404 : : * this page shortly, so we may
1405 : : * increment nr_reclaimed here (and
1406 : : * leave it off the LRU).
1407 : : */
1408 : 0 : nr_reclaimed++;
1409 : 0 : continue;
1410 : : }
1411 : : }
1412 : : }
1413 : :
1414 [ # # # # : 0 : if (PageAnon(page) && !PageSwapBacked(page)) {
# # ]
1415 : : /* follow __remove_mapping for reference */
1416 [ # # ]: 0 : if (!page_ref_freeze(page, 1))
1417 : 0 : goto keep_locked;
1418 [ # # # # ]: 0 : if (PageDirty(page)) {
1419 : 0 : page_ref_unfreeze(page, 1);
1420 : 0 : goto keep_locked;
1421 : : }
1422 : :
1423 : 0 : count_vm_event(PGLAZYFREED);
1424 : : count_memcg_page_event(page, PGLAZYFREED);
1425 [ # # # # ]: 0 : } else if (!mapping || !__remove_mapping(mapping, page, true,
1426 : : sc->target_mem_cgroup))
1427 : 0 : goto keep_locked;
1428 : :
1429 : 0 : unlock_page(page);
1430 : 0 : free_it:
1431 : : /*
1432 : : * THP may get swapped out in a whole, need account
1433 : : * all base pages.
1434 : : */
1435 : 0 : nr_reclaimed += nr_pages;
1436 : :
1437 : : /*
1438 : : * Is there need to periodically free_page_list? It would
1439 : : * appear not as the counts should be low
1440 : : */
1441 : 0 : if (unlikely(PageTransHuge(page)))
1442 : : (*get_compound_page_dtor(page))(page);
1443 : : else
1444 : 0 : list_add(&page->lru, &free_pages);
1445 : 0 : continue;
1446 : :
1447 : : activate_locked_split:
1448 : : /*
1449 : : * The tail pages that are failed to add into swap cache
1450 : : * reach here. Fixup nr_scanned and nr_pages.
1451 : : */
1452 [ # # ]: 0 : if (nr_pages > 1) {
1453 : 0 : sc->nr_scanned -= (nr_pages - 1);
1454 : 0 : nr_pages = 1;
1455 : : }
1456 : 0 : activate_locked:
1457 : : /* Not a candidate for swapping, so reclaim swap space. */
1458 [ # # # # : 0 : if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
# # ]
1459 : : PageMlocked(page)))
1460 : 0 : try_to_free_swap(page);
1461 : 0 : VM_BUG_ON_PAGE(PageActive(page), page);
1462 [ # # # # ]: 0 : if (!PageMlocked(page)) {
1463 : 0 : int type = page_is_file_cache(page);
1464 [ # # ]: 0 : SetPageActive(page);
1465 : 0 : stat->nr_activate[type] += nr_pages;
1466 : 0 : count_memcg_page_event(page, PGACTIVATE);
1467 : : }
1468 : 0 : keep_locked:
1469 : 0 : unlock_page(page);
1470 : 0 : keep:
1471 : 0 : list_add(&page->lru, &ret_pages);
1472 : 0 : VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1473 : : }
1474 : :
1475 : 0 : pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1476 : :
1477 : 0 : mem_cgroup_uncharge_list(&free_pages);
1478 : 0 : try_to_unmap_flush();
1479 : 0 : free_unref_page_list(&free_pages);
1480 : :
1481 [ # # ]: 0 : list_splice(&ret_pages, page_list);
1482 [ # # ]: 0 : count_vm_events(PGACTIVATE, pgactivate);
1483 : :
1484 : 0 : return nr_reclaimed;
1485 : : }
1486 : :
1487 : 0 : unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1488 : : struct list_head *page_list)
1489 : : {
1490 : 0 : struct scan_control sc = {
1491 : : .gfp_mask = GFP_KERNEL,
1492 : : .priority = DEF_PRIORITY,
1493 : : .may_unmap = 1,
1494 : : };
1495 : 0 : struct reclaim_stat dummy_stat;
1496 : 0 : unsigned long ret;
1497 : 0 : struct page *page, *next;
1498 : 0 : LIST_HEAD(clean_pages);
1499 : :
1500 [ # # ]: 0 : list_for_each_entry_safe(page, next, page_list, lru) {
1501 [ # # # # : 0 : if (page_is_file_cache(page) && !PageDirty(page) &&
# # ]
1502 [ # # ]: 0 : !__PageMovable(page) && !PageUnevictable(page)) {
1503 [ # # ]: 0 : ClearPageActive(page);
1504 : 0 : list_move(&page->lru, &clean_pages);
1505 : : }
1506 : : }
1507 : :
1508 : 0 : ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1509 : : TTU_IGNORE_ACCESS, &dummy_stat, true);
1510 [ # # ]: 0 : list_splice(&clean_pages, page_list);
1511 : 0 : mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1512 : 0 : return ret;
1513 : : }
1514 : :
1515 : : /*
1516 : : * Attempt to remove the specified page from its LRU. Only take this page
1517 : : * if it is of the appropriate PageActive status. Pages which are being
1518 : : * freed elsewhere are also ignored.
1519 : : *
1520 : : * page: page to consider
1521 : : * mode: one of the LRU isolation modes defined above
1522 : : *
1523 : : * returns 0 on success, -ve errno on failure.
1524 : : */
1525 : 0 : int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1526 : : {
1527 : 0 : int ret = -EINVAL;
1528 : :
1529 : : /* Only take pages on the LRU. */
1530 [ # # # # ]: 0 : if (!PageLRU(page))
1531 : : return ret;
1532 : :
1533 : : /* Compaction should not handle unevictable pages but CMA can do so */
1534 [ # # # # : 0 : if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
# # ]
1535 : : return ret;
1536 : :
1537 : 0 : ret = -EBUSY;
1538 : :
1539 : : /*
1540 : : * To minimise LRU disruption, the caller can indicate that it only
1541 : : * wants to isolate pages it will be able to operate on without
1542 : : * blocking - clean pages for the most part.
1543 : : *
1544 : : * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1545 : : * that it is possible to migrate without blocking
1546 : : */
1547 [ # # ]: 0 : if (mode & ISOLATE_ASYNC_MIGRATE) {
1548 : : /* All the caller can do on PageWriteback is block */
1549 [ # # # # ]: 0 : if (PageWriteback(page))
1550 : : return ret;
1551 : :
1552 [ # # # # ]: 0 : if (PageDirty(page)) {
1553 : 0 : struct address_space *mapping;
1554 : 0 : bool migrate_dirty;
1555 : :
1556 : : /*
1557 : : * Only pages without mappings or that have a
1558 : : * ->migratepage callback are possible to migrate
1559 : : * without blocking. However, we can be racing with
1560 : : * truncation so it's necessary to lock the page
1561 : : * to stabilise the mapping as truncation holds
1562 : : * the page lock until after the page is removed
1563 : : * from the page cache.
1564 : : */
1565 [ # # # # ]: 0 : if (!trylock_page(page))
1566 : : return ret;
1567 : :
1568 : 0 : mapping = page_mapping(page);
1569 [ # # # # ]: 0 : migrate_dirty = !mapping || mapping->a_ops->migratepage;
1570 : 0 : unlock_page(page);
1571 [ # # ]: 0 : if (!migrate_dirty)
1572 : : return ret;
1573 : : }
1574 : : }
1575 : :
1576 [ # # # # ]: 0 : if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1577 : : return ret;
1578 : :
1579 [ # # ]: 0 : if (likely(get_page_unless_zero(page))) {
1580 : : /*
1581 : : * Be careful not to clear PageLRU until after we're
1582 : : * sure the page is not being freed elsewhere -- the
1583 : : * page release code relies on it.
1584 : : */
1585 [ # # ]: 0 : ClearPageLRU(page);
1586 : 0 : ret = 0;
1587 : : }
1588 : :
1589 : : return ret;
1590 : : }
1591 : :
1592 : :
1593 : : /*
1594 : : * Update LRU sizes after isolating pages. The LRU size updates must
1595 : : * be complete before mem_cgroup_update_lru_size due to a santity check.
1596 : : */
1597 : 0 : static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1598 : : enum lru_list lru, unsigned long *nr_zone_taken)
1599 : : {
1600 : 0 : int zid;
1601 : :
1602 [ # # ]: 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1603 [ # # ]: 0 : if (!nr_zone_taken[zid])
1604 : 0 : continue;
1605 : :
1606 : 0 : __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1607 : : #ifdef CONFIG_MEMCG
1608 : : mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1609 : : #endif
1610 : : }
1611 : :
1612 : : }
1613 : :
1614 : : /**
1615 : : * pgdat->lru_lock is heavily contended. Some of the functions that
1616 : : * shrink the lists perform better by taking out a batch of pages
1617 : : * and working on them outside the LRU lock.
1618 : : *
1619 : : * For pagecache intensive workloads, this function is the hottest
1620 : : * spot in the kernel (apart from copy_*_user functions).
1621 : : *
1622 : : * Appropriate locks must be held before calling this function.
1623 : : *
1624 : : * @nr_to_scan: The number of eligible pages to look through on the list.
1625 : : * @lruvec: The LRU vector to pull pages from.
1626 : : * @dst: The temp list to put pages on to.
1627 : : * @nr_scanned: The number of pages that were scanned.
1628 : : * @sc: The scan_control struct for this reclaim session
1629 : : * @mode: One of the LRU isolation modes
1630 : : * @lru: LRU list id for isolating
1631 : : *
1632 : : * returns how many pages were moved onto *@dst.
1633 : : */
1634 : 0 : static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1635 : : struct lruvec *lruvec, struct list_head *dst,
1636 : : unsigned long *nr_scanned, struct scan_control *sc,
1637 : : enum lru_list lru)
1638 : : {
1639 : 0 : struct list_head *src = &lruvec->lists[lru];
1640 : 0 : unsigned long nr_taken = 0;
1641 : 0 : unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1642 : 0 : unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1643 : 0 : unsigned long skipped = 0;
1644 : 0 : unsigned long scan, total_scan, nr_pages;
1645 : 0 : LIST_HEAD(pages_skipped);
1646 [ # # ]: 0 : isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1647 : :
1648 : 0 : total_scan = 0;
1649 : 0 : scan = 0;
1650 [ # # ]: 0 : while (scan < nr_to_scan && !list_empty(src)) {
1651 : 0 : struct page *page;
1652 : :
1653 : 0 : page = lru_to_page(src);
1654 [ # # ]: 0 : prefetchw_prev_lru_page(page, src, flags);
1655 : :
1656 : 0 : VM_BUG_ON_PAGE(!PageLRU(page), page);
1657 : :
1658 : 0 : nr_pages = compound_nr(page);
1659 : 0 : total_scan += nr_pages;
1660 : :
1661 [ # # ]: 0 : if (page_zonenum(page) > sc->reclaim_idx) {
1662 : 0 : list_move(&page->lru, &pages_skipped);
1663 : 0 : nr_skipped[page_zonenum(page)] += nr_pages;
1664 : 0 : continue;
1665 : : }
1666 : :
1667 : : /*
1668 : : * Do not count skipped pages because that makes the function
1669 : : * return with no isolated pages if the LRU mostly contains
1670 : : * ineligible pages. This causes the VM to not reclaim any
1671 : : * pages, triggering a premature OOM.
1672 : : *
1673 : : * Account all tail pages of THP. This would not cause
1674 : : * premature OOM since __isolate_lru_page() returns -EBUSY
1675 : : * only when the page is being freed somewhere else.
1676 : : */
1677 : 0 : scan += nr_pages;
1678 [ # # # ]: 0 : switch (__isolate_lru_page(page, mode)) {
1679 : 0 : case 0:
1680 : 0 : nr_taken += nr_pages;
1681 : 0 : nr_zone_taken[page_zonenum(page)] += nr_pages;
1682 : 0 : list_move(&page->lru, dst);
1683 : : break;
1684 : :
1685 : 0 : case -EBUSY:
1686 : : /* else it is being freed elsewhere */
1687 : 0 : list_move(&page->lru, src);
1688 : 0 : continue;
1689 : :
1690 : 0 : default:
1691 [ # # ]: 0 : BUG();
1692 : : }
1693 : : }
1694 : :
1695 : : /*
1696 : : * Splice any skipped pages to the start of the LRU list. Note that
1697 : : * this disrupts the LRU order when reclaiming for lower zones but
1698 : : * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1699 : : * scanning would soon rescan the same pages to skip and put the
1700 : : * system at risk of premature OOM.
1701 : : */
1702 [ # # ]: 0 : if (!list_empty(&pages_skipped)) {
1703 : 0 : int zid;
1704 : :
1705 [ # # ]: 0 : list_splice(&pages_skipped, src);
1706 [ # # ]: 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1707 [ # # ]: 0 : if (!nr_skipped[zid])
1708 : 0 : continue;
1709 : :
1710 [ # # ]: 0 : __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1711 : 0 : skipped += nr_skipped[zid];
1712 : : }
1713 : : }
1714 : 0 : *nr_scanned = total_scan;
1715 : 0 : trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1716 : : total_scan, skipped, nr_taken, mode, lru);
1717 : 0 : update_lru_sizes(lruvec, lru, nr_zone_taken);
1718 : 0 : return nr_taken;
1719 : : }
1720 : :
1721 : : /**
1722 : : * isolate_lru_page - tries to isolate a page from its LRU list
1723 : : * @page: page to isolate from its LRU list
1724 : : *
1725 : : * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1726 : : * vmstat statistic corresponding to whatever LRU list the page was on.
1727 : : *
1728 : : * Returns 0 if the page was removed from an LRU list.
1729 : : * Returns -EBUSY if the page was not on an LRU list.
1730 : : *
1731 : : * The returned page will have PageLRU() cleared. If it was found on
1732 : : * the active list, it will have PageActive set. If it was found on
1733 : : * the unevictable list, it will have the PageUnevictable bit set. That flag
1734 : : * may need to be cleared by the caller before letting the page go.
1735 : : *
1736 : : * The vmstat statistic corresponding to the list on which the page was
1737 : : * found will be decremented.
1738 : : *
1739 : : * Restrictions:
1740 : : *
1741 : : * (1) Must be called with an elevated refcount on the page. This is a
1742 : : * fundamentnal difference from isolate_lru_pages (which is called
1743 : : * without a stable reference).
1744 : : * (2) the lru_lock must not be held.
1745 : : * (3) interrupts must be enabled.
1746 : : */
1747 : 0 : int isolate_lru_page(struct page *page)
1748 : : {
1749 : 0 : int ret = -EBUSY;
1750 : :
1751 : 0 : VM_BUG_ON_PAGE(!page_count(page), page);
1752 [ # # # # ]: 0 : WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1753 : :
1754 [ # # # # ]: 0 : if (PageLRU(page)) {
1755 : 0 : pg_data_t *pgdat = page_pgdat(page);
1756 : 0 : struct lruvec *lruvec;
1757 : :
1758 : 0 : spin_lock_irq(&pgdat->lru_lock);
1759 [ # # ]: 0 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
1760 [ # # # # ]: 0 : if (PageLRU(page)) {
1761 [ # # ]: 0 : int lru = page_lru(page);
1762 [ # # ]: 0 : get_page(page);
1763 [ # # ]: 0 : ClearPageLRU(page);
1764 : 0 : del_page_from_lru_list(page, lruvec, lru);
1765 : 0 : ret = 0;
1766 : : }
1767 : 0 : spin_unlock_irq(&pgdat->lru_lock);
1768 : : }
1769 : 0 : return ret;
1770 : : }
1771 : :
1772 : : /*
1773 : : * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1774 : : * then get rescheduled. When there are massive number of tasks doing page
1775 : : * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1776 : : * the LRU list will go small and be scanned faster than necessary, leading to
1777 : : * unnecessary swapping, thrashing and OOM.
1778 : : */
1779 : : static int too_many_isolated(struct pglist_data *pgdat, int file,
1780 : : struct scan_control *sc)
1781 : : {
1782 : : unsigned long inactive, isolated;
1783 : :
1784 : : if (current_is_kswapd())
1785 : : return 0;
1786 : :
1787 : : if (!writeback_throttling_sane(sc))
1788 : : return 0;
1789 : :
1790 : : if (file) {
1791 : : inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1792 : : isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1793 : : } else {
1794 : : inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1795 : : isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1796 : : }
1797 : :
1798 : : /*
1799 : : * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1800 : : * won't get blocked by normal direct-reclaimers, forming a circular
1801 : : * deadlock.
1802 : : */
1803 : : if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1804 : : inactive >>= 3;
1805 : :
1806 : : return isolated > inactive;
1807 : : }
1808 : :
1809 : : /*
1810 : : * This moves pages from @list to corresponding LRU list.
1811 : : *
1812 : : * We move them the other way if the page is referenced by one or more
1813 : : * processes, from rmap.
1814 : : *
1815 : : * If the pages are mostly unmapped, the processing is fast and it is
1816 : : * appropriate to hold zone_lru_lock across the whole operation. But if
1817 : : * the pages are mapped, the processing is slow (page_referenced()) so we
1818 : : * should drop zone_lru_lock around each page. It's impossible to balance
1819 : : * this, so instead we remove the pages from the LRU while processing them.
1820 : : * It is safe to rely on PG_active against the non-LRU pages in here because
1821 : : * nobody will play with that bit on a non-LRU page.
1822 : : *
1823 : : * The downside is that we have to touch page->_refcount against each page.
1824 : : * But we had to alter page->flags anyway.
1825 : : *
1826 : : * Returns the number of pages moved to the given lruvec.
1827 : : */
1828 : :
1829 : 0 : static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1830 : : struct list_head *list)
1831 : : {
1832 : 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1833 : 0 : int nr_pages, nr_moved = 0;
1834 : 0 : LIST_HEAD(pages_to_free);
1835 : 0 : struct page *page;
1836 : 0 : enum lru_list lru;
1837 : :
1838 [ # # ]: 0 : while (!list_empty(list)) {
1839 : 0 : page = lru_to_page(list);
1840 : 0 : VM_BUG_ON_PAGE(PageLRU(page), page);
1841 [ # # ]: 0 : if (unlikely(!page_evictable(page))) {
1842 : 0 : list_del(&page->lru);
1843 : 0 : spin_unlock_irq(&pgdat->lru_lock);
1844 : 0 : putback_lru_page(page);
1845 : 0 : spin_lock_irq(&pgdat->lru_lock);
1846 : 0 : continue;
1847 : : }
1848 [ # # ]: 0 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
1849 : :
1850 [ # # ]: 0 : SetPageLRU(page);
1851 [ # # ]: 0 : lru = page_lru(page);
1852 : :
1853 : 0 : nr_pages = hpage_nr_pages(page);
1854 : 0 : update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1855 : 0 : list_move(&page->lru, &lruvec->lists[lru]);
1856 : :
1857 [ # # ]: 0 : if (put_page_testzero(page)) {
1858 [ # # ]: 0 : __ClearPageLRU(page);
1859 [ # # ]: 0 : __ClearPageActive(page);
1860 : 0 : del_page_from_lru_list(page, lruvec, lru);
1861 : :
1862 [ # # ]: 0 : if (unlikely(PageCompound(page))) {
1863 : 0 : spin_unlock_irq(&pgdat->lru_lock);
1864 : 0 : (*get_compound_page_dtor(page))(page);
1865 : 0 : spin_lock_irq(&pgdat->lru_lock);
1866 : : } else
1867 : 0 : list_add(&page->lru, &pages_to_free);
1868 : : } else {
1869 : 0 : nr_moved += nr_pages;
1870 : : }
1871 : : }
1872 : :
1873 : : /*
1874 : : * To save our caller's stack, now use input list for pages to free.
1875 : : */
1876 [ # # ]: 0 : list_splice(&pages_to_free, list);
1877 : :
1878 : 0 : return nr_moved;
1879 : : }
1880 : :
1881 : : /*
1882 : : * If a kernel thread (such as nfsd for loop-back mounts) services
1883 : : * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1884 : : * In that case we should only throttle if the backing device it is
1885 : : * writing to is congested. In other cases it is safe to throttle.
1886 : : */
1887 : 0 : static int current_may_throttle(void)
1888 : : {
1889 [ # # ]: 0 : return !(current->flags & PF_LESS_THROTTLE) ||
1890 [ # # # # : 0 : current->backing_dev_info == NULL ||
# # ]
1891 [ # # ]: 0 : bdi_write_congested(current->backing_dev_info);
1892 : : }
1893 : :
1894 : : /*
1895 : : * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1896 : : * of reclaimed pages
1897 : : */
1898 : : static noinline_for_stack unsigned long
1899 : 0 : shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1900 : : struct scan_control *sc, enum lru_list lru)
1901 : : {
1902 : 0 : LIST_HEAD(page_list);
1903 : 0 : unsigned long nr_scanned;
1904 : 0 : unsigned long nr_reclaimed = 0;
1905 : 0 : unsigned long nr_taken;
1906 : 0 : struct reclaim_stat stat;
1907 : 0 : int file = is_file_lru(lru);
1908 : 0 : enum vm_event_item item;
1909 : 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1910 : 0 : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1911 : 0 : bool stalled = false;
1912 : :
1913 [ # # ]: 0 : while (unlikely(too_many_isolated(pgdat, file, sc))) {
1914 [ # # ]: 0 : if (stalled)
1915 : : return 0;
1916 : :
1917 : : /* wait a bit for the reclaimer. */
1918 : 0 : msleep(100);
1919 : 0 : stalled = true;
1920 : :
1921 : : /* We are about to die and free our memory. Return now. */
1922 [ # # ]: 0 : if (fatal_signal_pending(current))
1923 : : return SWAP_CLUSTER_MAX;
1924 : : }
1925 : :
1926 : 0 : lru_add_drain();
1927 : :
1928 : 0 : spin_lock_irq(&pgdat->lru_lock);
1929 : :
1930 : 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1931 : : &nr_scanned, sc, lru);
1932 : :
1933 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1934 : 0 : reclaim_stat->recent_scanned[file] += nr_taken;
1935 : :
1936 [ # # ]: 0 : item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1937 : 0 : if (!cgroup_reclaim(sc))
1938 [ # # ]: 0 : __count_vm_events(item, nr_scanned);
1939 : 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1940 : 0 : spin_unlock_irq(&pgdat->lru_lock);
1941 : :
1942 [ # # ]: 0 : if (nr_taken == 0)
1943 : : return 0;
1944 : :
1945 : 0 : nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1946 : : &stat, false);
1947 : :
1948 : 0 : spin_lock_irq(&pgdat->lru_lock);
1949 : :
1950 [ # # ]: 0 : item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1951 : 0 : if (!cgroup_reclaim(sc))
1952 [ # # ]: 0 : __count_vm_events(item, nr_reclaimed);
1953 : 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1954 : 0 : reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
1955 : 0 : reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
1956 : :
1957 : 0 : move_pages_to_lru(lruvec, &page_list);
1958 : :
1959 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1960 : :
1961 : 0 : spin_unlock_irq(&pgdat->lru_lock);
1962 : :
1963 : 0 : mem_cgroup_uncharge_list(&page_list);
1964 : 0 : free_unref_page_list(&page_list);
1965 : :
1966 : : /*
1967 : : * If dirty pages are scanned that are not queued for IO, it
1968 : : * implies that flushers are not doing their job. This can
1969 : : * happen when memory pressure pushes dirty pages to the end of
1970 : : * the LRU before the dirty limits are breached and the dirty
1971 : : * data has expired. It can also happen when the proportion of
1972 : : * dirty pages grows not through writes but through memory
1973 : : * pressure reclaiming all the clean cache. And in some cases,
1974 : : * the flushers simply cannot keep up with the allocation
1975 : : * rate. Nudge the flusher threads in case they are asleep.
1976 : : */
1977 [ # # ]: 0 : if (stat.nr_unqueued_dirty == nr_taken)
1978 : 0 : wakeup_flusher_threads(WB_REASON_VMSCAN);
1979 : :
1980 : 0 : sc->nr.dirty += stat.nr_dirty;
1981 : 0 : sc->nr.congested += stat.nr_congested;
1982 : 0 : sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1983 : 0 : sc->nr.writeback += stat.nr_writeback;
1984 : 0 : sc->nr.immediate += stat.nr_immediate;
1985 : 0 : sc->nr.taken += nr_taken;
1986 [ # # ]: 0 : if (file)
1987 : 0 : sc->nr.file_taken += nr_taken;
1988 : :
1989 : 0 : trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1990 : 0 : nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1991 : 0 : return nr_reclaimed;
1992 : : }
1993 : :
1994 : 0 : static void shrink_active_list(unsigned long nr_to_scan,
1995 : : struct lruvec *lruvec,
1996 : : struct scan_control *sc,
1997 : : enum lru_list lru)
1998 : : {
1999 : 0 : unsigned long nr_taken;
2000 : 0 : unsigned long nr_scanned;
2001 : 0 : unsigned long vm_flags;
2002 : 0 : LIST_HEAD(l_hold); /* The pages which were snipped off */
2003 : 0 : LIST_HEAD(l_active);
2004 : 0 : LIST_HEAD(l_inactive);
2005 : 0 : struct page *page;
2006 : 0 : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2007 : 0 : unsigned nr_deactivate, nr_activate;
2008 : 0 : unsigned nr_rotated = 0;
2009 : 0 : int file = is_file_lru(lru);
2010 : 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2011 : :
2012 : 0 : lru_add_drain();
2013 : :
2014 : 0 : spin_lock_irq(&pgdat->lru_lock);
2015 : :
2016 : 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2017 : : &nr_scanned, sc, lru);
2018 : :
2019 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2020 : 0 : reclaim_stat->recent_scanned[file] += nr_taken;
2021 : :
2022 [ # # ]: 0 : __count_vm_events(PGREFILL, nr_scanned);
2023 : 0 : __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2024 : :
2025 : 0 : spin_unlock_irq(&pgdat->lru_lock);
2026 : :
2027 [ # # ]: 0 : while (!list_empty(&l_hold)) {
2028 : 0 : cond_resched();
2029 : 0 : page = lru_to_page(&l_hold);
2030 : 0 : list_del(&page->lru);
2031 : :
2032 [ # # ]: 0 : if (unlikely(!page_evictable(page))) {
2033 : 0 : putback_lru_page(page);
2034 : 0 : continue;
2035 : : }
2036 : :
2037 [ # # ]: 0 : if (unlikely(buffer_heads_over_limit)) {
2038 [ # # # # ]: 0 : if (page_has_private(page) && trylock_page(page)) {
2039 [ # # ]: 0 : if (page_has_private(page))
2040 : 0 : try_to_release_page(page, 0);
2041 : 0 : unlock_page(page);
2042 : : }
2043 : : }
2044 : :
2045 [ # # ]: 0 : if (page_referenced(page, 0, sc->target_mem_cgroup,
2046 : : &vm_flags)) {
2047 : 0 : nr_rotated += hpage_nr_pages(page);
2048 : : /*
2049 : : * Identify referenced, file-backed active pages and
2050 : : * give them one more trip around the active list. So
2051 : : * that executable code get better chances to stay in
2052 : : * memory under moderate memory pressure. Anon pages
2053 : : * are not likely to be evicted by use-once streaming
2054 : : * IO, plus JVM can create lots of anon VM_EXEC pages,
2055 : : * so we ignore them here.
2056 : : */
2057 [ # # # # ]: 0 : if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2058 : 0 : list_add(&page->lru, &l_active);
2059 : 0 : continue;
2060 : : }
2061 : : }
2062 : :
2063 [ # # ]: 0 : ClearPageActive(page); /* we are de-activating */
2064 [ # # ]: 0 : SetPageWorkingset(page);
2065 : 0 : list_add(&page->lru, &l_inactive);
2066 : : }
2067 : :
2068 : : /*
2069 : : * Move pages back to the lru list.
2070 : : */
2071 : 0 : spin_lock_irq(&pgdat->lru_lock);
2072 : : /*
2073 : : * Count referenced pages from currently used mappings as rotated,
2074 : : * even though only some of them are actually re-activated. This
2075 : : * helps balance scan pressure between file and anonymous pages in
2076 : : * get_scan_count.
2077 : : */
2078 : 0 : reclaim_stat->recent_rotated[file] += nr_rotated;
2079 : :
2080 : 0 : nr_activate = move_pages_to_lru(lruvec, &l_active);
2081 : 0 : nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2082 : : /* Keep all free pages in l_active list */
2083 [ # # ]: 0 : list_splice(&l_inactive, &l_active);
2084 : :
2085 [ # # ]: 0 : __count_vm_events(PGDEACTIVATE, nr_deactivate);
2086 : 0 : __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2087 : :
2088 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2089 : 0 : spin_unlock_irq(&pgdat->lru_lock);
2090 : :
2091 : 0 : mem_cgroup_uncharge_list(&l_active);
2092 : 0 : free_unref_page_list(&l_active);
2093 : 0 : trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2094 : 0 : nr_deactivate, nr_rotated, sc->priority, file);
2095 : 0 : }
2096 : :
2097 : 0 : unsigned long reclaim_pages(struct list_head *page_list)
2098 : : {
2099 : 0 : int nid = -1;
2100 : 0 : unsigned long nr_reclaimed = 0;
2101 : 0 : LIST_HEAD(node_page_list);
2102 : 0 : struct reclaim_stat dummy_stat;
2103 : 0 : struct page *page;
2104 : 0 : struct scan_control sc = {
2105 : : .gfp_mask = GFP_KERNEL,
2106 : : .priority = DEF_PRIORITY,
2107 : : .may_writepage = 1,
2108 : : .may_unmap = 1,
2109 : : .may_swap = 1,
2110 : : };
2111 : :
2112 [ # # ]: 0 : while (!list_empty(page_list)) {
2113 : 0 : page = lru_to_page(page_list);
2114 [ # # ]: 0 : if (nid == -1) {
2115 : 0 : nid = page_to_nid(page);
2116 : 0 : INIT_LIST_HEAD(&node_page_list);
2117 : : }
2118 : :
2119 [ # # ]: 0 : if (nid == page_to_nid(page)) {
2120 [ # # ]: 0 : ClearPageActive(page);
2121 : 0 : list_move(&page->lru, &node_page_list);
2122 : 0 : continue;
2123 : : }
2124 : :
2125 : 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2126 : : NODE_DATA(nid),
2127 : : &sc, 0,
2128 : : &dummy_stat, false);
2129 [ # # ]: 0 : while (!list_empty(&node_page_list)) {
2130 : 0 : page = lru_to_page(&node_page_list);
2131 : 0 : list_del(&page->lru);
2132 : 0 : putback_lru_page(page);
2133 : : }
2134 : :
2135 : : nid = -1;
2136 : : }
2137 : :
2138 [ # # ]: 0 : if (!list_empty(&node_page_list)) {
2139 : 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2140 : : NODE_DATA(nid),
2141 : : &sc, 0,
2142 : : &dummy_stat, false);
2143 [ # # ]: 0 : while (!list_empty(&node_page_list)) {
2144 : 0 : page = lru_to_page(&node_page_list);
2145 : 0 : list_del(&page->lru);
2146 : 0 : putback_lru_page(page);
2147 : : }
2148 : : }
2149 : :
2150 : 0 : return nr_reclaimed;
2151 : : }
2152 : :
2153 : 0 : static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2154 : : struct lruvec *lruvec, struct scan_control *sc)
2155 : : {
2156 [ # # ]: 0 : if (is_active_lru(lru)) {
2157 [ # # ]: 0 : if (sc->may_deactivate & (1 << is_file_lru(lru)))
2158 : 0 : shrink_active_list(nr_to_scan, lruvec, sc, lru);
2159 : : else
2160 : 0 : sc->skipped_deactivate = 1;
2161 : 0 : return 0;
2162 : : }
2163 : :
2164 : 0 : return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2165 : : }
2166 : :
2167 : : /*
2168 : : * The inactive anon list should be small enough that the VM never has
2169 : : * to do too much work.
2170 : : *
2171 : : * The inactive file list should be small enough to leave most memory
2172 : : * to the established workingset on the scan-resistant active list,
2173 : : * but large enough to avoid thrashing the aggregate readahead window.
2174 : : *
2175 : : * Both inactive lists should also be large enough that each inactive
2176 : : * page has a chance to be referenced again before it is reclaimed.
2177 : : *
2178 : : * If that fails and refaulting is observed, the inactive list grows.
2179 : : *
2180 : : * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2181 : : * on this LRU, maintained by the pageout code. An inactive_ratio
2182 : : * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2183 : : *
2184 : : * total target max
2185 : : * memory ratio inactive
2186 : : * -------------------------------------
2187 : : * 10MB 1 5MB
2188 : : * 100MB 1 50MB
2189 : : * 1GB 3 250MB
2190 : : * 10GB 10 0.9GB
2191 : : * 100GB 31 3GB
2192 : : * 1TB 101 10GB
2193 : : * 10TB 320 32GB
2194 : : */
2195 : 0 : static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2196 : : {
2197 : 0 : enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2198 : 0 : unsigned long inactive, active;
2199 : 0 : unsigned long inactive_ratio;
2200 : 0 : unsigned long gb;
2201 : :
2202 : 0 : inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2203 : 0 : active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2204 : :
2205 : 0 : gb = (inactive + active) >> (30 - PAGE_SHIFT);
2206 [ # # ]: 0 : if (gb)
2207 : 0 : inactive_ratio = int_sqrt(10 * gb);
2208 : : else
2209 : : inactive_ratio = 1;
2210 : :
2211 : 0 : return inactive * inactive_ratio < active;
2212 : : }
2213 : :
2214 : : enum scan_balance {
2215 : : SCAN_EQUAL,
2216 : : SCAN_FRACT,
2217 : : SCAN_ANON,
2218 : : SCAN_FILE,
2219 : : };
2220 : :
2221 : : /*
2222 : : * Determine how aggressively the anon and file LRU lists should be
2223 : : * scanned. The relative value of each set of LRU lists is determined
2224 : : * by looking at the fraction of the pages scanned we did rotate back
2225 : : * onto the active list instead of evict.
2226 : : *
2227 : : * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2228 : : * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2229 : : */
2230 : 0 : static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2231 : : unsigned long *nr)
2232 : : {
2233 [ # # ]: 0 : struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2234 [ # # ]: 0 : int swappiness = mem_cgroup_swappiness(memcg);
2235 : 0 : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2236 : 0 : u64 fraction[2];
2237 : 0 : u64 denominator = 0; /* gcc */
2238 [ # # ]: 0 : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2239 : 0 : unsigned long anon_prio, file_prio;
2240 : 0 : enum scan_balance scan_balance;
2241 : 0 : unsigned long anon, file;
2242 : 0 : unsigned long ap, fp;
2243 : 0 : enum lru_list lru;
2244 : :
2245 : : /* If we have no swap space, do not bother scanning anon pages. */
2246 [ # # # # ]: 0 : if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2247 : 0 : scan_balance = SCAN_FILE;
2248 : 0 : goto out;
2249 : : }
2250 : :
2251 : : /*
2252 : : * Global reclaim will swap to prevent OOM even with no
2253 : : * swappiness, but memcg users want to use this knob to
2254 : : * disable swapping for individual groups completely when
2255 : : * using the memory controller's swap limit feature would be
2256 : : * too expensive.
2257 : : */
2258 : 0 : if (cgroup_reclaim(sc) && !swappiness) {
2259 : : scan_balance = SCAN_FILE;
2260 : : goto out;
2261 : : }
2262 : :
2263 : : /*
2264 : : * Do not apply any pressure balancing cleverness when the
2265 : : * system is close to OOM, scan both anon and file equally
2266 : : * (unless the swappiness setting disagrees with swapping).
2267 : : */
2268 [ # # # # ]: 0 : if (!sc->priority && swappiness) {
2269 : 0 : scan_balance = SCAN_EQUAL;
2270 : 0 : goto out;
2271 : : }
2272 : :
2273 : : /*
2274 : : * If the system is almost out of file pages, force-scan anon.
2275 : : */
2276 [ # # ]: 0 : if (sc->file_is_tiny) {
2277 : 0 : scan_balance = SCAN_ANON;
2278 : 0 : goto out;
2279 : : }
2280 : :
2281 : : /*
2282 : : * If there is enough inactive page cache, we do not reclaim
2283 : : * anything from the anonymous working right now.
2284 : : */
2285 [ # # ]: 0 : if (sc->cache_trim_mode) {
2286 : 0 : scan_balance = SCAN_FILE;
2287 : 0 : goto out;
2288 : : }
2289 : :
2290 : 0 : scan_balance = SCAN_FRACT;
2291 : :
2292 : : /*
2293 : : * With swappiness at 100, anonymous and file have the same priority.
2294 : : * This scanning priority is essentially the inverse of IO cost.
2295 : : */
2296 : 0 : anon_prio = swappiness;
2297 : 0 : file_prio = 200 - anon_prio;
2298 : :
2299 : : /*
2300 : : * OK, so we have swap space and a fair amount of page cache
2301 : : * pages. We use the recently rotated / recently scanned
2302 : : * ratios to determine how valuable each cache is.
2303 : : *
2304 : : * Because workloads change over time (and to avoid overflow)
2305 : : * we keep these statistics as a floating average, which ends
2306 : : * up weighing recent references more than old ones.
2307 : : *
2308 : : * anon in [0], file in [1]
2309 : : */
2310 : :
2311 : 0 : anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2312 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2313 : 0 : file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2314 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2315 : :
2316 : 0 : spin_lock_irq(&pgdat->lru_lock);
2317 [ # # ]: 0 : if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2318 : 0 : reclaim_stat->recent_scanned[0] /= 2;
2319 : 0 : reclaim_stat->recent_rotated[0] /= 2;
2320 : : }
2321 : :
2322 [ # # ]: 0 : if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2323 : 0 : reclaim_stat->recent_scanned[1] /= 2;
2324 : 0 : reclaim_stat->recent_rotated[1] /= 2;
2325 : : }
2326 : :
2327 : : /*
2328 : : * The amount of pressure on anon vs file pages is inversely
2329 : : * proportional to the fraction of recently scanned pages on
2330 : : * each list that were recently referenced and in active use.
2331 : : */
2332 : 0 : ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2333 : 0 : ap /= reclaim_stat->recent_rotated[0] + 1;
2334 : :
2335 : 0 : fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2336 : 0 : fp /= reclaim_stat->recent_rotated[1] + 1;
2337 : 0 : spin_unlock_irq(&pgdat->lru_lock);
2338 : :
2339 : 0 : fraction[0] = ap;
2340 : 0 : fraction[1] = fp;
2341 : 0 : denominator = ap + fp + 1;
2342 : 0 : out:
2343 [ # # ]: 0 : for_each_evictable_lru(lru) {
2344 : 0 : int file = is_file_lru(lru);
2345 : 0 : unsigned long lruvec_size;
2346 : 0 : unsigned long scan;
2347 : 0 : unsigned long protection;
2348 : :
2349 : 0 : lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2350 [ # # # ]: 0 : protection = mem_cgroup_protection(memcg,
2351 : : sc->memcg_low_reclaim);
2352 : :
2353 : 0 : if (protection) {
2354 : : /*
2355 : : * Scale a cgroup's reclaim pressure by proportioning
2356 : : * its current usage to its memory.low or memory.min
2357 : : * setting.
2358 : : *
2359 : : * This is important, as otherwise scanning aggression
2360 : : * becomes extremely binary -- from nothing as we
2361 : : * approach the memory protection threshold, to totally
2362 : : * nominal as we exceed it. This results in requiring
2363 : : * setting extremely liberal protection thresholds. It
2364 : : * also means we simply get no protection at all if we
2365 : : * set it too low, which is not ideal.
2366 : : *
2367 : : * If there is any protection in place, we reduce scan
2368 : : * pressure by how much of the total memory used is
2369 : : * within protection thresholds.
2370 : : *
2371 : : * There is one special case: in the first reclaim pass,
2372 : : * we skip over all groups that are within their low
2373 : : * protection. If that fails to reclaim enough pages to
2374 : : * satisfy the reclaim goal, we come back and override
2375 : : * the best-effort low protection. However, we still
2376 : : * ideally want to honor how well-behaved groups are in
2377 : : * that case instead of simply punishing them all
2378 : : * equally. As such, we reclaim them based on how much
2379 : : * memory they are using, reducing the scan pressure
2380 : : * again by how much of the total memory used is under
2381 : : * hard protection.
2382 : : */
2383 : : unsigned long cgroup_size = mem_cgroup_size(memcg);
2384 : :
2385 : : /* Avoid TOCTOU with earlier protection check */
2386 : : cgroup_size = max(cgroup_size, protection);
2387 : :
2388 : : scan = lruvec_size - lruvec_size * protection /
2389 : : cgroup_size;
2390 : :
2391 : : /*
2392 : : * Minimally target SWAP_CLUSTER_MAX pages to keep
2393 : : * reclaim moving forwards, avoiding decremeting
2394 : : * sc->priority further than desirable.
2395 : : */
2396 : : scan = max(scan, SWAP_CLUSTER_MAX);
2397 : : } else {
2398 : 0 : scan = lruvec_size;
2399 : : }
2400 : :
2401 : 0 : scan >>= sc->priority;
2402 : :
2403 : : /*
2404 : : * If the cgroup's already been deleted, make sure to
2405 : : * scrape out the remaining cache.
2406 : : */
2407 : 0 : if (!scan && !mem_cgroup_online(memcg))
2408 : : scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2409 : :
2410 [ # # # ]: 0 : switch (scan_balance) {
2411 : : case SCAN_EQUAL:
2412 : : /* Scan lists relative to size */
2413 : : break;
2414 : : case SCAN_FRACT:
2415 : : /*
2416 : : * Scan types proportional to swappiness and
2417 : : * their relative recent reclaim efficiency.
2418 : : * Make sure we don't miss the last page on
2419 : : * the offlined memory cgroups because of a
2420 : : * round-off error.
2421 : : */
2422 : 0 : scan = mem_cgroup_online(memcg) ?
2423 : 0 : div64_u64(scan * fraction[file], denominator) :
2424 : : DIV64_U64_ROUND_UP(scan * fraction[file],
2425 : : denominator);
2426 : 0 : break;
2427 : 0 : case SCAN_FILE:
2428 : : case SCAN_ANON:
2429 : : /* Scan one type exclusively */
2430 [ # # ]: 0 : if ((scan_balance == SCAN_FILE) != file) {
2431 : 0 : lruvec_size = 0;
2432 : 0 : scan = 0;
2433 : : }
2434 : : break;
2435 : : default:
2436 : : /* Look ma, no brain */
2437 : 0 : BUG();
2438 : : }
2439 : :
2440 : 0 : nr[lru] = scan;
2441 : : }
2442 : 0 : }
2443 : :
2444 : 0 : static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2445 : : {
2446 : 0 : unsigned long nr[NR_LRU_LISTS];
2447 : 0 : unsigned long targets[NR_LRU_LISTS];
2448 : 0 : unsigned long nr_to_scan;
2449 : 0 : enum lru_list lru;
2450 : 0 : unsigned long nr_reclaimed = 0;
2451 : 0 : unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2452 : 0 : struct blk_plug plug;
2453 : 0 : bool scan_adjusted;
2454 : :
2455 : 0 : get_scan_count(lruvec, sc, nr);
2456 : :
2457 : : /* Record the original scan target for proportional adjustments later */
2458 : 0 : memcpy(targets, nr, sizeof(nr));
2459 : :
2460 : : /*
2461 : : * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2462 : : * event that can occur when there is little memory pressure e.g.
2463 : : * multiple streaming readers/writers. Hence, we do not abort scanning
2464 : : * when the requested number of pages are reclaimed when scanning at
2465 : : * DEF_PRIORITY on the assumption that the fact we are direct
2466 : : * reclaiming implies that kswapd is not keeping up and it is best to
2467 : : * do a batch of work at once. For memcg reclaim one check is made to
2468 : : * abort proportional reclaim if either the file or anon lru has already
2469 : : * dropped to zero at the first pass.
2470 : : */
2471 [ # # ]: 0 : scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2472 [ # # ]: 0 : sc->priority == DEF_PRIORITY);
2473 : :
2474 : 0 : blk_start_plug(&plug);
2475 [ # # # # ]: 0 : while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2476 [ # # ]: 0 : nr[LRU_INACTIVE_FILE]) {
2477 : : unsigned long nr_anon, nr_file, percentage;
2478 : : unsigned long nr_scanned;
2479 : :
2480 [ # # ]: 0 : for_each_evictable_lru(lru) {
2481 [ # # ]: 0 : if (nr[lru]) {
2482 : 0 : nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2483 : 0 : nr[lru] -= nr_to_scan;
2484 : :
2485 : 0 : nr_reclaimed += shrink_list(lru, nr_to_scan,
2486 : : lruvec, sc);
2487 : : }
2488 : : }
2489 : :
2490 : 0 : cond_resched();
2491 : :
2492 [ # # ]: 0 : if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2493 : 0 : continue;
2494 : :
2495 : : /*
2496 : : * For kswapd and memcg, reclaim at least the number of pages
2497 : : * requested. Ensure that the anon and file LRUs are scanned
2498 : : * proportionally what was requested by get_scan_count(). We
2499 : : * stop reclaiming one LRU and reduce the amount scanning
2500 : : * proportional to the original scan target.
2501 : : */
2502 : 0 : nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2503 : 0 : nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2504 : :
2505 : : /*
2506 : : * It's just vindictive to attack the larger once the smaller
2507 : : * has gone to zero. And given the way we stop scanning the
2508 : : * smaller below, this makes sure that we only make one nudge
2509 : : * towards proportionality once we've got nr_to_reclaim.
2510 : : */
2511 [ # # ]: 0 : if (!nr_file || !nr_anon)
2512 : : break;
2513 : :
2514 [ # # ]: 0 : if (nr_file > nr_anon) {
2515 : 0 : unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2516 : 0 : targets[LRU_ACTIVE_ANON] + 1;
2517 : 0 : lru = LRU_BASE;
2518 : 0 : percentage = nr_anon * 100 / scan_target;
2519 : : } else {
2520 : 0 : unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2521 : 0 : targets[LRU_ACTIVE_FILE] + 1;
2522 : 0 : lru = LRU_FILE;
2523 : 0 : percentage = nr_file * 100 / scan_target;
2524 : : }
2525 : :
2526 : : /* Stop scanning the smaller of the LRU */
2527 : 0 : nr[lru] = 0;
2528 : 0 : nr[lru + LRU_ACTIVE] = 0;
2529 : :
2530 : : /*
2531 : : * Recalculate the other LRU scan count based on its original
2532 : : * scan target and the percentage scanning already complete
2533 : : */
2534 [ # # ]: 0 : lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2535 : 0 : nr_scanned = targets[lru] - nr[lru];
2536 : 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2537 : 0 : nr[lru] -= min(nr[lru], nr_scanned);
2538 : :
2539 : 0 : lru += LRU_ACTIVE;
2540 : 0 : nr_scanned = targets[lru] - nr[lru];
2541 : 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2542 : 0 : nr[lru] -= min(nr[lru], nr_scanned);
2543 : :
2544 : 0 : scan_adjusted = true;
2545 : : }
2546 : 0 : blk_finish_plug(&plug);
2547 : 0 : sc->nr_reclaimed += nr_reclaimed;
2548 : :
2549 : : /*
2550 : : * Even if we did not try to evict anon pages at all, we want to
2551 : : * rebalance the anon lru active/inactive ratio.
2552 : : */
2553 [ # # # # ]: 0 : if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2554 : 0 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2555 : : sc, LRU_ACTIVE_ANON);
2556 : 0 : }
2557 : :
2558 : : /* Use reclaim/compaction for costly allocs or under memory pressure */
2559 : 0 : static bool in_reclaim_compaction(struct scan_control *sc)
2560 : : {
2561 [ # # ]: 0 : if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2562 : 0 : (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2563 [ # # ]: 0 : sc->priority < DEF_PRIORITY - 2))
2564 : 0 : return true;
2565 : :
2566 : : return false;
2567 : : }
2568 : :
2569 : : /*
2570 : : * Reclaim/compaction is used for high-order allocation requests. It reclaims
2571 : : * order-0 pages before compacting the zone. should_continue_reclaim() returns
2572 : : * true if more pages should be reclaimed such that when the page allocator
2573 : : * calls try_to_compact_zone() that it will have enough free pages to succeed.
2574 : : * It will give up earlier than that if there is difficulty reclaiming pages.
2575 : : */
2576 : 0 : static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2577 : : unsigned long nr_reclaimed,
2578 : : struct scan_control *sc)
2579 : : {
2580 : 0 : unsigned long pages_for_compaction;
2581 : 0 : unsigned long inactive_lru_pages;
2582 : 0 : int z;
2583 : :
2584 : : /* If not in reclaim/compaction mode, stop */
2585 [ # # ]: 0 : if (!in_reclaim_compaction(sc))
2586 : : return false;
2587 : :
2588 : : /*
2589 : : * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2590 : : * number of pages that were scanned. This will return to the caller
2591 : : * with the risk reclaim/compaction and the resulting allocation attempt
2592 : : * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2593 : : * allocations through requiring that the full LRU list has been scanned
2594 : : * first, by assuming that zero delta of sc->nr_scanned means full LRU
2595 : : * scan, but that approximation was wrong, and there were corner cases
2596 : : * where always a non-zero amount of pages were scanned.
2597 : : */
2598 [ # # ]: 0 : if (!nr_reclaimed)
2599 : : return false;
2600 : :
2601 : : /* If compaction would go ahead or the allocation would succeed, stop */
2602 [ # # ]: 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
2603 : 0 : struct zone *zone = &pgdat->node_zones[z];
2604 [ # # ]: 0 : if (!managed_zone(zone))
2605 : 0 : continue;
2606 : :
2607 [ # # ]: 0 : switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2608 : : case COMPACT_SUCCESS:
2609 : : case COMPACT_CONTINUE:
2610 : : return false;
2611 : 0 : default:
2612 : : /* check next zone */
2613 : 0 : ;
2614 : : }
2615 : : }
2616 : :
2617 : : /*
2618 : : * If we have not reclaimed enough pages for compaction and the
2619 : : * inactive lists are large enough, continue reclaiming
2620 : : */
2621 : 0 : pages_for_compaction = compact_gap(sc->order);
2622 : 0 : inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2623 [ # # ]: 0 : if (get_nr_swap_pages() > 0)
2624 : 0 : inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2625 : :
2626 : 0 : return inactive_lru_pages > pages_for_compaction;
2627 : : }
2628 : :
2629 : 0 : static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2630 : : {
2631 : 0 : struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2632 : 0 : struct mem_cgroup *memcg;
2633 : :
2634 : 0 : memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2635 : 0 : do {
2636 : 0 : struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2637 : 0 : unsigned long reclaimed;
2638 : 0 : unsigned long scanned;
2639 : :
2640 : 0 : switch (mem_cgroup_protected(target_memcg, memcg)) {
2641 : : case MEMCG_PROT_MIN:
2642 : : /*
2643 : : * Hard protection.
2644 : : * If there is no reclaimable memory, OOM.
2645 : : */
2646 : : continue;
2647 : : case MEMCG_PROT_LOW:
2648 : : /*
2649 : : * Soft protection.
2650 : : * Respect the protection only as long as
2651 : : * there is an unprotected supply
2652 : : * of reclaimable memory from other cgroups.
2653 : : */
2654 : : if (!sc->memcg_low_reclaim) {
2655 : : sc->memcg_low_skipped = 1;
2656 : : continue;
2657 : : }
2658 : : memcg_memory_event(memcg, MEMCG_LOW);
2659 : : break;
2660 : : case MEMCG_PROT_NONE:
2661 : : /*
2662 : : * All protection thresholds breached. We may
2663 : : * still choose to vary the scan pressure
2664 : : * applied based on by how much the cgroup in
2665 : : * question has exceeded its protection
2666 : : * thresholds (see get_scan_count).
2667 : : */
2668 : : break;
2669 : : }
2670 : :
2671 : 0 : reclaimed = sc->nr_reclaimed;
2672 : 0 : scanned = sc->nr_scanned;
2673 : :
2674 : 0 : shrink_lruvec(lruvec, sc);
2675 : :
2676 : 0 : shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2677 : 0 : sc->priority);
2678 : :
2679 : : /* Record the group's reclaim efficiency */
2680 : 0 : vmpressure(sc->gfp_mask, memcg, false,
2681 : : sc->nr_scanned - scanned,
2682 : : sc->nr_reclaimed - reclaimed);
2683 : :
2684 : 0 : } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2685 : 0 : }
2686 : :
2687 : 0 : static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2688 : : {
2689 : 0 : struct reclaim_state *reclaim_state = current->reclaim_state;
2690 : 0 : unsigned long nr_reclaimed, nr_scanned;
2691 : 0 : struct lruvec *target_lruvec;
2692 : 0 : bool reclaimable = false;
2693 : 0 : unsigned long file;
2694 : :
2695 : 0 : target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2696 : :
2697 : 0 : again:
2698 : 0 : memset(&sc->nr, 0, sizeof(sc->nr));
2699 : :
2700 : 0 : nr_reclaimed = sc->nr_reclaimed;
2701 : 0 : nr_scanned = sc->nr_scanned;
2702 : :
2703 : : /*
2704 : : * Target desirable inactive:active list ratios for the anon
2705 : : * and file LRU lists.
2706 : : */
2707 [ # # ]: 0 : if (!sc->force_deactivate) {
2708 : 0 : unsigned long refaults;
2709 : :
2710 [ # # ]: 0 : if (inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2711 : 0 : sc->may_deactivate |= DEACTIVATE_ANON;
2712 : : else
2713 : 0 : sc->may_deactivate &= ~DEACTIVATE_ANON;
2714 : :
2715 : : /*
2716 : : * When refaults are being observed, it means a new
2717 : : * workingset is being established. Deactivate to get
2718 : : * rid of any stale active pages quickly.
2719 : : */
2720 : 0 : refaults = lruvec_page_state(target_lruvec,
2721 : : WORKINGSET_ACTIVATE);
2722 [ # # # # ]: 0 : if (refaults != target_lruvec->refaults ||
2723 : 0 : inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2724 : 0 : sc->may_deactivate |= DEACTIVATE_FILE;
2725 : : else
2726 : 0 : sc->may_deactivate &= ~DEACTIVATE_FILE;
2727 : : } else
2728 : 0 : sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2729 : :
2730 : : /*
2731 : : * If we have plenty of inactive file pages that aren't
2732 : : * thrashing, try to reclaim those first before touching
2733 : : * anonymous pages.
2734 : : */
2735 : 0 : file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2736 [ # # # # ]: 0 : if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2737 : 0 : sc->cache_trim_mode = 1;
2738 : : else
2739 : 0 : sc->cache_trim_mode = 0;
2740 : :
2741 : : /*
2742 : : * Prevent the reclaimer from falling into the cache trap: as
2743 : : * cache pages start out inactive, every cache fault will tip
2744 : : * the scan balance towards the file LRU. And as the file LRU
2745 : : * shrinks, so does the window for rotation from references.
2746 : : * This means we have a runaway feedback loop where a tiny
2747 : : * thrashing file LRU becomes infinitely more attractive than
2748 : : * anon pages. Try to detect this based on file LRU size.
2749 : : */
2750 : 0 : if (!cgroup_reclaim(sc)) {
2751 : 0 : unsigned long total_high_wmark = 0;
2752 : 0 : unsigned long free, anon;
2753 : 0 : int z;
2754 : :
2755 : 0 : free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2756 : 0 : file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2757 : 0 : node_page_state(pgdat, NR_INACTIVE_FILE);
2758 : :
2759 [ # # ]: 0 : for (z = 0; z < MAX_NR_ZONES; z++) {
2760 : 0 : struct zone *zone = &pgdat->node_zones[z];
2761 [ # # ]: 0 : if (!managed_zone(zone))
2762 : 0 : continue;
2763 : :
2764 : 0 : total_high_wmark += high_wmark_pages(zone);
2765 : : }
2766 : :
2767 : : /*
2768 : : * Consider anon: if that's low too, this isn't a
2769 : : * runaway file reclaim problem, but rather just
2770 : : * extreme pressure. Reclaim as per usual then.
2771 : : */
2772 : 0 : anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2773 : :
2774 : 0 : sc->file_is_tiny =
2775 : 0 : file + free <= total_high_wmark &&
2776 [ # # # # ]: 0 : !(sc->may_deactivate & DEACTIVATE_ANON) &&
2777 [ # # ]: 0 : anon >> sc->priority;
2778 : : }
2779 : :
2780 : 0 : shrink_node_memcgs(pgdat, sc);
2781 : :
2782 [ # # ]: 0 : if (reclaim_state) {
2783 : 0 : sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2784 : 0 : reclaim_state->reclaimed_slab = 0;
2785 : : }
2786 : :
2787 : : /* Record the subtree's reclaim efficiency */
2788 : 0 : vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2789 : : sc->nr_scanned - nr_scanned,
2790 [ # # ]: 0 : sc->nr_reclaimed - nr_reclaimed);
2791 : :
2792 [ # # ]: 0 : if (sc->nr_reclaimed - nr_reclaimed)
2793 : 0 : reclaimable = true;
2794 : :
2795 [ # # ]: 0 : if (current_is_kswapd()) {
2796 : : /*
2797 : : * If reclaim is isolating dirty pages under writeback,
2798 : : * it implies that the long-lived page allocation rate
2799 : : * is exceeding the page laundering rate. Either the
2800 : : * global limits are not being effective at throttling
2801 : : * processes due to the page distribution throughout
2802 : : * zones or there is heavy usage of a slow backing
2803 : : * device. The only option is to throttle from reclaim
2804 : : * context which is not ideal as there is no guarantee
2805 : : * the dirtying process is throttled in the same way
2806 : : * balance_dirty_pages() manages.
2807 : : *
2808 : : * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2809 : : * count the number of pages under pages flagged for
2810 : : * immediate reclaim and stall if any are encountered
2811 : : * in the nr_immediate check below.
2812 : : */
2813 [ # # # # ]: 0 : if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2814 : 0 : set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2815 : :
2816 : : /* Allow kswapd to start writing pages during reclaim.*/
2817 [ # # ]: 0 : if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2818 : 0 : set_bit(PGDAT_DIRTY, &pgdat->flags);
2819 : :
2820 : : /*
2821 : : * If kswapd scans pages marked marked for immediate
2822 : : * reclaim and under writeback (nr_immediate), it
2823 : : * implies that pages are cycling through the LRU
2824 : : * faster than they are written so also forcibly stall.
2825 : : */
2826 [ # # ]: 0 : if (sc->nr.immediate)
2827 : 0 : congestion_wait(BLK_RW_ASYNC, HZ/10);
2828 : : }
2829 : :
2830 : : /*
2831 : : * Tag a node/memcg as congested if all the dirty pages
2832 : : * scanned were backed by a congested BDI and
2833 : : * wait_iff_congested will stall.
2834 : : *
2835 : : * Legacy memcg will stall in page writeback so avoid forcibly
2836 : : * stalling in wait_iff_congested().
2837 : : */
2838 [ # # ]: 0 : if ((current_is_kswapd() ||
2839 : 0 : (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2840 [ # # # # ]: 0 : sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2841 : 0 : set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2842 : :
2843 : : /*
2844 : : * Stall direct reclaim for IO completions if underlying BDIs
2845 : : * and node is congested. Allow kswapd to continue until it
2846 : : * starts encountering unqueued dirty pages or cycling through
2847 : : * the LRU too quickly.
2848 : : */
2849 [ # # # # ]: 0 : if (!current_is_kswapd() && current_may_throttle() &&
2850 [ # # # # ]: 0 : !sc->hibernation_mode &&
2851 : 0 : test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2852 : 0 : wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2853 : :
2854 [ # # ]: 0 : if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2855 : : sc))
2856 : 0 : goto again;
2857 : :
2858 : : /*
2859 : : * Kswapd gives up on balancing particular nodes after too
2860 : : * many failures to reclaim anything from them and goes to
2861 : : * sleep. On reclaim progress, reset the failure counter. A
2862 : : * successful direct reclaim run will revive a dormant kswapd.
2863 : : */
2864 [ # # ]: 0 : if (reclaimable)
2865 : 0 : pgdat->kswapd_failures = 0;
2866 : 0 : }
2867 : :
2868 : : /*
2869 : : * Returns true if compaction should go ahead for a costly-order request, or
2870 : : * the allocation would already succeed without compaction. Return false if we
2871 : : * should reclaim first.
2872 : : */
2873 : : static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2874 : : {
2875 : : unsigned long watermark;
2876 : : enum compact_result suitable;
2877 : :
2878 : : suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2879 : : if (suitable == COMPACT_SUCCESS)
2880 : : /* Allocation should succeed already. Don't reclaim. */
2881 : : return true;
2882 : : if (suitable == COMPACT_SKIPPED)
2883 : : /* Compaction cannot yet proceed. Do reclaim. */
2884 : : return false;
2885 : :
2886 : : /*
2887 : : * Compaction is already possible, but it takes time to run and there
2888 : : * are potentially other callers using the pages just freed. So proceed
2889 : : * with reclaim to make a buffer of free pages available to give
2890 : : * compaction a reasonable chance of completing and allocating the page.
2891 : : * Note that we won't actually reclaim the whole buffer in one attempt
2892 : : * as the target watermark in should_continue_reclaim() is lower. But if
2893 : : * we are already above the high+gap watermark, don't reclaim at all.
2894 : : */
2895 : : watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2896 : :
2897 : : return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2898 : : }
2899 : :
2900 : : /*
2901 : : * This is the direct reclaim path, for page-allocating processes. We only
2902 : : * try to reclaim pages from zones which will satisfy the caller's allocation
2903 : : * request.
2904 : : *
2905 : : * If a zone is deemed to be full of pinned pages then just give it a light
2906 : : * scan then give up on it.
2907 : : */
2908 : 0 : static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2909 : : {
2910 : 0 : struct zoneref *z;
2911 : 0 : struct zone *zone;
2912 : 0 : unsigned long nr_soft_reclaimed;
2913 : 0 : unsigned long nr_soft_scanned;
2914 : 0 : gfp_t orig_mask;
2915 : 0 : pg_data_t *last_pgdat = NULL;
2916 : :
2917 : : /*
2918 : : * If the number of buffer_heads in the machine exceeds the maximum
2919 : : * allowed level, force direct reclaim to scan the highmem zone as
2920 : : * highmem pages could be pinning lowmem pages storing buffer_heads
2921 : : */
2922 : 0 : orig_mask = sc->gfp_mask;
2923 [ # # ]: 0 : if (buffer_heads_over_limit) {
2924 : 0 : sc->gfp_mask |= __GFP_HIGHMEM;
2925 : 0 : sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2926 : : }
2927 : :
2928 [ # # # # : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
# # ]
2929 : : sc->reclaim_idx, sc->nodemask) {
2930 : : /*
2931 : : * Take care memory controller reclaiming has small influence
2932 : : * to global LRU.
2933 : : */
2934 : 0 : if (!cgroup_reclaim(sc)) {
2935 [ # # ]: 0 : if (!cpuset_zone_allowed(zone,
2936 : : GFP_KERNEL | __GFP_HARDWALL))
2937 : 0 : continue;
2938 : :
2939 : : /*
2940 : : * If we already have plenty of memory free for
2941 : : * compaction in this zone, don't free any more.
2942 : : * Even though compaction is invoked for any
2943 : : * non-zero order, only frequent costly order
2944 : : * reclamation is disruptive enough to become a
2945 : : * noticeable problem, like transparent huge
2946 : : * page allocations.
2947 : : */
2948 : 0 : if (IS_ENABLED(CONFIG_COMPACTION) &&
2949 [ # # # # ]: 0 : sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2950 : 0 : compaction_ready(zone, sc)) {
2951 : 0 : sc->compaction_ready = true;
2952 : 0 : continue;
2953 : : }
2954 : :
2955 : : /*
2956 : : * Shrink each node in the zonelist once. If the
2957 : : * zonelist is ordered by zone (not the default) then a
2958 : : * node may be shrunk multiple times but in that case
2959 : : * the user prefers lower zones being preserved.
2960 : : */
2961 [ # # ]: 0 : if (zone->zone_pgdat == last_pgdat)
2962 : 0 : continue;
2963 : :
2964 : : /*
2965 : : * This steals pages from memory cgroups over softlimit
2966 : : * and returns the number of reclaimed pages and
2967 : : * scanned pages. This works for global memory pressure
2968 : : * and balancing, not for a memcg's limit.
2969 : : */
2970 : 0 : nr_soft_scanned = 0;
2971 : 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2972 : : sc->order, sc->gfp_mask,
2973 : : &nr_soft_scanned);
2974 : 0 : sc->nr_reclaimed += nr_soft_reclaimed;
2975 : 0 : sc->nr_scanned += nr_soft_scanned;
2976 : : /* need some check for avoid more shrink_zone() */
2977 : : }
2978 : :
2979 : : /* See comment about same check for global reclaim above */
2980 : 0 : if (zone->zone_pgdat == last_pgdat)
2981 : : continue;
2982 : 0 : last_pgdat = zone->zone_pgdat;
2983 : 0 : shrink_node(zone->zone_pgdat, sc);
2984 : : }
2985 : :
2986 : : /*
2987 : : * Restore to original mask to avoid the impact on the caller if we
2988 : : * promoted it to __GFP_HIGHMEM.
2989 : : */
2990 : 0 : sc->gfp_mask = orig_mask;
2991 : 0 : }
2992 : :
2993 : 0 : static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2994 : : {
2995 : 0 : struct lruvec *target_lruvec;
2996 : 0 : unsigned long refaults;
2997 : :
2998 : 0 : target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2999 : 0 : refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE);
3000 : 0 : target_lruvec->refaults = refaults;
3001 : : }
3002 : :
3003 : : /*
3004 : : * This is the main entry point to direct page reclaim.
3005 : : *
3006 : : * If a full scan of the inactive list fails to free enough memory then we
3007 : : * are "out of memory" and something needs to be killed.
3008 : : *
3009 : : * If the caller is !__GFP_FS then the probability of a failure is reasonably
3010 : : * high - the zone may be full of dirty or under-writeback pages, which this
3011 : : * caller can't do much about. We kick the writeback threads and take explicit
3012 : : * naps in the hope that some of these pages can be written. But if the
3013 : : * allocating task holds filesystem locks which prevent writeout this might not
3014 : : * work, and the allocation attempt will fail.
3015 : : *
3016 : : * returns: 0, if no pages reclaimed
3017 : : * else, the number of pages reclaimed
3018 : : */
3019 : 0 : static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3020 : : struct scan_control *sc)
3021 : : {
3022 : 0 : int initial_priority = sc->priority;
3023 : 0 : pg_data_t *last_pgdat;
3024 : 0 : struct zoneref *z;
3025 : 0 : struct zone *zone;
3026 : : retry:
3027 [ # # ]: 0 : delayacct_freepages_start();
3028 : :
3029 : 0 : if (!cgroup_reclaim(sc))
3030 : 0 : __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3031 : :
3032 : 0 : do {
3033 : 0 : vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3034 : : sc->priority);
3035 : 0 : sc->nr_scanned = 0;
3036 : 0 : shrink_zones(zonelist, sc);
3037 : :
3038 [ # # ]: 0 : if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3039 : : break;
3040 : :
3041 [ # # ]: 0 : if (sc->compaction_ready)
3042 : : break;
3043 : :
3044 : : /*
3045 : : * If we're getting trouble reclaiming, start doing
3046 : : * writepage even in laptop mode.
3047 : : */
3048 [ # # ]: 0 : if (sc->priority < DEF_PRIORITY - 2)
3049 : 0 : sc->may_writepage = 1;
3050 [ # # ]: 0 : } while (--sc->priority >= 0);
3051 : :
3052 : 0 : last_pgdat = NULL;
3053 [ # # # # : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
# # ]
3054 : : sc->nodemask) {
3055 [ # # ]: 0 : if (zone->zone_pgdat == last_pgdat)
3056 : 0 : continue;
3057 : 0 : last_pgdat = zone->zone_pgdat;
3058 : :
3059 : 0 : snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3060 : :
3061 : 0 : if (cgroup_reclaim(sc)) {
3062 : : struct lruvec *lruvec;
3063 : :
3064 : : lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3065 : : zone->zone_pgdat);
3066 : : clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3067 : : }
3068 : : }
3069 : :
3070 [ # # ]: 0 : delayacct_freepages_end();
3071 : :
3072 [ # # ]: 0 : if (sc->nr_reclaimed)
3073 : 0 : return sc->nr_reclaimed;
3074 : :
3075 : : /* Aborted reclaim to try compaction? don't OOM, then */
3076 [ # # ]: 0 : if (sc->compaction_ready)
3077 : : return 1;
3078 : :
3079 : : /*
3080 : : * We make inactive:active ratio decisions based on the node's
3081 : : * composition of memory, but a restrictive reclaim_idx or a
3082 : : * memory.low cgroup setting can exempt large amounts of
3083 : : * memory from reclaim. Neither of which are very common, so
3084 : : * instead of doing costly eligibility calculations of the
3085 : : * entire cgroup subtree up front, we assume the estimates are
3086 : : * good, and retry with forcible deactivation if that fails.
3087 : : */
3088 [ # # ]: 0 : if (sc->skipped_deactivate) {
3089 : 0 : sc->priority = initial_priority;
3090 : 0 : sc->force_deactivate = 1;
3091 : 0 : sc->skipped_deactivate = 0;
3092 : 0 : goto retry;
3093 : : }
3094 : :
3095 : : /* Untapped cgroup reserves? Don't OOM, retry. */
3096 [ # # ]: 0 : if (sc->memcg_low_skipped) {
3097 : 0 : sc->priority = initial_priority;
3098 : 0 : sc->force_deactivate = 0;
3099 : 0 : sc->skipped_deactivate = 0;
3100 : 0 : sc->memcg_low_reclaim = 1;
3101 : 0 : sc->memcg_low_skipped = 0;
3102 : 0 : goto retry;
3103 : : }
3104 : :
3105 : : return 0;
3106 : : }
3107 : :
3108 : 0 : static bool allow_direct_reclaim(pg_data_t *pgdat)
3109 : : {
3110 : 0 : struct zone *zone;
3111 : 0 : unsigned long pfmemalloc_reserve = 0;
3112 : 0 : unsigned long free_pages = 0;
3113 : 0 : int i;
3114 : 0 : bool wmark_ok;
3115 : :
3116 [ # # ]: 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3117 : : return true;
3118 : :
3119 [ # # ]: 0 : for (i = 0; i <= ZONE_NORMAL; i++) {
3120 : 0 : zone = &pgdat->node_zones[i];
3121 [ # # ]: 0 : if (!managed_zone(zone))
3122 : 0 : continue;
3123 : :
3124 [ # # ]: 0 : if (!zone_reclaimable_pages(zone))
3125 : 0 : continue;
3126 : :
3127 : 0 : pfmemalloc_reserve += min_wmark_pages(zone);
3128 : 0 : free_pages += zone_page_state(zone, NR_FREE_PAGES);
3129 : : }
3130 : :
3131 : : /* If there are no reserves (unexpected config) then do not throttle */
3132 [ # # ]: 0 : if (!pfmemalloc_reserve)
3133 : : return true;
3134 : :
3135 : 0 : wmark_ok = free_pages > pfmemalloc_reserve / 2;
3136 : :
3137 : : /* kswapd must be awake if processes are being throttled */
3138 [ # # # # ]: 0 : if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3139 : 0 : pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3140 : : (enum zone_type)ZONE_NORMAL);
3141 : 0 : wake_up_interruptible(&pgdat->kswapd_wait);
3142 : : }
3143 : :
3144 : : return wmark_ok;
3145 : : }
3146 : :
3147 : : /*
3148 : : * Throttle direct reclaimers if backing storage is backed by the network
3149 : : * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3150 : : * depleted. kswapd will continue to make progress and wake the processes
3151 : : * when the low watermark is reached.
3152 : : *
3153 : : * Returns true if a fatal signal was delivered during throttling. If this
3154 : : * happens, the page allocator should not consider triggering the OOM killer.
3155 : : */
3156 : 0 : static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3157 : : nodemask_t *nodemask)
3158 : : {
3159 : 0 : struct zoneref *z;
3160 : 0 : struct zone *zone;
3161 : 0 : pg_data_t *pgdat = NULL;
3162 : :
3163 : : /*
3164 : : * Kernel threads should not be throttled as they may be indirectly
3165 : : * responsible for cleaning pages necessary for reclaim to make forward
3166 : : * progress. kjournald for example may enter direct reclaim while
3167 : : * committing a transaction where throttling it could forcing other
3168 : : * processes to block on log_wait_commit().
3169 : : */
3170 [ # # ]: 0 : if (current->flags & PF_KTHREAD)
3171 : 0 : goto out;
3172 : :
3173 : : /*
3174 : : * If a fatal signal is pending, this process should not throttle.
3175 : : * It should return quickly so it can exit and free its memory
3176 : : */
3177 [ # # ]: 0 : if (fatal_signal_pending(current))
3178 : 0 : goto out;
3179 : :
3180 : : /*
3181 : : * Check if the pfmemalloc reserves are ok by finding the first node
3182 : : * with a usable ZONE_NORMAL or lower zone. The expectation is that
3183 : : * GFP_KERNEL will be required for allocating network buffers when
3184 : : * swapping over the network so ZONE_HIGHMEM is unusable.
3185 : : *
3186 : : * Throttling is based on the first usable node and throttled processes
3187 : : * wait on a queue until kswapd makes progress and wakes them. There
3188 : : * is an affinity then between processes waking up and where reclaim
3189 : : * progress has been made assuming the process wakes on the same node.
3190 : : * More importantly, processes running on remote nodes will not compete
3191 : : * for remote pfmemalloc reserves and processes on different nodes
3192 : : * should make reasonable progress.
3193 : : */
3194 [ # # # # ]: 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
3195 : : gfp_zone(gfp_mask), nodemask) {
3196 [ # # ]: 0 : if (zone_idx(zone) > ZONE_NORMAL)
3197 [ # # ]: 0 : continue;
3198 : :
3199 : : /* Throttle based on the first usable node */
3200 : 0 : pgdat = zone->zone_pgdat;
3201 [ # # ]: 0 : if (allow_direct_reclaim(pgdat))
3202 : 0 : goto out;
3203 : : break;
3204 : : }
3205 : :
3206 : : /* If no zone was usable by the allocation flags then do not throttle */
3207 [ # # ]: 0 : if (!pgdat)
3208 : 0 : goto out;
3209 : :
3210 : : /* Account for the throttling */
3211 : 0 : count_vm_event(PGSCAN_DIRECT_THROTTLE);
3212 : :
3213 : : /*
3214 : : * If the caller cannot enter the filesystem, it's possible that it
3215 : : * is due to the caller holding an FS lock or performing a journal
3216 : : * transaction in the case of a filesystem like ext[3|4]. In this case,
3217 : : * it is not safe to block on pfmemalloc_wait as kswapd could be
3218 : : * blocked waiting on the same lock. Instead, throttle for up to a
3219 : : * second before continuing.
3220 : : */
3221 [ # # ]: 0 : if (!(gfp_mask & __GFP_FS)) {
3222 [ # # # # : 0 : wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
# # # # ]
3223 : : allow_direct_reclaim(pgdat), HZ);
3224 : :
3225 : 0 : goto check_pending;
3226 : : }
3227 : :
3228 : : /* Throttle until kswapd wakes the process */
3229 [ # # # # : 0 : wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
# # ]
3230 : : allow_direct_reclaim(pgdat));
3231 : :
3232 : 0 : check_pending:
3233 [ # # ]: 0 : if (fatal_signal_pending(current))
3234 : 0 : return true;
3235 : :
3236 : 0 : out:
3237 : : return false;
3238 : : }
3239 : :
3240 : 0 : unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3241 : : gfp_t gfp_mask, nodemask_t *nodemask)
3242 : : {
3243 : 0 : unsigned long nr_reclaimed;
3244 : 0 : struct scan_control sc = {
3245 : : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3246 : : .gfp_mask = current_gfp_context(gfp_mask),
3247 : : .reclaim_idx = gfp_zone(gfp_mask),
3248 : : .order = order,
3249 : : .nodemask = nodemask,
3250 : : .priority = DEF_PRIORITY,
3251 [ # # ]: 0 : .may_writepage = !laptop_mode,
3252 : : .may_unmap = 1,
3253 : : .may_swap = 1,
3254 : : };
3255 : :
3256 : : /*
3257 : : * scan_control uses s8 fields for order, priority, and reclaim_idx.
3258 : : * Confirm they are large enough for max values.
3259 : : */
3260 : 0 : BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3261 : 0 : BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3262 : 0 : BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3263 : :
3264 : : /*
3265 : : * Do not enter reclaim if fatal signal was delivered while throttled.
3266 : : * 1 is returned so that the page allocator does not OOM kill at this
3267 : : * point.
3268 : : */
3269 [ # # ]: 0 : if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3270 : : return 1;
3271 : :
3272 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3273 : 0 : trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3274 : :
3275 : 0 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3276 : :
3277 : 0 : trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3278 : 0 : set_task_reclaim_state(current, NULL);
3279 : :
3280 : 0 : return nr_reclaimed;
3281 : : }
3282 : :
3283 : : #ifdef CONFIG_MEMCG
3284 : :
3285 : : /* Only used by soft limit reclaim. Do not reuse for anything else. */
3286 : : unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3287 : : gfp_t gfp_mask, bool noswap,
3288 : : pg_data_t *pgdat,
3289 : : unsigned long *nr_scanned)
3290 : : {
3291 : : struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3292 : : struct scan_control sc = {
3293 : : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3294 : : .target_mem_cgroup = memcg,
3295 : : .may_writepage = !laptop_mode,
3296 : : .may_unmap = 1,
3297 : : .reclaim_idx = MAX_NR_ZONES - 1,
3298 : : .may_swap = !noswap,
3299 : : };
3300 : :
3301 : : WARN_ON_ONCE(!current->reclaim_state);
3302 : :
3303 : : sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3304 : : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3305 : :
3306 : : trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3307 : : sc.gfp_mask);
3308 : :
3309 : : /*
3310 : : * NOTE: Although we can get the priority field, using it
3311 : : * here is not a good idea, since it limits the pages we can scan.
3312 : : * if we don't reclaim here, the shrink_node from balance_pgdat
3313 : : * will pick up pages from other mem cgroup's as well. We hack
3314 : : * the priority and make it zero.
3315 : : */
3316 : : shrink_lruvec(lruvec, &sc);
3317 : :
3318 : : trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3319 : :
3320 : : *nr_scanned = sc.nr_scanned;
3321 : :
3322 : : return sc.nr_reclaimed;
3323 : : }
3324 : :
3325 : : unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3326 : : unsigned long nr_pages,
3327 : : gfp_t gfp_mask,
3328 : : bool may_swap)
3329 : : {
3330 : : unsigned long nr_reclaimed;
3331 : : unsigned long pflags;
3332 : : unsigned int noreclaim_flag;
3333 : : struct scan_control sc = {
3334 : : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3335 : : .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3336 : : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3337 : : .reclaim_idx = MAX_NR_ZONES - 1,
3338 : : .target_mem_cgroup = memcg,
3339 : : .priority = DEF_PRIORITY,
3340 : : .may_writepage = !laptop_mode,
3341 : : .may_unmap = 1,
3342 : : .may_swap = may_swap,
3343 : : };
3344 : : /*
3345 : : * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3346 : : * equal pressure on all the nodes. This is based on the assumption that
3347 : : * the reclaim does not bail out early.
3348 : : */
3349 : : struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3350 : :
3351 : : set_task_reclaim_state(current, &sc.reclaim_state);
3352 : :
3353 : : trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3354 : :
3355 : : psi_memstall_enter(&pflags);
3356 : : noreclaim_flag = memalloc_noreclaim_save();
3357 : :
3358 : : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3359 : :
3360 : : memalloc_noreclaim_restore(noreclaim_flag);
3361 : : psi_memstall_leave(&pflags);
3362 : :
3363 : : trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3364 : : set_task_reclaim_state(current, NULL);
3365 : :
3366 : : return nr_reclaimed;
3367 : : }
3368 : : #endif
3369 : :
3370 : 0 : static void age_active_anon(struct pglist_data *pgdat,
3371 : : struct scan_control *sc)
3372 : : {
3373 : 0 : struct mem_cgroup *memcg;
3374 : 0 : struct lruvec *lruvec;
3375 : :
3376 [ # # ]: 0 : if (!total_swap_pages)
3377 : : return;
3378 : :
3379 : 0 : lruvec = mem_cgroup_lruvec(NULL, pgdat);
3380 [ # # ]: 0 : if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3381 : : return;
3382 : :
3383 : 0 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
3384 : 0 : do {
3385 : 0 : lruvec = mem_cgroup_lruvec(memcg, pgdat);
3386 : 0 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3387 : : sc, LRU_ACTIVE_ANON);
3388 : 0 : memcg = mem_cgroup_iter(NULL, memcg, NULL);
3389 : 0 : } while (memcg);
3390 : : }
3391 : :
3392 : 0 : static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3393 : : {
3394 : 0 : int i;
3395 : 0 : struct zone *zone;
3396 : :
3397 : : /*
3398 : : * Check for watermark boosts top-down as the higher zones
3399 : : * are more likely to be boosted. Both watermarks and boosts
3400 : : * should not be checked at the time time as reclaim would
3401 : : * start prematurely when there is no boosting and a lower
3402 : : * zone is balanced.
3403 : : */
3404 [ # # ]: 0 : for (i = classzone_idx; i >= 0; i--) {
3405 : 0 : zone = pgdat->node_zones + i;
3406 [ # # ]: 0 : if (!managed_zone(zone))
3407 : 0 : continue;
3408 : :
3409 [ # # ]: 0 : if (zone->watermark_boost)
3410 : : return true;
3411 : : }
3412 : :
3413 : : return false;
3414 : : }
3415 : :
3416 : : /*
3417 : : * Returns true if there is an eligible zone balanced for the request order
3418 : : * and classzone_idx
3419 : : */
3420 : 56 : static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3421 : : {
3422 : 56 : int i;
3423 : 56 : unsigned long mark = -1;
3424 : 56 : struct zone *zone;
3425 : :
3426 : : /*
3427 : : * Check watermarks bottom-up as lower zones are more likely to
3428 : : * meet watermarks.
3429 : : */
3430 [ + - ]: 56 : for (i = 0; i <= classzone_idx; i++) {
3431 : 56 : zone = pgdat->node_zones + i;
3432 : :
3433 [ - + ]: 56 : if (!managed_zone(zone))
3434 : 0 : continue;
3435 : :
3436 : 56 : mark = high_wmark_pages(zone);
3437 [ - + ]: 56 : if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3438 : : return true;
3439 : : }
3440 : :
3441 : : /*
3442 : : * If a node has no populated zone within classzone_idx, it does not
3443 : : * need balancing by definition. This can happen if a zone-restricted
3444 : : * allocation tries to wake a remote kswapd.
3445 : : */
3446 [ # # ]: 0 : if (mark == -1)
3447 : 0 : return true;
3448 : :
3449 : : return false;
3450 : : }
3451 : :
3452 : : /* Clear pgdat state for congested, dirty or under writeback. */
3453 : 56 : static void clear_pgdat_congested(pg_data_t *pgdat)
3454 : : {
3455 : 56 : struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3456 : :
3457 : 56 : clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3458 : 56 : clear_bit(PGDAT_DIRTY, &pgdat->flags);
3459 : 56 : clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3460 : 56 : }
3461 : :
3462 : : /*
3463 : : * Prepare kswapd for sleeping. This verifies that there are no processes
3464 : : * waiting in throttle_direct_reclaim() and that watermarks have been met.
3465 : : *
3466 : : * Returns true if kswapd is ready to sleep
3467 : : */
3468 : 56 : static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3469 : : {
3470 : : /*
3471 : : * The throttled processes are normally woken up in balance_pgdat() as
3472 : : * soon as allow_direct_reclaim() is true. But there is a potential
3473 : : * race between when kswapd checks the watermarks and a process gets
3474 : : * throttled. There is also a potential race if processes get
3475 : : * throttled, kswapd wakes, a large process exits thereby balancing the
3476 : : * zones, which causes kswapd to exit balance_pgdat() before reaching
3477 : : * the wake up checks. If kswapd is going to sleep, no process should
3478 : : * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3479 : : * the wake up is premature, processes will wake kswapd and get
3480 : : * throttled again. The difference from wake ups in balance_pgdat() is
3481 : : * that here we are under prepare_to_wait().
3482 : : */
3483 [ - + ]: 56 : if (waitqueue_active(&pgdat->pfmemalloc_wait))
3484 : 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3485 : :
3486 : : /* Hopeless node, leave it to direct reclaim */
3487 [ + - ]: 56 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3488 : : return true;
3489 : :
3490 [ + - ]: 56 : if (pgdat_balanced(pgdat, order, classzone_idx)) {
3491 : 56 : clear_pgdat_congested(pgdat);
3492 : 56 : return true;
3493 : : }
3494 : :
3495 : : return false;
3496 : : }
3497 : :
3498 : : /*
3499 : : * kswapd shrinks a node of pages that are at or below the highest usable
3500 : : * zone that is currently unbalanced.
3501 : : *
3502 : : * Returns true if kswapd scanned at least the requested number of pages to
3503 : : * reclaim or if the lack of progress was due to pages under writeback.
3504 : : * This is used to determine if the scanning priority needs to be raised.
3505 : : */
3506 : 0 : static bool kswapd_shrink_node(pg_data_t *pgdat,
3507 : : struct scan_control *sc)
3508 : : {
3509 : 0 : struct zone *zone;
3510 : 0 : int z;
3511 : :
3512 : : /* Reclaim a number of pages proportional to the number of zones */
3513 : 0 : sc->nr_to_reclaim = 0;
3514 [ # # ]: 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
3515 : 0 : zone = pgdat->node_zones + z;
3516 [ # # ]: 0 : if (!managed_zone(zone))
3517 : 0 : continue;
3518 : :
3519 : 0 : sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3520 : : }
3521 : :
3522 : : /*
3523 : : * Historically care was taken to put equal pressure on all zones but
3524 : : * now pressure is applied based on node LRU order.
3525 : : */
3526 : 0 : shrink_node(pgdat, sc);
3527 : :
3528 : : /*
3529 : : * Fragmentation may mean that the system cannot be rebalanced for
3530 : : * high-order allocations. If twice the allocation size has been
3531 : : * reclaimed then recheck watermarks only at order-0 to prevent
3532 : : * excessive reclaim. Assume that a process requested a high-order
3533 : : * can direct reclaim/compact.
3534 : : */
3535 [ # # # # ]: 0 : if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3536 : 0 : sc->order = 0;
3537 : :
3538 : 0 : return sc->nr_scanned >= sc->nr_to_reclaim;
3539 : : }
3540 : :
3541 : : /*
3542 : : * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3543 : : * that are eligible for use by the caller until at least one zone is
3544 : : * balanced.
3545 : : *
3546 : : * Returns the order kswapd finished reclaiming at.
3547 : : *
3548 : : * kswapd scans the zones in the highmem->normal->dma direction. It skips
3549 : : * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3550 : : * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3551 : : * or lower is eligible for reclaim until at least one usable zone is
3552 : : * balanced.
3553 : : */
3554 : 0 : static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3555 : : {
3556 : 0 : int i;
3557 : 0 : unsigned long nr_soft_reclaimed;
3558 : 0 : unsigned long nr_soft_scanned;
3559 : 0 : unsigned long pflags;
3560 : 0 : unsigned long nr_boost_reclaim;
3561 : 0 : unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3562 : 0 : bool boosted;
3563 : 0 : struct zone *zone;
3564 : 0 : struct scan_control sc = {
3565 : : .gfp_mask = GFP_KERNEL,
3566 : : .order = order,
3567 : : .may_unmap = 1,
3568 : : };
3569 : :
3570 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3571 : 0 : psi_memstall_enter(&pflags);
3572 : 0 : __fs_reclaim_acquire();
3573 : :
3574 : 0 : count_vm_event(PAGEOUTRUN);
3575 : :
3576 : : /*
3577 : : * Account for the reclaim boost. Note that the zone boost is left in
3578 : : * place so that parallel allocations that are near the watermark will
3579 : : * stall or direct reclaim until kswapd is finished.
3580 : : */
3581 : 0 : nr_boost_reclaim = 0;
3582 [ # # ]: 0 : for (i = 0; i <= classzone_idx; i++) {
3583 : 0 : zone = pgdat->node_zones + i;
3584 [ # # ]: 0 : if (!managed_zone(zone))
3585 : 0 : continue;
3586 : :
3587 : 0 : nr_boost_reclaim += zone->watermark_boost;
3588 : 0 : zone_boosts[i] = zone->watermark_boost;
3589 : : }
3590 : : boosted = nr_boost_reclaim;
3591 : :
3592 : 0 : restart:
3593 : 0 : sc.priority = DEF_PRIORITY;
3594 : 0 : do {
3595 : 0 : unsigned long nr_reclaimed = sc.nr_reclaimed;
3596 : 0 : bool raise_priority = true;
3597 : 0 : bool balanced;
3598 : 0 : bool ret;
3599 : :
3600 : 0 : sc.reclaim_idx = classzone_idx;
3601 : :
3602 : : /*
3603 : : * If the number of buffer_heads exceeds the maximum allowed
3604 : : * then consider reclaiming from all zones. This has a dual
3605 : : * purpose -- on 64-bit systems it is expected that
3606 : : * buffer_heads are stripped during active rotation. On 32-bit
3607 : : * systems, highmem pages can pin lowmem memory and shrinking
3608 : : * buffers can relieve lowmem pressure. Reclaim may still not
3609 : : * go ahead if all eligible zones for the original allocation
3610 : : * request are balanced to avoid excessive reclaim from kswapd.
3611 : : */
3612 [ # # ]: 0 : if (buffer_heads_over_limit) {
3613 [ # # ]: 0 : for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3614 : 0 : zone = pgdat->node_zones + i;
3615 [ # # ]: 0 : if (!managed_zone(zone))
3616 : 0 : continue;
3617 : :
3618 : 0 : sc.reclaim_idx = i;
3619 : 0 : break;
3620 : : }
3621 : : }
3622 : :
3623 : : /*
3624 : : * If the pgdat is imbalanced then ignore boosting and preserve
3625 : : * the watermarks for a later time and restart. Note that the
3626 : : * zone watermarks will be still reset at the end of balancing
3627 : : * on the grounds that the normal reclaim should be enough to
3628 : : * re-evaluate if boosting is required when kswapd next wakes.
3629 : : */
3630 : 0 : balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3631 [ # # ]: 0 : if (!balanced && nr_boost_reclaim) {
3632 : 0 : nr_boost_reclaim = 0;
3633 : 0 : goto restart;
3634 : : }
3635 : :
3636 : : /*
3637 : : * If boosting is not active then only reclaim if there are no
3638 : : * eligible zones. Note that sc.reclaim_idx is not used as
3639 : : * buffer_heads_over_limit may have adjusted it.
3640 : : */
3641 [ # # ]: 0 : if (!nr_boost_reclaim && balanced)
3642 : 0 : goto out;
3643 : :
3644 : : /* Limit the priority of boosting to avoid reclaim writeback */
3645 [ # # # # ]: 0 : if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3646 : 0 : raise_priority = false;
3647 : :
3648 : : /*
3649 : : * Do not writeback or swap pages for boosted reclaim. The
3650 : : * intent is to relieve pressure not issue sub-optimal IO
3651 : : * from reclaim context. If no pages are reclaimed, the
3652 : : * reclaim will be aborted.
3653 : : */
3654 [ # # # # ]: 0 : sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3655 : 0 : sc.may_swap = !nr_boost_reclaim;
3656 : :
3657 : : /*
3658 : : * Do some background aging of the anon list, to give
3659 : : * pages a chance to be referenced before reclaiming. All
3660 : : * pages are rotated regardless of classzone as this is
3661 : : * about consistent aging.
3662 : : */
3663 : 0 : age_active_anon(pgdat, &sc);
3664 : :
3665 : : /*
3666 : : * If we're getting trouble reclaiming, start doing writepage
3667 : : * even in laptop mode.
3668 : : */
3669 [ # # ]: 0 : if (sc.priority < DEF_PRIORITY - 2)
3670 : 0 : sc.may_writepage = 1;
3671 : :
3672 : : /* Call soft limit reclaim before calling shrink_node. */
3673 : 0 : sc.nr_scanned = 0;
3674 : 0 : nr_soft_scanned = 0;
3675 : 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3676 : : sc.gfp_mask, &nr_soft_scanned);
3677 : 0 : sc.nr_reclaimed += nr_soft_reclaimed;
3678 : :
3679 : : /*
3680 : : * There should be no need to raise the scanning priority if
3681 : : * enough pages are already being scanned that that high
3682 : : * watermark would be met at 100% efficiency.
3683 : : */
3684 [ # # ]: 0 : if (kswapd_shrink_node(pgdat, &sc))
3685 : 0 : raise_priority = false;
3686 : :
3687 : : /*
3688 : : * If the low watermark is met there is no need for processes
3689 : : * to be throttled on pfmemalloc_wait as they should not be
3690 : : * able to safely make forward progress. Wake them
3691 : : */
3692 [ # # # # ]: 0 : if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3693 : 0 : allow_direct_reclaim(pgdat))
3694 : 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3695 : :
3696 : : /* Check if kswapd should be suspending */
3697 : 0 : __fs_reclaim_release();
3698 : 0 : ret = try_to_freeze();
3699 [ # # ]: 0 : __fs_reclaim_acquire();
3700 [ # # # # ]: 0 : if (ret || kthread_should_stop())
3701 : : break;
3702 : :
3703 : : /*
3704 : : * Raise priority if scanning rate is too low or there was no
3705 : : * progress in reclaiming pages
3706 : : */
3707 : 0 : nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3708 : 0 : nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3709 : :
3710 : : /*
3711 : : * If reclaim made no progress for a boost, stop reclaim as
3712 : : * IO cannot be queued and it could be an infinite loop in
3713 : : * extreme circumstances.
3714 : : */
3715 [ # # ]: 0 : if (nr_boost_reclaim && !nr_reclaimed)
3716 : : break;
3717 : :
3718 [ # # ]: 0 : if (raise_priority || !nr_reclaimed)
3719 : 0 : sc.priority--;
3720 [ # # ]: 0 : } while (sc.priority >= 1);
3721 : :
3722 [ # # ]: 0 : if (!sc.nr_reclaimed)
3723 : 0 : pgdat->kswapd_failures++;
3724 : :
3725 : 0 : out:
3726 : : /* If reclaim was boosted, account for the reclaim done in this pass */
3727 [ # # ]: 0 : if (boosted) {
3728 : : unsigned long flags;
3729 : :
3730 [ # # ]: 0 : for (i = 0; i <= classzone_idx; i++) {
3731 [ # # ]: 0 : if (!zone_boosts[i])
3732 : 0 : continue;
3733 : :
3734 : : /* Increments are under the zone lock */
3735 : 0 : zone = pgdat->node_zones + i;
3736 : 0 : spin_lock_irqsave(&zone->lock, flags);
3737 : 0 : zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3738 : 0 : spin_unlock_irqrestore(&zone->lock, flags);
3739 : : }
3740 : :
3741 : : /*
3742 : : * As there is now likely space, wakeup kcompact to defragment
3743 : : * pageblocks.
3744 : : */
3745 : 0 : wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3746 : : }
3747 : :
3748 : 0 : snapshot_refaults(NULL, pgdat);
3749 : 0 : __fs_reclaim_release();
3750 : 0 : psi_memstall_leave(&pflags);
3751 : 0 : set_task_reclaim_state(current, NULL);
3752 : :
3753 : : /*
3754 : : * Return the order kswapd stopped reclaiming at as
3755 : : * prepare_kswapd_sleep() takes it into account. If another caller
3756 : : * entered the allocator slow path while kswapd was awake, order will
3757 : : * remain at the higher level.
3758 : : */
3759 : 0 : return sc.order;
3760 : : }
3761 : :
3762 : : /*
3763 : : * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3764 : : * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3765 : : * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3766 : : * after previous reclaim attempt (node is still unbalanced). In that case
3767 : : * return the zone index of the previous kswapd reclaim cycle.
3768 : : */
3769 : 28 : static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3770 : : enum zone_type prev_classzone_idx)
3771 : : {
3772 : 28 : if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3773 : 28 : return prev_classzone_idx;
3774 : : return pgdat->kswapd_classzone_idx;
3775 : : }
3776 : :
3777 : 28 : static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3778 : : unsigned int classzone_idx)
3779 : : {
3780 : 28 : long remaining = 0;
3781 : 28 : DEFINE_WAIT(wait);
3782 : :
3783 [ + - - + ]: 28 : if (freezing(current) || kthread_should_stop())
3784 : 0 : return;
3785 : :
3786 : 28 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3787 : :
3788 : : /*
3789 : : * Try to sleep for a short interval. Note that kcompactd will only be
3790 : : * woken if it is possible to sleep for a short interval. This is
3791 : : * deliberate on the assumption that if reclaim cannot keep an
3792 : : * eligible zone balanced that it's also unlikely that compaction will
3793 : : * succeed.
3794 : : */
3795 [ + - ]: 28 : if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3796 : : /*
3797 : : * Compaction records what page blocks it recently failed to
3798 : : * isolate pages from and skips them in the future scanning.
3799 : : * When kswapd is going to sleep, it is reasonable to assume
3800 : : * that pages and compaction may succeed so reset the cache.
3801 : : */
3802 : 28 : reset_isolation_suitable(pgdat);
3803 : :
3804 : : /*
3805 : : * We have freed the memory, now we should compact it to make
3806 : : * allocation of the requested order possible.
3807 : : */
3808 : 28 : wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3809 : :
3810 : 28 : remaining = schedule_timeout(HZ/10);
3811 : :
3812 : : /*
3813 : : * If woken prematurely then reset kswapd_classzone_idx and
3814 : : * order. The values will either be from a wakeup request or
3815 : : * the previous request that slept prematurely.
3816 : : */
3817 [ - + ]: 28 : if (remaining) {
3818 [ # # ]: 0 : pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3819 : 0 : pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3820 : : }
3821 : :
3822 : 28 : finish_wait(&pgdat->kswapd_wait, &wait);
3823 : 28 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3824 : : }
3825 : :
3826 : : /*
3827 : : * After a short sleep, check if it was a premature sleep. If not, then
3828 : : * go fully to sleep until explicitly woken up.
3829 : : */
3830 [ + - + - ]: 56 : if (!remaining &&
3831 : 28 : prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3832 : 28 : trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3833 : :
3834 : : /*
3835 : : * vmstat counters are not perfectly accurate and the estimated
3836 : : * value for counters such as NR_FREE_PAGES can deviate from the
3837 : : * true value by nr_online_cpus * threshold. To avoid the zone
3838 : : * watermarks being breached while under pressure, we reduce the
3839 : : * per-cpu vmstat threshold while kswapd is awake and restore
3840 : : * them before going back to sleep.
3841 : : */
3842 : 28 : set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3843 : :
3844 [ + - ]: 28 : if (!kthread_should_stop())
3845 : 28 : schedule();
3846 : :
3847 : 0 : set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3848 : : } else {
3849 [ # # ]: 0 : if (remaining)
3850 : 0 : count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3851 : : else
3852 : 0 : count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3853 : : }
3854 : 0 : finish_wait(&pgdat->kswapd_wait, &wait);
3855 : : }
3856 : :
3857 : : /*
3858 : : * The background pageout daemon, started as a kernel thread
3859 : : * from the init process.
3860 : : *
3861 : : * This basically trickles out pages so that we have _some_
3862 : : * free memory available even if there is no other activity
3863 : : * that frees anything up. This is needed for things like routing
3864 : : * etc, where we otherwise might have all activity going on in
3865 : : * asynchronous contexts that cannot page things out.
3866 : : *
3867 : : * If there are applications that are active memory-allocators
3868 : : * (most normal use), this basically shouldn't matter.
3869 : : */
3870 : 28 : static int kswapd(void *p)
3871 : : {
3872 : 28 : unsigned int alloc_order, reclaim_order;
3873 : 28 : unsigned int classzone_idx = MAX_NR_ZONES - 1;
3874 : 28 : pg_data_t *pgdat = (pg_data_t*)p;
3875 [ + - ]: 28 : struct task_struct *tsk = current;
3876 [ + - ]: 28 : const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3877 : :
3878 [ + - ]: 28 : if (!cpumask_empty(cpumask))
3879 : 28 : set_cpus_allowed_ptr(tsk, cpumask);
3880 : :
3881 : : /*
3882 : : * Tell the memory management that we're a "memory allocator",
3883 : : * and that if we need more memory we should get access to it
3884 : : * regardless (see "__alloc_pages()"). "kswapd" should
3885 : : * never get caught in the normal page freeing logic.
3886 : : *
3887 : : * (Kswapd normally doesn't need memory anyway, but sometimes
3888 : : * you need a small amount of memory in order to be able to
3889 : : * page out something else, and this flag essentially protects
3890 : : * us from recursively trying to free more memory as we're
3891 : : * trying to free the first piece of memory in the first place).
3892 : : */
3893 : 28 : tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3894 : 28 : set_freezable();
3895 : :
3896 : 28 : pgdat->kswapd_order = 0;
3897 : 28 : pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3898 : 28 : for ( ; ; ) {
3899 : 28 : bool ret;
3900 : :
3901 : 28 : alloc_order = reclaim_order = pgdat->kswapd_order;
3902 [ + - ]: 28 : classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3903 : :
3904 : 28 : kswapd_try_sleep:
3905 : 28 : kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3906 : : classzone_idx);
3907 : :
3908 : : /* Read the new order and classzone_idx */
3909 : 0 : alloc_order = reclaim_order = pgdat->kswapd_order;
3910 [ # # ]: 0 : classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3911 : 0 : pgdat->kswapd_order = 0;
3912 : 0 : pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3913 : :
3914 : 0 : ret = try_to_freeze();
3915 [ # # ]: 0 : if (kthread_should_stop())
3916 : : break;
3917 : :
3918 : : /*
3919 : : * We can speed up thawing tasks if we don't call balance_pgdat
3920 : : * after returning from the refrigerator
3921 : : */
3922 [ # # ]: 0 : if (ret)
3923 : 0 : continue;
3924 : :
3925 : : /*
3926 : : * Reclaim begins at the requested order but if a high-order
3927 : : * reclaim fails then kswapd falls back to reclaiming for
3928 : : * order-0. If that happens, kswapd will consider sleeping
3929 : : * for the order it finished reclaiming at (reclaim_order)
3930 : : * but kcompactd is woken to compact for the original
3931 : : * request (alloc_order).
3932 : : */
3933 : 0 : trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3934 : : alloc_order);
3935 : 0 : reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3936 [ # # ]: 0 : if (reclaim_order < alloc_order)
3937 : 0 : goto kswapd_try_sleep;
3938 : : }
3939 : :
3940 : 0 : tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3941 : :
3942 : 0 : return 0;
3943 : : }
3944 : :
3945 : : /*
3946 : : * A zone is low on free memory or too fragmented for high-order memory. If
3947 : : * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3948 : : * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3949 : : * has failed or is not needed, still wake up kcompactd if only compaction is
3950 : : * needed.
3951 : : */
3952 : 0 : void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3953 : : enum zone_type classzone_idx)
3954 : : {
3955 : 0 : pg_data_t *pgdat;
3956 : :
3957 [ # # ]: 0 : if (!managed_zone(zone))
3958 : : return;
3959 : :
3960 [ # # ]: 0 : if (!cpuset_zone_allowed(zone, gfp_flags))
3961 : : return;
3962 : 0 : pgdat = zone->zone_pgdat;
3963 : :
3964 [ # # ]: 0 : if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3965 : 0 : pgdat->kswapd_classzone_idx = classzone_idx;
3966 : : else
3967 : 0 : pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3968 : : classzone_idx);
3969 : 0 : pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3970 [ # # ]: 0 : if (!waitqueue_active(&pgdat->kswapd_wait))
3971 : : return;
3972 : :
3973 : : /* Hopeless node, leave it to direct reclaim if possible */
3974 [ # # # # ]: 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3975 [ # # ]: 0 : (pgdat_balanced(pgdat, order, classzone_idx) &&
3976 : 0 : !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3977 : : /*
3978 : : * There may be plenty of free memory available, but it's too
3979 : : * fragmented for high-order allocations. Wake up kcompactd
3980 : : * and rely on compaction_suitable() to determine if it's
3981 : : * needed. If it fails, it will defer subsequent attempts to
3982 : : * ratelimit its work.
3983 : : */
3984 [ # # ]: 0 : if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3985 : 0 : wakeup_kcompactd(pgdat, order, classzone_idx);
3986 : 0 : return;
3987 : : }
3988 : :
3989 : 0 : trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3990 : : gfp_flags);
3991 : 0 : wake_up_interruptible(&pgdat->kswapd_wait);
3992 : : }
3993 : :
3994 : : #ifdef CONFIG_HIBERNATION
3995 : : /*
3996 : : * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3997 : : * freed pages.
3998 : : *
3999 : : * Rather than trying to age LRUs the aim is to preserve the overall
4000 : : * LRU order by reclaiming preferentially
4001 : : * inactive > active > active referenced > active mapped
4002 : : */
4003 : 0 : unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4004 : : {
4005 : 0 : struct scan_control sc = {
4006 : : .nr_to_reclaim = nr_to_reclaim,
4007 : : .gfp_mask = GFP_HIGHUSER_MOVABLE,
4008 : : .reclaim_idx = MAX_NR_ZONES - 1,
4009 : : .priority = DEF_PRIORITY,
4010 : : .may_writepage = 1,
4011 : : .may_unmap = 1,
4012 : : .may_swap = 1,
4013 : : .hibernation_mode = 1,
4014 : : };
4015 : 0 : struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4016 : 0 : unsigned long nr_reclaimed;
4017 : 0 : unsigned int noreclaim_flag;
4018 : :
4019 : 0 : fs_reclaim_acquire(sc.gfp_mask);
4020 : 0 : noreclaim_flag = memalloc_noreclaim_save();
4021 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
4022 : :
4023 : 0 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4024 : :
4025 : 0 : set_task_reclaim_state(current, NULL);
4026 : 0 : memalloc_noreclaim_restore(noreclaim_flag);
4027 : 0 : fs_reclaim_release(sc.gfp_mask);
4028 : :
4029 : 0 : return nr_reclaimed;
4030 : : }
4031 : : #endif /* CONFIG_HIBERNATION */
4032 : :
4033 : : /* It's optimal to keep kswapds on the same CPUs as their memory, but
4034 : : not required for correctness. So if the last cpu in a node goes
4035 : : away, we get changed to run anywhere: as the first one comes back,
4036 : : restore their cpu bindings. */
4037 : 0 : static int kswapd_cpu_online(unsigned int cpu)
4038 : : {
4039 : 0 : int nid;
4040 : :
4041 [ # # ]: 0 : for_each_node_state(nid, N_MEMORY) {
4042 : 0 : pg_data_t *pgdat = NODE_DATA(nid);
4043 : 0 : const struct cpumask *mask;
4044 : :
4045 : 0 : mask = cpumask_of_node(pgdat->node_id);
4046 : :
4047 [ # # ]: 0 : if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4048 : : /* One of our CPUs online: restore mask */
4049 : 0 : set_cpus_allowed_ptr(pgdat->kswapd, mask);
4050 : : }
4051 : 0 : return 0;
4052 : : }
4053 : :
4054 : : /*
4055 : : * This kswapd start function will be called by init and node-hot-add.
4056 : : * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4057 : : */
4058 : 28 : int kswapd_run(int nid)
4059 : : {
4060 : 28 : pg_data_t *pgdat = NODE_DATA(nid);
4061 : 28 : int ret = 0;
4062 : :
4063 [ + - ]: 28 : if (pgdat->kswapd)
4064 : : return 0;
4065 : :
4066 [ + - ]: 28 : pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4067 [ - + ]: 28 : if (IS_ERR(pgdat->kswapd)) {
4068 : : /* failure at boot is fatal */
4069 [ # # ]: 0 : BUG_ON(system_state < SYSTEM_RUNNING);
4070 : 0 : pr_err("Failed to start kswapd on node %d\n", nid);
4071 : 0 : ret = PTR_ERR(pgdat->kswapd);
4072 : 0 : pgdat->kswapd = NULL;
4073 : : }
4074 : : return ret;
4075 : : }
4076 : :
4077 : : /*
4078 : : * Called by memory hotplug when all memory in a node is offlined. Caller must
4079 : : * hold mem_hotplug_begin/end().
4080 : : */
4081 : 0 : void kswapd_stop(int nid)
4082 : : {
4083 : 0 : struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4084 : :
4085 [ # # ]: 0 : if (kswapd) {
4086 : 0 : kthread_stop(kswapd);
4087 : 0 : NODE_DATA(nid)->kswapd = NULL;
4088 : : }
4089 : 0 : }
4090 : :
4091 : 28 : static int __init kswapd_init(void)
4092 : : {
4093 : 28 : int nid, ret;
4094 : :
4095 : 28 : swap_setup();
4096 [ + + ]: 112 : for_each_node_state(nid, N_MEMORY)
4097 : 28 : kswapd_run(nid);
4098 : 28 : ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4099 : : "mm/vmscan:online", kswapd_cpu_online,
4100 : : NULL);
4101 [ - + ]: 28 : WARN_ON(ret < 0);
4102 : 28 : return 0;
4103 : : }
4104 : :
4105 : : module_init(kswapd_init)
4106 : :
4107 : : #ifdef CONFIG_NUMA
4108 : : /*
4109 : : * Node reclaim mode
4110 : : *
4111 : : * If non-zero call node_reclaim when the number of free pages falls below
4112 : : * the watermarks.
4113 : : */
4114 : : int node_reclaim_mode __read_mostly;
4115 : :
4116 : : #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4117 : : #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4118 : :
4119 : : /*
4120 : : * Priority for NODE_RECLAIM. This determines the fraction of pages
4121 : : * of a node considered for each zone_reclaim. 4 scans 1/16th of
4122 : : * a zone.
4123 : : */
4124 : : #define NODE_RECLAIM_PRIORITY 4
4125 : :
4126 : : /*
4127 : : * Percentage of pages in a zone that must be unmapped for node_reclaim to
4128 : : * occur.
4129 : : */
4130 : : int sysctl_min_unmapped_ratio = 1;
4131 : :
4132 : : /*
4133 : : * If the number of slab pages in a zone grows beyond this percentage then
4134 : : * slab reclaim needs to occur.
4135 : : */
4136 : : int sysctl_min_slab_ratio = 5;
4137 : :
4138 : 0 : static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4139 : : {
4140 : 0 : unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4141 : 0 : unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4142 : 0 : node_page_state(pgdat, NR_ACTIVE_FILE);
4143 : :
4144 : : /*
4145 : : * It's possible for there to be more file mapped pages than
4146 : : * accounted for by the pages on the file LRU lists because
4147 : : * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4148 : : */
4149 [ # # ]: 0 : return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4150 : : }
4151 : :
4152 : : /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4153 : 0 : static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4154 : : {
4155 : 0 : unsigned long nr_pagecache_reclaimable;
4156 : 0 : unsigned long delta = 0;
4157 : :
4158 : : /*
4159 : : * If RECLAIM_UNMAP is set, then all file pages are considered
4160 : : * potentially reclaimable. Otherwise, we have to worry about
4161 : : * pages like swapcache and node_unmapped_file_pages() provides
4162 : : * a better estimate
4163 : : */
4164 [ # # ]: 0 : if (node_reclaim_mode & RECLAIM_UNMAP)
4165 : 0 : nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4166 : : else
4167 : 0 : nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4168 : :
4169 : : /* If we can't clean pages, remove dirty pages from consideration */
4170 [ # # ]: 0 : if (!(node_reclaim_mode & RECLAIM_WRITE))
4171 : 0 : delta += node_page_state(pgdat, NR_FILE_DIRTY);
4172 : :
4173 : : /* Watch for any possible underflows due to delta */
4174 [ # # ]: 0 : if (unlikely(delta > nr_pagecache_reclaimable))
4175 : 0 : delta = nr_pagecache_reclaimable;
4176 : :
4177 : 0 : return nr_pagecache_reclaimable - delta;
4178 : : }
4179 : :
4180 : : /*
4181 : : * Try to free up some pages from this node through reclaim.
4182 : : */
4183 : 0 : static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4184 : : {
4185 : : /* Minimum pages needed in order to stay on node */
4186 : 0 : const unsigned long nr_pages = 1 << order;
4187 [ # # ]: 0 : struct task_struct *p = current;
4188 : 0 : unsigned int noreclaim_flag;
4189 : 0 : struct scan_control sc = {
4190 : 0 : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4191 : : .gfp_mask = current_gfp_context(gfp_mask),
4192 : : .order = order,
4193 : : .priority = NODE_RECLAIM_PRIORITY,
4194 : 0 : .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4195 [ # # ]: 0 : .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4196 : : .may_swap = 1,
4197 : : .reclaim_idx = gfp_zone(gfp_mask),
4198 : : };
4199 : :
4200 : 0 : trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4201 : : sc.gfp_mask);
4202 : :
4203 : 0 : cond_resched();
4204 : 0 : fs_reclaim_acquire(sc.gfp_mask);
4205 : : /*
4206 : : * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4207 : : * and we also need to be able to write out pages for RECLAIM_WRITE
4208 : : * and RECLAIM_UNMAP.
4209 : : */
4210 : 0 : noreclaim_flag = memalloc_noreclaim_save();
4211 : 0 : p->flags |= PF_SWAPWRITE;
4212 : 0 : set_task_reclaim_state(p, &sc.reclaim_state);
4213 : :
4214 [ # # ]: 0 : if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4215 : : /*
4216 : : * Free memory by calling shrink node with increasing
4217 : : * priorities until we have enough memory freed.
4218 : : */
4219 : 0 : do {
4220 : 0 : shrink_node(pgdat, &sc);
4221 [ # # # # ]: 0 : } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4222 : : }
4223 : :
4224 : 0 : set_task_reclaim_state(p, NULL);
4225 : 0 : current->flags &= ~PF_SWAPWRITE;
4226 : 0 : memalloc_noreclaim_restore(noreclaim_flag);
4227 : 0 : fs_reclaim_release(sc.gfp_mask);
4228 : :
4229 : 0 : trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4230 : :
4231 : 0 : return sc.nr_reclaimed >= nr_pages;
4232 : : }
4233 : :
4234 : 0 : int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4235 : : {
4236 : 0 : int ret;
4237 : :
4238 : : /*
4239 : : * Node reclaim reclaims unmapped file backed pages and
4240 : : * slab pages if we are over the defined limits.
4241 : : *
4242 : : * A small portion of unmapped file backed pages is needed for
4243 : : * file I/O otherwise pages read by file I/O will be immediately
4244 : : * thrown out if the node is overallocated. So we do not reclaim
4245 : : * if less than a specified percentage of the node is used by
4246 : : * unmapped file backed pages.
4247 : : */
4248 [ # # ]: 0 : if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4249 [ # # ]: 0 : node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4250 : : return NODE_RECLAIM_FULL;
4251 : :
4252 : : /*
4253 : : * Do not scan if the allocation should not be delayed.
4254 : : */
4255 [ # # # # ]: 0 : if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4256 : : return NODE_RECLAIM_NOSCAN;
4257 : :
4258 : : /*
4259 : : * Only run node reclaim on the local node or on nodes that do not
4260 : : * have associated processors. This will favor the local processor
4261 : : * over remote processors and spread off node memory allocations
4262 : : * as wide as possible.
4263 : : */
4264 [ # # # # ]: 0 : if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4265 : : return NODE_RECLAIM_NOSCAN;
4266 : :
4267 [ # # ]: 0 : if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4268 : : return NODE_RECLAIM_NOSCAN;
4269 : :
4270 : 0 : ret = __node_reclaim(pgdat, gfp_mask, order);
4271 : 0 : clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4272 : :
4273 [ # # ]: 0 : if (!ret)
4274 : 0 : count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4275 : :
4276 : : return ret;
4277 : : }
4278 : : #endif
4279 : :
4280 : : /*
4281 : : * page_evictable - test whether a page is evictable
4282 : : * @page: the page to test
4283 : : *
4284 : : * Test whether page is evictable--i.e., should be placed on active/inactive
4285 : : * lists vs unevictable list.
4286 : : *
4287 : : * Reasons page might not be evictable:
4288 : : * (1) page's mapping marked unevictable
4289 : : * (2) page is part of an mlocked VMA
4290 : : *
4291 : : */
4292 : 1507617 : int page_evictable(struct page *page)
4293 : : {
4294 : 1507617 : int ret;
4295 : :
4296 : : /* Prevent address_space of inode and swap cache from being freed */
4297 : 1507617 : rcu_read_lock();
4298 [ + - - + ]: 3500868 : ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4299 : 1507617 : rcu_read_unlock();
4300 : 1507617 : return ret;
4301 : : }
4302 : :
4303 : : /**
4304 : : * check_move_unevictable_pages - check pages for evictability and move to
4305 : : * appropriate zone lru list
4306 : : * @pvec: pagevec with lru pages to check
4307 : : *
4308 : : * Checks pages for evictability, if an evictable page is in the unevictable
4309 : : * lru list, moves it to the appropriate evictable lru list. This function
4310 : : * should be only used for lru pages.
4311 : : */
4312 : 0 : void check_move_unevictable_pages(struct pagevec *pvec)
4313 : : {
4314 : 0 : struct lruvec *lruvec;
4315 : 0 : struct pglist_data *pgdat = NULL;
4316 : 0 : int pgscanned = 0;
4317 : 0 : int pgrescued = 0;
4318 : 0 : int i;
4319 : :
4320 [ # # ]: 0 : for (i = 0; i < pvec->nr; i++) {
4321 : 0 : struct page *page = pvec->pages[i];
4322 [ # # ]: 0 : struct pglist_data *pagepgdat = page_pgdat(page);
4323 : :
4324 : 0 : pgscanned++;
4325 [ # # ]: 0 : if (pagepgdat != pgdat) {
4326 [ # # ]: 0 : if (pgdat)
4327 : 0 : spin_unlock_irq(&pgdat->lru_lock);
4328 : 0 : pgdat = pagepgdat;
4329 : 0 : spin_lock_irq(&pgdat->lru_lock);
4330 : : }
4331 [ # # ]: 0 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
4332 : :
4333 [ # # # # : 0 : if (!PageLRU(page) || !PageUnevictable(page))
# # ]
4334 : 0 : continue;
4335 : :
4336 [ # # ]: 0 : if (page_evictable(page)) {
4337 : 0 : enum lru_list lru = page_lru_base_type(page);
4338 : :
4339 : 0 : VM_BUG_ON_PAGE(PageActive(page), page);
4340 [ # # ]: 0 : ClearPageUnevictable(page);
4341 : 0 : del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4342 : 0 : add_page_to_lru_list(page, lruvec, lru);
4343 : 0 : pgrescued++;
4344 : : }
4345 : : }
4346 : :
4347 [ # # ]: 0 : if (pgdat) {
4348 [ # # ]: 0 : __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4349 [ # # ]: 0 : __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4350 : 0 : spin_unlock_irq(&pgdat->lru_lock);
4351 : : }
4352 : 0 : }
4353 : : EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
|
