1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4 *
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
11 */
12
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15 #include <linux/mm.h>
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for buffer_heads_over_limit */
30 #include <linux/mm_inline.h>
31 #include <linux/backing-dev.h>
32 #include <linux/rmap.h>
33 #include <linux/topology.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/compaction.h>
37 #include <linux/notifier.h>
38 #include <linux/rwsem.h>
39 #include <linux/delay.h>
40 #include <linux/kthread.h>
41 #include <linux/freezer.h>
42 #include <linux/memcontrol.h>
43 #include <linux/migrate.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/memory-tiers.h>
47 #include <linux/oom.h>
48 #include <linux/pagevec.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52 #include <linux/psi.h>
53 #include <linux/pagewalk.h>
54 #include <linux/shmem_fs.h>
55 #include <linux/ctype.h>
56 #include <linux/debugfs.h>
57 #include <linux/khugepaged.h>
58 #include <linux/rculist_nulls.h>
59 #include <linux/random.h>
60
61 #include <asm/tlbflush.h>
62 #include <asm/div64.h>
63
64 #include <linux/swapops.h>
65 #include <linux/balloon_compaction.h>
66 #include <linux/sched/sysctl.h>
67
68 #include "internal.h"
69 #include "swap.h"
70
71 #define CREATE_TRACE_POINTS
72 #include <trace/events/vmscan.h>
73
74 struct scan_control {
75 /* How many pages shrink_list() should reclaim */
76 unsigned long nr_to_reclaim;
77
78 /*
79 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * are scanned.
81 */
82 nodemask_t *nodemask;
83
84 /*
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
87 */
88 struct mem_cgroup *target_mem_cgroup;
89
90 /*
91 * Scan pressure balancing between anon and file LRUs
92 */
93 unsigned long anon_cost;
94 unsigned long file_cost;
95
96 /* Can active folios be deactivated as part of reclaim? */
97 #define DEACTIVATE_ANON 1
98 #define DEACTIVATE_FILE 2
99 unsigned int may_deactivate:2;
100 unsigned int force_deactivate:1;
101 unsigned int skipped_deactivate:1;
102
103 /* Writepage batching in laptop mode; RECLAIM_WRITE */
104 unsigned int may_writepage:1;
105
106 /* Can mapped folios be reclaimed? */
107 unsigned int may_unmap:1;
108
109 /* Can folios be swapped as part of reclaim? */
110 unsigned int may_swap:1;
111
112 /* Proactive reclaim invoked by userspace through memory.reclaim */
113 unsigned int proactive:1;
114
115 /*
116 * Cgroup memory below memory.low is protected as long as we
117 * don't threaten to OOM. If any cgroup is reclaimed at
118 * reduced force or passed over entirely due to its memory.low
119 * setting (memcg_low_skipped), and nothing is reclaimed as a
120 * result, then go back for one more cycle that reclaims the protected
121 * memory (memcg_low_reclaim) to avert OOM.
122 */
123 unsigned int memcg_low_reclaim:1;
124 unsigned int memcg_low_skipped:1;
125
126 unsigned int hibernation_mode:1;
127
128 /* One of the zones is ready for compaction */
129 unsigned int compaction_ready:1;
130
131 /* There is easily reclaimable cold cache in the current node */
132 unsigned int cache_trim_mode:1;
133
134 /* The file folios on the current node are dangerously low */
135 unsigned int file_is_tiny:1;
136
137 /* Always discard instead of demoting to lower tier memory */
138 unsigned int no_demotion:1;
139
140 /* Allocation order */
141 s8 order;
142
143 /* Scan (total_size >> priority) pages at once */
144 s8 priority;
145
146 /* The highest zone to isolate folios for reclaim from */
147 s8 reclaim_idx;
148
149 /* This context's GFP mask */
150 gfp_t gfp_mask;
151
152 /* Incremented by the number of inactive pages that were scanned */
153 unsigned long nr_scanned;
154
155 /* Number of pages freed so far during a call to shrink_zones() */
156 unsigned long nr_reclaimed;
157
158 struct {
159 unsigned int dirty;
160 unsigned int unqueued_dirty;
161 unsigned int congested;
162 unsigned int writeback;
163 unsigned int immediate;
164 unsigned int file_taken;
165 unsigned int taken;
166 } nr;
167
168 /* for recording the reclaimed slab by now */
169 struct reclaim_state reclaim_state;
170 };
171
172 #ifdef ARCH_HAS_PREFETCHW
173 #define prefetchw_prev_lru_folio(_folio, _base, _field) \
174 do { \
175 if ((_folio)->lru.prev != _base) { \
176 struct folio *prev; \
177 \
178 prev = lru_to_folio(&(_folio->lru)); \
179 prefetchw(&prev->_field); \
180 } \
181 } while (0)
182 #else
183 #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
184 #endif
185
186 /*
187 * From 0 .. 200. Higher means more swappy.
188 */
189 int vm_swappiness = 60;
190
191 LIST_HEAD(shrinker_list);
192 DECLARE_RWSEM(shrinker_rwsem);
193
194 #ifdef CONFIG_MEMCG
195 static int shrinker_nr_max;
196
197 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
shrinker_map_size(int nr_items)198 static inline int shrinker_map_size(int nr_items)
199 {
200 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
201 }
202
shrinker_defer_size(int nr_items)203 static inline int shrinker_defer_size(int nr_items)
204 {
205 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
206 }
207
shrinker_info_protected(struct mem_cgroup * memcg,int nid)208 static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
209 int nid)
210 {
211 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
212 lockdep_is_held(&shrinker_rwsem));
213 }
214
expand_one_shrinker_info(struct mem_cgroup * memcg,int map_size,int defer_size,int old_map_size,int old_defer_size,int new_nr_max)215 static int expand_one_shrinker_info(struct mem_cgroup *memcg,
216 int map_size, int defer_size,
217 int old_map_size, int old_defer_size,
218 int new_nr_max)
219 {
220 struct shrinker_info *new, *old;
221 struct mem_cgroup_per_node *pn;
222 int nid;
223 int size = map_size + defer_size;
224
225 for_each_node(nid) {
226 pn = memcg->nodeinfo[nid];
227 old = shrinker_info_protected(memcg, nid);
228 /* Not yet online memcg */
229 if (!old)
230 return 0;
231
232 /* Already expanded this shrinker_info */
233 if (new_nr_max <= old->map_nr_max)
234 continue;
235
236 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
237 if (!new)
238 return -ENOMEM;
239
240 new->nr_deferred = (atomic_long_t *)(new + 1);
241 new->map = (void *)new->nr_deferred + defer_size;
242 new->map_nr_max = new_nr_max;
243
244 /* map: set all old bits, clear all new bits */
245 memset(new->map, (int)0xff, old_map_size);
246 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
247 /* nr_deferred: copy old values, clear all new values */
248 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
249 memset((void *)new->nr_deferred + old_defer_size, 0,
250 defer_size - old_defer_size);
251
252 rcu_assign_pointer(pn->shrinker_info, new);
253 kvfree_rcu(old, rcu);
254 }
255
256 return 0;
257 }
258
free_shrinker_info(struct mem_cgroup * memcg)259 void free_shrinker_info(struct mem_cgroup *memcg)
260 {
261 struct mem_cgroup_per_node *pn;
262 struct shrinker_info *info;
263 int nid;
264
265 for_each_node(nid) {
266 pn = memcg->nodeinfo[nid];
267 info = rcu_dereference_protected(pn->shrinker_info, true);
268 kvfree(info);
269 rcu_assign_pointer(pn->shrinker_info, NULL);
270 }
271 }
272
alloc_shrinker_info(struct mem_cgroup * memcg)273 int alloc_shrinker_info(struct mem_cgroup *memcg)
274 {
275 struct shrinker_info *info;
276 int nid, size, ret = 0;
277 int map_size, defer_size = 0;
278
279 down_write(&shrinker_rwsem);
280 map_size = shrinker_map_size(shrinker_nr_max);
281 defer_size = shrinker_defer_size(shrinker_nr_max);
282 size = map_size + defer_size;
283 for_each_node(nid) {
284 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
285 if (!info) {
286 free_shrinker_info(memcg);
287 ret = -ENOMEM;
288 break;
289 }
290 info->nr_deferred = (atomic_long_t *)(info + 1);
291 info->map = (void *)info->nr_deferred + defer_size;
292 info->map_nr_max = shrinker_nr_max;
293 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
294 }
295 up_write(&shrinker_rwsem);
296
297 return ret;
298 }
299
expand_shrinker_info(int new_id)300 static int expand_shrinker_info(int new_id)
301 {
302 int ret = 0;
303 int new_nr_max = round_up(new_id + 1, BITS_PER_LONG);
304 int map_size, defer_size = 0;
305 int old_map_size, old_defer_size = 0;
306 struct mem_cgroup *memcg;
307
308 if (!root_mem_cgroup)
309 goto out;
310
311 lockdep_assert_held(&shrinker_rwsem);
312
313 map_size = shrinker_map_size(new_nr_max);
314 defer_size = shrinker_defer_size(new_nr_max);
315 old_map_size = shrinker_map_size(shrinker_nr_max);
316 old_defer_size = shrinker_defer_size(shrinker_nr_max);
317
318 memcg = mem_cgroup_iter(NULL, NULL, NULL);
319 do {
320 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
321 old_map_size, old_defer_size,
322 new_nr_max);
323 if (ret) {
324 mem_cgroup_iter_break(NULL, memcg);
325 goto out;
326 }
327 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
328 out:
329 if (!ret)
330 shrinker_nr_max = new_nr_max;
331
332 return ret;
333 }
334
set_shrinker_bit(struct mem_cgroup * memcg,int nid,int shrinker_id)335 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
336 {
337 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
338 struct shrinker_info *info;
339
340 rcu_read_lock();
341 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
342 if (!WARN_ON_ONCE(shrinker_id >= info->map_nr_max)) {
343 /* Pairs with smp mb in shrink_slab() */
344 smp_mb__before_atomic();
345 set_bit(shrinker_id, info->map);
346 }
347 rcu_read_unlock();
348 }
349 }
350
351 static DEFINE_IDR(shrinker_idr);
352
prealloc_memcg_shrinker(struct shrinker * shrinker)353 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
354 {
355 int id, ret = -ENOMEM;
356
357 if (mem_cgroup_disabled())
358 return -ENOSYS;
359
360 down_write(&shrinker_rwsem);
361 /* This may call shrinker, so it must use down_read_trylock() */
362 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
363 if (id < 0)
364 goto unlock;
365
366 if (id >= shrinker_nr_max) {
367 if (expand_shrinker_info(id)) {
368 idr_remove(&shrinker_idr, id);
369 goto unlock;
370 }
371 }
372 shrinker->id = id;
373 ret = 0;
374 unlock:
375 up_write(&shrinker_rwsem);
376 return ret;
377 }
378
unregister_memcg_shrinker(struct shrinker * shrinker)379 static void unregister_memcg_shrinker(struct shrinker *shrinker)
380 {
381 int id = shrinker->id;
382
383 BUG_ON(id < 0);
384
385 lockdep_assert_held(&shrinker_rwsem);
386
387 idr_remove(&shrinker_idr, id);
388 }
389
xchg_nr_deferred_memcg(int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)390 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
391 struct mem_cgroup *memcg)
392 {
393 struct shrinker_info *info;
394
395 info = shrinker_info_protected(memcg, nid);
396 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
397 }
398
add_nr_deferred_memcg(long nr,int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)399 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
400 struct mem_cgroup *memcg)
401 {
402 struct shrinker_info *info;
403
404 info = shrinker_info_protected(memcg, nid);
405 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
406 }
407
reparent_shrinker_deferred(struct mem_cgroup * memcg)408 void reparent_shrinker_deferred(struct mem_cgroup *memcg)
409 {
410 int i, nid;
411 long nr;
412 struct mem_cgroup *parent;
413 struct shrinker_info *child_info, *parent_info;
414
415 parent = parent_mem_cgroup(memcg);
416 if (!parent)
417 parent = root_mem_cgroup;
418
419 /* Prevent from concurrent shrinker_info expand */
420 down_read(&shrinker_rwsem);
421 for_each_node(nid) {
422 child_info = shrinker_info_protected(memcg, nid);
423 parent_info = shrinker_info_protected(parent, nid);
424 for (i = 0; i < child_info->map_nr_max; i++) {
425 nr = atomic_long_read(&child_info->nr_deferred[i]);
426 atomic_long_add(nr, &parent_info->nr_deferred[i]);
427 }
428 }
429 up_read(&shrinker_rwsem);
430 }
431
432 /* Returns true for reclaim through cgroup limits or cgroup interfaces. */
cgroup_reclaim(struct scan_control * sc)433 static bool cgroup_reclaim(struct scan_control *sc)
434 {
435 return sc->target_mem_cgroup;
436 }
437
438 /*
439 * Returns true for reclaim on the root cgroup. This is true for direct
440 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
441 */
root_reclaim(struct scan_control * sc)442 static bool root_reclaim(struct scan_control *sc)
443 {
444 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
445 }
446
447 /**
448 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
449 * @sc: scan_control in question
450 *
451 * The normal page dirty throttling mechanism in balance_dirty_pages() is
452 * completely broken with the legacy memcg and direct stalling in
453 * shrink_folio_list() is used for throttling instead, which lacks all the
454 * niceties such as fairness, adaptive pausing, bandwidth proportional
455 * allocation and configurability.
456 *
457 * This function tests whether the vmscan currently in progress can assume
458 * that the normal dirty throttling mechanism is operational.
459 */
writeback_throttling_sane(struct scan_control * sc)460 static bool writeback_throttling_sane(struct scan_control *sc)
461 {
462 if (!cgroup_reclaim(sc))
463 return true;
464 #ifdef CONFIG_CGROUP_WRITEBACK
465 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
466 return true;
467 #endif
468 return false;
469 }
470 #else
prealloc_memcg_shrinker(struct shrinker * shrinker)471 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
472 {
473 return -ENOSYS;
474 }
475
unregister_memcg_shrinker(struct shrinker * shrinker)476 static void unregister_memcg_shrinker(struct shrinker *shrinker)
477 {
478 }
479
xchg_nr_deferred_memcg(int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)480 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
481 struct mem_cgroup *memcg)
482 {
483 return 0;
484 }
485
add_nr_deferred_memcg(long nr,int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)486 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
487 struct mem_cgroup *memcg)
488 {
489 return 0;
490 }
491
cgroup_reclaim(struct scan_control * sc)492 static bool cgroup_reclaim(struct scan_control *sc)
493 {
494 return false;
495 }
496
root_reclaim(struct scan_control * sc)497 static bool root_reclaim(struct scan_control *sc)
498 {
499 return true;
500 }
501
writeback_throttling_sane(struct scan_control * sc)502 static bool writeback_throttling_sane(struct scan_control *sc)
503 {
504 return true;
505 }
506 #endif
507
set_task_reclaim_state(struct task_struct * task,struct reclaim_state * rs)508 static void set_task_reclaim_state(struct task_struct *task,
509 struct reclaim_state *rs)
510 {
511 /* Check for an overwrite */
512 WARN_ON_ONCE(rs && task->reclaim_state);
513
514 /* Check for the nulling of an already-nulled member */
515 WARN_ON_ONCE(!rs && !task->reclaim_state);
516
517 task->reclaim_state = rs;
518 }
519
520 /*
521 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
522 * scan_control->nr_reclaimed.
523 */
flush_reclaim_state(struct scan_control * sc)524 static void flush_reclaim_state(struct scan_control *sc)
525 {
526 /*
527 * Currently, reclaim_state->reclaimed includes three types of pages
528 * freed outside of vmscan:
529 * (1) Slab pages.
530 * (2) Clean file pages from pruned inodes (on highmem systems).
531 * (3) XFS freed buffer pages.
532 *
533 * For all of these cases, we cannot universally link the pages to a
534 * single memcg. For example, a memcg-aware shrinker can free one object
535 * charged to the target memcg, causing an entire page to be freed.
536 * If we count the entire page as reclaimed from the memcg, we end up
537 * overestimating the reclaimed amount (potentially under-reclaiming).
538 *
539 * Only count such pages for global reclaim to prevent under-reclaiming
540 * from the target memcg; preventing unnecessary retries during memcg
541 * charging and false positives from proactive reclaim.
542 *
543 * For uncommon cases where the freed pages were actually mostly
544 * charged to the target memcg, we end up underestimating the reclaimed
545 * amount. This should be fine. The freed pages will be uncharged
546 * anyway, even if they are not counted here properly, and we will be
547 * able to make forward progress in charging (which is usually in a
548 * retry loop).
549 *
550 * We can go one step further, and report the uncharged objcg pages in
551 * memcg reclaim, to make reporting more accurate and reduce
552 * underestimation, but it's probably not worth the complexity for now.
553 */
554 if (current->reclaim_state && root_reclaim(sc)) {
555 sc->nr_reclaimed += current->reclaim_state->reclaimed;
556 current->reclaim_state->reclaimed = 0;
557 }
558 }
559
xchg_nr_deferred(struct shrinker * shrinker,struct shrink_control * sc)560 static long xchg_nr_deferred(struct shrinker *shrinker,
561 struct shrink_control *sc)
562 {
563 int nid = sc->nid;
564
565 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
566 nid = 0;
567
568 if (sc->memcg &&
569 (shrinker->flags & SHRINKER_MEMCG_AWARE))
570 return xchg_nr_deferred_memcg(nid, shrinker,
571 sc->memcg);
572
573 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
574 }
575
576
add_nr_deferred(long nr,struct shrinker * shrinker,struct shrink_control * sc)577 static long add_nr_deferred(long nr, struct shrinker *shrinker,
578 struct shrink_control *sc)
579 {
580 int nid = sc->nid;
581
582 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
583 nid = 0;
584
585 if (sc->memcg &&
586 (shrinker->flags & SHRINKER_MEMCG_AWARE))
587 return add_nr_deferred_memcg(nr, nid, shrinker,
588 sc->memcg);
589
590 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
591 }
592
can_demote(int nid,struct scan_control * sc)593 static bool can_demote(int nid, struct scan_control *sc)
594 {
595 if (!numa_demotion_enabled)
596 return false;
597 if (sc && sc->no_demotion)
598 return false;
599 if (next_demotion_node(nid) == NUMA_NO_NODE)
600 return false;
601
602 return true;
603 }
604
can_reclaim_anon_pages(struct mem_cgroup * memcg,int nid,struct scan_control * sc)605 static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
606 int nid,
607 struct scan_control *sc)
608 {
609 if (memcg == NULL) {
610 /*
611 * For non-memcg reclaim, is there
612 * space in any swap device?
613 */
614 if (get_nr_swap_pages() > 0)
615 return true;
616 } else {
617 /* Is the memcg below its swap limit? */
618 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
619 return true;
620 }
621
622 /*
623 * The page can not be swapped.
624 *
625 * Can it be reclaimed from this node via demotion?
626 */
627 return can_demote(nid, sc);
628 }
629
630 /*
631 * This misses isolated folios which are not accounted for to save counters.
632 * As the data only determines if reclaim or compaction continues, it is
633 * not expected that isolated folios will be a dominating factor.
634 */
zone_reclaimable_pages(struct zone * zone)635 unsigned long zone_reclaimable_pages(struct zone *zone)
636 {
637 unsigned long nr;
638
639 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
640 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
641 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
642 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
643 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
644
645 return nr;
646 }
647
648 /**
649 * lruvec_lru_size - Returns the number of pages on the given LRU list.
650 * @lruvec: lru vector
651 * @lru: lru to use
652 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
653 */
lruvec_lru_size(struct lruvec * lruvec,enum lru_list lru,int zone_idx)654 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
655 int zone_idx)
656 {
657 unsigned long size = 0;
658 int zid;
659
660 for (zid = 0; zid <= zone_idx; zid++) {
661 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
662
663 if (!managed_zone(zone))
664 continue;
665
666 if (!mem_cgroup_disabled())
667 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
668 else
669 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
670 }
671 return size;
672 }
673
674 /*
675 * Add a shrinker callback to be called from the vm.
676 */
__prealloc_shrinker(struct shrinker * shrinker)677 static int __prealloc_shrinker(struct shrinker *shrinker)
678 {
679 unsigned int size;
680 int err;
681
682 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
683 err = prealloc_memcg_shrinker(shrinker);
684 if (err != -ENOSYS)
685 return err;
686
687 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
688 }
689
690 size = sizeof(*shrinker->nr_deferred);
691 if (shrinker->flags & SHRINKER_NUMA_AWARE)
692 size *= nr_node_ids;
693
694 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
695 if (!shrinker->nr_deferred)
696 return -ENOMEM;
697
698 return 0;
699 }
700
701 #ifdef CONFIG_SHRINKER_DEBUG
prealloc_shrinker(struct shrinker * shrinker,const char * fmt,...)702 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
703 {
704 va_list ap;
705 int err;
706
707 va_start(ap, fmt);
708 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
709 va_end(ap);
710 if (!shrinker->name)
711 return -ENOMEM;
712
713 err = __prealloc_shrinker(shrinker);
714 if (err) {
715 kfree_const(shrinker->name);
716 shrinker->name = NULL;
717 }
718
719 return err;
720 }
721 #else
prealloc_shrinker(struct shrinker * shrinker,const char * fmt,...)722 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
723 {
724 return __prealloc_shrinker(shrinker);
725 }
726 #endif
727
free_prealloced_shrinker(struct shrinker * shrinker)728 void free_prealloced_shrinker(struct shrinker *shrinker)
729 {
730 #ifdef CONFIG_SHRINKER_DEBUG
731 kfree_const(shrinker->name);
732 shrinker->name = NULL;
733 #endif
734 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
735 down_write(&shrinker_rwsem);
736 unregister_memcg_shrinker(shrinker);
737 up_write(&shrinker_rwsem);
738 return;
739 }
740
741 kfree(shrinker->nr_deferred);
742 shrinker->nr_deferred = NULL;
743 }
744
register_shrinker_prepared(struct shrinker * shrinker)745 void register_shrinker_prepared(struct shrinker *shrinker)
746 {
747 down_write(&shrinker_rwsem);
748 list_add_tail(&shrinker->list, &shrinker_list);
749 shrinker->flags |= SHRINKER_REGISTERED;
750 shrinker_debugfs_add(shrinker);
751 up_write(&shrinker_rwsem);
752 }
753
__register_shrinker(struct shrinker * shrinker)754 static int __register_shrinker(struct shrinker *shrinker)
755 {
756 int err = __prealloc_shrinker(shrinker);
757
758 if (err)
759 return err;
760 register_shrinker_prepared(shrinker);
761 return 0;
762 }
763
764 #ifdef CONFIG_SHRINKER_DEBUG
register_shrinker(struct shrinker * shrinker,const char * fmt,...)765 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
766 {
767 va_list ap;
768 int err;
769
770 va_start(ap, fmt);
771 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
772 va_end(ap);
773 if (!shrinker->name)
774 return -ENOMEM;
775
776 err = __register_shrinker(shrinker);
777 if (err) {
778 kfree_const(shrinker->name);
779 shrinker->name = NULL;
780 }
781 return err;
782 }
783 #else
register_shrinker(struct shrinker * shrinker,const char * fmt,...)784 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
785 {
786 return __register_shrinker(shrinker);
787 }
788 #endif
789 EXPORT_SYMBOL(register_shrinker);
790
791 /*
792 * Remove one
793 */
unregister_shrinker(struct shrinker * shrinker)794 void unregister_shrinker(struct shrinker *shrinker)
795 {
796 struct dentry *debugfs_entry;
797 int debugfs_id;
798
799 if (!(shrinker->flags & SHRINKER_REGISTERED))
800 return;
801
802 down_write(&shrinker_rwsem);
803 list_del(&shrinker->list);
804 shrinker->flags &= ~SHRINKER_REGISTERED;
805 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
806 unregister_memcg_shrinker(shrinker);
807 debugfs_entry = shrinker_debugfs_detach(shrinker, &debugfs_id);
808 up_write(&shrinker_rwsem);
809
810 shrinker_debugfs_remove(debugfs_entry, debugfs_id);
811
812 kfree(shrinker->nr_deferred);
813 shrinker->nr_deferred = NULL;
814 }
815 EXPORT_SYMBOL(unregister_shrinker);
816
817 /**
818 * synchronize_shrinkers - Wait for all running shrinkers to complete.
819 *
820 * This is equivalent to calling unregister_shrink() and register_shrinker(),
821 * but atomically and with less overhead. This is useful to guarantee that all
822 * shrinker invocations have seen an update, before freeing memory, similar to
823 * rcu.
824 */
synchronize_shrinkers(void)825 void synchronize_shrinkers(void)
826 {
827 down_write(&shrinker_rwsem);
828 up_write(&shrinker_rwsem);
829 }
830 EXPORT_SYMBOL(synchronize_shrinkers);
831
832 #define SHRINK_BATCH 128
833
do_shrink_slab(struct shrink_control * shrinkctl,struct shrinker * shrinker,int priority)834 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
835 struct shrinker *shrinker, int priority)
836 {
837 unsigned long freed = 0;
838 unsigned long long delta;
839 long total_scan;
840 long freeable;
841 long nr;
842 long new_nr;
843 long batch_size = shrinker->batch ? shrinker->batch
844 : SHRINK_BATCH;
845 long scanned = 0, next_deferred;
846
847 freeable = shrinker->count_objects(shrinker, shrinkctl);
848 if (freeable == 0 || freeable == SHRINK_EMPTY)
849 return freeable;
850
851 /*
852 * copy the current shrinker scan count into a local variable
853 * and zero it so that other concurrent shrinker invocations
854 * don't also do this scanning work.
855 */
856 nr = xchg_nr_deferred(shrinker, shrinkctl);
857
858 if (shrinker->seeks) {
859 delta = freeable >> priority;
860 delta *= 4;
861 do_div(delta, shrinker->seeks);
862 } else {
863 /*
864 * These objects don't require any IO to create. Trim
865 * them aggressively under memory pressure to keep
866 * them from causing refetches in the IO caches.
867 */
868 delta = freeable / 2;
869 }
870
871 total_scan = nr >> priority;
872 total_scan += delta;
873 total_scan = min(total_scan, (2 * freeable));
874
875 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
876 freeable, delta, total_scan, priority);
877
878 /*
879 * Normally, we should not scan less than batch_size objects in one
880 * pass to avoid too frequent shrinker calls, but if the slab has less
881 * than batch_size objects in total and we are really tight on memory,
882 * we will try to reclaim all available objects, otherwise we can end
883 * up failing allocations although there are plenty of reclaimable
884 * objects spread over several slabs with usage less than the
885 * batch_size.
886 *
887 * We detect the "tight on memory" situations by looking at the total
888 * number of objects we want to scan (total_scan). If it is greater
889 * than the total number of objects on slab (freeable), we must be
890 * scanning at high prio and therefore should try to reclaim as much as
891 * possible.
892 */
893 while (total_scan >= batch_size ||
894 total_scan >= freeable) {
895 unsigned long ret;
896 unsigned long nr_to_scan = min(batch_size, total_scan);
897
898 shrinkctl->nr_to_scan = nr_to_scan;
899 shrinkctl->nr_scanned = nr_to_scan;
900 ret = shrinker->scan_objects(shrinker, shrinkctl);
901 if (ret == SHRINK_STOP)
902 break;
903 freed += ret;
904
905 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
906 total_scan -= shrinkctl->nr_scanned;
907 scanned += shrinkctl->nr_scanned;
908
909 cond_resched();
910 }
911
912 /*
913 * The deferred work is increased by any new work (delta) that wasn't
914 * done, decreased by old deferred work that was done now.
915 *
916 * And it is capped to two times of the freeable items.
917 */
918 next_deferred = max_t(long, (nr + delta - scanned), 0);
919 next_deferred = min(next_deferred, (2 * freeable));
920
921 /*
922 * move the unused scan count back into the shrinker in a
923 * manner that handles concurrent updates.
924 */
925 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
926
927 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
928 return freed;
929 }
930
931 #ifdef CONFIG_MEMCG
shrink_slab_memcg(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)932 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
933 struct mem_cgroup *memcg, int priority)
934 {
935 struct shrinker_info *info;
936 unsigned long ret, freed = 0;
937 int i;
938
939 if (!mem_cgroup_online(memcg))
940 return 0;
941
942 if (!down_read_trylock(&shrinker_rwsem))
943 return 0;
944
945 info = shrinker_info_protected(memcg, nid);
946 if (unlikely(!info))
947 goto unlock;
948
949 for_each_set_bit(i, info->map, info->map_nr_max) {
950 struct shrink_control sc = {
951 .gfp_mask = gfp_mask,
952 .nid = nid,
953 .memcg = memcg,
954 };
955 struct shrinker *shrinker;
956
957 shrinker = idr_find(&shrinker_idr, i);
958 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
959 if (!shrinker)
960 clear_bit(i, info->map);
961 continue;
962 }
963
964 /* Call non-slab shrinkers even though kmem is disabled */
965 if (!memcg_kmem_online() &&
966 !(shrinker->flags & SHRINKER_NONSLAB))
967 continue;
968
969 ret = do_shrink_slab(&sc, shrinker, priority);
970 if (ret == SHRINK_EMPTY) {
971 clear_bit(i, info->map);
972 /*
973 * After the shrinker reported that it had no objects to
974 * free, but before we cleared the corresponding bit in
975 * the memcg shrinker map, a new object might have been
976 * added. To make sure, we have the bit set in this
977 * case, we invoke the shrinker one more time and reset
978 * the bit if it reports that it is not empty anymore.
979 * The memory barrier here pairs with the barrier in
980 * set_shrinker_bit():
981 *
982 * list_lru_add() shrink_slab_memcg()
983 * list_add_tail() clear_bit()
984 * <MB> <MB>
985 * set_bit() do_shrink_slab()
986 */
987 smp_mb__after_atomic();
988 ret = do_shrink_slab(&sc, shrinker, priority);
989 if (ret == SHRINK_EMPTY)
990 ret = 0;
991 else
992 set_shrinker_bit(memcg, nid, i);
993 }
994 freed += ret;
995
996 if (rwsem_is_contended(&shrinker_rwsem)) {
997 freed = freed ? : 1;
998 break;
999 }
1000 }
1001 unlock:
1002 up_read(&shrinker_rwsem);
1003 return freed;
1004 }
1005 #else /* CONFIG_MEMCG */
shrink_slab_memcg(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)1006 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
1007 struct mem_cgroup *memcg, int priority)
1008 {
1009 return 0;
1010 }
1011 #endif /* CONFIG_MEMCG */
1012
1013 /**
1014 * shrink_slab - shrink slab caches
1015 * @gfp_mask: allocation context
1016 * @nid: node whose slab caches to target
1017 * @memcg: memory cgroup whose slab caches to target
1018 * @priority: the reclaim priority
1019 *
1020 * Call the shrink functions to age shrinkable caches.
1021 *
1022 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
1023 * unaware shrinkers will receive a node id of 0 instead.
1024 *
1025 * @memcg specifies the memory cgroup to target. Unaware shrinkers
1026 * are called only if it is the root cgroup.
1027 *
1028 * @priority is sc->priority, we take the number of objects and >> by priority
1029 * in order to get the scan target.
1030 *
1031 * Returns the number of reclaimed slab objects.
1032 */
shrink_slab(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)1033 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
1034 struct mem_cgroup *memcg,
1035 int priority)
1036 {
1037 unsigned long ret, freed = 0;
1038 struct shrinker *shrinker;
1039
1040 /*
1041 * The root memcg might be allocated even though memcg is disabled
1042 * via "cgroup_disable=memory" boot parameter. This could make
1043 * mem_cgroup_is_root() return false, then just run memcg slab
1044 * shrink, but skip global shrink. This may result in premature
1045 * oom.
1046 */
1047 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
1048 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
1049
1050 if (!down_read_trylock(&shrinker_rwsem))
1051 goto out;
1052
1053 list_for_each_entry(shrinker, &shrinker_list, list) {
1054 struct shrink_control sc = {
1055 .gfp_mask = gfp_mask,
1056 .nid = nid,
1057 .memcg = memcg,
1058 };
1059
1060 ret = do_shrink_slab(&sc, shrinker, priority);
1061 if (ret == SHRINK_EMPTY)
1062 ret = 0;
1063 freed += ret;
1064 /*
1065 * Bail out if someone want to register a new shrinker to
1066 * prevent the registration from being stalled for long periods
1067 * by parallel ongoing shrinking.
1068 */
1069 if (rwsem_is_contended(&shrinker_rwsem)) {
1070 freed = freed ? : 1;
1071 break;
1072 }
1073 }
1074
1075 up_read(&shrinker_rwsem);
1076 out:
1077 cond_resched();
1078 return freed;
1079 }
1080
drop_slab_node(int nid)1081 static unsigned long drop_slab_node(int nid)
1082 {
1083 unsigned long freed = 0;
1084 struct mem_cgroup *memcg = NULL;
1085
1086 memcg = mem_cgroup_iter(NULL, NULL, NULL);
1087 do {
1088 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1089 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1090
1091 return freed;
1092 }
1093
drop_slab(void)1094 void drop_slab(void)
1095 {
1096 int nid;
1097 int shift = 0;
1098 unsigned long freed;
1099
1100 do {
1101 freed = 0;
1102 for_each_online_node(nid) {
1103 if (fatal_signal_pending(current))
1104 return;
1105
1106 freed += drop_slab_node(nid);
1107 }
1108 } while ((freed >> shift++) > 1);
1109 }
1110
reclaimer_offset(void)1111 static int reclaimer_offset(void)
1112 {
1113 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
1114 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
1115 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
1116 PGSCAN_DIRECT - PGSCAN_KSWAPD);
1117 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
1118 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
1119 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
1120 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
1121
1122 if (current_is_kswapd())
1123 return 0;
1124 if (current_is_khugepaged())
1125 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
1126 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
1127 }
1128
is_page_cache_freeable(struct folio * folio)1129 static inline int is_page_cache_freeable(struct folio *folio)
1130 {
1131 /*
1132 * A freeable page cache folio is referenced only by the caller
1133 * that isolated the folio, the page cache and optional filesystem
1134 * private data at folio->private.
1135 */
1136 return folio_ref_count(folio) - folio_test_private(folio) ==
1137 1 + folio_nr_pages(folio);
1138 }
1139
1140 /*
1141 * We detected a synchronous write error writing a folio out. Probably
1142 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1143 * fsync(), msync() or close().
1144 *
1145 * The tricky part is that after writepage we cannot touch the mapping: nothing
1146 * prevents it from being freed up. But we have a ref on the folio and once
1147 * that folio is locked, the mapping is pinned.
1148 *
1149 * We're allowed to run sleeping folio_lock() here because we know the caller has
1150 * __GFP_FS.
1151 */
handle_write_error(struct address_space * mapping,struct folio * folio,int error)1152 static void handle_write_error(struct address_space *mapping,
1153 struct folio *folio, int error)
1154 {
1155 folio_lock(folio);
1156 if (folio_mapping(folio) == mapping)
1157 mapping_set_error(mapping, error);
1158 folio_unlock(folio);
1159 }
1160
skip_throttle_noprogress(pg_data_t * pgdat)1161 static bool skip_throttle_noprogress(pg_data_t *pgdat)
1162 {
1163 int reclaimable = 0, write_pending = 0;
1164 int i;
1165
1166 /*
1167 * If kswapd is disabled, reschedule if necessary but do not
1168 * throttle as the system is likely near OOM.
1169 */
1170 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1171 return true;
1172
1173 /*
1174 * If there are a lot of dirty/writeback folios then do not
1175 * throttle as throttling will occur when the folios cycle
1176 * towards the end of the LRU if still under writeback.
1177 */
1178 for (i = 0; i < MAX_NR_ZONES; i++) {
1179 struct zone *zone = pgdat->node_zones + i;
1180
1181 if (!managed_zone(zone))
1182 continue;
1183
1184 reclaimable += zone_reclaimable_pages(zone);
1185 write_pending += zone_page_state_snapshot(zone,
1186 NR_ZONE_WRITE_PENDING);
1187 }
1188 if (2 * write_pending <= reclaimable)
1189 return true;
1190
1191 return false;
1192 }
1193
reclaim_throttle(pg_data_t * pgdat,enum vmscan_throttle_state reason)1194 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1195 {
1196 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1197 long timeout, ret;
1198 DEFINE_WAIT(wait);
1199
1200 /*
1201 * Do not throttle user workers, kthreads other than kswapd or
1202 * workqueues. They may be required for reclaim to make
1203 * forward progress (e.g. journalling workqueues or kthreads).
1204 */
1205 if (!current_is_kswapd() &&
1206 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
1207 cond_resched();
1208 return;
1209 }
1210
1211 /*
1212 * These figures are pulled out of thin air.
1213 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1214 * parallel reclaimers which is a short-lived event so the timeout is
1215 * short. Failing to make progress or waiting on writeback are
1216 * potentially long-lived events so use a longer timeout. This is shaky
1217 * logic as a failure to make progress could be due to anything from
1218 * writeback to a slow device to excessive referenced folios at the tail
1219 * of the inactive LRU.
1220 */
1221 switch(reason) {
1222 case VMSCAN_THROTTLE_WRITEBACK:
1223 timeout = HZ/10;
1224
1225 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1226 WRITE_ONCE(pgdat->nr_reclaim_start,
1227 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1228 }
1229
1230 break;
1231 case VMSCAN_THROTTLE_CONGESTED:
1232 fallthrough;
1233 case VMSCAN_THROTTLE_NOPROGRESS:
1234 if (skip_throttle_noprogress(pgdat)) {
1235 cond_resched();
1236 return;
1237 }
1238
1239 timeout = 1;
1240
1241 break;
1242 case VMSCAN_THROTTLE_ISOLATED:
1243 timeout = HZ/50;
1244 break;
1245 default:
1246 WARN_ON_ONCE(1);
1247 timeout = HZ;
1248 break;
1249 }
1250
1251 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1252 ret = schedule_timeout(timeout);
1253 finish_wait(wqh, &wait);
1254
1255 if (reason == VMSCAN_THROTTLE_WRITEBACK)
1256 atomic_dec(&pgdat->nr_writeback_throttled);
1257
1258 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1259 jiffies_to_usecs(timeout - ret),
1260 reason);
1261 }
1262
1263 /*
1264 * Account for folios written if tasks are throttled waiting on dirty
1265 * folios to clean. If enough folios have been cleaned since throttling
1266 * started then wakeup the throttled tasks.
1267 */
__acct_reclaim_writeback(pg_data_t * pgdat,struct folio * folio,int nr_throttled)1268 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1269 int nr_throttled)
1270 {
1271 unsigned long nr_written;
1272
1273 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1274
1275 /*
1276 * This is an inaccurate read as the per-cpu deltas may not
1277 * be synchronised. However, given that the system is
1278 * writeback throttled, it is not worth taking the penalty
1279 * of getting an accurate count. At worst, the throttle
1280 * timeout guarantees forward progress.
1281 */
1282 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1283 READ_ONCE(pgdat->nr_reclaim_start);
1284
1285 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1286 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1287 }
1288
1289 /* possible outcome of pageout() */
1290 typedef enum {
1291 /* failed to write folio out, folio is locked */
1292 PAGE_KEEP,
1293 /* move folio to the active list, folio is locked */
1294 PAGE_ACTIVATE,
1295 /* folio has been sent to the disk successfully, folio is unlocked */
1296 PAGE_SUCCESS,
1297 /* folio is clean and locked */
1298 PAGE_CLEAN,
1299 } pageout_t;
1300
1301 /*
1302 * pageout is called by shrink_folio_list() for each dirty folio.
1303 * Calls ->writepage().
1304 */
pageout(struct folio * folio,struct address_space * mapping,struct swap_iocb ** plug)1305 static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1306 struct swap_iocb **plug)
1307 {
1308 /*
1309 * If the folio is dirty, only perform writeback if that write
1310 * will be non-blocking. To prevent this allocation from being
1311 * stalled by pagecache activity. But note that there may be
1312 * stalls if we need to run get_block(). We could test
1313 * PagePrivate for that.
1314 *
1315 * If this process is currently in __generic_file_write_iter() against
1316 * this folio's queue, we can perform writeback even if that
1317 * will block.
1318 *
1319 * If the folio is swapcache, write it back even if that would
1320 * block, for some throttling. This happens by accident, because
1321 * swap_backing_dev_info is bust: it doesn't reflect the
1322 * congestion state of the swapdevs. Easy to fix, if needed.
1323 */
1324 if (!is_page_cache_freeable(folio))
1325 return PAGE_KEEP;
1326 if (!mapping) {
1327 /*
1328 * Some data journaling orphaned folios can have
1329 * folio->mapping == NULL while being dirty with clean buffers.
1330 */
1331 if (folio_test_private(folio)) {
1332 if (try_to_free_buffers(folio)) {
1333 folio_clear_dirty(folio);
1334 pr_info("%s: orphaned folio\n", __func__);
1335 return PAGE_CLEAN;
1336 }
1337 }
1338 return PAGE_KEEP;
1339 }
1340 if (mapping->a_ops->writepage == NULL)
1341 return PAGE_ACTIVATE;
1342
1343 if (folio_clear_dirty_for_io(folio)) {
1344 int res;
1345 struct writeback_control wbc = {
1346 .sync_mode = WB_SYNC_NONE,
1347 .nr_to_write = SWAP_CLUSTER_MAX,
1348 .range_start = 0,
1349 .range_end = LLONG_MAX,
1350 .for_reclaim = 1,
1351 .swap_plug = plug,
1352 };
1353
1354 folio_set_reclaim(folio);
1355 res = mapping->a_ops->writepage(&folio->page, &wbc);
1356 if (res < 0)
1357 handle_write_error(mapping, folio, res);
1358 if (res == AOP_WRITEPAGE_ACTIVATE) {
1359 folio_clear_reclaim(folio);
1360 return PAGE_ACTIVATE;
1361 }
1362
1363 if (!folio_test_writeback(folio)) {
1364 /* synchronous write or broken a_ops? */
1365 folio_clear_reclaim(folio);
1366 }
1367 trace_mm_vmscan_write_folio(folio);
1368 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1369 return PAGE_SUCCESS;
1370 }
1371
1372 return PAGE_CLEAN;
1373 }
1374
1375 /*
1376 * Same as remove_mapping, but if the folio is removed from the mapping, it
1377 * gets returned with a refcount of 0.
1378 */
__remove_mapping(struct address_space * mapping,struct folio * folio,bool reclaimed,struct mem_cgroup * target_memcg)1379 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1380 bool reclaimed, struct mem_cgroup *target_memcg)
1381 {
1382 int refcount;
1383 void *shadow = NULL;
1384
1385 BUG_ON(!folio_test_locked(folio));
1386 BUG_ON(mapping != folio_mapping(folio));
1387
1388 if (!folio_test_swapcache(folio))
1389 spin_lock(&mapping->host->i_lock);
1390 xa_lock_irq(&mapping->i_pages);
1391 /*
1392 * The non racy check for a busy folio.
1393 *
1394 * Must be careful with the order of the tests. When someone has
1395 * a ref to the folio, it may be possible that they dirty it then
1396 * drop the reference. So if the dirty flag is tested before the
1397 * refcount here, then the following race may occur:
1398 *
1399 * get_user_pages(&page);
1400 * [user mapping goes away]
1401 * write_to(page);
1402 * !folio_test_dirty(folio) [good]
1403 * folio_set_dirty(folio);
1404 * folio_put(folio);
1405 * !refcount(folio) [good, discard it]
1406 *
1407 * [oops, our write_to data is lost]
1408 *
1409 * Reversing the order of the tests ensures such a situation cannot
1410 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
1411 * load is not satisfied before that of folio->_refcount.
1412 *
1413 * Note that if the dirty flag is always set via folio_mark_dirty,
1414 * and thus under the i_pages lock, then this ordering is not required.
1415 */
1416 refcount = 1 + folio_nr_pages(folio);
1417 if (!folio_ref_freeze(folio, refcount))
1418 goto cannot_free;
1419 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
1420 if (unlikely(folio_test_dirty(folio))) {
1421 folio_ref_unfreeze(folio, refcount);
1422 goto cannot_free;
1423 }
1424
1425 if (folio_test_swapcache(folio)) {
1426 swp_entry_t swap = folio->swap;
1427
1428 if (reclaimed && !mapping_exiting(mapping))
1429 shadow = workingset_eviction(folio, target_memcg);
1430 __delete_from_swap_cache(folio, swap, shadow);
1431 mem_cgroup_swapout(folio, swap);
1432 xa_unlock_irq(&mapping->i_pages);
1433 put_swap_folio(folio, swap);
1434 } else {
1435 void (*free_folio)(struct folio *);
1436
1437 free_folio = mapping->a_ops->free_folio;
1438 /*
1439 * Remember a shadow entry for reclaimed file cache in
1440 * order to detect refaults, thus thrashing, later on.
1441 *
1442 * But don't store shadows in an address space that is
1443 * already exiting. This is not just an optimization,
1444 * inode reclaim needs to empty out the radix tree or
1445 * the nodes are lost. Don't plant shadows behind its
1446 * back.
1447 *
1448 * We also don't store shadows for DAX mappings because the
1449 * only page cache folios found in these are zero pages
1450 * covering holes, and because we don't want to mix DAX
1451 * exceptional entries and shadow exceptional entries in the
1452 * same address_space.
1453 */
1454 if (reclaimed && folio_is_file_lru(folio) &&
1455 !mapping_exiting(mapping) && !dax_mapping(mapping))
1456 shadow = workingset_eviction(folio, target_memcg);
1457 __filemap_remove_folio(folio, shadow);
1458 xa_unlock_irq(&mapping->i_pages);
1459 if (mapping_shrinkable(mapping))
1460 inode_add_lru(mapping->host);
1461 spin_unlock(&mapping->host->i_lock);
1462
1463 if (free_folio)
1464 free_folio(folio);
1465 }
1466
1467 return 1;
1468
1469 cannot_free:
1470 xa_unlock_irq(&mapping->i_pages);
1471 if (!folio_test_swapcache(folio))
1472 spin_unlock(&mapping->host->i_lock);
1473 return 0;
1474 }
1475
1476 /**
1477 * remove_mapping() - Attempt to remove a folio from its mapping.
1478 * @mapping: The address space.
1479 * @folio: The folio to remove.
1480 *
1481 * If the folio is dirty, under writeback or if someone else has a ref
1482 * on it, removal will fail.
1483 * Return: The number of pages removed from the mapping. 0 if the folio
1484 * could not be removed.
1485 * Context: The caller should have a single refcount on the folio and
1486 * hold its lock.
1487 */
remove_mapping(struct address_space * mapping,struct folio * folio)1488 long remove_mapping(struct address_space *mapping, struct folio *folio)
1489 {
1490 if (__remove_mapping(mapping, folio, false, NULL)) {
1491 /*
1492 * Unfreezing the refcount with 1 effectively
1493 * drops the pagecache ref for us without requiring another
1494 * atomic operation.
1495 */
1496 folio_ref_unfreeze(folio, 1);
1497 return folio_nr_pages(folio);
1498 }
1499 return 0;
1500 }
1501
1502 /**
1503 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1504 * @folio: Folio to be returned to an LRU list.
1505 *
1506 * Add previously isolated @folio to appropriate LRU list.
1507 * The folio may still be unevictable for other reasons.
1508 *
1509 * Context: lru_lock must not be held, interrupts must be enabled.
1510 */
folio_putback_lru(struct folio * folio)1511 void folio_putback_lru(struct folio *folio)
1512 {
1513 folio_add_lru(folio);
1514 folio_put(folio); /* drop ref from isolate */
1515 }
1516
1517 enum folio_references {
1518 FOLIOREF_RECLAIM,
1519 FOLIOREF_RECLAIM_CLEAN,
1520 FOLIOREF_KEEP,
1521 FOLIOREF_ACTIVATE,
1522 };
1523
folio_check_references(struct folio * folio,struct scan_control * sc)1524 static enum folio_references folio_check_references(struct folio *folio,
1525 struct scan_control *sc)
1526 {
1527 int referenced_ptes, referenced_folio;
1528 unsigned long vm_flags;
1529
1530 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1531 &vm_flags);
1532 referenced_folio = folio_test_clear_referenced(folio);
1533
1534 /*
1535 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1536 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1537 */
1538 if (vm_flags & VM_LOCKED)
1539 return FOLIOREF_ACTIVATE;
1540
1541 /* rmap lock contention: rotate */
1542 if (referenced_ptes == -1)
1543 return FOLIOREF_KEEP;
1544
1545 if (referenced_ptes) {
1546 /*
1547 * All mapped folios start out with page table
1548 * references from the instantiating fault, so we need
1549 * to look twice if a mapped file/anon folio is used more
1550 * than once.
1551 *
1552 * Mark it and spare it for another trip around the
1553 * inactive list. Another page table reference will
1554 * lead to its activation.
1555 *
1556 * Note: the mark is set for activated folios as well
1557 * so that recently deactivated but used folios are
1558 * quickly recovered.
1559 */
1560 folio_set_referenced(folio);
1561
1562 if (referenced_folio || referenced_ptes > 1)
1563 return FOLIOREF_ACTIVATE;
1564
1565 /*
1566 * Activate file-backed executable folios after first usage.
1567 */
1568 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1569 return FOLIOREF_ACTIVATE;
1570
1571 return FOLIOREF_KEEP;
1572 }
1573
1574 /* Reclaim if clean, defer dirty folios to writeback */
1575 if (referenced_folio && folio_is_file_lru(folio))
1576 return FOLIOREF_RECLAIM_CLEAN;
1577
1578 return FOLIOREF_RECLAIM;
1579 }
1580
1581 /* Check if a folio is dirty or under writeback */
folio_check_dirty_writeback(struct folio * folio,bool * dirty,bool * writeback)1582 static void folio_check_dirty_writeback(struct folio *folio,
1583 bool *dirty, bool *writeback)
1584 {
1585 struct address_space *mapping;
1586
1587 /*
1588 * Anonymous folios are not handled by flushers and must be written
1589 * from reclaim context. Do not stall reclaim based on them.
1590 * MADV_FREE anonymous folios are put into inactive file list too.
1591 * They could be mistakenly treated as file lru. So further anon
1592 * test is needed.
1593 */
1594 if (!folio_is_file_lru(folio) ||
1595 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1596 *dirty = false;
1597 *writeback = false;
1598 return;
1599 }
1600
1601 /* By default assume that the folio flags are accurate */
1602 *dirty = folio_test_dirty(folio);
1603 *writeback = folio_test_writeback(folio);
1604
1605 /* Verify dirty/writeback state if the filesystem supports it */
1606 if (!folio_test_private(folio))
1607 return;
1608
1609 mapping = folio_mapping(folio);
1610 if (mapping && mapping->a_ops->is_dirty_writeback)
1611 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1612 }
1613
alloc_demote_folio(struct folio * src,unsigned long private)1614 static struct folio *alloc_demote_folio(struct folio *src,
1615 unsigned long private)
1616 {
1617 struct folio *dst;
1618 nodemask_t *allowed_mask;
1619 struct migration_target_control *mtc;
1620
1621 mtc = (struct migration_target_control *)private;
1622
1623 allowed_mask = mtc->nmask;
1624 /*
1625 * make sure we allocate from the target node first also trying to
1626 * demote or reclaim pages from the target node via kswapd if we are
1627 * low on free memory on target node. If we don't do this and if
1628 * we have free memory on the slower(lower) memtier, we would start
1629 * allocating pages from slower(lower) memory tiers without even forcing
1630 * a demotion of cold pages from the target memtier. This can result
1631 * in the kernel placing hot pages in slower(lower) memory tiers.
1632 */
1633 mtc->nmask = NULL;
1634 mtc->gfp_mask |= __GFP_THISNODE;
1635 dst = alloc_migration_target(src, (unsigned long)mtc);
1636 if (dst)
1637 return dst;
1638
1639 mtc->gfp_mask &= ~__GFP_THISNODE;
1640 mtc->nmask = allowed_mask;
1641
1642 return alloc_migration_target(src, (unsigned long)mtc);
1643 }
1644
1645 /*
1646 * Take folios on @demote_folios and attempt to demote them to another node.
1647 * Folios which are not demoted are left on @demote_folios.
1648 */
demote_folio_list(struct list_head * demote_folios,struct pglist_data * pgdat)1649 static unsigned int demote_folio_list(struct list_head *demote_folios,
1650 struct pglist_data *pgdat)
1651 {
1652 int target_nid = next_demotion_node(pgdat->node_id);
1653 unsigned int nr_succeeded;
1654 nodemask_t allowed_mask;
1655
1656 struct migration_target_control mtc = {
1657 /*
1658 * Allocate from 'node', or fail quickly and quietly.
1659 * When this happens, 'page' will likely just be discarded
1660 * instead of migrated.
1661 */
1662 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1663 __GFP_NOMEMALLOC | GFP_NOWAIT,
1664 .nid = target_nid,
1665 .nmask = &allowed_mask
1666 };
1667
1668 if (list_empty(demote_folios))
1669 return 0;
1670
1671 if (target_nid == NUMA_NO_NODE)
1672 return 0;
1673
1674 node_get_allowed_targets(pgdat, &allowed_mask);
1675
1676 /* Demotion ignores all cpuset and mempolicy settings */
1677 migrate_pages(demote_folios, alloc_demote_folio, NULL,
1678 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1679 &nr_succeeded);
1680
1681 __count_vm_events(PGDEMOTE_KSWAPD + reclaimer_offset(), nr_succeeded);
1682
1683 return nr_succeeded;
1684 }
1685
may_enter_fs(struct folio * folio,gfp_t gfp_mask)1686 static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1687 {
1688 if (gfp_mask & __GFP_FS)
1689 return true;
1690 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1691 return false;
1692 /*
1693 * We can "enter_fs" for swap-cache with only __GFP_IO
1694 * providing this isn't SWP_FS_OPS.
1695 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1696 * but that will never affect SWP_FS_OPS, so the data_race
1697 * is safe.
1698 */
1699 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1700 }
1701
1702 /*
1703 * shrink_folio_list() returns the number of reclaimed pages
1704 */
shrink_folio_list(struct list_head * folio_list,struct pglist_data * pgdat,struct scan_control * sc,struct reclaim_stat * stat,bool ignore_references)1705 static unsigned int shrink_folio_list(struct list_head *folio_list,
1706 struct pglist_data *pgdat, struct scan_control *sc,
1707 struct reclaim_stat *stat, bool ignore_references)
1708 {
1709 LIST_HEAD(ret_folios);
1710 LIST_HEAD(free_folios);
1711 LIST_HEAD(demote_folios);
1712 unsigned int nr_reclaimed = 0;
1713 unsigned int pgactivate = 0;
1714 bool do_demote_pass;
1715 struct swap_iocb *plug = NULL;
1716
1717 memset(stat, 0, sizeof(*stat));
1718 cond_resched();
1719 do_demote_pass = can_demote(pgdat->node_id, sc);
1720
1721 retry:
1722 while (!list_empty(folio_list)) {
1723 struct address_space *mapping;
1724 struct folio *folio;
1725 enum folio_references references = FOLIOREF_RECLAIM;
1726 bool dirty, writeback;
1727 unsigned int nr_pages;
1728
1729 cond_resched();
1730
1731 folio = lru_to_folio(folio_list);
1732 list_del(&folio->lru);
1733
1734 if (!folio_trylock(folio))
1735 goto keep;
1736
1737 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1738
1739 nr_pages = folio_nr_pages(folio);
1740
1741 /* Account the number of base pages */
1742 sc->nr_scanned += nr_pages;
1743
1744 if (unlikely(!folio_evictable(folio)))
1745 goto activate_locked;
1746
1747 if (!sc->may_unmap && folio_mapped(folio))
1748 goto keep_locked;
1749
1750 /* folio_update_gen() tried to promote this page? */
1751 if (lru_gen_enabled() && !ignore_references &&
1752 folio_mapped(folio) && folio_test_referenced(folio))
1753 goto keep_locked;
1754
1755 /*
1756 * The number of dirty pages determines if a node is marked
1757 * reclaim_congested. kswapd will stall and start writing
1758 * folios if the tail of the LRU is all dirty unqueued folios.
1759 */
1760 folio_check_dirty_writeback(folio, &dirty, &writeback);
1761 if (dirty || writeback)
1762 stat->nr_dirty += nr_pages;
1763
1764 if (dirty && !writeback)
1765 stat->nr_unqueued_dirty += nr_pages;
1766
1767 /*
1768 * Treat this folio as congested if folios are cycling
1769 * through the LRU so quickly that the folios marked
1770 * for immediate reclaim are making it to the end of
1771 * the LRU a second time.
1772 */
1773 if (writeback && folio_test_reclaim(folio))
1774 stat->nr_congested += nr_pages;
1775
1776 /*
1777 * If a folio at the tail of the LRU is under writeback, there
1778 * are three cases to consider.
1779 *
1780 * 1) If reclaim is encountering an excessive number
1781 * of folios under writeback and this folio has both
1782 * the writeback and reclaim flags set, then it
1783 * indicates that folios are being queued for I/O but
1784 * are being recycled through the LRU before the I/O
1785 * can complete. Waiting on the folio itself risks an
1786 * indefinite stall if it is impossible to writeback
1787 * the folio due to I/O error or disconnected storage
1788 * so instead note that the LRU is being scanned too
1789 * quickly and the caller can stall after the folio
1790 * list has been processed.
1791 *
1792 * 2) Global or new memcg reclaim encounters a folio that is
1793 * not marked for immediate reclaim, or the caller does not
1794 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1795 * not to fs). In this case mark the folio for immediate
1796 * reclaim and continue scanning.
1797 *
1798 * Require may_enter_fs() because we would wait on fs, which
1799 * may not have submitted I/O yet. And the loop driver might
1800 * enter reclaim, and deadlock if it waits on a folio for
1801 * which it is needed to do the write (loop masks off
1802 * __GFP_IO|__GFP_FS for this reason); but more thought
1803 * would probably show more reasons.
1804 *
1805 * 3) Legacy memcg encounters a folio that already has the
1806 * reclaim flag set. memcg does not have any dirty folio
1807 * throttling so we could easily OOM just because too many
1808 * folios are in writeback and there is nothing else to
1809 * reclaim. Wait for the writeback to complete.
1810 *
1811 * In cases 1) and 2) we activate the folios to get them out of
1812 * the way while we continue scanning for clean folios on the
1813 * inactive list and refilling from the active list. The
1814 * observation here is that waiting for disk writes is more
1815 * expensive than potentially causing reloads down the line.
1816 * Since they're marked for immediate reclaim, they won't put
1817 * memory pressure on the cache working set any longer than it
1818 * takes to write them to disk.
1819 */
1820 if (folio_test_writeback(folio)) {
1821 /* Case 1 above */
1822 if (current_is_kswapd() &&
1823 folio_test_reclaim(folio) &&
1824 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1825 stat->nr_immediate += nr_pages;
1826 goto activate_locked;
1827
1828 /* Case 2 above */
1829 } else if (writeback_throttling_sane(sc) ||
1830 !folio_test_reclaim(folio) ||
1831 !may_enter_fs(folio, sc->gfp_mask)) {
1832 /*
1833 * This is slightly racy -
1834 * folio_end_writeback() might have
1835 * just cleared the reclaim flag, then
1836 * setting the reclaim flag here ends up
1837 * interpreted as the readahead flag - but
1838 * that does not matter enough to care.
1839 * What we do want is for this folio to
1840 * have the reclaim flag set next time
1841 * memcg reclaim reaches the tests above,
1842 * so it will then wait for writeback to
1843 * avoid OOM; and it's also appropriate
1844 * in global reclaim.
1845 */
1846 folio_set_reclaim(folio);
1847 stat->nr_writeback += nr_pages;
1848 goto activate_locked;
1849
1850 /* Case 3 above */
1851 } else {
1852 folio_unlock(folio);
1853 folio_wait_writeback(folio);
1854 /* then go back and try same folio again */
1855 list_add_tail(&folio->lru, folio_list);
1856 continue;
1857 }
1858 }
1859
1860 if (!ignore_references)
1861 references = folio_check_references(folio, sc);
1862
1863 switch (references) {
1864 case FOLIOREF_ACTIVATE:
1865 goto activate_locked;
1866 case FOLIOREF_KEEP:
1867 stat->nr_ref_keep += nr_pages;
1868 goto keep_locked;
1869 case FOLIOREF_RECLAIM:
1870 case FOLIOREF_RECLAIM_CLEAN:
1871 ; /* try to reclaim the folio below */
1872 }
1873
1874 /*
1875 * Before reclaiming the folio, try to relocate
1876 * its contents to another node.
1877 */
1878 if (do_demote_pass &&
1879 (thp_migration_supported() || !folio_test_large(folio))) {
1880 list_add(&folio->lru, &demote_folios);
1881 folio_unlock(folio);
1882 continue;
1883 }
1884
1885 /*
1886 * Anonymous process memory has backing store?
1887 * Try to allocate it some swap space here.
1888 * Lazyfree folio could be freed directly
1889 */
1890 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1891 if (!folio_test_swapcache(folio)) {
1892 if (!(sc->gfp_mask & __GFP_IO))
1893 goto keep_locked;
1894 if (folio_maybe_dma_pinned(folio))
1895 goto keep_locked;
1896 if (folio_test_large(folio)) {
1897 /* cannot split folio, skip it */
1898 if (!can_split_folio(folio, NULL))
1899 goto activate_locked;
1900 /*
1901 * Split folios without a PMD map right
1902 * away. Chances are some or all of the
1903 * tail pages can be freed without IO.
1904 */
1905 if (!folio_entire_mapcount(folio) &&
1906 split_folio_to_list(folio,
1907 folio_list))
1908 goto activate_locked;
1909 }
1910 if (!add_to_swap(folio)) {
1911 if (!folio_test_large(folio))
1912 goto activate_locked_split;
1913 /* Fallback to swap normal pages */
1914 if (split_folio_to_list(folio,
1915 folio_list))
1916 goto activate_locked;
1917 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1918 count_vm_event(THP_SWPOUT_FALLBACK);
1919 #endif
1920 if (!add_to_swap(folio))
1921 goto activate_locked_split;
1922 }
1923 }
1924 } else if (folio_test_swapbacked(folio) &&
1925 folio_test_large(folio)) {
1926 /* Split shmem folio */
1927 if (split_folio_to_list(folio, folio_list))
1928 goto keep_locked;
1929 }
1930
1931 /*
1932 * If the folio was split above, the tail pages will make
1933 * their own pass through this function and be accounted
1934 * then.
1935 */
1936 if ((nr_pages > 1) && !folio_test_large(folio)) {
1937 sc->nr_scanned -= (nr_pages - 1);
1938 nr_pages = 1;
1939 }
1940
1941 /*
1942 * The folio is mapped into the page tables of one or more
1943 * processes. Try to unmap it here.
1944 */
1945 if (folio_mapped(folio)) {
1946 enum ttu_flags flags = TTU_BATCH_FLUSH;
1947 bool was_swapbacked = folio_test_swapbacked(folio);
1948
1949 if (folio_test_pmd_mappable(folio))
1950 flags |= TTU_SPLIT_HUGE_PMD;
1951
1952 try_to_unmap(folio, flags);
1953 if (folio_mapped(folio)) {
1954 stat->nr_unmap_fail += nr_pages;
1955 if (!was_swapbacked &&
1956 folio_test_swapbacked(folio))
1957 stat->nr_lazyfree_fail += nr_pages;
1958 goto activate_locked;
1959 }
1960 }
1961
1962 /*
1963 * Folio is unmapped now so it cannot be newly pinned anymore.
1964 * No point in trying to reclaim folio if it is pinned.
1965 * Furthermore we don't want to reclaim underlying fs metadata
1966 * if the folio is pinned and thus potentially modified by the
1967 * pinning process as that may upset the filesystem.
1968 */
1969 if (folio_maybe_dma_pinned(folio))
1970 goto activate_locked;
1971
1972 mapping = folio_mapping(folio);
1973 if (folio_test_dirty(folio)) {
1974 /*
1975 * Only kswapd can writeback filesystem folios
1976 * to avoid risk of stack overflow. But avoid
1977 * injecting inefficient single-folio I/O into
1978 * flusher writeback as much as possible: only
1979 * write folios when we've encountered many
1980 * dirty folios, and when we've already scanned
1981 * the rest of the LRU for clean folios and see
1982 * the same dirty folios again (with the reclaim
1983 * flag set).
1984 */
1985 if (folio_is_file_lru(folio) &&
1986 (!current_is_kswapd() ||
1987 !folio_test_reclaim(folio) ||
1988 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1989 /*
1990 * Immediately reclaim when written back.
1991 * Similar in principle to folio_deactivate()
1992 * except we already have the folio isolated
1993 * and know it's dirty
1994 */
1995 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1996 nr_pages);
1997 folio_set_reclaim(folio);
1998
1999 goto activate_locked;
2000 }
2001
2002 if (references == FOLIOREF_RECLAIM_CLEAN)
2003 goto keep_locked;
2004 if (!may_enter_fs(folio, sc->gfp_mask))
2005 goto keep_locked;
2006 if (!sc->may_writepage)
2007 goto keep_locked;
2008
2009 /*
2010 * Folio is dirty. Flush the TLB if a writable entry
2011 * potentially exists to avoid CPU writes after I/O
2012 * starts and then write it out here.
2013 */
2014 try_to_unmap_flush_dirty();
2015 switch (pageout(folio, mapping, &plug)) {
2016 case PAGE_KEEP:
2017 goto keep_locked;
2018 case PAGE_ACTIVATE:
2019 goto activate_locked;
2020 case PAGE_SUCCESS:
2021 stat->nr_pageout += nr_pages;
2022
2023 if (folio_test_writeback(folio))
2024 goto keep;
2025 if (folio_test_dirty(folio))
2026 goto keep;
2027
2028 /*
2029 * A synchronous write - probably a ramdisk. Go
2030 * ahead and try to reclaim the folio.
2031 */
2032 if (!folio_trylock(folio))
2033 goto keep;
2034 if (folio_test_dirty(folio) ||
2035 folio_test_writeback(folio))
2036 goto keep_locked;
2037 mapping = folio_mapping(folio);
2038 fallthrough;
2039 case PAGE_CLEAN:
2040 ; /* try to free the folio below */
2041 }
2042 }
2043
2044 /*
2045 * If the folio has buffers, try to free the buffer
2046 * mappings associated with this folio. If we succeed
2047 * we try to free the folio as well.
2048 *
2049 * We do this even if the folio is dirty.
2050 * filemap_release_folio() does not perform I/O, but it
2051 * is possible for a folio to have the dirty flag set,
2052 * but it is actually clean (all its buffers are clean).
2053 * This happens if the buffers were written out directly,
2054 * with submit_bh(). ext3 will do this, as well as
2055 * the blockdev mapping. filemap_release_folio() will
2056 * discover that cleanness and will drop the buffers
2057 * and mark the folio clean - it can be freed.
2058 *
2059 * Rarely, folios can have buffers and no ->mapping.
2060 * These are the folios which were not successfully
2061 * invalidated in truncate_cleanup_folio(). We try to
2062 * drop those buffers here and if that worked, and the
2063 * folio is no longer mapped into process address space
2064 * (refcount == 1) it can be freed. Otherwise, leave
2065 * the folio on the LRU so it is swappable.
2066 */
2067 if (folio_needs_release(folio)) {
2068 if (!filemap_release_folio(folio, sc->gfp_mask))
2069 goto activate_locked;
2070 if (!mapping && folio_ref_count(folio) == 1) {
2071 folio_unlock(folio);
2072 if (folio_put_testzero(folio))
2073 goto free_it;
2074 else {
2075 /*
2076 * rare race with speculative reference.
2077 * the speculative reference will free
2078 * this folio shortly, so we may
2079 * increment nr_reclaimed here (and
2080 * leave it off the LRU).
2081 */
2082 nr_reclaimed += nr_pages;
2083 continue;
2084 }
2085 }
2086 }
2087
2088 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
2089 /* follow __remove_mapping for reference */
2090 if (!folio_ref_freeze(folio, 1))
2091 goto keep_locked;
2092 /*
2093 * The folio has only one reference left, which is
2094 * from the isolation. After the caller puts the
2095 * folio back on the lru and drops the reference, the
2096 * folio will be freed anyway. It doesn't matter
2097 * which lru it goes on. So we don't bother checking
2098 * the dirty flag here.
2099 */
2100 count_vm_events(PGLAZYFREED, nr_pages);
2101 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
2102 } else if (!mapping || !__remove_mapping(mapping, folio, true,
2103 sc->target_mem_cgroup))
2104 goto keep_locked;
2105
2106 folio_unlock(folio);
2107 free_it:
2108 /*
2109 * Folio may get swapped out as a whole, need to account
2110 * all pages in it.
2111 */
2112 nr_reclaimed += nr_pages;
2113
2114 /*
2115 * Is there need to periodically free_folio_list? It would
2116 * appear not as the counts should be low
2117 */
2118 if (unlikely(folio_test_large(folio)))
2119 destroy_large_folio(folio);
2120 else
2121 list_add(&folio->lru, &free_folios);
2122 continue;
2123
2124 activate_locked_split:
2125 /*
2126 * The tail pages that are failed to add into swap cache
2127 * reach here. Fixup nr_scanned and nr_pages.
2128 */
2129 if (nr_pages > 1) {
2130 sc->nr_scanned -= (nr_pages - 1);
2131 nr_pages = 1;
2132 }
2133 activate_locked:
2134 /* Not a candidate for swapping, so reclaim swap space. */
2135 if (folio_test_swapcache(folio) &&
2136 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
2137 folio_free_swap(folio);
2138 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2139 if (!folio_test_mlocked(folio)) {
2140 int type = folio_is_file_lru(folio);
2141 folio_set_active(folio);
2142 stat->nr_activate[type] += nr_pages;
2143 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2144 }
2145 keep_locked:
2146 folio_unlock(folio);
2147 keep:
2148 list_add(&folio->lru, &ret_folios);
2149 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2150 folio_test_unevictable(folio), folio);
2151 }
2152 /* 'folio_list' is always empty here */
2153
2154 /* Migrate folios selected for demotion */
2155 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
2156 /* Folios that could not be demoted are still in @demote_folios */
2157 if (!list_empty(&demote_folios)) {
2158 /* Folios which weren't demoted go back on @folio_list */
2159 list_splice_init(&demote_folios, folio_list);
2160
2161 /*
2162 * goto retry to reclaim the undemoted folios in folio_list if
2163 * desired.
2164 *
2165 * Reclaiming directly from top tier nodes is not often desired
2166 * due to it breaking the LRU ordering: in general memory
2167 * should be reclaimed from lower tier nodes and demoted from
2168 * top tier nodes.
2169 *
2170 * However, disabling reclaim from top tier nodes entirely
2171 * would cause ooms in edge scenarios where lower tier memory
2172 * is unreclaimable for whatever reason, eg memory being
2173 * mlocked or too hot to reclaim. We can disable reclaim
2174 * from top tier nodes in proactive reclaim though as that is
2175 * not real memory pressure.
2176 */
2177 if (!sc->proactive) {
2178 do_demote_pass = false;
2179 goto retry;
2180 }
2181 }
2182
2183 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2184
2185 mem_cgroup_uncharge_list(&free_folios);
2186 try_to_unmap_flush();
2187 free_unref_page_list(&free_folios);
2188
2189 list_splice(&ret_folios, folio_list);
2190 count_vm_events(PGACTIVATE, pgactivate);
2191
2192 if (plug)
2193 swap_write_unplug(plug);
2194 return nr_reclaimed;
2195 }
2196
reclaim_clean_pages_from_list(struct zone * zone,struct list_head * folio_list)2197 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2198 struct list_head *folio_list)
2199 {
2200 struct scan_control sc = {
2201 .gfp_mask = GFP_KERNEL,
2202 .may_unmap = 1,
2203 };
2204 struct reclaim_stat stat;
2205 unsigned int nr_reclaimed;
2206 struct folio *folio, *next;
2207 LIST_HEAD(clean_folios);
2208 unsigned int noreclaim_flag;
2209
2210 list_for_each_entry_safe(folio, next, folio_list, lru) {
2211 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2212 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2213 !folio_test_unevictable(folio)) {
2214 folio_clear_active(folio);
2215 list_move(&folio->lru, &clean_folios);
2216 }
2217 }
2218
2219 /*
2220 * We should be safe here since we are only dealing with file pages and
2221 * we are not kswapd and therefore cannot write dirty file pages. But
2222 * call memalloc_noreclaim_save() anyway, just in case these conditions
2223 * change in the future.
2224 */
2225 noreclaim_flag = memalloc_noreclaim_save();
2226 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
2227 &stat, true);
2228 memalloc_noreclaim_restore(noreclaim_flag);
2229
2230 list_splice(&clean_folios, folio_list);
2231 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2232 -(long)nr_reclaimed);
2233 /*
2234 * Since lazyfree pages are isolated from file LRU from the beginning,
2235 * they will rotate back to anonymous LRU in the end if it failed to
2236 * discard so isolated count will be mismatched.
2237 * Compensate the isolated count for both LRU lists.
2238 */
2239 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2240 stat.nr_lazyfree_fail);
2241 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2242 -(long)stat.nr_lazyfree_fail);
2243 return nr_reclaimed;
2244 }
2245
2246 /*
2247 * Update LRU sizes after isolating pages. The LRU size updates must
2248 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2249 */
update_lru_sizes(struct lruvec * lruvec,enum lru_list lru,unsigned long * nr_zone_taken)2250 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2251 enum lru_list lru, unsigned long *nr_zone_taken)
2252 {
2253 int zid;
2254
2255 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2256 if (!nr_zone_taken[zid])
2257 continue;
2258
2259 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2260 }
2261
2262 }
2263
2264 #ifdef CONFIG_CMA
2265 /*
2266 * It is waste of effort to scan and reclaim CMA pages if it is not available
2267 * for current allocation context. Kswapd can not be enrolled as it can not
2268 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
2269 */
skip_cma(struct folio * folio,struct scan_control * sc)2270 static bool skip_cma(struct folio *folio, struct scan_control *sc)
2271 {
2272 return !current_is_kswapd() &&
2273 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
2274 get_pageblock_migratetype(&folio->page) == MIGRATE_CMA;
2275 }
2276 #else
skip_cma(struct folio * folio,struct scan_control * sc)2277 static bool skip_cma(struct folio *folio, struct scan_control *sc)
2278 {
2279 return false;
2280 }
2281 #endif
2282
2283 /*
2284 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2285 *
2286 * lruvec->lru_lock is heavily contended. Some of the functions that
2287 * shrink the lists perform better by taking out a batch of pages
2288 * and working on them outside the LRU lock.
2289 *
2290 * For pagecache intensive workloads, this function is the hottest
2291 * spot in the kernel (apart from copy_*_user functions).
2292 *
2293 * Lru_lock must be held before calling this function.
2294 *
2295 * @nr_to_scan: The number of eligible pages to look through on the list.
2296 * @lruvec: The LRU vector to pull pages from.
2297 * @dst: The temp list to put pages on to.
2298 * @nr_scanned: The number of pages that were scanned.
2299 * @sc: The scan_control struct for this reclaim session
2300 * @lru: LRU list id for isolating
2301 *
2302 * returns how many pages were moved onto *@dst.
2303 */
isolate_lru_folios(unsigned long nr_to_scan,struct lruvec * lruvec,struct list_head * dst,unsigned long * nr_scanned,struct scan_control * sc,enum lru_list lru)2304 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
2305 struct lruvec *lruvec, struct list_head *dst,
2306 unsigned long *nr_scanned, struct scan_control *sc,
2307 enum lru_list lru)
2308 {
2309 struct list_head *src = &lruvec->lists[lru];
2310 unsigned long nr_taken = 0;
2311 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2312 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2313 unsigned long skipped = 0;
2314 unsigned long scan, total_scan, nr_pages;
2315 LIST_HEAD(folios_skipped);
2316
2317 total_scan = 0;
2318 scan = 0;
2319 while (scan < nr_to_scan && !list_empty(src)) {
2320 struct list_head *move_to = src;
2321 struct folio *folio;
2322
2323 folio = lru_to_folio(src);
2324 prefetchw_prev_lru_folio(folio, src, flags);
2325
2326 nr_pages = folio_nr_pages(folio);
2327 total_scan += nr_pages;
2328
2329 if (folio_zonenum(folio) > sc->reclaim_idx ||
2330 skip_cma(folio, sc)) {
2331 nr_skipped[folio_zonenum(folio)] += nr_pages;
2332 move_to = &folios_skipped;
2333 goto move;
2334 }
2335
2336 /*
2337 * Do not count skipped folios because that makes the function
2338 * return with no isolated folios if the LRU mostly contains
2339 * ineligible folios. This causes the VM to not reclaim any
2340 * folios, triggering a premature OOM.
2341 * Account all pages in a folio.
2342 */
2343 scan += nr_pages;
2344
2345 if (!folio_test_lru(folio))
2346 goto move;
2347 if (!sc->may_unmap && folio_mapped(folio))
2348 goto move;
2349
2350 /*
2351 * Be careful not to clear the lru flag until after we're
2352 * sure the folio is not being freed elsewhere -- the
2353 * folio release code relies on it.
2354 */
2355 if (unlikely(!folio_try_get(folio)))
2356 goto move;
2357
2358 if (!folio_test_clear_lru(folio)) {
2359 /* Another thread is already isolating this folio */
2360 folio_put(folio);
2361 goto move;
2362 }
2363
2364 nr_taken += nr_pages;
2365 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2366 move_to = dst;
2367 move:
2368 list_move(&folio->lru, move_to);
2369 }
2370
2371 /*
2372 * Splice any skipped folios to the start of the LRU list. Note that
2373 * this disrupts the LRU order when reclaiming for lower zones but
2374 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2375 * scanning would soon rescan the same folios to skip and waste lots
2376 * of cpu cycles.
2377 */
2378 if (!list_empty(&folios_skipped)) {
2379 int zid;
2380
2381 list_splice(&folios_skipped, src);
2382 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2383 if (!nr_skipped[zid])
2384 continue;
2385
2386 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2387 skipped += nr_skipped[zid];
2388 }
2389 }
2390 *nr_scanned = total_scan;
2391 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2392 total_scan, skipped, nr_taken,
2393 sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2394 update_lru_sizes(lruvec, lru, nr_zone_taken);
2395 return nr_taken;
2396 }
2397
2398 /**
2399 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2400 * @folio: Folio to isolate from its LRU list.
2401 *
2402 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2403 * corresponding to whatever LRU list the folio was on.
2404 *
2405 * The folio will have its LRU flag cleared. If it was found on the
2406 * active list, it will have the Active flag set. If it was found on the
2407 * unevictable list, it will have the Unevictable flag set. These flags
2408 * may need to be cleared by the caller before letting the page go.
2409 *
2410 * Context:
2411 *
2412 * (1) Must be called with an elevated refcount on the folio. This is a
2413 * fundamental difference from isolate_lru_folios() (which is called
2414 * without a stable reference).
2415 * (2) The lru_lock must not be held.
2416 * (3) Interrupts must be enabled.
2417 *
2418 * Return: true if the folio was removed from an LRU list.
2419 * false if the folio was not on an LRU list.
2420 */
folio_isolate_lru(struct folio * folio)2421 bool folio_isolate_lru(struct folio *folio)
2422 {
2423 bool ret = false;
2424
2425 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2426
2427 if (folio_test_clear_lru(folio)) {
2428 struct lruvec *lruvec;
2429
2430 folio_get(folio);
2431 lruvec = folio_lruvec_lock_irq(folio);
2432 lruvec_del_folio(lruvec, folio);
2433 unlock_page_lruvec_irq(lruvec);
2434 ret = true;
2435 }
2436
2437 return ret;
2438 }
2439
2440 /*
2441 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2442 * then get rescheduled. When there are massive number of tasks doing page
2443 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2444 * the LRU list will go small and be scanned faster than necessary, leading to
2445 * unnecessary swapping, thrashing and OOM.
2446 */
too_many_isolated(struct pglist_data * pgdat,int file,struct scan_control * sc)2447 static int too_many_isolated(struct pglist_data *pgdat, int file,
2448 struct scan_control *sc)
2449 {
2450 unsigned long inactive, isolated;
2451 bool too_many;
2452
2453 if (current_is_kswapd())
2454 return 0;
2455
2456 if (!writeback_throttling_sane(sc))
2457 return 0;
2458
2459 if (file) {
2460 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2461 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2462 } else {
2463 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2464 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2465 }
2466
2467 /*
2468 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2469 * won't get blocked by normal direct-reclaimers, forming a circular
2470 * deadlock.
2471 */
2472 if (gfp_has_io_fs(sc->gfp_mask))
2473 inactive >>= 3;
2474
2475 too_many = isolated > inactive;
2476
2477 /* Wake up tasks throttled due to too_many_isolated. */
2478 if (!too_many)
2479 wake_throttle_isolated(pgdat);
2480
2481 return too_many;
2482 }
2483
2484 /*
2485 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
2486 * On return, @list is reused as a list of folios to be freed by the caller.
2487 *
2488 * Returns the number of pages moved to the given lruvec.
2489 */
move_folios_to_lru(struct lruvec * lruvec,struct list_head * list)2490 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
2491 struct list_head *list)
2492 {
2493 int nr_pages, nr_moved = 0;
2494 LIST_HEAD(folios_to_free);
2495
2496 while (!list_empty(list)) {
2497 struct folio *folio = lru_to_folio(list);
2498
2499 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2500 list_del(&folio->lru);
2501 if (unlikely(!folio_evictable(folio))) {
2502 spin_unlock_irq(&lruvec->lru_lock);
2503 folio_putback_lru(folio);
2504 spin_lock_irq(&lruvec->lru_lock);
2505 continue;
2506 }
2507
2508 /*
2509 * The folio_set_lru needs to be kept here for list integrity.
2510 * Otherwise:
2511 * #0 move_folios_to_lru #1 release_pages
2512 * if (!folio_put_testzero())
2513 * if (folio_put_testzero())
2514 * !lru //skip lru_lock
2515 * folio_set_lru()
2516 * list_add(&folio->lru,)
2517 * list_add(&folio->lru,)
2518 */
2519 folio_set_lru(folio);
2520
2521 if (unlikely(folio_put_testzero(folio))) {
2522 __folio_clear_lru_flags(folio);
2523
2524 if (unlikely(folio_test_large(folio))) {
2525 spin_unlock_irq(&lruvec->lru_lock);
2526 destroy_large_folio(folio);
2527 spin_lock_irq(&lruvec->lru_lock);
2528 } else
2529 list_add(&folio->lru, &folios_to_free);
2530
2531 continue;
2532 }
2533
2534 /*
2535 * All pages were isolated from the same lruvec (and isolation
2536 * inhibits memcg migration).
2537 */
2538 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2539 lruvec_add_folio(lruvec, folio);
2540 nr_pages = folio_nr_pages(folio);
2541 nr_moved += nr_pages;
2542 if (folio_test_active(folio))
2543 workingset_age_nonresident(lruvec, nr_pages);
2544 }
2545
2546 /*
2547 * To save our caller's stack, now use input list for pages to free.
2548 */
2549 list_splice(&folios_to_free, list);
2550
2551 return nr_moved;
2552 }
2553
2554 /*
2555 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2556 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2557 * we should not throttle. Otherwise it is safe to do so.
2558 */
current_may_throttle(void)2559 static int current_may_throttle(void)
2560 {
2561 return !(current->flags & PF_LOCAL_THROTTLE);
2562 }
2563
2564 /*
2565 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2566 * of reclaimed pages
2567 */
shrink_inactive_list(unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc,enum lru_list lru)2568 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2569 struct lruvec *lruvec, struct scan_control *sc,
2570 enum lru_list lru)
2571 {
2572 LIST_HEAD(folio_list);
2573 unsigned long nr_scanned;
2574 unsigned int nr_reclaimed = 0;
2575 unsigned long nr_taken;
2576 struct reclaim_stat stat;
2577 bool file = is_file_lru(lru);
2578 enum vm_event_item item;
2579 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2580 bool stalled = false;
2581
2582 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2583 if (stalled)
2584 return 0;
2585
2586 /* wait a bit for the reclaimer. */
2587 stalled = true;
2588 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2589
2590 /* We are about to die and free our memory. Return now. */
2591 if (fatal_signal_pending(current))
2592 return SWAP_CLUSTER_MAX;
2593 }
2594
2595 lru_add_drain();
2596
2597 spin_lock_irq(&lruvec->lru_lock);
2598
2599 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
2600 &nr_scanned, sc, lru);
2601
2602 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2603 item = PGSCAN_KSWAPD + reclaimer_offset();
2604 if (!cgroup_reclaim(sc))
2605 __count_vm_events(item, nr_scanned);
2606 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2607 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2608
2609 spin_unlock_irq(&lruvec->lru_lock);
2610
2611 if (nr_taken == 0)
2612 return 0;
2613
2614 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
2615
2616 spin_lock_irq(&lruvec->lru_lock);
2617 move_folios_to_lru(lruvec, &folio_list);
2618
2619 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2620 item = PGSTEAL_KSWAPD + reclaimer_offset();
2621 if (!cgroup_reclaim(sc))
2622 __count_vm_events(item, nr_reclaimed);
2623 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2624 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2625 spin_unlock_irq(&lruvec->lru_lock);
2626
2627 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
2628 mem_cgroup_uncharge_list(&folio_list);
2629 free_unref_page_list(&folio_list);
2630
2631 /*
2632 * If dirty folios are scanned that are not queued for IO, it
2633 * implies that flushers are not doing their job. This can
2634 * happen when memory pressure pushes dirty folios to the end of
2635 * the LRU before the dirty limits are breached and the dirty
2636 * data has expired. It can also happen when the proportion of
2637 * dirty folios grows not through writes but through memory
2638 * pressure reclaiming all the clean cache. And in some cases,
2639 * the flushers simply cannot keep up with the allocation
2640 * rate. Nudge the flusher threads in case they are asleep.
2641 */
2642 if (stat.nr_unqueued_dirty == nr_taken) {
2643 wakeup_flusher_threads(WB_REASON_VMSCAN);
2644 /*
2645 * For cgroupv1 dirty throttling is achieved by waking up
2646 * the kernel flusher here and later waiting on folios
2647 * which are in writeback to finish (see shrink_folio_list()).
2648 *
2649 * Flusher may not be able to issue writeback quickly
2650 * enough for cgroupv1 writeback throttling to work
2651 * on a large system.
2652 */
2653 if (!writeback_throttling_sane(sc))
2654 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
2655 }
2656
2657 sc->nr.dirty += stat.nr_dirty;
2658 sc->nr.congested += stat.nr_congested;
2659 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2660 sc->nr.writeback += stat.nr_writeback;
2661 sc->nr.immediate += stat.nr_immediate;
2662 sc->nr.taken += nr_taken;
2663 if (file)
2664 sc->nr.file_taken += nr_taken;
2665
2666 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2667 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2668 return nr_reclaimed;
2669 }
2670
2671 /*
2672 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2673 *
2674 * We move them the other way if the folio is referenced by one or more
2675 * processes.
2676 *
2677 * If the folios are mostly unmapped, the processing is fast and it is
2678 * appropriate to hold lru_lock across the whole operation. But if
2679 * the folios are mapped, the processing is slow (folio_referenced()), so
2680 * we should drop lru_lock around each folio. It's impossible to balance
2681 * this, so instead we remove the folios from the LRU while processing them.
2682 * It is safe to rely on the active flag against the non-LRU folios in here
2683 * because nobody will play with that bit on a non-LRU folio.
2684 *
2685 * The downside is that we have to touch folio->_refcount against each folio.
2686 * But we had to alter folio->flags anyway.
2687 */
shrink_active_list(unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc,enum lru_list lru)2688 static void shrink_active_list(unsigned long nr_to_scan,
2689 struct lruvec *lruvec,
2690 struct scan_control *sc,
2691 enum lru_list lru)
2692 {
2693 unsigned long nr_taken;
2694 unsigned long nr_scanned;
2695 unsigned long vm_flags;
2696 LIST_HEAD(l_hold); /* The folios which were snipped off */
2697 LIST_HEAD(l_active);
2698 LIST_HEAD(l_inactive);
2699 unsigned nr_deactivate, nr_activate;
2700 unsigned nr_rotated = 0;
2701 int file = is_file_lru(lru);
2702 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2703
2704 lru_add_drain();
2705
2706 spin_lock_irq(&lruvec->lru_lock);
2707
2708 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2709 &nr_scanned, sc, lru);
2710
2711 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2712
2713 if (!cgroup_reclaim(sc))
2714 __count_vm_events(PGREFILL, nr_scanned);
2715 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2716
2717 spin_unlock_irq(&lruvec->lru_lock);
2718
2719 while (!list_empty(&l_hold)) {
2720 struct folio *folio;
2721
2722 cond_resched();
2723 folio = lru_to_folio(&l_hold);
2724 list_del(&folio->lru);
2725
2726 if (unlikely(!folio_evictable(folio))) {
2727 folio_putback_lru(folio);
2728 continue;
2729 }
2730
2731 if (unlikely(buffer_heads_over_limit)) {
2732 if (folio_needs_release(folio) &&
2733 folio_trylock(folio)) {
2734 filemap_release_folio(folio, 0);
2735 folio_unlock(folio);
2736 }
2737 }
2738
2739 /* Referenced or rmap lock contention: rotate */
2740 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2741 &vm_flags) != 0) {
2742 /*
2743 * Identify referenced, file-backed active folios and
2744 * give them one more trip around the active list. So
2745 * that executable code get better chances to stay in
2746 * memory under moderate memory pressure. Anon folios
2747 * are not likely to be evicted by use-once streaming
2748 * IO, plus JVM can create lots of anon VM_EXEC folios,
2749 * so we ignore them here.
2750 */
2751 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2752 nr_rotated += folio_nr_pages(folio);
2753 list_add(&folio->lru, &l_active);
2754 continue;
2755 }
2756 }
2757
2758 folio_clear_active(folio); /* we are de-activating */
2759 folio_set_workingset(folio);
2760 list_add(&folio->lru, &l_inactive);
2761 }
2762
2763 /*
2764 * Move folios back to the lru list.
2765 */
2766 spin_lock_irq(&lruvec->lru_lock);
2767
2768 nr_activate = move_folios_to_lru(lruvec, &l_active);
2769 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2770 /* Keep all free folios in l_active list */
2771 list_splice(&l_inactive, &l_active);
2772
2773 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2774 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2775
2776 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2777 spin_unlock_irq(&lruvec->lru_lock);
2778
2779 if (nr_rotated)
2780 lru_note_cost(lruvec, file, 0, nr_rotated);
2781 mem_cgroup_uncharge_list(&l_active);
2782 free_unref_page_list(&l_active);
2783 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2784 nr_deactivate, nr_rotated, sc->priority, file);
2785 }
2786
reclaim_folio_list(struct list_head * folio_list,struct pglist_data * pgdat)2787 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2788 struct pglist_data *pgdat)
2789 {
2790 struct reclaim_stat dummy_stat;
2791 unsigned int nr_reclaimed;
2792 struct folio *folio;
2793 struct scan_control sc = {
2794 .gfp_mask = GFP_KERNEL,
2795 .may_writepage = 1,
2796 .may_unmap = 1,
2797 .may_swap = 1,
2798 .no_demotion = 1,
2799 };
2800
2801 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2802 while (!list_empty(folio_list)) {
2803 folio = lru_to_folio(folio_list);
2804 list_del(&folio->lru);
2805 folio_putback_lru(folio);
2806 }
2807
2808 return nr_reclaimed;
2809 }
2810
reclaim_pages(struct list_head * folio_list)2811 unsigned long reclaim_pages(struct list_head *folio_list)
2812 {
2813 int nid;
2814 unsigned int nr_reclaimed = 0;
2815 LIST_HEAD(node_folio_list);
2816 unsigned int noreclaim_flag;
2817
2818 if (list_empty(folio_list))
2819 return nr_reclaimed;
2820
2821 noreclaim_flag = memalloc_noreclaim_save();
2822
2823 nid = folio_nid(lru_to_folio(folio_list));
2824 do {
2825 struct folio *folio = lru_to_folio(folio_list);
2826
2827 if (nid == folio_nid(folio)) {
2828 folio_clear_active(folio);
2829 list_move(&folio->lru, &node_folio_list);
2830 continue;
2831 }
2832
2833 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2834 nid = folio_nid(lru_to_folio(folio_list));
2835 } while (!list_empty(folio_list));
2836
2837 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2838
2839 memalloc_noreclaim_restore(noreclaim_flag);
2840
2841 return nr_reclaimed;
2842 }
2843
shrink_list(enum lru_list lru,unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc)2844 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2845 struct lruvec *lruvec, struct scan_control *sc)
2846 {
2847 if (is_active_lru(lru)) {
2848 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2849 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2850 else
2851 sc->skipped_deactivate = 1;
2852 return 0;
2853 }
2854
2855 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2856 }
2857
2858 /*
2859 * The inactive anon list should be small enough that the VM never has
2860 * to do too much work.
2861 *
2862 * The inactive file list should be small enough to leave most memory
2863 * to the established workingset on the scan-resistant active list,
2864 * but large enough to avoid thrashing the aggregate readahead window.
2865 *
2866 * Both inactive lists should also be large enough that each inactive
2867 * folio has a chance to be referenced again before it is reclaimed.
2868 *
2869 * If that fails and refaulting is observed, the inactive list grows.
2870 *
2871 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2872 * on this LRU, maintained by the pageout code. An inactive_ratio
2873 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2874 *
2875 * total target max
2876 * memory ratio inactive
2877 * -------------------------------------
2878 * 10MB 1 5MB
2879 * 100MB 1 50MB
2880 * 1GB 3 250MB
2881 * 10GB 10 0.9GB
2882 * 100GB 31 3GB
2883 * 1TB 101 10GB
2884 * 10TB 320 32GB
2885 */
inactive_is_low(struct lruvec * lruvec,enum lru_list inactive_lru)2886 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2887 {
2888 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2889 unsigned long inactive, active;
2890 unsigned long inactive_ratio;
2891 unsigned long gb;
2892
2893 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2894 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2895
2896 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2897 if (gb)
2898 inactive_ratio = int_sqrt(10 * gb);
2899 else
2900 inactive_ratio = 1;
2901
2902 return inactive * inactive_ratio < active;
2903 }
2904
2905 enum scan_balance {
2906 SCAN_EQUAL,
2907 SCAN_FRACT,
2908 SCAN_ANON,
2909 SCAN_FILE,
2910 };
2911
prepare_scan_count(pg_data_t * pgdat,struct scan_control * sc)2912 static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2913 {
2914 unsigned long file;
2915 struct lruvec *target_lruvec;
2916
2917 if (lru_gen_enabled())
2918 return;
2919
2920 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2921
2922 /*
2923 * Flush the memory cgroup stats, so that we read accurate per-memcg
2924 * lruvec stats for heuristics.
2925 */
2926 mem_cgroup_flush_stats();
2927
2928 /*
2929 * Determine the scan balance between anon and file LRUs.
2930 */
2931 spin_lock_irq(&target_lruvec->lru_lock);
2932 sc->anon_cost = target_lruvec->anon_cost;
2933 sc->file_cost = target_lruvec->file_cost;
2934 spin_unlock_irq(&target_lruvec->lru_lock);
2935
2936 /*
2937 * Target desirable inactive:active list ratios for the anon
2938 * and file LRU lists.
2939 */
2940 if (!sc->force_deactivate) {
2941 unsigned long refaults;
2942
2943 /*
2944 * When refaults are being observed, it means a new
2945 * workingset is being established. Deactivate to get
2946 * rid of any stale active pages quickly.
2947 */
2948 refaults = lruvec_page_state(target_lruvec,
2949 WORKINGSET_ACTIVATE_ANON);
2950 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2951 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2952 sc->may_deactivate |= DEACTIVATE_ANON;
2953 else
2954 sc->may_deactivate &= ~DEACTIVATE_ANON;
2955
2956 refaults = lruvec_page_state(target_lruvec,
2957 WORKINGSET_ACTIVATE_FILE);
2958 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2959 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2960 sc->may_deactivate |= DEACTIVATE_FILE;
2961 else
2962 sc->may_deactivate &= ~DEACTIVATE_FILE;
2963 } else
2964 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2965
2966 /*
2967 * If we have plenty of inactive file pages that aren't
2968 * thrashing, try to reclaim those first before touching
2969 * anonymous pages.
2970 */
2971 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2972 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2973 sc->cache_trim_mode = 1;
2974 else
2975 sc->cache_trim_mode = 0;
2976
2977 /*
2978 * Prevent the reclaimer from falling into the cache trap: as
2979 * cache pages start out inactive, every cache fault will tip
2980 * the scan balance towards the file LRU. And as the file LRU
2981 * shrinks, so does the window for rotation from references.
2982 * This means we have a runaway feedback loop where a tiny
2983 * thrashing file LRU becomes infinitely more attractive than
2984 * anon pages. Try to detect this based on file LRU size.
2985 */
2986 if (!cgroup_reclaim(sc)) {
2987 unsigned long total_high_wmark = 0;
2988 unsigned long free, anon;
2989 int z;
2990
2991 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2992 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2993 node_page_state(pgdat, NR_INACTIVE_FILE);
2994
2995 for (z = 0; z < MAX_NR_ZONES; z++) {
2996 struct zone *zone = &pgdat->node_zones[z];
2997
2998 if (!managed_zone(zone))
2999 continue;
3000
3001 total_high_wmark += high_wmark_pages(zone);
3002 }
3003
3004 /*
3005 * Consider anon: if that's low too, this isn't a
3006 * runaway file reclaim problem, but rather just
3007 * extreme pressure. Reclaim as per usual then.
3008 */
3009 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
3010
3011 sc->file_is_tiny =
3012 file + free <= total_high_wmark &&
3013 !(sc->may_deactivate & DEACTIVATE_ANON) &&
3014 anon >> sc->priority;
3015 }
3016 }
3017
3018 /*
3019 * Determine how aggressively the anon and file LRU lists should be
3020 * scanned.
3021 *
3022 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
3023 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
3024 */
get_scan_count(struct lruvec * lruvec,struct scan_control * sc,unsigned long * nr)3025 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
3026 unsigned long *nr)
3027 {
3028 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3029 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3030 unsigned long anon_cost, file_cost, total_cost;
3031 int swappiness = mem_cgroup_swappiness(memcg);
3032 u64 fraction[ANON_AND_FILE];
3033 u64 denominator = 0; /* gcc */
3034 enum scan_balance scan_balance;
3035 unsigned long ap, fp;
3036 enum lru_list lru;
3037
3038 /* If we have no swap space, do not bother scanning anon folios. */
3039 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
3040 scan_balance = SCAN_FILE;
3041 goto out;
3042 }
3043
3044 /*
3045 * Global reclaim will swap to prevent OOM even with no
3046 * swappiness, but memcg users want to use this knob to
3047 * disable swapping for individual groups completely when
3048 * using the memory controller's swap limit feature would be
3049 * too expensive.
3050 */
3051 if (cgroup_reclaim(sc) && !swappiness) {
3052 scan_balance = SCAN_FILE;
3053 goto out;
3054 }
3055
3056 /*
3057 * Do not apply any pressure balancing cleverness when the
3058 * system is close to OOM, scan both anon and file equally
3059 * (unless the swappiness setting disagrees with swapping).
3060 */
3061 if (!sc->priority && swappiness) {
3062 scan_balance = SCAN_EQUAL;
3063 goto out;
3064 }
3065
3066 /*
3067 * If the system is almost out of file pages, force-scan anon.
3068 */
3069 if (sc->file_is_tiny) {
3070 scan_balance = SCAN_ANON;
3071 goto out;
3072 }
3073
3074 /*
3075 * If there is enough inactive page cache, we do not reclaim
3076 * anything from the anonymous working right now.
3077 */
3078 if (sc->cache_trim_mode) {
3079 scan_balance = SCAN_FILE;
3080 goto out;
3081 }
3082
3083 scan_balance = SCAN_FRACT;
3084 /*
3085 * Calculate the pressure balance between anon and file pages.
3086 *
3087 * The amount of pressure we put on each LRU is inversely
3088 * proportional to the cost of reclaiming each list, as
3089 * determined by the share of pages that are refaulting, times
3090 * the relative IO cost of bringing back a swapped out
3091 * anonymous page vs reloading a filesystem page (swappiness).
3092 *
3093 * Although we limit that influence to ensure no list gets
3094 * left behind completely: at least a third of the pressure is
3095 * applied, before swappiness.
3096 *
3097 * With swappiness at 100, anon and file have equal IO cost.
3098 */
3099 total_cost = sc->anon_cost + sc->file_cost;
3100 anon_cost = total_cost + sc->anon_cost;
3101 file_cost = total_cost + sc->file_cost;
3102 total_cost = anon_cost + file_cost;
3103
3104 ap = swappiness * (total_cost + 1);
3105 ap /= anon_cost + 1;
3106
3107 fp = (200 - swappiness) * (total_cost + 1);
3108 fp /= file_cost + 1;
3109
3110 fraction[0] = ap;
3111 fraction[1] = fp;
3112 denominator = ap + fp;
3113 out:
3114 for_each_evictable_lru(lru) {
3115 int file = is_file_lru(lru);
3116 unsigned long lruvec_size;
3117 unsigned long low, min;
3118 unsigned long scan;
3119
3120 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
3121 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
3122 &min, &low);
3123
3124 if (min || low) {
3125 /*
3126 * Scale a cgroup's reclaim pressure by proportioning
3127 * its current usage to its memory.low or memory.min
3128 * setting.
3129 *
3130 * This is important, as otherwise scanning aggression
3131 * becomes extremely binary -- from nothing as we
3132 * approach the memory protection threshold, to totally
3133 * nominal as we exceed it. This results in requiring
3134 * setting extremely liberal protection thresholds. It
3135 * also means we simply get no protection at all if we
3136 * set it too low, which is not ideal.
3137 *
3138 * If there is any protection in place, we reduce scan
3139 * pressure by how much of the total memory used is
3140 * within protection thresholds.
3141 *
3142 * There is one special case: in the first reclaim pass,
3143 * we skip over all groups that are within their low
3144 * protection. If that fails to reclaim enough pages to
3145 * satisfy the reclaim goal, we come back and override
3146 * the best-effort low protection. However, we still
3147 * ideally want to honor how well-behaved groups are in
3148 * that case instead of simply punishing them all
3149 * equally. As such, we reclaim them based on how much
3150 * memory they are using, reducing the scan pressure
3151 * again by how much of the total memory used is under
3152 * hard protection.
3153 */
3154 unsigned long cgroup_size = mem_cgroup_size(memcg);
3155 unsigned long protection;
3156
3157 /* memory.low scaling, make sure we retry before OOM */
3158 if (!sc->memcg_low_reclaim && low > min) {
3159 protection = low;
3160 sc->memcg_low_skipped = 1;
3161 } else {
3162 protection = min;
3163 }
3164
3165 /* Avoid TOCTOU with earlier protection check */
3166 cgroup_size = max(cgroup_size, protection);
3167
3168 scan = lruvec_size - lruvec_size * protection /
3169 (cgroup_size + 1);
3170
3171 /*
3172 * Minimally target SWAP_CLUSTER_MAX pages to keep
3173 * reclaim moving forwards, avoiding decrementing
3174 * sc->priority further than desirable.
3175 */
3176 scan = max(scan, SWAP_CLUSTER_MAX);
3177 } else {
3178 scan = lruvec_size;
3179 }
3180
3181 scan >>= sc->priority;
3182
3183 /*
3184 * If the cgroup's already been deleted, make sure to
3185 * scrape out the remaining cache.
3186 */
3187 if (!scan && !mem_cgroup_online(memcg))
3188 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
3189
3190 switch (scan_balance) {
3191 case SCAN_EQUAL:
3192 /* Scan lists relative to size */
3193 break;
3194 case SCAN_FRACT:
3195 /*
3196 * Scan types proportional to swappiness and
3197 * their relative recent reclaim efficiency.
3198 * Make sure we don't miss the last page on
3199 * the offlined memory cgroups because of a
3200 * round-off error.
3201 */
3202 scan = mem_cgroup_online(memcg) ?
3203 div64_u64(scan * fraction[file], denominator) :
3204 DIV64_U64_ROUND_UP(scan * fraction[file],
3205 denominator);
3206 break;
3207 case SCAN_FILE:
3208 case SCAN_ANON:
3209 /* Scan one type exclusively */
3210 if ((scan_balance == SCAN_FILE) != file)
3211 scan = 0;
3212 break;
3213 default:
3214 /* Look ma, no brain */
3215 BUG();
3216 }
3217
3218 nr[lru] = scan;
3219 }
3220 }
3221
3222 /*
3223 * Anonymous LRU management is a waste if there is
3224 * ultimately no way to reclaim the memory.
3225 */
can_age_anon_pages(struct pglist_data * pgdat,struct scan_control * sc)3226 static bool can_age_anon_pages(struct pglist_data *pgdat,
3227 struct scan_control *sc)
3228 {
3229 /* Aging the anon LRU is valuable if swap is present: */
3230 if (total_swap_pages > 0)
3231 return true;
3232
3233 /* Also valuable if anon pages can be demoted: */
3234 return can_demote(pgdat->node_id, sc);
3235 }
3236
3237 #ifdef CONFIG_LRU_GEN
3238
3239 #ifdef CONFIG_LRU_GEN_ENABLED
3240 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
3241 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
3242 #else
3243 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
3244 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
3245 #endif
3246
should_walk_mmu(void)3247 static bool should_walk_mmu(void)
3248 {
3249 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
3250 }
3251
should_clear_pmd_young(void)3252 static bool should_clear_pmd_young(void)
3253 {
3254 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
3255 }
3256
3257 /******************************************************************************
3258 * shorthand helpers
3259 ******************************************************************************/
3260
3261 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
3262
3263 #define DEFINE_MAX_SEQ(lruvec) \
3264 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
3265
3266 #define DEFINE_MIN_SEQ(lruvec) \
3267 unsigned long min_seq[ANON_AND_FILE] = { \
3268 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
3269 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
3270 }
3271
3272 #define for_each_gen_type_zone(gen, type, zone) \
3273 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
3274 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
3275 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
3276
3277 #define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
3278 #define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
3279
get_lruvec(struct mem_cgroup * memcg,int nid)3280 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
3281 {
3282 struct pglist_data *pgdat = NODE_DATA(nid);
3283
3284 #ifdef CONFIG_MEMCG
3285 if (memcg) {
3286 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
3287
3288 /* see the comment in mem_cgroup_lruvec() */
3289 if (!lruvec->pgdat)
3290 lruvec->pgdat = pgdat;
3291
3292 return lruvec;
3293 }
3294 #endif
3295 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3296
3297 return &pgdat->__lruvec;
3298 }
3299
get_swappiness(struct lruvec * lruvec,struct scan_control * sc)3300 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
3301 {
3302 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3303 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3304
3305 if (!sc->may_swap)
3306 return 0;
3307
3308 if (!can_demote(pgdat->node_id, sc) &&
3309 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3310 return 0;
3311
3312 return mem_cgroup_swappiness(memcg);
3313 }
3314
get_nr_gens(struct lruvec * lruvec,int type)3315 static int get_nr_gens(struct lruvec *lruvec, int type)
3316 {
3317 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3318 }
3319
seq_is_valid(struct lruvec * lruvec)3320 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3321 {
3322 /* see the comment on lru_gen_folio */
3323 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3324 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3325 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3326 }
3327
3328 /******************************************************************************
3329 * Bloom filters
3330 ******************************************************************************/
3331
3332 /*
3333 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3334 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3335 * bits in a bitmap, k is the number of hash functions and n is the number of
3336 * inserted items.
3337 *
3338 * Page table walkers use one of the two filters to reduce their search space.
3339 * To get rid of non-leaf entries that no longer have enough leaf entries, the
3340 * aging uses the double-buffering technique to flip to the other filter each
3341 * time it produces a new generation. For non-leaf entries that have enough
3342 * leaf entries, the aging carries them over to the next generation in
3343 * walk_pmd_range(); the eviction also report them when walking the rmap
3344 * in lru_gen_look_around().
3345 *
3346 * For future optimizations:
3347 * 1. It's not necessary to keep both filters all the time. The spare one can be
3348 * freed after the RCU grace period and reallocated if needed again.
3349 * 2. And when reallocating, it's worth scaling its size according to the number
3350 * of inserted entries in the other filter, to reduce the memory overhead on
3351 * small systems and false positives on large systems.
3352 * 3. Jenkins' hash function is an alternative to Knuth's.
3353 */
3354 #define BLOOM_FILTER_SHIFT 15
3355
filter_gen_from_seq(unsigned long seq)3356 static inline int filter_gen_from_seq(unsigned long seq)
3357 {
3358 return seq % NR_BLOOM_FILTERS;
3359 }
3360
get_item_key(void * item,int * key)3361 static void get_item_key(void *item, int *key)
3362 {
3363 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3364
3365 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3366
3367 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3368 key[1] = hash >> BLOOM_FILTER_SHIFT;
3369 }
3370
test_bloom_filter(struct lruvec * lruvec,unsigned long seq,void * item)3371 static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3372 {
3373 int key[2];
3374 unsigned long *filter;
3375 int gen = filter_gen_from_seq(seq);
3376
3377 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3378 if (!filter)
3379 return true;
3380
3381 get_item_key(item, key);
3382
3383 return test_bit(key[0], filter) && test_bit(key[1], filter);
3384 }
3385
update_bloom_filter(struct lruvec * lruvec,unsigned long seq,void * item)3386 static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3387 {
3388 int key[2];
3389 unsigned long *filter;
3390 int gen = filter_gen_from_seq(seq);
3391
3392 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3393 if (!filter)
3394 return;
3395
3396 get_item_key(item, key);
3397
3398 if (!test_bit(key[0], filter))
3399 set_bit(key[0], filter);
3400 if (!test_bit(key[1], filter))
3401 set_bit(key[1], filter);
3402 }
3403
reset_bloom_filter(struct lruvec * lruvec,unsigned long seq)3404 static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3405 {
3406 unsigned long *filter;
3407 int gen = filter_gen_from_seq(seq);
3408
3409 filter = lruvec->mm_state.filters[gen];
3410 if (filter) {
3411 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3412 return;
3413 }
3414
3415 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
3416 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3417 WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3418 }
3419
3420 /******************************************************************************
3421 * mm_struct list
3422 ******************************************************************************/
3423
get_mm_list(struct mem_cgroup * memcg)3424 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3425 {
3426 static struct lru_gen_mm_list mm_list = {
3427 .fifo = LIST_HEAD_INIT(mm_list.fifo),
3428 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3429 };
3430
3431 #ifdef CONFIG_MEMCG
3432 if (memcg)
3433 return &memcg->mm_list;
3434 #endif
3435 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3436
3437 return &mm_list;
3438 }
3439
lru_gen_add_mm(struct mm_struct * mm)3440 void lru_gen_add_mm(struct mm_struct *mm)
3441 {
3442 int nid;
3443 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3444 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3445
3446 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
3447 #ifdef CONFIG_MEMCG
3448 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
3449 mm->lru_gen.memcg = memcg;
3450 #endif
3451 spin_lock(&mm_list->lock);
3452
3453 for_each_node_state(nid, N_MEMORY) {
3454 struct lruvec *lruvec = get_lruvec(memcg, nid);
3455
3456 /* the first addition since the last iteration */
3457 if (lruvec->mm_state.tail == &mm_list->fifo)
3458 lruvec->mm_state.tail = &mm->lru_gen.list;
3459 }
3460
3461 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3462
3463 spin_unlock(&mm_list->lock);
3464 }
3465
lru_gen_del_mm(struct mm_struct * mm)3466 void lru_gen_del_mm(struct mm_struct *mm)
3467 {
3468 int nid;
3469 struct lru_gen_mm_list *mm_list;
3470 struct mem_cgroup *memcg = NULL;
3471
3472 if (list_empty(&mm->lru_gen.list))
3473 return;
3474
3475 #ifdef CONFIG_MEMCG
3476 memcg = mm->lru_gen.memcg;
3477 #endif
3478 mm_list = get_mm_list(memcg);
3479
3480 spin_lock(&mm_list->lock);
3481
3482 for_each_node(nid) {
3483 struct lruvec *lruvec = get_lruvec(memcg, nid);
3484
3485 /* where the current iteration continues after */
3486 if (lruvec->mm_state.head == &mm->lru_gen.list)
3487 lruvec->mm_state.head = lruvec->mm_state.head->prev;
3488
3489 /* where the last iteration ended before */
3490 if (lruvec->mm_state.tail == &mm->lru_gen.list)
3491 lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3492 }
3493
3494 list_del_init(&mm->lru_gen.list);
3495
3496 spin_unlock(&mm_list->lock);
3497
3498 #ifdef CONFIG_MEMCG
3499 mem_cgroup_put(mm->lru_gen.memcg);
3500 mm->lru_gen.memcg = NULL;
3501 #endif
3502 }
3503
3504 #ifdef CONFIG_MEMCG
lru_gen_migrate_mm(struct mm_struct * mm)3505 void lru_gen_migrate_mm(struct mm_struct *mm)
3506 {
3507 struct mem_cgroup *memcg;
3508 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3509
3510 VM_WARN_ON_ONCE(task->mm != mm);
3511 lockdep_assert_held(&task->alloc_lock);
3512
3513 /* for mm_update_next_owner() */
3514 if (mem_cgroup_disabled())
3515 return;
3516
3517 /* migration can happen before addition */
3518 if (!mm->lru_gen.memcg)
3519 return;
3520
3521 rcu_read_lock();
3522 memcg = mem_cgroup_from_task(task);
3523 rcu_read_unlock();
3524 if (memcg == mm->lru_gen.memcg)
3525 return;
3526
3527 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3528
3529 lru_gen_del_mm(mm);
3530 lru_gen_add_mm(mm);
3531 }
3532 #endif
3533
reset_mm_stats(struct lruvec * lruvec,struct lru_gen_mm_walk * walk,bool last)3534 static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3535 {
3536 int i;
3537 int hist;
3538
3539 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3540
3541 if (walk) {
3542 hist = lru_hist_from_seq(walk->max_seq);
3543
3544 for (i = 0; i < NR_MM_STATS; i++) {
3545 WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3546 lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3547 walk->mm_stats[i] = 0;
3548 }
3549 }
3550
3551 if (NR_HIST_GENS > 1 && last) {
3552 hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3553
3554 for (i = 0; i < NR_MM_STATS; i++)
3555 WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3556 }
3557 }
3558
should_skip_mm(struct mm_struct * mm,struct lru_gen_mm_walk * walk)3559 static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3560 {
3561 int type;
3562 unsigned long size = 0;
3563 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3564 int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
3565
3566 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
3567 return true;
3568
3569 clear_bit(key, &mm->lru_gen.bitmap);
3570
3571 for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3572 size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3573 get_mm_counter(mm, MM_ANONPAGES) +
3574 get_mm_counter(mm, MM_SHMEMPAGES);
3575 }
3576
3577 if (size < MIN_LRU_BATCH)
3578 return true;
3579
3580 return !mmget_not_zero(mm);
3581 }
3582
iterate_mm_list(struct lruvec * lruvec,struct lru_gen_mm_walk * walk,struct mm_struct ** iter)3583 static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3584 struct mm_struct **iter)
3585 {
3586 bool first = false;
3587 bool last = false;
3588 struct mm_struct *mm = NULL;
3589 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3590 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3591 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3592
3593 /*
3594 * mm_state->seq is incremented after each iteration of mm_list. There
3595 * are three interesting cases for this page table walker:
3596 * 1. It tries to start a new iteration with a stale max_seq: there is
3597 * nothing left to do.
3598 * 2. It started the next iteration: it needs to reset the Bloom filter
3599 * so that a fresh set of PTE tables can be recorded.
3600 * 3. It ended the current iteration: it needs to reset the mm stats
3601 * counters and tell its caller to increment max_seq.
3602 */
3603 spin_lock(&mm_list->lock);
3604
3605 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
3606
3607 if (walk->max_seq <= mm_state->seq)
3608 goto done;
3609
3610 if (!mm_state->head)
3611 mm_state->head = &mm_list->fifo;
3612
3613 if (mm_state->head == &mm_list->fifo)
3614 first = true;
3615
3616 do {
3617 mm_state->head = mm_state->head->next;
3618 if (mm_state->head == &mm_list->fifo) {
3619 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3620 last = true;
3621 break;
3622 }
3623
3624 /* force scan for those added after the last iteration */
3625 if (!mm_state->tail || mm_state->tail == mm_state->head) {
3626 mm_state->tail = mm_state->head->next;
3627 walk->force_scan = true;
3628 }
3629
3630 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3631 if (should_skip_mm(mm, walk))
3632 mm = NULL;
3633 } while (!mm);
3634 done:
3635 if (*iter || last)
3636 reset_mm_stats(lruvec, walk, last);
3637
3638 spin_unlock(&mm_list->lock);
3639
3640 if (mm && first)
3641 reset_bloom_filter(lruvec, walk->max_seq + 1);
3642
3643 if (*iter)
3644 mmput_async(*iter);
3645
3646 *iter = mm;
3647
3648 return last;
3649 }
3650
iterate_mm_list_nowalk(struct lruvec * lruvec,unsigned long max_seq)3651 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3652 {
3653 bool success = false;
3654 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3655 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3656 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3657
3658 spin_lock(&mm_list->lock);
3659
3660 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
3661
3662 if (max_seq > mm_state->seq) {
3663 mm_state->head = NULL;
3664 mm_state->tail = NULL;
3665 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3666 reset_mm_stats(lruvec, NULL, true);
3667 success = true;
3668 }
3669
3670 spin_unlock(&mm_list->lock);
3671
3672 return success;
3673 }
3674
3675 /******************************************************************************
3676 * PID controller
3677 ******************************************************************************/
3678
3679 /*
3680 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3681 *
3682 * The P term is refaulted/(evicted+protected) from a tier in the generation
3683 * currently being evicted; the I term is the exponential moving average of the
3684 * P term over the generations previously evicted, using the smoothing factor
3685 * 1/2; the D term isn't supported.
3686 *
3687 * The setpoint (SP) is always the first tier of one type; the process variable
3688 * (PV) is either any tier of the other type or any other tier of the same
3689 * type.
3690 *
3691 * The error is the difference between the SP and the PV; the correction is to
3692 * turn off protection when SP>PV or turn on protection when SP<PV.
3693 *
3694 * For future optimizations:
3695 * 1. The D term may discount the other two terms over time so that long-lived
3696 * generations can resist stale information.
3697 */
3698 struct ctrl_pos {
3699 unsigned long refaulted;
3700 unsigned long total;
3701 int gain;
3702 };
3703
read_ctrl_pos(struct lruvec * lruvec,int type,int tier,int gain,struct ctrl_pos * pos)3704 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3705 struct ctrl_pos *pos)
3706 {
3707 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3708 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3709
3710 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3711 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3712 pos->total = lrugen->avg_total[type][tier] +
3713 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3714 if (tier)
3715 pos->total += lrugen->protected[hist][type][tier - 1];
3716 pos->gain = gain;
3717 }
3718
reset_ctrl_pos(struct lruvec * lruvec,int type,bool carryover)3719 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3720 {
3721 int hist, tier;
3722 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3723 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3724 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3725
3726 lockdep_assert_held(&lruvec->lru_lock);
3727
3728 if (!carryover && !clear)
3729 return;
3730
3731 hist = lru_hist_from_seq(seq);
3732
3733 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3734 if (carryover) {
3735 unsigned long sum;
3736
3737 sum = lrugen->avg_refaulted[type][tier] +
3738 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3739 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3740
3741 sum = lrugen->avg_total[type][tier] +
3742 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3743 if (tier)
3744 sum += lrugen->protected[hist][type][tier - 1];
3745 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3746 }
3747
3748 if (clear) {
3749 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3750 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3751 if (tier)
3752 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3753 }
3754 }
3755 }
3756
positive_ctrl_err(struct ctrl_pos * sp,struct ctrl_pos * pv)3757 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3758 {
3759 /*
3760 * Return true if the PV has a limited number of refaults or a lower
3761 * refaulted/total than the SP.
3762 */
3763 return pv->refaulted < MIN_LRU_BATCH ||
3764 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3765 (sp->refaulted + 1) * pv->total * pv->gain;
3766 }
3767
3768 /******************************************************************************
3769 * the aging
3770 ******************************************************************************/
3771
3772 /* promote pages accessed through page tables */
folio_update_gen(struct folio * folio,int gen)3773 static int folio_update_gen(struct folio *folio, int gen)
3774 {
3775 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3776
3777 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3778 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3779
3780 do {
3781 /* lru_gen_del_folio() has isolated this page? */
3782 if (!(old_flags & LRU_GEN_MASK)) {
3783 /* for shrink_folio_list() */
3784 new_flags = old_flags | BIT(PG_referenced);
3785 continue;
3786 }
3787
3788 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3789 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3790 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3791
3792 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3793 }
3794
3795 /* protect pages accessed multiple times through file descriptors */
folio_inc_gen(struct lruvec * lruvec,struct folio * folio,bool reclaiming)3796 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3797 {
3798 int type = folio_is_file_lru(folio);
3799 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3800 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3801 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3802
3803 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3804
3805 do {
3806 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3807 /* folio_update_gen() has promoted this page? */
3808 if (new_gen >= 0 && new_gen != old_gen)
3809 return new_gen;
3810
3811 new_gen = (old_gen + 1) % MAX_NR_GENS;
3812
3813 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3814 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3815 /* for folio_end_writeback() */
3816 if (reclaiming)
3817 new_flags |= BIT(PG_reclaim);
3818 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3819
3820 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3821
3822 return new_gen;
3823 }
3824
update_batch_size(struct lru_gen_mm_walk * walk,struct folio * folio,int old_gen,int new_gen)3825 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3826 int old_gen, int new_gen)
3827 {
3828 int type = folio_is_file_lru(folio);
3829 int zone = folio_zonenum(folio);
3830 int delta = folio_nr_pages(folio);
3831
3832 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3833 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3834
3835 walk->batched++;
3836
3837 walk->nr_pages[old_gen][type][zone] -= delta;
3838 walk->nr_pages[new_gen][type][zone] += delta;
3839 }
3840
reset_batch_size(struct lruvec * lruvec,struct lru_gen_mm_walk * walk)3841 static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3842 {
3843 int gen, type, zone;
3844 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3845
3846 walk->batched = 0;
3847
3848 for_each_gen_type_zone(gen, type, zone) {
3849 enum lru_list lru = type * LRU_INACTIVE_FILE;
3850 int delta = walk->nr_pages[gen][type][zone];
3851
3852 if (!delta)
3853 continue;
3854
3855 walk->nr_pages[gen][type][zone] = 0;
3856 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3857 lrugen->nr_pages[gen][type][zone] + delta);
3858
3859 if (lru_gen_is_active(lruvec, gen))
3860 lru += LRU_ACTIVE;
3861 __update_lru_size(lruvec, lru, zone, delta);
3862 }
3863 }
3864
should_skip_vma(unsigned long start,unsigned long end,struct mm_walk * args)3865 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3866 {
3867 struct address_space *mapping;
3868 struct vm_area_struct *vma = args->vma;
3869 struct lru_gen_mm_walk *walk = args->private;
3870
3871 if (!vma_is_accessible(vma))
3872 return true;
3873
3874 if (is_vm_hugetlb_page(vma))
3875 return true;
3876
3877 if (!vma_has_recency(vma))
3878 return true;
3879
3880 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3881 return true;
3882
3883 if (vma == get_gate_vma(vma->vm_mm))
3884 return true;
3885
3886 if (vma_is_anonymous(vma))
3887 return !walk->can_swap;
3888
3889 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3890 return true;
3891
3892 mapping = vma->vm_file->f_mapping;
3893 if (mapping_unevictable(mapping))
3894 return true;
3895
3896 if (shmem_mapping(mapping))
3897 return !walk->can_swap;
3898
3899 /* to exclude special mappings like dax, etc. */
3900 return !mapping->a_ops->read_folio;
3901 }
3902
3903 /*
3904 * Some userspace memory allocators map many single-page VMAs. Instead of
3905 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3906 * table to reduce zigzags and improve cache performance.
3907 */
get_next_vma(unsigned long mask,unsigned long size,struct mm_walk * args,unsigned long * vm_start,unsigned long * vm_end)3908 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3909 unsigned long *vm_start, unsigned long *vm_end)
3910 {
3911 unsigned long start = round_up(*vm_end, size);
3912 unsigned long end = (start | ~mask) + 1;
3913 VMA_ITERATOR(vmi, args->mm, start);
3914
3915 VM_WARN_ON_ONCE(mask & size);
3916 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3917
3918 for_each_vma(vmi, args->vma) {
3919 if (end && end <= args->vma->vm_start)
3920 return false;
3921
3922 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3923 continue;
3924
3925 *vm_start = max(start, args->vma->vm_start);
3926 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3927
3928 return true;
3929 }
3930
3931 return false;
3932 }
3933
get_pte_pfn(pte_t pte,struct vm_area_struct * vma,unsigned long addr)3934 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3935 {
3936 unsigned long pfn = pte_pfn(pte);
3937
3938 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3939
3940 if (!pte_present(pte) || is_zero_pfn(pfn))
3941 return -1;
3942
3943 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3944 return -1;
3945
3946 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3947 return -1;
3948
3949 return pfn;
3950 }
3951
3952 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
get_pmd_pfn(pmd_t pmd,struct vm_area_struct * vma,unsigned long addr)3953 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3954 {
3955 unsigned long pfn = pmd_pfn(pmd);
3956
3957 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3958
3959 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3960 return -1;
3961
3962 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3963 return -1;
3964
3965 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3966 return -1;
3967
3968 return pfn;
3969 }
3970 #endif
3971
get_pfn_folio(unsigned long pfn,struct mem_cgroup * memcg,struct pglist_data * pgdat,bool can_swap)3972 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3973 struct pglist_data *pgdat, bool can_swap)
3974 {
3975 struct folio *folio;
3976
3977 /* try to avoid unnecessary memory loads */
3978 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3979 return NULL;
3980
3981 folio = pfn_folio(pfn);
3982 if (folio_nid(folio) != pgdat->node_id)
3983 return NULL;
3984
3985 if (folio_memcg_rcu(folio) != memcg)
3986 return NULL;
3987
3988 /* file VMAs can contain anon pages from COW */
3989 if (!folio_is_file_lru(folio) && !can_swap)
3990 return NULL;
3991
3992 return folio;
3993 }
3994
suitable_to_scan(int total,int young)3995 static bool suitable_to_scan(int total, int young)
3996 {
3997 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3998
3999 /* suitable if the average number of young PTEs per cacheline is >=1 */
4000 return young * n >= total;
4001 }
4002
walk_pte_range(pmd_t * pmd,unsigned long start,unsigned long end,struct mm_walk * args)4003 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
4004 struct mm_walk *args)
4005 {
4006 int i;
4007 pte_t *pte;
4008 spinlock_t *ptl;
4009 unsigned long addr;
4010 int total = 0;
4011 int young = 0;
4012 struct lru_gen_mm_walk *walk = args->private;
4013 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
4014 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4015 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
4016
4017 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
4018 if (!pte)
4019 return false;
4020 if (!spin_trylock(ptl)) {
4021 pte_unmap(pte);
4022 return false;
4023 }
4024
4025 arch_enter_lazy_mmu_mode();
4026 restart:
4027 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
4028 unsigned long pfn;
4029 struct folio *folio;
4030 pte_t ptent = ptep_get(pte + i);
4031
4032 total++;
4033 walk->mm_stats[MM_LEAF_TOTAL]++;
4034
4035 pfn = get_pte_pfn(ptent, args->vma, addr);
4036 if (pfn == -1)
4037 continue;
4038
4039 if (!pte_young(ptent)) {
4040 walk->mm_stats[MM_LEAF_OLD]++;
4041 continue;
4042 }
4043
4044 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
4045 if (!folio)
4046 continue;
4047
4048 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
4049 VM_WARN_ON_ONCE(true);
4050
4051 young++;
4052 walk->mm_stats[MM_LEAF_YOUNG]++;
4053
4054 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4055 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4056 !folio_test_swapcache(folio)))
4057 folio_mark_dirty(folio);
4058
4059 old_gen = folio_update_gen(folio, new_gen);
4060 if (old_gen >= 0 && old_gen != new_gen)
4061 update_batch_size(walk, folio, old_gen, new_gen);
4062 }
4063
4064 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
4065 goto restart;
4066
4067 arch_leave_lazy_mmu_mode();
4068 pte_unmap_unlock(pte, ptl);
4069
4070 return suitable_to_scan(total, young);
4071 }
4072
4073 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
walk_pmd_range_locked(pud_t * pud,unsigned long addr,struct vm_area_struct * vma,struct mm_walk * args,unsigned long * bitmap,unsigned long * first)4074 static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
4075 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
4076 {
4077 int i;
4078 pmd_t *pmd;
4079 spinlock_t *ptl;
4080 struct lru_gen_mm_walk *walk = args->private;
4081 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
4082 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4083 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
4084
4085 VM_WARN_ON_ONCE(pud_leaf(*pud));
4086
4087 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
4088 if (*first == -1) {
4089 *first = addr;
4090 bitmap_zero(bitmap, MIN_LRU_BATCH);
4091 return;
4092 }
4093
4094 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
4095 if (i && i <= MIN_LRU_BATCH) {
4096 __set_bit(i - 1, bitmap);
4097 return;
4098 }
4099
4100 pmd = pmd_offset(pud, *first);
4101
4102 ptl = pmd_lockptr(args->mm, pmd);
4103 if (!spin_trylock(ptl))
4104 goto done;
4105
4106 arch_enter_lazy_mmu_mode();
4107
4108 do {
4109 unsigned long pfn;
4110 struct folio *folio;
4111
4112 /* don't round down the first address */
4113 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
4114
4115 pfn = get_pmd_pfn(pmd[i], vma, addr);
4116 if (pfn == -1)
4117 goto next;
4118
4119 if (!pmd_trans_huge(pmd[i])) {
4120 if (should_clear_pmd_young())
4121 pmdp_test_and_clear_young(vma, addr, pmd + i);
4122 goto next;
4123 }
4124
4125 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
4126 if (!folio)
4127 goto next;
4128
4129 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
4130 goto next;
4131
4132 walk->mm_stats[MM_LEAF_YOUNG]++;
4133
4134 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
4135 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4136 !folio_test_swapcache(folio)))
4137 folio_mark_dirty(folio);
4138
4139 old_gen = folio_update_gen(folio, new_gen);
4140 if (old_gen >= 0 && old_gen != new_gen)
4141 update_batch_size(walk, folio, old_gen, new_gen);
4142 next:
4143 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
4144 } while (i <= MIN_LRU_BATCH);
4145
4146 arch_leave_lazy_mmu_mode();
4147 spin_unlock(ptl);
4148 done:
4149 *first = -1;
4150 }
4151 #else
walk_pmd_range_locked(pud_t * pud,unsigned long addr,struct vm_area_struct * vma,struct mm_walk * args,unsigned long * bitmap,unsigned long * first)4152 static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
4153 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
4154 {
4155 }
4156 #endif
4157
walk_pmd_range(pud_t * pud,unsigned long start,unsigned long end,struct mm_walk * args)4158 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
4159 struct mm_walk *args)
4160 {
4161 int i;
4162 pmd_t *pmd;
4163 unsigned long next;
4164 unsigned long addr;
4165 struct vm_area_struct *vma;
4166 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
4167 unsigned long first = -1;
4168 struct lru_gen_mm_walk *walk = args->private;
4169
4170 VM_WARN_ON_ONCE(pud_leaf(*pud));
4171
4172 /*
4173 * Finish an entire PMD in two passes: the first only reaches to PTE
4174 * tables to avoid taking the PMD lock; the second, if necessary, takes
4175 * the PMD lock to clear the accessed bit in PMD entries.
4176 */
4177 pmd = pmd_offset(pud, start & PUD_MASK);
4178 restart:
4179 /* walk_pte_range() may call get_next_vma() */
4180 vma = args->vma;
4181 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
4182 pmd_t val = pmdp_get_lockless(pmd + i);
4183
4184 next = pmd_addr_end(addr, end);
4185
4186 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
4187 walk->mm_stats[MM_LEAF_TOTAL]++;
4188 continue;
4189 }
4190
4191 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4192 if (pmd_trans_huge(val)) {
4193 unsigned long pfn = pmd_pfn(val);
4194 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4195
4196 walk->mm_stats[MM_LEAF_TOTAL]++;
4197
4198 if (!pmd_young(val)) {
4199 walk->mm_stats[MM_LEAF_OLD]++;
4200 continue;
4201 }
4202
4203 /* try to avoid unnecessary memory loads */
4204 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4205 continue;
4206
4207 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
4208 continue;
4209 }
4210 #endif
4211 walk->mm_stats[MM_NONLEAF_TOTAL]++;
4212
4213 if (should_clear_pmd_young()) {
4214 if (!pmd_young(val))
4215 continue;
4216
4217 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
4218 }
4219
4220 if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
4221 continue;
4222
4223 walk->mm_stats[MM_NONLEAF_FOUND]++;
4224
4225 if (!walk_pte_range(&val, addr, next, args))
4226 continue;
4227
4228 walk->mm_stats[MM_NONLEAF_ADDED]++;
4229
4230 /* carry over to the next generation */
4231 update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
4232 }
4233
4234 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
4235
4236 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
4237 goto restart;
4238 }
4239
walk_pud_range(p4d_t * p4d,unsigned long start,unsigned long end,struct mm_walk * args)4240 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
4241 struct mm_walk *args)
4242 {
4243 int i;
4244 pud_t *pud;
4245 unsigned long addr;
4246 unsigned long next;
4247 struct lru_gen_mm_walk *walk = args->private;
4248
4249 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
4250
4251 pud = pud_offset(p4d, start & P4D_MASK);
4252 restart:
4253 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
4254 pud_t val = READ_ONCE(pud[i]);
4255
4256 next = pud_addr_end(addr, end);
4257
4258 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
4259 continue;
4260
4261 walk_pmd_range(&val, addr, next, args);
4262
4263 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
4264 end = (addr | ~PUD_MASK) + 1;
4265 goto done;
4266 }
4267 }
4268
4269 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
4270 goto restart;
4271
4272 end = round_up(end, P4D_SIZE);
4273 done:
4274 if (!end || !args->vma)
4275 return 1;
4276
4277 walk->next_addr = max(end, args->vma->vm_start);
4278
4279 return -EAGAIN;
4280 }
4281
walk_mm(struct lruvec * lruvec,struct mm_struct * mm,struct lru_gen_mm_walk * walk)4282 static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
4283 {
4284 static const struct mm_walk_ops mm_walk_ops = {
4285 .test_walk = should_skip_vma,
4286 .p4d_entry = walk_pud_range,
4287 .walk_lock = PGWALK_RDLOCK,
4288 };
4289
4290 int err;
4291 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4292
4293 walk->next_addr = FIRST_USER_ADDRESS;
4294
4295 do {
4296 DEFINE_MAX_SEQ(lruvec);
4297
4298 err = -EBUSY;
4299
4300 /* another thread might have called inc_max_seq() */
4301 if (walk->max_seq != max_seq)
4302 break;
4303
4304 /* folio_update_gen() requires stable folio_memcg() */
4305 if (!mem_cgroup_trylock_pages(memcg))
4306 break;
4307
4308 /* the caller might be holding the lock for write */
4309 if (mmap_read_trylock(mm)) {
4310 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
4311
4312 mmap_read_unlock(mm);
4313 }
4314
4315 mem_cgroup_unlock_pages();
4316
4317 if (walk->batched) {
4318 spin_lock_irq(&lruvec->lru_lock);
4319 reset_batch_size(lruvec, walk);
4320 spin_unlock_irq(&lruvec->lru_lock);
4321 }
4322
4323 cond_resched();
4324 } while (err == -EAGAIN);
4325 }
4326
set_mm_walk(struct pglist_data * pgdat,bool force_alloc)4327 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
4328 {
4329 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4330
4331 if (pgdat && current_is_kswapd()) {
4332 VM_WARN_ON_ONCE(walk);
4333
4334 walk = &pgdat->mm_walk;
4335 } else if (!walk && force_alloc) {
4336 VM_WARN_ON_ONCE(current_is_kswapd());
4337
4338 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4339 }
4340
4341 current->reclaim_state->mm_walk = walk;
4342
4343 return walk;
4344 }
4345
clear_mm_walk(void)4346 static void clear_mm_walk(void)
4347 {
4348 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4349
4350 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
4351 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
4352
4353 current->reclaim_state->mm_walk = NULL;
4354
4355 if (!current_is_kswapd())
4356 kfree(walk);
4357 }
4358
inc_min_seq(struct lruvec * lruvec,int type,bool can_swap)4359 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
4360 {
4361 int zone;
4362 int remaining = MAX_LRU_BATCH;
4363 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4364 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
4365
4366 if (type == LRU_GEN_ANON && !can_swap)
4367 goto done;
4368
4369 /* prevent cold/hot inversion if force_scan is true */
4370 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4371 struct list_head *head = &lrugen->folios[old_gen][type][zone];
4372
4373 while (!list_empty(head)) {
4374 struct folio *folio = lru_to_folio(head);
4375
4376 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4377 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4378 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4379 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4380
4381 new_gen = folio_inc_gen(lruvec, folio, false);
4382 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
4383
4384 if (!--remaining)
4385 return false;
4386 }
4387 }
4388 done:
4389 reset_ctrl_pos(lruvec, type, true);
4390 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4391
4392 return true;
4393 }
4394
try_to_inc_min_seq(struct lruvec * lruvec,bool can_swap)4395 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4396 {
4397 int gen, type, zone;
4398 bool success = false;
4399 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4400 DEFINE_MIN_SEQ(lruvec);
4401
4402 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4403
4404 /* find the oldest populated generation */
4405 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4406 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4407 gen = lru_gen_from_seq(min_seq[type]);
4408
4409 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4410 if (!list_empty(&lrugen->folios[gen][type][zone]))
4411 goto next;
4412 }
4413
4414 min_seq[type]++;
4415 }
4416 next:
4417 ;
4418 }
4419
4420 /* see the comment on lru_gen_folio */
4421 if (can_swap) {
4422 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4423 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4424 }
4425
4426 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4427 if (min_seq[type] == lrugen->min_seq[type])
4428 continue;
4429
4430 reset_ctrl_pos(lruvec, type, true);
4431 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4432 success = true;
4433 }
4434
4435 return success;
4436 }
4437
inc_max_seq(struct lruvec * lruvec,bool can_swap,bool force_scan)4438 static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
4439 {
4440 int prev, next;
4441 int type, zone;
4442 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4443 restart:
4444 spin_lock_irq(&lruvec->lru_lock);
4445
4446 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4447
4448 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
4449 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4450 continue;
4451
4452 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
4453
4454 if (inc_min_seq(lruvec, type, can_swap))
4455 continue;
4456
4457 spin_unlock_irq(&lruvec->lru_lock);
4458 cond_resched();
4459 goto restart;
4460 }
4461
4462 /*
4463 * Update the active/inactive LRU sizes for compatibility. Both sides of
4464 * the current max_seq need to be covered, since max_seq+1 can overlap
4465 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4466 * overlap, cold/hot inversion happens.
4467 */
4468 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4469 next = lru_gen_from_seq(lrugen->max_seq + 1);
4470
4471 for (type = 0; type < ANON_AND_FILE; type++) {
4472 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4473 enum lru_list lru = type * LRU_INACTIVE_FILE;
4474 long delta = lrugen->nr_pages[prev][type][zone] -
4475 lrugen->nr_pages[next][type][zone];
4476
4477 if (!delta)
4478 continue;
4479
4480 __update_lru_size(lruvec, lru, zone, delta);
4481 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4482 }
4483 }
4484
4485 for (type = 0; type < ANON_AND_FILE; type++)
4486 reset_ctrl_pos(lruvec, type, false);
4487
4488 WRITE_ONCE(lrugen->timestamps[next], jiffies);
4489 /* make sure preceding modifications appear */
4490 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4491
4492 spin_unlock_irq(&lruvec->lru_lock);
4493 }
4494
try_to_inc_max_seq(struct lruvec * lruvec,unsigned long max_seq,struct scan_control * sc,bool can_swap,bool force_scan)4495 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4496 struct scan_control *sc, bool can_swap, bool force_scan)
4497 {
4498 bool success;
4499 struct lru_gen_mm_walk *walk;
4500 struct mm_struct *mm = NULL;
4501 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4502
4503 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
4504
4505 /* see the comment in iterate_mm_list() */
4506 if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
4507 success = false;
4508 goto done;
4509 }
4510
4511 /*
4512 * If the hardware doesn't automatically set the accessed bit, fallback
4513 * to lru_gen_look_around(), which only clears the accessed bit in a
4514 * handful of PTEs. Spreading the work out over a period of time usually
4515 * is less efficient, but it avoids bursty page faults.
4516 */
4517 if (!should_walk_mmu()) {
4518 success = iterate_mm_list_nowalk(lruvec, max_seq);
4519 goto done;
4520 }
4521
4522 walk = set_mm_walk(NULL, true);
4523 if (!walk) {
4524 success = iterate_mm_list_nowalk(lruvec, max_seq);
4525 goto done;
4526 }
4527
4528 walk->lruvec = lruvec;
4529 walk->max_seq = max_seq;
4530 walk->can_swap = can_swap;
4531 walk->force_scan = force_scan;
4532
4533 do {
4534 success = iterate_mm_list(lruvec, walk, &mm);
4535 if (mm)
4536 walk_mm(lruvec, mm, walk);
4537 } while (mm);
4538 done:
4539 if (success)
4540 inc_max_seq(lruvec, can_swap, force_scan);
4541
4542 return success;
4543 }
4544
4545 /******************************************************************************
4546 * working set protection
4547 ******************************************************************************/
4548
set_initial_priority(struct pglist_data * pgdat,struct scan_control * sc)4549 static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4550 {
4551 int priority;
4552 unsigned long reclaimable;
4553
4554 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4555 return;
4556 /*
4557 * Determine the initial priority based on
4558 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4559 * where reclaimed_to_scanned_ratio = inactive / total.
4560 */
4561 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4562 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
4563 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4564
4565 /* round down reclaimable and round up sc->nr_to_reclaim */
4566 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4567
4568 /*
4569 * The estimation is based on LRU pages only, so cap it to prevent
4570 * overshoots of shrinker objects by large margins.
4571 */
4572 sc->priority = clamp(priority, DEF_PRIORITY / 2, DEF_PRIORITY);
4573 }
4574
lruvec_is_sizable(struct lruvec * lruvec,struct scan_control * sc)4575 static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
4576 {
4577 int gen, type, zone;
4578 unsigned long total = 0;
4579 bool can_swap = get_swappiness(lruvec, sc);
4580 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4581 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4582 DEFINE_MAX_SEQ(lruvec);
4583 DEFINE_MIN_SEQ(lruvec);
4584
4585 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4586 unsigned long seq;
4587
4588 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4589 gen = lru_gen_from_seq(seq);
4590
4591 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4592 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4593 }
4594 }
4595
4596 /* whether the size is big enough to be helpful */
4597 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4598 }
4599
lruvec_is_reclaimable(struct lruvec * lruvec,struct scan_control * sc,unsigned long min_ttl)4600 static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
4601 unsigned long min_ttl)
4602 {
4603 int gen;
4604 unsigned long birth;
4605 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4606 DEFINE_MIN_SEQ(lruvec);
4607
4608 if (mem_cgroup_below_min(NULL, memcg))
4609 return false;
4610
4611 if (!lruvec_is_sizable(lruvec, sc))
4612 return false;
4613
4614 /* see the comment on lru_gen_folio */
4615 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4616 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4617
4618 return time_is_before_jiffies(birth + min_ttl);
4619 }
4620
4621 /* to protect the working set of the last N jiffies */
4622 static unsigned long lru_gen_min_ttl __read_mostly;
4623
lru_gen_age_node(struct pglist_data * pgdat,struct scan_control * sc)4624 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4625 {
4626 struct mem_cgroup *memcg;
4627 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4628 bool reclaimable = !min_ttl;
4629
4630 VM_WARN_ON_ONCE(!current_is_kswapd());
4631
4632 set_initial_priority(pgdat, sc);
4633
4634 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4635 do {
4636 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4637
4638 mem_cgroup_calculate_protection(NULL, memcg);
4639
4640 if (!reclaimable)
4641 reclaimable = lruvec_is_reclaimable(lruvec, sc, min_ttl);
4642 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4643
4644 /*
4645 * The main goal is to OOM kill if every generation from all memcgs is
4646 * younger than min_ttl. However, another possibility is all memcgs are
4647 * either too small or below min.
4648 */
4649 if (!reclaimable && mutex_trylock(&oom_lock)) {
4650 struct oom_control oc = {
4651 .gfp_mask = sc->gfp_mask,
4652 };
4653
4654 out_of_memory(&oc);
4655
4656 mutex_unlock(&oom_lock);
4657 }
4658 }
4659
4660 /******************************************************************************
4661 * rmap/PT walk feedback
4662 ******************************************************************************/
4663
4664 /*
4665 * This function exploits spatial locality when shrink_folio_list() walks the
4666 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4667 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4668 * the PTE table to the Bloom filter. This forms a feedback loop between the
4669 * eviction and the aging.
4670 */
lru_gen_look_around(struct page_vma_mapped_walk * pvmw)4671 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4672 {
4673 int i;
4674 unsigned long start;
4675 unsigned long end;
4676 struct lru_gen_mm_walk *walk;
4677 int young = 0;
4678 pte_t *pte = pvmw->pte;
4679 unsigned long addr = pvmw->address;
4680 struct vm_area_struct *vma = pvmw->vma;
4681 struct folio *folio = pfn_folio(pvmw->pfn);
4682 bool can_swap = !folio_is_file_lru(folio);
4683 struct mem_cgroup *memcg = folio_memcg(folio);
4684 struct pglist_data *pgdat = folio_pgdat(folio);
4685 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4686 DEFINE_MAX_SEQ(lruvec);
4687 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4688
4689 lockdep_assert_held(pvmw->ptl);
4690 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4691
4692 if (spin_is_contended(pvmw->ptl))
4693 return;
4694
4695 /* exclude special VMAs containing anon pages from COW */
4696 if (vma->vm_flags & VM_SPECIAL)
4697 return;
4698
4699 /* avoid taking the LRU lock under the PTL when possible */
4700 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4701
4702 start = max(addr & PMD_MASK, vma->vm_start);
4703 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4704
4705 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4706 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4707 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4708 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4709 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4710 else {
4711 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4712 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4713 }
4714 }
4715
4716 /* folio_update_gen() requires stable folio_memcg() */
4717 if (!mem_cgroup_trylock_pages(memcg))
4718 return;
4719
4720 arch_enter_lazy_mmu_mode();
4721
4722 pte -= (addr - start) / PAGE_SIZE;
4723
4724 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4725 unsigned long pfn;
4726 pte_t ptent = ptep_get(pte + i);
4727
4728 pfn = get_pte_pfn(ptent, vma, addr);
4729 if (pfn == -1)
4730 continue;
4731
4732 if (!pte_young(ptent))
4733 continue;
4734
4735 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4736 if (!folio)
4737 continue;
4738
4739 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4740 VM_WARN_ON_ONCE(true);
4741
4742 young++;
4743
4744 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4745 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4746 !folio_test_swapcache(folio)))
4747 folio_mark_dirty(folio);
4748
4749 if (walk) {
4750 old_gen = folio_update_gen(folio, new_gen);
4751 if (old_gen >= 0 && old_gen != new_gen)
4752 update_batch_size(walk, folio, old_gen, new_gen);
4753
4754 continue;
4755 }
4756
4757 old_gen = folio_lru_gen(folio);
4758 if (old_gen < 0)
4759 folio_set_referenced(folio);
4760 else if (old_gen != new_gen)
4761 folio_activate(folio);
4762 }
4763
4764 arch_leave_lazy_mmu_mode();
4765 mem_cgroup_unlock_pages();
4766
4767 /* feedback from rmap walkers to page table walkers */
4768 if (suitable_to_scan(i, young))
4769 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4770 }
4771
4772 /******************************************************************************
4773 * memcg LRU
4774 ******************************************************************************/
4775
4776 /* see the comment on MEMCG_NR_GENS */
4777 enum {
4778 MEMCG_LRU_NOP,
4779 MEMCG_LRU_HEAD,
4780 MEMCG_LRU_TAIL,
4781 MEMCG_LRU_OLD,
4782 MEMCG_LRU_YOUNG,
4783 };
4784
4785 #ifdef CONFIG_MEMCG
4786
lru_gen_memcg_seg(struct lruvec * lruvec)4787 static int lru_gen_memcg_seg(struct lruvec *lruvec)
4788 {
4789 return READ_ONCE(lruvec->lrugen.seg);
4790 }
4791
lru_gen_rotate_memcg(struct lruvec * lruvec,int op)4792 static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4793 {
4794 int seg;
4795 int old, new;
4796 unsigned long flags;
4797 int bin = get_random_u32_below(MEMCG_NR_BINS);
4798 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4799
4800 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4801
4802 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4803
4804 seg = 0;
4805 new = old = lruvec->lrugen.gen;
4806
4807 /* see the comment on MEMCG_NR_GENS */
4808 if (op == MEMCG_LRU_HEAD)
4809 seg = MEMCG_LRU_HEAD;
4810 else if (op == MEMCG_LRU_TAIL)
4811 seg = MEMCG_LRU_TAIL;
4812 else if (op == MEMCG_LRU_OLD)
4813 new = get_memcg_gen(pgdat->memcg_lru.seq);
4814 else if (op == MEMCG_LRU_YOUNG)
4815 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4816 else
4817 VM_WARN_ON_ONCE(true);
4818
4819 WRITE_ONCE(lruvec->lrugen.seg, seg);
4820 WRITE_ONCE(lruvec->lrugen.gen, new);
4821
4822 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4823
4824 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4825 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4826 else
4827 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4828
4829 pgdat->memcg_lru.nr_memcgs[old]--;
4830 pgdat->memcg_lru.nr_memcgs[new]++;
4831
4832 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4833 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4834
4835 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4836 }
4837
lru_gen_online_memcg(struct mem_cgroup * memcg)4838 void lru_gen_online_memcg(struct mem_cgroup *memcg)
4839 {
4840 int gen;
4841 int nid;
4842 int bin = get_random_u32_below(MEMCG_NR_BINS);
4843
4844 for_each_node(nid) {
4845 struct pglist_data *pgdat = NODE_DATA(nid);
4846 struct lruvec *lruvec = get_lruvec(memcg, nid);
4847
4848 spin_lock_irq(&pgdat->memcg_lru.lock);
4849
4850 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4851
4852 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4853
4854 lruvec->lrugen.gen = gen;
4855
4856 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4857 pgdat->memcg_lru.nr_memcgs[gen]++;
4858
4859 spin_unlock_irq(&pgdat->memcg_lru.lock);
4860 }
4861 }
4862
lru_gen_offline_memcg(struct mem_cgroup * memcg)4863 void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4864 {
4865 int nid;
4866
4867 for_each_node(nid) {
4868 struct lruvec *lruvec = get_lruvec(memcg, nid);
4869
4870 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4871 }
4872 }
4873
lru_gen_release_memcg(struct mem_cgroup * memcg)4874 void lru_gen_release_memcg(struct mem_cgroup *memcg)
4875 {
4876 int gen;
4877 int nid;
4878
4879 for_each_node(nid) {
4880 struct pglist_data *pgdat = NODE_DATA(nid);
4881 struct lruvec *lruvec = get_lruvec(memcg, nid);
4882
4883 spin_lock_irq(&pgdat->memcg_lru.lock);
4884
4885 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4886 goto unlock;
4887
4888 gen = lruvec->lrugen.gen;
4889
4890 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4891 pgdat->memcg_lru.nr_memcgs[gen]--;
4892
4893 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4894 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4895 unlock:
4896 spin_unlock_irq(&pgdat->memcg_lru.lock);
4897 }
4898 }
4899
lru_gen_soft_reclaim(struct mem_cgroup * memcg,int nid)4900 void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4901 {
4902 struct lruvec *lruvec = get_lruvec(memcg, nid);
4903
4904 /* see the comment on MEMCG_NR_GENS */
4905 if (lru_gen_memcg_seg(lruvec) != MEMCG_LRU_HEAD)
4906 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4907 }
4908
4909 #else /* !CONFIG_MEMCG */
4910
lru_gen_memcg_seg(struct lruvec * lruvec)4911 static int lru_gen_memcg_seg(struct lruvec *lruvec)
4912 {
4913 return 0;
4914 }
4915
4916 #endif
4917
4918 /******************************************************************************
4919 * the eviction
4920 ******************************************************************************/
4921
sort_folio(struct lruvec * lruvec,struct folio * folio,struct scan_control * sc,int tier_idx)4922 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4923 int tier_idx)
4924 {
4925 bool success;
4926 int gen = folio_lru_gen(folio);
4927 int type = folio_is_file_lru(folio);
4928 int zone = folio_zonenum(folio);
4929 int delta = folio_nr_pages(folio);
4930 int refs = folio_lru_refs(folio);
4931 int tier = lru_tier_from_refs(refs);
4932 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4933
4934 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4935
4936 /* unevictable */
4937 if (!folio_evictable(folio)) {
4938 success = lru_gen_del_folio(lruvec, folio, true);
4939 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4940 folio_set_unevictable(folio);
4941 lruvec_add_folio(lruvec, folio);
4942 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4943 return true;
4944 }
4945
4946 /* dirty lazyfree */
4947 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4948 success = lru_gen_del_folio(lruvec, folio, true);
4949 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4950 folio_set_swapbacked(folio);
4951 lruvec_add_folio_tail(lruvec, folio);
4952 return true;
4953 }
4954
4955 /* promoted */
4956 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4957 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4958 return true;
4959 }
4960
4961 /* protected */
4962 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4963 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4964
4965 gen = folio_inc_gen(lruvec, folio, false);
4966 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4967
4968 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4969 lrugen->protected[hist][type][tier - 1] + delta);
4970 return true;
4971 }
4972
4973 /* ineligible */
4974 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4975 gen = folio_inc_gen(lruvec, folio, false);
4976 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4977 return true;
4978 }
4979
4980 /* waiting for writeback */
4981 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4982 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4983 gen = folio_inc_gen(lruvec, folio, true);
4984 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4985 return true;
4986 }
4987
4988 return false;
4989 }
4990
isolate_folio(struct lruvec * lruvec,struct folio * folio,struct scan_control * sc)4991 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4992 {
4993 bool success;
4994
4995 /* swapping inhibited */
4996 if (!(sc->gfp_mask & __GFP_IO) &&
4997 (folio_test_dirty(folio) ||
4998 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4999 return false;
5000
5001 /* raced with release_pages() */
5002 if (!folio_try_get(folio))
5003 return false;
5004
5005 /* raced with another isolation */
5006 if (!folio_test_clear_lru(folio)) {
5007 folio_put(folio);
5008 return false;
5009 }
5010
5011 /* see the comment on MAX_NR_TIERS */
5012 if (!folio_test_referenced(folio))
5013 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
5014
5015 /* for shrink_folio_list() */
5016 folio_clear_reclaim(folio);
5017 folio_clear_referenced(folio);
5018
5019 success = lru_gen_del_folio(lruvec, folio, true);
5020 VM_WARN_ON_ONCE_FOLIO(!success, folio);
5021
5022 return true;
5023 }
5024
scan_folios(struct lruvec * lruvec,struct scan_control * sc,int type,int tier,struct list_head * list)5025 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
5026 int type, int tier, struct list_head *list)
5027 {
5028 int i;
5029 int gen;
5030 enum vm_event_item item;
5031 int sorted = 0;
5032 int scanned = 0;
5033 int isolated = 0;
5034 int remaining = MAX_LRU_BATCH;
5035 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5036 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5037
5038 VM_WARN_ON_ONCE(!list_empty(list));
5039
5040 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
5041 return 0;
5042
5043 gen = lru_gen_from_seq(lrugen->min_seq[type]);
5044
5045 for (i = MAX_NR_ZONES; i > 0; i--) {
5046 LIST_HEAD(moved);
5047 int skipped = 0;
5048 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
5049 struct list_head *head = &lrugen->folios[gen][type][zone];
5050
5051 while (!list_empty(head)) {
5052 struct folio *folio = lru_to_folio(head);
5053 int delta = folio_nr_pages(folio);
5054
5055 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5056 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5057 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5058 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5059
5060 scanned += delta;
5061
5062 if (sort_folio(lruvec, folio, sc, tier))
5063 sorted += delta;
5064 else if (isolate_folio(lruvec, folio, sc)) {
5065 list_add(&folio->lru, list);
5066 isolated += delta;
5067 } else {
5068 list_move(&folio->lru, &moved);
5069 skipped += delta;
5070 }
5071
5072 if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
5073 break;
5074 }
5075
5076 if (skipped) {
5077 list_splice(&moved, head);
5078 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
5079 }
5080
5081 if (!remaining || isolated >= MIN_LRU_BATCH)
5082 break;
5083 }
5084
5085 item = PGSCAN_KSWAPD + reclaimer_offset();
5086 if (!cgroup_reclaim(sc)) {
5087 __count_vm_events(item, isolated);
5088 __count_vm_events(PGREFILL, sorted);
5089 }
5090 __count_memcg_events(memcg, item, isolated);
5091 __count_memcg_events(memcg, PGREFILL, sorted);
5092 __count_vm_events(PGSCAN_ANON + type, isolated);
5093
5094 /*
5095 * There might not be eligible folios due to reclaim_idx. Check the
5096 * remaining to prevent livelock if it's not making progress.
5097 */
5098 return isolated || !remaining ? scanned : 0;
5099 }
5100
get_tier_idx(struct lruvec * lruvec,int type)5101 static int get_tier_idx(struct lruvec *lruvec, int type)
5102 {
5103 int tier;
5104 struct ctrl_pos sp, pv;
5105
5106 /*
5107 * To leave a margin for fluctuations, use a larger gain factor (1:2).
5108 * This value is chosen because any other tier would have at least twice
5109 * as many refaults as the first tier.
5110 */
5111 read_ctrl_pos(lruvec, type, 0, 1, &sp);
5112 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
5113 read_ctrl_pos(lruvec, type, tier, 2, &pv);
5114 if (!positive_ctrl_err(&sp, &pv))
5115 break;
5116 }
5117
5118 return tier - 1;
5119 }
5120
get_type_to_scan(struct lruvec * lruvec,int swappiness,int * tier_idx)5121 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
5122 {
5123 int type, tier;
5124 struct ctrl_pos sp, pv;
5125 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
5126
5127 /*
5128 * Compare the first tier of anon with that of file to determine which
5129 * type to scan. Also need to compare other tiers of the selected type
5130 * with the first tier of the other type to determine the last tier (of
5131 * the selected type) to evict.
5132 */
5133 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
5134 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
5135 type = positive_ctrl_err(&sp, &pv);
5136
5137 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
5138 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
5139 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
5140 if (!positive_ctrl_err(&sp, &pv))
5141 break;
5142 }
5143
5144 *tier_idx = tier - 1;
5145
5146 return type;
5147 }
5148
isolate_folios(struct lruvec * lruvec,struct scan_control * sc,int swappiness,int * type_scanned,struct list_head * list)5149 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
5150 int *type_scanned, struct list_head *list)
5151 {
5152 int i;
5153 int type;
5154 int scanned;
5155 int tier = -1;
5156 DEFINE_MIN_SEQ(lruvec);
5157
5158 /*
5159 * Try to make the obvious choice first. When anon and file are both
5160 * available from the same generation, interpret swappiness 1 as file
5161 * first and 200 as anon first.
5162 */
5163 if (!swappiness)
5164 type = LRU_GEN_FILE;
5165 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
5166 type = LRU_GEN_ANON;
5167 else if (swappiness == 1)
5168 type = LRU_GEN_FILE;
5169 else if (swappiness == 200)
5170 type = LRU_GEN_ANON;
5171 else
5172 type = get_type_to_scan(lruvec, swappiness, &tier);
5173
5174 for (i = !swappiness; i < ANON_AND_FILE; i++) {
5175 if (tier < 0)
5176 tier = get_tier_idx(lruvec, type);
5177
5178 scanned = scan_folios(lruvec, sc, type, tier, list);
5179 if (scanned)
5180 break;
5181
5182 type = !type;
5183 tier = -1;
5184 }
5185
5186 *type_scanned = type;
5187
5188 return scanned;
5189 }
5190
evict_folios(struct lruvec * lruvec,struct scan_control * sc,int swappiness)5191 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
5192 {
5193 int type;
5194 int scanned;
5195 int reclaimed;
5196 LIST_HEAD(list);
5197 LIST_HEAD(clean);
5198 struct folio *folio;
5199 struct folio *next;
5200 enum vm_event_item item;
5201 struct reclaim_stat stat;
5202 struct lru_gen_mm_walk *walk;
5203 bool skip_retry = false;
5204 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5205 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
5206
5207 spin_lock_irq(&lruvec->lru_lock);
5208
5209 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
5210
5211 scanned += try_to_inc_min_seq(lruvec, swappiness);
5212
5213 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
5214 scanned = 0;
5215
5216 spin_unlock_irq(&lruvec->lru_lock);
5217
5218 if (list_empty(&list))
5219 return scanned;
5220 retry:
5221 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
5222 sc->nr_reclaimed += reclaimed;
5223
5224 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
5225 if (!folio_evictable(folio)) {
5226 list_del(&folio->lru);
5227 folio_putback_lru(folio);
5228 continue;
5229 }
5230
5231 if (folio_test_reclaim(folio) &&
5232 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
5233 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
5234 if (folio_test_workingset(folio))
5235 folio_set_referenced(folio);
5236 continue;
5237 }
5238
5239 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
5240 folio_mapped(folio) || folio_test_locked(folio) ||
5241 folio_test_dirty(folio) || folio_test_writeback(folio)) {
5242 /* don't add rejected folios to the oldest generation */
5243 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
5244 BIT(PG_active));
5245 continue;
5246 }
5247
5248 /* retry folios that may have missed folio_rotate_reclaimable() */
5249 list_move(&folio->lru, &clean);
5250 }
5251
5252 spin_lock_irq(&lruvec->lru_lock);
5253
5254 move_folios_to_lru(lruvec, &list);
5255
5256 walk = current->reclaim_state->mm_walk;
5257 if (walk && walk->batched)
5258 reset_batch_size(lruvec, walk);
5259
5260 item = PGSTEAL_KSWAPD + reclaimer_offset();
5261 if (!cgroup_reclaim(sc))
5262 __count_vm_events(item, reclaimed);
5263 __count_memcg_events(memcg, item, reclaimed);
5264 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
5265
5266 spin_unlock_irq(&lruvec->lru_lock);
5267
5268 mem_cgroup_uncharge_list(&list);
5269 free_unref_page_list(&list);
5270
5271 INIT_LIST_HEAD(&list);
5272 list_splice_init(&clean, &list);
5273
5274 if (!list_empty(&list)) {
5275 skip_retry = true;
5276 goto retry;
5277 }
5278
5279 return scanned;
5280 }
5281
should_run_aging(struct lruvec * lruvec,unsigned long max_seq,struct scan_control * sc,bool can_swap,unsigned long * nr_to_scan)5282 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
5283 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
5284 {
5285 int gen, type, zone;
5286 unsigned long old = 0;
5287 unsigned long young = 0;
5288 unsigned long total = 0;
5289 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5290 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5291 DEFINE_MIN_SEQ(lruvec);
5292
5293 /* whether this lruvec is completely out of cold folios */
5294 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
5295 *nr_to_scan = 0;
5296 return true;
5297 }
5298
5299 for (type = !can_swap; type < ANON_AND_FILE; type++) {
5300 unsigned long seq;
5301
5302 for (seq = min_seq[type]; seq <= max_seq; seq++) {
5303 unsigned long size = 0;
5304
5305 gen = lru_gen_from_seq(seq);
5306
5307 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5308 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5309
5310 total += size;
5311 if (seq == max_seq)
5312 young += size;
5313 else if (seq + MIN_NR_GENS == max_seq)
5314 old += size;
5315 }
5316 }
5317
5318 /* try to scrape all its memory if this memcg was deleted */
5319 if (!mem_cgroup_online(memcg)) {
5320 *nr_to_scan = total;
5321 return false;
5322 }
5323
5324 *nr_to_scan = total >> sc->priority;
5325
5326 /*
5327 * The aging tries to be lazy to reduce the overhead, while the eviction
5328 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
5329 * ideal number of generations is MIN_NR_GENS+1.
5330 */
5331 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
5332 return false;
5333
5334 /*
5335 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
5336 * of the total number of pages for each generation. A reasonable range
5337 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
5338 * aging cares about the upper bound of hot pages, while the eviction
5339 * cares about the lower bound of cold pages.
5340 */
5341 if (young * MIN_NR_GENS > total)
5342 return true;
5343 if (old * (MIN_NR_GENS + 2) < total)
5344 return true;
5345
5346 return false;
5347 }
5348
5349 /*
5350 * For future optimizations:
5351 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
5352 * reclaim.
5353 */
get_nr_to_scan(struct lruvec * lruvec,struct scan_control * sc,bool can_swap)5354 static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
5355 {
5356 unsigned long nr_to_scan;
5357 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5358 DEFINE_MAX_SEQ(lruvec);
5359
5360 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
5361 return -1;
5362
5363 if (!should_run_aging(lruvec, max_seq, sc, can_swap, &nr_to_scan))
5364 return nr_to_scan;
5365
5366 /* skip the aging path at the default priority */
5367 if (sc->priority == DEF_PRIORITY)
5368 return nr_to_scan;
5369
5370 /* skip this lruvec as it's low on cold folios */
5371 return try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false) ? -1 : 0;
5372 }
5373
should_abort_scan(struct lruvec * lruvec,struct scan_control * sc)5374 static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
5375 {
5376 int i;
5377 enum zone_watermarks mark;
5378
5379 /* don't abort memcg reclaim to ensure fairness */
5380 if (!root_reclaim(sc))
5381 return false;
5382
5383 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
5384 return true;
5385
5386 /* check the order to exclude compaction-induced reclaim */
5387 if (!current_is_kswapd() || sc->order)
5388 return false;
5389
5390 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
5391 WMARK_PROMO : WMARK_HIGH;
5392
5393 for (i = 0; i <= sc->reclaim_idx; i++) {
5394 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
5395 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
5396
5397 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
5398 return false;
5399 }
5400
5401 /* kswapd should abort if all eligible zones are safe */
5402 return true;
5403 }
5404
try_to_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5405 static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5406 {
5407 long nr_to_scan;
5408 unsigned long scanned = 0;
5409 int swappiness = get_swappiness(lruvec, sc);
5410
5411 /* clean file folios are more likely to exist */
5412 if (swappiness && !(sc->gfp_mask & __GFP_IO))
5413 swappiness = 1;
5414
5415 while (true) {
5416 int delta;
5417
5418 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
5419 if (nr_to_scan <= 0)
5420 break;
5421
5422 delta = evict_folios(lruvec, sc, swappiness);
5423 if (!delta)
5424 break;
5425
5426 scanned += delta;
5427 if (scanned >= nr_to_scan)
5428 break;
5429
5430 if (should_abort_scan(lruvec, sc))
5431 break;
5432
5433 cond_resched();
5434 }
5435
5436 /* whether this lruvec should be rotated */
5437 return nr_to_scan < 0;
5438 }
5439
shrink_one(struct lruvec * lruvec,struct scan_control * sc)5440 static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
5441 {
5442 bool success;
5443 unsigned long scanned = sc->nr_scanned;
5444 unsigned long reclaimed = sc->nr_reclaimed;
5445 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5446 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
5447
5448 /* lru_gen_age_node() called mem_cgroup_calculate_protection() */
5449 if (mem_cgroup_below_min(NULL, memcg))
5450 return MEMCG_LRU_YOUNG;
5451
5452 if (mem_cgroup_below_low(NULL, memcg)) {
5453 /* see the comment on MEMCG_NR_GENS */
5454 if (lru_gen_memcg_seg(lruvec) != MEMCG_LRU_TAIL)
5455 return MEMCG_LRU_TAIL;
5456
5457 memcg_memory_event(memcg, MEMCG_LOW);
5458 }
5459
5460 success = try_to_shrink_lruvec(lruvec, sc);
5461
5462 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
5463
5464 if (!sc->proactive)
5465 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
5466 sc->nr_reclaimed - reclaimed);
5467
5468 flush_reclaim_state(sc);
5469
5470 if (success && mem_cgroup_online(memcg))
5471 return MEMCG_LRU_YOUNG;
5472
5473 if (!success && lruvec_is_sizable(lruvec, sc))
5474 return 0;
5475
5476 /* one retry if offlined or too small */
5477 return lru_gen_memcg_seg(lruvec) != MEMCG_LRU_TAIL ?
5478 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
5479 }
5480
5481 #ifdef CONFIG_MEMCG
5482
shrink_many(struct pglist_data * pgdat,struct scan_control * sc)5483 static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
5484 {
5485 int op;
5486 int gen;
5487 int bin;
5488 int first_bin;
5489 struct lruvec *lruvec;
5490 struct lru_gen_folio *lrugen;
5491 struct mem_cgroup *memcg;
5492 struct hlist_nulls_node *pos;
5493
5494 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
5495 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
5496 restart:
5497 op = 0;
5498 memcg = NULL;
5499
5500 rcu_read_lock();
5501
5502 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
5503 if (op) {
5504 lru_gen_rotate_memcg(lruvec, op);
5505 op = 0;
5506 }
5507
5508 mem_cgroup_put(memcg);
5509 memcg = NULL;
5510
5511 if (gen != READ_ONCE(lrugen->gen))
5512 continue;
5513
5514 lruvec = container_of(lrugen, struct lruvec, lrugen);
5515 memcg = lruvec_memcg(lruvec);
5516
5517 if (!mem_cgroup_tryget(memcg)) {
5518 lru_gen_release_memcg(memcg);
5519 memcg = NULL;
5520 continue;
5521 }
5522
5523 rcu_read_unlock();
5524
5525 op = shrink_one(lruvec, sc);
5526
5527 rcu_read_lock();
5528
5529 if (should_abort_scan(lruvec, sc))
5530 break;
5531 }
5532
5533 rcu_read_unlock();
5534
5535 if (op)
5536 lru_gen_rotate_memcg(lruvec, op);
5537
5538 mem_cgroup_put(memcg);
5539
5540 if (!is_a_nulls(pos))
5541 return;
5542
5543 /* restart if raced with lru_gen_rotate_memcg() */
5544 if (gen != get_nulls_value(pos))
5545 goto restart;
5546
5547 /* try the rest of the bins of the current generation */
5548 bin = get_memcg_bin(bin + 1);
5549 if (bin != first_bin)
5550 goto restart;
5551 }
5552
lru_gen_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5553 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5554 {
5555 struct blk_plug plug;
5556
5557 VM_WARN_ON_ONCE(root_reclaim(sc));
5558 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
5559
5560 lru_add_drain();
5561
5562 blk_start_plug(&plug);
5563
5564 set_mm_walk(NULL, sc->proactive);
5565
5566 if (try_to_shrink_lruvec(lruvec, sc))
5567 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
5568
5569 clear_mm_walk();
5570
5571 blk_finish_plug(&plug);
5572 }
5573
5574 #else /* !CONFIG_MEMCG */
5575
shrink_many(struct pglist_data * pgdat,struct scan_control * sc)5576 static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
5577 {
5578 BUILD_BUG();
5579 }
5580
lru_gen_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5581 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5582 {
5583 BUILD_BUG();
5584 }
5585
5586 #endif
5587
lru_gen_shrink_node(struct pglist_data * pgdat,struct scan_control * sc)5588 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5589 {
5590 struct blk_plug plug;
5591 unsigned long reclaimed = sc->nr_reclaimed;
5592
5593 VM_WARN_ON_ONCE(!root_reclaim(sc));
5594
5595 /*
5596 * Unmapped clean folios are already prioritized. Scanning for more of
5597 * them is likely futile and can cause high reclaim latency when there
5598 * is a large number of memcgs.
5599 */
5600 if (!sc->may_writepage || !sc->may_unmap)
5601 goto done;
5602
5603 lru_add_drain();
5604
5605 blk_start_plug(&plug);
5606
5607 set_mm_walk(pgdat, sc->proactive);
5608
5609 set_initial_priority(pgdat, sc);
5610
5611 if (current_is_kswapd())
5612 sc->nr_reclaimed = 0;
5613
5614 if (mem_cgroup_disabled())
5615 shrink_one(&pgdat->__lruvec, sc);
5616 else
5617 shrink_many(pgdat, sc);
5618
5619 if (current_is_kswapd())
5620 sc->nr_reclaimed += reclaimed;
5621
5622 clear_mm_walk();
5623
5624 blk_finish_plug(&plug);
5625 done:
5626 /* kswapd should never fail */
5627 pgdat->kswapd_failures = 0;
5628 }
5629
5630 /******************************************************************************
5631 * state change
5632 ******************************************************************************/
5633
state_is_valid(struct lruvec * lruvec)5634 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5635 {
5636 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5637
5638 if (lrugen->enabled) {
5639 enum lru_list lru;
5640
5641 for_each_evictable_lru(lru) {
5642 if (!list_empty(&lruvec->lists[lru]))
5643 return false;
5644 }
5645 } else {
5646 int gen, type, zone;
5647
5648 for_each_gen_type_zone(gen, type, zone) {
5649 if (!list_empty(&lrugen->folios[gen][type][zone]))
5650 return false;
5651 }
5652 }
5653
5654 return true;
5655 }
5656
fill_evictable(struct lruvec * lruvec)5657 static bool fill_evictable(struct lruvec *lruvec)
5658 {
5659 enum lru_list lru;
5660 int remaining = MAX_LRU_BATCH;
5661
5662 for_each_evictable_lru(lru) {
5663 int type = is_file_lru(lru);
5664 bool active = is_active_lru(lru);
5665 struct list_head *head = &lruvec->lists[lru];
5666
5667 while (!list_empty(head)) {
5668 bool success;
5669 struct folio *folio = lru_to_folio(head);
5670
5671 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5672 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5673 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5674 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5675
5676 lruvec_del_folio(lruvec, folio);
5677 success = lru_gen_add_folio(lruvec, folio, false);
5678 VM_WARN_ON_ONCE(!success);
5679
5680 if (!--remaining)
5681 return false;
5682 }
5683 }
5684
5685 return true;
5686 }
5687
drain_evictable(struct lruvec * lruvec)5688 static bool drain_evictable(struct lruvec *lruvec)
5689 {
5690 int gen, type, zone;
5691 int remaining = MAX_LRU_BATCH;
5692
5693 for_each_gen_type_zone(gen, type, zone) {
5694 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5695
5696 while (!list_empty(head)) {
5697 bool success;
5698 struct folio *folio = lru_to_folio(head);
5699
5700 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5701 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5702 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5703 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5704
5705 success = lru_gen_del_folio(lruvec, folio, false);
5706 VM_WARN_ON_ONCE(!success);
5707 lruvec_add_folio(lruvec, folio);
5708
5709 if (!--remaining)
5710 return false;
5711 }
5712 }
5713
5714 return true;
5715 }
5716
lru_gen_change_state(bool enabled)5717 static void lru_gen_change_state(bool enabled)
5718 {
5719 static DEFINE_MUTEX(state_mutex);
5720
5721 struct mem_cgroup *memcg;
5722
5723 cgroup_lock();
5724 cpus_read_lock();
5725 get_online_mems();
5726 mutex_lock(&state_mutex);
5727
5728 if (enabled == lru_gen_enabled())
5729 goto unlock;
5730
5731 if (enabled)
5732 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5733 else
5734 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5735
5736 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5737 do {
5738 int nid;
5739
5740 for_each_node(nid) {
5741 struct lruvec *lruvec = get_lruvec(memcg, nid);
5742
5743 spin_lock_irq(&lruvec->lru_lock);
5744
5745 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5746 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5747
5748 lruvec->lrugen.enabled = enabled;
5749
5750 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5751 spin_unlock_irq(&lruvec->lru_lock);
5752 cond_resched();
5753 spin_lock_irq(&lruvec->lru_lock);
5754 }
5755
5756 spin_unlock_irq(&lruvec->lru_lock);
5757 }
5758
5759 cond_resched();
5760 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5761 unlock:
5762 mutex_unlock(&state_mutex);
5763 put_online_mems();
5764 cpus_read_unlock();
5765 cgroup_unlock();
5766 }
5767
5768 /******************************************************************************
5769 * sysfs interface
5770 ******************************************************************************/
5771
min_ttl_ms_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)5772 static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5773 {
5774 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5775 }
5776
5777 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
min_ttl_ms_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)5778 static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5779 const char *buf, size_t len)
5780 {
5781 unsigned int msecs;
5782
5783 if (kstrtouint(buf, 0, &msecs))
5784 return -EINVAL;
5785
5786 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5787
5788 return len;
5789 }
5790
5791 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5792
enabled_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)5793 static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5794 {
5795 unsigned int caps = 0;
5796
5797 if (get_cap(LRU_GEN_CORE))
5798 caps |= BIT(LRU_GEN_CORE);
5799
5800 if (should_walk_mmu())
5801 caps |= BIT(LRU_GEN_MM_WALK);
5802
5803 if (should_clear_pmd_young())
5804 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5805
5806 return sysfs_emit(buf, "0x%04x\n", caps);
5807 }
5808
5809 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
enabled_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)5810 static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5811 const char *buf, size_t len)
5812 {
5813 int i;
5814 unsigned int caps;
5815
5816 if (tolower(*buf) == 'n')
5817 caps = 0;
5818 else if (tolower(*buf) == 'y')
5819 caps = -1;
5820 else if (kstrtouint(buf, 0, &caps))
5821 return -EINVAL;
5822
5823 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5824 bool enabled = caps & BIT(i);
5825
5826 if (i == LRU_GEN_CORE)
5827 lru_gen_change_state(enabled);
5828 else if (enabled)
5829 static_branch_enable(&lru_gen_caps[i]);
5830 else
5831 static_branch_disable(&lru_gen_caps[i]);
5832 }
5833
5834 return len;
5835 }
5836
5837 static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5838
5839 static struct attribute *lru_gen_attrs[] = {
5840 &lru_gen_min_ttl_attr.attr,
5841 &lru_gen_enabled_attr.attr,
5842 NULL
5843 };
5844
5845 static const struct attribute_group lru_gen_attr_group = {
5846 .name = "lru_gen",
5847 .attrs = lru_gen_attrs,
5848 };
5849
5850 /******************************************************************************
5851 * debugfs interface
5852 ******************************************************************************/
5853
lru_gen_seq_start(struct seq_file * m,loff_t * pos)5854 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5855 {
5856 struct mem_cgroup *memcg;
5857 loff_t nr_to_skip = *pos;
5858
5859 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5860 if (!m->private)
5861 return ERR_PTR(-ENOMEM);
5862
5863 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5864 do {
5865 int nid;
5866
5867 for_each_node_state(nid, N_MEMORY) {
5868 if (!nr_to_skip--)
5869 return get_lruvec(memcg, nid);
5870 }
5871 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5872
5873 return NULL;
5874 }
5875
lru_gen_seq_stop(struct seq_file * m,void * v)5876 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5877 {
5878 if (!IS_ERR_OR_NULL(v))
5879 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5880
5881 kvfree(m->private);
5882 m->private = NULL;
5883 }
5884
lru_gen_seq_next(struct seq_file * m,void * v,loff_t * pos)5885 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5886 {
5887 int nid = lruvec_pgdat(v)->node_id;
5888 struct mem_cgroup *memcg = lruvec_memcg(v);
5889
5890 ++*pos;
5891
5892 nid = next_memory_node(nid);
5893 if (nid == MAX_NUMNODES) {
5894 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5895 if (!memcg)
5896 return NULL;
5897
5898 nid = first_memory_node;
5899 }
5900
5901 return get_lruvec(memcg, nid);
5902 }
5903
lru_gen_seq_show_full(struct seq_file * m,struct lruvec * lruvec,unsigned long max_seq,unsigned long * min_seq,unsigned long seq)5904 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5905 unsigned long max_seq, unsigned long *min_seq,
5906 unsigned long seq)
5907 {
5908 int i;
5909 int type, tier;
5910 int hist = lru_hist_from_seq(seq);
5911 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5912
5913 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5914 seq_printf(m, " %10d", tier);
5915 for (type = 0; type < ANON_AND_FILE; type++) {
5916 const char *s = " ";
5917 unsigned long n[3] = {};
5918
5919 if (seq == max_seq) {
5920 s = "RT ";
5921 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5922 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5923 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5924 s = "rep";
5925 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5926 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5927 if (tier)
5928 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5929 }
5930
5931 for (i = 0; i < 3; i++)
5932 seq_printf(m, " %10lu%c", n[i], s[i]);
5933 }
5934 seq_putc(m, '\n');
5935 }
5936
5937 seq_puts(m, " ");
5938 for (i = 0; i < NR_MM_STATS; i++) {
5939 const char *s = " ";
5940 unsigned long n = 0;
5941
5942 if (seq == max_seq && NR_HIST_GENS == 1) {
5943 s = "LOYNFA";
5944 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5945 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5946 s = "loynfa";
5947 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5948 }
5949
5950 seq_printf(m, " %10lu%c", n, s[i]);
5951 }
5952 seq_putc(m, '\n');
5953 }
5954
5955 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
lru_gen_seq_show(struct seq_file * m,void * v)5956 static int lru_gen_seq_show(struct seq_file *m, void *v)
5957 {
5958 unsigned long seq;
5959 bool full = !debugfs_real_fops(m->file)->write;
5960 struct lruvec *lruvec = v;
5961 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5962 int nid = lruvec_pgdat(lruvec)->node_id;
5963 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5964 DEFINE_MAX_SEQ(lruvec);
5965 DEFINE_MIN_SEQ(lruvec);
5966
5967 if (nid == first_memory_node) {
5968 const char *path = memcg ? m->private : "";
5969
5970 #ifdef CONFIG_MEMCG
5971 if (memcg)
5972 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5973 #endif
5974 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5975 }
5976
5977 seq_printf(m, " node %5d\n", nid);
5978
5979 if (!full)
5980 seq = min_seq[LRU_GEN_ANON];
5981 else if (max_seq >= MAX_NR_GENS)
5982 seq = max_seq - MAX_NR_GENS + 1;
5983 else
5984 seq = 0;
5985
5986 for (; seq <= max_seq; seq++) {
5987 int type, zone;
5988 int gen = lru_gen_from_seq(seq);
5989 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5990
5991 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5992
5993 for (type = 0; type < ANON_AND_FILE; type++) {
5994 unsigned long size = 0;
5995 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5996
5997 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5998 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5999
6000 seq_printf(m, " %10lu%c", size, mark);
6001 }
6002
6003 seq_putc(m, '\n');
6004
6005 if (full)
6006 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
6007 }
6008
6009 return 0;
6010 }
6011
6012 static const struct seq_operations lru_gen_seq_ops = {
6013 .start = lru_gen_seq_start,
6014 .stop = lru_gen_seq_stop,
6015 .next = lru_gen_seq_next,
6016 .show = lru_gen_seq_show,
6017 };
6018
run_aging(struct lruvec * lruvec,unsigned long seq,struct scan_control * sc,bool can_swap,bool force_scan)6019 static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
6020 bool can_swap, bool force_scan)
6021 {
6022 DEFINE_MAX_SEQ(lruvec);
6023 DEFINE_MIN_SEQ(lruvec);
6024
6025 if (seq < max_seq)
6026 return 0;
6027
6028 if (seq > max_seq)
6029 return -EINVAL;
6030
6031 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
6032 return -ERANGE;
6033
6034 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
6035
6036 return 0;
6037 }
6038
run_eviction(struct lruvec * lruvec,unsigned long seq,struct scan_control * sc,int swappiness,unsigned long nr_to_reclaim)6039 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
6040 int swappiness, unsigned long nr_to_reclaim)
6041 {
6042 DEFINE_MAX_SEQ(lruvec);
6043
6044 if (seq + MIN_NR_GENS > max_seq)
6045 return -EINVAL;
6046
6047 sc->nr_reclaimed = 0;
6048
6049 while (!signal_pending(current)) {
6050 DEFINE_MIN_SEQ(lruvec);
6051
6052 if (seq < min_seq[!swappiness])
6053 return 0;
6054
6055 if (sc->nr_reclaimed >= nr_to_reclaim)
6056 return 0;
6057
6058 if (!evict_folios(lruvec, sc, swappiness))
6059 return 0;
6060
6061 cond_resched();
6062 }
6063
6064 return -EINTR;
6065 }
6066
run_cmd(char cmd,int memcg_id,int nid,unsigned long seq,struct scan_control * sc,int swappiness,unsigned long opt)6067 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
6068 struct scan_control *sc, int swappiness, unsigned long opt)
6069 {
6070 struct lruvec *lruvec;
6071 int err = -EINVAL;
6072 struct mem_cgroup *memcg = NULL;
6073
6074 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
6075 return -EINVAL;
6076
6077 if (!mem_cgroup_disabled()) {
6078 rcu_read_lock();
6079
6080 memcg = mem_cgroup_from_id(memcg_id);
6081 if (!mem_cgroup_tryget(memcg))
6082 memcg = NULL;
6083
6084 rcu_read_unlock();
6085
6086 if (!memcg)
6087 return -EINVAL;
6088 }
6089
6090 if (memcg_id != mem_cgroup_id(memcg))
6091 goto done;
6092
6093 lruvec = get_lruvec(memcg, nid);
6094
6095 if (swappiness < 0)
6096 swappiness = get_swappiness(lruvec, sc);
6097 else if (swappiness > 200)
6098 goto done;
6099
6100 switch (cmd) {
6101 case '+':
6102 err = run_aging(lruvec, seq, sc, swappiness, opt);
6103 break;
6104 case '-':
6105 err = run_eviction(lruvec, seq, sc, swappiness, opt);
6106 break;
6107 }
6108 done:
6109 mem_cgroup_put(memcg);
6110
6111 return err;
6112 }
6113
6114 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
lru_gen_seq_write(struct file * file,const char __user * src,size_t len,loff_t * pos)6115 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
6116 size_t len, loff_t *pos)
6117 {
6118 void *buf;
6119 char *cur, *next;
6120 unsigned int flags;
6121 struct blk_plug plug;
6122 int err = -EINVAL;
6123 struct scan_control sc = {
6124 .may_writepage = true,
6125 .may_unmap = true,
6126 .may_swap = true,
6127 .reclaim_idx = MAX_NR_ZONES - 1,
6128 .gfp_mask = GFP_KERNEL,
6129 };
6130
6131 buf = kvmalloc(len + 1, GFP_KERNEL);
6132 if (!buf)
6133 return -ENOMEM;
6134
6135 if (copy_from_user(buf, src, len)) {
6136 kvfree(buf);
6137 return -EFAULT;
6138 }
6139
6140 set_task_reclaim_state(current, &sc.reclaim_state);
6141 flags = memalloc_noreclaim_save();
6142 blk_start_plug(&plug);
6143 if (!set_mm_walk(NULL, true)) {
6144 err = -ENOMEM;
6145 goto done;
6146 }
6147
6148 next = buf;
6149 next[len] = '\0';
6150
6151 while ((cur = strsep(&next, ",;\n"))) {
6152 int n;
6153 int end;
6154 char cmd;
6155 unsigned int memcg_id;
6156 unsigned int nid;
6157 unsigned long seq;
6158 unsigned int swappiness = -1;
6159 unsigned long opt = -1;
6160
6161 cur = skip_spaces(cur);
6162 if (!*cur)
6163 continue;
6164
6165 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
6166 &seq, &end, &swappiness, &end, &opt, &end);
6167 if (n < 4 || cur[end]) {
6168 err = -EINVAL;
6169 break;
6170 }
6171
6172 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
6173 if (err)
6174 break;
6175 }
6176 done:
6177 clear_mm_walk();
6178 blk_finish_plug(&plug);
6179 memalloc_noreclaim_restore(flags);
6180 set_task_reclaim_state(current, NULL);
6181
6182 kvfree(buf);
6183
6184 return err ? : len;
6185 }
6186
lru_gen_seq_open(struct inode * inode,struct file * file)6187 static int lru_gen_seq_open(struct inode *inode, struct file *file)
6188 {
6189 return seq_open(file, &lru_gen_seq_ops);
6190 }
6191
6192 static const struct file_operations lru_gen_rw_fops = {
6193 .open = lru_gen_seq_open,
6194 .read = seq_read,
6195 .write = lru_gen_seq_write,
6196 .llseek = seq_lseek,
6197 .release = seq_release,
6198 };
6199
6200 static const struct file_operations lru_gen_ro_fops = {
6201 .open = lru_gen_seq_open,
6202 .read = seq_read,
6203 .llseek = seq_lseek,
6204 .release = seq_release,
6205 };
6206
6207 /******************************************************************************
6208 * initialization
6209 ******************************************************************************/
6210
lru_gen_init_lruvec(struct lruvec * lruvec)6211 void lru_gen_init_lruvec(struct lruvec *lruvec)
6212 {
6213 int i;
6214 int gen, type, zone;
6215 struct lru_gen_folio *lrugen = &lruvec->lrugen;
6216
6217 lrugen->max_seq = MIN_NR_GENS + 1;
6218 lrugen->enabled = lru_gen_enabled();
6219
6220 for (i = 0; i <= MIN_NR_GENS + 1; i++)
6221 lrugen->timestamps[i] = jiffies;
6222
6223 for_each_gen_type_zone(gen, type, zone)
6224 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
6225
6226 lruvec->mm_state.seq = MIN_NR_GENS;
6227 }
6228
6229 #ifdef CONFIG_MEMCG
6230
lru_gen_init_pgdat(struct pglist_data * pgdat)6231 void lru_gen_init_pgdat(struct pglist_data *pgdat)
6232 {
6233 int i, j;
6234
6235 spin_lock_init(&pgdat->memcg_lru.lock);
6236
6237 for (i = 0; i < MEMCG_NR_GENS; i++) {
6238 for (j = 0; j < MEMCG_NR_BINS; j++)
6239 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
6240 }
6241 }
6242
lru_gen_init_memcg(struct mem_cgroup * memcg)6243 void lru_gen_init_memcg(struct mem_cgroup *memcg)
6244 {
6245 INIT_LIST_HEAD(&memcg->mm_list.fifo);
6246 spin_lock_init(&memcg->mm_list.lock);
6247 }
6248
lru_gen_exit_memcg(struct mem_cgroup * memcg)6249 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
6250 {
6251 int i;
6252 int nid;
6253
6254 VM_WARN_ON_ONCE(!list_empty(&memcg->mm_list.fifo));
6255
6256 for_each_node(nid) {
6257 struct lruvec *lruvec = get_lruvec(memcg, nid);
6258
6259 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
6260 sizeof(lruvec->lrugen.nr_pages)));
6261
6262 lruvec->lrugen.list.next = LIST_POISON1;
6263
6264 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
6265 bitmap_free(lruvec->mm_state.filters[i]);
6266 lruvec->mm_state.filters[i] = NULL;
6267 }
6268 }
6269 }
6270
6271 #endif /* CONFIG_MEMCG */
6272
init_lru_gen(void)6273 static int __init init_lru_gen(void)
6274 {
6275 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
6276 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
6277
6278 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
6279 pr_err("lru_gen: failed to create sysfs group\n");
6280
6281 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
6282 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
6283
6284 return 0;
6285 };
6286 late_initcall(init_lru_gen);
6287
6288 #else /* !CONFIG_LRU_GEN */
6289
lru_gen_age_node(struct pglist_data * pgdat,struct scan_control * sc)6290 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6291 {
6292 }
6293
lru_gen_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)6294 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
6295 {
6296 }
6297
lru_gen_shrink_node(struct pglist_data * pgdat,struct scan_control * sc)6298 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
6299 {
6300 }
6301
6302 #endif /* CONFIG_LRU_GEN */
6303
shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)6304 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
6305 {
6306 unsigned long nr[NR_LRU_LISTS];
6307 unsigned long targets[NR_LRU_LISTS];
6308 unsigned long nr_to_scan;
6309 enum lru_list lru;
6310 unsigned long nr_reclaimed = 0;
6311 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
6312 bool proportional_reclaim;
6313 struct blk_plug plug;
6314
6315 if (lru_gen_enabled() && !root_reclaim(sc)) {
6316 lru_gen_shrink_lruvec(lruvec, sc);
6317 return;
6318 }
6319
6320 get_scan_count(lruvec, sc, nr);
6321
6322 /* Record the original scan target for proportional adjustments later */
6323 memcpy(targets, nr, sizeof(nr));
6324
6325 /*
6326 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
6327 * event that can occur when there is little memory pressure e.g.
6328 * multiple streaming readers/writers. Hence, we do not abort scanning
6329 * when the requested number of pages are reclaimed when scanning at
6330 * DEF_PRIORITY on the assumption that the fact we are direct
6331 * reclaiming implies that kswapd is not keeping up and it is best to
6332 * do a batch of work at once. For memcg reclaim one check is made to
6333 * abort proportional reclaim if either the file or anon lru has already
6334 * dropped to zero at the first pass.
6335 */
6336 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
6337 sc->priority == DEF_PRIORITY);
6338
6339 blk_start_plug(&plug);
6340 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
6341 nr[LRU_INACTIVE_FILE]) {
6342 unsigned long nr_anon, nr_file, percentage;
6343 unsigned long nr_scanned;
6344
6345 for_each_evictable_lru(lru) {
6346 if (nr[lru]) {
6347 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
6348 nr[lru] -= nr_to_scan;
6349
6350 nr_reclaimed += shrink_list(lru, nr_to_scan,
6351 lruvec, sc);
6352 }
6353 }
6354
6355 cond_resched();
6356
6357 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
6358 continue;
6359
6360 /*
6361 * For kswapd and memcg, reclaim at least the number of pages
6362 * requested. Ensure that the anon and file LRUs are scanned
6363 * proportionally what was requested by get_scan_count(). We
6364 * stop reclaiming one LRU and reduce the amount scanning
6365 * proportional to the original scan target.
6366 */
6367 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
6368 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
6369
6370 /*
6371 * It's just vindictive to attack the larger once the smaller
6372 * has gone to zero. And given the way we stop scanning the
6373 * smaller below, this makes sure that we only make one nudge
6374 * towards proportionality once we've got nr_to_reclaim.
6375 */
6376 if (!nr_file || !nr_anon)
6377 break;
6378
6379 if (nr_file > nr_anon) {
6380 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
6381 targets[LRU_ACTIVE_ANON] + 1;
6382 lru = LRU_BASE;
6383 percentage = nr_anon * 100 / scan_target;
6384 } else {
6385 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
6386 targets[LRU_ACTIVE_FILE] + 1;
6387 lru = LRU_FILE;
6388 percentage = nr_file * 100 / scan_target;
6389 }
6390
6391 /* Stop scanning the smaller of the LRU */
6392 nr[lru] = 0;
6393 nr[lru + LRU_ACTIVE] = 0;
6394
6395 /*
6396 * Recalculate the other LRU scan count based on its original
6397 * scan target and the percentage scanning already complete
6398 */
6399 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
6400 nr_scanned = targets[lru] - nr[lru];
6401 nr[lru] = targets[lru] * (100 - percentage) / 100;
6402 nr[lru] -= min(nr[lru], nr_scanned);
6403
6404 lru += LRU_ACTIVE;
6405 nr_scanned = targets[lru] - nr[lru];
6406 nr[lru] = targets[lru] * (100 - percentage) / 100;
6407 nr[lru] -= min(nr[lru], nr_scanned);
6408 }
6409 blk_finish_plug(&plug);
6410 sc->nr_reclaimed += nr_reclaimed;
6411
6412 /*
6413 * Even if we did not try to evict anon pages at all, we want to
6414 * rebalance the anon lru active/inactive ratio.
6415 */
6416 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
6417 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6418 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6419 sc, LRU_ACTIVE_ANON);
6420 }
6421
6422 /* Use reclaim/compaction for costly allocs or under memory pressure */
in_reclaim_compaction(struct scan_control * sc)6423 static bool in_reclaim_compaction(struct scan_control *sc)
6424 {
6425 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
6426 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
6427 sc->priority < DEF_PRIORITY - 2))
6428 return true;
6429
6430 return false;
6431 }
6432
6433 /*
6434 * Reclaim/compaction is used for high-order allocation requests. It reclaims
6435 * order-0 pages before compacting the zone. should_continue_reclaim() returns
6436 * true if more pages should be reclaimed such that when the page allocator
6437 * calls try_to_compact_pages() that it will have enough free pages to succeed.
6438 * It will give up earlier than that if there is difficulty reclaiming pages.
6439 */
should_continue_reclaim(struct pglist_data * pgdat,unsigned long nr_reclaimed,struct scan_control * sc)6440 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
6441 unsigned long nr_reclaimed,
6442 struct scan_control *sc)
6443 {
6444 unsigned long pages_for_compaction;
6445 unsigned long inactive_lru_pages;
6446 int z;
6447
6448 /* If not in reclaim/compaction mode, stop */
6449 if (!in_reclaim_compaction(sc))
6450 return false;
6451
6452 /*
6453 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
6454 * number of pages that were scanned. This will return to the caller
6455 * with the risk reclaim/compaction and the resulting allocation attempt
6456 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
6457 * allocations through requiring that the full LRU list has been scanned
6458 * first, by assuming that zero delta of sc->nr_scanned means full LRU
6459 * scan, but that approximation was wrong, and there were corner cases
6460 * where always a non-zero amount of pages were scanned.
6461 */
6462 if (!nr_reclaimed)
6463 return false;
6464
6465 /* If compaction would go ahead or the allocation would succeed, stop */
6466 for (z = 0; z <= sc->reclaim_idx; z++) {
6467 struct zone *zone = &pgdat->node_zones[z];
6468 if (!managed_zone(zone))
6469 continue;
6470
6471 /* Allocation can already succeed, nothing to do */
6472 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6473 sc->reclaim_idx, 0))
6474 return false;
6475
6476 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
6477 return false;
6478 }
6479
6480 /*
6481 * If we have not reclaimed enough pages for compaction and the
6482 * inactive lists are large enough, continue reclaiming
6483 */
6484 pages_for_compaction = compact_gap(sc->order);
6485 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
6486 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
6487 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
6488
6489 return inactive_lru_pages > pages_for_compaction;
6490 }
6491
shrink_node_memcgs(pg_data_t * pgdat,struct scan_control * sc)6492 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
6493 {
6494 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
6495 struct mem_cgroup *memcg;
6496
6497 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
6498 do {
6499 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6500 unsigned long reclaimed;
6501 unsigned long scanned;
6502
6503 /*
6504 * This loop can become CPU-bound when target memcgs
6505 * aren't eligible for reclaim - either because they
6506 * don't have any reclaimable pages, or because their
6507 * memory is explicitly protected. Avoid soft lockups.
6508 */
6509 cond_resched();
6510
6511 mem_cgroup_calculate_protection(target_memcg, memcg);
6512
6513 if (mem_cgroup_below_min(target_memcg, memcg)) {
6514 /*
6515 * Hard protection.
6516 * If there is no reclaimable memory, OOM.
6517 */
6518 continue;
6519 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
6520 /*
6521 * Soft protection.
6522 * Respect the protection only as long as
6523 * there is an unprotected supply
6524 * of reclaimable memory from other cgroups.
6525 */
6526 if (!sc->memcg_low_reclaim) {
6527 sc->memcg_low_skipped = 1;
6528 continue;
6529 }
6530 memcg_memory_event(memcg, MEMCG_LOW);
6531 }
6532
6533 reclaimed = sc->nr_reclaimed;
6534 scanned = sc->nr_scanned;
6535
6536 shrink_lruvec(lruvec, sc);
6537
6538 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
6539 sc->priority);
6540
6541 /* Record the group's reclaim efficiency */
6542 if (!sc->proactive)
6543 vmpressure(sc->gfp_mask, memcg, false,
6544 sc->nr_scanned - scanned,
6545 sc->nr_reclaimed - reclaimed);
6546
6547 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
6548 }
6549
shrink_node(pg_data_t * pgdat,struct scan_control * sc)6550 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6551 {
6552 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
6553 struct lruvec *target_lruvec;
6554 bool reclaimable = false;
6555
6556 if (lru_gen_enabled() && root_reclaim(sc)) {
6557 lru_gen_shrink_node(pgdat, sc);
6558 return;
6559 }
6560
6561 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
6562
6563 again:
6564 memset(&sc->nr, 0, sizeof(sc->nr));
6565
6566 nr_reclaimed = sc->nr_reclaimed;
6567 nr_scanned = sc->nr_scanned;
6568
6569 prepare_scan_count(pgdat, sc);
6570
6571 shrink_node_memcgs(pgdat, sc);
6572
6573 flush_reclaim_state(sc);
6574
6575 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
6576
6577 /* Record the subtree's reclaim efficiency */
6578 if (!sc->proactive)
6579 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
6580 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
6581
6582 if (nr_node_reclaimed)
6583 reclaimable = true;
6584
6585 if (current_is_kswapd()) {
6586 /*
6587 * If reclaim is isolating dirty pages under writeback,
6588 * it implies that the long-lived page allocation rate
6589 * is exceeding the page laundering rate. Either the
6590 * global limits are not being effective at throttling
6591 * processes due to the page distribution throughout
6592 * zones or there is heavy usage of a slow backing
6593 * device. The only option is to throttle from reclaim
6594 * context which is not ideal as there is no guarantee
6595 * the dirtying process is throttled in the same way
6596 * balance_dirty_pages() manages.
6597 *
6598 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6599 * count the number of pages under pages flagged for
6600 * immediate reclaim and stall if any are encountered
6601 * in the nr_immediate check below.
6602 */
6603 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6604 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6605
6606 /* Allow kswapd to start writing pages during reclaim.*/
6607 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
6608 set_bit(PGDAT_DIRTY, &pgdat->flags);
6609
6610 /*
6611 * If kswapd scans pages marked for immediate
6612 * reclaim and under writeback (nr_immediate), it
6613 * implies that pages are cycling through the LRU
6614 * faster than they are written so forcibly stall
6615 * until some pages complete writeback.
6616 */
6617 if (sc->nr.immediate)
6618 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6619 }
6620
6621 /*
6622 * Tag a node/memcg as congested if all the dirty pages were marked
6623 * for writeback and immediate reclaim (counted in nr.congested).
6624 *
6625 * Legacy memcg will stall in page writeback so avoid forcibly
6626 * stalling in reclaim_throttle().
6627 */
6628 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
6629 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
6630 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
6631
6632 if (current_is_kswapd())
6633 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
6634 }
6635
6636 /*
6637 * Stall direct reclaim for IO completions if the lruvec is
6638 * node is congested. Allow kswapd to continue until it
6639 * starts encountering unqueued dirty pages or cycling through
6640 * the LRU too quickly.
6641 */
6642 if (!current_is_kswapd() && current_may_throttle() &&
6643 !sc->hibernation_mode &&
6644 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
6645 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
6646 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6647
6648 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
6649 goto again;
6650
6651 /*
6652 * Kswapd gives up on balancing particular nodes after too
6653 * many failures to reclaim anything from them and goes to
6654 * sleep. On reclaim progress, reset the failure counter. A
6655 * successful direct reclaim run will revive a dormant kswapd.
6656 */
6657 if (reclaimable)
6658 pgdat->kswapd_failures = 0;
6659 }
6660
6661 /*
6662 * Returns true if compaction should go ahead for a costly-order request, or
6663 * the allocation would already succeed without compaction. Return false if we
6664 * should reclaim first.
6665 */
compaction_ready(struct zone * zone,struct scan_control * sc)6666 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6667 {
6668 unsigned long watermark;
6669
6670 if (!gfp_compaction_allowed(sc->gfp_mask))
6671 return false;
6672
6673 /* Allocation can already succeed, nothing to do */
6674 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6675 sc->reclaim_idx, 0))
6676 return true;
6677
6678 /* Compaction cannot yet proceed. Do reclaim. */
6679 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6680 return false;
6681
6682 /*
6683 * Compaction is already possible, but it takes time to run and there
6684 * are potentially other callers using the pages just freed. So proceed
6685 * with reclaim to make a buffer of free pages available to give
6686 * compaction a reasonable chance of completing and allocating the page.
6687 * Note that we won't actually reclaim the whole buffer in one attempt
6688 * as the target watermark in should_continue_reclaim() is lower. But if
6689 * we are already above the high+gap watermark, don't reclaim at all.
6690 */
6691 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6692
6693 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6694 }
6695
consider_reclaim_throttle(pg_data_t * pgdat,struct scan_control * sc)6696 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6697 {
6698 /*
6699 * If reclaim is making progress greater than 12% efficiency then
6700 * wake all the NOPROGRESS throttled tasks.
6701 */
6702 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6703 wait_queue_head_t *wqh;
6704
6705 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6706 if (waitqueue_active(wqh))
6707 wake_up(wqh);
6708
6709 return;
6710 }
6711
6712 /*
6713 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6714 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6715 * under writeback and marked for immediate reclaim at the tail of the
6716 * LRU.
6717 */
6718 if (current_is_kswapd() || cgroup_reclaim(sc))
6719 return;
6720
6721 /* Throttle if making no progress at high prioities. */
6722 if (sc->priority == 1 && !sc->nr_reclaimed)
6723 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6724 }
6725
6726 /*
6727 * This is the direct reclaim path, for page-allocating processes. We only
6728 * try to reclaim pages from zones which will satisfy the caller's allocation
6729 * request.
6730 *
6731 * If a zone is deemed to be full of pinned pages then just give it a light
6732 * scan then give up on it.
6733 */
shrink_zones(struct zonelist * zonelist,struct scan_control * sc)6734 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6735 {
6736 struct zoneref *z;
6737 struct zone *zone;
6738 unsigned long nr_soft_reclaimed;
6739 unsigned long nr_soft_scanned;
6740 gfp_t orig_mask;
6741 pg_data_t *last_pgdat = NULL;
6742 pg_data_t *first_pgdat = NULL;
6743
6744 /*
6745 * If the number of buffer_heads in the machine exceeds the maximum
6746 * allowed level, force direct reclaim to scan the highmem zone as
6747 * highmem pages could be pinning lowmem pages storing buffer_heads
6748 */
6749 orig_mask = sc->gfp_mask;
6750 if (buffer_heads_over_limit) {
6751 sc->gfp_mask |= __GFP_HIGHMEM;
6752 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6753 }
6754
6755 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6756 sc->reclaim_idx, sc->nodemask) {
6757 /*
6758 * Take care memory controller reclaiming has small influence
6759 * to global LRU.
6760 */
6761 if (!cgroup_reclaim(sc)) {
6762 if (!cpuset_zone_allowed(zone,
6763 GFP_KERNEL | __GFP_HARDWALL))
6764 continue;
6765
6766 /*
6767 * If we already have plenty of memory free for
6768 * compaction in this zone, don't free any more.
6769 * Even though compaction is invoked for any
6770 * non-zero order, only frequent costly order
6771 * reclamation is disruptive enough to become a
6772 * noticeable problem, like transparent huge
6773 * page allocations.
6774 */
6775 if (IS_ENABLED(CONFIG_COMPACTION) &&
6776 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6777 compaction_ready(zone, sc)) {
6778 sc->compaction_ready = true;
6779 continue;
6780 }
6781
6782 /*
6783 * Shrink each node in the zonelist once. If the
6784 * zonelist is ordered by zone (not the default) then a
6785 * node may be shrunk multiple times but in that case
6786 * the user prefers lower zones being preserved.
6787 */
6788 if (zone->zone_pgdat == last_pgdat)
6789 continue;
6790
6791 /*
6792 * This steals pages from memory cgroups over softlimit
6793 * and returns the number of reclaimed pages and
6794 * scanned pages. This works for global memory pressure
6795 * and balancing, not for a memcg's limit.
6796 */
6797 nr_soft_scanned = 0;
6798 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6799 sc->order, sc->gfp_mask,
6800 &nr_soft_scanned);
6801 sc->nr_reclaimed += nr_soft_reclaimed;
6802 sc->nr_scanned += nr_soft_scanned;
6803 /* need some check for avoid more shrink_zone() */
6804 }
6805
6806 if (!first_pgdat)
6807 first_pgdat = zone->zone_pgdat;
6808
6809 /* See comment about same check for global reclaim above */
6810 if (zone->zone_pgdat == last_pgdat)
6811 continue;
6812 last_pgdat = zone->zone_pgdat;
6813 shrink_node(zone->zone_pgdat, sc);
6814 }
6815
6816 if (first_pgdat)
6817 consider_reclaim_throttle(first_pgdat, sc);
6818
6819 /*
6820 * Restore to original mask to avoid the impact on the caller if we
6821 * promoted it to __GFP_HIGHMEM.
6822 */
6823 sc->gfp_mask = orig_mask;
6824 }
6825
snapshot_refaults(struct mem_cgroup * target_memcg,pg_data_t * pgdat)6826 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6827 {
6828 struct lruvec *target_lruvec;
6829 unsigned long refaults;
6830
6831 if (lru_gen_enabled())
6832 return;
6833
6834 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6835 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6836 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6837 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6838 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6839 }
6840
6841 /*
6842 * This is the main entry point to direct page reclaim.
6843 *
6844 * If a full scan of the inactive list fails to free enough memory then we
6845 * are "out of memory" and something needs to be killed.
6846 *
6847 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6848 * high - the zone may be full of dirty or under-writeback pages, which this
6849 * caller can't do much about. We kick the writeback threads and take explicit
6850 * naps in the hope that some of these pages can be written. But if the
6851 * allocating task holds filesystem locks which prevent writeout this might not
6852 * work, and the allocation attempt will fail.
6853 *
6854 * returns: 0, if no pages reclaimed
6855 * else, the number of pages reclaimed
6856 */
do_try_to_free_pages(struct zonelist * zonelist,struct scan_control * sc)6857 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6858 struct scan_control *sc)
6859 {
6860 int initial_priority = sc->priority;
6861 pg_data_t *last_pgdat;
6862 struct zoneref *z;
6863 struct zone *zone;
6864 retry:
6865 delayacct_freepages_start();
6866
6867 if (!cgroup_reclaim(sc))
6868 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6869
6870 do {
6871 if (!sc->proactive)
6872 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6873 sc->priority);
6874 sc->nr_scanned = 0;
6875 shrink_zones(zonelist, sc);
6876
6877 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6878 break;
6879
6880 if (sc->compaction_ready)
6881 break;
6882
6883 /*
6884 * If we're getting trouble reclaiming, start doing
6885 * writepage even in laptop mode.
6886 */
6887 if (sc->priority < DEF_PRIORITY - 2)
6888 sc->may_writepage = 1;
6889 } while (--sc->priority >= 0);
6890
6891 last_pgdat = NULL;
6892 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6893 sc->nodemask) {
6894 if (zone->zone_pgdat == last_pgdat)
6895 continue;
6896 last_pgdat = zone->zone_pgdat;
6897
6898 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6899
6900 if (cgroup_reclaim(sc)) {
6901 struct lruvec *lruvec;
6902
6903 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6904 zone->zone_pgdat);
6905 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6906 }
6907 }
6908
6909 delayacct_freepages_end();
6910
6911 if (sc->nr_reclaimed)
6912 return sc->nr_reclaimed;
6913
6914 /* Aborted reclaim to try compaction? don't OOM, then */
6915 if (sc->compaction_ready)
6916 return 1;
6917
6918 /*
6919 * We make inactive:active ratio decisions based on the node's
6920 * composition of memory, but a restrictive reclaim_idx or a
6921 * memory.low cgroup setting can exempt large amounts of
6922 * memory from reclaim. Neither of which are very common, so
6923 * instead of doing costly eligibility calculations of the
6924 * entire cgroup subtree up front, we assume the estimates are
6925 * good, and retry with forcible deactivation if that fails.
6926 */
6927 if (sc->skipped_deactivate) {
6928 sc->priority = initial_priority;
6929 sc->force_deactivate = 1;
6930 sc->skipped_deactivate = 0;
6931 goto retry;
6932 }
6933
6934 /* Untapped cgroup reserves? Don't OOM, retry. */
6935 if (sc->memcg_low_skipped) {
6936 sc->priority = initial_priority;
6937 sc->force_deactivate = 0;
6938 sc->memcg_low_reclaim = 1;
6939 sc->memcg_low_skipped = 0;
6940 goto retry;
6941 }
6942
6943 return 0;
6944 }
6945
allow_direct_reclaim(pg_data_t * pgdat)6946 static bool allow_direct_reclaim(pg_data_t *pgdat)
6947 {
6948 struct zone *zone;
6949 unsigned long pfmemalloc_reserve = 0;
6950 unsigned long free_pages = 0;
6951 int i;
6952 bool wmark_ok;
6953
6954 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6955 return true;
6956
6957 for (i = 0; i <= ZONE_NORMAL; i++) {
6958 zone = &pgdat->node_zones[i];
6959 if (!managed_zone(zone))
6960 continue;
6961
6962 if (!zone_reclaimable_pages(zone))
6963 continue;
6964
6965 pfmemalloc_reserve += min_wmark_pages(zone);
6966 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6967 }
6968
6969 /* If there are no reserves (unexpected config) then do not throttle */
6970 if (!pfmemalloc_reserve)
6971 return true;
6972
6973 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6974
6975 /* kswapd must be awake if processes are being throttled */
6976 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6977 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6978 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6979
6980 wake_up_interruptible(&pgdat->kswapd_wait);
6981 }
6982
6983 return wmark_ok;
6984 }
6985
6986 /*
6987 * Throttle direct reclaimers if backing storage is backed by the network
6988 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6989 * depleted. kswapd will continue to make progress and wake the processes
6990 * when the low watermark is reached.
6991 *
6992 * Returns true if a fatal signal was delivered during throttling. If this
6993 * happens, the page allocator should not consider triggering the OOM killer.
6994 */
throttle_direct_reclaim(gfp_t gfp_mask,struct zonelist * zonelist,nodemask_t * nodemask)6995 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6996 nodemask_t *nodemask)
6997 {
6998 struct zoneref *z;
6999 struct zone *zone;
7000 pg_data_t *pgdat = NULL;
7001
7002 /*
7003 * Kernel threads should not be throttled as they may be indirectly
7004 * responsible for cleaning pages necessary for reclaim to make forward
7005 * progress. kjournald for example may enter direct reclaim while
7006 * committing a transaction where throttling it could forcing other
7007 * processes to block on log_wait_commit().
7008 */
7009 if (current->flags & PF_KTHREAD)
7010 goto out;
7011
7012 /*
7013 * If a fatal signal is pending, this process should not throttle.
7014 * It should return quickly so it can exit and free its memory
7015 */
7016 if (fatal_signal_pending(current))
7017 goto out;
7018
7019 /*
7020 * Check if the pfmemalloc reserves are ok by finding the first node
7021 * with a usable ZONE_NORMAL or lower zone. The expectation is that
7022 * GFP_KERNEL will be required for allocating network buffers when
7023 * swapping over the network so ZONE_HIGHMEM is unusable.
7024 *
7025 * Throttling is based on the first usable node and throttled processes
7026 * wait on a queue until kswapd makes progress and wakes them. There
7027 * is an affinity then between processes waking up and where reclaim
7028 * progress has been made assuming the process wakes on the same node.
7029 * More importantly, processes running on remote nodes will not compete
7030 * for remote pfmemalloc reserves and processes on different nodes
7031 * should make reasonable progress.
7032 */
7033 for_each_zone_zonelist_nodemask(zone, z, zonelist,
7034 gfp_zone(gfp_mask), nodemask) {
7035 if (zone_idx(zone) > ZONE_NORMAL)
7036 continue;
7037
7038 /* Throttle based on the first usable node */
7039 pgdat = zone->zone_pgdat;
7040 if (allow_direct_reclaim(pgdat))
7041 goto out;
7042 break;
7043 }
7044
7045 /* If no zone was usable by the allocation flags then do not throttle */
7046 if (!pgdat)
7047 goto out;
7048
7049 /* Account for the throttling */
7050 count_vm_event(PGSCAN_DIRECT_THROTTLE);
7051
7052 /*
7053 * If the caller cannot enter the filesystem, it's possible that it
7054 * is due to the caller holding an FS lock or performing a journal
7055 * transaction in the case of a filesystem like ext[3|4]. In this case,
7056 * it is not safe to block on pfmemalloc_wait as kswapd could be
7057 * blocked waiting on the same lock. Instead, throttle for up to a
7058 * second before continuing.
7059 */
7060 if (!(gfp_mask & __GFP_FS))
7061 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
7062 allow_direct_reclaim(pgdat), HZ);
7063 else
7064 /* Throttle until kswapd wakes the process */
7065 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
7066 allow_direct_reclaim(pgdat));
7067
7068 if (fatal_signal_pending(current))
7069 return true;
7070
7071 out:
7072 return false;
7073 }
7074
try_to_free_pages(struct zonelist * zonelist,int order,gfp_t gfp_mask,nodemask_t * nodemask)7075 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
7076 gfp_t gfp_mask, nodemask_t *nodemask)
7077 {
7078 unsigned long nr_reclaimed;
7079 struct scan_control sc = {
7080 .nr_to_reclaim = SWAP_CLUSTER_MAX,
7081 .gfp_mask = current_gfp_context(gfp_mask),
7082 .reclaim_idx = gfp_zone(gfp_mask),
7083 .order = order,
7084 .nodemask = nodemask,
7085 .priority = DEF_PRIORITY,
7086 .may_writepage = !laptop_mode,
7087 .may_unmap = 1,
7088 .may_swap = 1,
7089 };
7090
7091 /*
7092 * scan_control uses s8 fields for order, priority, and reclaim_idx.
7093 * Confirm they are large enough for max values.
7094 */
7095 BUILD_BUG_ON(MAX_ORDER >= S8_MAX);
7096 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
7097 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
7098
7099 /*
7100 * Do not enter reclaim if fatal signal was delivered while throttled.
7101 * 1 is returned so that the page allocator does not OOM kill at this
7102 * point.
7103 */
7104 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
7105 return 1;
7106
7107 set_task_reclaim_state(current, &sc.reclaim_state);
7108 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
7109
7110 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7111
7112 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
7113 set_task_reclaim_state(current, NULL);
7114
7115 return nr_reclaimed;
7116 }
7117
7118 #ifdef CONFIG_MEMCG
7119
7120 /* Only used by soft limit reclaim. Do not reuse for anything else. */
mem_cgroup_shrink_node(struct mem_cgroup * memcg,gfp_t gfp_mask,bool noswap,pg_data_t * pgdat,unsigned long * nr_scanned)7121 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
7122 gfp_t gfp_mask, bool noswap,
7123 pg_data_t *pgdat,
7124 unsigned long *nr_scanned)
7125 {
7126 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
7127 struct scan_control sc = {
7128 .nr_to_reclaim = SWAP_CLUSTER_MAX,
7129 .target_mem_cgroup = memcg,
7130 .may_writepage = !laptop_mode,
7131 .may_unmap = 1,
7132 .reclaim_idx = MAX_NR_ZONES - 1,
7133 .may_swap = !noswap,
7134 };
7135
7136 WARN_ON_ONCE(!current->reclaim_state);
7137
7138 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
7139 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
7140
7141 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
7142 sc.gfp_mask);
7143
7144 /*
7145 * NOTE: Although we can get the priority field, using it
7146 * here is not a good idea, since it limits the pages we can scan.
7147 * if we don't reclaim here, the shrink_node from balance_pgdat
7148 * will pick up pages from other mem cgroup's as well. We hack
7149 * the priority and make it zero.
7150 */
7151 shrink_lruvec(lruvec, &sc);
7152
7153 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
7154
7155 *nr_scanned = sc.nr_scanned;
7156
7157 return sc.nr_reclaimed;
7158 }
7159
try_to_free_mem_cgroup_pages(struct mem_cgroup * memcg,unsigned long nr_pages,gfp_t gfp_mask,unsigned int reclaim_options)7160 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
7161 unsigned long nr_pages,
7162 gfp_t gfp_mask,
7163 unsigned int reclaim_options)
7164 {
7165 unsigned long nr_reclaimed;
7166 unsigned int noreclaim_flag;
7167 struct scan_control sc = {
7168 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7169 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
7170 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
7171 .reclaim_idx = MAX_NR_ZONES - 1,
7172 .target_mem_cgroup = memcg,
7173 .priority = DEF_PRIORITY,
7174 .may_writepage = !laptop_mode,
7175 .may_unmap = 1,
7176 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
7177 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
7178 };
7179 /*
7180 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
7181 * equal pressure on all the nodes. This is based on the assumption that
7182 * the reclaim does not bail out early.
7183 */
7184 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7185
7186 set_task_reclaim_state(current, &sc.reclaim_state);
7187 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
7188 noreclaim_flag = memalloc_noreclaim_save();
7189
7190 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7191
7192 memalloc_noreclaim_restore(noreclaim_flag);
7193 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
7194 set_task_reclaim_state(current, NULL);
7195
7196 return nr_reclaimed;
7197 }
7198 #endif
7199
kswapd_age_node(struct pglist_data * pgdat,struct scan_control * sc)7200 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
7201 {
7202 struct mem_cgroup *memcg;
7203 struct lruvec *lruvec;
7204
7205 if (lru_gen_enabled()) {
7206 lru_gen_age_node(pgdat, sc);
7207 return;
7208 }
7209
7210 if (!can_age_anon_pages(pgdat, sc))
7211 return;
7212
7213 lruvec = mem_cgroup_lruvec(NULL, pgdat);
7214 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
7215 return;
7216
7217 memcg = mem_cgroup_iter(NULL, NULL, NULL);
7218 do {
7219 lruvec = mem_cgroup_lruvec(memcg, pgdat);
7220 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
7221 sc, LRU_ACTIVE_ANON);
7222 memcg = mem_cgroup_iter(NULL, memcg, NULL);
7223 } while (memcg);
7224 }
7225
pgdat_watermark_boosted(pg_data_t * pgdat,int highest_zoneidx)7226 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
7227 {
7228 int i;
7229 struct zone *zone;
7230
7231 /*
7232 * Check for watermark boosts top-down as the higher zones
7233 * are more likely to be boosted. Both watermarks and boosts
7234 * should not be checked at the same time as reclaim would
7235 * start prematurely when there is no boosting and a lower
7236 * zone is balanced.
7237 */
7238 for (i = highest_zoneidx; i >= 0; i--) {
7239 zone = pgdat->node_zones + i;
7240 if (!managed_zone(zone))
7241 continue;
7242
7243 if (zone->watermark_boost)
7244 return true;
7245 }
7246
7247 return false;
7248 }
7249
7250 /*
7251 * Returns true if there is an eligible zone balanced for the request order
7252 * and highest_zoneidx
7253 */
pgdat_balanced(pg_data_t * pgdat,int order,int highest_zoneidx)7254 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
7255 {
7256 int i;
7257 unsigned long mark = -1;
7258 struct zone *zone;
7259
7260 /*
7261 * Check watermarks bottom-up as lower zones are more likely to
7262 * meet watermarks.
7263 */
7264 for (i = 0; i <= highest_zoneidx; i++) {
7265 zone = pgdat->node_zones + i;
7266
7267 if (!managed_zone(zone))
7268 continue;
7269
7270 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
7271 mark = wmark_pages(zone, WMARK_PROMO);
7272 else
7273 mark = high_wmark_pages(zone);
7274 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
7275 return true;
7276 }
7277
7278 /*
7279 * If a node has no managed zone within highest_zoneidx, it does not
7280 * need balancing by definition. This can happen if a zone-restricted
7281 * allocation tries to wake a remote kswapd.
7282 */
7283 if (mark == -1)
7284 return true;
7285
7286 return false;
7287 }
7288
7289 /* Clear pgdat state for congested, dirty or under writeback. */
clear_pgdat_congested(pg_data_t * pgdat)7290 static void clear_pgdat_congested(pg_data_t *pgdat)
7291 {
7292 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
7293
7294 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
7295 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
7296 clear_bit(PGDAT_DIRTY, &pgdat->flags);
7297 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
7298 }
7299
7300 /*
7301 * Prepare kswapd for sleeping. This verifies that there are no processes
7302 * waiting in throttle_direct_reclaim() and that watermarks have been met.
7303 *
7304 * Returns true if kswapd is ready to sleep
7305 */
prepare_kswapd_sleep(pg_data_t * pgdat,int order,int highest_zoneidx)7306 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
7307 int highest_zoneidx)
7308 {
7309 /*
7310 * The throttled processes are normally woken up in balance_pgdat() as
7311 * soon as allow_direct_reclaim() is true. But there is a potential
7312 * race between when kswapd checks the watermarks and a process gets
7313 * throttled. There is also a potential race if processes get
7314 * throttled, kswapd wakes, a large process exits thereby balancing the
7315 * zones, which causes kswapd to exit balance_pgdat() before reaching
7316 * the wake up checks. If kswapd is going to sleep, no process should
7317 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
7318 * the wake up is premature, processes will wake kswapd and get
7319 * throttled again. The difference from wake ups in balance_pgdat() is
7320 * that here we are under prepare_to_wait().
7321 */
7322 if (waitqueue_active(&pgdat->pfmemalloc_wait))
7323 wake_up_all(&pgdat->pfmemalloc_wait);
7324
7325 /* Hopeless node, leave it to direct reclaim */
7326 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
7327 return true;
7328
7329 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
7330 clear_pgdat_congested(pgdat);
7331 return true;
7332 }
7333
7334 return false;
7335 }
7336
7337 /*
7338 * kswapd shrinks a node of pages that are at or below the highest usable
7339 * zone that is currently unbalanced.
7340 *
7341 * Returns true if kswapd scanned at least the requested number of pages to
7342 * reclaim or if the lack of progress was due to pages under writeback.
7343 * This is used to determine if the scanning priority needs to be raised.
7344 */
kswapd_shrink_node(pg_data_t * pgdat,struct scan_control * sc)7345 static bool kswapd_shrink_node(pg_data_t *pgdat,
7346 struct scan_control *sc)
7347 {
7348 struct zone *zone;
7349 int z;
7350 unsigned long nr_reclaimed = sc->nr_reclaimed;
7351
7352 /* Reclaim a number of pages proportional to the number of zones */
7353 sc->nr_to_reclaim = 0;
7354 for (z = 0; z <= sc->reclaim_idx; z++) {
7355 zone = pgdat->node_zones + z;
7356 if (!managed_zone(zone))
7357 continue;
7358
7359 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
7360 }
7361
7362 /*
7363 * Historically care was taken to put equal pressure on all zones but
7364 * now pressure is applied based on node LRU order.
7365 */
7366 shrink_node(pgdat, sc);
7367
7368 /*
7369 * Fragmentation may mean that the system cannot be rebalanced for
7370 * high-order allocations. If twice the allocation size has been
7371 * reclaimed then recheck watermarks only at order-0 to prevent
7372 * excessive reclaim. Assume that a process requested a high-order
7373 * can direct reclaim/compact.
7374 */
7375 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
7376 sc->order = 0;
7377
7378 /* account for progress from mm_account_reclaimed_pages() */
7379 return max(sc->nr_scanned, sc->nr_reclaimed - nr_reclaimed) >= sc->nr_to_reclaim;
7380 }
7381
7382 /* Page allocator PCP high watermark is lowered if reclaim is active. */
7383 static inline void
update_reclaim_active(pg_data_t * pgdat,int highest_zoneidx,bool active)7384 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
7385 {
7386 int i;
7387 struct zone *zone;
7388
7389 for (i = 0; i <= highest_zoneidx; i++) {
7390 zone = pgdat->node_zones + i;
7391
7392 if (!managed_zone(zone))
7393 continue;
7394
7395 if (active)
7396 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
7397 else
7398 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
7399 }
7400 }
7401
7402 static inline void
set_reclaim_active(pg_data_t * pgdat,int highest_zoneidx)7403 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
7404 {
7405 update_reclaim_active(pgdat, highest_zoneidx, true);
7406 }
7407
7408 static inline void
clear_reclaim_active(pg_data_t * pgdat,int highest_zoneidx)7409 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
7410 {
7411 update_reclaim_active(pgdat, highest_zoneidx, false);
7412 }
7413
7414 /*
7415 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
7416 * that are eligible for use by the caller until at least one zone is
7417 * balanced.
7418 *
7419 * Returns the order kswapd finished reclaiming at.
7420 *
7421 * kswapd scans the zones in the highmem->normal->dma direction. It skips
7422 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
7423 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
7424 * or lower is eligible for reclaim until at least one usable zone is
7425 * balanced.
7426 */
balance_pgdat(pg_data_t * pgdat,int order,int highest_zoneidx)7427 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
7428 {
7429 int i;
7430 unsigned long nr_soft_reclaimed;
7431 unsigned long nr_soft_scanned;
7432 unsigned long pflags;
7433 unsigned long nr_boost_reclaim;
7434 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
7435 bool boosted;
7436 struct zone *zone;
7437 struct scan_control sc = {
7438 .gfp_mask = GFP_KERNEL,
7439 .order = order,
7440 .may_unmap = 1,
7441 };
7442
7443 set_task_reclaim_state(current, &sc.reclaim_state);
7444 psi_memstall_enter(&pflags);
7445 __fs_reclaim_acquire(_THIS_IP_);
7446
7447 count_vm_event(PAGEOUTRUN);
7448
7449 /*
7450 * Account for the reclaim boost. Note that the zone boost is left in
7451 * place so that parallel allocations that are near the watermark will
7452 * stall or direct reclaim until kswapd is finished.
7453 */
7454 nr_boost_reclaim = 0;
7455 for (i = 0; i <= highest_zoneidx; i++) {
7456 zone = pgdat->node_zones + i;
7457 if (!managed_zone(zone))
7458 continue;
7459
7460 nr_boost_reclaim += zone->watermark_boost;
7461 zone_boosts[i] = zone->watermark_boost;
7462 }
7463 boosted = nr_boost_reclaim;
7464
7465 restart:
7466 set_reclaim_active(pgdat, highest_zoneidx);
7467 sc.priority = DEF_PRIORITY;
7468 do {
7469 unsigned long nr_reclaimed = sc.nr_reclaimed;
7470 bool raise_priority = true;
7471 bool balanced;
7472 bool ret;
7473
7474 sc.reclaim_idx = highest_zoneidx;
7475
7476 /*
7477 * If the number of buffer_heads exceeds the maximum allowed
7478 * then consider reclaiming from all zones. This has a dual
7479 * purpose -- on 64-bit systems it is expected that
7480 * buffer_heads are stripped during active rotation. On 32-bit
7481 * systems, highmem pages can pin lowmem memory and shrinking
7482 * buffers can relieve lowmem pressure. Reclaim may still not
7483 * go ahead if all eligible zones for the original allocation
7484 * request are balanced to avoid excessive reclaim from kswapd.
7485 */
7486 if (buffer_heads_over_limit) {
7487 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7488 zone = pgdat->node_zones + i;
7489 if (!managed_zone(zone))
7490 continue;
7491
7492 sc.reclaim_idx = i;
7493 break;
7494 }
7495 }
7496
7497 /*
7498 * If the pgdat is imbalanced then ignore boosting and preserve
7499 * the watermarks for a later time and restart. Note that the
7500 * zone watermarks will be still reset at the end of balancing
7501 * on the grounds that the normal reclaim should be enough to
7502 * re-evaluate if boosting is required when kswapd next wakes.
7503 */
7504 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
7505 if (!balanced && nr_boost_reclaim) {
7506 nr_boost_reclaim = 0;
7507 goto restart;
7508 }
7509
7510 /*
7511 * If boosting is not active then only reclaim if there are no
7512 * eligible zones. Note that sc.reclaim_idx is not used as
7513 * buffer_heads_over_limit may have adjusted it.
7514 */
7515 if (!nr_boost_reclaim && balanced)
7516 goto out;
7517
7518 /* Limit the priority of boosting to avoid reclaim writeback */
7519 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7520 raise_priority = false;
7521
7522 /*
7523 * Do not writeback or swap pages for boosted reclaim. The
7524 * intent is to relieve pressure not issue sub-optimal IO
7525 * from reclaim context. If no pages are reclaimed, the
7526 * reclaim will be aborted.
7527 */
7528 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7529 sc.may_swap = !nr_boost_reclaim;
7530
7531 /*
7532 * Do some background aging, to give pages a chance to be
7533 * referenced before reclaiming. All pages are rotated
7534 * regardless of classzone as this is about consistent aging.
7535 */
7536 kswapd_age_node(pgdat, &sc);
7537
7538 /*
7539 * If we're getting trouble reclaiming, start doing writepage
7540 * even in laptop mode.
7541 */
7542 if (sc.priority < DEF_PRIORITY - 2)
7543 sc.may_writepage = 1;
7544
7545 /* Call soft limit reclaim before calling shrink_node. */
7546 sc.nr_scanned = 0;
7547 nr_soft_scanned = 0;
7548 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
7549 sc.gfp_mask, &nr_soft_scanned);
7550 sc.nr_reclaimed += nr_soft_reclaimed;
7551
7552 /*
7553 * There should be no need to raise the scanning priority if
7554 * enough pages are already being scanned that that high
7555 * watermark would be met at 100% efficiency.
7556 */
7557 if (kswapd_shrink_node(pgdat, &sc))
7558 raise_priority = false;
7559
7560 /*
7561 * If the low watermark is met there is no need for processes
7562 * to be throttled on pfmemalloc_wait as they should not be
7563 * able to safely make forward progress. Wake them
7564 */
7565 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
7566 allow_direct_reclaim(pgdat))
7567 wake_up_all(&pgdat->pfmemalloc_wait);
7568
7569 /* Check if kswapd should be suspending */
7570 __fs_reclaim_release(_THIS_IP_);
7571 ret = try_to_freeze();
7572 __fs_reclaim_acquire(_THIS_IP_);
7573 if (ret || kthread_should_stop())
7574 break;
7575
7576 /*
7577 * Raise priority if scanning rate is too low or there was no
7578 * progress in reclaiming pages
7579 */
7580 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7581 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7582
7583 /*
7584 * If reclaim made no progress for a boost, stop reclaim as
7585 * IO cannot be queued and it could be an infinite loop in
7586 * extreme circumstances.
7587 */
7588 if (nr_boost_reclaim && !nr_reclaimed)
7589 break;
7590
7591 if (raise_priority || !nr_reclaimed)
7592 sc.priority--;
7593 } while (sc.priority >= 1);
7594
7595 if (!sc.nr_reclaimed)
7596 pgdat->kswapd_failures++;
7597
7598 out:
7599 clear_reclaim_active(pgdat, highest_zoneidx);
7600
7601 /* If reclaim was boosted, account for the reclaim done in this pass */
7602 if (boosted) {
7603 unsigned long flags;
7604
7605 for (i = 0; i <= highest_zoneidx; i++) {
7606 if (!zone_boosts[i])
7607 continue;
7608
7609 /* Increments are under the zone lock */
7610 zone = pgdat->node_zones + i;
7611 spin_lock_irqsave(&zone->lock, flags);
7612 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7613 spin_unlock_irqrestore(&zone->lock, flags);
7614 }
7615
7616 /*
7617 * As there is now likely space, wakeup kcompact to defragment
7618 * pageblocks.
7619 */
7620 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7621 }
7622
7623 snapshot_refaults(NULL, pgdat);
7624 __fs_reclaim_release(_THIS_IP_);
7625 psi_memstall_leave(&pflags);
7626 set_task_reclaim_state(current, NULL);
7627
7628 /*
7629 * Return the order kswapd stopped reclaiming at as
7630 * prepare_kswapd_sleep() takes it into account. If another caller
7631 * entered the allocator slow path while kswapd was awake, order will
7632 * remain at the higher level.
7633 */
7634 return sc.order;
7635 }
7636
7637 /*
7638 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7639 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7640 * not a valid index then either kswapd runs for first time or kswapd couldn't
7641 * sleep after previous reclaim attempt (node is still unbalanced). In that
7642 * case return the zone index of the previous kswapd reclaim cycle.
7643 */
kswapd_highest_zoneidx(pg_data_t * pgdat,enum zone_type prev_highest_zoneidx)7644 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7645 enum zone_type prev_highest_zoneidx)
7646 {
7647 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7648
7649 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7650 }
7651
kswapd_try_to_sleep(pg_data_t * pgdat,int alloc_order,int reclaim_order,unsigned int highest_zoneidx)7652 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7653 unsigned int highest_zoneidx)
7654 {
7655 long remaining = 0;
7656 DEFINE_WAIT(wait);
7657
7658 if (freezing(current) || kthread_should_stop())
7659 return;
7660
7661 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7662
7663 /*
7664 * Try to sleep for a short interval. Note that kcompactd will only be
7665 * woken if it is possible to sleep for a short interval. This is
7666 * deliberate on the assumption that if reclaim cannot keep an
7667 * eligible zone balanced that it's also unlikely that compaction will
7668 * succeed.
7669 */
7670 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7671 /*
7672 * Compaction records what page blocks it recently failed to
7673 * isolate pages from and skips them in the future scanning.
7674 * When kswapd is going to sleep, it is reasonable to assume
7675 * that pages and compaction may succeed so reset the cache.
7676 */
7677 reset_isolation_suitable(pgdat);
7678
7679 /*
7680 * We have freed the memory, now we should compact it to make
7681 * allocation of the requested order possible.
7682 */
7683 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7684
7685 remaining = schedule_timeout(HZ/10);
7686
7687 /*
7688 * If woken prematurely then reset kswapd_highest_zoneidx and
7689 * order. The values will either be from a wakeup request or
7690 * the previous request that slept prematurely.
7691 */
7692 if (remaining) {
7693 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7694 kswapd_highest_zoneidx(pgdat,
7695 highest_zoneidx));
7696
7697 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7698 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7699 }
7700
7701 finish_wait(&pgdat->kswapd_wait, &wait);
7702 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7703 }
7704
7705 /*
7706 * After a short sleep, check if it was a premature sleep. If not, then
7707 * go fully to sleep until explicitly woken up.
7708 */
7709 if (!remaining &&
7710 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7711 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7712
7713 /*
7714 * vmstat counters are not perfectly accurate and the estimated
7715 * value for counters such as NR_FREE_PAGES can deviate from the
7716 * true value by nr_online_cpus * threshold. To avoid the zone
7717 * watermarks being breached while under pressure, we reduce the
7718 * per-cpu vmstat threshold while kswapd is awake and restore
7719 * them before going back to sleep.
7720 */
7721 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7722
7723 if (!kthread_should_stop())
7724 schedule();
7725
7726 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7727 } else {
7728 if (remaining)
7729 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7730 else
7731 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7732 }
7733 finish_wait(&pgdat->kswapd_wait, &wait);
7734 }
7735
7736 /*
7737 * The background pageout daemon, started as a kernel thread
7738 * from the init process.
7739 *
7740 * This basically trickles out pages so that we have _some_
7741 * free memory available even if there is no other activity
7742 * that frees anything up. This is needed for things like routing
7743 * etc, where we otherwise might have all activity going on in
7744 * asynchronous contexts that cannot page things out.
7745 *
7746 * If there are applications that are active memory-allocators
7747 * (most normal use), this basically shouldn't matter.
7748 */
kswapd(void * p)7749 static int kswapd(void *p)
7750 {
7751 unsigned int alloc_order, reclaim_order;
7752 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7753 pg_data_t *pgdat = (pg_data_t *)p;
7754 struct task_struct *tsk = current;
7755 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7756
7757 if (!cpumask_empty(cpumask))
7758 set_cpus_allowed_ptr(tsk, cpumask);
7759
7760 /*
7761 * Tell the memory management that we're a "memory allocator",
7762 * and that if we need more memory we should get access to it
7763 * regardless (see "__alloc_pages()"). "kswapd" should
7764 * never get caught in the normal page freeing logic.
7765 *
7766 * (Kswapd normally doesn't need memory anyway, but sometimes
7767 * you need a small amount of memory in order to be able to
7768 * page out something else, and this flag essentially protects
7769 * us from recursively trying to free more memory as we're
7770 * trying to free the first piece of memory in the first place).
7771 */
7772 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7773 set_freezable();
7774
7775 WRITE_ONCE(pgdat->kswapd_order, 0);
7776 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7777 atomic_set(&pgdat->nr_writeback_throttled, 0);
7778 for ( ; ; ) {
7779 bool ret;
7780
7781 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7782 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7783 highest_zoneidx);
7784
7785 kswapd_try_sleep:
7786 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7787 highest_zoneidx);
7788
7789 /* Read the new order and highest_zoneidx */
7790 alloc_order = READ_ONCE(pgdat->kswapd_order);
7791 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7792 highest_zoneidx);
7793 WRITE_ONCE(pgdat->kswapd_order, 0);
7794 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7795
7796 ret = try_to_freeze();
7797 if (kthread_should_stop())
7798 break;
7799
7800 /*
7801 * We can speed up thawing tasks if we don't call balance_pgdat
7802 * after returning from the refrigerator
7803 */
7804 if (ret)
7805 continue;
7806
7807 /*
7808 * Reclaim begins at the requested order but if a high-order
7809 * reclaim fails then kswapd falls back to reclaiming for
7810 * order-0. If that happens, kswapd will consider sleeping
7811 * for the order it finished reclaiming at (reclaim_order)
7812 * but kcompactd is woken to compact for the original
7813 * request (alloc_order).
7814 */
7815 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7816 alloc_order);
7817 reclaim_order = balance_pgdat(pgdat, alloc_order,
7818 highest_zoneidx);
7819 if (reclaim_order < alloc_order)
7820 goto kswapd_try_sleep;
7821 }
7822
7823 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7824
7825 return 0;
7826 }
7827
7828 /*
7829 * A zone is low on free memory or too fragmented for high-order memory. If
7830 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7831 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7832 * has failed or is not needed, still wake up kcompactd if only compaction is
7833 * needed.
7834 */
wakeup_kswapd(struct zone * zone,gfp_t gfp_flags,int order,enum zone_type highest_zoneidx)7835 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7836 enum zone_type highest_zoneidx)
7837 {
7838 pg_data_t *pgdat;
7839 enum zone_type curr_idx;
7840
7841 if (!managed_zone(zone))
7842 return;
7843
7844 if (!cpuset_zone_allowed(zone, gfp_flags))
7845 return;
7846
7847 pgdat = zone->zone_pgdat;
7848 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7849
7850 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7851 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7852
7853 if (READ_ONCE(pgdat->kswapd_order) < order)
7854 WRITE_ONCE(pgdat->kswapd_order, order);
7855
7856 if (!waitqueue_active(&pgdat->kswapd_wait))
7857 return;
7858
7859 /* Hopeless node, leave it to direct reclaim if possible */
7860 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7861 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7862 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7863 /*
7864 * There may be plenty of free memory available, but it's too
7865 * fragmented for high-order allocations. Wake up kcompactd
7866 * and rely on compaction_suitable() to determine if it's
7867 * needed. If it fails, it will defer subsequent attempts to
7868 * ratelimit its work.
7869 */
7870 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7871 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7872 return;
7873 }
7874
7875 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7876 gfp_flags);
7877 wake_up_interruptible(&pgdat->kswapd_wait);
7878 }
7879
7880 #ifdef CONFIG_HIBERNATION
7881 /*
7882 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7883 * freed pages.
7884 *
7885 * Rather than trying to age LRUs the aim is to preserve the overall
7886 * LRU order by reclaiming preferentially
7887 * inactive > active > active referenced > active mapped
7888 */
shrink_all_memory(unsigned long nr_to_reclaim)7889 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7890 {
7891 struct scan_control sc = {
7892 .nr_to_reclaim = nr_to_reclaim,
7893 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7894 .reclaim_idx = MAX_NR_ZONES - 1,
7895 .priority = DEF_PRIORITY,
7896 .may_writepage = 1,
7897 .may_unmap = 1,
7898 .may_swap = 1,
7899 .hibernation_mode = 1,
7900 };
7901 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7902 unsigned long nr_reclaimed;
7903 unsigned int noreclaim_flag;
7904
7905 fs_reclaim_acquire(sc.gfp_mask);
7906 noreclaim_flag = memalloc_noreclaim_save();
7907 set_task_reclaim_state(current, &sc.reclaim_state);
7908
7909 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7910
7911 set_task_reclaim_state(current, NULL);
7912 memalloc_noreclaim_restore(noreclaim_flag);
7913 fs_reclaim_release(sc.gfp_mask);
7914
7915 return nr_reclaimed;
7916 }
7917 #endif /* CONFIG_HIBERNATION */
7918
7919 /*
7920 * This kswapd start function will be called by init and node-hot-add.
7921 */
kswapd_run(int nid)7922 void __meminit kswapd_run(int nid)
7923 {
7924 pg_data_t *pgdat = NODE_DATA(nid);
7925
7926 pgdat_kswapd_lock(pgdat);
7927 if (!pgdat->kswapd) {
7928 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7929 if (IS_ERR(pgdat->kswapd)) {
7930 /* failure at boot is fatal */
7931 BUG_ON(system_state < SYSTEM_RUNNING);
7932 pr_err("Failed to start kswapd on node %d\n", nid);
7933 pgdat->kswapd = NULL;
7934 }
7935 }
7936 pgdat_kswapd_unlock(pgdat);
7937 }
7938
7939 /*
7940 * Called by memory hotplug when all memory in a node is offlined. Caller must
7941 * be holding mem_hotplug_begin/done().
7942 */
kswapd_stop(int nid)7943 void __meminit kswapd_stop(int nid)
7944 {
7945 pg_data_t *pgdat = NODE_DATA(nid);
7946 struct task_struct *kswapd;
7947
7948 pgdat_kswapd_lock(pgdat);
7949 kswapd = pgdat->kswapd;
7950 if (kswapd) {
7951 kthread_stop(kswapd);
7952 pgdat->kswapd = NULL;
7953 }
7954 pgdat_kswapd_unlock(pgdat);
7955 }
7956
kswapd_init(void)7957 static int __init kswapd_init(void)
7958 {
7959 int nid;
7960
7961 swap_setup();
7962 for_each_node_state(nid, N_MEMORY)
7963 kswapd_run(nid);
7964 return 0;
7965 }
7966
7967 module_init(kswapd_init)
7968
7969 #ifdef CONFIG_NUMA
7970 /*
7971 * Node reclaim mode
7972 *
7973 * If non-zero call node_reclaim when the number of free pages falls below
7974 * the watermarks.
7975 */
7976 int node_reclaim_mode __read_mostly;
7977
7978 /*
7979 * Priority for NODE_RECLAIM. This determines the fraction of pages
7980 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7981 * a zone.
7982 */
7983 #define NODE_RECLAIM_PRIORITY 4
7984
7985 /*
7986 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7987 * occur.
7988 */
7989 int sysctl_min_unmapped_ratio = 1;
7990
7991 /*
7992 * If the number of slab pages in a zone grows beyond this percentage then
7993 * slab reclaim needs to occur.
7994 */
7995 int sysctl_min_slab_ratio = 5;
7996
node_unmapped_file_pages(struct pglist_data * pgdat)7997 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7998 {
7999 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
8000 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
8001 node_page_state(pgdat, NR_ACTIVE_FILE);
8002
8003 /*
8004 * It's possible for there to be more file mapped pages than
8005 * accounted for by the pages on the file LRU lists because
8006 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
8007 */
8008 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
8009 }
8010
8011 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
node_pagecache_reclaimable(struct pglist_data * pgdat)8012 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
8013 {
8014 unsigned long nr_pagecache_reclaimable;
8015 unsigned long delta = 0;
8016
8017 /*
8018 * If RECLAIM_UNMAP is set, then all file pages are considered
8019 * potentially reclaimable. Otherwise, we have to worry about
8020 * pages like swapcache and node_unmapped_file_pages() provides
8021 * a better estimate
8022 */
8023 if (node_reclaim_mode & RECLAIM_UNMAP)
8024 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
8025 else
8026 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
8027
8028 /* If we can't clean pages, remove dirty pages from consideration */
8029 if (!(node_reclaim_mode & RECLAIM_WRITE))
8030 delta += node_page_state(pgdat, NR_FILE_DIRTY);
8031
8032 /* Watch for any possible underflows due to delta */
8033 if (unlikely(delta > nr_pagecache_reclaimable))
8034 delta = nr_pagecache_reclaimable;
8035
8036 return nr_pagecache_reclaimable - delta;
8037 }
8038
8039 /*
8040 * Try to free up some pages from this node through reclaim.
8041 */
__node_reclaim(struct pglist_data * pgdat,gfp_t gfp_mask,unsigned int order)8042 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
8043 {
8044 /* Minimum pages needed in order to stay on node */
8045 const unsigned long nr_pages = 1 << order;
8046 struct task_struct *p = current;
8047 unsigned int noreclaim_flag;
8048 struct scan_control sc = {
8049 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
8050 .gfp_mask = current_gfp_context(gfp_mask),
8051 .order = order,
8052 .priority = NODE_RECLAIM_PRIORITY,
8053 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
8054 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
8055 .may_swap = 1,
8056 .reclaim_idx = gfp_zone(gfp_mask),
8057 };
8058 unsigned long pflags;
8059
8060 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
8061 sc.gfp_mask);
8062
8063 cond_resched();
8064 psi_memstall_enter(&pflags);
8065 fs_reclaim_acquire(sc.gfp_mask);
8066 /*
8067 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
8068 */
8069 noreclaim_flag = memalloc_noreclaim_save();
8070 set_task_reclaim_state(p, &sc.reclaim_state);
8071
8072 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
8073 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
8074 /*
8075 * Free memory by calling shrink node with increasing
8076 * priorities until we have enough memory freed.
8077 */
8078 do {
8079 shrink_node(pgdat, &sc);
8080 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
8081 }
8082
8083 set_task_reclaim_state(p, NULL);
8084 memalloc_noreclaim_restore(noreclaim_flag);
8085 fs_reclaim_release(sc.gfp_mask);
8086 psi_memstall_leave(&pflags);
8087
8088 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
8089
8090 return sc.nr_reclaimed >= nr_pages;
8091 }
8092
node_reclaim(struct pglist_data * pgdat,gfp_t gfp_mask,unsigned int order)8093 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
8094 {
8095 int ret;
8096
8097 /*
8098 * Node reclaim reclaims unmapped file backed pages and
8099 * slab pages if we are over the defined limits.
8100 *
8101 * A small portion of unmapped file backed pages is needed for
8102 * file I/O otherwise pages read by file I/O will be immediately
8103 * thrown out if the node is overallocated. So we do not reclaim
8104 * if less than a specified percentage of the node is used by
8105 * unmapped file backed pages.
8106 */
8107 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
8108 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
8109 pgdat->min_slab_pages)
8110 return NODE_RECLAIM_FULL;
8111
8112 /*
8113 * Do not scan if the allocation should not be delayed.
8114 */
8115 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
8116 return NODE_RECLAIM_NOSCAN;
8117
8118 /*
8119 * Only run node reclaim on the local node or on nodes that do not
8120 * have associated processors. This will favor the local processor
8121 * over remote processors and spread off node memory allocations
8122 * as wide as possible.
8123 */
8124 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
8125 return NODE_RECLAIM_NOSCAN;
8126
8127 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
8128 return NODE_RECLAIM_NOSCAN;
8129
8130 ret = __node_reclaim(pgdat, gfp_mask, order);
8131 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
8132
8133 if (!ret)
8134 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
8135
8136 return ret;
8137 }
8138 #endif
8139
8140 /**
8141 * check_move_unevictable_folios - Move evictable folios to appropriate zone
8142 * lru list
8143 * @fbatch: Batch of lru folios to check.
8144 *
8145 * Checks folios for evictability, if an evictable folio is in the unevictable
8146 * lru list, moves it to the appropriate evictable lru list. This function
8147 * should be only used for lru folios.
8148 */
check_move_unevictable_folios(struct folio_batch * fbatch)8149 void check_move_unevictable_folios(struct folio_batch *fbatch)
8150 {
8151 struct lruvec *lruvec = NULL;
8152 int pgscanned = 0;
8153 int pgrescued = 0;
8154 int i;
8155
8156 for (i = 0; i < fbatch->nr; i++) {
8157 struct folio *folio = fbatch->folios[i];
8158 int nr_pages = folio_nr_pages(folio);
8159
8160 pgscanned += nr_pages;
8161
8162 /* block memcg migration while the folio moves between lrus */
8163 if (!folio_test_clear_lru(folio))
8164 continue;
8165
8166 lruvec = folio_lruvec_relock_irq(folio, lruvec);
8167 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
8168 lruvec_del_folio(lruvec, folio);
8169 folio_clear_unevictable(folio);
8170 lruvec_add_folio(lruvec, folio);
8171 pgrescued += nr_pages;
8172 }
8173 folio_set_lru(folio);
8174 }
8175
8176 if (lruvec) {
8177 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
8178 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
8179 unlock_page_lruvec_irq(lruvec);
8180 } else if (pgscanned) {
8181 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
8182 }
8183 }
8184 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
8185