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 : : /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 : : unsigned int may_writepage:1;
84 : :
85 : : /* Can mapped pages be reclaimed? */
86 : : unsigned int may_unmap:1;
87 : :
88 : : /* Can pages be swapped as part of reclaim? */
89 : : unsigned int may_swap:1;
90 : :
91 : : /*
92 : : * Cgroups are not reclaimed below their configured memory.low,
93 : : * unless we threaten to OOM. If any cgroups are skipped due to
94 : : * memory.low and nothing was reclaimed, go back for memory.low.
95 : : */
96 : : unsigned int memcg_low_reclaim:1;
97 : : unsigned int memcg_low_skipped:1;
98 : :
99 : : unsigned int hibernation_mode:1;
100 : :
101 : : /* One of the zones is ready for compaction */
102 : : unsigned int compaction_ready:1;
103 : :
104 : : /* Allocation order */
105 : : s8 order;
106 : :
107 : : /* Scan (total_size >> priority) pages at once */
108 : : s8 priority;
109 : :
110 : : /* The highest zone to isolate pages for reclaim from */
111 : : s8 reclaim_idx;
112 : :
113 : : /* This context's GFP mask */
114 : : gfp_t gfp_mask;
115 : :
116 : : /* Incremented by the number of inactive pages that were scanned */
117 : : unsigned long nr_scanned;
118 : :
119 : : /* Number of pages freed so far during a call to shrink_zones() */
120 : : unsigned long nr_reclaimed;
121 : :
122 : : struct {
123 : : unsigned int dirty;
124 : : unsigned int unqueued_dirty;
125 : : unsigned int congested;
126 : : unsigned int writeback;
127 : : unsigned int immediate;
128 : : unsigned int file_taken;
129 : : unsigned int taken;
130 : : } nr;
131 : :
132 : : /* for recording the reclaimed slab by now */
133 : : struct reclaim_state reclaim_state;
134 : : };
135 : :
136 : : #ifdef ARCH_HAS_PREFETCH
137 : : #define prefetch_prev_lru_page(_page, _base, _field) \
138 : : do { \
139 : : if ((_page)->lru.prev != _base) { \
140 : : struct page *prev; \
141 : : \
142 : : prev = lru_to_page(&(_page->lru)); \
143 : : prefetch(&prev->_field); \
144 : : } \
145 : : } while (0)
146 : : #else
147 : : #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148 : : #endif
149 : :
150 : : #ifdef ARCH_HAS_PREFETCHW
151 : : #define prefetchw_prev_lru_page(_page, _base, _field) \
152 : : do { \
153 : : if ((_page)->lru.prev != _base) { \
154 : : struct page *prev; \
155 : : \
156 : : prev = lru_to_page(&(_page->lru)); \
157 : : prefetchw(&prev->_field); \
158 : : } \
159 : : } while (0)
160 : : #else
161 : : #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162 : : #endif
163 : :
164 : : /*
165 : : * From 0 .. 100. Higher means more swappy.
166 : : */
167 : : int vm_swappiness = 60;
168 : : /*
169 : : * The total number of pages which are beyond the high watermark within all
170 : : * zones.
171 : : */
172 : : unsigned long vm_total_pages;
173 : :
174 : 0 : static void set_task_reclaim_state(struct task_struct *task,
175 : : struct reclaim_state *rs)
176 : : {
177 : : /* Check for an overwrite */
178 : 0 : WARN_ON_ONCE(rs && task->reclaim_state);
179 : :
180 : : /* Check for the nulling of an already-nulled member */
181 : 0 : WARN_ON_ONCE(!rs && !task->reclaim_state);
182 : :
183 : 0 : task->reclaim_state = rs;
184 : 0 : }
185 : :
186 : : static LIST_HEAD(shrinker_list);
187 : : static DECLARE_RWSEM(shrinker_rwsem);
188 : :
189 : : #ifdef CONFIG_MEMCG
190 : : /*
191 : : * We allow subsystems to populate their shrinker-related
192 : : * LRU lists before register_shrinker_prepared() is called
193 : : * for the shrinker, since we don't want to impose
194 : : * restrictions on their internal registration order.
195 : : * In this case shrink_slab_memcg() may find corresponding
196 : : * bit is set in the shrinkers map.
197 : : *
198 : : * This value is used by the function to detect registering
199 : : * shrinkers and to skip do_shrink_slab() calls for them.
200 : : */
201 : : #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202 : :
203 : : static DEFINE_IDR(shrinker_idr);
204 : : static int shrinker_nr_max;
205 : :
206 : 3 : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207 : : {
208 : : int id, ret = -ENOMEM;
209 : :
210 : 3 : down_write(&shrinker_rwsem);
211 : : /* This may call shrinker, so it must use down_read_trylock() */
212 : 3 : id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
213 : 3 : if (id < 0)
214 : : goto unlock;
215 : :
216 : 3 : if (id >= shrinker_nr_max) {
217 : 3 : if (memcg_expand_shrinker_maps(id)) {
218 : 0 : idr_remove(&shrinker_idr, id);
219 : 0 : goto unlock;
220 : : }
221 : :
222 : 3 : shrinker_nr_max = id + 1;
223 : : }
224 : 3 : shrinker->id = id;
225 : : ret = 0;
226 : : unlock:
227 : 3 : up_write(&shrinker_rwsem);
228 : 3 : return ret;
229 : : }
230 : :
231 : 3 : static void unregister_memcg_shrinker(struct shrinker *shrinker)
232 : : {
233 : 3 : int id = shrinker->id;
234 : :
235 : 3 : BUG_ON(id < 0);
236 : :
237 : 3 : down_write(&shrinker_rwsem);
238 : 3 : idr_remove(&shrinker_idr, id);
239 : 3 : up_write(&shrinker_rwsem);
240 : 3 : }
241 : :
242 : : static bool global_reclaim(struct scan_control *sc)
243 : : {
244 : 0 : return !sc->target_mem_cgroup;
245 : : }
246 : :
247 : : /**
248 : : * sane_reclaim - is the usual dirty throttling mechanism operational?
249 : : * @sc: scan_control in question
250 : : *
251 : : * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 : : * completely broken with the legacy memcg and direct stalling in
253 : : * shrink_page_list() is used for throttling instead, which lacks all the
254 : : * niceties such as fairness, adaptive pausing, bandwidth proportional
255 : : * allocation and configurability.
256 : : *
257 : : * This function tests whether the vmscan currently in progress can assume
258 : : * that the normal dirty throttling mechanism is operational.
259 : : */
260 : : static bool sane_reclaim(struct scan_control *sc)
261 : : {
262 : 0 : struct mem_cgroup *memcg = sc->target_mem_cgroup;
263 : :
264 : 0 : if (!memcg)
265 : : return true;
266 : : #ifdef CONFIG_CGROUP_WRITEBACK
267 : 0 : if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
268 : : return true;
269 : : #endif
270 : : return false;
271 : : }
272 : :
273 : : static void set_memcg_congestion(pg_data_t *pgdat,
274 : : struct mem_cgroup *memcg,
275 : : bool congested)
276 : : {
277 : : struct mem_cgroup_per_node *mn;
278 : :
279 : 0 : if (!memcg)
280 : : return;
281 : :
282 : 0 : mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 : : WRITE_ONCE(mn->congested, congested);
284 : : }
285 : :
286 : : static bool memcg_congested(pg_data_t *pgdat,
287 : : struct mem_cgroup *memcg)
288 : : {
289 : : struct mem_cgroup_per_node *mn;
290 : :
291 : 0 : mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 : 0 : return READ_ONCE(mn->congested);
293 : :
294 : : }
295 : : #else
296 : : static int prealloc_memcg_shrinker(struct shrinker *shrinker)
297 : : {
298 : : return 0;
299 : : }
300 : :
301 : : static void unregister_memcg_shrinker(struct shrinker *shrinker)
302 : : {
303 : : }
304 : :
305 : : static bool global_reclaim(struct scan_control *sc)
306 : : {
307 : : return true;
308 : : }
309 : :
310 : : static bool sane_reclaim(struct scan_control *sc)
311 : : {
312 : : return true;
313 : : }
314 : :
315 : : static inline void set_memcg_congestion(struct pglist_data *pgdat,
316 : : struct mem_cgroup *memcg, bool congested)
317 : : {
318 : : }
319 : :
320 : : static inline bool memcg_congested(struct pglist_data *pgdat,
321 : : struct mem_cgroup *memcg)
322 : : {
323 : : return false;
324 : :
325 : : }
326 : : #endif
327 : :
328 : : /*
329 : : * This misses isolated pages which are not accounted for to save counters.
330 : : * As the data only determines if reclaim or compaction continues, it is
331 : : * not expected that isolated pages will be a dominating factor.
332 : : */
333 : 0 : unsigned long zone_reclaimable_pages(struct zone *zone)
334 : : {
335 : : unsigned long nr;
336 : :
337 : 0 : nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
338 : 0 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
339 : 0 : if (get_nr_swap_pages() > 0)
340 : 0 : nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
341 : 0 : zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
342 : :
343 : 0 : return nr;
344 : : }
345 : :
346 : : /**
347 : : * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 : : * @lruvec: lru vector
349 : : * @lru: lru to use
350 : : * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
351 : : */
352 : 0 : unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
353 : : {
354 : : unsigned long lru_size = 0;
355 : : int zid;
356 : :
357 : 0 : if (!mem_cgroup_disabled()) {
358 : 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++)
359 : 0 : lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
360 : : } else
361 : : lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
362 : :
363 : 0 : for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
364 : 0 : struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
365 : : unsigned long size;
366 : :
367 : 0 : if (!managed_zone(zone))
368 : 0 : continue;
369 : :
370 : 0 : if (!mem_cgroup_disabled())
371 : : size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
372 : : else
373 : : size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
374 : 0 : NR_ZONE_LRU_BASE + lru);
375 : 0 : lru_size -= min(size, lru_size);
376 : : }
377 : :
378 : 0 : return lru_size;
379 : :
380 : : }
381 : :
382 : : /*
383 : : * Add a shrinker callback to be called from the vm.
384 : : */
385 : 3 : int prealloc_shrinker(struct shrinker *shrinker)
386 : : {
387 : : unsigned int size = sizeof(*shrinker->nr_deferred);
388 : :
389 : : if (shrinker->flags & SHRINKER_NUMA_AWARE)
390 : : size *= nr_node_ids;
391 : :
392 : 3 : shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
393 : 3 : if (!shrinker->nr_deferred)
394 : : return -ENOMEM;
395 : :
396 : 3 : if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
397 : 3 : if (prealloc_memcg_shrinker(shrinker))
398 : : goto free_deferred;
399 : : }
400 : :
401 : : return 0;
402 : :
403 : : free_deferred:
404 : 0 : kfree(shrinker->nr_deferred);
405 : 0 : shrinker->nr_deferred = NULL;
406 : 0 : return -ENOMEM;
407 : : }
408 : :
409 : 0 : void free_prealloced_shrinker(struct shrinker *shrinker)
410 : : {
411 : 0 : if (!shrinker->nr_deferred)
412 : 0 : return;
413 : :
414 : 0 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
415 : 0 : unregister_memcg_shrinker(shrinker);
416 : :
417 : 0 : kfree(shrinker->nr_deferred);
418 : 0 : shrinker->nr_deferred = NULL;
419 : : }
420 : :
421 : 3 : void register_shrinker_prepared(struct shrinker *shrinker)
422 : : {
423 : 3 : down_write(&shrinker_rwsem);
424 : 3 : list_add_tail(&shrinker->list, &shrinker_list);
425 : : #ifdef CONFIG_MEMCG
426 : 3 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
427 : 3 : idr_replace(&shrinker_idr, shrinker, shrinker->id);
428 : : #endif
429 : 3 : up_write(&shrinker_rwsem);
430 : 3 : }
431 : :
432 : 3 : int register_shrinker(struct shrinker *shrinker)
433 : : {
434 : 3 : int err = prealloc_shrinker(shrinker);
435 : :
436 : 3 : if (err)
437 : : return err;
438 : 3 : register_shrinker_prepared(shrinker);
439 : 3 : return 0;
440 : : }
441 : : EXPORT_SYMBOL(register_shrinker);
442 : :
443 : : /*
444 : : * Remove one
445 : : */
446 : 3 : void unregister_shrinker(struct shrinker *shrinker)
447 : : {
448 : 3 : if (!shrinker->nr_deferred)
449 : 3 : return;
450 : 3 : if (shrinker->flags & SHRINKER_MEMCG_AWARE)
451 : 3 : unregister_memcg_shrinker(shrinker);
452 : 3 : down_write(&shrinker_rwsem);
453 : : list_del(&shrinker->list);
454 : 3 : up_write(&shrinker_rwsem);
455 : 3 : kfree(shrinker->nr_deferred);
456 : 3 : shrinker->nr_deferred = NULL;
457 : : }
458 : : EXPORT_SYMBOL(unregister_shrinker);
459 : :
460 : : #define SHRINK_BATCH 128
461 : :
462 : 0 : static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
463 : : struct shrinker *shrinker, int priority)
464 : : {
465 : : unsigned long freed = 0;
466 : : unsigned long long delta;
467 : : long total_scan;
468 : : long freeable;
469 : : long nr;
470 : : long new_nr;
471 : 0 : int nid = shrinkctl->nid;
472 : 0 : long batch_size = shrinker->batch ? shrinker->batch
473 : 0 : : SHRINK_BATCH;
474 : : long scanned = 0, next_deferred;
475 : :
476 : 0 : if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
477 : : nid = 0;
478 : :
479 : 0 : freeable = shrinker->count_objects(shrinker, shrinkctl);
480 : 0 : if (freeable == 0 || freeable == SHRINK_EMPTY)
481 : : return freeable;
482 : :
483 : : /*
484 : : * copy the current shrinker scan count into a local variable
485 : : * and zero it so that other concurrent shrinker invocations
486 : : * don't also do this scanning work.
487 : : */
488 : 0 : nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
489 : :
490 : : total_scan = nr;
491 : 0 : if (shrinker->seeks) {
492 : 0 : delta = freeable >> priority;
493 : 0 : delta *= 4;
494 : 0 : do_div(delta, shrinker->seeks);
495 : : } else {
496 : : /*
497 : : * These objects don't require any IO to create. Trim
498 : : * them aggressively under memory pressure to keep
499 : : * them from causing refetches in the IO caches.
500 : : */
501 : 0 : delta = freeable / 2;
502 : : }
503 : :
504 : 0 : total_scan += delta;
505 : 0 : if (total_scan < 0) {
506 : 0 : pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 : : shrinker->scan_objects, total_scan);
508 : : total_scan = freeable;
509 : : next_deferred = nr;
510 : : } else
511 : : next_deferred = total_scan;
512 : :
513 : : /*
514 : : * We need to avoid excessive windup on filesystem shrinkers
515 : : * due to large numbers of GFP_NOFS allocations causing the
516 : : * shrinkers to return -1 all the time. This results in a large
517 : : * nr being built up so when a shrink that can do some work
518 : : * comes along it empties the entire cache due to nr >>>
519 : : * freeable. This is bad for sustaining a working set in
520 : : * memory.
521 : : *
522 : : * Hence only allow the shrinker to scan the entire cache when
523 : : * a large delta change is calculated directly.
524 : : */
525 : 0 : if (delta < freeable / 4)
526 : 0 : total_scan = min(total_scan, freeable / 2);
527 : :
528 : : /*
529 : : * Avoid risking looping forever due to too large nr value:
530 : : * never try to free more than twice the estimate number of
531 : : * freeable entries.
532 : : */
533 : 0 : if (total_scan > freeable * 2)
534 : : total_scan = freeable * 2;
535 : :
536 : 0 : trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 : : freeable, delta, total_scan, priority);
538 : :
539 : : /*
540 : : * Normally, we should not scan less than batch_size objects in one
541 : : * pass to avoid too frequent shrinker calls, but if the slab has less
542 : : * than batch_size objects in total and we are really tight on memory,
543 : : * we will try to reclaim all available objects, otherwise we can end
544 : : * up failing allocations although there are plenty of reclaimable
545 : : * objects spread over several slabs with usage less than the
546 : : * batch_size.
547 : : *
548 : : * We detect the "tight on memory" situations by looking at the total
549 : : * number of objects we want to scan (total_scan). If it is greater
550 : : * than the total number of objects on slab (freeable), we must be
551 : : * scanning at high prio and therefore should try to reclaim as much as
552 : : * possible.
553 : : */
554 : 0 : while (total_scan >= batch_size ||
555 : 0 : total_scan >= freeable) {
556 : : unsigned long ret;
557 : 0 : unsigned long nr_to_scan = min(batch_size, total_scan);
558 : :
559 : 0 : shrinkctl->nr_to_scan = nr_to_scan;
560 : 0 : shrinkctl->nr_scanned = nr_to_scan;
561 : 0 : ret = shrinker->scan_objects(shrinker, shrinkctl);
562 : 0 : if (ret == SHRINK_STOP)
563 : : break;
564 : 0 : freed += ret;
565 : :
566 : 0 : count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 : 0 : total_scan -= shrinkctl->nr_scanned;
568 : 0 : scanned += shrinkctl->nr_scanned;
569 : :
570 : 0 : cond_resched();
571 : : }
572 : :
573 : 0 : if (next_deferred >= scanned)
574 : 0 : next_deferred -= scanned;
575 : : else
576 : : next_deferred = 0;
577 : : /*
578 : : * move the unused scan count back into the shrinker in a
579 : : * manner that handles concurrent updates. If we exhausted the
580 : : * scan, there is no need to do an update.
581 : : */
582 : 0 : if (next_deferred > 0)
583 : 0 : new_nr = atomic_long_add_return(next_deferred,
584 : 0 : &shrinker->nr_deferred[nid]);
585 : : else
586 : 0 : new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
587 : :
588 : 0 : trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
589 : 0 : return freed;
590 : : }
591 : :
592 : : #ifdef CONFIG_MEMCG
593 : 0 : static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 : : struct mem_cgroup *memcg, int priority)
595 : : {
596 : : struct memcg_shrinker_map *map;
597 : : unsigned long ret, freed = 0;
598 : : int i;
599 : :
600 : 0 : if (!mem_cgroup_online(memcg))
601 : : return 0;
602 : :
603 : 0 : if (!down_read_trylock(&shrinker_rwsem))
604 : : return 0;
605 : :
606 : 0 : map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
607 : : true);
608 : 0 : if (unlikely(!map))
609 : : goto unlock;
610 : :
611 : 0 : for_each_set_bit(i, map->map, shrinker_nr_max) {
612 : 0 : struct shrink_control sc = {
613 : : .gfp_mask = gfp_mask,
614 : : .nid = nid,
615 : : .memcg = memcg,
616 : : };
617 : : struct shrinker *shrinker;
618 : :
619 : 0 : shrinker = idr_find(&shrinker_idr, i);
620 : 0 : if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
621 : 0 : if (!shrinker)
622 : 0 : clear_bit(i, map->map);
623 : 0 : continue;
624 : : }
625 : :
626 : : /* Call non-slab shrinkers even though kmem is disabled */
627 : 0 : if (!memcg_kmem_enabled() &&
628 : 0 : !(shrinker->flags & SHRINKER_NONSLAB))
629 : 0 : continue;
630 : :
631 : 0 : ret = do_shrink_slab(&sc, shrinker, priority);
632 : 0 : if (ret == SHRINK_EMPTY) {
633 : 0 : clear_bit(i, map->map);
634 : : /*
635 : : * After the shrinker reported that it had no objects to
636 : : * free, but before we cleared the corresponding bit in
637 : : * the memcg shrinker map, a new object might have been
638 : : * added. To make sure, we have the bit set in this
639 : : * case, we invoke the shrinker one more time and reset
640 : : * the bit if it reports that it is not empty anymore.
641 : : * The memory barrier here pairs with the barrier in
642 : : * memcg_set_shrinker_bit():
643 : : *
644 : : * list_lru_add() shrink_slab_memcg()
645 : : * list_add_tail() clear_bit()
646 : : * <MB> <MB>
647 : : * set_bit() do_shrink_slab()
648 : : */
649 : 0 : smp_mb__after_atomic();
650 : 0 : ret = do_shrink_slab(&sc, shrinker, priority);
651 : 0 : if (ret == SHRINK_EMPTY)
652 : : ret = 0;
653 : : else
654 : 0 : memcg_set_shrinker_bit(memcg, nid, i);
655 : : }
656 : 0 : freed += ret;
657 : :
658 : 0 : if (rwsem_is_contended(&shrinker_rwsem)) {
659 : 0 : freed = freed ? : 1;
660 : 0 : break;
661 : : }
662 : : }
663 : : unlock:
664 : 0 : up_read(&shrinker_rwsem);
665 : 0 : return freed;
666 : : }
667 : : #else /* CONFIG_MEMCG */
668 : : static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 : : struct mem_cgroup *memcg, int priority)
670 : : {
671 : : return 0;
672 : : }
673 : : #endif /* CONFIG_MEMCG */
674 : :
675 : : /**
676 : : * shrink_slab - shrink slab caches
677 : : * @gfp_mask: allocation context
678 : : * @nid: node whose slab caches to target
679 : : * @memcg: memory cgroup whose slab caches to target
680 : : * @priority: the reclaim priority
681 : : *
682 : : * Call the shrink functions to age shrinkable caches.
683 : : *
684 : : * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 : : * unaware shrinkers will receive a node id of 0 instead.
686 : : *
687 : : * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 : : * are called only if it is the root cgroup.
689 : : *
690 : : * @priority is sc->priority, we take the number of objects and >> by priority
691 : : * in order to get the scan target.
692 : : *
693 : : * Returns the number of reclaimed slab objects.
694 : : */
695 : 0 : static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 : : struct mem_cgroup *memcg,
697 : : int priority)
698 : : {
699 : : unsigned long ret, freed = 0;
700 : : struct shrinker *shrinker;
701 : :
702 : : /*
703 : : * The root memcg might be allocated even though memcg is disabled
704 : : * via "cgroup_disable=memory" boot parameter. This could make
705 : : * mem_cgroup_is_root() return false, then just run memcg slab
706 : : * shrink, but skip global shrink. This may result in premature
707 : : * oom.
708 : : */
709 : 0 : if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 : 0 : return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
711 : :
712 : 0 : if (!down_read_trylock(&shrinker_rwsem))
713 : : goto out;
714 : :
715 : 0 : list_for_each_entry(shrinker, &shrinker_list, list) {
716 : 0 : struct shrink_control sc = {
717 : : .gfp_mask = gfp_mask,
718 : : .nid = nid,
719 : : .memcg = memcg,
720 : : };
721 : :
722 : 0 : ret = do_shrink_slab(&sc, shrinker, priority);
723 : 0 : if (ret == SHRINK_EMPTY)
724 : : ret = 0;
725 : 0 : freed += ret;
726 : : /*
727 : : * Bail out if someone want to register a new shrinker to
728 : : * prevent the regsitration from being stalled for long periods
729 : : * by parallel ongoing shrinking.
730 : : */
731 : 0 : if (rwsem_is_contended(&shrinker_rwsem)) {
732 : 0 : freed = freed ? : 1;
733 : 0 : break;
734 : : }
735 : : }
736 : :
737 : 0 : up_read(&shrinker_rwsem);
738 : : out:
739 : 0 : cond_resched();
740 : 0 : return freed;
741 : : }
742 : :
743 : 0 : void drop_slab_node(int nid)
744 : : {
745 : : unsigned long freed;
746 : :
747 : : do {
748 : : struct mem_cgroup *memcg = NULL;
749 : :
750 : : freed = 0;
751 : 0 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
752 : : do {
753 : 0 : freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 : 0 : } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 : 0 : } while (freed > 10);
756 : 0 : }
757 : :
758 : 0 : void drop_slab(void)
759 : : {
760 : : int nid;
761 : :
762 : 0 : for_each_online_node(nid)
763 : 0 : drop_slab_node(nid);
764 : 0 : }
765 : :
766 : : static inline int is_page_cache_freeable(struct page *page)
767 : : {
768 : : /*
769 : : * A freeable page cache page is referenced only by the caller
770 : : * that isolated the page, the page cache and optional buffer
771 : : * heads at page->private.
772 : : */
773 : : int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
774 : : HPAGE_PMD_NR : 1;
775 : 0 : return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
776 : : }
777 : :
778 : 0 : static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
779 : : {
780 : 0 : if (current->flags & PF_SWAPWRITE)
781 : : return 1;
782 : 0 : if (!inode_write_congested(inode))
783 : : return 1;
784 : 0 : if (inode_to_bdi(inode) == current->backing_dev_info)
785 : : return 1;
786 : 0 : return 0;
787 : : }
788 : :
789 : : /*
790 : : * We detected a synchronous write error writing a page out. Probably
791 : : * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 : : * fsync(), msync() or close().
793 : : *
794 : : * The tricky part is that after writepage we cannot touch the mapping: nothing
795 : : * prevents it from being freed up. But we have a ref on the page and once
796 : : * that page is locked, the mapping is pinned.
797 : : *
798 : : * We're allowed to run sleeping lock_page() here because we know the caller has
799 : : * __GFP_FS.
800 : : */
801 : 0 : static void handle_write_error(struct address_space *mapping,
802 : : struct page *page, int error)
803 : : {
804 : 0 : lock_page(page);
805 : 0 : if (page_mapping(page) == mapping)
806 : 0 : mapping_set_error(mapping, error);
807 : 0 : unlock_page(page);
808 : 0 : }
809 : :
810 : : /* possible outcome of pageout() */
811 : : typedef enum {
812 : : /* failed to write page out, page is locked */
813 : : PAGE_KEEP,
814 : : /* move page to the active list, page is locked */
815 : : PAGE_ACTIVATE,
816 : : /* page has been sent to the disk successfully, page is unlocked */
817 : : PAGE_SUCCESS,
818 : : /* page is clean and locked */
819 : : PAGE_CLEAN,
820 : : } pageout_t;
821 : :
822 : : /*
823 : : * pageout is called by shrink_page_list() for each dirty page.
824 : : * Calls ->writepage().
825 : : */
826 : 0 : static pageout_t pageout(struct page *page, struct address_space *mapping,
827 : : struct scan_control *sc)
828 : : {
829 : : /*
830 : : * If the page is dirty, only perform writeback if that write
831 : : * will be non-blocking. To prevent this allocation from being
832 : : * stalled by pagecache activity. But note that there may be
833 : : * stalls if we need to run get_block(). We could test
834 : : * PagePrivate for that.
835 : : *
836 : : * If this process is currently in __generic_file_write_iter() against
837 : : * this page's queue, we can perform writeback even if that
838 : : * will block.
839 : : *
840 : : * If the page is swapcache, write it back even if that would
841 : : * block, for some throttling. This happens by accident, because
842 : : * swap_backing_dev_info is bust: it doesn't reflect the
843 : : * congestion state of the swapdevs. Easy to fix, if needed.
844 : : */
845 : 0 : if (!is_page_cache_freeable(page))
846 : : return PAGE_KEEP;
847 : 0 : if (!mapping) {
848 : : /*
849 : : * Some data journaling orphaned pages can have
850 : : * page->mapping == NULL while being dirty with clean buffers.
851 : : */
852 : 0 : if (page_has_private(page)) {
853 : 0 : if (try_to_free_buffers(page)) {
854 : : ClearPageDirty(page);
855 : 0 : pr_info("%s: orphaned page\n", __func__);
856 : 0 : return PAGE_CLEAN;
857 : : }
858 : : }
859 : : return PAGE_KEEP;
860 : : }
861 : 0 : if (mapping->a_ops->writepage == NULL)
862 : : return PAGE_ACTIVATE;
863 : 0 : if (!may_write_to_inode(mapping->host, sc))
864 : : return PAGE_KEEP;
865 : :
866 : 0 : if (clear_page_dirty_for_io(page)) {
867 : : int res;
868 : 0 : struct writeback_control wbc = {
869 : : .sync_mode = WB_SYNC_NONE,
870 : : .nr_to_write = SWAP_CLUSTER_MAX,
871 : : .range_start = 0,
872 : : .range_end = LLONG_MAX,
873 : : .for_reclaim = 1,
874 : : };
875 : :
876 : : SetPageReclaim(page);
877 : 0 : res = mapping->a_ops->writepage(page, &wbc);
878 : 0 : if (res < 0)
879 : 0 : handle_write_error(mapping, page, res);
880 : 0 : if (res == AOP_WRITEPAGE_ACTIVATE) {
881 : : ClearPageReclaim(page);
882 : 0 : return PAGE_ACTIVATE;
883 : : }
884 : :
885 : 0 : if (!PageWriteback(page)) {
886 : : /* synchronous write or broken a_ops? */
887 : : ClearPageReclaim(page);
888 : : }
889 : 0 : trace_mm_vmscan_writepage(page);
890 : 0 : inc_node_page_state(page, NR_VMSCAN_WRITE);
891 : 0 : return PAGE_SUCCESS;
892 : : }
893 : :
894 : : return PAGE_CLEAN;
895 : : }
896 : :
897 : : /*
898 : : * Same as remove_mapping, but if the page is removed from the mapping, it
899 : : * gets returned with a refcount of 0.
900 : : */
901 : 0 : static int __remove_mapping(struct address_space *mapping, struct page *page,
902 : : bool reclaimed)
903 : : {
904 : : unsigned long flags;
905 : : int refcount;
906 : :
907 : 0 : BUG_ON(!PageLocked(page));
908 : 0 : BUG_ON(mapping != page_mapping(page));
909 : :
910 : 0 : xa_lock_irqsave(&mapping->i_pages, flags);
911 : : /*
912 : : * The non racy check for a busy page.
913 : : *
914 : : * Must be careful with the order of the tests. When someone has
915 : : * a ref to the page, it may be possible that they dirty it then
916 : : * drop the reference. So if PageDirty is tested before page_count
917 : : * here, then the following race may occur:
918 : : *
919 : : * get_user_pages(&page);
920 : : * [user mapping goes away]
921 : : * write_to(page);
922 : : * !PageDirty(page) [good]
923 : : * SetPageDirty(page);
924 : : * put_page(page);
925 : : * !page_count(page) [good, discard it]
926 : : *
927 : : * [oops, our write_to data is lost]
928 : : *
929 : : * Reversing the order of the tests ensures such a situation cannot
930 : : * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 : : * load is not satisfied before that of page->_refcount.
932 : : *
933 : : * Note that if SetPageDirty is always performed via set_page_dirty,
934 : : * and thus under the i_pages lock, then this ordering is not required.
935 : : */
936 : 0 : refcount = 1 + compound_nr(page);
937 : 0 : if (!page_ref_freeze(page, refcount))
938 : : goto cannot_free;
939 : : /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 : 0 : if (unlikely(PageDirty(page))) {
941 : : page_ref_unfreeze(page, refcount);
942 : : goto cannot_free;
943 : : }
944 : :
945 : 0 : if (PageSwapCache(page)) {
946 : 0 : swp_entry_t swap = { .val = page_private(page) };
947 : : mem_cgroup_swapout(page, swap);
948 : 0 : __delete_from_swap_cache(page, swap);
949 : : xa_unlock_irqrestore(&mapping->i_pages, flags);
950 : 0 : put_swap_page(page, swap);
951 : : } else {
952 : : void (*freepage)(struct page *);
953 : : void *shadow = NULL;
954 : :
955 : 0 : freepage = mapping->a_ops->freepage;
956 : : /*
957 : : * Remember a shadow entry for reclaimed file cache in
958 : : * order to detect refaults, thus thrashing, later on.
959 : : *
960 : : * But don't store shadows in an address space that is
961 : : * already exiting. This is not just an optizimation,
962 : : * inode reclaim needs to empty out the radix tree or
963 : : * the nodes are lost. Don't plant shadows behind its
964 : : * back.
965 : : *
966 : : * We also don't store shadows for DAX mappings because the
967 : : * only page cache pages found in these are zero pages
968 : : * covering holes, and because we don't want to mix DAX
969 : : * exceptional entries and shadow exceptional entries in the
970 : : * same address_space.
971 : : */
972 : 0 : if (reclaimed && page_is_file_cache(page) &&
973 : : !mapping_exiting(mapping) && !dax_mapping(mapping))
974 : 0 : shadow = workingset_eviction(page);
975 : 0 : __delete_from_page_cache(page, shadow);
976 : : xa_unlock_irqrestore(&mapping->i_pages, flags);
977 : :
978 : 0 : if (freepage != NULL)
979 : 0 : freepage(page);
980 : : }
981 : :
982 : : return 1;
983 : :
984 : : cannot_free:
985 : : xa_unlock_irqrestore(&mapping->i_pages, flags);
986 : 0 : return 0;
987 : : }
988 : :
989 : : /*
990 : : * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 : : * someone else has a ref on the page, abort and return 0. If it was
992 : : * successfully detached, return 1. Assumes the caller has a single ref on
993 : : * this page.
994 : : */
995 : 0 : int remove_mapping(struct address_space *mapping, struct page *page)
996 : : {
997 : 0 : if (__remove_mapping(mapping, page, false)) {
998 : : /*
999 : : * Unfreezing the refcount with 1 rather than 2 effectively
1000 : : * drops the pagecache ref for us without requiring another
1001 : : * atomic operation.
1002 : : */
1003 : : page_ref_unfreeze(page, 1);
1004 : 0 : return 1;
1005 : : }
1006 : : return 0;
1007 : : }
1008 : :
1009 : : /**
1010 : : * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 : : * @page: page to be put back to appropriate lru list
1012 : : *
1013 : : * Add previously isolated @page to appropriate LRU list.
1014 : : * Page may still be unevictable for other reasons.
1015 : : *
1016 : : * lru_lock must not be held, interrupts must be enabled.
1017 : : */
1018 : 3 : void putback_lru_page(struct page *page)
1019 : : {
1020 : 3 : lru_cache_add(page);
1021 : 3 : put_page(page); /* drop ref from isolate */
1022 : 3 : }
1023 : :
1024 : : enum page_references {
1025 : : PAGEREF_RECLAIM,
1026 : : PAGEREF_RECLAIM_CLEAN,
1027 : : PAGEREF_KEEP,
1028 : : PAGEREF_ACTIVATE,
1029 : : };
1030 : :
1031 : 0 : static enum page_references page_check_references(struct page *page,
1032 : : struct scan_control *sc)
1033 : : {
1034 : : int referenced_ptes, referenced_page;
1035 : : unsigned long vm_flags;
1036 : :
1037 : 0 : referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038 : : &vm_flags);
1039 : : referenced_page = TestClearPageReferenced(page);
1040 : :
1041 : : /*
1042 : : * Mlock lost the isolation race with us. Let try_to_unmap()
1043 : : * move the page to the unevictable list.
1044 : : */
1045 : 0 : if (vm_flags & VM_LOCKED)
1046 : : return PAGEREF_RECLAIM;
1047 : :
1048 : 0 : if (referenced_ptes) {
1049 : 0 : if (PageSwapBacked(page))
1050 : : return PAGEREF_ACTIVATE;
1051 : : /*
1052 : : * All mapped pages start out with page table
1053 : : * references from the instantiating fault, so we need
1054 : : * to look twice if a mapped file page is used more
1055 : : * than once.
1056 : : *
1057 : : * Mark it and spare it for another trip around the
1058 : : * inactive list. Another page table reference will
1059 : : * lead to its activation.
1060 : : *
1061 : : * Note: the mark is set for activated pages as well
1062 : : * so that recently deactivated but used pages are
1063 : : * quickly recovered.
1064 : : */
1065 : : SetPageReferenced(page);
1066 : :
1067 : 0 : if (referenced_page || referenced_ptes > 1)
1068 : : return PAGEREF_ACTIVATE;
1069 : :
1070 : : /*
1071 : : * Activate file-backed executable pages after first usage.
1072 : : */
1073 : 0 : if (vm_flags & VM_EXEC)
1074 : : return PAGEREF_ACTIVATE;
1075 : :
1076 : 0 : return PAGEREF_KEEP;
1077 : : }
1078 : :
1079 : : /* Reclaim if clean, defer dirty pages to writeback */
1080 : 0 : if (referenced_page && !PageSwapBacked(page))
1081 : : return PAGEREF_RECLAIM_CLEAN;
1082 : :
1083 : : return PAGEREF_RECLAIM;
1084 : : }
1085 : :
1086 : : /* Check if a page is dirty or under writeback */
1087 : 0 : static void page_check_dirty_writeback(struct page *page,
1088 : : bool *dirty, bool *writeback)
1089 : : {
1090 : : struct address_space *mapping;
1091 : :
1092 : : /*
1093 : : * Anonymous pages are not handled by flushers and must be written
1094 : : * from reclaim context. Do not stall reclaim based on them
1095 : : */
1096 : 0 : if (!page_is_file_cache(page) ||
1097 : 0 : (PageAnon(page) && !PageSwapBacked(page))) {
1098 : 0 : *dirty = false;
1099 : 0 : *writeback = false;
1100 : 0 : return;
1101 : : }
1102 : :
1103 : : /* By default assume that the page flags are accurate */
1104 : 0 : *dirty = PageDirty(page);
1105 : 0 : *writeback = PageWriteback(page);
1106 : :
1107 : : /* Verify dirty/writeback state if the filesystem supports it */
1108 : 0 : if (!page_has_private(page))
1109 : : return;
1110 : :
1111 : 0 : mapping = page_mapping(page);
1112 : 0 : if (mapping && mapping->a_ops->is_dirty_writeback)
1113 : 0 : mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1114 : : }
1115 : :
1116 : : /*
1117 : : * shrink_page_list() returns the number of reclaimed pages
1118 : : */
1119 : 3 : static unsigned long shrink_page_list(struct list_head *page_list,
1120 : : struct pglist_data *pgdat,
1121 : : struct scan_control *sc,
1122 : : enum ttu_flags ttu_flags,
1123 : : struct reclaim_stat *stat,
1124 : : bool ignore_references)
1125 : : {
1126 : 3 : LIST_HEAD(ret_pages);
1127 : 3 : LIST_HEAD(free_pages);
1128 : : unsigned nr_reclaimed = 0;
1129 : : unsigned pgactivate = 0;
1130 : :
1131 : 3 : memset(stat, 0, sizeof(*stat));
1132 : 3 : cond_resched();
1133 : :
1134 : 3 : while (!list_empty(page_list)) {
1135 : : struct address_space *mapping;
1136 : : struct page *page;
1137 : : int may_enter_fs;
1138 : : enum page_references references = PAGEREF_RECLAIM;
1139 : : bool dirty, writeback;
1140 : : unsigned int nr_pages;
1141 : :
1142 : 0 : cond_resched();
1143 : :
1144 : 0 : page = lru_to_page(page_list);
1145 : : list_del(&page->lru);
1146 : :
1147 : 0 : if (!trylock_page(page))
1148 : : goto keep;
1149 : :
1150 : : VM_BUG_ON_PAGE(PageActive(page), page);
1151 : :
1152 : : nr_pages = compound_nr(page);
1153 : :
1154 : : /* Account the number of base pages even though THP */
1155 : 0 : sc->nr_scanned += nr_pages;
1156 : :
1157 : 0 : if (unlikely(!page_evictable(page)))
1158 : : goto activate_locked;
1159 : :
1160 : 0 : if (!sc->may_unmap && page_mapped(page))
1161 : : goto keep_locked;
1162 : :
1163 : 0 : may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164 : 0 : (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165 : :
1166 : : /*
1167 : : * The number of dirty pages determines if a node is marked
1168 : : * reclaim_congested which affects wait_iff_congested. kswapd
1169 : : * will stall and start writing pages if the tail of the LRU
1170 : : * is all dirty unqueued pages.
1171 : : */
1172 : 0 : page_check_dirty_writeback(page, &dirty, &writeback);
1173 : 0 : if (dirty || writeback)
1174 : 0 : stat->nr_dirty++;
1175 : :
1176 : 0 : if (dirty && !writeback)
1177 : 0 : stat->nr_unqueued_dirty++;
1178 : :
1179 : : /*
1180 : : * Treat this page as congested if the underlying BDI is or if
1181 : : * pages are cycling through the LRU so quickly that the
1182 : : * pages marked for immediate reclaim are making it to the
1183 : : * end of the LRU a second time.
1184 : : */
1185 : 0 : mapping = page_mapping(page);
1186 : 0 : if (((dirty || writeback) && mapping &&
1187 : 0 : inode_write_congested(mapping->host)) ||
1188 : 0 : (writeback && PageReclaim(page)))
1189 : 0 : stat->nr_congested++;
1190 : :
1191 : : /*
1192 : : * If a page at the tail of the LRU is under writeback, there
1193 : : * are three cases to consider.
1194 : : *
1195 : : * 1) If reclaim is encountering an excessive number of pages
1196 : : * under writeback and this page is both under writeback and
1197 : : * PageReclaim then it indicates that pages are being queued
1198 : : * for IO but are being recycled through the LRU before the
1199 : : * IO can complete. Waiting on the page itself risks an
1200 : : * indefinite stall if it is impossible to writeback the
1201 : : * page due to IO error or disconnected storage so instead
1202 : : * note that the LRU is being scanned too quickly and the
1203 : : * caller can stall after page list has been processed.
1204 : : *
1205 : : * 2) Global or new memcg reclaim encounters a page that is
1206 : : * not marked for immediate reclaim, or the caller does not
1207 : : * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 : : * not to fs). In this case mark the page for immediate
1209 : : * reclaim and continue scanning.
1210 : : *
1211 : : * Require may_enter_fs because we would wait on fs, which
1212 : : * may not have submitted IO yet. And the loop driver might
1213 : : * enter reclaim, and deadlock if it waits on a page for
1214 : : * which it is needed to do the write (loop masks off
1215 : : * __GFP_IO|__GFP_FS for this reason); but more thought
1216 : : * would probably show more reasons.
1217 : : *
1218 : : * 3) Legacy memcg encounters a page that is already marked
1219 : : * PageReclaim. memcg does not have any dirty pages
1220 : : * throttling so we could easily OOM just because too many
1221 : : * pages are in writeback and there is nothing else to
1222 : : * reclaim. Wait for the writeback to complete.
1223 : : *
1224 : : * In cases 1) and 2) we activate the pages to get them out of
1225 : : * the way while we continue scanning for clean pages on the
1226 : : * inactive list and refilling from the active list. The
1227 : : * observation here is that waiting for disk writes is more
1228 : : * expensive than potentially causing reloads down the line.
1229 : : * Since they're marked for immediate reclaim, they won't put
1230 : : * memory pressure on the cache working set any longer than it
1231 : : * takes to write them to disk.
1232 : : */
1233 : 0 : if (PageWriteback(page)) {
1234 : : /* Case 1 above */
1235 : 0 : if (current_is_kswapd() &&
1236 : 0 : PageReclaim(page) &&
1237 : : test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238 : 0 : stat->nr_immediate++;
1239 : 0 : goto activate_locked;
1240 : :
1241 : : /* Case 2 above */
1242 : 0 : } else if (sane_reclaim(sc) ||
1243 : 0 : !PageReclaim(page) || !may_enter_fs) {
1244 : : /*
1245 : : * This is slightly racy - end_page_writeback()
1246 : : * might have just cleared PageReclaim, then
1247 : : * setting PageReclaim here end up interpreted
1248 : : * as PageReadahead - but that does not matter
1249 : : * enough to care. What we do want is for this
1250 : : * page to have PageReclaim set next time memcg
1251 : : * reclaim reaches the tests above, so it will
1252 : : * then wait_on_page_writeback() to avoid OOM;
1253 : : * and it's also appropriate in global reclaim.
1254 : : */
1255 : : SetPageReclaim(page);
1256 : 0 : stat->nr_writeback++;
1257 : 0 : goto activate_locked;
1258 : :
1259 : : /* Case 3 above */
1260 : : } else {
1261 : 0 : unlock_page(page);
1262 : 0 : wait_on_page_writeback(page);
1263 : : /* then go back and try same page again */
1264 : 0 : list_add_tail(&page->lru, page_list);
1265 : 0 : continue;
1266 : : }
1267 : : }
1268 : :
1269 : 0 : if (!ignore_references)
1270 : 0 : references = page_check_references(page, sc);
1271 : :
1272 : 0 : switch (references) {
1273 : : case PAGEREF_ACTIVATE:
1274 : : goto activate_locked;
1275 : : case PAGEREF_KEEP:
1276 : 0 : stat->nr_ref_keep += nr_pages;
1277 : 0 : goto keep_locked;
1278 : : case PAGEREF_RECLAIM:
1279 : : case PAGEREF_RECLAIM_CLEAN:
1280 : : ; /* try to reclaim the page below */
1281 : : }
1282 : :
1283 : : /*
1284 : : * Anonymous process memory has backing store?
1285 : : * Try to allocate it some swap space here.
1286 : : * Lazyfree page could be freed directly
1287 : : */
1288 : 0 : if (PageAnon(page) && PageSwapBacked(page)) {
1289 : 0 : if (!PageSwapCache(page)) {
1290 : 0 : if (!(sc->gfp_mask & __GFP_IO))
1291 : : goto keep_locked;
1292 : : if (PageTransHuge(page)) {
1293 : : /* cannot split THP, skip it */
1294 : : if (!can_split_huge_page(page, NULL))
1295 : : goto activate_locked;
1296 : : /*
1297 : : * Split pages without a PMD map right
1298 : : * away. Chances are some or all of the
1299 : : * tail pages can be freed without IO.
1300 : : */
1301 : : if (!compound_mapcount(page) &&
1302 : : split_huge_page_to_list(page,
1303 : : page_list))
1304 : : goto activate_locked;
1305 : : }
1306 : 0 : if (!add_to_swap(page)) {
1307 : : if (!PageTransHuge(page))
1308 : : goto activate_locked_split;
1309 : : /* Fallback to swap normal pages */
1310 : : if (split_huge_page_to_list(page,
1311 : : page_list))
1312 : : goto activate_locked;
1313 : : #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 : : count_vm_event(THP_SWPOUT_FALLBACK);
1315 : : #endif
1316 : : if (!add_to_swap(page))
1317 : : goto activate_locked_split;
1318 : : }
1319 : :
1320 : : may_enter_fs = 1;
1321 : :
1322 : : /* Adding to swap updated mapping */
1323 : 0 : mapping = page_mapping(page);
1324 : : }
1325 : : } else if (unlikely(PageTransHuge(page))) {
1326 : : /* Split file THP */
1327 : : if (split_huge_page_to_list(page, page_list))
1328 : : goto keep_locked;
1329 : : }
1330 : :
1331 : : /*
1332 : : * THP may get split above, need minus tail pages and update
1333 : : * nr_pages to avoid accounting tail pages twice.
1334 : : *
1335 : : * The tail pages that are added into swap cache successfully
1336 : : * reach here.
1337 : : */
1338 : 0 : if ((nr_pages > 1) && !PageTransHuge(page)) {
1339 : 0 : sc->nr_scanned -= (nr_pages - 1);
1340 : : nr_pages = 1;
1341 : : }
1342 : :
1343 : : /*
1344 : : * The page is mapped into the page tables of one or more
1345 : : * processes. Try to unmap it here.
1346 : : */
1347 : 0 : if (page_mapped(page)) {
1348 : 0 : enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349 : :
1350 : : if (unlikely(PageTransHuge(page)))
1351 : : flags |= TTU_SPLIT_HUGE_PMD;
1352 : 0 : if (!try_to_unmap(page, flags)) {
1353 : 0 : stat->nr_unmap_fail += nr_pages;
1354 : 0 : goto activate_locked;
1355 : : }
1356 : : }
1357 : :
1358 : 0 : if (PageDirty(page)) {
1359 : : /*
1360 : : * Only kswapd can writeback filesystem pages
1361 : : * to avoid risk of stack overflow. But avoid
1362 : : * injecting inefficient single-page IO into
1363 : : * flusher writeback as much as possible: only
1364 : : * write pages when we've encountered many
1365 : : * dirty pages, and when we've already scanned
1366 : : * the rest of the LRU for clean pages and see
1367 : : * the same dirty pages again (PageReclaim).
1368 : : */
1369 : 0 : if (page_is_file_cache(page) &&
1370 : 0 : (!current_is_kswapd() || !PageReclaim(page) ||
1371 : : !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372 : : /*
1373 : : * Immediately reclaim when written back.
1374 : : * Similar in principal to deactivate_page()
1375 : : * except we already have the page isolated
1376 : : * and know it's dirty
1377 : : */
1378 : 0 : inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379 : : SetPageReclaim(page);
1380 : :
1381 : : goto activate_locked;
1382 : : }
1383 : :
1384 : 0 : if (references == PAGEREF_RECLAIM_CLEAN)
1385 : : goto keep_locked;
1386 : 0 : if (!may_enter_fs)
1387 : : goto keep_locked;
1388 : 0 : if (!sc->may_writepage)
1389 : : goto keep_locked;
1390 : :
1391 : : /*
1392 : : * Page is dirty. Flush the TLB if a writable entry
1393 : : * potentially exists to avoid CPU writes after IO
1394 : : * starts and then write it out here.
1395 : : */
1396 : : try_to_unmap_flush_dirty();
1397 : 0 : switch (pageout(page, mapping, sc)) {
1398 : : case PAGE_KEEP:
1399 : : goto keep_locked;
1400 : : case PAGE_ACTIVATE:
1401 : : goto activate_locked;
1402 : : case PAGE_SUCCESS:
1403 : 0 : if (PageWriteback(page))
1404 : : goto keep;
1405 : 0 : if (PageDirty(page))
1406 : : goto keep;
1407 : :
1408 : : /*
1409 : : * A synchronous write - probably a ramdisk. Go
1410 : : * ahead and try to reclaim the page.
1411 : : */
1412 : 0 : if (!trylock_page(page))
1413 : : goto keep;
1414 : 0 : if (PageDirty(page) || PageWriteback(page))
1415 : : goto keep_locked;
1416 : 0 : mapping = page_mapping(page);
1417 : : case PAGE_CLEAN:
1418 : : ; /* try to free the page below */
1419 : : }
1420 : : }
1421 : :
1422 : : /*
1423 : : * If the page has buffers, try to free the buffer mappings
1424 : : * associated with this page. If we succeed we try to free
1425 : : * the page as well.
1426 : : *
1427 : : * We do this even if the page is PageDirty().
1428 : : * try_to_release_page() does not perform I/O, but it is
1429 : : * possible for a page to have PageDirty set, but it is actually
1430 : : * clean (all its buffers are clean). This happens if the
1431 : : * buffers were written out directly, with submit_bh(). ext3
1432 : : * will do this, as well as the blockdev mapping.
1433 : : * try_to_release_page() will discover that cleanness and will
1434 : : * drop the buffers and mark the page clean - it can be freed.
1435 : : *
1436 : : * Rarely, pages can have buffers and no ->mapping. These are
1437 : : * the pages which were not successfully invalidated in
1438 : : * truncate_complete_page(). We try to drop those buffers here
1439 : : * and if that worked, and the page is no longer mapped into
1440 : : * process address space (page_count == 1) it can be freed.
1441 : : * Otherwise, leave the page on the LRU so it is swappable.
1442 : : */
1443 : 0 : if (page_has_private(page)) {
1444 : 0 : if (!try_to_release_page(page, sc->gfp_mask))
1445 : : goto activate_locked;
1446 : 0 : if (!mapping && page_count(page) == 1) {
1447 : 0 : unlock_page(page);
1448 : 0 : if (put_page_testzero(page))
1449 : : goto free_it;
1450 : : else {
1451 : : /*
1452 : : * rare race with speculative reference.
1453 : : * the speculative reference will free
1454 : : * this page shortly, so we may
1455 : : * increment nr_reclaimed here (and
1456 : : * leave it off the LRU).
1457 : : */
1458 : 0 : nr_reclaimed++;
1459 : 0 : continue;
1460 : : }
1461 : : }
1462 : : }
1463 : :
1464 : 0 : if (PageAnon(page) && !PageSwapBacked(page)) {
1465 : : /* follow __remove_mapping for reference */
1466 : 0 : if (!page_ref_freeze(page, 1))
1467 : : goto keep_locked;
1468 : 0 : if (PageDirty(page)) {
1469 : : page_ref_unfreeze(page, 1);
1470 : : goto keep_locked;
1471 : : }
1472 : :
1473 : : count_vm_event(PGLAZYFREED);
1474 : : count_memcg_page_event(page, PGLAZYFREED);
1475 : 0 : } else if (!mapping || !__remove_mapping(mapping, page, true))
1476 : : goto keep_locked;
1477 : :
1478 : 0 : unlock_page(page);
1479 : : free_it:
1480 : : /*
1481 : : * THP may get swapped out in a whole, need account
1482 : : * all base pages.
1483 : : */
1484 : 0 : nr_reclaimed += nr_pages;
1485 : :
1486 : : /*
1487 : : * Is there need to periodically free_page_list? It would
1488 : : * appear not as the counts should be low
1489 : : */
1490 : : if (unlikely(PageTransHuge(page)))
1491 : : (*get_compound_page_dtor(page))(page);
1492 : : else
1493 : 0 : list_add(&page->lru, &free_pages);
1494 : 0 : continue;
1495 : :
1496 : : activate_locked_split:
1497 : : /*
1498 : : * The tail pages that are failed to add into swap cache
1499 : : * reach here. Fixup nr_scanned and nr_pages.
1500 : : */
1501 : 0 : if (nr_pages > 1) {
1502 : 0 : sc->nr_scanned -= (nr_pages - 1);
1503 : : nr_pages = 1;
1504 : : }
1505 : : activate_locked:
1506 : : /* Not a candidate for swapping, so reclaim swap space. */
1507 : 0 : if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508 : : PageMlocked(page)))
1509 : 0 : try_to_free_swap(page);
1510 : : VM_BUG_ON_PAGE(PageActive(page), page);
1511 : 0 : if (!PageMlocked(page)) {
1512 : 0 : int type = page_is_file_cache(page);
1513 : : SetPageActive(page);
1514 : 0 : stat->nr_activate[type] += nr_pages;
1515 : : count_memcg_page_event(page, PGACTIVATE);
1516 : : }
1517 : : keep_locked:
1518 : 0 : unlock_page(page);
1519 : : keep:
1520 : 0 : list_add(&page->lru, &ret_pages);
1521 : : VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1522 : : }
1523 : :
1524 : 3 : pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525 : :
1526 : 3 : mem_cgroup_uncharge_list(&free_pages);
1527 : : try_to_unmap_flush();
1528 : 3 : free_unref_page_list(&free_pages);
1529 : :
1530 : : list_splice(&ret_pages, page_list);
1531 : : count_vm_events(PGACTIVATE, pgactivate);
1532 : :
1533 : 3 : return nr_reclaimed;
1534 : : }
1535 : :
1536 : 3 : unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537 : : struct list_head *page_list)
1538 : : {
1539 : 3 : struct scan_control sc = {
1540 : : .gfp_mask = GFP_KERNEL,
1541 : : .priority = DEF_PRIORITY,
1542 : : .may_unmap = 1,
1543 : : };
1544 : : struct reclaim_stat dummy_stat;
1545 : : unsigned long ret;
1546 : : struct page *page, *next;
1547 : 3 : LIST_HEAD(clean_pages);
1548 : :
1549 : 3 : list_for_each_entry_safe(page, next, page_list, lru) {
1550 : 0 : if (page_is_file_cache(page) && !PageDirty(page) &&
1551 : 0 : !__PageMovable(page) && !PageUnevictable(page)) {
1552 : : ClearPageActive(page);
1553 : : list_move(&page->lru, &clean_pages);
1554 : : }
1555 : : }
1556 : :
1557 : 3 : ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558 : : TTU_IGNORE_ACCESS, &dummy_stat, true);
1559 : : list_splice(&clean_pages, page_list);
1560 : 3 : mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1561 : 3 : return ret;
1562 : : }
1563 : :
1564 : : /*
1565 : : * Attempt to remove the specified page from its LRU. Only take this page
1566 : : * if it is of the appropriate PageActive status. Pages which are being
1567 : : * freed elsewhere are also ignored.
1568 : : *
1569 : : * page: page to consider
1570 : : * mode: one of the LRU isolation modes defined above
1571 : : *
1572 : : * returns 0 on success, -ve errno on failure.
1573 : : */
1574 : 0 : int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1575 : : {
1576 : : int ret = -EINVAL;
1577 : :
1578 : : /* Only take pages on the LRU. */
1579 : 0 : if (!PageLRU(page))
1580 : : return ret;
1581 : :
1582 : : /* Compaction should not handle unevictable pages but CMA can do so */
1583 : 0 : if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1584 : : return ret;
1585 : :
1586 : : ret = -EBUSY;
1587 : :
1588 : : /*
1589 : : * To minimise LRU disruption, the caller can indicate that it only
1590 : : * wants to isolate pages it will be able to operate on without
1591 : : * blocking - clean pages for the most part.
1592 : : *
1593 : : * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 : : * that it is possible to migrate without blocking
1595 : : */
1596 : 0 : if (mode & ISOLATE_ASYNC_MIGRATE) {
1597 : : /* All the caller can do on PageWriteback is block */
1598 : 0 : if (PageWriteback(page))
1599 : : return ret;
1600 : :
1601 : 0 : if (PageDirty(page)) {
1602 : : struct address_space *mapping;
1603 : : bool migrate_dirty;
1604 : :
1605 : : /*
1606 : : * Only pages without mappings or that have a
1607 : : * ->migratepage callback are possible to migrate
1608 : : * without blocking. However, we can be racing with
1609 : : * truncation so it's necessary to lock the page
1610 : : * to stabilise the mapping as truncation holds
1611 : : * the page lock until after the page is removed
1612 : : * from the page cache.
1613 : : */
1614 : 0 : if (!trylock_page(page))
1615 : : return ret;
1616 : :
1617 : 0 : mapping = page_mapping(page);
1618 : 0 : migrate_dirty = !mapping || mapping->a_ops->migratepage;
1619 : 0 : unlock_page(page);
1620 : 0 : if (!migrate_dirty)
1621 : : return ret;
1622 : : }
1623 : : }
1624 : :
1625 : 0 : if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1626 : : return ret;
1627 : :
1628 : 0 : if (likely(get_page_unless_zero(page))) {
1629 : : /*
1630 : : * Be careful not to clear PageLRU until after we're
1631 : : * sure the page is not being freed elsewhere -- the
1632 : : * page release code relies on it.
1633 : : */
1634 : : ClearPageLRU(page);
1635 : : ret = 0;
1636 : : }
1637 : :
1638 : 0 : return ret;
1639 : : }
1640 : :
1641 : :
1642 : : /*
1643 : : * Update LRU sizes after isolating pages. The LRU size updates must
1644 : : * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 : : */
1646 : : static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647 : : enum lru_list lru, unsigned long *nr_zone_taken)
1648 : : {
1649 : : int zid;
1650 : :
1651 : 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652 : 0 : if (!nr_zone_taken[zid])
1653 : 0 : continue;
1654 : :
1655 : 0 : __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656 : : #ifdef CONFIG_MEMCG
1657 : 0 : mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1658 : : #endif
1659 : : }
1660 : :
1661 : : }
1662 : :
1663 : : /**
1664 : : * pgdat->lru_lock is heavily contended. Some of the functions that
1665 : : * shrink the lists perform better by taking out a batch of pages
1666 : : * and working on them outside the LRU lock.
1667 : : *
1668 : : * For pagecache intensive workloads, this function is the hottest
1669 : : * spot in the kernel (apart from copy_*_user functions).
1670 : : *
1671 : : * Appropriate locks must be held before calling this function.
1672 : : *
1673 : : * @nr_to_scan: The number of eligible pages to look through on the list.
1674 : : * @lruvec: The LRU vector to pull pages from.
1675 : : * @dst: The temp list to put pages on to.
1676 : : * @nr_scanned: The number of pages that were scanned.
1677 : : * @sc: The scan_control struct for this reclaim session
1678 : : * @mode: One of the LRU isolation modes
1679 : : * @lru: LRU list id for isolating
1680 : : *
1681 : : * returns how many pages were moved onto *@dst.
1682 : : */
1683 : 0 : static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684 : : struct lruvec *lruvec, struct list_head *dst,
1685 : : unsigned long *nr_scanned, struct scan_control *sc,
1686 : : enum lru_list lru)
1687 : : {
1688 : 0 : struct list_head *src = &lruvec->lists[lru];
1689 : : unsigned long nr_taken = 0;
1690 : 0 : unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691 : 0 : unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692 : : unsigned long skipped = 0;
1693 : : unsigned long scan, total_scan, nr_pages;
1694 : 0 : LIST_HEAD(pages_skipped);
1695 : 0 : isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1696 : :
1697 : : total_scan = 0;
1698 : : scan = 0;
1699 : 0 : while (scan < nr_to_scan && !list_empty(src)) {
1700 : : struct page *page;
1701 : :
1702 : 0 : page = lru_to_page(src);
1703 : 0 : prefetchw_prev_lru_page(page, src, flags);
1704 : :
1705 : : VM_BUG_ON_PAGE(!PageLRU(page), page);
1706 : :
1707 : : nr_pages = compound_nr(page);
1708 : 0 : total_scan += nr_pages;
1709 : :
1710 : 0 : if (page_zonenum(page) > sc->reclaim_idx) {
1711 : 0 : list_move(&page->lru, &pages_skipped);
1712 : 0 : nr_skipped[page_zonenum(page)] += nr_pages;
1713 : 0 : continue;
1714 : : }
1715 : :
1716 : : /*
1717 : : * Do not count skipped pages because that makes the function
1718 : : * return with no isolated pages if the LRU mostly contains
1719 : : * ineligible pages. This causes the VM to not reclaim any
1720 : : * pages, triggering a premature OOM.
1721 : : *
1722 : : * Account all tail pages of THP. This would not cause
1723 : : * premature OOM since __isolate_lru_page() returns -EBUSY
1724 : : * only when the page is being freed somewhere else.
1725 : : */
1726 : 0 : scan += nr_pages;
1727 : 0 : switch (__isolate_lru_page(page, mode)) {
1728 : : case 0:
1729 : 0 : nr_taken += nr_pages;
1730 : 0 : nr_zone_taken[page_zonenum(page)] += nr_pages;
1731 : 0 : list_move(&page->lru, dst);
1732 : : break;
1733 : :
1734 : : case -EBUSY:
1735 : : /* else it is being freed elsewhere */
1736 : 0 : list_move(&page->lru, src);
1737 : 0 : continue;
1738 : :
1739 : : default:
1740 : 0 : BUG();
1741 : : }
1742 : : }
1743 : :
1744 : : /*
1745 : : * Splice any skipped pages to the start of the LRU list. Note that
1746 : : * this disrupts the LRU order when reclaiming for lower zones but
1747 : : * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 : : * scanning would soon rescan the same pages to skip and put the
1749 : : * system at risk of premature OOM.
1750 : : */
1751 : 0 : if (!list_empty(&pages_skipped)) {
1752 : : int zid;
1753 : :
1754 : : list_splice(&pages_skipped, src);
1755 : 0 : for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756 : 0 : if (!nr_skipped[zid])
1757 : 0 : continue;
1758 : :
1759 : 0 : __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760 : 0 : skipped += nr_skipped[zid];
1761 : : }
1762 : : }
1763 : 0 : *nr_scanned = total_scan;
1764 : 0 : trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765 : : total_scan, skipped, nr_taken, mode, lru);
1766 : : update_lru_sizes(lruvec, lru, nr_zone_taken);
1767 : 0 : return nr_taken;
1768 : : }
1769 : :
1770 : : /**
1771 : : * isolate_lru_page - tries to isolate a page from its LRU list
1772 : : * @page: page to isolate from its LRU list
1773 : : *
1774 : : * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 : : * vmstat statistic corresponding to whatever LRU list the page was on.
1776 : : *
1777 : : * Returns 0 if the page was removed from an LRU list.
1778 : : * Returns -EBUSY if the page was not on an LRU list.
1779 : : *
1780 : : * The returned page will have PageLRU() cleared. If it was found on
1781 : : * the active list, it will have PageActive set. If it was found on
1782 : : * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 : : * may need to be cleared by the caller before letting the page go.
1784 : : *
1785 : : * The vmstat statistic corresponding to the list on which the page was
1786 : : * found will be decremented.
1787 : : *
1788 : : * Restrictions:
1789 : : *
1790 : : * (1) Must be called with an elevated refcount on the page. This is a
1791 : : * fundamentnal difference from isolate_lru_pages (which is called
1792 : : * without a stable reference).
1793 : : * (2) the lru_lock must not be held.
1794 : : * (3) interrupts must be enabled.
1795 : : */
1796 : 3 : int isolate_lru_page(struct page *page)
1797 : : {
1798 : : int ret = -EBUSY;
1799 : :
1800 : : VM_BUG_ON_PAGE(!page_count(page), page);
1801 : 3 : WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802 : :
1803 : 3 : if (PageLRU(page)) {
1804 : : pg_data_t *pgdat = page_pgdat(page);
1805 : : struct lruvec *lruvec;
1806 : :
1807 : : spin_lock_irq(&pgdat->lru_lock);
1808 : 3 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809 : 3 : if (PageLRU(page)) {
1810 : : int lru = page_lru(page);
1811 : 3 : get_page(page);
1812 : : ClearPageLRU(page);
1813 : : del_page_from_lru_list(page, lruvec, lru);
1814 : : ret = 0;
1815 : : }
1816 : : spin_unlock_irq(&pgdat->lru_lock);
1817 : : }
1818 : 3 : return ret;
1819 : : }
1820 : :
1821 : : /*
1822 : : * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 : : * then get resheduled. When there are massive number of tasks doing page
1824 : : * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 : : * the LRU list will go small and be scanned faster than necessary, leading to
1826 : : * unnecessary swapping, thrashing and OOM.
1827 : : */
1828 : 0 : static int too_many_isolated(struct pglist_data *pgdat, int file,
1829 : : struct scan_control *sc)
1830 : : {
1831 : : unsigned long inactive, isolated;
1832 : :
1833 : 0 : if (current_is_kswapd())
1834 : : return 0;
1835 : :
1836 : 0 : if (!sane_reclaim(sc))
1837 : : return 0;
1838 : :
1839 : 0 : if (file) {
1840 : : inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841 : : isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842 : : } else {
1843 : : inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844 : : isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1845 : : }
1846 : :
1847 : : /*
1848 : : * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 : : * won't get blocked by normal direct-reclaimers, forming a circular
1850 : : * deadlock.
1851 : : */
1852 : 0 : if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1853 : 0 : inactive >>= 3;
1854 : :
1855 : 0 : return isolated > inactive;
1856 : : }
1857 : :
1858 : : /*
1859 : : * This moves pages from @list to corresponding LRU list.
1860 : : *
1861 : : * We move them the other way if the page is referenced by one or more
1862 : : * processes, from rmap.
1863 : : *
1864 : : * If the pages are mostly unmapped, the processing is fast and it is
1865 : : * appropriate to hold zone_lru_lock across the whole operation. But if
1866 : : * the pages are mapped, the processing is slow (page_referenced()) so we
1867 : : * should drop zone_lru_lock around each page. It's impossible to balance
1868 : : * this, so instead we remove the pages from the LRU while processing them.
1869 : : * It is safe to rely on PG_active against the non-LRU pages in here because
1870 : : * nobody will play with that bit on a non-LRU page.
1871 : : *
1872 : : * The downside is that we have to touch page->_refcount against each page.
1873 : : * But we had to alter page->flags anyway.
1874 : : *
1875 : : * Returns the number of pages moved to the given lruvec.
1876 : : */
1877 : :
1878 : 0 : static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879 : : struct list_head *list)
1880 : : {
1881 : : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882 : : int nr_pages, nr_moved = 0;
1883 : 0 : LIST_HEAD(pages_to_free);
1884 : : struct page *page;
1885 : : enum lru_list lru;
1886 : :
1887 : 0 : while (!list_empty(list)) {
1888 : 0 : page = lru_to_page(list);
1889 : : VM_BUG_ON_PAGE(PageLRU(page), page);
1890 : 0 : if (unlikely(!page_evictable(page))) {
1891 : : list_del(&page->lru);
1892 : : spin_unlock_irq(&pgdat->lru_lock);
1893 : : putback_lru_page(page);
1894 : : spin_lock_irq(&pgdat->lru_lock);
1895 : 0 : continue;
1896 : : }
1897 : 0 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
1898 : :
1899 : : SetPageLRU(page);
1900 : : lru = page_lru(page);
1901 : :
1902 : : nr_pages = hpage_nr_pages(page);
1903 : : update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904 : 0 : list_move(&page->lru, &lruvec->lists[lru]);
1905 : :
1906 : 0 : if (put_page_testzero(page)) {
1907 : : __ClearPageLRU(page);
1908 : : __ClearPageActive(page);
1909 : : del_page_from_lru_list(page, lruvec, lru);
1910 : :
1911 : 0 : if (unlikely(PageCompound(page))) {
1912 : : spin_unlock_irq(&pgdat->lru_lock);
1913 : 0 : (*get_compound_page_dtor(page))(page);
1914 : : spin_lock_irq(&pgdat->lru_lock);
1915 : : } else
1916 : : list_add(&page->lru, &pages_to_free);
1917 : : } else {
1918 : 0 : nr_moved += nr_pages;
1919 : : }
1920 : : }
1921 : :
1922 : : /*
1923 : : * To save our caller's stack, now use input list for pages to free.
1924 : : */
1925 : : list_splice(&pages_to_free, list);
1926 : :
1927 : 0 : return nr_moved;
1928 : : }
1929 : :
1930 : : /*
1931 : : * If a kernel thread (such as nfsd for loop-back mounts) services
1932 : : * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 : : * In that case we should only throttle if the backing device it is
1934 : : * writing to is congested. In other cases it is safe to throttle.
1935 : : */
1936 : 0 : static int current_may_throttle(void)
1937 : : {
1938 : 0 : return !(current->flags & PF_LESS_THROTTLE) ||
1939 : 0 : current->backing_dev_info == NULL ||
1940 : : bdi_write_congested(current->backing_dev_info);
1941 : : }
1942 : :
1943 : : /*
1944 : : * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 : : * of reclaimed pages
1946 : : */
1947 : : static noinline_for_stack unsigned long
1948 : 0 : shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949 : : struct scan_control *sc, enum lru_list lru)
1950 : : {
1951 : 0 : LIST_HEAD(page_list);
1952 : : unsigned long nr_scanned;
1953 : : unsigned long nr_reclaimed = 0;
1954 : : unsigned long nr_taken;
1955 : : struct reclaim_stat stat;
1956 : : int file = is_file_lru(lru);
1957 : : enum vm_event_item item;
1958 : : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959 : : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960 : : bool stalled = false;
1961 : :
1962 : 0 : while (unlikely(too_many_isolated(pgdat, file, sc))) {
1963 : 0 : if (stalled)
1964 : : return 0;
1965 : :
1966 : : /* wait a bit for the reclaimer. */
1967 : 0 : msleep(100);
1968 : : stalled = true;
1969 : :
1970 : : /* We are about to die and free our memory. Return now. */
1971 : 0 : if (fatal_signal_pending(current))
1972 : : return SWAP_CLUSTER_MAX;
1973 : : }
1974 : :
1975 : 0 : lru_add_drain();
1976 : :
1977 : : spin_lock_irq(&pgdat->lru_lock);
1978 : :
1979 : 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980 : : &nr_scanned, sc, lru);
1981 : :
1982 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 : 0 : reclaim_stat->recent_scanned[file] += nr_taken;
1984 : :
1985 : 0 : item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986 : 0 : if (global_reclaim(sc))
1987 : 0 : __count_vm_events(item, nr_scanned);
1988 : 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1989 : : spin_unlock_irq(&pgdat->lru_lock);
1990 : :
1991 : 0 : if (nr_taken == 0)
1992 : : return 0;
1993 : :
1994 : 0 : nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1995 : : &stat, false);
1996 : :
1997 : : spin_lock_irq(&pgdat->lru_lock);
1998 : :
1999 : 0 : item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000 : 0 : if (global_reclaim(sc))
2001 : : __count_vm_events(item, nr_reclaimed);
2002 : 0 : __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003 : 0 : reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004 : 0 : reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2005 : :
2006 : 0 : move_pages_to_lru(lruvec, &page_list);
2007 : :
2008 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009 : :
2010 : : spin_unlock_irq(&pgdat->lru_lock);
2011 : :
2012 : 0 : mem_cgroup_uncharge_list(&page_list);
2013 : 0 : free_unref_page_list(&page_list);
2014 : :
2015 : : /*
2016 : : * If dirty pages are scanned that are not queued for IO, it
2017 : : * implies that flushers are not doing their job. This can
2018 : : * happen when memory pressure pushes dirty pages to the end of
2019 : : * the LRU before the dirty limits are breached and the dirty
2020 : : * data has expired. It can also happen when the proportion of
2021 : : * dirty pages grows not through writes but through memory
2022 : : * pressure reclaiming all the clean cache. And in some cases,
2023 : : * the flushers simply cannot keep up with the allocation
2024 : : * rate. Nudge the flusher threads in case they are asleep.
2025 : : */
2026 : 0 : if (stat.nr_unqueued_dirty == nr_taken)
2027 : 0 : wakeup_flusher_threads(WB_REASON_VMSCAN);
2028 : :
2029 : 0 : sc->nr.dirty += stat.nr_dirty;
2030 : 0 : sc->nr.congested += stat.nr_congested;
2031 : 0 : sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032 : 0 : sc->nr.writeback += stat.nr_writeback;
2033 : 0 : sc->nr.immediate += stat.nr_immediate;
2034 : 0 : sc->nr.taken += nr_taken;
2035 : 0 : if (file)
2036 : 0 : sc->nr.file_taken += nr_taken;
2037 : :
2038 : 0 : trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039 : 0 : nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040 : 0 : return nr_reclaimed;
2041 : : }
2042 : :
2043 : 0 : static void shrink_active_list(unsigned long nr_to_scan,
2044 : : struct lruvec *lruvec,
2045 : : struct scan_control *sc,
2046 : : enum lru_list lru)
2047 : : {
2048 : : unsigned long nr_taken;
2049 : : unsigned long nr_scanned;
2050 : : unsigned long vm_flags;
2051 : 0 : LIST_HEAD(l_hold); /* The pages which were snipped off */
2052 : 0 : LIST_HEAD(l_active);
2053 : 0 : LIST_HEAD(l_inactive);
2054 : : struct page *page;
2055 : : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056 : : unsigned nr_deactivate, nr_activate;
2057 : : unsigned nr_rotated = 0;
2058 : : int file = is_file_lru(lru);
2059 : : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2060 : :
2061 : 0 : lru_add_drain();
2062 : :
2063 : : spin_lock_irq(&pgdat->lru_lock);
2064 : :
2065 : 0 : nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066 : : &nr_scanned, sc, lru);
2067 : :
2068 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069 : 0 : reclaim_stat->recent_scanned[file] += nr_taken;
2070 : :
2071 : 0 : __count_vm_events(PGREFILL, nr_scanned);
2072 : 0 : __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073 : :
2074 : : spin_unlock_irq(&pgdat->lru_lock);
2075 : :
2076 : 0 : while (!list_empty(&l_hold)) {
2077 : 0 : cond_resched();
2078 : 0 : page = lru_to_page(&l_hold);
2079 : : list_del(&page->lru);
2080 : :
2081 : 0 : if (unlikely(!page_evictable(page))) {
2082 : : putback_lru_page(page);
2083 : 0 : continue;
2084 : : }
2085 : :
2086 : 0 : if (unlikely(buffer_heads_over_limit)) {
2087 : 0 : if (page_has_private(page) && trylock_page(page)) {
2088 : 0 : if (page_has_private(page))
2089 : 0 : try_to_release_page(page, 0);
2090 : 0 : unlock_page(page);
2091 : : }
2092 : : }
2093 : :
2094 : 0 : if (page_referenced(page, 0, sc->target_mem_cgroup,
2095 : : &vm_flags)) {
2096 : 0 : nr_rotated += hpage_nr_pages(page);
2097 : : /*
2098 : : * Identify referenced, file-backed active pages and
2099 : : * give them one more trip around the active list. So
2100 : : * that executable code get better chances to stay in
2101 : : * memory under moderate memory pressure. Anon pages
2102 : : * are not likely to be evicted by use-once streaming
2103 : : * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 : : * so we ignore them here.
2105 : : */
2106 : 0 : if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107 : 0 : list_add(&page->lru, &l_active);
2108 : 0 : continue;
2109 : : }
2110 : : }
2111 : :
2112 : : ClearPageActive(page); /* we are de-activating */
2113 : : SetPageWorkingset(page);
2114 : 0 : list_add(&page->lru, &l_inactive);
2115 : : }
2116 : :
2117 : : /*
2118 : : * Move pages back to the lru list.
2119 : : */
2120 : : spin_lock_irq(&pgdat->lru_lock);
2121 : : /*
2122 : : * Count referenced pages from currently used mappings as rotated,
2123 : : * even though only some of them are actually re-activated. This
2124 : : * helps balance scan pressure between file and anonymous pages in
2125 : : * get_scan_count.
2126 : : */
2127 : 0 : reclaim_stat->recent_rotated[file] += nr_rotated;
2128 : :
2129 : 0 : nr_activate = move_pages_to_lru(lruvec, &l_active);
2130 : 0 : nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131 : : /* Keep all free pages in l_active list */
2132 : : list_splice(&l_inactive, &l_active);
2133 : :
2134 : : __count_vm_events(PGDEACTIVATE, nr_deactivate);
2135 : 0 : __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136 : :
2137 : 0 : __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138 : : spin_unlock_irq(&pgdat->lru_lock);
2139 : :
2140 : 0 : mem_cgroup_uncharge_list(&l_active);
2141 : 0 : free_unref_page_list(&l_active);
2142 : 0 : trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143 : 0 : nr_deactivate, nr_rotated, sc->priority, file);
2144 : 0 : }
2145 : :
2146 : 0 : unsigned long reclaim_pages(struct list_head *page_list)
2147 : : {
2148 : : int nid = -1;
2149 : : unsigned long nr_reclaimed = 0;
2150 : 0 : LIST_HEAD(node_page_list);
2151 : : struct reclaim_stat dummy_stat;
2152 : : struct page *page;
2153 : 0 : struct scan_control sc = {
2154 : : .gfp_mask = GFP_KERNEL,
2155 : : .priority = DEF_PRIORITY,
2156 : : .may_writepage = 1,
2157 : : .may_unmap = 1,
2158 : : .may_swap = 1,
2159 : : };
2160 : :
2161 : 0 : while (!list_empty(page_list)) {
2162 : 0 : page = lru_to_page(page_list);
2163 : 0 : if (nid == -1) {
2164 : : nid = page_to_nid(page);
2165 : : INIT_LIST_HEAD(&node_page_list);
2166 : : }
2167 : :
2168 : 0 : if (nid == page_to_nid(page)) {
2169 : : ClearPageActive(page);
2170 : 0 : list_move(&page->lru, &node_page_list);
2171 : 0 : continue;
2172 : : }
2173 : :
2174 : 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2175 : : NODE_DATA(nid),
2176 : : &sc, 0,
2177 : : &dummy_stat, false);
2178 : 0 : while (!list_empty(&node_page_list)) {
2179 : 0 : page = lru_to_page(&node_page_list);
2180 : : list_del(&page->lru);
2181 : : putback_lru_page(page);
2182 : : }
2183 : :
2184 : : nid = -1;
2185 : : }
2186 : :
2187 : 0 : if (!list_empty(&node_page_list)) {
2188 : 0 : nr_reclaimed += shrink_page_list(&node_page_list,
2189 : : NODE_DATA(nid),
2190 : : &sc, 0,
2191 : : &dummy_stat, false);
2192 : 0 : while (!list_empty(&node_page_list)) {
2193 : 0 : page = lru_to_page(&node_page_list);
2194 : : list_del(&page->lru);
2195 : : putback_lru_page(page);
2196 : : }
2197 : : }
2198 : :
2199 : 0 : return nr_reclaimed;
2200 : : }
2201 : :
2202 : : /*
2203 : : * The inactive anon list should be small enough that the VM never has
2204 : : * to do too much work.
2205 : : *
2206 : : * The inactive file list should be small enough to leave most memory
2207 : : * to the established workingset on the scan-resistant active list,
2208 : : * but large enough to avoid thrashing the aggregate readahead window.
2209 : : *
2210 : : * Both inactive lists should also be large enough that each inactive
2211 : : * page has a chance to be referenced again before it is reclaimed.
2212 : : *
2213 : : * If that fails and refaulting is observed, the inactive list grows.
2214 : : *
2215 : : * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 : : * on this LRU, maintained by the pageout code. An inactive_ratio
2217 : : * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2218 : : *
2219 : : * total target max
2220 : : * memory ratio inactive
2221 : : * -------------------------------------
2222 : : * 10MB 1 5MB
2223 : : * 100MB 1 50MB
2224 : : * 1GB 3 250MB
2225 : : * 10GB 10 0.9GB
2226 : : * 100GB 31 3GB
2227 : : * 1TB 101 10GB
2228 : : * 10TB 320 32GB
2229 : : */
2230 : 0 : static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2231 : : struct scan_control *sc, bool trace)
2232 : : {
2233 : 0 : enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2234 : : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2235 : 0 : enum lru_list inactive_lru = file * LRU_FILE;
2236 : : unsigned long inactive, active;
2237 : : unsigned long inactive_ratio;
2238 : : unsigned long refaults;
2239 : : unsigned long gb;
2240 : :
2241 : : /*
2242 : : * If we don't have swap space, anonymous page deactivation
2243 : : * is pointless.
2244 : : */
2245 : 0 : if (!file && !total_swap_pages)
2246 : : return false;
2247 : :
2248 : 0 : inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2249 : 0 : active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2250 : :
2251 : : /*
2252 : : * When refaults are being observed, it means a new workingset
2253 : : * is being established. Disable active list protection to get
2254 : : * rid of the stale workingset quickly.
2255 : : */
2256 : 0 : refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2257 : 0 : if (file && lruvec->refaults != refaults) {
2258 : : inactive_ratio = 0;
2259 : : } else {
2260 : 0 : gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261 : 0 : if (gb)
2262 : 0 : inactive_ratio = int_sqrt(10 * gb);
2263 : : else
2264 : : inactive_ratio = 1;
2265 : : }
2266 : :
2267 : 0 : if (trace)
2268 : 0 : trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2269 : : lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2270 : : lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2271 : : inactive_ratio, file);
2272 : :
2273 : 0 : return inactive * inactive_ratio < active;
2274 : : }
2275 : :
2276 : 0 : static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277 : : struct lruvec *lruvec, struct scan_control *sc)
2278 : : {
2279 : 0 : if (is_active_lru(lru)) {
2280 : 0 : if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2281 : 0 : shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282 : : return 0;
2283 : : }
2284 : :
2285 : 0 : return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2286 : : }
2287 : :
2288 : : enum scan_balance {
2289 : : SCAN_EQUAL,
2290 : : SCAN_FRACT,
2291 : : SCAN_ANON,
2292 : : SCAN_FILE,
2293 : : };
2294 : :
2295 : : /*
2296 : : * Determine how aggressively the anon and file LRU lists should be
2297 : : * scanned. The relative value of each set of LRU lists is determined
2298 : : * by looking at the fraction of the pages scanned we did rotate back
2299 : : * onto the active list instead of evict.
2300 : : *
2301 : : * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 : : * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 : : */
2304 : 0 : static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2305 : : struct scan_control *sc, unsigned long *nr,
2306 : : unsigned long *lru_pages)
2307 : : {
2308 : 0 : int swappiness = mem_cgroup_swappiness(memcg);
2309 : : struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2310 : : u64 fraction[2];
2311 : : u64 denominator = 0; /* gcc */
2312 : : struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2313 : : unsigned long anon_prio, file_prio;
2314 : : enum scan_balance scan_balance;
2315 : : unsigned long anon, file;
2316 : : unsigned long ap, fp;
2317 : : enum lru_list lru;
2318 : :
2319 : : /* If we have no swap space, do not bother scanning anon pages. */
2320 : 0 : if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2321 : : scan_balance = SCAN_FILE;
2322 : : goto out;
2323 : : }
2324 : :
2325 : : /*
2326 : : * Global reclaim will swap to prevent OOM even with no
2327 : : * swappiness, but memcg users want to use this knob to
2328 : : * disable swapping for individual groups completely when
2329 : : * using the memory controller's swap limit feature would be
2330 : : * too expensive.
2331 : : */
2332 : 0 : if (!global_reclaim(sc) && !swappiness) {
2333 : : scan_balance = SCAN_FILE;
2334 : : goto out;
2335 : : }
2336 : :
2337 : : /*
2338 : : * Do not apply any pressure balancing cleverness when the
2339 : : * system is close to OOM, scan both anon and file equally
2340 : : * (unless the swappiness setting disagrees with swapping).
2341 : : */
2342 : 0 : if (!sc->priority && swappiness) {
2343 : : scan_balance = SCAN_EQUAL;
2344 : : goto out;
2345 : : }
2346 : :
2347 : : /*
2348 : : * Prevent the reclaimer from falling into the cache trap: as
2349 : : * cache pages start out inactive, every cache fault will tip
2350 : : * the scan balance towards the file LRU. And as the file LRU
2351 : : * shrinks, so does the window for rotation from references.
2352 : : * This means we have a runaway feedback loop where a tiny
2353 : : * thrashing file LRU becomes infinitely more attractive than
2354 : : * anon pages. Try to detect this based on file LRU size.
2355 : : */
2356 : 0 : if (global_reclaim(sc)) {
2357 : : unsigned long pgdatfile;
2358 : : unsigned long pgdatfree;
2359 : : int z;
2360 : : unsigned long total_high_wmark = 0;
2361 : :
2362 : : pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2363 : 0 : pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2364 : : node_page_state(pgdat, NR_INACTIVE_FILE);
2365 : :
2366 : 0 : for (z = 0; z < MAX_NR_ZONES; z++) {
2367 : 0 : struct zone *zone = &pgdat->node_zones[z];
2368 : 0 : if (!managed_zone(zone))
2369 : 0 : continue;
2370 : :
2371 : 0 : total_high_wmark += high_wmark_pages(zone);
2372 : : }
2373 : :
2374 : 0 : if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2375 : : /*
2376 : : * Force SCAN_ANON if there are enough inactive
2377 : : * anonymous pages on the LRU in eligible zones.
2378 : : * Otherwise, the small LRU gets thrashed.
2379 : : */
2380 : 0 : if (!inactive_list_is_low(lruvec, false, sc, false) &&
2381 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2382 : 0 : >> sc->priority) {
2383 : : scan_balance = SCAN_ANON;
2384 : : goto out;
2385 : : }
2386 : : }
2387 : : }
2388 : :
2389 : : /*
2390 : : * If there is enough inactive page cache, i.e. if the size of the
2391 : : * inactive list is greater than that of the active list *and* the
2392 : : * inactive list actually has some pages to scan on this priority, we
2393 : : * do not reclaim anything from the anonymous working set right now.
2394 : : * Without the second condition we could end up never scanning an
2395 : : * lruvec even if it has plenty of old anonymous pages unless the
2396 : : * system is under heavy pressure.
2397 : : */
2398 : 0 : if (!inactive_list_is_low(lruvec, true, sc, false) &&
2399 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2400 : : scan_balance = SCAN_FILE;
2401 : : goto out;
2402 : : }
2403 : :
2404 : : scan_balance = SCAN_FRACT;
2405 : :
2406 : : /*
2407 : : * With swappiness at 100, anonymous and file have the same priority.
2408 : : * This scanning priority is essentially the inverse of IO cost.
2409 : : */
2410 : 0 : anon_prio = swappiness;
2411 : 0 : file_prio = 200 - anon_prio;
2412 : :
2413 : : /*
2414 : : * OK, so we have swap space and a fair amount of page cache
2415 : : * pages. We use the recently rotated / recently scanned
2416 : : * ratios to determine how valuable each cache is.
2417 : : *
2418 : : * Because workloads change over time (and to avoid overflow)
2419 : : * we keep these statistics as a floating average, which ends
2420 : : * up weighing recent references more than old ones.
2421 : : *
2422 : : * anon in [0], file in [1]
2423 : : */
2424 : :
2425 : 0 : anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2426 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2427 : 0 : file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2428 : 0 : lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2429 : :
2430 : : spin_lock_irq(&pgdat->lru_lock);
2431 : 0 : if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2432 : 0 : reclaim_stat->recent_scanned[0] /= 2;
2433 : 0 : reclaim_stat->recent_rotated[0] /= 2;
2434 : : }
2435 : :
2436 : 0 : if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2437 : 0 : reclaim_stat->recent_scanned[1] /= 2;
2438 : 0 : reclaim_stat->recent_rotated[1] /= 2;
2439 : : }
2440 : :
2441 : : /*
2442 : : * The amount of pressure on anon vs file pages is inversely
2443 : : * proportional to the fraction of recently scanned pages on
2444 : : * each list that were recently referenced and in active use.
2445 : : */
2446 : 0 : ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2447 : 0 : ap /= reclaim_stat->recent_rotated[0] + 1;
2448 : :
2449 : 0 : fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2450 : 0 : fp /= reclaim_stat->recent_rotated[1] + 1;
2451 : : spin_unlock_irq(&pgdat->lru_lock);
2452 : :
2453 : 0 : fraction[0] = ap;
2454 : 0 : fraction[1] = fp;
2455 : 0 : denominator = ap + fp + 1;
2456 : : out:
2457 : 0 : *lru_pages = 0;
2458 : 0 : for_each_evictable_lru(lru) {
2459 : : int file = is_file_lru(lru);
2460 : : unsigned long lruvec_size;
2461 : : unsigned long scan;
2462 : : unsigned long protection;
2463 : :
2464 : 0 : lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2465 : : protection = mem_cgroup_protection(memcg,
2466 : 0 : sc->memcg_low_reclaim);
2467 : :
2468 : 0 : if (protection) {
2469 : : /*
2470 : : * Scale a cgroup's reclaim pressure by proportioning
2471 : : * its current usage to its memory.low or memory.min
2472 : : * setting.
2473 : : *
2474 : : * This is important, as otherwise scanning aggression
2475 : : * becomes extremely binary -- from nothing as we
2476 : : * approach the memory protection threshold, to totally
2477 : : * nominal as we exceed it. This results in requiring
2478 : : * setting extremely liberal protection thresholds. It
2479 : : * also means we simply get no protection at all if we
2480 : : * set it too low, which is not ideal.
2481 : : *
2482 : : * If there is any protection in place, we reduce scan
2483 : : * pressure by how much of the total memory used is
2484 : : * within protection thresholds.
2485 : : *
2486 : : * There is one special case: in the first reclaim pass,
2487 : : * we skip over all groups that are within their low
2488 : : * protection. If that fails to reclaim enough pages to
2489 : : * satisfy the reclaim goal, we come back and override
2490 : : * the best-effort low protection. However, we still
2491 : : * ideally want to honor how well-behaved groups are in
2492 : : * that case instead of simply punishing them all
2493 : : * equally. As such, we reclaim them based on how much
2494 : : * memory they are using, reducing the scan pressure
2495 : : * again by how much of the total memory used is under
2496 : : * hard protection.
2497 : : */
2498 : 0 : unsigned long cgroup_size = mem_cgroup_size(memcg);
2499 : :
2500 : : /* Avoid TOCTOU with earlier protection check */
2501 : 0 : cgroup_size = max(cgroup_size, protection);
2502 : :
2503 : 0 : scan = lruvec_size - lruvec_size * protection /
2504 : : cgroup_size;
2505 : :
2506 : : /*
2507 : : * Minimally target SWAP_CLUSTER_MAX pages to keep
2508 : : * reclaim moving forwards, avoiding decremeting
2509 : : * sc->priority further than desirable.
2510 : : */
2511 : 0 : scan = max(scan, SWAP_CLUSTER_MAX);
2512 : : } else {
2513 : : scan = lruvec_size;
2514 : : }
2515 : :
2516 : 0 : scan >>= sc->priority;
2517 : :
2518 : : /*
2519 : : * If the cgroup's already been deleted, make sure to
2520 : : * scrape out the remaining cache.
2521 : : */
2522 : 0 : if (!scan && !mem_cgroup_online(memcg))
2523 : 0 : scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2524 : :
2525 : 0 : switch (scan_balance) {
2526 : : case SCAN_EQUAL:
2527 : : /* Scan lists relative to size */
2528 : : break;
2529 : : case SCAN_FRACT:
2530 : : /*
2531 : : * Scan types proportional to swappiness and
2532 : : * their relative recent reclaim efficiency.
2533 : : * Make sure we don't miss the last page on
2534 : : * the offlined memory cgroups because of a
2535 : : * round-off error.
2536 : : */
2537 : 0 : scan = mem_cgroup_online(memcg) ?
2538 : 0 : div64_u64(scan * fraction[file], denominator) :
2539 : 0 : DIV64_U64_ROUND_UP(scan * fraction[file],
2540 : : denominator);
2541 : 0 : break;
2542 : : case SCAN_FILE:
2543 : : case SCAN_ANON:
2544 : : /* Scan one type exclusively */
2545 : 0 : if ((scan_balance == SCAN_FILE) != file) {
2546 : : lruvec_size = 0;
2547 : : scan = 0;
2548 : : }
2549 : : break;
2550 : : default:
2551 : : /* Look ma, no brain */
2552 : 0 : BUG();
2553 : : }
2554 : :
2555 : 0 : *lru_pages += lruvec_size;
2556 : 0 : nr[lru] = scan;
2557 : : }
2558 : 0 : }
2559 : :
2560 : : /*
2561 : : * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2562 : : */
2563 : 0 : static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2564 : : struct scan_control *sc, unsigned long *lru_pages)
2565 : : {
2566 : : struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2567 : : unsigned long nr[NR_LRU_LISTS];
2568 : : unsigned long targets[NR_LRU_LISTS];
2569 : : unsigned long nr_to_scan;
2570 : : enum lru_list lru;
2571 : : unsigned long nr_reclaimed = 0;
2572 : 0 : unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2573 : : struct blk_plug plug;
2574 : : bool scan_adjusted;
2575 : :
2576 : 0 : get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2577 : :
2578 : : /* Record the original scan target for proportional adjustments later */
2579 : 0 : memcpy(targets, nr, sizeof(nr));
2580 : :
2581 : : /*
2582 : : * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2583 : : * event that can occur when there is little memory pressure e.g.
2584 : : * multiple streaming readers/writers. Hence, we do not abort scanning
2585 : : * when the requested number of pages are reclaimed when scanning at
2586 : : * DEF_PRIORITY on the assumption that the fact we are direct
2587 : : * reclaiming implies that kswapd is not keeping up and it is best to
2588 : : * do a batch of work at once. For memcg reclaim one check is made to
2589 : : * abort proportional reclaim if either the file or anon lru has already
2590 : : * dropped to zero at the first pass.
2591 : : */
2592 : 0 : scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2593 : 0 : sc->priority == DEF_PRIORITY);
2594 : :
2595 : 0 : blk_start_plug(&plug);
2596 : 0 : while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2597 : 0 : nr[LRU_INACTIVE_FILE]) {
2598 : : unsigned long nr_anon, nr_file, percentage;
2599 : : unsigned long nr_scanned;
2600 : :
2601 : 0 : for_each_evictable_lru(lru) {
2602 : 0 : if (nr[lru]) {
2603 : 0 : nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2604 : 0 : nr[lru] -= nr_to_scan;
2605 : :
2606 : 0 : nr_reclaimed += shrink_list(lru, nr_to_scan,
2607 : : lruvec, sc);
2608 : : }
2609 : : }
2610 : :
2611 : 0 : cond_resched();
2612 : :
2613 : 0 : if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2614 : 0 : continue;
2615 : :
2616 : : /*
2617 : : * For kswapd and memcg, reclaim at least the number of pages
2618 : : * requested. Ensure that the anon and file LRUs are scanned
2619 : : * proportionally what was requested by get_scan_count(). We
2620 : : * stop reclaiming one LRU and reduce the amount scanning
2621 : : * proportional to the original scan target.
2622 : : */
2623 : 0 : nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2624 : 0 : nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2625 : :
2626 : : /*
2627 : : * It's just vindictive to attack the larger once the smaller
2628 : : * has gone to zero. And given the way we stop scanning the
2629 : : * smaller below, this makes sure that we only make one nudge
2630 : : * towards proportionality once we've got nr_to_reclaim.
2631 : : */
2632 : 0 : if (!nr_file || !nr_anon)
2633 : : break;
2634 : :
2635 : 0 : if (nr_file > nr_anon) {
2636 : 0 : unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2637 : 0 : targets[LRU_ACTIVE_ANON] + 1;
2638 : : lru = LRU_BASE;
2639 : 0 : percentage = nr_anon * 100 / scan_target;
2640 : : } else {
2641 : 0 : unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2642 : 0 : targets[LRU_ACTIVE_FILE] + 1;
2643 : : lru = LRU_FILE;
2644 : 0 : percentage = nr_file * 100 / scan_target;
2645 : : }
2646 : :
2647 : : /* Stop scanning the smaller of the LRU */
2648 : 0 : nr[lru] = 0;
2649 : 0 : nr[lru + LRU_ACTIVE] = 0;
2650 : :
2651 : : /*
2652 : : * Recalculate the other LRU scan count based on its original
2653 : : * scan target and the percentage scanning already complete
2654 : : */
2655 : 0 : lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2656 : 0 : nr_scanned = targets[lru] - nr[lru];
2657 : 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2658 : 0 : nr[lru] -= min(nr[lru], nr_scanned);
2659 : :
2660 : 0 : lru += LRU_ACTIVE;
2661 : 0 : nr_scanned = targets[lru] - nr[lru];
2662 : 0 : nr[lru] = targets[lru] * (100 - percentage) / 100;
2663 : 0 : nr[lru] -= min(nr[lru], nr_scanned);
2664 : :
2665 : : scan_adjusted = true;
2666 : : }
2667 : 0 : blk_finish_plug(&plug);
2668 : 0 : sc->nr_reclaimed += nr_reclaimed;
2669 : :
2670 : : /*
2671 : : * Even if we did not try to evict anon pages at all, we want to
2672 : : * rebalance the anon lru active/inactive ratio.
2673 : : */
2674 : 0 : if (inactive_list_is_low(lruvec, false, sc, true))
2675 : 0 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2676 : : sc, LRU_ACTIVE_ANON);
2677 : 0 : }
2678 : :
2679 : : /* Use reclaim/compaction for costly allocs or under memory pressure */
2680 : : static bool in_reclaim_compaction(struct scan_control *sc)
2681 : : {
2682 : 0 : if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2683 : 0 : (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2684 : 0 : sc->priority < DEF_PRIORITY - 2))
2685 : : return true;
2686 : :
2687 : : return false;
2688 : : }
2689 : :
2690 : : /*
2691 : : * Reclaim/compaction is used for high-order allocation requests. It reclaims
2692 : : * order-0 pages before compacting the zone. should_continue_reclaim() returns
2693 : : * true if more pages should be reclaimed such that when the page allocator
2694 : : * calls try_to_compact_zone() that it will have enough free pages to succeed.
2695 : : * It will give up earlier than that if there is difficulty reclaiming pages.
2696 : : */
2697 : 0 : static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2698 : : unsigned long nr_reclaimed,
2699 : : struct scan_control *sc)
2700 : : {
2701 : : unsigned long pages_for_compaction;
2702 : : unsigned long inactive_lru_pages;
2703 : : int z;
2704 : :
2705 : : /* If not in reclaim/compaction mode, stop */
2706 : 0 : if (!in_reclaim_compaction(sc))
2707 : : return false;
2708 : :
2709 : : /*
2710 : : * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2711 : : * number of pages that were scanned. This will return to the caller
2712 : : * with the risk reclaim/compaction and the resulting allocation attempt
2713 : : * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2714 : : * allocations through requiring that the full LRU list has been scanned
2715 : : * first, by assuming that zero delta of sc->nr_scanned means full LRU
2716 : : * scan, but that approximation was wrong, and there were corner cases
2717 : : * where always a non-zero amount of pages were scanned.
2718 : : */
2719 : 0 : if (!nr_reclaimed)
2720 : : return false;
2721 : :
2722 : : /* If compaction would go ahead or the allocation would succeed, stop */
2723 : 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
2724 : 0 : struct zone *zone = &pgdat->node_zones[z];
2725 : 0 : if (!managed_zone(zone))
2726 : 0 : continue;
2727 : :
2728 : 0 : switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2729 : : case COMPACT_SUCCESS:
2730 : : case COMPACT_CONTINUE:
2731 : : return false;
2732 : : default:
2733 : : /* check next zone */
2734 : : ;
2735 : : }
2736 : : }
2737 : :
2738 : : /*
2739 : : * If we have not reclaimed enough pages for compaction and the
2740 : : * inactive lists are large enough, continue reclaiming
2741 : : */
2742 : 0 : pages_for_compaction = compact_gap(sc->order);
2743 : : inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2744 : 0 : if (get_nr_swap_pages() > 0)
2745 : 0 : inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2746 : :
2747 : 0 : return inactive_lru_pages > pages_for_compaction;
2748 : : }
2749 : :
2750 : : static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2751 : : {
2752 : 0 : return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2753 : 0 : (memcg && memcg_congested(pgdat, memcg));
2754 : : }
2755 : :
2756 : 0 : static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2757 : : {
2758 : 0 : struct reclaim_state *reclaim_state = current->reclaim_state;
2759 : : unsigned long nr_reclaimed, nr_scanned;
2760 : : bool reclaimable = false;
2761 : :
2762 : : do {
2763 : 0 : struct mem_cgroup *root = sc->target_mem_cgroup;
2764 : : unsigned long node_lru_pages = 0;
2765 : : struct mem_cgroup *memcg;
2766 : :
2767 : 0 : memset(&sc->nr, 0, sizeof(sc->nr));
2768 : :
2769 : 0 : nr_reclaimed = sc->nr_reclaimed;
2770 : 0 : nr_scanned = sc->nr_scanned;
2771 : :
2772 : 0 : memcg = mem_cgroup_iter(root, NULL, NULL);
2773 : : do {
2774 : : unsigned long lru_pages;
2775 : : unsigned long reclaimed;
2776 : : unsigned long scanned;
2777 : :
2778 : 0 : switch (mem_cgroup_protected(root, memcg)) {
2779 : : case MEMCG_PROT_MIN:
2780 : : /*
2781 : : * Hard protection.
2782 : : * If there is no reclaimable memory, OOM.
2783 : : */
2784 : 0 : continue;
2785 : : case MEMCG_PROT_LOW:
2786 : : /*
2787 : : * Soft protection.
2788 : : * Respect the protection only as long as
2789 : : * there is an unprotected supply
2790 : : * of reclaimable memory from other cgroups.
2791 : : */
2792 : 0 : if (!sc->memcg_low_reclaim) {
2793 : 0 : sc->memcg_low_skipped = 1;
2794 : 0 : continue;
2795 : : }
2796 : 0 : memcg_memory_event(memcg, MEMCG_LOW);
2797 : 0 : break;
2798 : : case MEMCG_PROT_NONE:
2799 : : /*
2800 : : * All protection thresholds breached. We may
2801 : : * still choose to vary the scan pressure
2802 : : * applied based on by how much the cgroup in
2803 : : * question has exceeded its protection
2804 : : * thresholds (see get_scan_count).
2805 : : */
2806 : : break;
2807 : : }
2808 : :
2809 : 0 : reclaimed = sc->nr_reclaimed;
2810 : 0 : scanned = sc->nr_scanned;
2811 : 0 : shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2812 : : node_lru_pages += lru_pages;
2813 : :
2814 : 0 : shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2815 : 0 : sc->priority);
2816 : :
2817 : : /* Record the group's reclaim efficiency */
2818 : 0 : vmpressure(sc->gfp_mask, memcg, false,
2819 : 0 : sc->nr_scanned - scanned,
2820 : 0 : sc->nr_reclaimed - reclaimed);
2821 : :
2822 : 0 : } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2823 : :
2824 : 0 : if (reclaim_state) {
2825 : 0 : sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2826 : 0 : reclaim_state->reclaimed_slab = 0;
2827 : : }
2828 : :
2829 : : /* Record the subtree's reclaim efficiency */
2830 : 0 : vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2831 : 0 : sc->nr_scanned - nr_scanned,
2832 : 0 : sc->nr_reclaimed - nr_reclaimed);
2833 : :
2834 : 0 : if (sc->nr_reclaimed - nr_reclaimed)
2835 : : reclaimable = true;
2836 : :
2837 : 0 : if (current_is_kswapd()) {
2838 : : /*
2839 : : * If reclaim is isolating dirty pages under writeback,
2840 : : * it implies that the long-lived page allocation rate
2841 : : * is exceeding the page laundering rate. Either the
2842 : : * global limits are not being effective at throttling
2843 : : * processes due to the page distribution throughout
2844 : : * zones or there is heavy usage of a slow backing
2845 : : * device. The only option is to throttle from reclaim
2846 : : * context which is not ideal as there is no guarantee
2847 : : * the dirtying process is throttled in the same way
2848 : : * balance_dirty_pages() manages.
2849 : : *
2850 : : * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2851 : : * count the number of pages under pages flagged for
2852 : : * immediate reclaim and stall if any are encountered
2853 : : * in the nr_immediate check below.
2854 : : */
2855 : 0 : if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2856 : 0 : set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2857 : :
2858 : : /*
2859 : : * Tag a node as congested if all the dirty pages
2860 : : * scanned were backed by a congested BDI and
2861 : : * wait_iff_congested will stall.
2862 : : */
2863 : 0 : if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2864 : 0 : set_bit(PGDAT_CONGESTED, &pgdat->flags);
2865 : :
2866 : : /* Allow kswapd to start writing pages during reclaim.*/
2867 : 0 : if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2868 : 0 : set_bit(PGDAT_DIRTY, &pgdat->flags);
2869 : :
2870 : : /*
2871 : : * If kswapd scans pages marked marked for immediate
2872 : : * reclaim and under writeback (nr_immediate), it
2873 : : * implies that pages are cycling through the LRU
2874 : : * faster than they are written so also forcibly stall.
2875 : : */
2876 : 0 : if (sc->nr.immediate)
2877 : 0 : congestion_wait(BLK_RW_ASYNC, HZ/10);
2878 : : }
2879 : :
2880 : : /*
2881 : : * Legacy memcg will stall in page writeback so avoid forcibly
2882 : : * stalling in wait_iff_congested().
2883 : : */
2884 : 0 : if (!global_reclaim(sc) && sane_reclaim(sc) &&
2885 : 0 : sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2886 : : set_memcg_congestion(pgdat, root, true);
2887 : :
2888 : : /*
2889 : : * Stall direct reclaim for IO completions if underlying BDIs
2890 : : * and node is congested. Allow kswapd to continue until it
2891 : : * starts encountering unqueued dirty pages or cycling through
2892 : : * the LRU too quickly.
2893 : : */
2894 : 0 : if (!sc->hibernation_mode && !current_is_kswapd() &&
2895 : 0 : current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2896 : 0 : wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2897 : :
2898 : 0 : } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2899 : 0 : sc));
2900 : :
2901 : : /*
2902 : : * Kswapd gives up on balancing particular nodes after too
2903 : : * many failures to reclaim anything from them and goes to
2904 : : * sleep. On reclaim progress, reset the failure counter. A
2905 : : * successful direct reclaim run will revive a dormant kswapd.
2906 : : */
2907 : 0 : if (reclaimable)
2908 : 0 : pgdat->kswapd_failures = 0;
2909 : :
2910 : 0 : return reclaimable;
2911 : : }
2912 : :
2913 : : /*
2914 : : * Returns true if compaction should go ahead for a costly-order request, or
2915 : : * the allocation would already succeed without compaction. Return false if we
2916 : : * should reclaim first.
2917 : : */
2918 : 0 : static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2919 : : {
2920 : : unsigned long watermark;
2921 : : enum compact_result suitable;
2922 : :
2923 : 0 : suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2924 : 0 : if (suitable == COMPACT_SUCCESS)
2925 : : /* Allocation should succeed already. Don't reclaim. */
2926 : : return true;
2927 : 0 : if (suitable == COMPACT_SKIPPED)
2928 : : /* Compaction cannot yet proceed. Do reclaim. */
2929 : : return false;
2930 : :
2931 : : /*
2932 : : * Compaction is already possible, but it takes time to run and there
2933 : : * are potentially other callers using the pages just freed. So proceed
2934 : : * with reclaim to make a buffer of free pages available to give
2935 : : * compaction a reasonable chance of completing and allocating the page.
2936 : : * Note that we won't actually reclaim the whole buffer in one attempt
2937 : : * as the target watermark in should_continue_reclaim() is lower. But if
2938 : : * we are already above the high+gap watermark, don't reclaim at all.
2939 : : */
2940 : 0 : watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2941 : :
2942 : 0 : return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2943 : : }
2944 : :
2945 : : /*
2946 : : * This is the direct reclaim path, for page-allocating processes. We only
2947 : : * try to reclaim pages from zones which will satisfy the caller's allocation
2948 : : * request.
2949 : : *
2950 : : * If a zone is deemed to be full of pinned pages then just give it a light
2951 : : * scan then give up on it.
2952 : : */
2953 : 0 : static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2954 : : {
2955 : : struct zoneref *z;
2956 : : struct zone *zone;
2957 : : unsigned long nr_soft_reclaimed;
2958 : : unsigned long nr_soft_scanned;
2959 : : gfp_t orig_mask;
2960 : : pg_data_t *last_pgdat = NULL;
2961 : :
2962 : : /*
2963 : : * If the number of buffer_heads in the machine exceeds the maximum
2964 : : * allowed level, force direct reclaim to scan the highmem zone as
2965 : : * highmem pages could be pinning lowmem pages storing buffer_heads
2966 : : */
2967 : 0 : orig_mask = sc->gfp_mask;
2968 : 0 : if (buffer_heads_over_limit) {
2969 : 0 : sc->gfp_mask |= __GFP_HIGHMEM;
2970 : 0 : sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2971 : : }
2972 : :
2973 : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
2974 : : sc->reclaim_idx, sc->nodemask) {
2975 : : /*
2976 : : * Take care memory controller reclaiming has small influence
2977 : : * to global LRU.
2978 : : */
2979 : 0 : if (global_reclaim(sc)) {
2980 : 0 : if (!cpuset_zone_allowed(zone,
2981 : : GFP_KERNEL | __GFP_HARDWALL))
2982 : 0 : continue;
2983 : :
2984 : : /*
2985 : : * If we already have plenty of memory free for
2986 : : * compaction in this zone, don't free any more.
2987 : : * Even though compaction is invoked for any
2988 : : * non-zero order, only frequent costly order
2989 : : * reclamation is disruptive enough to become a
2990 : : * noticeable problem, like transparent huge
2991 : : * page allocations.
2992 : : */
2993 : 0 : if (IS_ENABLED(CONFIG_COMPACTION) &&
2994 : 0 : sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2995 : 0 : compaction_ready(zone, sc)) {
2996 : 0 : sc->compaction_ready = true;
2997 : 0 : continue;
2998 : : }
2999 : :
3000 : : /*
3001 : : * Shrink each node in the zonelist once. If the
3002 : : * zonelist is ordered by zone (not the default) then a
3003 : : * node may be shrunk multiple times but in that case
3004 : : * the user prefers lower zones being preserved.
3005 : : */
3006 : 0 : if (zone->zone_pgdat == last_pgdat)
3007 : 0 : continue;
3008 : :
3009 : : /*
3010 : : * This steals pages from memory cgroups over softlimit
3011 : : * and returns the number of reclaimed pages and
3012 : : * scanned pages. This works for global memory pressure
3013 : : * and balancing, not for a memcg's limit.
3014 : : */
3015 : 0 : nr_soft_scanned = 0;
3016 : 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3017 : 0 : sc->order, sc->gfp_mask,
3018 : : &nr_soft_scanned);
3019 : 0 : sc->nr_reclaimed += nr_soft_reclaimed;
3020 : 0 : sc->nr_scanned += nr_soft_scanned;
3021 : : /* need some check for avoid more shrink_zone() */
3022 : : }
3023 : :
3024 : : /* See comment about same check for global reclaim above */
3025 : 0 : if (zone->zone_pgdat == last_pgdat)
3026 : 0 : continue;
3027 : : last_pgdat = zone->zone_pgdat;
3028 : 0 : shrink_node(zone->zone_pgdat, sc);
3029 : : }
3030 : :
3031 : : /*
3032 : : * Restore to original mask to avoid the impact on the caller if we
3033 : : * promoted it to __GFP_HIGHMEM.
3034 : : */
3035 : 0 : sc->gfp_mask = orig_mask;
3036 : 0 : }
3037 : :
3038 : 0 : static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3039 : : {
3040 : : struct mem_cgroup *memcg;
3041 : :
3042 : 0 : memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3043 : : do {
3044 : : unsigned long refaults;
3045 : : struct lruvec *lruvec;
3046 : :
3047 : : lruvec = mem_cgroup_lruvec(pgdat, memcg);
3048 : 0 : refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3049 : 0 : lruvec->refaults = refaults;
3050 : 0 : } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3051 : 0 : }
3052 : :
3053 : : /*
3054 : : * This is the main entry point to direct page reclaim.
3055 : : *
3056 : : * If a full scan of the inactive list fails to free enough memory then we
3057 : : * are "out of memory" and something needs to be killed.
3058 : : *
3059 : : * If the caller is !__GFP_FS then the probability of a failure is reasonably
3060 : : * high - the zone may be full of dirty or under-writeback pages, which this
3061 : : * caller can't do much about. We kick the writeback threads and take explicit
3062 : : * naps in the hope that some of these pages can be written. But if the
3063 : : * allocating task holds filesystem locks which prevent writeout this might not
3064 : : * work, and the allocation attempt will fail.
3065 : : *
3066 : : * returns: 0, if no pages reclaimed
3067 : : * else, the number of pages reclaimed
3068 : : */
3069 : 0 : static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3070 : : struct scan_control *sc)
3071 : : {
3072 : 0 : int initial_priority = sc->priority;
3073 : : pg_data_t *last_pgdat;
3074 : : struct zoneref *z;
3075 : : struct zone *zone;
3076 : : retry:
3077 : 0 : delayacct_freepages_start();
3078 : :
3079 : 0 : if (global_reclaim(sc))
3080 : 0 : __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3081 : :
3082 : : do {
3083 : 0 : vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3084 : 0 : sc->priority);
3085 : 0 : sc->nr_scanned = 0;
3086 : 0 : shrink_zones(zonelist, sc);
3087 : :
3088 : 0 : if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3089 : : break;
3090 : :
3091 : 0 : if (sc->compaction_ready)
3092 : : break;
3093 : :
3094 : : /*
3095 : : * If we're getting trouble reclaiming, start doing
3096 : : * writepage even in laptop mode.
3097 : : */
3098 : 0 : if (sc->priority < DEF_PRIORITY - 2)
3099 : 0 : sc->may_writepage = 1;
3100 : 0 : } while (--sc->priority >= 0);
3101 : :
3102 : : last_pgdat = NULL;
3103 : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3104 : : sc->nodemask) {
3105 : 0 : if (zone->zone_pgdat == last_pgdat)
3106 : 0 : continue;
3107 : : last_pgdat = zone->zone_pgdat;
3108 : 0 : snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3109 : 0 : set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3110 : : }
3111 : :
3112 : 0 : delayacct_freepages_end();
3113 : :
3114 : 0 : if (sc->nr_reclaimed)
3115 : 0 : return sc->nr_reclaimed;
3116 : :
3117 : : /* Aborted reclaim to try compaction? don't OOM, then */
3118 : 0 : if (sc->compaction_ready)
3119 : : return 1;
3120 : :
3121 : : /* Untapped cgroup reserves? Don't OOM, retry. */
3122 : 0 : if (sc->memcg_low_skipped) {
3123 : 0 : sc->priority = initial_priority;
3124 : 0 : sc->memcg_low_reclaim = 1;
3125 : 0 : sc->memcg_low_skipped = 0;
3126 : 0 : goto retry;
3127 : : }
3128 : :
3129 : : return 0;
3130 : : }
3131 : :
3132 : 0 : static bool allow_direct_reclaim(pg_data_t *pgdat)
3133 : : {
3134 : : struct zone *zone;
3135 : : unsigned long pfmemalloc_reserve = 0;
3136 : : unsigned long free_pages = 0;
3137 : : int i;
3138 : : bool wmark_ok;
3139 : :
3140 : 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3141 : : return true;
3142 : :
3143 : 0 : for (i = 0; i <= ZONE_NORMAL; i++) {
3144 : 0 : zone = &pgdat->node_zones[i];
3145 : 0 : if (!managed_zone(zone))
3146 : 0 : continue;
3147 : :
3148 : 0 : if (!zone_reclaimable_pages(zone))
3149 : 0 : continue;
3150 : :
3151 : 0 : pfmemalloc_reserve += min_wmark_pages(zone);
3152 : 0 : free_pages += zone_page_state(zone, NR_FREE_PAGES);
3153 : : }
3154 : :
3155 : : /* If there are no reserves (unexpected config) then do not throttle */
3156 : 0 : if (!pfmemalloc_reserve)
3157 : : return true;
3158 : :
3159 : 0 : wmark_ok = free_pages > pfmemalloc_reserve / 2;
3160 : :
3161 : : /* kswapd must be awake if processes are being throttled */
3162 : 0 : if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3163 : 0 : pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3164 : : (enum zone_type)ZONE_NORMAL);
3165 : 0 : wake_up_interruptible(&pgdat->kswapd_wait);
3166 : : }
3167 : :
3168 : 0 : return wmark_ok;
3169 : : }
3170 : :
3171 : : /*
3172 : : * Throttle direct reclaimers if backing storage is backed by the network
3173 : : * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3174 : : * depleted. kswapd will continue to make progress and wake the processes
3175 : : * when the low watermark is reached.
3176 : : *
3177 : : * Returns true if a fatal signal was delivered during throttling. If this
3178 : : * happens, the page allocator should not consider triggering the OOM killer.
3179 : : */
3180 : 0 : static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3181 : : nodemask_t *nodemask)
3182 : : {
3183 : : struct zoneref *z;
3184 : : struct zone *zone;
3185 : : pg_data_t *pgdat = NULL;
3186 : :
3187 : : /*
3188 : : * Kernel threads should not be throttled as they may be indirectly
3189 : : * responsible for cleaning pages necessary for reclaim to make forward
3190 : : * progress. kjournald for example may enter direct reclaim while
3191 : : * committing a transaction where throttling it could forcing other
3192 : : * processes to block on log_wait_commit().
3193 : : */
3194 : 0 : if (current->flags & PF_KTHREAD)
3195 : : goto out;
3196 : :
3197 : : /*
3198 : : * If a fatal signal is pending, this process should not throttle.
3199 : : * It should return quickly so it can exit and free its memory
3200 : : */
3201 : 0 : if (fatal_signal_pending(current))
3202 : : goto out;
3203 : :
3204 : : /*
3205 : : * Check if the pfmemalloc reserves are ok by finding the first node
3206 : : * with a usable ZONE_NORMAL or lower zone. The expectation is that
3207 : : * GFP_KERNEL will be required for allocating network buffers when
3208 : : * swapping over the network so ZONE_HIGHMEM is unusable.
3209 : : *
3210 : : * Throttling is based on the first usable node and throttled processes
3211 : : * wait on a queue until kswapd makes progress and wakes them. There
3212 : : * is an affinity then between processes waking up and where reclaim
3213 : : * progress has been made assuming the process wakes on the same node.
3214 : : * More importantly, processes running on remote nodes will not compete
3215 : : * for remote pfmemalloc reserves and processes on different nodes
3216 : : * should make reasonable progress.
3217 : : */
3218 : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist,
3219 : : gfp_zone(gfp_mask), nodemask) {
3220 : 0 : if (zone_idx(zone) > ZONE_NORMAL)
3221 : 0 : continue;
3222 : :
3223 : : /* Throttle based on the first usable node */
3224 : 0 : pgdat = zone->zone_pgdat;
3225 : 0 : if (allow_direct_reclaim(pgdat))
3226 : : goto out;
3227 : : break;
3228 : : }
3229 : :
3230 : : /* If no zone was usable by the allocation flags then do not throttle */
3231 : 0 : if (!pgdat)
3232 : : goto out;
3233 : :
3234 : : /* Account for the throttling */
3235 : : count_vm_event(PGSCAN_DIRECT_THROTTLE);
3236 : :
3237 : : /*
3238 : : * If the caller cannot enter the filesystem, it's possible that it
3239 : : * is due to the caller holding an FS lock or performing a journal
3240 : : * transaction in the case of a filesystem like ext[3|4]. In this case,
3241 : : * it is not safe to block on pfmemalloc_wait as kswapd could be
3242 : : * blocked waiting on the same lock. Instead, throttle for up to a
3243 : : * second before continuing.
3244 : : */
3245 : 0 : if (!(gfp_mask & __GFP_FS)) {
3246 : 0 : wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3247 : : allow_direct_reclaim(pgdat), HZ);
3248 : :
3249 : : goto check_pending;
3250 : : }
3251 : :
3252 : : /* Throttle until kswapd wakes the process */
3253 : 0 : wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3254 : : allow_direct_reclaim(pgdat));
3255 : :
3256 : : check_pending:
3257 : 0 : if (fatal_signal_pending(current))
3258 : : return true;
3259 : :
3260 : : out:
3261 : : return false;
3262 : : }
3263 : :
3264 : 0 : unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3265 : : gfp_t gfp_mask, nodemask_t *nodemask)
3266 : : {
3267 : : unsigned long nr_reclaimed;
3268 : 0 : struct scan_control sc = {
3269 : : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3270 : 0 : .gfp_mask = current_gfp_context(gfp_mask),
3271 : : .reclaim_idx = gfp_zone(gfp_mask),
3272 : : .order = order,
3273 : : .nodemask = nodemask,
3274 : : .priority = DEF_PRIORITY,
3275 : 0 : .may_writepage = !laptop_mode,
3276 : : .may_unmap = 1,
3277 : : .may_swap = 1,
3278 : : };
3279 : :
3280 : : /*
3281 : : * scan_control uses s8 fields for order, priority, and reclaim_idx.
3282 : : * Confirm they are large enough for max values.
3283 : : */
3284 : : BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3285 : : BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3286 : : BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3287 : :
3288 : : /*
3289 : : * Do not enter reclaim if fatal signal was delivered while throttled.
3290 : : * 1 is returned so that the page allocator does not OOM kill at this
3291 : : * point.
3292 : : */
3293 : 0 : if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3294 : : return 1;
3295 : :
3296 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3297 : 0 : trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3298 : :
3299 : 0 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3300 : :
3301 : 0 : trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3302 : 0 : set_task_reclaim_state(current, NULL);
3303 : :
3304 : 0 : return nr_reclaimed;
3305 : : }
3306 : :
3307 : : #ifdef CONFIG_MEMCG
3308 : :
3309 : : /* Only used by soft limit reclaim. Do not reuse for anything else. */
3310 : 0 : unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3311 : : gfp_t gfp_mask, bool noswap,
3312 : : pg_data_t *pgdat,
3313 : : unsigned long *nr_scanned)
3314 : : {
3315 : 0 : struct scan_control sc = {
3316 : : .nr_to_reclaim = SWAP_CLUSTER_MAX,
3317 : : .target_mem_cgroup = memcg,
3318 : 0 : .may_writepage = !laptop_mode,
3319 : : .may_unmap = 1,
3320 : : .reclaim_idx = MAX_NR_ZONES - 1,
3321 : 0 : .may_swap = !noswap,
3322 : : };
3323 : : unsigned long lru_pages;
3324 : :
3325 : 0 : WARN_ON_ONCE(!current->reclaim_state);
3326 : :
3327 : 0 : sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3328 : : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3329 : :
3330 : 0 : trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3331 : : sc.gfp_mask);
3332 : :
3333 : : /*
3334 : : * NOTE: Although we can get the priority field, using it
3335 : : * here is not a good idea, since it limits the pages we can scan.
3336 : : * if we don't reclaim here, the shrink_node from balance_pgdat
3337 : : * will pick up pages from other mem cgroup's as well. We hack
3338 : : * the priority and make it zero.
3339 : : */
3340 : 0 : shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3341 : :
3342 : 0 : trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3343 : :
3344 : 0 : *nr_scanned = sc.nr_scanned;
3345 : :
3346 : 0 : return sc.nr_reclaimed;
3347 : : }
3348 : :
3349 : 0 : unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3350 : : unsigned long nr_pages,
3351 : : gfp_t gfp_mask,
3352 : : bool may_swap)
3353 : : {
3354 : : struct zonelist *zonelist;
3355 : : unsigned long nr_reclaimed;
3356 : : unsigned long pflags;
3357 : : int nid;
3358 : : unsigned int noreclaim_flag;
3359 : 0 : struct scan_control sc = {
3360 : 0 : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3361 : 0 : .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3362 : : (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3363 : : .reclaim_idx = MAX_NR_ZONES - 1,
3364 : : .target_mem_cgroup = memcg,
3365 : : .priority = DEF_PRIORITY,
3366 : 0 : .may_writepage = !laptop_mode,
3367 : : .may_unmap = 1,
3368 : : .may_swap = may_swap,
3369 : : };
3370 : :
3371 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3372 : : /*
3373 : : * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3374 : : * take care of from where we get pages. So the node where we start the
3375 : : * scan does not need to be the current node.
3376 : : */
3377 : 0 : nid = mem_cgroup_select_victim_node(memcg);
3378 : :
3379 : : zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3380 : :
3381 : 0 : trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3382 : :
3383 : : psi_memstall_enter(&pflags);
3384 : : noreclaim_flag = memalloc_noreclaim_save();
3385 : :
3386 : 0 : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3387 : :
3388 : : memalloc_noreclaim_restore(noreclaim_flag);
3389 : : psi_memstall_leave(&pflags);
3390 : :
3391 : 0 : trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3392 : 0 : set_task_reclaim_state(current, NULL);
3393 : :
3394 : 0 : return nr_reclaimed;
3395 : : }
3396 : : #endif
3397 : :
3398 : 0 : static void age_active_anon(struct pglist_data *pgdat,
3399 : : struct scan_control *sc)
3400 : : {
3401 : : struct mem_cgroup *memcg;
3402 : :
3403 : 0 : if (!total_swap_pages)
3404 : 0 : return;
3405 : :
3406 : 0 : memcg = mem_cgroup_iter(NULL, NULL, NULL);
3407 : : do {
3408 : : struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3409 : :
3410 : 0 : if (inactive_list_is_low(lruvec, false, sc, true))
3411 : 0 : shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3412 : : sc, LRU_ACTIVE_ANON);
3413 : :
3414 : 0 : memcg = mem_cgroup_iter(NULL, memcg, NULL);
3415 : 0 : } while (memcg);
3416 : : }
3417 : :
3418 : : static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3419 : : {
3420 : : int i;
3421 : : struct zone *zone;
3422 : :
3423 : : /*
3424 : : * Check for watermark boosts top-down as the higher zones
3425 : : * are more likely to be boosted. Both watermarks and boosts
3426 : : * should not be checked at the time time as reclaim would
3427 : : * start prematurely when there is no boosting and a lower
3428 : : * zone is balanced.
3429 : : */
3430 : 0 : for (i = classzone_idx; i >= 0; i--) {
3431 : 0 : zone = pgdat->node_zones + i;
3432 : 0 : if (!managed_zone(zone))
3433 : 0 : continue;
3434 : :
3435 : 0 : if (zone->watermark_boost)
3436 : : return true;
3437 : : }
3438 : :
3439 : : return false;
3440 : : }
3441 : :
3442 : : /*
3443 : : * Returns true if there is an eligible zone balanced for the request order
3444 : : * and classzone_idx
3445 : : */
3446 : 3 : static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3447 : : {
3448 : : int i;
3449 : : unsigned long mark = -1;
3450 : : struct zone *zone;
3451 : :
3452 : : /*
3453 : : * Check watermarks bottom-up as lower zones are more likely to
3454 : : * meet watermarks.
3455 : : */
3456 : 3 : for (i = 0; i <= classzone_idx; i++) {
3457 : 3 : zone = pgdat->node_zones + i;
3458 : :
3459 : 3 : if (!managed_zone(zone))
3460 : 0 : continue;
3461 : :
3462 : 3 : mark = high_wmark_pages(zone);
3463 : 3 : if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3464 : : return true;
3465 : : }
3466 : :
3467 : : /*
3468 : : * If a node has no populated zone within classzone_idx, it does not
3469 : : * need balancing by definition. This can happen if a zone-restricted
3470 : : * allocation tries to wake a remote kswapd.
3471 : : */
3472 : 0 : if (mark == -1)
3473 : : return true;
3474 : :
3475 : 0 : return false;
3476 : : }
3477 : :
3478 : : /* Clear pgdat state for congested, dirty or under writeback. */
3479 : 3 : static void clear_pgdat_congested(pg_data_t *pgdat)
3480 : : {
3481 : 3 : clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3482 : 3 : clear_bit(PGDAT_DIRTY, &pgdat->flags);
3483 : 3 : clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3484 : 3 : }
3485 : :
3486 : : /*
3487 : : * Prepare kswapd for sleeping. This verifies that there are no processes
3488 : : * waiting in throttle_direct_reclaim() and that watermarks have been met.
3489 : : *
3490 : : * Returns true if kswapd is ready to sleep
3491 : : */
3492 : 3 : static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3493 : : {
3494 : : /*
3495 : : * The throttled processes are normally woken up in balance_pgdat() as
3496 : : * soon as allow_direct_reclaim() is true. But there is a potential
3497 : : * race between when kswapd checks the watermarks and a process gets
3498 : : * throttled. There is also a potential race if processes get
3499 : : * throttled, kswapd wakes, a large process exits thereby balancing the
3500 : : * zones, which causes kswapd to exit balance_pgdat() before reaching
3501 : : * the wake up checks. If kswapd is going to sleep, no process should
3502 : : * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3503 : : * the wake up is premature, processes will wake kswapd and get
3504 : : * throttled again. The difference from wake ups in balance_pgdat() is
3505 : : * that here we are under prepare_to_wait().
3506 : : */
3507 : 3 : if (waitqueue_active(&pgdat->pfmemalloc_wait))
3508 : 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3509 : :
3510 : : /* Hopeless node, leave it to direct reclaim */
3511 : 3 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3512 : : return true;
3513 : :
3514 : 3 : if (pgdat_balanced(pgdat, order, classzone_idx)) {
3515 : 3 : clear_pgdat_congested(pgdat);
3516 : 3 : return true;
3517 : : }
3518 : :
3519 : : return false;
3520 : : }
3521 : :
3522 : : /*
3523 : : * kswapd shrinks a node of pages that are at or below the highest usable
3524 : : * zone that is currently unbalanced.
3525 : : *
3526 : : * Returns true if kswapd scanned at least the requested number of pages to
3527 : : * reclaim or if the lack of progress was due to pages under writeback.
3528 : : * This is used to determine if the scanning priority needs to be raised.
3529 : : */
3530 : 0 : static bool kswapd_shrink_node(pg_data_t *pgdat,
3531 : : struct scan_control *sc)
3532 : : {
3533 : : struct zone *zone;
3534 : : int z;
3535 : :
3536 : : /* Reclaim a number of pages proportional to the number of zones */
3537 : 0 : sc->nr_to_reclaim = 0;
3538 : 0 : for (z = 0; z <= sc->reclaim_idx; z++) {
3539 : 0 : zone = pgdat->node_zones + z;
3540 : 0 : if (!managed_zone(zone))
3541 : 0 : continue;
3542 : :
3543 : 0 : sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3544 : : }
3545 : :
3546 : : /*
3547 : : * Historically care was taken to put equal pressure on all zones but
3548 : : * now pressure is applied based on node LRU order.
3549 : : */
3550 : 0 : shrink_node(pgdat, sc);
3551 : :
3552 : : /*
3553 : : * Fragmentation may mean that the system cannot be rebalanced for
3554 : : * high-order allocations. If twice the allocation size has been
3555 : : * reclaimed then recheck watermarks only at order-0 to prevent
3556 : : * excessive reclaim. Assume that a process requested a high-order
3557 : : * can direct reclaim/compact.
3558 : : */
3559 : 0 : if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3560 : 0 : sc->order = 0;
3561 : :
3562 : 0 : return sc->nr_scanned >= sc->nr_to_reclaim;
3563 : : }
3564 : :
3565 : : /*
3566 : : * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3567 : : * that are eligible for use by the caller until at least one zone is
3568 : : * balanced.
3569 : : *
3570 : : * Returns the order kswapd finished reclaiming at.
3571 : : *
3572 : : * kswapd scans the zones in the highmem->normal->dma direction. It skips
3573 : : * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3574 : : * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3575 : : * or lower is eligible for reclaim until at least one usable zone is
3576 : : * balanced.
3577 : : */
3578 : 0 : static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3579 : : {
3580 : : int i;
3581 : : unsigned long nr_soft_reclaimed;
3582 : : unsigned long nr_soft_scanned;
3583 : : unsigned long pflags;
3584 : : unsigned long nr_boost_reclaim;
3585 : 0 : unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3586 : : bool boosted;
3587 : : struct zone *zone;
3588 : 0 : struct scan_control sc = {
3589 : : .gfp_mask = GFP_KERNEL,
3590 : : .order = order,
3591 : : .may_unmap = 1,
3592 : : };
3593 : :
3594 : 0 : set_task_reclaim_state(current, &sc.reclaim_state);
3595 : : psi_memstall_enter(&pflags);
3596 : : __fs_reclaim_acquire();
3597 : :
3598 : : count_vm_event(PAGEOUTRUN);
3599 : :
3600 : : /*
3601 : : * Account for the reclaim boost. Note that the zone boost is left in
3602 : : * place so that parallel allocations that are near the watermark will
3603 : : * stall or direct reclaim until kswapd is finished.
3604 : : */
3605 : : nr_boost_reclaim = 0;
3606 : 0 : for (i = 0; i <= classzone_idx; i++) {
3607 : 0 : zone = pgdat->node_zones + i;
3608 : 0 : if (!managed_zone(zone))
3609 : 0 : continue;
3610 : :
3611 : 0 : nr_boost_reclaim += zone->watermark_boost;
3612 : 0 : zone_boosts[i] = zone->watermark_boost;
3613 : : }
3614 : : boosted = nr_boost_reclaim;
3615 : :
3616 : : restart:
3617 : 0 : sc.priority = DEF_PRIORITY;
3618 : : do {
3619 : 0 : unsigned long nr_reclaimed = sc.nr_reclaimed;
3620 : : bool raise_priority = true;
3621 : : bool balanced;
3622 : : bool ret;
3623 : :
3624 : 0 : sc.reclaim_idx = classzone_idx;
3625 : :
3626 : : /*
3627 : : * If the number of buffer_heads exceeds the maximum allowed
3628 : : * then consider reclaiming from all zones. This has a dual
3629 : : * purpose -- on 64-bit systems it is expected that
3630 : : * buffer_heads are stripped during active rotation. On 32-bit
3631 : : * systems, highmem pages can pin lowmem memory and shrinking
3632 : : * buffers can relieve lowmem pressure. Reclaim may still not
3633 : : * go ahead if all eligible zones for the original allocation
3634 : : * request are balanced to avoid excessive reclaim from kswapd.
3635 : : */
3636 : 0 : if (buffer_heads_over_limit) {
3637 : 0 : for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3638 : 0 : zone = pgdat->node_zones + i;
3639 : 0 : if (!managed_zone(zone))
3640 : 0 : continue;
3641 : :
3642 : 0 : sc.reclaim_idx = i;
3643 : 0 : break;
3644 : : }
3645 : : }
3646 : :
3647 : : /*
3648 : : * If the pgdat is imbalanced then ignore boosting and preserve
3649 : : * the watermarks for a later time and restart. Note that the
3650 : : * zone watermarks will be still reset at the end of balancing
3651 : : * on the grounds that the normal reclaim should be enough to
3652 : : * re-evaluate if boosting is required when kswapd next wakes.
3653 : : */
3654 : 0 : balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3655 : 0 : if (!balanced && nr_boost_reclaim) {
3656 : : nr_boost_reclaim = 0;
3657 : : goto restart;
3658 : : }
3659 : :
3660 : : /*
3661 : : * If boosting is not active then only reclaim if there are no
3662 : : * eligible zones. Note that sc.reclaim_idx is not used as
3663 : : * buffer_heads_over_limit may have adjusted it.
3664 : : */
3665 : 0 : if (!nr_boost_reclaim && balanced)
3666 : : goto out;
3667 : :
3668 : : /* Limit the priority of boosting to avoid reclaim writeback */
3669 : 0 : if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3670 : : raise_priority = false;
3671 : :
3672 : : /*
3673 : : * Do not writeback or swap pages for boosted reclaim. The
3674 : : * intent is to relieve pressure not issue sub-optimal IO
3675 : : * from reclaim context. If no pages are reclaimed, the
3676 : : * reclaim will be aborted.
3677 : : */
3678 : 0 : sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3679 : 0 : sc.may_swap = !nr_boost_reclaim;
3680 : :
3681 : : /*
3682 : : * Do some background aging of the anon list, to give
3683 : : * pages a chance to be referenced before reclaiming. All
3684 : : * pages are rotated regardless of classzone as this is
3685 : : * about consistent aging.
3686 : : */
3687 : 0 : age_active_anon(pgdat, &sc);
3688 : :
3689 : : /*
3690 : : * If we're getting trouble reclaiming, start doing writepage
3691 : : * even in laptop mode.
3692 : : */
3693 : 0 : if (sc.priority < DEF_PRIORITY - 2)
3694 : 0 : sc.may_writepage = 1;
3695 : :
3696 : : /* Call soft limit reclaim before calling shrink_node. */
3697 : 0 : sc.nr_scanned = 0;
3698 : 0 : nr_soft_scanned = 0;
3699 : 0 : nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3700 : : sc.gfp_mask, &nr_soft_scanned);
3701 : 0 : sc.nr_reclaimed += nr_soft_reclaimed;
3702 : :
3703 : : /*
3704 : : * There should be no need to raise the scanning priority if
3705 : : * enough pages are already being scanned that that high
3706 : : * watermark would be met at 100% efficiency.
3707 : : */
3708 : 0 : if (kswapd_shrink_node(pgdat, &sc))
3709 : : raise_priority = false;
3710 : :
3711 : : /*
3712 : : * If the low watermark is met there is no need for processes
3713 : : * to be throttled on pfmemalloc_wait as they should not be
3714 : : * able to safely make forward progress. Wake them
3715 : : */
3716 : 0 : if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3717 : 0 : allow_direct_reclaim(pgdat))
3718 : 0 : wake_up_all(&pgdat->pfmemalloc_wait);
3719 : :
3720 : : /* Check if kswapd should be suspending */
3721 : : __fs_reclaim_release();
3722 : : ret = try_to_freeze();
3723 : : __fs_reclaim_acquire();
3724 : 0 : if (ret || kthread_should_stop())
3725 : : break;
3726 : :
3727 : : /*
3728 : : * Raise priority if scanning rate is too low or there was no
3729 : : * progress in reclaiming pages
3730 : : */
3731 : 0 : nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3732 : 0 : nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3733 : :
3734 : : /*
3735 : : * If reclaim made no progress for a boost, stop reclaim as
3736 : : * IO cannot be queued and it could be an infinite loop in
3737 : : * extreme circumstances.
3738 : : */
3739 : 0 : if (nr_boost_reclaim && !nr_reclaimed)
3740 : : break;
3741 : :
3742 : 0 : if (raise_priority || !nr_reclaimed)
3743 : 0 : sc.priority--;
3744 : 0 : } while (sc.priority >= 1);
3745 : :
3746 : 0 : if (!sc.nr_reclaimed)
3747 : 0 : pgdat->kswapd_failures++;
3748 : :
3749 : : out:
3750 : : /* If reclaim was boosted, account for the reclaim done in this pass */
3751 : 0 : if (boosted) {
3752 : : unsigned long flags;
3753 : :
3754 : 0 : for (i = 0; i <= classzone_idx; i++) {
3755 : 0 : if (!zone_boosts[i])
3756 : 0 : continue;
3757 : :
3758 : : /* Increments are under the zone lock */
3759 : 0 : zone = pgdat->node_zones + i;
3760 : 0 : spin_lock_irqsave(&zone->lock, flags);
3761 : 0 : zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3762 : : spin_unlock_irqrestore(&zone->lock, flags);
3763 : : }
3764 : :
3765 : : /*
3766 : : * As there is now likely space, wakeup kcompact to defragment
3767 : : * pageblocks.
3768 : : */
3769 : 0 : wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3770 : : }
3771 : :
3772 : 0 : snapshot_refaults(NULL, pgdat);
3773 : : __fs_reclaim_release();
3774 : : psi_memstall_leave(&pflags);
3775 : 0 : set_task_reclaim_state(current, NULL);
3776 : :
3777 : : /*
3778 : : * Return the order kswapd stopped reclaiming at as
3779 : : * prepare_kswapd_sleep() takes it into account. If another caller
3780 : : * entered the allocator slow path while kswapd was awake, order will
3781 : : * remain at the higher level.
3782 : : */
3783 : 0 : return sc.order;
3784 : : }
3785 : :
3786 : : /*
3787 : : * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3788 : : * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3789 : : * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3790 : : * after previous reclaim attempt (node is still unbalanced). In that case
3791 : : * return the zone index of the previous kswapd reclaim cycle.
3792 : : */
3793 : : static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3794 : : enum zone_type prev_classzone_idx)
3795 : : {
3796 : 3 : if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3797 : : return prev_classzone_idx;
3798 : : return pgdat->kswapd_classzone_idx;
3799 : : }
3800 : :
3801 : 3 : static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3802 : : unsigned int classzone_idx)
3803 : : {
3804 : : long remaining = 0;
3805 : 3 : DEFINE_WAIT(wait);
3806 : :
3807 : 3 : if (freezing(current) || kthread_should_stop())
3808 : 0 : return;
3809 : :
3810 : 3 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3811 : :
3812 : : /*
3813 : : * Try to sleep for a short interval. Note that kcompactd will only be
3814 : : * woken if it is possible to sleep for a short interval. This is
3815 : : * deliberate on the assumption that if reclaim cannot keep an
3816 : : * eligible zone balanced that it's also unlikely that compaction will
3817 : : * succeed.
3818 : : */
3819 : 3 : if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3820 : : /*
3821 : : * Compaction records what page blocks it recently failed to
3822 : : * isolate pages from and skips them in the future scanning.
3823 : : * When kswapd is going to sleep, it is reasonable to assume
3824 : : * that pages and compaction may succeed so reset the cache.
3825 : : */
3826 : 3 : reset_isolation_suitable(pgdat);
3827 : :
3828 : : /*
3829 : : * We have freed the memory, now we should compact it to make
3830 : : * allocation of the requested order possible.
3831 : : */
3832 : 3 : wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3833 : :
3834 : 3 : remaining = schedule_timeout(HZ/10);
3835 : :
3836 : : /*
3837 : : * If woken prematurely then reset kswapd_classzone_idx and
3838 : : * order. The values will either be from a wakeup request or
3839 : : * the previous request that slept prematurely.
3840 : : */
3841 : 3 : if (remaining) {
3842 : 0 : pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3843 : 0 : pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3844 : : }
3845 : :
3846 : 3 : finish_wait(&pgdat->kswapd_wait, &wait);
3847 : 3 : prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3848 : : }
3849 : :
3850 : : /*
3851 : : * After a short sleep, check if it was a premature sleep. If not, then
3852 : : * go fully to sleep until explicitly woken up.
3853 : : */
3854 : 3 : if (!remaining &&
3855 : 3 : prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3856 : 3 : trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3857 : :
3858 : : /*
3859 : : * vmstat counters are not perfectly accurate and the estimated
3860 : : * value for counters such as NR_FREE_PAGES can deviate from the
3861 : : * true value by nr_online_cpus * threshold. To avoid the zone
3862 : : * watermarks being breached while under pressure, we reduce the
3863 : : * per-cpu vmstat threshold while kswapd is awake and restore
3864 : : * them before going back to sleep.
3865 : : */
3866 : 3 : set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3867 : :
3868 : 3 : if (!kthread_should_stop())
3869 : 3 : schedule();
3870 : :
3871 : 0 : set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3872 : : } else {
3873 : 0 : if (remaining)
3874 : : count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3875 : : else
3876 : : count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3877 : : }
3878 : 0 : finish_wait(&pgdat->kswapd_wait, &wait);
3879 : : }
3880 : :
3881 : : /*
3882 : : * The background pageout daemon, started as a kernel thread
3883 : : * from the init process.
3884 : : *
3885 : : * This basically trickles out pages so that we have _some_
3886 : : * free memory available even if there is no other activity
3887 : : * that frees anything up. This is needed for things like routing
3888 : : * etc, where we otherwise might have all activity going on in
3889 : : * asynchronous contexts that cannot page things out.
3890 : : *
3891 : : * If there are applications that are active memory-allocators
3892 : : * (most normal use), this basically shouldn't matter.
3893 : : */
3894 : 3 : static int kswapd(void *p)
3895 : : {
3896 : : unsigned int alloc_order, reclaim_order;
3897 : : unsigned int classzone_idx = MAX_NR_ZONES - 1;
3898 : : pg_data_t *pgdat = (pg_data_t*)p;
3899 : 3 : struct task_struct *tsk = current;
3900 : : const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3901 : :
3902 : 3 : if (!cpumask_empty(cpumask))
3903 : 3 : set_cpus_allowed_ptr(tsk, cpumask);
3904 : :
3905 : : /*
3906 : : * Tell the memory management that we're a "memory allocator",
3907 : : * and that if we need more memory we should get access to it
3908 : : * regardless (see "__alloc_pages()"). "kswapd" should
3909 : : * never get caught in the normal page freeing logic.
3910 : : *
3911 : : * (Kswapd normally doesn't need memory anyway, but sometimes
3912 : : * you need a small amount of memory in order to be able to
3913 : : * page out something else, and this flag essentially protects
3914 : : * us from recursively trying to free more memory as we're
3915 : : * trying to free the first piece of memory in the first place).
3916 : : */
3917 : 3 : tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3918 : 3 : set_freezable();
3919 : :
3920 : 3 : pgdat->kswapd_order = 0;
3921 : 3 : pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3922 : : for ( ; ; ) {
3923 : : bool ret;
3924 : :
3925 : 3 : alloc_order = reclaim_order = pgdat->kswapd_order;
3926 : : classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3927 : :
3928 : : kswapd_try_sleep:
3929 : 3 : kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3930 : : classzone_idx);
3931 : :
3932 : : /* Read the new order and classzone_idx */
3933 : 0 : alloc_order = reclaim_order = pgdat->kswapd_order;
3934 : : classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3935 : 0 : pgdat->kswapd_order = 0;
3936 : 0 : pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3937 : :
3938 : : ret = try_to_freeze();
3939 : 0 : if (kthread_should_stop())
3940 : : break;
3941 : :
3942 : : /*
3943 : : * We can speed up thawing tasks if we don't call balance_pgdat
3944 : : * after returning from the refrigerator
3945 : : */
3946 : 0 : if (ret)
3947 : 0 : continue;
3948 : :
3949 : : /*
3950 : : * Reclaim begins at the requested order but if a high-order
3951 : : * reclaim fails then kswapd falls back to reclaiming for
3952 : : * order-0. If that happens, kswapd will consider sleeping
3953 : : * for the order it finished reclaiming at (reclaim_order)
3954 : : * but kcompactd is woken to compact for the original
3955 : : * request (alloc_order).
3956 : : */
3957 : 0 : trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3958 : : alloc_order);
3959 : 0 : reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3960 : 0 : if (reclaim_order < alloc_order)
3961 : : goto kswapd_try_sleep;
3962 : : }
3963 : :
3964 : 0 : tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3965 : :
3966 : 0 : return 0;
3967 : : }
3968 : :
3969 : : /*
3970 : : * A zone is low on free memory or too fragmented for high-order memory. If
3971 : : * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3972 : : * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3973 : : * has failed or is not needed, still wake up kcompactd if only compaction is
3974 : : * needed.
3975 : : */
3976 : 0 : void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3977 : : enum zone_type classzone_idx)
3978 : : {
3979 : : pg_data_t *pgdat;
3980 : :
3981 : 0 : if (!managed_zone(zone))
3982 : : return;
3983 : :
3984 : 0 : if (!cpuset_zone_allowed(zone, gfp_flags))
3985 : : return;
3986 : 0 : pgdat = zone->zone_pgdat;
3987 : :
3988 : 0 : if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3989 : 0 : pgdat->kswapd_classzone_idx = classzone_idx;
3990 : : else
3991 : 0 : pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3992 : : classzone_idx);
3993 : 0 : pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3994 : 0 : if (!waitqueue_active(&pgdat->kswapd_wait))
3995 : : return;
3996 : :
3997 : : /* Hopeless node, leave it to direct reclaim if possible */
3998 : 0 : if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3999 : 0 : (pgdat_balanced(pgdat, order, classzone_idx) &&
4000 : : !pgdat_watermark_boosted(pgdat, classzone_idx))) {
4001 : : /*
4002 : : * There may be plenty of free memory available, but it's too
4003 : : * fragmented for high-order allocations. Wake up kcompactd
4004 : : * and rely on compaction_suitable() to determine if it's
4005 : : * needed. If it fails, it will defer subsequent attempts to
4006 : : * ratelimit its work.
4007 : : */
4008 : 0 : if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4009 : 0 : wakeup_kcompactd(pgdat, order, classzone_idx);
4010 : : return;
4011 : : }
4012 : :
4013 : 0 : trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4014 : : gfp_flags);
4015 : 0 : wake_up_interruptible(&pgdat->kswapd_wait);
4016 : : }
4017 : :
4018 : : #ifdef CONFIG_HIBERNATION
4019 : : /*
4020 : : * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4021 : : * freed pages.
4022 : : *
4023 : : * Rather than trying to age LRUs the aim is to preserve the overall
4024 : : * LRU order by reclaiming preferentially
4025 : : * inactive > active > active referenced > active mapped
4026 : : */
4027 : : unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4028 : : {
4029 : : struct scan_control sc = {
4030 : : .nr_to_reclaim = nr_to_reclaim,
4031 : : .gfp_mask = GFP_HIGHUSER_MOVABLE,
4032 : : .reclaim_idx = MAX_NR_ZONES - 1,
4033 : : .priority = DEF_PRIORITY,
4034 : : .may_writepage = 1,
4035 : : .may_unmap = 1,
4036 : : .may_swap = 1,
4037 : : .hibernation_mode = 1,
4038 : : };
4039 : : struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4040 : : unsigned long nr_reclaimed;
4041 : : unsigned int noreclaim_flag;
4042 : :
4043 : : fs_reclaim_acquire(sc.gfp_mask);
4044 : : noreclaim_flag = memalloc_noreclaim_save();
4045 : : set_task_reclaim_state(current, &sc.reclaim_state);
4046 : :
4047 : : nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4048 : :
4049 : : set_task_reclaim_state(current, NULL);
4050 : : memalloc_noreclaim_restore(noreclaim_flag);
4051 : : fs_reclaim_release(sc.gfp_mask);
4052 : :
4053 : : return nr_reclaimed;
4054 : : }
4055 : : #endif /* CONFIG_HIBERNATION */
4056 : :
4057 : : /* It's optimal to keep kswapds on the same CPUs as their memory, but
4058 : : not required for correctness. So if the last cpu in a node goes
4059 : : away, we get changed to run anywhere: as the first one comes back,
4060 : : restore their cpu bindings. */
4061 : 0 : static int kswapd_cpu_online(unsigned int cpu)
4062 : : {
4063 : : int nid;
4064 : :
4065 : 0 : for_each_node_state(nid, N_MEMORY) {
4066 : : pg_data_t *pgdat = NODE_DATA(nid);
4067 : : const struct cpumask *mask;
4068 : :
4069 : : mask = cpumask_of_node(pgdat->node_id);
4070 : :
4071 : 0 : if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4072 : : /* One of our CPUs online: restore mask */
4073 : 0 : set_cpus_allowed_ptr(pgdat->kswapd, mask);
4074 : : }
4075 : 0 : return 0;
4076 : : }
4077 : :
4078 : : /*
4079 : : * This kswapd start function will be called by init and node-hot-add.
4080 : : * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4081 : : */
4082 : 3 : int kswapd_run(int nid)
4083 : : {
4084 : : pg_data_t *pgdat = NODE_DATA(nid);
4085 : : int ret = 0;
4086 : :
4087 : 3 : if (pgdat->kswapd)
4088 : : return 0;
4089 : :
4090 : 3 : pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4091 : 3 : if (IS_ERR(pgdat->kswapd)) {
4092 : : /* failure at boot is fatal */
4093 : 0 : BUG_ON(system_state < SYSTEM_RUNNING);
4094 : 0 : pr_err("Failed to start kswapd on node %d\n", nid);
4095 : 0 : ret = PTR_ERR(pgdat->kswapd);
4096 : 0 : pgdat->kswapd = NULL;
4097 : : }
4098 : 3 : return ret;
4099 : : }
4100 : :
4101 : : /*
4102 : : * Called by memory hotplug when all memory in a node is offlined. Caller must
4103 : : * hold mem_hotplug_begin/end().
4104 : : */
4105 : 0 : void kswapd_stop(int nid)
4106 : : {
4107 : 0 : struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4108 : :
4109 : 0 : if (kswapd) {
4110 : 0 : kthread_stop(kswapd);
4111 : 0 : NODE_DATA(nid)->kswapd = NULL;
4112 : : }
4113 : 0 : }
4114 : :
4115 : 3 : static int __init kswapd_init(void)
4116 : : {
4117 : : int nid, ret;
4118 : :
4119 : 3 : swap_setup();
4120 : 3 : for_each_node_state(nid, N_MEMORY)
4121 : 3 : kswapd_run(nid);
4122 : : ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4123 : : "mm/vmscan:online", kswapd_cpu_online,
4124 : : NULL);
4125 : 3 : WARN_ON(ret < 0);
4126 : 3 : return 0;
4127 : : }
4128 : :
4129 : : module_init(kswapd_init)
4130 : :
4131 : : #ifdef CONFIG_NUMA
4132 : : /*
4133 : : * Node reclaim mode
4134 : : *
4135 : : * If non-zero call node_reclaim when the number of free pages falls below
4136 : : * the watermarks.
4137 : : */
4138 : : int node_reclaim_mode __read_mostly;
4139 : :
4140 : : #define RECLAIM_OFF 0
4141 : : #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4142 : : #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4143 : : #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4144 : :
4145 : : /*
4146 : : * Priority for NODE_RECLAIM. This determines the fraction of pages
4147 : : * of a node considered for each zone_reclaim. 4 scans 1/16th of
4148 : : * a zone.
4149 : : */
4150 : : #define NODE_RECLAIM_PRIORITY 4
4151 : :
4152 : : /*
4153 : : * Percentage of pages in a zone that must be unmapped for node_reclaim to
4154 : : * occur.
4155 : : */
4156 : : int sysctl_min_unmapped_ratio = 1;
4157 : :
4158 : : /*
4159 : : * If the number of slab pages in a zone grows beyond this percentage then
4160 : : * slab reclaim needs to occur.
4161 : : */
4162 : : int sysctl_min_slab_ratio = 5;
4163 : :
4164 : : static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4165 : : {
4166 : : unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4167 : : unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4168 : : node_page_state(pgdat, NR_ACTIVE_FILE);
4169 : :
4170 : : /*
4171 : : * It's possible for there to be more file mapped pages than
4172 : : * accounted for by the pages on the file LRU lists because
4173 : : * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4174 : : */
4175 : : return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4176 : : }
4177 : :
4178 : : /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4179 : : static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4180 : : {
4181 : : unsigned long nr_pagecache_reclaimable;
4182 : : unsigned long delta = 0;
4183 : :
4184 : : /*
4185 : : * If RECLAIM_UNMAP is set, then all file pages are considered
4186 : : * potentially reclaimable. Otherwise, we have to worry about
4187 : : * pages like swapcache and node_unmapped_file_pages() provides
4188 : : * a better estimate
4189 : : */
4190 : : if (node_reclaim_mode & RECLAIM_UNMAP)
4191 : : nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4192 : : else
4193 : : nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4194 : :
4195 : : /* If we can't clean pages, remove dirty pages from consideration */
4196 : : if (!(node_reclaim_mode & RECLAIM_WRITE))
4197 : : delta += node_page_state(pgdat, NR_FILE_DIRTY);
4198 : :
4199 : : /* Watch for any possible underflows due to delta */
4200 : : if (unlikely(delta > nr_pagecache_reclaimable))
4201 : : delta = nr_pagecache_reclaimable;
4202 : :
4203 : : return nr_pagecache_reclaimable - delta;
4204 : : }
4205 : :
4206 : : /*
4207 : : * Try to free up some pages from this node through reclaim.
4208 : : */
4209 : : static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4210 : : {
4211 : : /* Minimum pages needed in order to stay on node */
4212 : : const unsigned long nr_pages = 1 << order;
4213 : : struct task_struct *p = current;
4214 : : unsigned int noreclaim_flag;
4215 : : struct scan_control sc = {
4216 : : .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4217 : : .gfp_mask = current_gfp_context(gfp_mask),
4218 : : .order = order,
4219 : : .priority = NODE_RECLAIM_PRIORITY,
4220 : : .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4221 : : .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4222 : : .may_swap = 1,
4223 : : .reclaim_idx = gfp_zone(gfp_mask),
4224 : : };
4225 : :
4226 : : trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4227 : : sc.gfp_mask);
4228 : :
4229 : : cond_resched();
4230 : : fs_reclaim_acquire(sc.gfp_mask);
4231 : : /*
4232 : : * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4233 : : * and we also need to be able to write out pages for RECLAIM_WRITE
4234 : : * and RECLAIM_UNMAP.
4235 : : */
4236 : : noreclaim_flag = memalloc_noreclaim_save();
4237 : : p->flags |= PF_SWAPWRITE;
4238 : : set_task_reclaim_state(p, &sc.reclaim_state);
4239 : :
4240 : : if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4241 : : /*
4242 : : * Free memory by calling shrink node with increasing
4243 : : * priorities until we have enough memory freed.
4244 : : */
4245 : : do {
4246 : : shrink_node(pgdat, &sc);
4247 : : } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4248 : : }
4249 : :
4250 : : set_task_reclaim_state(p, NULL);
4251 : : current->flags &= ~PF_SWAPWRITE;
4252 : : memalloc_noreclaim_restore(noreclaim_flag);
4253 : : fs_reclaim_release(sc.gfp_mask);
4254 : :
4255 : : trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4256 : :
4257 : : return sc.nr_reclaimed >= nr_pages;
4258 : : }
4259 : :
4260 : : int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4261 : : {
4262 : : int ret;
4263 : :
4264 : : /*
4265 : : * Node reclaim reclaims unmapped file backed pages and
4266 : : * slab pages if we are over the defined limits.
4267 : : *
4268 : : * A small portion of unmapped file backed pages is needed for
4269 : : * file I/O otherwise pages read by file I/O will be immediately
4270 : : * thrown out if the node is overallocated. So we do not reclaim
4271 : : * if less than a specified percentage of the node is used by
4272 : : * unmapped file backed pages.
4273 : : */
4274 : : if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4275 : : node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4276 : : return NODE_RECLAIM_FULL;
4277 : :
4278 : : /*
4279 : : * Do not scan if the allocation should not be delayed.
4280 : : */
4281 : : if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4282 : : return NODE_RECLAIM_NOSCAN;
4283 : :
4284 : : /*
4285 : : * Only run node reclaim on the local node or on nodes that do not
4286 : : * have associated processors. This will favor the local processor
4287 : : * over remote processors and spread off node memory allocations
4288 : : * as wide as possible.
4289 : : */
4290 : : if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4291 : : return NODE_RECLAIM_NOSCAN;
4292 : :
4293 : : if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4294 : : return NODE_RECLAIM_NOSCAN;
4295 : :
4296 : : ret = __node_reclaim(pgdat, gfp_mask, order);
4297 : : clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4298 : :
4299 : : if (!ret)
4300 : : count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4301 : :
4302 : : return ret;
4303 : : }
4304 : : #endif
4305 : :
4306 : : /*
4307 : : * page_evictable - test whether a page is evictable
4308 : : * @page: the page to test
4309 : : *
4310 : : * Test whether page is evictable--i.e., should be placed on active/inactive
4311 : : * lists vs unevictable list.
4312 : : *
4313 : : * Reasons page might not be evictable:
4314 : : * (1) page's mapping marked unevictable
4315 : : * (2) page is part of an mlocked VMA
4316 : : *
4317 : : */
4318 : 3 : int page_evictable(struct page *page)
4319 : : {
4320 : : int ret;
4321 : :
4322 : : /* Prevent address_space of inode and swap cache from being freed */
4323 : : rcu_read_lock();
4324 : 3 : ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4325 : : rcu_read_unlock();
4326 : 3 : return ret;
4327 : : }
4328 : :
4329 : : /**
4330 : : * check_move_unevictable_pages - check pages for evictability and move to
4331 : : * appropriate zone lru list
4332 : : * @pvec: pagevec with lru pages to check
4333 : : *
4334 : : * Checks pages for evictability, if an evictable page is in the unevictable
4335 : : * lru list, moves it to the appropriate evictable lru list. This function
4336 : : * should be only used for lru pages.
4337 : : */
4338 : 0 : void check_move_unevictable_pages(struct pagevec *pvec)
4339 : : {
4340 : : struct lruvec *lruvec;
4341 : : struct pglist_data *pgdat = NULL;
4342 : : int pgscanned = 0;
4343 : : int pgrescued = 0;
4344 : : int i;
4345 : :
4346 : 0 : for (i = 0; i < pvec->nr; i++) {
4347 : 0 : struct page *page = pvec->pages[i];
4348 : : struct pglist_data *pagepgdat = page_pgdat(page);
4349 : :
4350 : 0 : pgscanned++;
4351 : 0 : if (pagepgdat != pgdat) {
4352 : 0 : if (pgdat)
4353 : : spin_unlock_irq(&pgdat->lru_lock);
4354 : : pgdat = pagepgdat;
4355 : : spin_lock_irq(&pgdat->lru_lock);
4356 : : }
4357 : 0 : lruvec = mem_cgroup_page_lruvec(page, pgdat);
4358 : :
4359 : 0 : if (!PageLRU(page) || !PageUnevictable(page))
4360 : 0 : continue;
4361 : :
4362 : 0 : if (page_evictable(page)) {
4363 : : enum lru_list lru = page_lru_base_type(page);
4364 : :
4365 : : VM_BUG_ON_PAGE(PageActive(page), page);
4366 : : ClearPageUnevictable(page);
4367 : : del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4368 : : add_page_to_lru_list(page, lruvec, lru);
4369 : 0 : pgrescued++;
4370 : : }
4371 : : }
4372 : :
4373 : 0 : if (pgdat) {
4374 : : __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4375 : : __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4376 : : spin_unlock_irq(&pgdat->lru_lock);
4377 : : }
4378 : 0 : }
4379 : : EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
|
