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/delay.h>
39#include <linux/kthread.h>
40#include <linux/freezer.h>
41#include <linux/memcontrol.h>
42#include <linux/migrate.h>
43#include <linux/delayacct.h>
44#include <linux/sysctl.h>
45#include <linux/memory-tiers.h>
46#include <linux/oom.h>
47#include <linux/pagevec.h>
48#include <linux/prefetch.h>
49#include <linux/printk.h>
50#include <linux/dax.h>
51#include <linux/psi.h>
52#include <linux/pagewalk.h>
53#include <linux/shmem_fs.h>
54#include <linux/ctype.h>
55#include <linux/debugfs.h>
56#include <linux/khugepaged.h>
57#include <linux/rculist_nulls.h>
58#include <linux/random.h>
59#include <linux/mmu_notifier.h>
60#include <linux/parser.h>
61
62#include <asm/tlbflush.h>
63#include <asm/div64.h>
64
65#include <linux/swapops.h>
66#include <linux/balloon_compaction.h>
67#include <linux/sched/sysctl.h>
68
69#include "internal.h"
70#include "swap.h"
71
72#define CREATE_TRACE_POINTS
73#include <trace/events/vmscan.h>
74
75struct scan_control {
76 /* How many pages shrink_list() should reclaim */
77 unsigned long nr_to_reclaim;
78
79 /*
80 * Nodemask of nodes allowed by the caller. If NULL, all nodes
81 * are scanned.
82 */
83 nodemask_t *nodemask;
84
85 /*
86 * The memory cgroup that hit its limit and as a result is the
87 * primary target of this reclaim invocation.
88 */
89 struct mem_cgroup *target_mem_cgroup;
90
91 /*
92 * Scan pressure balancing between anon and file LRUs
93 */
94 unsigned long anon_cost;
95 unsigned long file_cost;
96
97 /* Swappiness value for proactive reclaim. Always use sc_swappiness()! */
98 int *proactive_swappiness;
99
100 /* Can active folios be deactivated as part of reclaim? */
101#define DEACTIVATE_ANON 1
102#define DEACTIVATE_FILE 2
103 unsigned int may_deactivate:2;
104 unsigned int force_deactivate:1;
105 unsigned int skipped_deactivate:1;
106
107 /* Writepage batching in laptop mode; RECLAIM_WRITE */
108 unsigned int may_writepage:1;
109
110 /* Can mapped folios be reclaimed? */
111 unsigned int may_unmap:1;
112
113 /* Can folios be swapped as part of reclaim? */
114 unsigned int may_swap:1;
115
116 /* Not allow cache_trim_mode to be turned on as part of reclaim? */
117 unsigned int no_cache_trim_mode:1;
118
119 /* Has cache_trim_mode failed at least once? */
120 unsigned int cache_trim_mode_failed:1;
121
122 /* Proactive reclaim invoked by userspace */
123 unsigned int proactive:1;
124
125 /*
126 * Cgroup memory below memory.low is protected as long as we
127 * don't threaten to OOM. If any cgroup is reclaimed at
128 * reduced force or passed over entirely due to its memory.low
129 * setting (memcg_low_skipped), and nothing is reclaimed as a
130 * result, then go back for one more cycle that reclaims the protected
131 * memory (memcg_low_reclaim) to avert OOM.
132 */
133 unsigned int memcg_low_reclaim:1;
134 unsigned int memcg_low_skipped:1;
135
136 /* Shared cgroup tree walk failed, rescan the whole tree */
137 unsigned int memcg_full_walk:1;
138
139 unsigned int hibernation_mode:1;
140
141 /* One of the zones is ready for compaction */
142 unsigned int compaction_ready:1;
143
144 /* There is easily reclaimable cold cache in the current node */
145 unsigned int cache_trim_mode:1;
146
147 /* The file folios on the current node are dangerously low */
148 unsigned int file_is_tiny:1;
149
150 /* Always discard instead of demoting to lower tier memory */
151 unsigned int no_demotion:1;
152
153 /* Allocation order */
154 s8 order;
155
156 /* Scan (total_size >> priority) pages at once */
157 s8 priority;
158
159 /* The highest zone to isolate folios for reclaim from */
160 s8 reclaim_idx;
161
162 /* This context's GFP mask */
163 gfp_t gfp_mask;
164
165 /* Incremented by the number of inactive pages that were scanned */
166 unsigned long nr_scanned;
167
168 /* Number of pages freed so far during a call to shrink_zones() */
169 unsigned long nr_reclaimed;
170
171 struct {
172 unsigned int dirty;
173 unsigned int unqueued_dirty;
174 unsigned int congested;
175 unsigned int writeback;
176 unsigned int immediate;
177 unsigned int file_taken;
178 unsigned int taken;
179 } nr;
180
181 /* for recording the reclaimed slab by now */
182 struct reclaim_state reclaim_state;
183};
184
185#ifdef ARCH_HAS_PREFETCHW
186#define prefetchw_prev_lru_folio(_folio, _base, _field) \
187 do { \
188 if ((_folio)->lru.prev != _base) { \
189 struct folio *prev; \
190 \
191 prev = lru_to_folio(&(_folio->lru)); \
192 prefetchw(&prev->_field); \
193 } \
194 } while (0)
195#else
196#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
197#endif
198
199/*
200 * From 0 .. MAX_SWAPPINESS. Higher means more swappy.
201 */
202int vm_swappiness = 60;
203
204#ifdef CONFIG_MEMCG
205
206/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
207static bool cgroup_reclaim(struct scan_control *sc)
208{
209 return sc->target_mem_cgroup;
210}
211
212/*
213 * Returns true for reclaim on the root cgroup. This is true for direct
214 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
215 */
216static bool root_reclaim(struct scan_control *sc)
217{
218 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
219}
220
221/**
222 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
223 * @sc: scan_control in question
224 *
225 * The normal page dirty throttling mechanism in balance_dirty_pages() is
226 * completely broken with the legacy memcg and direct stalling in
227 * shrink_folio_list() is used for throttling instead, which lacks all the
228 * niceties such as fairness, adaptive pausing, bandwidth proportional
229 * allocation and configurability.
230 *
231 * This function tests whether the vmscan currently in progress can assume
232 * that the normal dirty throttling mechanism is operational.
233 */
234static bool writeback_throttling_sane(struct scan_control *sc)
235{
236 if (!cgroup_reclaim(sc))
237 return true;
238#ifdef CONFIG_CGROUP_WRITEBACK
239 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
240 return true;
241#endif
242 return false;
243}
244
245static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
246{
247 if (sc->proactive && sc->proactive_swappiness)
248 return *sc->proactive_swappiness;
249 return mem_cgroup_swappiness(memcg);
250}
251#else
252static bool cgroup_reclaim(struct scan_control *sc)
253{
254 return false;
255}
256
257static bool root_reclaim(struct scan_control *sc)
258{
259 return true;
260}
261
262static bool writeback_throttling_sane(struct scan_control *sc)
263{
264 return true;
265}
266
267static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
268{
269 return READ_ONCE(vm_swappiness);
270}
271#endif
272
273/* for_each_managed_zone_pgdat - helper macro to iterate over all managed zones in a pgdat up to
274 * and including the specified highidx
275 * @zone: The current zone in the iterator
276 * @pgdat: The pgdat which node_zones are being iterated
277 * @idx: The index variable
278 * @highidx: The index of the highest zone to return
279 *
280 * This macro iterates through all managed zones up to and including the specified highidx.
281 * The zone iterator enters an invalid state after macro call and must be reinitialized
282 * before it can be used again.
283 */
284#define for_each_managed_zone_pgdat(zone, pgdat, idx, highidx) \
285 for ((idx) = 0, (zone) = (pgdat)->node_zones; \
286 (idx) <= (highidx); \
287 (idx)++, (zone)++) \
288 if (!managed_zone(zone)) \
289 continue; \
290 else
291
292static void set_task_reclaim_state(struct task_struct *task,
293 struct reclaim_state *rs)
294{
295 /* Check for an overwrite */
296 WARN_ON_ONCE(rs && task->reclaim_state);
297
298 /* Check for the nulling of an already-nulled member */
299 WARN_ON_ONCE(!rs && !task->reclaim_state);
300
301 task->reclaim_state = rs;
302}
303
304/*
305 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
306 * scan_control->nr_reclaimed.
307 */
308static void flush_reclaim_state(struct scan_control *sc)
309{
310 /*
311 * Currently, reclaim_state->reclaimed includes three types of pages
312 * freed outside of vmscan:
313 * (1) Slab pages.
314 * (2) Clean file pages from pruned inodes (on highmem systems).
315 * (3) XFS freed buffer pages.
316 *
317 * For all of these cases, we cannot universally link the pages to a
318 * single memcg. For example, a memcg-aware shrinker can free one object
319 * charged to the target memcg, causing an entire page to be freed.
320 * If we count the entire page as reclaimed from the memcg, we end up
321 * overestimating the reclaimed amount (potentially under-reclaiming).
322 *
323 * Only count such pages for global reclaim to prevent under-reclaiming
324 * from the target memcg; preventing unnecessary retries during memcg
325 * charging and false positives from proactive reclaim.
326 *
327 * For uncommon cases where the freed pages were actually mostly
328 * charged to the target memcg, we end up underestimating the reclaimed
329 * amount. This should be fine. The freed pages will be uncharged
330 * anyway, even if they are not counted here properly, and we will be
331 * able to make forward progress in charging (which is usually in a
332 * retry loop).
333 *
334 * We can go one step further, and report the uncharged objcg pages in
335 * memcg reclaim, to make reporting more accurate and reduce
336 * underestimation, but it's probably not worth the complexity for now.
337 */
338 if (current->reclaim_state && root_reclaim(sc)) {
339 sc->nr_reclaimed += current->reclaim_state->reclaimed;
340 current->reclaim_state->reclaimed = 0;
341 }
342}
343
344static bool can_demote(int nid, struct scan_control *sc,
345 struct mem_cgroup *memcg)
346{
347 int demotion_nid;
348
349 if (!numa_demotion_enabled)
350 return false;
351 if (sc && sc->no_demotion)
352 return false;
353
354 demotion_nid = next_demotion_node(node: nid);
355 if (demotion_nid == NUMA_NO_NODE)
356 return false;
357
358 /* If demotion node isn't in the cgroup's mems_allowed, fall back */
359 return mem_cgroup_node_allowed(memcg, nid: demotion_nid);
360}
361
362static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
363 int nid,
364 struct scan_control *sc)
365{
366 if (memcg == NULL) {
367 /*
368 * For non-memcg reclaim, is there
369 * space in any swap device?
370 */
371 if (get_nr_swap_pages() > 0)
372 return true;
373 } else {
374 /* Is the memcg below its swap limit? */
375 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
376 return true;
377 }
378
379 /*
380 * The page can not be swapped.
381 *
382 * Can it be reclaimed from this node via demotion?
383 */
384 return can_demote(nid, sc, memcg);
385}
386
387/*
388 * This misses isolated folios which are not accounted for to save counters.
389 * As the data only determines if reclaim or compaction continues, it is
390 * not expected that isolated folios will be a dominating factor.
391 */
392unsigned long zone_reclaimable_pages(struct zone *zone)
393{
394 unsigned long nr;
395
396 nr = zone_page_state_snapshot(zone, item: NR_ZONE_INACTIVE_FILE) +
397 zone_page_state_snapshot(zone, item: NR_ZONE_ACTIVE_FILE);
398 if (can_reclaim_anon_pages(NULL, nid: zone_to_nid(zone), NULL))
399 nr += zone_page_state_snapshot(zone, item: NR_ZONE_INACTIVE_ANON) +
400 zone_page_state_snapshot(zone, item: NR_ZONE_ACTIVE_ANON);
401
402 return nr;
403}
404
405/**
406 * lruvec_lru_size - Returns the number of pages on the given LRU list.
407 * @lruvec: lru vector
408 * @lru: lru to use
409 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
410 */
411static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
412 int zone_idx)
413{
414 unsigned long size = 0;
415 int zid;
416 struct zone *zone;
417
418 for_each_managed_zone_pgdat(zone, lruvec_pgdat(lruvec), zid, zone_idx) {
419 if (!mem_cgroup_disabled())
420 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zone_idx: zid);
421 else
422 size += zone_page_state(zone, item: NR_ZONE_LRU_BASE + lru);
423 }
424 return size;
425}
426
427static unsigned long drop_slab_node(int nid)
428{
429 unsigned long freed = 0;
430 struct mem_cgroup *memcg = NULL;
431
432 memcg = mem_cgroup_iter(NULL, NULL, NULL);
433 do {
434 freed += shrink_slab(GFP_KERNEL, nid, memcg, priority: 0);
435 } while ((memcg = mem_cgroup_iter(NULL, prev: memcg, NULL)) != NULL);
436
437 return freed;
438}
439
440void drop_slab(void)
441{
442 int nid;
443 int shift = 0;
444 unsigned long freed;
445
446 do {
447 freed = 0;
448 for_each_online_node(nid) {
449 if (fatal_signal_pending(current))
450 return;
451
452 freed += drop_slab_node(nid);
453 }
454 } while ((freed >> shift++) > 1);
455}
456
457#define CHECK_RECLAIMER_OFFSET(type) \
458 do { \
459 BUILD_BUG_ON(PGSTEAL_##type - PGSTEAL_KSWAPD != \
460 PGDEMOTE_##type - PGDEMOTE_KSWAPD); \
461 BUILD_BUG_ON(PGSTEAL_##type - PGSTEAL_KSWAPD != \
462 PGSCAN_##type - PGSCAN_KSWAPD); \
463 } while (0)
464
465static int reclaimer_offset(struct scan_control *sc)
466{
467 CHECK_RECLAIMER_OFFSET(DIRECT);
468 CHECK_RECLAIMER_OFFSET(KHUGEPAGED);
469 CHECK_RECLAIMER_OFFSET(PROACTIVE);
470
471 if (current_is_kswapd())
472 return 0;
473 if (current_is_khugepaged())
474 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
475 if (sc->proactive)
476 return PGSTEAL_PROACTIVE - PGSTEAL_KSWAPD;
477 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
478}
479
480static inline int is_page_cache_freeable(struct folio *folio)
481{
482 /*
483 * A freeable page cache folio is referenced only by the caller
484 * that isolated the folio, the page cache and optional filesystem
485 * private data at folio->private.
486 */
487 return folio_ref_count(folio) - folio_test_private(folio) ==
488 1 + folio_nr_pages(folio);
489}
490
491/*
492 * We detected a synchronous write error writing a folio out. Probably
493 * -ENOSPC. We need to propagate that into the address_space for a subsequent
494 * fsync(), msync() or close().
495 *
496 * The tricky part is that after writepage we cannot touch the mapping: nothing
497 * prevents it from being freed up. But we have a ref on the folio and once
498 * that folio is locked, the mapping is pinned.
499 *
500 * We're allowed to run sleeping folio_lock() here because we know the caller has
501 * __GFP_FS.
502 */
503static void handle_write_error(struct address_space *mapping,
504 struct folio *folio, int error)
505{
506 folio_lock(folio);
507 if (folio_mapping(folio) == mapping)
508 mapping_set_error(mapping, error);
509 folio_unlock(folio);
510}
511
512static bool skip_throttle_noprogress(pg_data_t *pgdat)
513{
514 int reclaimable = 0, write_pending = 0;
515 int i;
516 struct zone *zone;
517 /*
518 * If kswapd is disabled, reschedule if necessary but do not
519 * throttle as the system is likely near OOM.
520 */
521 if (atomic_read(v: &pgdat->kswapd_failures) >= MAX_RECLAIM_RETRIES)
522 return true;
523
524 /*
525 * If there are a lot of dirty/writeback folios then do not
526 * throttle as throttling will occur when the folios cycle
527 * towards the end of the LRU if still under writeback.
528 */
529 for_each_managed_zone_pgdat(zone, pgdat, i, MAX_NR_ZONES - 1) {
530 reclaimable += zone_reclaimable_pages(zone);
531 write_pending += zone_page_state_snapshot(zone,
532 item: NR_ZONE_WRITE_PENDING);
533 }
534 if (2 * write_pending <= reclaimable)
535 return true;
536
537 return false;
538}
539
540void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
541{
542 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
543 long timeout, ret;
544 DEFINE_WAIT(wait);
545
546 /*
547 * Do not throttle user workers, kthreads other than kswapd or
548 * workqueues. They may be required for reclaim to make
549 * forward progress (e.g. journalling workqueues or kthreads).
550 */
551 if (!current_is_kswapd() &&
552 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
553 cond_resched();
554 return;
555 }
556
557 /*
558 * These figures are pulled out of thin air.
559 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
560 * parallel reclaimers which is a short-lived event so the timeout is
561 * short. Failing to make progress or waiting on writeback are
562 * potentially long-lived events so use a longer timeout. This is shaky
563 * logic as a failure to make progress could be due to anything from
564 * writeback to a slow device to excessive referenced folios at the tail
565 * of the inactive LRU.
566 */
567 switch(reason) {
568 case VMSCAN_THROTTLE_WRITEBACK:
569 timeout = HZ/10;
570
571 if (atomic_inc_return(v: &pgdat->nr_writeback_throttled) == 1) {
572 WRITE_ONCE(pgdat->nr_reclaim_start,
573 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
574 }
575
576 break;
577 case VMSCAN_THROTTLE_CONGESTED:
578 fallthrough;
579 case VMSCAN_THROTTLE_NOPROGRESS:
580 if (skip_throttle_noprogress(pgdat)) {
581 cond_resched();
582 return;
583 }
584
585 timeout = 1;
586
587 break;
588 case VMSCAN_THROTTLE_ISOLATED:
589 timeout = HZ/50;
590 break;
591 default:
592 WARN_ON_ONCE(1);
593 timeout = HZ;
594 break;
595 }
596
597 prepare_to_wait(wq_head: wqh, wq_entry: &wait, TASK_UNINTERRUPTIBLE);
598 ret = schedule_timeout(timeout);
599 finish_wait(wq_head: wqh, wq_entry: &wait);
600
601 if (reason == VMSCAN_THROTTLE_WRITEBACK)
602 atomic_dec(v: &pgdat->nr_writeback_throttled);
603
604 trace_mm_vmscan_throttled(nid: pgdat->node_id, usec_timeout: jiffies_to_usecs(j: timeout),
605 usec_delayed: jiffies_to_usecs(j: timeout - ret),
606 reason);
607}
608
609/*
610 * Account for folios written if tasks are throttled waiting on dirty
611 * folios to clean. If enough folios have been cleaned since throttling
612 * started then wakeup the throttled tasks.
613 */
614void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
615 int nr_throttled)
616{
617 unsigned long nr_written;
618
619 node_stat_add_folio(folio, item: NR_THROTTLED_WRITTEN);
620
621 /*
622 * This is an inaccurate read as the per-cpu deltas may not
623 * be synchronised. However, given that the system is
624 * writeback throttled, it is not worth taking the penalty
625 * of getting an accurate count. At worst, the throttle
626 * timeout guarantees forward progress.
627 */
628 nr_written = node_page_state(pgdat, item: NR_THROTTLED_WRITTEN) -
629 READ_ONCE(pgdat->nr_reclaim_start);
630
631 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
632 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
633}
634
635/* possible outcome of pageout() */
636typedef enum {
637 /* failed to write folio out, folio is locked */
638 PAGE_KEEP,
639 /* move folio to the active list, folio is locked */
640 PAGE_ACTIVATE,
641 /* folio has been sent to the disk successfully, folio is unlocked */
642 PAGE_SUCCESS,
643 /* folio is clean and locked */
644 PAGE_CLEAN,
645} pageout_t;
646
647static pageout_t writeout(struct folio *folio, struct address_space *mapping,
648 struct swap_iocb **plug, struct list_head *folio_list)
649{
650 int res;
651
652 folio_set_reclaim(folio);
653
654 /*
655 * The large shmem folio can be split if CONFIG_THP_SWAP is not enabled
656 * or we failed to allocate contiguous swap entries, in which case
657 * the split out folios get added back to folio_list.
658 */
659 if (shmem_mapping(mapping))
660 res = shmem_writeout(folio, plug, folio_list);
661 else
662 res = swap_writeout(folio, swap_plug: plug);
663
664 if (res < 0)
665 handle_write_error(mapping, folio, error: res);
666 if (res == AOP_WRITEPAGE_ACTIVATE) {
667 folio_clear_reclaim(folio);
668 return PAGE_ACTIVATE;
669 }
670
671 /* synchronous write? */
672 if (!folio_test_writeback(folio))
673 folio_clear_reclaim(folio);
674
675 trace_mm_vmscan_write_folio(folio);
676 node_stat_add_folio(folio, item: NR_VMSCAN_WRITE);
677 return PAGE_SUCCESS;
678}
679
680/*
681 * pageout is called by shrink_folio_list() for each dirty folio.
682 */
683static pageout_t pageout(struct folio *folio, struct address_space *mapping,
684 struct swap_iocb **plug, struct list_head *folio_list)
685{
686 /*
687 * We no longer attempt to writeback filesystem folios here, other
688 * than tmpfs/shmem. That's taken care of in page-writeback.
689 * If we find a dirty filesystem folio at the end of the LRU list,
690 * typically that means the filesystem is saturating the storage
691 * with contiguous writes and telling it to write a folio here
692 * would only make the situation worse by injecting an element
693 * of random access.
694 *
695 * If the folio is swapcache, write it back even if that would
696 * block, for some throttling. This happens by accident, because
697 * swap_backing_dev_info is bust: it doesn't reflect the
698 * congestion state of the swapdevs. Easy to fix, if needed.
699 */
700 if (!is_page_cache_freeable(folio))
701 return PAGE_KEEP;
702 if (!mapping) {
703 /*
704 * Some data journaling orphaned folios can have
705 * folio->mapping == NULL while being dirty with clean buffers.
706 */
707 if (folio_test_private(folio)) {
708 if (try_to_free_buffers(folio)) {
709 folio_clear_dirty(folio);
710 pr_info("%s: orphaned folio\n", __func__);
711 return PAGE_CLEAN;
712 }
713 }
714 return PAGE_KEEP;
715 }
716
717 if (!shmem_mapping(mapping) && !folio_test_anon(folio))
718 return PAGE_ACTIVATE;
719 if (!folio_clear_dirty_for_io(folio))
720 return PAGE_CLEAN;
721 return writeout(folio, mapping, plug, folio_list);
722}
723
724/*
725 * Same as remove_mapping, but if the folio is removed from the mapping, it
726 * gets returned with a refcount of 0.
727 */
728static int __remove_mapping(struct address_space *mapping, struct folio *folio,
729 bool reclaimed, struct mem_cgroup *target_memcg)
730{
731 int refcount;
732 void *shadow = NULL;
733 struct swap_cluster_info *ci;
734
735 BUG_ON(!folio_test_locked(folio));
736 BUG_ON(mapping != folio_mapping(folio));
737
738 if (folio_test_swapcache(folio)) {
739 ci = swap_cluster_get_and_lock_irq(folio);
740 } else {
741 spin_lock(lock: &mapping->host->i_lock);
742 xa_lock_irq(&mapping->i_pages);
743 }
744
745 /*
746 * The non racy check for a busy folio.
747 *
748 * Must be careful with the order of the tests. When someone has
749 * a ref to the folio, it may be possible that they dirty it then
750 * drop the reference. So if the dirty flag is tested before the
751 * refcount here, then the following race may occur:
752 *
753 * get_user_pages(&page);
754 * [user mapping goes away]
755 * write_to(page);
756 * !folio_test_dirty(folio) [good]
757 * folio_set_dirty(folio);
758 * folio_put(folio);
759 * !refcount(folio) [good, discard it]
760 *
761 * [oops, our write_to data is lost]
762 *
763 * Reversing the order of the tests ensures such a situation cannot
764 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
765 * load is not satisfied before that of folio->_refcount.
766 *
767 * Note that if the dirty flag is always set via folio_mark_dirty,
768 * and thus under the i_pages lock, then this ordering is not required.
769 */
770 refcount = 1 + folio_nr_pages(folio);
771 if (!folio_ref_freeze(folio, count: refcount))
772 goto cannot_free;
773 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
774 if (unlikely(folio_test_dirty(folio))) {
775 folio_ref_unfreeze(folio, count: refcount);
776 goto cannot_free;
777 }
778
779 if (folio_test_swapcache(folio)) {
780 swp_entry_t swap = folio->swap;
781
782 if (reclaimed && !mapping_exiting(mapping))
783 shadow = workingset_eviction(folio, target_memcg);
784 __swap_cache_del_folio(ci, folio, entry: swap, shadow);
785 memcg1_swapout(folio, entry: swap);
786 swap_cluster_unlock_irq(ci);
787 put_swap_folio(folio, entry: swap);
788 } else {
789 void (*free_folio)(struct folio *);
790
791 free_folio = mapping->a_ops->free_folio;
792 /*
793 * Remember a shadow entry for reclaimed file cache in
794 * order to detect refaults, thus thrashing, later on.
795 *
796 * But don't store shadows in an address space that is
797 * already exiting. This is not just an optimization,
798 * inode reclaim needs to empty out the radix tree or
799 * the nodes are lost. Don't plant shadows behind its
800 * back.
801 *
802 * We also don't store shadows for DAX mappings because the
803 * only page cache folios found in these are zero pages
804 * covering holes, and because we don't want to mix DAX
805 * exceptional entries and shadow exceptional entries in the
806 * same address_space.
807 */
808 if (reclaimed && folio_is_file_lru(folio) &&
809 !mapping_exiting(mapping) && !dax_mapping(mapping))
810 shadow = workingset_eviction(folio, target_memcg);
811 __filemap_remove_folio(folio, shadow);
812 xa_unlock_irq(&mapping->i_pages);
813 if (mapping_shrinkable(mapping))
814 inode_add_lru(inode: mapping->host);
815 spin_unlock(lock: &mapping->host->i_lock);
816
817 if (free_folio)
818 free_folio(folio);
819 }
820
821 return 1;
822
823cannot_free:
824 if (folio_test_swapcache(folio)) {
825 swap_cluster_unlock_irq(ci);
826 } else {
827 xa_unlock_irq(&mapping->i_pages);
828 spin_unlock(lock: &mapping->host->i_lock);
829 }
830 return 0;
831}
832
833/**
834 * remove_mapping() - Attempt to remove a folio from its mapping.
835 * @mapping: The address space.
836 * @folio: The folio to remove.
837 *
838 * If the folio is dirty, under writeback or if someone else has a ref
839 * on it, removal will fail.
840 * Return: The number of pages removed from the mapping. 0 if the folio
841 * could not be removed.
842 * Context: The caller should have a single refcount on the folio and
843 * hold its lock.
844 */
845long remove_mapping(struct address_space *mapping, struct folio *folio)
846{
847 if (__remove_mapping(mapping, folio, reclaimed: false, NULL)) {
848 /*
849 * Unfreezing the refcount with 1 effectively
850 * drops the pagecache ref for us without requiring another
851 * atomic operation.
852 */
853 folio_ref_unfreeze(folio, count: 1);
854 return folio_nr_pages(folio);
855 }
856 return 0;
857}
858
859/**
860 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
861 * @folio: Folio to be returned to an LRU list.
862 *
863 * Add previously isolated @folio to appropriate LRU list.
864 * The folio may still be unevictable for other reasons.
865 *
866 * Context: lru_lock must not be held, interrupts must be enabled.
867 */
868void folio_putback_lru(struct folio *folio)
869{
870 folio_add_lru(folio);
871 folio_put(folio); /* drop ref from isolate */
872}
873
874enum folio_references {
875 FOLIOREF_RECLAIM,
876 FOLIOREF_RECLAIM_CLEAN,
877 FOLIOREF_KEEP,
878 FOLIOREF_ACTIVATE,
879};
880
881#ifdef CONFIG_LRU_GEN
882/*
883 * Only used on a mapped folio in the eviction (rmap walk) path, where promotion
884 * needs to be done by taking the folio off the LRU list and then adding it back
885 * with PG_active set. In contrast, the aging (page table walk) path uses
886 * folio_update_gen().
887 */
888static bool lru_gen_set_refs(struct folio *folio)
889{
890 /* see the comment on LRU_REFS_FLAGS */
891 if (!folio_test_referenced(folio) && !folio_test_workingset(folio)) {
892 set_mask_bits(&folio->flags.f, LRU_REFS_MASK, BIT(PG_referenced));
893 return false;
894 }
895
896 set_mask_bits(&folio->flags.f, LRU_REFS_FLAGS, BIT(PG_workingset));
897 return true;
898}
899#else
900static bool lru_gen_set_refs(struct folio *folio)
901{
902 return false;
903}
904#endif /* CONFIG_LRU_GEN */
905
906static enum folio_references folio_check_references(struct folio *folio,
907 struct scan_control *sc)
908{
909 int referenced_ptes, referenced_folio;
910 vm_flags_t vm_flags;
911
912 referenced_ptes = folio_referenced(folio, is_locked: 1, memcg: sc->target_mem_cgroup,
913 vm_flags: &vm_flags);
914
915 /*
916 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
917 * Let the folio, now marked Mlocked, be moved to the unevictable list.
918 */
919 if (vm_flags & VM_LOCKED)
920 return FOLIOREF_ACTIVATE;
921
922 /*
923 * There are two cases to consider.
924 * 1) Rmap lock contention: rotate.
925 * 2) Skip the non-shared swapbacked folio mapped solely by
926 * the exiting or OOM-reaped process.
927 */
928 if (referenced_ptes == -1)
929 return FOLIOREF_KEEP;
930
931 if (lru_gen_enabled()) {
932 if (!referenced_ptes)
933 return FOLIOREF_RECLAIM;
934
935 return lru_gen_set_refs(folio) ? FOLIOREF_ACTIVATE : FOLIOREF_KEEP;
936 }
937
938 referenced_folio = folio_test_clear_referenced(folio);
939
940 if (referenced_ptes) {
941 /*
942 * All mapped folios start out with page table
943 * references from the instantiating fault, so we need
944 * to look twice if a mapped file/anon folio is used more
945 * than once.
946 *
947 * Mark it and spare it for another trip around the
948 * inactive list. Another page table reference will
949 * lead to its activation.
950 *
951 * Note: the mark is set for activated folios as well
952 * so that recently deactivated but used folios are
953 * quickly recovered.
954 */
955 folio_set_referenced(folio);
956
957 if (referenced_folio || referenced_ptes > 1)
958 return FOLIOREF_ACTIVATE;
959
960 /*
961 * Activate file-backed executable folios after first usage.
962 */
963 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
964 return FOLIOREF_ACTIVATE;
965
966 return FOLIOREF_KEEP;
967 }
968
969 /* Reclaim if clean, defer dirty folios to writeback */
970 if (referenced_folio && folio_is_file_lru(folio))
971 return FOLIOREF_RECLAIM_CLEAN;
972
973 return FOLIOREF_RECLAIM;
974}
975
976/* Check if a folio is dirty or under writeback */
977static void folio_check_dirty_writeback(struct folio *folio,
978 bool *dirty, bool *writeback)
979{
980 struct address_space *mapping;
981
982 /*
983 * Anonymous folios are not handled by flushers and must be written
984 * from reclaim context. Do not stall reclaim based on them.
985 * MADV_FREE anonymous folios are put into inactive file list too.
986 * They could be mistakenly treated as file lru. So further anon
987 * test is needed.
988 */
989 if (!folio_is_file_lru(folio) ||
990 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
991 *dirty = false;
992 *writeback = false;
993 return;
994 }
995
996 /* By default assume that the folio flags are accurate */
997 *dirty = folio_test_dirty(folio);
998 *writeback = folio_test_writeback(folio);
999
1000 /* Verify dirty/writeback state if the filesystem supports it */
1001 if (!folio_test_private(folio))
1002 return;
1003
1004 mapping = folio_mapping(folio);
1005 if (mapping && mapping->a_ops->is_dirty_writeback)
1006 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1007}
1008
1009static struct folio *alloc_demote_folio(struct folio *src,
1010 unsigned long private)
1011{
1012 struct folio *dst;
1013 nodemask_t *allowed_mask;
1014 struct migration_target_control *mtc;
1015
1016 mtc = (struct migration_target_control *)private;
1017
1018 allowed_mask = mtc->nmask;
1019 /*
1020 * make sure we allocate from the target node first also trying to
1021 * demote or reclaim pages from the target node via kswapd if we are
1022 * low on free memory on target node. If we don't do this and if
1023 * we have free memory on the slower(lower) memtier, we would start
1024 * allocating pages from slower(lower) memory tiers without even forcing
1025 * a demotion of cold pages from the target memtier. This can result
1026 * in the kernel placing hot pages in slower(lower) memory tiers.
1027 */
1028 mtc->nmask = NULL;
1029 mtc->gfp_mask |= __GFP_THISNODE;
1030 dst = alloc_migration_target(src, private: (unsigned long)mtc);
1031 if (dst)
1032 return dst;
1033
1034 mtc->gfp_mask &= ~__GFP_THISNODE;
1035 mtc->nmask = allowed_mask;
1036
1037 return alloc_migration_target(src, private: (unsigned long)mtc);
1038}
1039
1040/*
1041 * Take folios on @demote_folios and attempt to demote them to another node.
1042 * Folios which are not demoted are left on @demote_folios.
1043 */
1044static unsigned int demote_folio_list(struct list_head *demote_folios,
1045 struct pglist_data *pgdat)
1046{
1047 int target_nid = next_demotion_node(node: pgdat->node_id);
1048 unsigned int nr_succeeded;
1049 nodemask_t allowed_mask;
1050
1051 struct migration_target_control mtc = {
1052 /*
1053 * Allocate from 'node', or fail quickly and quietly.
1054 * When this happens, 'page' will likely just be discarded
1055 * instead of migrated.
1056 */
1057 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1058 __GFP_NOMEMALLOC | GFP_NOWAIT,
1059 .nid = target_nid,
1060 .nmask = &allowed_mask,
1061 .reason = MR_DEMOTION,
1062 };
1063
1064 if (list_empty(head: demote_folios))
1065 return 0;
1066
1067 if (target_nid == NUMA_NO_NODE)
1068 return 0;
1069
1070 node_get_allowed_targets(pgdat, targets: &allowed_mask);
1071
1072 /* Demotion ignores all cpuset and mempolicy settings */
1073 migrate_pages(l: demote_folios, new: alloc_demote_folio, NULL,
1074 private: (unsigned long)&mtc, mode: MIGRATE_ASYNC, reason: MR_DEMOTION,
1075 ret_succeeded: &nr_succeeded);
1076
1077 return nr_succeeded;
1078}
1079
1080static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1081{
1082 if (gfp_mask & __GFP_FS)
1083 return true;
1084 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1085 return false;
1086 /*
1087 * We can "enter_fs" for swap-cache with only __GFP_IO
1088 * providing this isn't SWP_FS_OPS.
1089 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1090 * but that will never affect SWP_FS_OPS, so the data_race
1091 * is safe.
1092 */
1093 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1094}
1095
1096/*
1097 * shrink_folio_list() returns the number of reclaimed pages
1098 */
1099static unsigned int shrink_folio_list(struct list_head *folio_list,
1100 struct pglist_data *pgdat, struct scan_control *sc,
1101 struct reclaim_stat *stat, bool ignore_references,
1102 struct mem_cgroup *memcg)
1103{
1104 struct folio_batch free_folios;
1105 LIST_HEAD(ret_folios);
1106 LIST_HEAD(demote_folios);
1107 unsigned int nr_reclaimed = 0, nr_demoted = 0;
1108 unsigned int pgactivate = 0;
1109 bool do_demote_pass;
1110 struct swap_iocb *plug = NULL;
1111
1112 folio_batch_init(fbatch: &free_folios);
1113 memset(s: stat, c: 0, n: sizeof(*stat));
1114 cond_resched();
1115 do_demote_pass = can_demote(nid: pgdat->node_id, sc, memcg);
1116
1117retry:
1118 while (!list_empty(head: folio_list)) {
1119 struct address_space *mapping;
1120 struct folio *folio;
1121 enum folio_references references = FOLIOREF_RECLAIM;
1122 bool dirty, writeback;
1123 unsigned int nr_pages;
1124
1125 cond_resched();
1126
1127 folio = lru_to_folio(head: folio_list);
1128 list_del(entry: &folio->lru);
1129
1130 if (!folio_trylock(folio))
1131 goto keep;
1132
1133 if (folio_contain_hwpoisoned_page(folio)) {
1134 /*
1135 * unmap_poisoned_folio() can't handle large
1136 * folio, just skip it. memory_failure() will
1137 * handle it if the UCE is triggered again.
1138 */
1139 if (folio_test_large(folio))
1140 goto keep_locked;
1141
1142 unmap_poisoned_folio(folio, pfn: folio_pfn(folio), must_kill: false);
1143 folio_unlock(folio);
1144 folio_put(folio);
1145 continue;
1146 }
1147
1148 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1149
1150 nr_pages = folio_nr_pages(folio);
1151
1152 /* Account the number of base pages */
1153 sc->nr_scanned += nr_pages;
1154
1155 if (unlikely(!folio_evictable(folio)))
1156 goto activate_locked;
1157
1158 if (!sc->may_unmap && folio_mapped(folio))
1159 goto keep_locked;
1160
1161 /*
1162 * The number of dirty pages determines if a node is marked
1163 * reclaim_congested. kswapd will stall and start writing
1164 * folios if the tail of the LRU is all dirty unqueued folios.
1165 */
1166 folio_check_dirty_writeback(folio, dirty: &dirty, writeback: &writeback);
1167 if (dirty || writeback)
1168 stat->nr_dirty += nr_pages;
1169
1170 if (dirty && !writeback)
1171 stat->nr_unqueued_dirty += nr_pages;
1172
1173 /*
1174 * Treat this folio as congested if folios are cycling
1175 * through the LRU so quickly that the folios marked
1176 * for immediate reclaim are making it to the end of
1177 * the LRU a second time.
1178 */
1179 if (writeback && folio_test_reclaim(folio))
1180 stat->nr_congested += nr_pages;
1181
1182 /*
1183 * If a folio at the tail of the LRU is under writeback, there
1184 * are three cases to consider.
1185 *
1186 * 1) If reclaim is encountering an excessive number
1187 * of folios under writeback and this folio has both
1188 * the writeback and reclaim flags set, then it
1189 * indicates that folios are being queued for I/O but
1190 * are being recycled through the LRU before the I/O
1191 * can complete. Waiting on the folio itself risks an
1192 * indefinite stall if it is impossible to writeback
1193 * the folio due to I/O error or disconnected storage
1194 * so instead note that the LRU is being scanned too
1195 * quickly and the caller can stall after the folio
1196 * list has been processed.
1197 *
1198 * 2) Global or new memcg reclaim encounters a folio that is
1199 * not marked for immediate reclaim, or the caller does not
1200 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1201 * not to fs), or the folio belongs to a mapping where
1202 * waiting on writeback during reclaim may lead to a deadlock.
1203 * In this case mark the folio for immediate reclaim and
1204 * continue scanning.
1205 *
1206 * Require may_enter_fs() because we would wait on fs, which
1207 * may not have submitted I/O yet. And the loop driver might
1208 * enter reclaim, and deadlock if it waits on a folio for
1209 * which it is needed to do the write (loop masks off
1210 * __GFP_IO|__GFP_FS for this reason); but more thought
1211 * would probably show more reasons.
1212 *
1213 * 3) Legacy memcg encounters a folio that already has the
1214 * reclaim flag set. memcg does not have any dirty folio
1215 * throttling so we could easily OOM just because too many
1216 * folios are in writeback and there is nothing else to
1217 * reclaim. Wait for the writeback to complete.
1218 *
1219 * In cases 1) and 2) we activate the folios to get them out of
1220 * the way while we continue scanning for clean folios on the
1221 * inactive list and refilling from the active list. The
1222 * observation here is that waiting for disk writes is more
1223 * expensive than potentially causing reloads down the line.
1224 * Since they're marked for immediate reclaim, they won't put
1225 * memory pressure on the cache working set any longer than it
1226 * takes to write them to disk.
1227 */
1228 if (folio_test_writeback(folio)) {
1229 mapping = folio_mapping(folio);
1230
1231 /* Case 1 above */
1232 if (current_is_kswapd() &&
1233 folio_test_reclaim(folio) &&
1234 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1235 stat->nr_immediate += nr_pages;
1236 goto activate_locked;
1237
1238 /* Case 2 above */
1239 } else if (writeback_throttling_sane(sc) ||
1240 !folio_test_reclaim(folio) ||
1241 !may_enter_fs(folio, gfp_mask: sc->gfp_mask) ||
1242 (mapping &&
1243 mapping_writeback_may_deadlock_on_reclaim(mapping))) {
1244 /*
1245 * This is slightly racy -
1246 * folio_end_writeback() might have
1247 * just cleared the reclaim flag, then
1248 * setting the reclaim flag here ends up
1249 * interpreted as the readahead flag - but
1250 * that does not matter enough to care.
1251 * What we do want is for this folio to
1252 * have the reclaim flag set next time
1253 * memcg reclaim reaches the tests above,
1254 * so it will then wait for writeback to
1255 * avoid OOM; and it's also appropriate
1256 * in global reclaim.
1257 */
1258 folio_set_reclaim(folio);
1259 stat->nr_writeback += nr_pages;
1260 goto activate_locked;
1261
1262 /* Case 3 above */
1263 } else {
1264 folio_unlock(folio);
1265 folio_wait_writeback(folio);
1266 /* then go back and try same folio again */
1267 list_add_tail(new: &folio->lru, head: folio_list);
1268 continue;
1269 }
1270 }
1271
1272 if (!ignore_references)
1273 references = folio_check_references(folio, sc);
1274
1275 switch (references) {
1276 case FOLIOREF_ACTIVATE:
1277 goto activate_locked;
1278 case FOLIOREF_KEEP:
1279 stat->nr_ref_keep += nr_pages;
1280 goto keep_locked;
1281 case FOLIOREF_RECLAIM:
1282 case FOLIOREF_RECLAIM_CLEAN:
1283 ; /* try to reclaim the folio below */
1284 }
1285
1286 /*
1287 * Before reclaiming the folio, try to relocate
1288 * its contents to another node.
1289 */
1290 if (do_demote_pass &&
1291 (thp_migration_supported() || !folio_test_large(folio))) {
1292 list_add(new: &folio->lru, head: &demote_folios);
1293 folio_unlock(folio);
1294 continue;
1295 }
1296
1297 /*
1298 * Anonymous process memory has backing store?
1299 * Try to allocate it some swap space here.
1300 * Lazyfree folio could be freed directly
1301 */
1302 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1303 if (!folio_test_swapcache(folio)) {
1304 if (!(sc->gfp_mask & __GFP_IO))
1305 goto keep_locked;
1306 if (folio_maybe_dma_pinned(folio))
1307 goto keep_locked;
1308 if (folio_test_large(folio)) {
1309 /* cannot split folio, skip it */
1310 if (!can_split_folio(folio, caller_pins: 1, NULL))
1311 goto activate_locked;
1312 /*
1313 * Split partially mapped folios right away.
1314 * We can free the unmapped pages without IO.
1315 */
1316 if (data_race(!list_empty(&folio->_deferred_list) &&
1317 folio_test_partially_mapped(folio)) &&
1318 split_folio_to_list(folio, list: folio_list))
1319 goto activate_locked;
1320 }
1321 if (folio_alloc_swap(folio, __GFP_HIGH | __GFP_NOWARN)) {
1322 int __maybe_unused order = folio_order(folio);
1323
1324 if (!folio_test_large(folio))
1325 goto activate_locked_split;
1326 /* Fallback to swap normal pages */
1327 if (split_folio_to_list(folio, list: folio_list))
1328 goto activate_locked;
1329#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1330 if (nr_pages >= HPAGE_PMD_NR) {
1331 count_memcg_folio_events(folio,
1332 THP_SWPOUT_FALLBACK, 1);
1333 count_vm_event(THP_SWPOUT_FALLBACK);
1334 }
1335#endif
1336 count_mthp_stat(order, item: MTHP_STAT_SWPOUT_FALLBACK);
1337 if (folio_alloc_swap(folio, __GFP_HIGH | __GFP_NOWARN))
1338 goto activate_locked_split;
1339 }
1340 /*
1341 * Normally the folio will be dirtied in unmap because its
1342 * pte should be dirty. A special case is MADV_FREE page. The
1343 * page's pte could have dirty bit cleared but the folio's
1344 * SwapBacked flag is still set because clearing the dirty bit
1345 * and SwapBacked flag has no lock protected. For such folio,
1346 * unmap will not set dirty bit for it, so folio reclaim will
1347 * not write the folio out. This can cause data corruption when
1348 * the folio is swapped in later. Always setting the dirty flag
1349 * for the folio solves the problem.
1350 */
1351 folio_mark_dirty(folio);
1352 }
1353 }
1354
1355 /*
1356 * If the folio was split above, the tail pages will make
1357 * their own pass through this function and be accounted
1358 * then.
1359 */
1360 if ((nr_pages > 1) && !folio_test_large(folio)) {
1361 sc->nr_scanned -= (nr_pages - 1);
1362 nr_pages = 1;
1363 }
1364
1365 /*
1366 * The folio is mapped into the page tables of one or more
1367 * processes. Try to unmap it here.
1368 */
1369 if (folio_mapped(folio)) {
1370 enum ttu_flags flags = TTU_BATCH_FLUSH;
1371 bool was_swapbacked = folio_test_swapbacked(folio);
1372
1373 if (folio_test_pmd_mappable(folio))
1374 flags |= TTU_SPLIT_HUGE_PMD;
1375 /*
1376 * Without TTU_SYNC, try_to_unmap will only begin to
1377 * hold PTL from the first present PTE within a large
1378 * folio. Some initial PTEs might be skipped due to
1379 * races with parallel PTE writes in which PTEs can be
1380 * cleared temporarily before being written new present
1381 * values. This will lead to a large folio is still
1382 * mapped while some subpages have been partially
1383 * unmapped after try_to_unmap; TTU_SYNC helps
1384 * try_to_unmap acquire PTL from the first PTE,
1385 * eliminating the influence of temporary PTE values.
1386 */
1387 if (folio_test_large(folio))
1388 flags |= TTU_SYNC;
1389
1390 try_to_unmap(folio, flags);
1391 if (folio_mapped(folio)) {
1392 stat->nr_unmap_fail += nr_pages;
1393 if (!was_swapbacked &&
1394 folio_test_swapbacked(folio))
1395 stat->nr_lazyfree_fail += nr_pages;
1396 goto activate_locked;
1397 }
1398 }
1399
1400 /*
1401 * Folio is unmapped now so it cannot be newly pinned anymore.
1402 * No point in trying to reclaim folio if it is pinned.
1403 * Furthermore we don't want to reclaim underlying fs metadata
1404 * if the folio is pinned and thus potentially modified by the
1405 * pinning process as that may upset the filesystem.
1406 */
1407 if (folio_maybe_dma_pinned(folio))
1408 goto activate_locked;
1409
1410 mapping = folio_mapping(folio);
1411 if (folio_test_dirty(folio)) {
1412 /*
1413 * Only kswapd can writeback filesystem folios
1414 * to avoid risk of stack overflow. But avoid
1415 * injecting inefficient single-folio I/O into
1416 * flusher writeback as much as possible: only
1417 * write folios when we've encountered many
1418 * dirty folios, and when we've already scanned
1419 * the rest of the LRU for clean folios and see
1420 * the same dirty folios again (with the reclaim
1421 * flag set).
1422 */
1423 if (folio_is_file_lru(folio) &&
1424 (!current_is_kswapd() ||
1425 !folio_test_reclaim(folio) ||
1426 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1427 /*
1428 * Immediately reclaim when written back.
1429 * Similar in principle to folio_deactivate()
1430 * except we already have the folio isolated
1431 * and know it's dirty
1432 */
1433 node_stat_mod_folio(folio, item: NR_VMSCAN_IMMEDIATE,
1434 nr: nr_pages);
1435 folio_set_reclaim(folio);
1436
1437 goto activate_locked;
1438 }
1439
1440 if (references == FOLIOREF_RECLAIM_CLEAN)
1441 goto keep_locked;
1442 if (!may_enter_fs(folio, gfp_mask: sc->gfp_mask))
1443 goto keep_locked;
1444 if (!sc->may_writepage)
1445 goto keep_locked;
1446
1447 /*
1448 * Folio is dirty. Flush the TLB if a writable entry
1449 * potentially exists to avoid CPU writes after I/O
1450 * starts and then write it out here.
1451 */
1452 try_to_unmap_flush_dirty();
1453 switch (pageout(folio, mapping, plug: &plug, folio_list)) {
1454 case PAGE_KEEP:
1455 goto keep_locked;
1456 case PAGE_ACTIVATE:
1457 /*
1458 * If shmem folio is split when writeback to swap,
1459 * the tail pages will make their own pass through
1460 * this function and be accounted then.
1461 */
1462 if (nr_pages > 1 && !folio_test_large(folio)) {
1463 sc->nr_scanned -= (nr_pages - 1);
1464 nr_pages = 1;
1465 }
1466 goto activate_locked;
1467 case PAGE_SUCCESS:
1468 if (nr_pages > 1 && !folio_test_large(folio)) {
1469 sc->nr_scanned -= (nr_pages - 1);
1470 nr_pages = 1;
1471 }
1472 stat->nr_pageout += nr_pages;
1473
1474 if (folio_test_writeback(folio))
1475 goto keep;
1476 if (folio_test_dirty(folio))
1477 goto keep;
1478
1479 /*
1480 * A synchronous write - probably a ramdisk. Go
1481 * ahead and try to reclaim the folio.
1482 */
1483 if (!folio_trylock(folio))
1484 goto keep;
1485 if (folio_test_dirty(folio) ||
1486 folio_test_writeback(folio))
1487 goto keep_locked;
1488 mapping = folio_mapping(folio);
1489 fallthrough;
1490 case PAGE_CLEAN:
1491 ; /* try to free the folio below */
1492 }
1493 }
1494
1495 /*
1496 * If the folio has buffers, try to free the buffer
1497 * mappings associated with this folio. If we succeed
1498 * we try to free the folio as well.
1499 *
1500 * We do this even if the folio is dirty.
1501 * filemap_release_folio() does not perform I/O, but it
1502 * is possible for a folio to have the dirty flag set,
1503 * but it is actually clean (all its buffers are clean).
1504 * This happens if the buffers were written out directly,
1505 * with submit_bh(). ext3 will do this, as well as
1506 * the blockdev mapping. filemap_release_folio() will
1507 * discover that cleanness and will drop the buffers
1508 * and mark the folio clean - it can be freed.
1509 *
1510 * Rarely, folios can have buffers and no ->mapping.
1511 * These are the folios which were not successfully
1512 * invalidated in truncate_cleanup_folio(). We try to
1513 * drop those buffers here and if that worked, and the
1514 * folio is no longer mapped into process address space
1515 * (refcount == 1) it can be freed. Otherwise, leave
1516 * the folio on the LRU so it is swappable.
1517 */
1518 if (folio_needs_release(folio)) {
1519 if (!filemap_release_folio(folio, gfp: sc->gfp_mask))
1520 goto activate_locked;
1521 if (!mapping && folio_ref_count(folio) == 1) {
1522 folio_unlock(folio);
1523 if (folio_put_testzero(folio))
1524 goto free_it;
1525 else {
1526 /*
1527 * rare race with speculative reference.
1528 * the speculative reference will free
1529 * this folio shortly, so we may
1530 * increment nr_reclaimed here (and
1531 * leave it off the LRU).
1532 */
1533 nr_reclaimed += nr_pages;
1534 continue;
1535 }
1536 }
1537 }
1538
1539 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1540 /* follow __remove_mapping for reference */
1541 if (!folio_ref_freeze(folio, count: 1))
1542 goto keep_locked;
1543 /*
1544 * The folio has only one reference left, which is
1545 * from the isolation. After the caller puts the
1546 * folio back on the lru and drops the reference, the
1547 * folio will be freed anyway. It doesn't matter
1548 * which lru it goes on. So we don't bother checking
1549 * the dirty flag here.
1550 */
1551 count_vm_events(item: PGLAZYFREED, delta: nr_pages);
1552 count_memcg_folio_events(folio, idx: PGLAZYFREED, nr: nr_pages);
1553 } else if (!mapping || !__remove_mapping(mapping, folio, reclaimed: true,
1554 target_memcg: sc->target_mem_cgroup))
1555 goto keep_locked;
1556
1557 folio_unlock(folio);
1558free_it:
1559 /*
1560 * Folio may get swapped out as a whole, need to account
1561 * all pages in it.
1562 */
1563 nr_reclaimed += nr_pages;
1564
1565 folio_unqueue_deferred_split(folio);
1566 if (folio_batch_add(fbatch: &free_folios, folio) == 0) {
1567 mem_cgroup_uncharge_folios(folios: &free_folios);
1568 try_to_unmap_flush();
1569 free_unref_folios(fbatch: &free_folios);
1570 }
1571 continue;
1572
1573activate_locked_split:
1574 /*
1575 * The tail pages that are failed to add into swap cache
1576 * reach here. Fixup nr_scanned and nr_pages.
1577 */
1578 if (nr_pages > 1) {
1579 sc->nr_scanned -= (nr_pages - 1);
1580 nr_pages = 1;
1581 }
1582activate_locked:
1583 /* Not a candidate for swapping, so reclaim swap space. */
1584 if (folio_test_swapcache(folio) &&
1585 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1586 folio_free_swap(folio);
1587 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1588 if (!folio_test_mlocked(folio)) {
1589 int type = folio_is_file_lru(folio);
1590 folio_set_active(folio);
1591 stat->nr_activate[type] += nr_pages;
1592 count_memcg_folio_events(folio, idx: PGACTIVATE, nr: nr_pages);
1593 }
1594keep_locked:
1595 folio_unlock(folio);
1596keep:
1597 list_add(new: &folio->lru, head: &ret_folios);
1598 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1599 folio_test_unevictable(folio), folio);
1600 }
1601 /* 'folio_list' is always empty here */
1602
1603 /* Migrate folios selected for demotion */
1604 nr_demoted = demote_folio_list(demote_folios: &demote_folios, pgdat);
1605 nr_reclaimed += nr_demoted;
1606 stat->nr_demoted += nr_demoted;
1607 /* Folios that could not be demoted are still in @demote_folios */
1608 if (!list_empty(head: &demote_folios)) {
1609 /* Folios which weren't demoted go back on @folio_list */
1610 list_splice_init(list: &demote_folios, head: folio_list);
1611
1612 /*
1613 * goto retry to reclaim the undemoted folios in folio_list if
1614 * desired.
1615 *
1616 * Reclaiming directly from top tier nodes is not often desired
1617 * due to it breaking the LRU ordering: in general memory
1618 * should be reclaimed from lower tier nodes and demoted from
1619 * top tier nodes.
1620 *
1621 * However, disabling reclaim from top tier nodes entirely
1622 * would cause ooms in edge scenarios where lower tier memory
1623 * is unreclaimable for whatever reason, eg memory being
1624 * mlocked or too hot to reclaim. We can disable reclaim
1625 * from top tier nodes in proactive reclaim though as that is
1626 * not real memory pressure.
1627 */
1628 if (!sc->proactive) {
1629 do_demote_pass = false;
1630 goto retry;
1631 }
1632 }
1633
1634 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1635
1636 mem_cgroup_uncharge_folios(folios: &free_folios);
1637 try_to_unmap_flush();
1638 free_unref_folios(fbatch: &free_folios);
1639
1640 list_splice(list: &ret_folios, head: folio_list);
1641 count_vm_events(item: PGACTIVATE, delta: pgactivate);
1642
1643 if (plug)
1644 swap_write_unplug(sio: plug);
1645 return nr_reclaimed;
1646}
1647
1648unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1649 struct list_head *folio_list)
1650{
1651 struct scan_control sc = {
1652 .gfp_mask = GFP_KERNEL,
1653 .may_unmap = 1,
1654 };
1655 struct reclaim_stat stat;
1656 unsigned int nr_reclaimed;
1657 struct folio *folio, *next;
1658 LIST_HEAD(clean_folios);
1659 unsigned int noreclaim_flag;
1660
1661 list_for_each_entry_safe(folio, next, folio_list, lru) {
1662 /* TODO: these pages should not even appear in this list. */
1663 if (page_has_movable_ops(page: &folio->page))
1664 continue;
1665 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1666 !folio_test_dirty(folio) && !folio_test_unevictable(folio)) {
1667 folio_clear_active(folio);
1668 list_move(list: &folio->lru, head: &clean_folios);
1669 }
1670 }
1671
1672 /*
1673 * We should be safe here since we are only dealing with file pages and
1674 * we are not kswapd and therefore cannot write dirty file pages. But
1675 * call memalloc_noreclaim_save() anyway, just in case these conditions
1676 * change in the future.
1677 */
1678 noreclaim_flag = memalloc_noreclaim_save();
1679 nr_reclaimed = shrink_folio_list(folio_list: &clean_folios, pgdat: zone->zone_pgdat, sc: &sc,
1680 stat: &stat, ignore_references: true, NULL);
1681 memalloc_noreclaim_restore(flags: noreclaim_flag);
1682
1683 list_splice(list: &clean_folios, head: folio_list);
1684 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1685 -(long)nr_reclaimed);
1686 /*
1687 * Since lazyfree pages are isolated from file LRU from the beginning,
1688 * they will rotate back to anonymous LRU in the end if it failed to
1689 * discard so isolated count will be mismatched.
1690 * Compensate the isolated count for both LRU lists.
1691 */
1692 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1693 stat.nr_lazyfree_fail);
1694 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1695 -(long)stat.nr_lazyfree_fail);
1696 return nr_reclaimed;
1697}
1698
1699/*
1700 * Update LRU sizes after isolating pages. The LRU size updates must
1701 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1702 */
1703static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1704 enum lru_list lru, unsigned long *nr_zone_taken)
1705{
1706 int zid;
1707
1708 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1709 if (!nr_zone_taken[zid])
1710 continue;
1711
1712 update_lru_size(lruvec, lru, zid, nr_pages: -nr_zone_taken[zid]);
1713 }
1714
1715}
1716
1717/*
1718 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1719 *
1720 * lruvec->lru_lock is heavily contended. Some of the functions that
1721 * shrink the lists perform better by taking out a batch of pages
1722 * and working on them outside the LRU lock.
1723 *
1724 * For pagecache intensive workloads, this function is the hottest
1725 * spot in the kernel (apart from copy_*_user functions).
1726 *
1727 * Lru_lock must be held before calling this function.
1728 *
1729 * @nr_to_scan: The number of eligible pages to look through on the list.
1730 * @lruvec: The LRU vector to pull pages from.
1731 * @dst: The temp list to put pages on to.
1732 * @nr_scanned: The number of pages that were scanned.
1733 * @sc: The scan_control struct for this reclaim session
1734 * @lru: LRU list id for isolating
1735 *
1736 * returns how many pages were moved onto *@dst.
1737 */
1738static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1739 struct lruvec *lruvec, struct list_head *dst,
1740 unsigned long *nr_scanned, struct scan_control *sc,
1741 enum lru_list lru)
1742{
1743 struct list_head *src = &lruvec->lists[lru];
1744 unsigned long nr_taken = 0;
1745 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1746 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1747 unsigned long skipped = 0, total_scan = 0, scan = 0;
1748 unsigned long nr_pages;
1749 unsigned long max_nr_skipped = 0;
1750 LIST_HEAD(folios_skipped);
1751
1752 while (scan < nr_to_scan && !list_empty(head: src)) {
1753 struct list_head *move_to = src;
1754 struct folio *folio;
1755
1756 folio = lru_to_folio(head: src);
1757 prefetchw_prev_lru_folio(folio, src, flags);
1758
1759 nr_pages = folio_nr_pages(folio);
1760 total_scan += nr_pages;
1761
1762 /* Using max_nr_skipped to prevent hard LOCKUP*/
1763 if (max_nr_skipped < SWAP_CLUSTER_MAX_SKIPPED &&
1764 (folio_zonenum(folio) > sc->reclaim_idx)) {
1765 nr_skipped[folio_zonenum(folio)] += nr_pages;
1766 move_to = &folios_skipped;
1767 max_nr_skipped++;
1768 goto move;
1769 }
1770
1771 /*
1772 * Do not count skipped folios because that makes the function
1773 * return with no isolated folios if the LRU mostly contains
1774 * ineligible folios. This causes the VM to not reclaim any
1775 * folios, triggering a premature OOM.
1776 * Account all pages in a folio.
1777 */
1778 scan += nr_pages;
1779
1780 if (!folio_test_lru(folio))
1781 goto move;
1782 if (!sc->may_unmap && folio_mapped(folio))
1783 goto move;
1784
1785 /*
1786 * Be careful not to clear the lru flag until after we're
1787 * sure the folio is not being freed elsewhere -- the
1788 * folio release code relies on it.
1789 */
1790 if (unlikely(!folio_try_get(folio)))
1791 goto move;
1792
1793 if (!folio_test_clear_lru(folio)) {
1794 /* Another thread is already isolating this folio */
1795 folio_put(folio);
1796 goto move;
1797 }
1798
1799 nr_taken += nr_pages;
1800 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1801 move_to = dst;
1802move:
1803 list_move(list: &folio->lru, head: move_to);
1804 }
1805
1806 /*
1807 * Splice any skipped folios to the start of the LRU list. Note that
1808 * this disrupts the LRU order when reclaiming for lower zones but
1809 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1810 * scanning would soon rescan the same folios to skip and waste lots
1811 * of cpu cycles.
1812 */
1813 if (!list_empty(head: &folios_skipped)) {
1814 int zid;
1815
1816 list_splice(list: &folios_skipped, head: src);
1817 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1818 if (!nr_skipped[zid])
1819 continue;
1820
1821 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1822 skipped += nr_skipped[zid];
1823 }
1824 }
1825 *nr_scanned = total_scan;
1826 trace_mm_vmscan_lru_isolate(highest_zoneidx: sc->reclaim_idx, order: sc->order, nr_requested: nr_to_scan,
1827 nr_scanned: total_scan, nr_skipped: skipped, nr_taken, lru);
1828 update_lru_sizes(lruvec, lru, nr_zone_taken);
1829 return nr_taken;
1830}
1831
1832/**
1833 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1834 * @folio: Folio to isolate from its LRU list.
1835 *
1836 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1837 * corresponding to whatever LRU list the folio was on.
1838 *
1839 * The folio will have its LRU flag cleared. If it was found on the
1840 * active list, it will have the Active flag set. If it was found on the
1841 * unevictable list, it will have the Unevictable flag set. These flags
1842 * may need to be cleared by the caller before letting the page go.
1843 *
1844 * Context:
1845 *
1846 * (1) Must be called with an elevated refcount on the folio. This is a
1847 * fundamental difference from isolate_lru_folios() (which is called
1848 * without a stable reference).
1849 * (2) The lru_lock must not be held.
1850 * (3) Interrupts must be enabled.
1851 *
1852 * Return: true if the folio was removed from an LRU list.
1853 * false if the folio was not on an LRU list.
1854 */
1855bool folio_isolate_lru(struct folio *folio)
1856{
1857 bool ret = false;
1858
1859 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1860
1861 if (folio_test_clear_lru(folio)) {
1862 struct lruvec *lruvec;
1863
1864 folio_get(folio);
1865 lruvec = folio_lruvec_lock_irq(folio);
1866 lruvec_del_folio(lruvec, folio);
1867 unlock_page_lruvec_irq(lruvec);
1868 ret = true;
1869 }
1870
1871 return ret;
1872}
1873
1874/*
1875 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1876 * then get rescheduled. When there are massive number of tasks doing page
1877 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1878 * the LRU list will go small and be scanned faster than necessary, leading to
1879 * unnecessary swapping, thrashing and OOM.
1880 */
1881static bool too_many_isolated(struct pglist_data *pgdat, int file,
1882 struct scan_control *sc)
1883{
1884 unsigned long inactive, isolated;
1885 bool too_many;
1886
1887 if (current_is_kswapd())
1888 return false;
1889
1890 if (!writeback_throttling_sane(sc))
1891 return false;
1892
1893 if (file) {
1894 inactive = node_page_state(pgdat, item: NR_INACTIVE_FILE);
1895 isolated = node_page_state(pgdat, item: NR_ISOLATED_FILE);
1896 } else {
1897 inactive = node_page_state(pgdat, item: NR_INACTIVE_ANON);
1898 isolated = node_page_state(pgdat, item: NR_ISOLATED_ANON);
1899 }
1900
1901 /*
1902 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1903 * won't get blocked by normal direct-reclaimers, forming a circular
1904 * deadlock.
1905 */
1906 if (gfp_has_io_fs(gfp: sc->gfp_mask))
1907 inactive >>= 3;
1908
1909 too_many = isolated > inactive;
1910
1911 /* Wake up tasks throttled due to too_many_isolated. */
1912 if (!too_many)
1913 wake_throttle_isolated(pgdat);
1914
1915 return too_many;
1916}
1917
1918/*
1919 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1920 *
1921 * Returns the number of pages moved to the given lruvec.
1922 */
1923static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1924 struct list_head *list)
1925{
1926 int nr_pages, nr_moved = 0;
1927 struct folio_batch free_folios;
1928
1929 folio_batch_init(fbatch: &free_folios);
1930 while (!list_empty(head: list)) {
1931 struct folio *folio = lru_to_folio(head: list);
1932
1933 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1934 list_del(entry: &folio->lru);
1935 if (unlikely(!folio_evictable(folio))) {
1936 spin_unlock_irq(lock: &lruvec->lru_lock);
1937 folio_putback_lru(folio);
1938 spin_lock_irq(lock: &lruvec->lru_lock);
1939 continue;
1940 }
1941
1942 /*
1943 * The folio_set_lru needs to be kept here for list integrity.
1944 * Otherwise:
1945 * #0 move_folios_to_lru #1 release_pages
1946 * if (!folio_put_testzero())
1947 * if (folio_put_testzero())
1948 * !lru //skip lru_lock
1949 * folio_set_lru()
1950 * list_add(&folio->lru,)
1951 * list_add(&folio->lru,)
1952 */
1953 folio_set_lru(folio);
1954
1955 if (unlikely(folio_put_testzero(folio))) {
1956 __folio_clear_lru_flags(folio);
1957
1958 folio_unqueue_deferred_split(folio);
1959 if (folio_batch_add(fbatch: &free_folios, folio) == 0) {
1960 spin_unlock_irq(lock: &lruvec->lru_lock);
1961 mem_cgroup_uncharge_folios(folios: &free_folios);
1962 free_unref_folios(fbatch: &free_folios);
1963 spin_lock_irq(lock: &lruvec->lru_lock);
1964 }
1965
1966 continue;
1967 }
1968
1969 /*
1970 * All pages were isolated from the same lruvec (and isolation
1971 * inhibits memcg migration).
1972 */
1973 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1974 lruvec_add_folio(lruvec, folio);
1975 nr_pages = folio_nr_pages(folio);
1976 nr_moved += nr_pages;
1977 if (folio_test_active(folio))
1978 workingset_age_nonresident(lruvec, nr_pages);
1979 }
1980
1981 if (free_folios.nr) {
1982 spin_unlock_irq(lock: &lruvec->lru_lock);
1983 mem_cgroup_uncharge_folios(folios: &free_folios);
1984 free_unref_folios(fbatch: &free_folios);
1985 spin_lock_irq(lock: &lruvec->lru_lock);
1986 }
1987
1988 return nr_moved;
1989}
1990
1991/*
1992 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1993 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1994 * we should not throttle. Otherwise it is safe to do so.
1995 */
1996static int current_may_throttle(void)
1997{
1998 return !(current->flags & PF_LOCAL_THROTTLE);
1999}
2000
2001/*
2002 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2003 * of reclaimed pages
2004 */
2005static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2006 struct lruvec *lruvec, struct scan_control *sc,
2007 enum lru_list lru)
2008{
2009 LIST_HEAD(folio_list);
2010 unsigned long nr_scanned;
2011 unsigned int nr_reclaimed = 0;
2012 unsigned long nr_taken;
2013 struct reclaim_stat stat;
2014 bool file = is_file_lru(lru);
2015 enum vm_event_item item;
2016 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2017 bool stalled = false;
2018
2019 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2020 if (stalled)
2021 return 0;
2022
2023 /* wait a bit for the reclaimer. */
2024 stalled = true;
2025 reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_ISOLATED);
2026
2027 /* We are about to die and free our memory. Return now. */
2028 if (fatal_signal_pending(current))
2029 return SWAP_CLUSTER_MAX;
2030 }
2031
2032 lru_add_drain();
2033
2034 spin_lock_irq(lock: &lruvec->lru_lock);
2035
2036 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, dst: &folio_list,
2037 nr_scanned: &nr_scanned, sc, lru);
2038
2039 __mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, nr_taken);
2040 item = PGSCAN_KSWAPD + reclaimer_offset(sc);
2041 if (!cgroup_reclaim(sc))
2042 __count_vm_events(item, delta: nr_scanned);
2043 count_memcg_events(memcg: lruvec_memcg(lruvec), idx: item, count: nr_scanned);
2044 __count_vm_events(item: PGSCAN_ANON + file, delta: nr_scanned);
2045
2046 spin_unlock_irq(lock: &lruvec->lru_lock);
2047
2048 if (nr_taken == 0)
2049 return 0;
2050
2051 nr_reclaimed = shrink_folio_list(folio_list: &folio_list, pgdat, sc, stat: &stat, ignore_references: false,
2052 memcg: lruvec_memcg(lruvec));
2053
2054 spin_lock_irq(lock: &lruvec->lru_lock);
2055 move_folios_to_lru(lruvec, list: &folio_list);
2056
2057 __mod_lruvec_state(lruvec, idx: PGDEMOTE_KSWAPD + reclaimer_offset(sc),
2058 val: stat.nr_demoted);
2059 __mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, -nr_taken);
2060 item = PGSTEAL_KSWAPD + reclaimer_offset(sc);
2061 if (!cgroup_reclaim(sc))
2062 __count_vm_events(item, delta: nr_reclaimed);
2063 count_memcg_events(memcg: lruvec_memcg(lruvec), idx: item, count: nr_reclaimed);
2064 __count_vm_events(item: PGSTEAL_ANON + file, delta: nr_reclaimed);
2065
2066 lru_note_cost_unlock_irq(lruvec, file, nr_io: stat.nr_pageout,
2067 nr_rotated: nr_scanned - nr_reclaimed);
2068
2069 /*
2070 * If dirty folios are scanned that are not queued for IO, it
2071 * implies that flushers are not doing their job. This can
2072 * happen when memory pressure pushes dirty folios to the end of
2073 * the LRU before the dirty limits are breached and the dirty
2074 * data has expired. It can also happen when the proportion of
2075 * dirty folios grows not through writes but through memory
2076 * pressure reclaiming all the clean cache. And in some cases,
2077 * the flushers simply cannot keep up with the allocation
2078 * rate. Nudge the flusher threads in case they are asleep.
2079 */
2080 if (stat.nr_unqueued_dirty == nr_taken) {
2081 wakeup_flusher_threads(reason: WB_REASON_VMSCAN);
2082 /*
2083 * For cgroupv1 dirty throttling is achieved by waking up
2084 * the kernel flusher here and later waiting on folios
2085 * which are in writeback to finish (see shrink_folio_list()).
2086 *
2087 * Flusher may not be able to issue writeback quickly
2088 * enough for cgroupv1 writeback throttling to work
2089 * on a large system.
2090 */
2091 if (!writeback_throttling_sane(sc))
2092 reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_WRITEBACK);
2093 }
2094
2095 sc->nr.dirty += stat.nr_dirty;
2096 sc->nr.congested += stat.nr_congested;
2097 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2098 sc->nr.writeback += stat.nr_writeback;
2099 sc->nr.immediate += stat.nr_immediate;
2100 sc->nr.taken += nr_taken;
2101 if (file)
2102 sc->nr.file_taken += nr_taken;
2103
2104 trace_mm_vmscan_lru_shrink_inactive(nid: pgdat->node_id,
2105 nr_scanned, nr_reclaimed, stat: &stat, priority: sc->priority, file);
2106 return nr_reclaimed;
2107}
2108
2109/*
2110 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2111 *
2112 * We move them the other way if the folio is referenced by one or more
2113 * processes.
2114 *
2115 * If the folios are mostly unmapped, the processing is fast and it is
2116 * appropriate to hold lru_lock across the whole operation. But if
2117 * the folios are mapped, the processing is slow (folio_referenced()), so
2118 * we should drop lru_lock around each folio. It's impossible to balance
2119 * this, so instead we remove the folios from the LRU while processing them.
2120 * It is safe to rely on the active flag against the non-LRU folios in here
2121 * because nobody will play with that bit on a non-LRU folio.
2122 *
2123 * The downside is that we have to touch folio->_refcount against each folio.
2124 * But we had to alter folio->flags anyway.
2125 */
2126static void shrink_active_list(unsigned long nr_to_scan,
2127 struct lruvec *lruvec,
2128 struct scan_control *sc,
2129 enum lru_list lru)
2130{
2131 unsigned long nr_taken;
2132 unsigned long nr_scanned;
2133 vm_flags_t vm_flags;
2134 LIST_HEAD(l_hold); /* The folios which were snipped off */
2135 LIST_HEAD(l_active);
2136 LIST_HEAD(l_inactive);
2137 unsigned nr_deactivate, nr_activate;
2138 unsigned nr_rotated = 0;
2139 bool file = is_file_lru(lru);
2140 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2141
2142 lru_add_drain();
2143
2144 spin_lock_irq(lock: &lruvec->lru_lock);
2145
2146 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, dst: &l_hold,
2147 nr_scanned: &nr_scanned, sc, lru);
2148
2149 __mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, nr_taken);
2150
2151 if (!cgroup_reclaim(sc))
2152 __count_vm_events(item: PGREFILL, delta: nr_scanned);
2153 count_memcg_events(memcg: lruvec_memcg(lruvec), idx: PGREFILL, count: nr_scanned);
2154
2155 spin_unlock_irq(lock: &lruvec->lru_lock);
2156
2157 while (!list_empty(head: &l_hold)) {
2158 struct folio *folio;
2159
2160 cond_resched();
2161 folio = lru_to_folio(head: &l_hold);
2162 list_del(entry: &folio->lru);
2163
2164 if (unlikely(!folio_evictable(folio))) {
2165 folio_putback_lru(folio);
2166 continue;
2167 }
2168
2169 if (unlikely(buffer_heads_over_limit)) {
2170 if (folio_needs_release(folio) &&
2171 folio_trylock(folio)) {
2172 filemap_release_folio(folio, gfp: 0);
2173 folio_unlock(folio);
2174 }
2175 }
2176
2177 /* Referenced or rmap lock contention: rotate */
2178 if (folio_referenced(folio, is_locked: 0, memcg: sc->target_mem_cgroup,
2179 vm_flags: &vm_flags) != 0) {
2180 /*
2181 * Identify referenced, file-backed active folios and
2182 * give them one more trip around the active list. So
2183 * that executable code get better chances to stay in
2184 * memory under moderate memory pressure. Anon folios
2185 * are not likely to be evicted by use-once streaming
2186 * IO, plus JVM can create lots of anon VM_EXEC folios,
2187 * so we ignore them here.
2188 */
2189 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2190 nr_rotated += folio_nr_pages(folio);
2191 list_add(new: &folio->lru, head: &l_active);
2192 continue;
2193 }
2194 }
2195
2196 folio_clear_active(folio); /* we are de-activating */
2197 folio_set_workingset(folio);
2198 list_add(new: &folio->lru, head: &l_inactive);
2199 }
2200
2201 /*
2202 * Move folios back to the lru list.
2203 */
2204 spin_lock_irq(lock: &lruvec->lru_lock);
2205
2206 nr_activate = move_folios_to_lru(lruvec, list: &l_active);
2207 nr_deactivate = move_folios_to_lru(lruvec, list: &l_inactive);
2208
2209 __count_vm_events(item: PGDEACTIVATE, delta: nr_deactivate);
2210 count_memcg_events(memcg: lruvec_memcg(lruvec), idx: PGDEACTIVATE, count: nr_deactivate);
2211
2212 __mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, -nr_taken);
2213
2214 lru_note_cost_unlock_irq(lruvec, file, nr_io: 0, nr_rotated);
2215 trace_mm_vmscan_lru_shrink_active(nid: pgdat->node_id, nr_taken, nr_active: nr_activate,
2216 nr_deactivated: nr_deactivate, nr_referenced: nr_rotated, priority: sc->priority, file);
2217}
2218
2219static unsigned int reclaim_folio_list(struct list_head *folio_list,
2220 struct pglist_data *pgdat)
2221{
2222 struct reclaim_stat stat;
2223 unsigned int nr_reclaimed;
2224 struct folio *folio;
2225 struct scan_control sc = {
2226 .gfp_mask = GFP_KERNEL,
2227 .may_writepage = 1,
2228 .may_unmap = 1,
2229 .may_swap = 1,
2230 .no_demotion = 1,
2231 };
2232
2233 nr_reclaimed = shrink_folio_list(folio_list, pgdat, sc: &sc, stat: &stat, ignore_references: true, NULL);
2234 while (!list_empty(head: folio_list)) {
2235 folio = lru_to_folio(head: folio_list);
2236 list_del(entry: &folio->lru);
2237 folio_putback_lru(folio);
2238 }
2239 trace_mm_vmscan_reclaim_pages(nid: pgdat->node_id, nr_scanned: sc.nr_scanned, nr_reclaimed, stat: &stat);
2240
2241 return nr_reclaimed;
2242}
2243
2244unsigned long reclaim_pages(struct list_head *folio_list)
2245{
2246 int nid;
2247 unsigned int nr_reclaimed = 0;
2248 LIST_HEAD(node_folio_list);
2249 unsigned int noreclaim_flag;
2250
2251 if (list_empty(head: folio_list))
2252 return nr_reclaimed;
2253
2254 noreclaim_flag = memalloc_noreclaim_save();
2255
2256 nid = folio_nid(folio: lru_to_folio(head: folio_list));
2257 do {
2258 struct folio *folio = lru_to_folio(head: folio_list);
2259
2260 if (nid == folio_nid(folio)) {
2261 folio_clear_active(folio);
2262 list_move(list: &folio->lru, head: &node_folio_list);
2263 continue;
2264 }
2265
2266 nr_reclaimed += reclaim_folio_list(folio_list: &node_folio_list, NODE_DATA(nid));
2267 nid = folio_nid(folio: lru_to_folio(head: folio_list));
2268 } while (!list_empty(head: folio_list));
2269
2270 nr_reclaimed += reclaim_folio_list(folio_list: &node_folio_list, NODE_DATA(nid));
2271
2272 memalloc_noreclaim_restore(flags: noreclaim_flag);
2273
2274 return nr_reclaimed;
2275}
2276
2277static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2278 struct lruvec *lruvec, struct scan_control *sc)
2279{
2280 if (is_active_lru(lru)) {
2281 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2282 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2283 else
2284 sc->skipped_deactivate = 1;
2285 return 0;
2286 }
2287
2288 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2289}
2290
2291/*
2292 * The inactive anon list should be small enough that the VM never has
2293 * to do too much work.
2294 *
2295 * The inactive file list should be small enough to leave most memory
2296 * to the established workingset on the scan-resistant active list,
2297 * but large enough to avoid thrashing the aggregate readahead window.
2298 *
2299 * Both inactive lists should also be large enough that each inactive
2300 * folio has a chance to be referenced again before it is reclaimed.
2301 *
2302 * If that fails and refaulting is observed, the inactive list grows.
2303 *
2304 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2305 * on this LRU, maintained by the pageout code. An inactive_ratio
2306 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2307 *
2308 * total target max
2309 * memory ratio inactive
2310 * -------------------------------------
2311 * 10MB 1 5MB
2312 * 100MB 1 50MB
2313 * 1GB 3 250MB
2314 * 10GB 10 0.9GB
2315 * 100GB 31 3GB
2316 * 1TB 101 10GB
2317 * 10TB 320 32GB
2318 */
2319static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2320{
2321 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2322 unsigned long inactive, active;
2323 unsigned long inactive_ratio;
2324 unsigned long gb;
2325
2326 inactive = lruvec_page_state(lruvec, idx: NR_LRU_BASE + inactive_lru);
2327 active = lruvec_page_state(lruvec, idx: NR_LRU_BASE + active_lru);
2328
2329 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2330 if (gb)
2331 inactive_ratio = int_sqrt(10 * gb);
2332 else
2333 inactive_ratio = 1;
2334
2335 return inactive * inactive_ratio < active;
2336}
2337
2338enum scan_balance {
2339 SCAN_EQUAL,
2340 SCAN_FRACT,
2341 SCAN_ANON,
2342 SCAN_FILE,
2343};
2344
2345static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2346{
2347 unsigned long file;
2348 struct lruvec *target_lruvec;
2349
2350 if (lru_gen_enabled())
2351 return;
2352
2353 target_lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup, pgdat);
2354
2355 /*
2356 * Flush the memory cgroup stats in rate-limited way as we don't need
2357 * most accurate stats here. We may switch to regular stats flushing
2358 * in the future once it is cheap enough.
2359 */
2360 mem_cgroup_flush_stats_ratelimited(memcg: sc->target_mem_cgroup);
2361
2362 /*
2363 * Determine the scan balance between anon and file LRUs.
2364 */
2365 spin_lock_irq(lock: &target_lruvec->lru_lock);
2366 sc->anon_cost = target_lruvec->anon_cost;
2367 sc->file_cost = target_lruvec->file_cost;
2368 spin_unlock_irq(lock: &target_lruvec->lru_lock);
2369
2370 /*
2371 * Target desirable inactive:active list ratios for the anon
2372 * and file LRU lists.
2373 */
2374 if (!sc->force_deactivate) {
2375 unsigned long refaults;
2376
2377 /*
2378 * When refaults are being observed, it means a new
2379 * workingset is being established. Deactivate to get
2380 * rid of any stale active pages quickly.
2381 */
2382 refaults = lruvec_page_state(lruvec: target_lruvec,
2383 idx: WORKINGSET_ACTIVATE_ANON);
2384 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2385 inactive_is_low(lruvec: target_lruvec, inactive_lru: LRU_INACTIVE_ANON))
2386 sc->may_deactivate |= DEACTIVATE_ANON;
2387 else
2388 sc->may_deactivate &= ~DEACTIVATE_ANON;
2389
2390 refaults = lruvec_page_state(lruvec: target_lruvec,
2391 idx: WORKINGSET_ACTIVATE_FILE);
2392 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2393 inactive_is_low(lruvec: target_lruvec, inactive_lru: LRU_INACTIVE_FILE))
2394 sc->may_deactivate |= DEACTIVATE_FILE;
2395 else
2396 sc->may_deactivate &= ~DEACTIVATE_FILE;
2397 } else
2398 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2399
2400 /*
2401 * If we have plenty of inactive file pages that aren't
2402 * thrashing, try to reclaim those first before touching
2403 * anonymous pages.
2404 */
2405 file = lruvec_page_state(lruvec: target_lruvec, idx: NR_INACTIVE_FILE);
2406 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2407 !sc->no_cache_trim_mode)
2408 sc->cache_trim_mode = 1;
2409 else
2410 sc->cache_trim_mode = 0;
2411
2412 /*
2413 * Prevent the reclaimer from falling into the cache trap: as
2414 * cache pages start out inactive, every cache fault will tip
2415 * the scan balance towards the file LRU. And as the file LRU
2416 * shrinks, so does the window for rotation from references.
2417 * This means we have a runaway feedback loop where a tiny
2418 * thrashing file LRU becomes infinitely more attractive than
2419 * anon pages. Try to detect this based on file LRU size.
2420 */
2421 if (!cgroup_reclaim(sc)) {
2422 unsigned long total_high_wmark = 0;
2423 unsigned long free, anon;
2424 int z;
2425 struct zone *zone;
2426
2427 free = sum_zone_node_page_state(node: pgdat->node_id, item: NR_FREE_PAGES);
2428 file = node_page_state(pgdat, item: NR_ACTIVE_FILE) +
2429 node_page_state(pgdat, item: NR_INACTIVE_FILE);
2430
2431 for_each_managed_zone_pgdat(zone, pgdat, z, MAX_NR_ZONES - 1) {
2432 total_high_wmark += high_wmark_pages(z: zone);
2433 }
2434
2435 /*
2436 * Consider anon: if that's low too, this isn't a
2437 * runaway file reclaim problem, but rather just
2438 * extreme pressure. Reclaim as per usual then.
2439 */
2440 anon = node_page_state(pgdat, item: NR_INACTIVE_ANON);
2441
2442 sc->file_is_tiny =
2443 file + free <= total_high_wmark &&
2444 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2445 anon >> sc->priority;
2446 }
2447}
2448
2449static inline void calculate_pressure_balance(struct scan_control *sc,
2450 int swappiness, u64 *fraction, u64 *denominator)
2451{
2452 unsigned long anon_cost, file_cost, total_cost;
2453 unsigned long ap, fp;
2454
2455 /*
2456 * Calculate the pressure balance between anon and file pages.
2457 *
2458 * The amount of pressure we put on each LRU is inversely
2459 * proportional to the cost of reclaiming each list, as
2460 * determined by the share of pages that are refaulting, times
2461 * the relative IO cost of bringing back a swapped out
2462 * anonymous page vs reloading a filesystem page (swappiness).
2463 *
2464 * Although we limit that influence to ensure no list gets
2465 * left behind completely: at least a third of the pressure is
2466 * applied, before swappiness.
2467 *
2468 * With swappiness at 100, anon and file have equal IO cost.
2469 */
2470 total_cost = sc->anon_cost + sc->file_cost;
2471 anon_cost = total_cost + sc->anon_cost;
2472 file_cost = total_cost + sc->file_cost;
2473 total_cost = anon_cost + file_cost;
2474
2475 ap = swappiness * (total_cost + 1);
2476 ap /= anon_cost + 1;
2477
2478 fp = (MAX_SWAPPINESS - swappiness) * (total_cost + 1);
2479 fp /= file_cost + 1;
2480
2481 fraction[WORKINGSET_ANON] = ap;
2482 fraction[WORKINGSET_FILE] = fp;
2483 *denominator = ap + fp;
2484}
2485
2486static unsigned long apply_proportional_protection(struct mem_cgroup *memcg,
2487 struct scan_control *sc, unsigned long scan)
2488{
2489 unsigned long min, low;
2490
2491 mem_cgroup_protection(root: sc->target_mem_cgroup, memcg, min: &min, low: &low);
2492
2493 if (min || low) {
2494 /*
2495 * Scale a cgroup's reclaim pressure by proportioning
2496 * its current usage to its memory.low or memory.min
2497 * setting.
2498 *
2499 * This is important, as otherwise scanning aggression
2500 * becomes extremely binary -- from nothing as we
2501 * approach the memory protection threshold, to totally
2502 * nominal as we exceed it. This results in requiring
2503 * setting extremely liberal protection thresholds. It
2504 * also means we simply get no protection at all if we
2505 * set it too low, which is not ideal.
2506 *
2507 * If there is any protection in place, we reduce scan
2508 * pressure by how much of the total memory used is
2509 * within protection thresholds.
2510 *
2511 * There is one special case: in the first reclaim pass,
2512 * we skip over all groups that are within their low
2513 * protection. If that fails to reclaim enough pages to
2514 * satisfy the reclaim goal, we come back and override
2515 * the best-effort low protection. However, we still
2516 * ideally want to honor how well-behaved groups are in
2517 * that case instead of simply punishing them all
2518 * equally. As such, we reclaim them based on how much
2519 * memory they are using, reducing the scan pressure
2520 * again by how much of the total memory used is under
2521 * hard protection.
2522 */
2523 unsigned long cgroup_size = mem_cgroup_size(memcg);
2524 unsigned long protection;
2525
2526 /* memory.low scaling, make sure we retry before OOM */
2527 if (!sc->memcg_low_reclaim && low > min) {
2528 protection = low;
2529 sc->memcg_low_skipped = 1;
2530 } else {
2531 protection = min;
2532 }
2533
2534 /* Avoid TOCTOU with earlier protection check */
2535 cgroup_size = max(cgroup_size, protection);
2536
2537 scan -= scan * protection / (cgroup_size + 1);
2538
2539 /*
2540 * Minimally target SWAP_CLUSTER_MAX pages to keep
2541 * reclaim moving forwards, avoiding decrementing
2542 * sc->priority further than desirable.
2543 */
2544 scan = max(scan, SWAP_CLUSTER_MAX);
2545 }
2546 return scan;
2547}
2548
2549/*
2550 * Determine how aggressively the anon and file LRU lists should be
2551 * scanned.
2552 *
2553 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2554 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2555 */
2556static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2557 unsigned long *nr)
2558{
2559 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2560 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2561 int swappiness = sc_swappiness(sc, memcg);
2562 u64 fraction[ANON_AND_FILE];
2563 u64 denominator = 0; /* gcc */
2564 enum scan_balance scan_balance;
2565 enum lru_list lru;
2566
2567 /* If we have no swap space, do not bother scanning anon folios. */
2568 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, nid: pgdat->node_id, sc)) {
2569 scan_balance = SCAN_FILE;
2570 goto out;
2571 }
2572
2573 /*
2574 * Global reclaim will swap to prevent OOM even with no
2575 * swappiness, but memcg users want to use this knob to
2576 * disable swapping for individual groups completely when
2577 * using the memory controller's swap limit feature would be
2578 * too expensive.
2579 */
2580 if (cgroup_reclaim(sc) && !swappiness) {
2581 scan_balance = SCAN_FILE;
2582 goto out;
2583 }
2584
2585 /* Proactive reclaim initiated by userspace for anonymous memory only */
2586 if (swappiness == SWAPPINESS_ANON_ONLY) {
2587 WARN_ON_ONCE(!sc->proactive);
2588 scan_balance = SCAN_ANON;
2589 goto out;
2590 }
2591
2592 /*
2593 * Do not apply any pressure balancing cleverness when the
2594 * system is close to OOM, scan both anon and file equally
2595 * (unless the swappiness setting disagrees with swapping).
2596 */
2597 if (!sc->priority && swappiness) {
2598 scan_balance = SCAN_EQUAL;
2599 goto out;
2600 }
2601
2602 /*
2603 * If the system is almost out of file pages, force-scan anon.
2604 */
2605 if (sc->file_is_tiny) {
2606 scan_balance = SCAN_ANON;
2607 goto out;
2608 }
2609
2610 /*
2611 * If there is enough inactive page cache, we do not reclaim
2612 * anything from the anonymous working right now to make sure
2613 * a streaming file access pattern doesn't cause swapping.
2614 */
2615 if (sc->cache_trim_mode) {
2616 scan_balance = SCAN_FILE;
2617 goto out;
2618 }
2619
2620 scan_balance = SCAN_FRACT;
2621 calculate_pressure_balance(sc, swappiness, fraction, denominator: &denominator);
2622
2623out:
2624 for_each_evictable_lru(lru) {
2625 bool file = is_file_lru(lru);
2626 unsigned long lruvec_size;
2627 unsigned long scan;
2628
2629 lruvec_size = lruvec_lru_size(lruvec, lru, zone_idx: sc->reclaim_idx);
2630 scan = apply_proportional_protection(memcg, sc, scan: lruvec_size);
2631 scan >>= sc->priority;
2632
2633 /*
2634 * If the cgroup's already been deleted, make sure to
2635 * scrape out the remaining cache.
2636 */
2637 if (!scan && !mem_cgroup_online(memcg))
2638 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2639
2640 switch (scan_balance) {
2641 case SCAN_EQUAL:
2642 /* Scan lists relative to size */
2643 break;
2644 case SCAN_FRACT:
2645 /*
2646 * Scan types proportional to swappiness and
2647 * their relative recent reclaim efficiency.
2648 * Make sure we don't miss the last page on
2649 * the offlined memory cgroups because of a
2650 * round-off error.
2651 */
2652 scan = mem_cgroup_online(memcg) ?
2653 div64_u64(dividend: scan * fraction[file], divisor: denominator) :
2654 DIV64_U64_ROUND_UP(scan * fraction[file],
2655 denominator);
2656 break;
2657 case SCAN_FILE:
2658 case SCAN_ANON:
2659 /* Scan one type exclusively */
2660 if ((scan_balance == SCAN_FILE) != file)
2661 scan = 0;
2662 break;
2663 default:
2664 /* Look ma, no brain */
2665 BUG();
2666 }
2667
2668 nr[lru] = scan;
2669 }
2670}
2671
2672/*
2673 * Anonymous LRU management is a waste if there is
2674 * ultimately no way to reclaim the memory.
2675 */
2676static bool can_age_anon_pages(struct lruvec *lruvec,
2677 struct scan_control *sc)
2678{
2679 /* Aging the anon LRU is valuable if swap is present: */
2680 if (total_swap_pages > 0)
2681 return true;
2682
2683 /* Also valuable if anon pages can be demoted: */
2684 return can_demote(nid: lruvec_pgdat(lruvec)->node_id, sc,
2685 memcg: lruvec_memcg(lruvec));
2686}
2687
2688#ifdef CONFIG_LRU_GEN
2689
2690#ifdef CONFIG_LRU_GEN_ENABLED
2691DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2692#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2693#else
2694DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2695#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2696#endif
2697
2698static bool should_walk_mmu(void)
2699{
2700 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2701}
2702
2703static bool should_clear_pmd_young(void)
2704{
2705 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2706}
2707
2708/******************************************************************************
2709 * shorthand helpers
2710 ******************************************************************************/
2711
2712#define DEFINE_MAX_SEQ(lruvec) \
2713 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2714
2715#define DEFINE_MIN_SEQ(lruvec) \
2716 unsigned long min_seq[ANON_AND_FILE] = { \
2717 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2718 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2719 }
2720
2721/* Get the min/max evictable type based on swappiness */
2722#define min_type(swappiness) (!(swappiness))
2723#define max_type(swappiness) ((swappiness) < SWAPPINESS_ANON_ONLY)
2724
2725#define evictable_min_seq(min_seq, swappiness) \
2726 min((min_seq)[min_type(swappiness)], (min_seq)[max_type(swappiness)])
2727
2728#define for_each_gen_type_zone(gen, type, zone) \
2729 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2730 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2731 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2732
2733#define for_each_evictable_type(type, swappiness) \
2734 for ((type) = min_type(swappiness); (type) <= max_type(swappiness); (type)++)
2735
2736#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2737#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2738
2739static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2740{
2741 struct pglist_data *pgdat = NODE_DATA(nid);
2742
2743#ifdef CONFIG_MEMCG
2744 if (memcg) {
2745 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2746
2747 /* see the comment in mem_cgroup_lruvec() */
2748 if (!lruvec->pgdat)
2749 lruvec->pgdat = pgdat;
2750
2751 return lruvec;
2752 }
2753#endif
2754 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2755
2756 return &pgdat->__lruvec;
2757}
2758
2759static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2760{
2761 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2762 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2763
2764 if (!sc->may_swap)
2765 return 0;
2766
2767 if (!can_demote(pgdat->node_id, sc, memcg) &&
2768 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2769 return 0;
2770
2771 return sc_swappiness(sc, memcg);
2772}
2773
2774static int get_nr_gens(struct lruvec *lruvec, int type)
2775{
2776 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2777}
2778
2779static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2780{
2781 int type;
2782
2783 for (type = 0; type < ANON_AND_FILE; type++) {
2784 int n = get_nr_gens(lruvec, type);
2785
2786 if (n < MIN_NR_GENS || n > MAX_NR_GENS)
2787 return false;
2788 }
2789
2790 return true;
2791}
2792
2793/******************************************************************************
2794 * Bloom filters
2795 ******************************************************************************/
2796
2797/*
2798 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2799 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2800 * bits in a bitmap, k is the number of hash functions and n is the number of
2801 * inserted items.
2802 *
2803 * Page table walkers use one of the two filters to reduce their search space.
2804 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2805 * aging uses the double-buffering technique to flip to the other filter each
2806 * time it produces a new generation. For non-leaf entries that have enough
2807 * leaf entries, the aging carries them over to the next generation in
2808 * walk_pmd_range(); the eviction also report them when walking the rmap
2809 * in lru_gen_look_around().
2810 *
2811 * For future optimizations:
2812 * 1. It's not necessary to keep both filters all the time. The spare one can be
2813 * freed after the RCU grace period and reallocated if needed again.
2814 * 2. And when reallocating, it's worth scaling its size according to the number
2815 * of inserted entries in the other filter, to reduce the memory overhead on
2816 * small systems and false positives on large systems.
2817 * 3. Jenkins' hash function is an alternative to Knuth's.
2818 */
2819#define BLOOM_FILTER_SHIFT 15
2820
2821static inline int filter_gen_from_seq(unsigned long seq)
2822{
2823 return seq % NR_BLOOM_FILTERS;
2824}
2825
2826static void get_item_key(void *item, int *key)
2827{
2828 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2829
2830 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2831
2832 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2833 key[1] = hash >> BLOOM_FILTER_SHIFT;
2834}
2835
2836static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2837 void *item)
2838{
2839 int key[2];
2840 unsigned long *filter;
2841 int gen = filter_gen_from_seq(seq);
2842
2843 filter = READ_ONCE(mm_state->filters[gen]);
2844 if (!filter)
2845 return true;
2846
2847 get_item_key(item, key);
2848
2849 return test_bit(key[0], filter) && test_bit(key[1], filter);
2850}
2851
2852static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2853 void *item)
2854{
2855 int key[2];
2856 unsigned long *filter;
2857 int gen = filter_gen_from_seq(seq);
2858
2859 filter = READ_ONCE(mm_state->filters[gen]);
2860 if (!filter)
2861 return;
2862
2863 get_item_key(item, key);
2864
2865 if (!test_bit(key[0], filter))
2866 set_bit(key[0], filter);
2867 if (!test_bit(key[1], filter))
2868 set_bit(key[1], filter);
2869}
2870
2871static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2872{
2873 unsigned long *filter;
2874 int gen = filter_gen_from_seq(seq);
2875
2876 filter = mm_state->filters[gen];
2877 if (filter) {
2878 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2879 return;
2880 }
2881
2882 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2883 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2884 WRITE_ONCE(mm_state->filters[gen], filter);
2885}
2886
2887/******************************************************************************
2888 * mm_struct list
2889 ******************************************************************************/
2890
2891#ifdef CONFIG_LRU_GEN_WALKS_MMU
2892
2893static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2894{
2895 static struct lru_gen_mm_list mm_list = {
2896 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2897 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2898 };
2899
2900#ifdef CONFIG_MEMCG
2901 if (memcg)
2902 return &memcg->mm_list;
2903#endif
2904 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2905
2906 return &mm_list;
2907}
2908
2909static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2910{
2911 return &lruvec->mm_state;
2912}
2913
2914static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2915{
2916 int key;
2917 struct mm_struct *mm;
2918 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2919 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2920
2921 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2922 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2923
2924 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2925 return NULL;
2926
2927 clear_bit(key, &mm->lru_gen.bitmap);
2928
2929 return mmget_not_zero(mm) ? mm : NULL;
2930}
2931
2932void lru_gen_add_mm(struct mm_struct *mm)
2933{
2934 int nid;
2935 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2936 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2937
2938 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2939#ifdef CONFIG_MEMCG
2940 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2941 mm->lru_gen.memcg = memcg;
2942#endif
2943 spin_lock(&mm_list->lock);
2944
2945 for_each_node_state(nid, N_MEMORY) {
2946 struct lruvec *lruvec = get_lruvec(memcg, nid);
2947 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2948
2949 /* the first addition since the last iteration */
2950 if (mm_state->tail == &mm_list->fifo)
2951 mm_state->tail = &mm->lru_gen.list;
2952 }
2953
2954 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2955
2956 spin_unlock(&mm_list->lock);
2957}
2958
2959void lru_gen_del_mm(struct mm_struct *mm)
2960{
2961 int nid;
2962 struct lru_gen_mm_list *mm_list;
2963 struct mem_cgroup *memcg = NULL;
2964
2965 if (list_empty(&mm->lru_gen.list))
2966 return;
2967
2968#ifdef CONFIG_MEMCG
2969 memcg = mm->lru_gen.memcg;
2970#endif
2971 mm_list = get_mm_list(memcg);
2972
2973 spin_lock(&mm_list->lock);
2974
2975 for_each_node(nid) {
2976 struct lruvec *lruvec = get_lruvec(memcg, nid);
2977 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2978
2979 /* where the current iteration continues after */
2980 if (mm_state->head == &mm->lru_gen.list)
2981 mm_state->head = mm_state->head->prev;
2982
2983 /* where the last iteration ended before */
2984 if (mm_state->tail == &mm->lru_gen.list)
2985 mm_state->tail = mm_state->tail->next;
2986 }
2987
2988 list_del_init(&mm->lru_gen.list);
2989
2990 spin_unlock(&mm_list->lock);
2991
2992#ifdef CONFIG_MEMCG
2993 mem_cgroup_put(mm->lru_gen.memcg);
2994 mm->lru_gen.memcg = NULL;
2995#endif
2996}
2997
2998#ifdef CONFIG_MEMCG
2999void lru_gen_migrate_mm(struct mm_struct *mm)
3000{
3001 struct mem_cgroup *memcg;
3002 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3003
3004 VM_WARN_ON_ONCE(task->mm != mm);
3005 lockdep_assert_held(&task->alloc_lock);
3006
3007 /* for mm_update_next_owner() */
3008 if (mem_cgroup_disabled())
3009 return;
3010
3011 /* migration can happen before addition */
3012 if (!mm->lru_gen.memcg)
3013 return;
3014
3015 rcu_read_lock();
3016 memcg = mem_cgroup_from_task(task);
3017 rcu_read_unlock();
3018 if (memcg == mm->lru_gen.memcg)
3019 return;
3020
3021 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3022
3023 lru_gen_del_mm(mm);
3024 lru_gen_add_mm(mm);
3025}
3026#endif
3027
3028#else /* !CONFIG_LRU_GEN_WALKS_MMU */
3029
3030static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3031{
3032 return NULL;
3033}
3034
3035static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
3036{
3037 return NULL;
3038}
3039
3040static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
3041{
3042 return NULL;
3043}
3044
3045#endif
3046
3047static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
3048{
3049 int i;
3050 int hist;
3051 struct lruvec *lruvec = walk->lruvec;
3052 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3053
3054 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3055
3056 hist = lru_hist_from_seq(walk->seq);
3057
3058 for (i = 0; i < NR_MM_STATS; i++) {
3059 WRITE_ONCE(mm_state->stats[hist][i],
3060 mm_state->stats[hist][i] + walk->mm_stats[i]);
3061 walk->mm_stats[i] = 0;
3062 }
3063
3064 if (NR_HIST_GENS > 1 && last) {
3065 hist = lru_hist_from_seq(walk->seq + 1);
3066
3067 for (i = 0; i < NR_MM_STATS; i++)
3068 WRITE_ONCE(mm_state->stats[hist][i], 0);
3069 }
3070}
3071
3072static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
3073{
3074 bool first = false;
3075 bool last = false;
3076 struct mm_struct *mm = NULL;
3077 struct lruvec *lruvec = walk->lruvec;
3078 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3079 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3080 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3081
3082 /*
3083 * mm_state->seq is incremented after each iteration of mm_list. There
3084 * are three interesting cases for this page table walker:
3085 * 1. It tries to start a new iteration with a stale max_seq: there is
3086 * nothing left to do.
3087 * 2. It started the next iteration: it needs to reset the Bloom filter
3088 * so that a fresh set of PTE tables can be recorded.
3089 * 3. It ended the current iteration: it needs to reset the mm stats
3090 * counters and tell its caller to increment max_seq.
3091 */
3092 spin_lock(&mm_list->lock);
3093
3094 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
3095
3096 if (walk->seq <= mm_state->seq)
3097 goto done;
3098
3099 if (!mm_state->head)
3100 mm_state->head = &mm_list->fifo;
3101
3102 if (mm_state->head == &mm_list->fifo)
3103 first = true;
3104
3105 do {
3106 mm_state->head = mm_state->head->next;
3107 if (mm_state->head == &mm_list->fifo) {
3108 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3109 last = true;
3110 break;
3111 }
3112
3113 /* force scan for those added after the last iteration */
3114 if (!mm_state->tail || mm_state->tail == mm_state->head) {
3115 mm_state->tail = mm_state->head->next;
3116 walk->force_scan = true;
3117 }
3118 } while (!(mm = get_next_mm(walk)));
3119done:
3120 if (*iter || last)
3121 reset_mm_stats(walk, last);
3122
3123 spin_unlock(&mm_list->lock);
3124
3125 if (mm && first)
3126 reset_bloom_filter(mm_state, walk->seq + 1);
3127
3128 if (*iter)
3129 mmput_async(*iter);
3130
3131 *iter = mm;
3132
3133 return last;
3134}
3135
3136static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
3137{
3138 bool success = false;
3139 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3140 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3141 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3142
3143 spin_lock(&mm_list->lock);
3144
3145 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
3146
3147 if (seq > mm_state->seq) {
3148 mm_state->head = NULL;
3149 mm_state->tail = NULL;
3150 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3151 success = true;
3152 }
3153
3154 spin_unlock(&mm_list->lock);
3155
3156 return success;
3157}
3158
3159/******************************************************************************
3160 * PID controller
3161 ******************************************************************************/
3162
3163/*
3164 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3165 *
3166 * The P term is refaulted/(evicted+protected) from a tier in the generation
3167 * currently being evicted; the I term is the exponential moving average of the
3168 * P term over the generations previously evicted, using the smoothing factor
3169 * 1/2; the D term isn't supported.
3170 *
3171 * The setpoint (SP) is always the first tier of one type; the process variable
3172 * (PV) is either any tier of the other type or any other tier of the same
3173 * type.
3174 *
3175 * The error is the difference between the SP and the PV; the correction is to
3176 * turn off protection when SP>PV or turn on protection when SP<PV.
3177 *
3178 * For future optimizations:
3179 * 1. The D term may discount the other two terms over time so that long-lived
3180 * generations can resist stale information.
3181 */
3182struct ctrl_pos {
3183 unsigned long refaulted;
3184 unsigned long total;
3185 int gain;
3186};
3187
3188static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3189 struct ctrl_pos *pos)
3190{
3191 int i;
3192 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3193 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3194
3195 pos->gain = gain;
3196 pos->refaulted = pos->total = 0;
3197
3198 for (i = tier % MAX_NR_TIERS; i <= min(tier, MAX_NR_TIERS - 1); i++) {
3199 pos->refaulted += lrugen->avg_refaulted[type][i] +
3200 atomic_long_read(&lrugen->refaulted[hist][type][i]);
3201 pos->total += lrugen->avg_total[type][i] +
3202 lrugen->protected[hist][type][i] +
3203 atomic_long_read(&lrugen->evicted[hist][type][i]);
3204 }
3205}
3206
3207static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3208{
3209 int hist, tier;
3210 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3211 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3212 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3213
3214 lockdep_assert_held(&lruvec->lru_lock);
3215
3216 if (!carryover && !clear)
3217 return;
3218
3219 hist = lru_hist_from_seq(seq);
3220
3221 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3222 if (carryover) {
3223 unsigned long sum;
3224
3225 sum = lrugen->avg_refaulted[type][tier] +
3226 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3227 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3228
3229 sum = lrugen->avg_total[type][tier] +
3230 lrugen->protected[hist][type][tier] +
3231 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3232 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3233 }
3234
3235 if (clear) {
3236 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3237 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3238 WRITE_ONCE(lrugen->protected[hist][type][tier], 0);
3239 }
3240 }
3241}
3242
3243static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3244{
3245 /*
3246 * Return true if the PV has a limited number of refaults or a lower
3247 * refaulted/total than the SP.
3248 */
3249 return pv->refaulted < MIN_LRU_BATCH ||
3250 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3251 (sp->refaulted + 1) * pv->total * pv->gain;
3252}
3253
3254/******************************************************************************
3255 * the aging
3256 ******************************************************************************/
3257
3258/* promote pages accessed through page tables */
3259static int folio_update_gen(struct folio *folio, int gen)
3260{
3261 unsigned long new_flags, old_flags = READ_ONCE(folio->flags.f);
3262
3263 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3264
3265 /* see the comment on LRU_REFS_FLAGS */
3266 if (!folio_test_referenced(folio) && !folio_test_workingset(folio)) {
3267 set_mask_bits(&folio->flags.f, LRU_REFS_MASK, BIT(PG_referenced));
3268 return -1;
3269 }
3270
3271 do {
3272 /* lru_gen_del_folio() has isolated this page? */
3273 if (!(old_flags & LRU_GEN_MASK))
3274 return -1;
3275
3276 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_FLAGS);
3277 new_flags |= ((gen + 1UL) << LRU_GEN_PGOFF) | BIT(PG_workingset);
3278 } while (!try_cmpxchg(&folio->flags.f, &old_flags, new_flags));
3279
3280 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3281}
3282
3283/* protect pages accessed multiple times through file descriptors */
3284static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3285{
3286 int type = folio_is_file_lru(folio);
3287 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3288 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3289 unsigned long new_flags, old_flags = READ_ONCE(folio->flags.f);
3290
3291 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3292
3293 do {
3294 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3295 /* folio_update_gen() has promoted this page? */
3296 if (new_gen >= 0 && new_gen != old_gen)
3297 return new_gen;
3298
3299 new_gen = (old_gen + 1) % MAX_NR_GENS;
3300
3301 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_FLAGS);
3302 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3303 /* for folio_end_writeback() */
3304 if (reclaiming)
3305 new_flags |= BIT(PG_reclaim);
3306 } while (!try_cmpxchg(&folio->flags.f, &old_flags, new_flags));
3307
3308 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3309
3310 return new_gen;
3311}
3312
3313static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3314 int old_gen, int new_gen)
3315{
3316 int type = folio_is_file_lru(folio);
3317 int zone = folio_zonenum(folio);
3318 int delta = folio_nr_pages(folio);
3319
3320 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3321 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3322
3323 walk->batched++;
3324
3325 walk->nr_pages[old_gen][type][zone] -= delta;
3326 walk->nr_pages[new_gen][type][zone] += delta;
3327}
3328
3329static void reset_batch_size(struct lru_gen_mm_walk *walk)
3330{
3331 int gen, type, zone;
3332 struct lruvec *lruvec = walk->lruvec;
3333 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3334
3335 walk->batched = 0;
3336
3337 for_each_gen_type_zone(gen, type, zone) {
3338 enum lru_list lru = type * LRU_INACTIVE_FILE;
3339 int delta = walk->nr_pages[gen][type][zone];
3340
3341 if (!delta)
3342 continue;
3343
3344 walk->nr_pages[gen][type][zone] = 0;
3345 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3346 lrugen->nr_pages[gen][type][zone] + delta);
3347
3348 if (lru_gen_is_active(lruvec, gen))
3349 lru += LRU_ACTIVE;
3350 __update_lru_size(lruvec, lru, zone, delta);
3351 }
3352}
3353
3354static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3355{
3356 struct address_space *mapping;
3357 struct vm_area_struct *vma = args->vma;
3358 struct lru_gen_mm_walk *walk = args->private;
3359
3360 if (!vma_is_accessible(vma))
3361 return true;
3362
3363 if (is_vm_hugetlb_page(vma))
3364 return true;
3365
3366 if (!vma_has_recency(vma))
3367 return true;
3368
3369 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3370 return true;
3371
3372 if (vma == get_gate_vma(vma->vm_mm))
3373 return true;
3374
3375 if (vma_is_anonymous(vma))
3376 return !walk->swappiness;
3377
3378 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3379 return true;
3380
3381 mapping = vma->vm_file->f_mapping;
3382 if (mapping_unevictable(mapping))
3383 return true;
3384
3385 if (shmem_mapping(mapping))
3386 return !walk->swappiness;
3387
3388 if (walk->swappiness > MAX_SWAPPINESS)
3389 return true;
3390
3391 /* to exclude special mappings like dax, etc. */
3392 return !mapping->a_ops->read_folio;
3393}
3394
3395/*
3396 * Some userspace memory allocators map many single-page VMAs. Instead of
3397 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3398 * table to reduce zigzags and improve cache performance.
3399 */
3400static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3401 unsigned long *vm_start, unsigned long *vm_end)
3402{
3403 unsigned long start = round_up(*vm_end, size);
3404 unsigned long end = (start | ~mask) + 1;
3405 VMA_ITERATOR(vmi, args->mm, start);
3406
3407 VM_WARN_ON_ONCE(mask & size);
3408 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3409
3410 for_each_vma(vmi, args->vma) {
3411 if (end && end <= args->vma->vm_start)
3412 return false;
3413
3414 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3415 continue;
3416
3417 *vm_start = max(start, args->vma->vm_start);
3418 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3419
3420 return true;
3421 }
3422
3423 return false;
3424}
3425
3426static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr,
3427 struct pglist_data *pgdat)
3428{
3429 unsigned long pfn = pte_pfn(pte);
3430
3431 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3432
3433 if (!pte_present(pte) || is_zero_pfn(pfn))
3434 return -1;
3435
3436 if (WARN_ON_ONCE(pte_special(pte)))
3437 return -1;
3438
3439 if (!pte_young(pte) && !mm_has_notifiers(vma->vm_mm))
3440 return -1;
3441
3442 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3443 return -1;
3444
3445 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3446 return -1;
3447
3448 return pfn;
3449}
3450
3451static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr,
3452 struct pglist_data *pgdat)
3453{
3454 unsigned long pfn = pmd_pfn(pmd);
3455
3456 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3457
3458 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3459 return -1;
3460
3461 if (!pmd_young(pmd) && !mm_has_notifiers(vma->vm_mm))
3462 return -1;
3463
3464 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3465 return -1;
3466
3467 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3468 return -1;
3469
3470 return pfn;
3471}
3472
3473static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3474 struct pglist_data *pgdat)
3475{
3476 struct folio *folio = pfn_folio(pfn);
3477
3478 if (folio_lru_gen(folio) < 0)
3479 return NULL;
3480
3481 if (folio_nid(folio) != pgdat->node_id)
3482 return NULL;
3483
3484 if (folio_memcg(folio) != memcg)
3485 return NULL;
3486
3487 return folio;
3488}
3489
3490static bool suitable_to_scan(int total, int young)
3491{
3492 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3493
3494 /* suitable if the average number of young PTEs per cacheline is >=1 */
3495 return young * n >= total;
3496}
3497
3498static void walk_update_folio(struct lru_gen_mm_walk *walk, struct folio *folio,
3499 int new_gen, bool dirty)
3500{
3501 int old_gen;
3502
3503 if (!folio)
3504 return;
3505
3506 if (dirty && !folio_test_dirty(folio) &&
3507 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3508 !folio_test_swapcache(folio)))
3509 folio_mark_dirty(folio);
3510
3511 if (walk) {
3512 old_gen = folio_update_gen(folio, new_gen);
3513 if (old_gen >= 0 && old_gen != new_gen)
3514 update_batch_size(walk, folio, old_gen, new_gen);
3515 } else if (lru_gen_set_refs(folio)) {
3516 old_gen = folio_lru_gen(folio);
3517 if (old_gen >= 0 && old_gen != new_gen)
3518 folio_activate(folio);
3519 }
3520}
3521
3522static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3523 struct mm_walk *args)
3524{
3525 int i;
3526 bool dirty;
3527 pte_t *pte;
3528 spinlock_t *ptl;
3529 unsigned long addr;
3530 int total = 0;
3531 int young = 0;
3532 struct folio *last = NULL;
3533 struct lru_gen_mm_walk *walk = args->private;
3534 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3535 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3536 DEFINE_MAX_SEQ(walk->lruvec);
3537 int gen = lru_gen_from_seq(max_seq);
3538 pmd_t pmdval;
3539
3540 pte = pte_offset_map_rw_nolock(args->mm, pmd, start & PMD_MASK, &pmdval, &ptl);
3541 if (!pte)
3542 return false;
3543
3544 if (!spin_trylock(ptl)) {
3545 pte_unmap(pte);
3546 return true;
3547 }
3548
3549 if (unlikely(!pmd_same(pmdval, pmdp_get_lockless(pmd)))) {
3550 pte_unmap_unlock(pte, ptl);
3551 return false;
3552 }
3553
3554 arch_enter_lazy_mmu_mode();
3555restart:
3556 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3557 unsigned long pfn;
3558 struct folio *folio;
3559 pte_t ptent = ptep_get(pte + i);
3560
3561 total++;
3562 walk->mm_stats[MM_LEAF_TOTAL]++;
3563
3564 pfn = get_pte_pfn(ptent, args->vma, addr, pgdat);
3565 if (pfn == -1)
3566 continue;
3567
3568 folio = get_pfn_folio(pfn, memcg, pgdat);
3569 if (!folio)
3570 continue;
3571
3572 if (!ptep_clear_young_notify(args->vma, addr, pte + i))
3573 continue;
3574
3575 if (last != folio) {
3576 walk_update_folio(walk, last, gen, dirty);
3577
3578 last = folio;
3579 dirty = false;
3580 }
3581
3582 if (pte_dirty(ptent))
3583 dirty = true;
3584
3585 young++;
3586 walk->mm_stats[MM_LEAF_YOUNG]++;
3587 }
3588
3589 walk_update_folio(walk, last, gen, dirty);
3590 last = NULL;
3591
3592 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3593 goto restart;
3594
3595 arch_leave_lazy_mmu_mode();
3596 pte_unmap_unlock(pte, ptl);
3597
3598 return suitable_to_scan(total, young);
3599}
3600
3601static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3602 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3603{
3604 int i;
3605 bool dirty;
3606 pmd_t *pmd;
3607 spinlock_t *ptl;
3608 struct folio *last = NULL;
3609 struct lru_gen_mm_walk *walk = args->private;
3610 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3611 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3612 DEFINE_MAX_SEQ(walk->lruvec);
3613 int gen = lru_gen_from_seq(max_seq);
3614
3615 VM_WARN_ON_ONCE(pud_leaf(*pud));
3616
3617 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3618 if (*first == -1) {
3619 *first = addr;
3620 bitmap_zero(bitmap, MIN_LRU_BATCH);
3621 return;
3622 }
3623
3624 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3625 if (i && i <= MIN_LRU_BATCH) {
3626 __set_bit(i - 1, bitmap);
3627 return;
3628 }
3629
3630 pmd = pmd_offset(pud, *first);
3631
3632 ptl = pmd_lockptr(args->mm, pmd);
3633 if (!spin_trylock(ptl))
3634 goto done;
3635
3636 arch_enter_lazy_mmu_mode();
3637
3638 do {
3639 unsigned long pfn;
3640 struct folio *folio;
3641
3642 /* don't round down the first address */
3643 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3644
3645 if (!pmd_present(pmd[i]))
3646 goto next;
3647
3648 if (!pmd_trans_huge(pmd[i])) {
3649 if (!walk->force_scan && should_clear_pmd_young() &&
3650 !mm_has_notifiers(args->mm))
3651 pmdp_test_and_clear_young(vma, addr, pmd + i);
3652 goto next;
3653 }
3654
3655 pfn = get_pmd_pfn(pmd[i], vma, addr, pgdat);
3656 if (pfn == -1)
3657 goto next;
3658
3659 folio = get_pfn_folio(pfn, memcg, pgdat);
3660 if (!folio)
3661 goto next;
3662
3663 if (!pmdp_clear_young_notify(vma, addr, pmd + i))
3664 goto next;
3665
3666 if (last != folio) {
3667 walk_update_folio(walk, last, gen, dirty);
3668
3669 last = folio;
3670 dirty = false;
3671 }
3672
3673 if (pmd_dirty(pmd[i]))
3674 dirty = true;
3675
3676 walk->mm_stats[MM_LEAF_YOUNG]++;
3677next:
3678 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3679 } while (i <= MIN_LRU_BATCH);
3680
3681 walk_update_folio(walk, last, gen, dirty);
3682
3683 arch_leave_lazy_mmu_mode();
3684 spin_unlock(ptl);
3685done:
3686 *first = -1;
3687}
3688
3689static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3690 struct mm_walk *args)
3691{
3692 int i;
3693 pmd_t *pmd;
3694 unsigned long next;
3695 unsigned long addr;
3696 struct vm_area_struct *vma;
3697 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3698 unsigned long first = -1;
3699 struct lru_gen_mm_walk *walk = args->private;
3700 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3701
3702 VM_WARN_ON_ONCE(pud_leaf(*pud));
3703
3704 /*
3705 * Finish an entire PMD in two passes: the first only reaches to PTE
3706 * tables to avoid taking the PMD lock; the second, if necessary, takes
3707 * the PMD lock to clear the accessed bit in PMD entries.
3708 */
3709 pmd = pmd_offset(pud, start & PUD_MASK);
3710restart:
3711 /* walk_pte_range() may call get_next_vma() */
3712 vma = args->vma;
3713 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3714 pmd_t val = pmdp_get_lockless(pmd + i);
3715
3716 next = pmd_addr_end(addr, end);
3717
3718 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3719 walk->mm_stats[MM_LEAF_TOTAL]++;
3720 continue;
3721 }
3722
3723 if (pmd_trans_huge(val)) {
3724 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3725 unsigned long pfn = get_pmd_pfn(val, vma, addr, pgdat);
3726
3727 walk->mm_stats[MM_LEAF_TOTAL]++;
3728
3729 if (pfn != -1)
3730 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3731 continue;
3732 }
3733
3734 if (!walk->force_scan && should_clear_pmd_young() &&
3735 !mm_has_notifiers(args->mm)) {
3736 if (!pmd_young(val))
3737 continue;
3738
3739 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3740 }
3741
3742 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3743 continue;
3744
3745 walk->mm_stats[MM_NONLEAF_FOUND]++;
3746
3747 if (!walk_pte_range(&val, addr, next, args))
3748 continue;
3749
3750 walk->mm_stats[MM_NONLEAF_ADDED]++;
3751
3752 /* carry over to the next generation */
3753 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3754 }
3755
3756 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3757
3758 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3759 goto restart;
3760}
3761
3762static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3763 struct mm_walk *args)
3764{
3765 int i;
3766 pud_t *pud;
3767 unsigned long addr;
3768 unsigned long next;
3769 struct lru_gen_mm_walk *walk = args->private;
3770
3771 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3772
3773 pud = pud_offset(p4d, start & P4D_MASK);
3774restart:
3775 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3776 pud_t val = READ_ONCE(pud[i]);
3777
3778 next = pud_addr_end(addr, end);
3779
3780 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3781 continue;
3782
3783 walk_pmd_range(&val, addr, next, args);
3784
3785 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3786 end = (addr | ~PUD_MASK) + 1;
3787 goto done;
3788 }
3789 }
3790
3791 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3792 goto restart;
3793
3794 end = round_up(end, P4D_SIZE);
3795done:
3796 if (!end || !args->vma)
3797 return 1;
3798
3799 walk->next_addr = max(end, args->vma->vm_start);
3800
3801 return -EAGAIN;
3802}
3803
3804static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3805{
3806 static const struct mm_walk_ops mm_walk_ops = {
3807 .test_walk = should_skip_vma,
3808 .p4d_entry = walk_pud_range,
3809 .walk_lock = PGWALK_RDLOCK,
3810 };
3811 int err;
3812 struct lruvec *lruvec = walk->lruvec;
3813
3814 walk->next_addr = FIRST_USER_ADDRESS;
3815
3816 do {
3817 DEFINE_MAX_SEQ(lruvec);
3818
3819 err = -EBUSY;
3820
3821 /* another thread might have called inc_max_seq() */
3822 if (walk->seq != max_seq)
3823 break;
3824
3825 /* the caller might be holding the lock for write */
3826 if (mmap_read_trylock(mm)) {
3827 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3828
3829 mmap_read_unlock(mm);
3830 }
3831
3832 if (walk->batched) {
3833 spin_lock_irq(&lruvec->lru_lock);
3834 reset_batch_size(walk);
3835 spin_unlock_irq(&lruvec->lru_lock);
3836 }
3837
3838 cond_resched();
3839 } while (err == -EAGAIN);
3840}
3841
3842static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3843{
3844 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3845
3846 if (pgdat && current_is_kswapd()) {
3847 VM_WARN_ON_ONCE(walk);
3848
3849 walk = &pgdat->mm_walk;
3850 } else if (!walk && force_alloc) {
3851 VM_WARN_ON_ONCE(current_is_kswapd());
3852
3853 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3854 }
3855
3856 current->reclaim_state->mm_walk = walk;
3857
3858 return walk;
3859}
3860
3861static void clear_mm_walk(void)
3862{
3863 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3864
3865 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3866 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3867
3868 current->reclaim_state->mm_walk = NULL;
3869
3870 if (!current_is_kswapd())
3871 kfree(walk);
3872}
3873
3874static bool inc_min_seq(struct lruvec *lruvec, int type, int swappiness)
3875{
3876 int zone;
3877 int remaining = MAX_LRU_BATCH;
3878 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3879 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3880 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3881
3882 /* For file type, skip the check if swappiness is anon only */
3883 if (type && (swappiness == SWAPPINESS_ANON_ONLY))
3884 goto done;
3885
3886 /* For anon type, skip the check if swappiness is zero (file only) */
3887 if (!type && !swappiness)
3888 goto done;
3889
3890 /* prevent cold/hot inversion if the type is evictable */
3891 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3892 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3893
3894 while (!list_empty(head)) {
3895 struct folio *folio = lru_to_folio(head);
3896 int refs = folio_lru_refs(folio);
3897 bool workingset = folio_test_workingset(folio);
3898
3899 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3900 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3901 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3902 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3903
3904 new_gen = folio_inc_gen(lruvec, folio, false);
3905 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3906
3907 /* don't count the workingset being lazily promoted */
3908 if (refs + workingset != BIT(LRU_REFS_WIDTH) + 1) {
3909 int tier = lru_tier_from_refs(refs, workingset);
3910 int delta = folio_nr_pages(folio);
3911
3912 WRITE_ONCE(lrugen->protected[hist][type][tier],
3913 lrugen->protected[hist][type][tier] + delta);
3914 }
3915
3916 if (!--remaining)
3917 return false;
3918 }
3919 }
3920done:
3921 reset_ctrl_pos(lruvec, type, true);
3922 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3923
3924 return true;
3925}
3926
3927static bool try_to_inc_min_seq(struct lruvec *lruvec, int swappiness)
3928{
3929 int gen, type, zone;
3930 bool success = false;
3931 bool seq_inc_flag = false;
3932 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3933 DEFINE_MIN_SEQ(lruvec);
3934
3935 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3936
3937 /* find the oldest populated generation */
3938 for_each_evictable_type(type, swappiness) {
3939 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3940 gen = lru_gen_from_seq(min_seq[type]);
3941
3942 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3943 if (!list_empty(&lrugen->folios[gen][type][zone]))
3944 goto next;
3945 }
3946
3947 min_seq[type]++;
3948 seq_inc_flag = true;
3949 }
3950next:
3951 ;
3952 }
3953
3954 /*
3955 * If min_seq[type] of both anonymous and file is not increased,
3956 * we can directly return false to avoid unnecessary checking
3957 * overhead later.
3958 */
3959 if (!seq_inc_flag)
3960 return success;
3961
3962 /* see the comment on lru_gen_folio */
3963 if (swappiness && swappiness <= MAX_SWAPPINESS) {
3964 unsigned long seq = lrugen->max_seq - MIN_NR_GENS;
3965
3966 if (min_seq[LRU_GEN_ANON] > seq && min_seq[LRU_GEN_FILE] < seq)
3967 min_seq[LRU_GEN_ANON] = seq;
3968 else if (min_seq[LRU_GEN_FILE] > seq && min_seq[LRU_GEN_ANON] < seq)
3969 min_seq[LRU_GEN_FILE] = seq;
3970 }
3971
3972 for_each_evictable_type(type, swappiness) {
3973 if (min_seq[type] <= lrugen->min_seq[type])
3974 continue;
3975
3976 reset_ctrl_pos(lruvec, type, true);
3977 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3978 success = true;
3979 }
3980
3981 return success;
3982}
3983
3984static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq, int swappiness)
3985{
3986 bool success;
3987 int prev, next;
3988 int type, zone;
3989 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3990restart:
3991 if (seq < READ_ONCE(lrugen->max_seq))
3992 return false;
3993
3994 spin_lock_irq(&lruvec->lru_lock);
3995
3996 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3997
3998 success = seq == lrugen->max_seq;
3999 if (!success)
4000 goto unlock;
4001
4002 for (type = 0; type < ANON_AND_FILE; type++) {
4003 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4004 continue;
4005
4006 if (inc_min_seq(lruvec, type, swappiness))
4007 continue;
4008
4009 spin_unlock_irq(&lruvec->lru_lock);
4010 cond_resched();
4011 goto restart;
4012 }
4013
4014 /*
4015 * Update the active/inactive LRU sizes for compatibility. Both sides of
4016 * the current max_seq need to be covered, since max_seq+1 can overlap
4017 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4018 * overlap, cold/hot inversion happens.
4019 */
4020 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4021 next = lru_gen_from_seq(lrugen->max_seq + 1);
4022
4023 for (type = 0; type < ANON_AND_FILE; type++) {
4024 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4025 enum lru_list lru = type * LRU_INACTIVE_FILE;
4026 long delta = lrugen->nr_pages[prev][type][zone] -
4027 lrugen->nr_pages[next][type][zone];
4028
4029 if (!delta)
4030 continue;
4031
4032 __update_lru_size(lruvec, lru, zone, delta);
4033 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4034 }
4035 }
4036
4037 for (type = 0; type < ANON_AND_FILE; type++)
4038 reset_ctrl_pos(lruvec, type, false);
4039
4040 WRITE_ONCE(lrugen->timestamps[next], jiffies);
4041 /* make sure preceding modifications appear */
4042 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4043unlock:
4044 spin_unlock_irq(&lruvec->lru_lock);
4045
4046 return success;
4047}
4048
4049static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
4050 int swappiness, bool force_scan)
4051{
4052 bool success;
4053 struct lru_gen_mm_walk *walk;
4054 struct mm_struct *mm = NULL;
4055 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4056 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4057
4058 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
4059
4060 if (!mm_state)
4061 return inc_max_seq(lruvec, seq, swappiness);
4062
4063 /* see the comment in iterate_mm_list() */
4064 if (seq <= READ_ONCE(mm_state->seq))
4065 return false;
4066
4067 /*
4068 * If the hardware doesn't automatically set the accessed bit, fallback
4069 * to lru_gen_look_around(), which only clears the accessed bit in a
4070 * handful of PTEs. Spreading the work out over a period of time usually
4071 * is less efficient, but it avoids bursty page faults.
4072 */
4073 if (!should_walk_mmu()) {
4074 success = iterate_mm_list_nowalk(lruvec, seq);
4075 goto done;
4076 }
4077
4078 walk = set_mm_walk(NULL, true);
4079 if (!walk) {
4080 success = iterate_mm_list_nowalk(lruvec, seq);
4081 goto done;
4082 }
4083
4084 walk->lruvec = lruvec;
4085 walk->seq = seq;
4086 walk->swappiness = swappiness;
4087 walk->force_scan = force_scan;
4088
4089 do {
4090 success = iterate_mm_list(walk, &mm);
4091 if (mm)
4092 walk_mm(mm, walk);
4093 } while (mm);
4094done:
4095 if (success) {
4096 success = inc_max_seq(lruvec, seq, swappiness);
4097 WARN_ON_ONCE(!success);
4098 }
4099
4100 return success;
4101}
4102
4103/******************************************************************************
4104 * working set protection
4105 ******************************************************************************/
4106
4107static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4108{
4109 int priority;
4110 unsigned long reclaimable;
4111
4112 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4113 return;
4114 /*
4115 * Determine the initial priority based on
4116 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4117 * where reclaimed_to_scanned_ratio = inactive / total.
4118 */
4119 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4120 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
4121 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4122
4123 /* round down reclaimable and round up sc->nr_to_reclaim */
4124 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4125
4126 /*
4127 * The estimation is based on LRU pages only, so cap it to prevent
4128 * overshoots of shrinker objects by large margins.
4129 */
4130 sc->priority = clamp(priority, DEF_PRIORITY / 2, DEF_PRIORITY);
4131}
4132
4133static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
4134{
4135 int gen, type, zone;
4136 unsigned long total = 0;
4137 int swappiness = get_swappiness(lruvec, sc);
4138 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4139 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4140 DEFINE_MAX_SEQ(lruvec);
4141 DEFINE_MIN_SEQ(lruvec);
4142
4143 for_each_evictable_type(type, swappiness) {
4144 unsigned long seq;
4145
4146 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4147 gen = lru_gen_from_seq(seq);
4148
4149 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4150 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4151 }
4152 }
4153
4154 /* whether the size is big enough to be helpful */
4155 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4156}
4157
4158static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
4159 unsigned long min_ttl)
4160{
4161 int gen;
4162 unsigned long birth;
4163 int swappiness = get_swappiness(lruvec, sc);
4164 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4165 DEFINE_MIN_SEQ(lruvec);
4166
4167 if (mem_cgroup_below_min(NULL, memcg))
4168 return false;
4169
4170 if (!lruvec_is_sizable(lruvec, sc))
4171 return false;
4172
4173 gen = lru_gen_from_seq(evictable_min_seq(min_seq, swappiness));
4174 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4175
4176 return time_is_before_jiffies(birth + min_ttl);
4177}
4178
4179/* to protect the working set of the last N jiffies */
4180static unsigned long lru_gen_min_ttl __read_mostly;
4181
4182static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4183{
4184 struct mem_cgroup *memcg;
4185 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4186 bool reclaimable = !min_ttl;
4187
4188 VM_WARN_ON_ONCE(!current_is_kswapd());
4189
4190 set_initial_priority(pgdat, sc);
4191
4192 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4193 do {
4194 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4195
4196 mem_cgroup_calculate_protection(NULL, memcg);
4197
4198 if (!reclaimable)
4199 reclaimable = lruvec_is_reclaimable(lruvec, sc, min_ttl);
4200 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4201
4202 /*
4203 * The main goal is to OOM kill if every generation from all memcgs is
4204 * younger than min_ttl. However, another possibility is all memcgs are
4205 * either too small or below min.
4206 */
4207 if (!reclaimable && mutex_trylock(&oom_lock)) {
4208 struct oom_control oc = {
4209 .gfp_mask = sc->gfp_mask,
4210 };
4211
4212 out_of_memory(&oc);
4213
4214 mutex_unlock(&oom_lock);
4215 }
4216}
4217
4218/******************************************************************************
4219 * rmap/PT walk feedback
4220 ******************************************************************************/
4221
4222/*
4223 * This function exploits spatial locality when shrink_folio_list() walks the
4224 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4225 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4226 * the PTE table to the Bloom filter. This forms a feedback loop between the
4227 * eviction and the aging.
4228 */
4229bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4230{
4231 int i;
4232 bool dirty;
4233 unsigned long start;
4234 unsigned long end;
4235 struct lru_gen_mm_walk *walk;
4236 struct folio *last = NULL;
4237 int young = 1;
4238 pte_t *pte = pvmw->pte;
4239 unsigned long addr = pvmw->address;
4240 struct vm_area_struct *vma = pvmw->vma;
4241 struct folio *folio = pfn_folio(pvmw->pfn);
4242 struct mem_cgroup *memcg = folio_memcg(folio);
4243 struct pglist_data *pgdat = folio_pgdat(folio);
4244 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4245 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4246 DEFINE_MAX_SEQ(lruvec);
4247 int gen = lru_gen_from_seq(max_seq);
4248
4249 lockdep_assert_held(pvmw->ptl);
4250 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4251
4252 if (!ptep_clear_young_notify(vma, addr, pte))
4253 return false;
4254
4255 if (spin_is_contended(pvmw->ptl))
4256 return true;
4257
4258 /* exclude special VMAs containing anon pages from COW */
4259 if (vma->vm_flags & VM_SPECIAL)
4260 return true;
4261
4262 /* avoid taking the LRU lock under the PTL when possible */
4263 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4264
4265 start = max(addr & PMD_MASK, vma->vm_start);
4266 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4267
4268 if (end - start == PAGE_SIZE)
4269 return true;
4270
4271 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4272 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4273 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4274 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4275 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4276 else {
4277 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4278 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4279 }
4280 }
4281
4282 arch_enter_lazy_mmu_mode();
4283
4284 pte -= (addr - start) / PAGE_SIZE;
4285
4286 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4287 unsigned long pfn;
4288 pte_t ptent = ptep_get(pte + i);
4289
4290 pfn = get_pte_pfn(ptent, vma, addr, pgdat);
4291 if (pfn == -1)
4292 continue;
4293
4294 folio = get_pfn_folio(pfn, memcg, pgdat);
4295 if (!folio)
4296 continue;
4297
4298 if (!ptep_clear_young_notify(vma, addr, pte + i))
4299 continue;
4300
4301 if (last != folio) {
4302 walk_update_folio(walk, last, gen, dirty);
4303
4304 last = folio;
4305 dirty = false;
4306 }
4307
4308 if (pte_dirty(ptent))
4309 dirty = true;
4310
4311 young++;
4312 }
4313
4314 walk_update_folio(walk, last, gen, dirty);
4315
4316 arch_leave_lazy_mmu_mode();
4317
4318 /* feedback from rmap walkers to page table walkers */
4319 if (mm_state && suitable_to_scan(i, young))
4320 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4321
4322 return true;
4323}
4324
4325/******************************************************************************
4326 * memcg LRU
4327 ******************************************************************************/
4328
4329/* see the comment on MEMCG_NR_GENS */
4330enum {
4331 MEMCG_LRU_NOP,
4332 MEMCG_LRU_HEAD,
4333 MEMCG_LRU_TAIL,
4334 MEMCG_LRU_OLD,
4335 MEMCG_LRU_YOUNG,
4336};
4337
4338static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4339{
4340 int seg;
4341 int old, new;
4342 unsigned long flags;
4343 int bin = get_random_u32_below(MEMCG_NR_BINS);
4344 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4345
4346 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4347
4348 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4349
4350 seg = 0;
4351 new = old = lruvec->lrugen.gen;
4352
4353 /* see the comment on MEMCG_NR_GENS */
4354 if (op == MEMCG_LRU_HEAD)
4355 seg = MEMCG_LRU_HEAD;
4356 else if (op == MEMCG_LRU_TAIL)
4357 seg = MEMCG_LRU_TAIL;
4358 else if (op == MEMCG_LRU_OLD)
4359 new = get_memcg_gen(pgdat->memcg_lru.seq);
4360 else if (op == MEMCG_LRU_YOUNG)
4361 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4362 else
4363 VM_WARN_ON_ONCE(true);
4364
4365 WRITE_ONCE(lruvec->lrugen.seg, seg);
4366 WRITE_ONCE(lruvec->lrugen.gen, new);
4367
4368 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4369
4370 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4371 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4372 else
4373 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4374
4375 pgdat->memcg_lru.nr_memcgs[old]--;
4376 pgdat->memcg_lru.nr_memcgs[new]++;
4377
4378 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4379 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4380
4381 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4382}
4383
4384#ifdef CONFIG_MEMCG
4385
4386void lru_gen_online_memcg(struct mem_cgroup *memcg)
4387{
4388 int gen;
4389 int nid;
4390 int bin = get_random_u32_below(MEMCG_NR_BINS);
4391
4392 for_each_node(nid) {
4393 struct pglist_data *pgdat = NODE_DATA(nid);
4394 struct lruvec *lruvec = get_lruvec(memcg, nid);
4395
4396 spin_lock_irq(&pgdat->memcg_lru.lock);
4397
4398 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4399
4400 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4401
4402 lruvec->lrugen.gen = gen;
4403
4404 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4405 pgdat->memcg_lru.nr_memcgs[gen]++;
4406
4407 spin_unlock_irq(&pgdat->memcg_lru.lock);
4408 }
4409}
4410
4411void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4412{
4413 int nid;
4414
4415 for_each_node(nid) {
4416 struct lruvec *lruvec = get_lruvec(memcg, nid);
4417
4418 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4419 }
4420}
4421
4422void lru_gen_release_memcg(struct mem_cgroup *memcg)
4423{
4424 int gen;
4425 int nid;
4426
4427 for_each_node(nid) {
4428 struct pglist_data *pgdat = NODE_DATA(nid);
4429 struct lruvec *lruvec = get_lruvec(memcg, nid);
4430
4431 spin_lock_irq(&pgdat->memcg_lru.lock);
4432
4433 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4434 goto unlock;
4435
4436 gen = lruvec->lrugen.gen;
4437
4438 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4439 pgdat->memcg_lru.nr_memcgs[gen]--;
4440
4441 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4442 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4443unlock:
4444 spin_unlock_irq(&pgdat->memcg_lru.lock);
4445 }
4446}
4447
4448void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4449{
4450 struct lruvec *lruvec = get_lruvec(memcg, nid);
4451
4452 /* see the comment on MEMCG_NR_GENS */
4453 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4454 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4455}
4456
4457#endif /* CONFIG_MEMCG */
4458
4459/******************************************************************************
4460 * the eviction
4461 ******************************************************************************/
4462
4463static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4464 int tier_idx)
4465{
4466 bool success;
4467 bool dirty, writeback;
4468 int gen = folio_lru_gen(folio);
4469 int type = folio_is_file_lru(folio);
4470 int zone = folio_zonenum(folio);
4471 int delta = folio_nr_pages(folio);
4472 int refs = folio_lru_refs(folio);
4473 bool workingset = folio_test_workingset(folio);
4474 int tier = lru_tier_from_refs(refs, workingset);
4475 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4476
4477 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4478
4479 /* unevictable */
4480 if (!folio_evictable(folio)) {
4481 success = lru_gen_del_folio(lruvec, folio, true);
4482 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4483 folio_set_unevictable(folio);
4484 lruvec_add_folio(lruvec, folio);
4485 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4486 return true;
4487 }
4488
4489 /* promoted */
4490 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4491 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4492 return true;
4493 }
4494
4495 /* protected */
4496 if (tier > tier_idx || refs + workingset == BIT(LRU_REFS_WIDTH) + 1) {
4497 gen = folio_inc_gen(lruvec, folio, false);
4498 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4499
4500 /* don't count the workingset being lazily promoted */
4501 if (refs + workingset != BIT(LRU_REFS_WIDTH) + 1) {
4502 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4503
4504 WRITE_ONCE(lrugen->protected[hist][type][tier],
4505 lrugen->protected[hist][type][tier] + delta);
4506 }
4507 return true;
4508 }
4509
4510 /* ineligible */
4511 if (zone > sc->reclaim_idx) {
4512 gen = folio_inc_gen(lruvec, folio, false);
4513 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4514 return true;
4515 }
4516
4517 dirty = folio_test_dirty(folio);
4518 writeback = folio_test_writeback(folio);
4519 if (type == LRU_GEN_FILE && dirty) {
4520 sc->nr.file_taken += delta;
4521 if (!writeback)
4522 sc->nr.unqueued_dirty += delta;
4523 }
4524
4525 /* waiting for writeback */
4526 if (writeback || (type == LRU_GEN_FILE && dirty)) {
4527 gen = folio_inc_gen(lruvec, folio, true);
4528 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4529 return true;
4530 }
4531
4532 return false;
4533}
4534
4535static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4536{
4537 bool success;
4538
4539 /* swap constrained */
4540 if (!(sc->gfp_mask & __GFP_IO) &&
4541 (folio_test_dirty(folio) ||
4542 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4543 return false;
4544
4545 /* raced with release_pages() */
4546 if (!folio_try_get(folio))
4547 return false;
4548
4549 /* raced with another isolation */
4550 if (!folio_test_clear_lru(folio)) {
4551 folio_put(folio);
4552 return false;
4553 }
4554
4555 /* see the comment on LRU_REFS_FLAGS */
4556 if (!folio_test_referenced(folio))
4557 set_mask_bits(&folio->flags.f, LRU_REFS_MASK, 0);
4558
4559 /* for shrink_folio_list() */
4560 folio_clear_reclaim(folio);
4561
4562 success = lru_gen_del_folio(lruvec, folio, true);
4563 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4564
4565 return true;
4566}
4567
4568static int scan_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
4569 struct scan_control *sc, int type, int tier,
4570 struct list_head *list)
4571{
4572 int i;
4573 int gen;
4574 enum vm_event_item item;
4575 int sorted = 0;
4576 int scanned = 0;
4577 int isolated = 0;
4578 int skipped = 0;
4579 int remaining = min(nr_to_scan, MAX_LRU_BATCH);
4580 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4581 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4582
4583 VM_WARN_ON_ONCE(!list_empty(list));
4584
4585 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4586 return 0;
4587
4588 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4589
4590 for (i = MAX_NR_ZONES; i > 0; i--) {
4591 LIST_HEAD(moved);
4592 int skipped_zone = 0;
4593 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4594 struct list_head *head = &lrugen->folios[gen][type][zone];
4595
4596 while (!list_empty(head)) {
4597 struct folio *folio = lru_to_folio(head);
4598 int delta = folio_nr_pages(folio);
4599
4600 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4601 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4602 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4603 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4604
4605 scanned += delta;
4606
4607 if (sort_folio(lruvec, folio, sc, tier))
4608 sorted += delta;
4609 else if (isolate_folio(lruvec, folio, sc)) {
4610 list_add(&folio->lru, list);
4611 isolated += delta;
4612 } else {
4613 list_move(&folio->lru, &moved);
4614 skipped_zone += delta;
4615 }
4616
4617 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4618 break;
4619 }
4620
4621 if (skipped_zone) {
4622 list_splice(&moved, head);
4623 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4624 skipped += skipped_zone;
4625 }
4626
4627 if (!remaining || isolated >= MIN_LRU_BATCH)
4628 break;
4629 }
4630
4631 item = PGSCAN_KSWAPD + reclaimer_offset(sc);
4632 if (!cgroup_reclaim(sc)) {
4633 __count_vm_events(item, isolated);
4634 __count_vm_events(PGREFILL, sorted);
4635 }
4636 count_memcg_events(memcg, item, isolated);
4637 count_memcg_events(memcg, PGREFILL, sorted);
4638 __count_vm_events(PGSCAN_ANON + type, isolated);
4639 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4640 scanned, skipped, isolated,
4641 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4642 if (type == LRU_GEN_FILE)
4643 sc->nr.file_taken += isolated;
4644 /*
4645 * There might not be eligible folios due to reclaim_idx. Check the
4646 * remaining to prevent livelock if it's not making progress.
4647 */
4648 return isolated || !remaining ? scanned : 0;
4649}
4650
4651static int get_tier_idx(struct lruvec *lruvec, int type)
4652{
4653 int tier;
4654 struct ctrl_pos sp, pv;
4655
4656 /*
4657 * To leave a margin for fluctuations, use a larger gain factor (2:3).
4658 * This value is chosen because any other tier would have at least twice
4659 * as many refaults as the first tier.
4660 */
4661 read_ctrl_pos(lruvec, type, 0, 2, &sp);
4662 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4663 read_ctrl_pos(lruvec, type, tier, 3, &pv);
4664 if (!positive_ctrl_err(&sp, &pv))
4665 break;
4666 }
4667
4668 return tier - 1;
4669}
4670
4671static int get_type_to_scan(struct lruvec *lruvec, int swappiness)
4672{
4673 struct ctrl_pos sp, pv;
4674
4675 if (swappiness <= MIN_SWAPPINESS + 1)
4676 return LRU_GEN_FILE;
4677
4678 if (swappiness >= MAX_SWAPPINESS)
4679 return LRU_GEN_ANON;
4680 /*
4681 * Compare the sum of all tiers of anon with that of file to determine
4682 * which type to scan.
4683 */
4684 read_ctrl_pos(lruvec, LRU_GEN_ANON, MAX_NR_TIERS, swappiness, &sp);
4685 read_ctrl_pos(lruvec, LRU_GEN_FILE, MAX_NR_TIERS, MAX_SWAPPINESS - swappiness, &pv);
4686
4687 return positive_ctrl_err(&sp, &pv);
4688}
4689
4690static int isolate_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
4691 struct scan_control *sc, int swappiness,
4692 int *type_scanned, struct list_head *list)
4693{
4694 int i;
4695 int type = get_type_to_scan(lruvec, swappiness);
4696
4697 for_each_evictable_type(i, swappiness) {
4698 int scanned;
4699 int tier = get_tier_idx(lruvec, type);
4700
4701 *type_scanned = type;
4702
4703 scanned = scan_folios(nr_to_scan, lruvec, sc, type, tier, list);
4704 if (scanned)
4705 return scanned;
4706
4707 type = !type;
4708 }
4709
4710 return 0;
4711}
4712
4713static int evict_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
4714 struct scan_control *sc, int swappiness)
4715{
4716 int type;
4717 int scanned;
4718 int reclaimed;
4719 LIST_HEAD(list);
4720 LIST_HEAD(clean);
4721 struct folio *folio;
4722 struct folio *next;
4723 enum vm_event_item item;
4724 struct reclaim_stat stat;
4725 struct lru_gen_mm_walk *walk;
4726 bool skip_retry = false;
4727 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4728 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4729 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4730
4731 spin_lock_irq(&lruvec->lru_lock);
4732
4733 scanned = isolate_folios(nr_to_scan, lruvec, sc, swappiness, &type, &list);
4734
4735 scanned += try_to_inc_min_seq(lruvec, swappiness);
4736
4737 if (evictable_min_seq(lrugen->min_seq, swappiness) + MIN_NR_GENS > lrugen->max_seq)
4738 scanned = 0;
4739
4740 spin_unlock_irq(&lruvec->lru_lock);
4741
4742 if (list_empty(&list))
4743 return scanned;
4744retry:
4745 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false, memcg);
4746 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
4747 sc->nr_reclaimed += reclaimed;
4748 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4749 scanned, reclaimed, &stat, sc->priority,
4750 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4751
4752 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4753 DEFINE_MIN_SEQ(lruvec);
4754
4755 if (!folio_evictable(folio)) {
4756 list_del(&folio->lru);
4757 folio_putback_lru(folio);
4758 continue;
4759 }
4760
4761 /* retry folios that may have missed folio_rotate_reclaimable() */
4762 if (!skip_retry && !folio_test_active(folio) && !folio_mapped(folio) &&
4763 !folio_test_dirty(folio) && !folio_test_writeback(folio)) {
4764 list_move(&folio->lru, &clean);
4765 continue;
4766 }
4767
4768 /* don't add rejected folios to the oldest generation */
4769 if (lru_gen_folio_seq(lruvec, folio, false) == min_seq[type])
4770 set_mask_bits(&folio->flags.f, LRU_REFS_FLAGS, BIT(PG_active));
4771 }
4772
4773 spin_lock_irq(&lruvec->lru_lock);
4774
4775 move_folios_to_lru(lruvec, &list);
4776
4777 walk = current->reclaim_state->mm_walk;
4778 if (walk && walk->batched) {
4779 walk->lruvec = lruvec;
4780 reset_batch_size(walk);
4781 }
4782
4783 __mod_lruvec_state(lruvec, PGDEMOTE_KSWAPD + reclaimer_offset(sc),
4784 stat.nr_demoted);
4785
4786 item = PGSTEAL_KSWAPD + reclaimer_offset(sc);
4787 if (!cgroup_reclaim(sc))
4788 __count_vm_events(item, reclaimed);
4789 count_memcg_events(memcg, item, reclaimed);
4790 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4791
4792 spin_unlock_irq(&lruvec->lru_lock);
4793
4794 list_splice_init(&clean, &list);
4795
4796 if (!list_empty(&list)) {
4797 skip_retry = true;
4798 goto retry;
4799 }
4800
4801 return scanned;
4802}
4803
4804static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4805 int swappiness, unsigned long *nr_to_scan)
4806{
4807 int gen, type, zone;
4808 unsigned long size = 0;
4809 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4810 DEFINE_MIN_SEQ(lruvec);
4811
4812 *nr_to_scan = 0;
4813 /* have to run aging, since eviction is not possible anymore */
4814 if (evictable_min_seq(min_seq, swappiness) + MIN_NR_GENS > max_seq)
4815 return true;
4816
4817 for_each_evictable_type(type, swappiness) {
4818 unsigned long seq;
4819
4820 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4821 gen = lru_gen_from_seq(seq);
4822
4823 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4824 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4825 }
4826 }
4827
4828 *nr_to_scan = size;
4829 /* better to run aging even though eviction is still possible */
4830 return evictable_min_seq(min_seq, swappiness) + MIN_NR_GENS == max_seq;
4831}
4832
4833/*
4834 * For future optimizations:
4835 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4836 * reclaim.
4837 */
4838static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4839{
4840 bool success;
4841 unsigned long nr_to_scan;
4842 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4843 DEFINE_MAX_SEQ(lruvec);
4844
4845 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4846 return -1;
4847
4848 success = should_run_aging(lruvec, max_seq, swappiness, &nr_to_scan);
4849
4850 /* try to scrape all its memory if this memcg was deleted */
4851 if (nr_to_scan && !mem_cgroup_online(memcg))
4852 return nr_to_scan;
4853
4854 nr_to_scan = apply_proportional_protection(memcg, sc, nr_to_scan);
4855
4856 /* try to get away with not aging at the default priority */
4857 if (!success || sc->priority == DEF_PRIORITY)
4858 return nr_to_scan >> sc->priority;
4859
4860 /* stop scanning this lruvec as it's low on cold folios */
4861 return try_to_inc_max_seq(lruvec, max_seq, swappiness, false) ? -1 : 0;
4862}
4863
4864static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4865{
4866 int i;
4867 enum zone_watermarks mark;
4868
4869 /* don't abort memcg reclaim to ensure fairness */
4870 if (!root_reclaim(sc))
4871 return false;
4872
4873 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4874 return true;
4875
4876 /* check the order to exclude compaction-induced reclaim */
4877 if (!current_is_kswapd() || sc->order)
4878 return false;
4879
4880 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4881 WMARK_PROMO : WMARK_HIGH;
4882
4883 for (i = 0; i <= sc->reclaim_idx; i++) {
4884 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4885 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4886
4887 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4888 return false;
4889 }
4890
4891 /* kswapd should abort if all eligible zones are safe */
4892 return true;
4893}
4894
4895static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4896{
4897 long nr_to_scan;
4898 unsigned long scanned = 0;
4899 int swappiness = get_swappiness(lruvec, sc);
4900
4901 while (true) {
4902 int delta;
4903
4904 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4905 if (nr_to_scan <= 0)
4906 break;
4907
4908 delta = evict_folios(nr_to_scan, lruvec, sc, swappiness);
4909 if (!delta)
4910 break;
4911
4912 scanned += delta;
4913 if (scanned >= nr_to_scan)
4914 break;
4915
4916 if (should_abort_scan(lruvec, sc))
4917 break;
4918
4919 cond_resched();
4920 }
4921
4922 /*
4923 * If too many file cache in the coldest generation can't be evicted
4924 * due to being dirty, wake up the flusher.
4925 */
4926 if (sc->nr.unqueued_dirty && sc->nr.unqueued_dirty == sc->nr.file_taken)
4927 wakeup_flusher_threads(WB_REASON_VMSCAN);
4928
4929 /* whether this lruvec should be rotated */
4930 return nr_to_scan < 0;
4931}
4932
4933static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4934{
4935 bool success;
4936 unsigned long scanned = sc->nr_scanned;
4937 unsigned long reclaimed = sc->nr_reclaimed;
4938 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4939 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4940
4941 /* lru_gen_age_node() called mem_cgroup_calculate_protection() */
4942 if (mem_cgroup_below_min(NULL, memcg))
4943 return MEMCG_LRU_YOUNG;
4944
4945 if (mem_cgroup_below_low(NULL, memcg)) {
4946 /* see the comment on MEMCG_NR_GENS */
4947 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4948 return MEMCG_LRU_TAIL;
4949
4950 memcg_memory_event(memcg, MEMCG_LOW);
4951 }
4952
4953 success = try_to_shrink_lruvec(lruvec, sc);
4954
4955 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4956
4957 if (!sc->proactive)
4958 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4959 sc->nr_reclaimed - reclaimed);
4960
4961 flush_reclaim_state(sc);
4962
4963 if (success && mem_cgroup_online(memcg))
4964 return MEMCG_LRU_YOUNG;
4965
4966 if (!success && lruvec_is_sizable(lruvec, sc))
4967 return 0;
4968
4969 /* one retry if offlined or too small */
4970 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4971 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4972}
4973
4974static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4975{
4976 int op;
4977 int gen;
4978 int bin;
4979 int first_bin;
4980 struct lruvec *lruvec;
4981 struct lru_gen_folio *lrugen;
4982 struct mem_cgroup *memcg;
4983 struct hlist_nulls_node *pos;
4984
4985 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4986 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4987restart:
4988 op = 0;
4989 memcg = NULL;
4990
4991 rcu_read_lock();
4992
4993 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4994 if (op) {
4995 lru_gen_rotate_memcg(lruvec, op);
4996 op = 0;
4997 }
4998
4999 mem_cgroup_put(memcg);
5000 memcg = NULL;
5001
5002 if (gen != READ_ONCE(lrugen->gen))
5003 continue;
5004
5005 lruvec = container_of(lrugen, struct lruvec, lrugen);
5006 memcg = lruvec_memcg(lruvec);
5007
5008 if (!mem_cgroup_tryget(memcg)) {
5009 lru_gen_release_memcg(memcg);
5010 memcg = NULL;
5011 continue;
5012 }
5013
5014 rcu_read_unlock();
5015
5016 op = shrink_one(lruvec, sc);
5017
5018 rcu_read_lock();
5019
5020 if (should_abort_scan(lruvec, sc))
5021 break;
5022 }
5023
5024 rcu_read_unlock();
5025
5026 if (op)
5027 lru_gen_rotate_memcg(lruvec, op);
5028
5029 mem_cgroup_put(memcg);
5030
5031 if (!is_a_nulls(pos))
5032 return;
5033
5034 /* restart if raced with lru_gen_rotate_memcg() */
5035 if (gen != get_nulls_value(pos))
5036 goto restart;
5037
5038 /* try the rest of the bins of the current generation */
5039 bin = get_memcg_bin(bin + 1);
5040 if (bin != first_bin)
5041 goto restart;
5042}
5043
5044static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5045{
5046 struct blk_plug plug;
5047
5048 VM_WARN_ON_ONCE(root_reclaim(sc));
5049 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
5050
5051 lru_add_drain();
5052
5053 blk_start_plug(&plug);
5054
5055 set_mm_walk(NULL, sc->proactive);
5056
5057 if (try_to_shrink_lruvec(lruvec, sc))
5058 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
5059
5060 clear_mm_walk();
5061
5062 blk_finish_plug(&plug);
5063}
5064
5065static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5066{
5067 struct blk_plug plug;
5068 unsigned long reclaimed = sc->nr_reclaimed;
5069
5070 VM_WARN_ON_ONCE(!root_reclaim(sc));
5071
5072 /*
5073 * Unmapped clean folios are already prioritized. Scanning for more of
5074 * them is likely futile and can cause high reclaim latency when there
5075 * is a large number of memcgs.
5076 */
5077 if (!sc->may_writepage || !sc->may_unmap)
5078 goto done;
5079
5080 lru_add_drain();
5081
5082 blk_start_plug(&plug);
5083
5084 set_mm_walk(pgdat, sc->proactive);
5085
5086 set_initial_priority(pgdat, sc);
5087
5088 if (current_is_kswapd())
5089 sc->nr_reclaimed = 0;
5090
5091 if (mem_cgroup_disabled())
5092 shrink_one(&pgdat->__lruvec, sc);
5093 else
5094 shrink_many(pgdat, sc);
5095
5096 if (current_is_kswapd())
5097 sc->nr_reclaimed += reclaimed;
5098
5099 clear_mm_walk();
5100
5101 blk_finish_plug(&plug);
5102done:
5103 if (sc->nr_reclaimed > reclaimed)
5104 atomic_set(&pgdat->kswapd_failures, 0);
5105}
5106
5107/******************************************************************************
5108 * state change
5109 ******************************************************************************/
5110
5111static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5112{
5113 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5114
5115 if (lrugen->enabled) {
5116 enum lru_list lru;
5117
5118 for_each_evictable_lru(lru) {
5119 if (!list_empty(&lruvec->lists[lru]))
5120 return false;
5121 }
5122 } else {
5123 int gen, type, zone;
5124
5125 for_each_gen_type_zone(gen, type, zone) {
5126 if (!list_empty(&lrugen->folios[gen][type][zone]))
5127 return false;
5128 }
5129 }
5130
5131 return true;
5132}
5133
5134static bool fill_evictable(struct lruvec *lruvec)
5135{
5136 enum lru_list lru;
5137 int remaining = MAX_LRU_BATCH;
5138
5139 for_each_evictable_lru(lru) {
5140 int type = is_file_lru(lru);
5141 bool active = is_active_lru(lru);
5142 struct list_head *head = &lruvec->lists[lru];
5143
5144 while (!list_empty(head)) {
5145 bool success;
5146 struct folio *folio = lru_to_folio(head);
5147
5148 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5149 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5150 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5151 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5152
5153 lruvec_del_folio(lruvec, folio);
5154 success = lru_gen_add_folio(lruvec, folio, false);
5155 VM_WARN_ON_ONCE(!success);
5156
5157 if (!--remaining)
5158 return false;
5159 }
5160 }
5161
5162 return true;
5163}
5164
5165static bool drain_evictable(struct lruvec *lruvec)
5166{
5167 int gen, type, zone;
5168 int remaining = MAX_LRU_BATCH;
5169
5170 for_each_gen_type_zone(gen, type, zone) {
5171 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5172
5173 while (!list_empty(head)) {
5174 bool success;
5175 struct folio *folio = lru_to_folio(head);
5176
5177 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5178 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5179 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5180 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5181
5182 success = lru_gen_del_folio(lruvec, folio, false);
5183 VM_WARN_ON_ONCE(!success);
5184 lruvec_add_folio(lruvec, folio);
5185
5186 if (!--remaining)
5187 return false;
5188 }
5189 }
5190
5191 return true;
5192}
5193
5194static void lru_gen_change_state(bool enabled)
5195{
5196 static DEFINE_MUTEX(state_mutex);
5197
5198 struct mem_cgroup *memcg;
5199
5200 cgroup_lock();
5201 cpus_read_lock();
5202 get_online_mems();
5203 mutex_lock(&state_mutex);
5204
5205 if (enabled == lru_gen_enabled())
5206 goto unlock;
5207
5208 if (enabled)
5209 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5210 else
5211 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5212
5213 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5214 do {
5215 int nid;
5216
5217 for_each_node(nid) {
5218 struct lruvec *lruvec = get_lruvec(memcg, nid);
5219
5220 spin_lock_irq(&lruvec->lru_lock);
5221
5222 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5223 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5224
5225 lruvec->lrugen.enabled = enabled;
5226
5227 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5228 spin_unlock_irq(&lruvec->lru_lock);
5229 cond_resched();
5230 spin_lock_irq(&lruvec->lru_lock);
5231 }
5232
5233 spin_unlock_irq(&lruvec->lru_lock);
5234 }
5235
5236 cond_resched();
5237 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5238unlock:
5239 mutex_unlock(&state_mutex);
5240 put_online_mems();
5241 cpus_read_unlock();
5242 cgroup_unlock();
5243}
5244
5245/******************************************************************************
5246 * sysfs interface
5247 ******************************************************************************/
5248
5249static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5250{
5251 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5252}
5253
5254/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5255static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5256 const char *buf, size_t len)
5257{
5258 unsigned int msecs;
5259
5260 if (kstrtouint(buf, 0, &msecs))
5261 return -EINVAL;
5262
5263 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5264
5265 return len;
5266}
5267
5268static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5269
5270static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5271{
5272 unsigned int caps = 0;
5273
5274 if (get_cap(LRU_GEN_CORE))
5275 caps |= BIT(LRU_GEN_CORE);
5276
5277 if (should_walk_mmu())
5278 caps |= BIT(LRU_GEN_MM_WALK);
5279
5280 if (should_clear_pmd_young())
5281 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5282
5283 return sysfs_emit(buf, "0x%04x\n", caps);
5284}
5285
5286/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5287static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5288 const char *buf, size_t len)
5289{
5290 int i;
5291 unsigned int caps;
5292
5293 if (tolower(*buf) == 'n')
5294 caps = 0;
5295 else if (tolower(*buf) == 'y')
5296 caps = -1;
5297 else if (kstrtouint(buf, 0, &caps))
5298 return -EINVAL;
5299
5300 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5301 bool enabled = caps & BIT(i);
5302
5303 if (i == LRU_GEN_CORE)
5304 lru_gen_change_state(enabled);
5305 else if (enabled)
5306 static_branch_enable(&lru_gen_caps[i]);
5307 else
5308 static_branch_disable(&lru_gen_caps[i]);
5309 }
5310
5311 return len;
5312}
5313
5314static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5315
5316static struct attribute *lru_gen_attrs[] = {
5317 &lru_gen_min_ttl_attr.attr,
5318 &lru_gen_enabled_attr.attr,
5319 NULL
5320};
5321
5322static const struct attribute_group lru_gen_attr_group = {
5323 .name = "lru_gen",
5324 .attrs = lru_gen_attrs,
5325};
5326
5327/******************************************************************************
5328 * debugfs interface
5329 ******************************************************************************/
5330
5331static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5332{
5333 struct mem_cgroup *memcg;
5334 loff_t nr_to_skip = *pos;
5335
5336 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5337 if (!m->private)
5338 return ERR_PTR(-ENOMEM);
5339
5340 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5341 do {
5342 int nid;
5343
5344 for_each_node_state(nid, N_MEMORY) {
5345 if (!nr_to_skip--)
5346 return get_lruvec(memcg, nid);
5347 }
5348 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5349
5350 return NULL;
5351}
5352
5353static void lru_gen_seq_stop(struct seq_file *m, void *v)
5354{
5355 if (!IS_ERR_OR_NULL(v))
5356 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5357
5358 kvfree(m->private);
5359 m->private = NULL;
5360}
5361
5362static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5363{
5364 int nid = lruvec_pgdat(v)->node_id;
5365 struct mem_cgroup *memcg = lruvec_memcg(v);
5366
5367 ++*pos;
5368
5369 nid = next_memory_node(nid);
5370 if (nid == MAX_NUMNODES) {
5371 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5372 if (!memcg)
5373 return NULL;
5374
5375 nid = first_memory_node;
5376 }
5377
5378 return get_lruvec(memcg, nid);
5379}
5380
5381static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5382 unsigned long max_seq, unsigned long *min_seq,
5383 unsigned long seq)
5384{
5385 int i;
5386 int type, tier;
5387 int hist = lru_hist_from_seq(seq);
5388 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5389 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5390
5391 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5392 seq_printf(m, " %10d", tier);
5393 for (type = 0; type < ANON_AND_FILE; type++) {
5394 const char *s = "xxx";
5395 unsigned long n[3] = {};
5396
5397 if (seq == max_seq) {
5398 s = "RTx";
5399 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5400 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5401 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5402 s = "rep";
5403 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5404 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5405 n[2] = READ_ONCE(lrugen->protected[hist][type][tier]);
5406 }
5407
5408 for (i = 0; i < 3; i++)
5409 seq_printf(m, " %10lu%c", n[i], s[i]);
5410 }
5411 seq_putc(m, '\n');
5412 }
5413
5414 if (!mm_state)
5415 return;
5416
5417 seq_puts(m, " ");
5418 for (i = 0; i < NR_MM_STATS; i++) {
5419 const char *s = "xxxx";
5420 unsigned long n = 0;
5421
5422 if (seq == max_seq && NR_HIST_GENS == 1) {
5423 s = "TYFA";
5424 n = READ_ONCE(mm_state->stats[hist][i]);
5425 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5426 s = "tyfa";
5427 n = READ_ONCE(mm_state->stats[hist][i]);
5428 }
5429
5430 seq_printf(m, " %10lu%c", n, s[i]);
5431 }
5432 seq_putc(m, '\n');
5433}
5434
5435/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5436static int lru_gen_seq_show(struct seq_file *m, void *v)
5437{
5438 unsigned long seq;
5439 bool full = debugfs_get_aux_num(m->file);
5440 struct lruvec *lruvec = v;
5441 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5442 int nid = lruvec_pgdat(lruvec)->node_id;
5443 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5444 DEFINE_MAX_SEQ(lruvec);
5445 DEFINE_MIN_SEQ(lruvec);
5446
5447 if (nid == first_memory_node) {
5448 const char *path = memcg ? m->private : "";
5449
5450#ifdef CONFIG_MEMCG
5451 if (memcg)
5452 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5453#endif
5454 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5455 }
5456
5457 seq_printf(m, " node %5d\n", nid);
5458
5459 if (!full)
5460 seq = evictable_min_seq(min_seq, MAX_SWAPPINESS / 2);
5461 else if (max_seq >= MAX_NR_GENS)
5462 seq = max_seq - MAX_NR_GENS + 1;
5463 else
5464 seq = 0;
5465
5466 for (; seq <= max_seq; seq++) {
5467 int type, zone;
5468 int gen = lru_gen_from_seq(seq);
5469 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5470
5471 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5472
5473 for (type = 0; type < ANON_AND_FILE; type++) {
5474 unsigned long size = 0;
5475 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5476
5477 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5478 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5479
5480 seq_printf(m, " %10lu%c", size, mark);
5481 }
5482
5483 seq_putc(m, '\n');
5484
5485 if (full)
5486 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5487 }
5488
5489 return 0;
5490}
5491
5492static const struct seq_operations lru_gen_seq_ops = {
5493 .start = lru_gen_seq_start,
5494 .stop = lru_gen_seq_stop,
5495 .next = lru_gen_seq_next,
5496 .show = lru_gen_seq_show,
5497};
5498
5499static int run_aging(struct lruvec *lruvec, unsigned long seq,
5500 int swappiness, bool force_scan)
5501{
5502 DEFINE_MAX_SEQ(lruvec);
5503
5504 if (seq > max_seq)
5505 return -EINVAL;
5506
5507 return try_to_inc_max_seq(lruvec, max_seq, swappiness, force_scan) ? 0 : -EEXIST;
5508}
5509
5510static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5511 int swappiness, unsigned long nr_to_reclaim)
5512{
5513 DEFINE_MAX_SEQ(lruvec);
5514
5515 if (seq + MIN_NR_GENS > max_seq)
5516 return -EINVAL;
5517
5518 sc->nr_reclaimed = 0;
5519
5520 while (!signal_pending(current)) {
5521 DEFINE_MIN_SEQ(lruvec);
5522
5523 if (seq < evictable_min_seq(min_seq, swappiness))
5524 return 0;
5525
5526 if (sc->nr_reclaimed >= nr_to_reclaim)
5527 return 0;
5528
5529 if (!evict_folios(nr_to_reclaim - sc->nr_reclaimed, lruvec, sc,
5530 swappiness))
5531 return 0;
5532
5533 cond_resched();
5534 }
5535
5536 return -EINTR;
5537}
5538
5539static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5540 struct scan_control *sc, int swappiness, unsigned long opt)
5541{
5542 struct lruvec *lruvec;
5543 int err = -EINVAL;
5544 struct mem_cgroup *memcg = NULL;
5545
5546 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5547 return -EINVAL;
5548
5549 if (!mem_cgroup_disabled()) {
5550 rcu_read_lock();
5551
5552 memcg = mem_cgroup_from_id(memcg_id);
5553 if (!mem_cgroup_tryget(memcg))
5554 memcg = NULL;
5555
5556 rcu_read_unlock();
5557
5558 if (!memcg)
5559 return -EINVAL;
5560 }
5561
5562 if (memcg_id != mem_cgroup_id(memcg))
5563 goto done;
5564
5565 sc->target_mem_cgroup = memcg;
5566 lruvec = get_lruvec(memcg, nid);
5567
5568 if (swappiness < MIN_SWAPPINESS)
5569 swappiness = get_swappiness(lruvec, sc);
5570 else if (swappiness > SWAPPINESS_ANON_ONLY)
5571 goto done;
5572
5573 switch (cmd) {
5574 case '+':
5575 err = run_aging(lruvec, seq, swappiness, opt);
5576 break;
5577 case '-':
5578 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5579 break;
5580 }
5581done:
5582 mem_cgroup_put(memcg);
5583
5584 return err;
5585}
5586
5587/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5588static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5589 size_t len, loff_t *pos)
5590{
5591 void *buf;
5592 char *cur, *next;
5593 unsigned int flags;
5594 struct blk_plug plug;
5595 int err = -EINVAL;
5596 struct scan_control sc = {
5597 .may_writepage = true,
5598 .may_unmap = true,
5599 .may_swap = true,
5600 .reclaim_idx = MAX_NR_ZONES - 1,
5601 .gfp_mask = GFP_KERNEL,
5602 .proactive = true,
5603 };
5604
5605 buf = kvmalloc(len + 1, GFP_KERNEL);
5606 if (!buf)
5607 return -ENOMEM;
5608
5609 if (copy_from_user(buf, src, len)) {
5610 kvfree(buf);
5611 return -EFAULT;
5612 }
5613
5614 set_task_reclaim_state(current, &sc.reclaim_state);
5615 flags = memalloc_noreclaim_save();
5616 blk_start_plug(&plug);
5617 if (!set_mm_walk(NULL, true)) {
5618 err = -ENOMEM;
5619 goto done;
5620 }
5621
5622 next = buf;
5623 next[len] = '\0';
5624
5625 while ((cur = strsep(&next, ",;\n"))) {
5626 int n;
5627 int end;
5628 char cmd, swap_string[5];
5629 unsigned int memcg_id;
5630 unsigned int nid;
5631 unsigned long seq;
5632 unsigned int swappiness;
5633 unsigned long opt = -1;
5634
5635 cur = skip_spaces(cur);
5636 if (!*cur)
5637 continue;
5638
5639 n = sscanf(cur, "%c %u %u %lu %n %4s %n %lu %n", &cmd, &memcg_id, &nid,
5640 &seq, &end, swap_string, &end, &opt, &end);
5641 if (n < 4 || cur[end]) {
5642 err = -EINVAL;
5643 break;
5644 }
5645
5646 if (n == 4) {
5647 swappiness = -1;
5648 } else if (!strcmp("max", swap_string)) {
5649 /* set by userspace for anonymous memory only */
5650 swappiness = SWAPPINESS_ANON_ONLY;
5651 } else {
5652 err = kstrtouint(swap_string, 0, &swappiness);
5653 if (err)
5654 break;
5655 }
5656
5657 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5658 if (err)
5659 break;
5660 }
5661done:
5662 clear_mm_walk();
5663 blk_finish_plug(&plug);
5664 memalloc_noreclaim_restore(flags);
5665 set_task_reclaim_state(current, NULL);
5666
5667 kvfree(buf);
5668
5669 return err ? : len;
5670}
5671
5672static int lru_gen_seq_open(struct inode *inode, struct file *file)
5673{
5674 return seq_open(file, &lru_gen_seq_ops);
5675}
5676
5677static const struct file_operations lru_gen_rw_fops = {
5678 .open = lru_gen_seq_open,
5679 .read = seq_read,
5680 .write = lru_gen_seq_write,
5681 .llseek = seq_lseek,
5682 .release = seq_release,
5683};
5684
5685static const struct file_operations lru_gen_ro_fops = {
5686 .open = lru_gen_seq_open,
5687 .read = seq_read,
5688 .llseek = seq_lseek,
5689 .release = seq_release,
5690};
5691
5692/******************************************************************************
5693 * initialization
5694 ******************************************************************************/
5695
5696void lru_gen_init_pgdat(struct pglist_data *pgdat)
5697{
5698 int i, j;
5699
5700 spin_lock_init(&pgdat->memcg_lru.lock);
5701
5702 for (i = 0; i < MEMCG_NR_GENS; i++) {
5703 for (j = 0; j < MEMCG_NR_BINS; j++)
5704 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5705 }
5706}
5707
5708void lru_gen_init_lruvec(struct lruvec *lruvec)
5709{
5710 int i;
5711 int gen, type, zone;
5712 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5713 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5714
5715 lrugen->max_seq = MIN_NR_GENS + 1;
5716 lrugen->enabled = lru_gen_enabled();
5717
5718 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5719 lrugen->timestamps[i] = jiffies;
5720
5721 for_each_gen_type_zone(gen, type, zone)
5722 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5723
5724 if (mm_state)
5725 mm_state->seq = MIN_NR_GENS;
5726}
5727
5728#ifdef CONFIG_MEMCG
5729
5730void lru_gen_init_memcg(struct mem_cgroup *memcg)
5731{
5732 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5733
5734 if (!mm_list)
5735 return;
5736
5737 INIT_LIST_HEAD(&mm_list->fifo);
5738 spin_lock_init(&mm_list->lock);
5739}
5740
5741void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5742{
5743 int i;
5744 int nid;
5745 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5746
5747 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5748
5749 for_each_node(nid) {
5750 struct lruvec *lruvec = get_lruvec(memcg, nid);
5751 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5752
5753 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5754 sizeof(lruvec->lrugen.nr_pages)));
5755
5756 lruvec->lrugen.list.next = LIST_POISON1;
5757
5758 if (!mm_state)
5759 continue;
5760
5761 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5762 bitmap_free(mm_state->filters[i]);
5763 mm_state->filters[i] = NULL;
5764 }
5765 }
5766}
5767
5768#endif /* CONFIG_MEMCG */
5769
5770static int __init init_lru_gen(void)
5771{
5772 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5773 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5774
5775 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5776 pr_err("lru_gen: failed to create sysfs group\n");
5777
5778 debugfs_create_file_aux_num("lru_gen", 0644, NULL, NULL, false,
5779 &lru_gen_rw_fops);
5780 debugfs_create_file_aux_num("lru_gen_full", 0444, NULL, NULL, true,
5781 &lru_gen_ro_fops);
5782
5783 return 0;
5784};
5785late_initcall(init_lru_gen);
5786
5787#else /* !CONFIG_LRU_GEN */
5788
5789static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5790{
5791 BUILD_BUG();
5792}
5793
5794static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5795{
5796 BUILD_BUG();
5797}
5798
5799static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5800{
5801 BUILD_BUG();
5802}
5803
5804#endif /* CONFIG_LRU_GEN */
5805
5806static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5807{
5808 unsigned long nr[NR_LRU_LISTS];
5809 unsigned long targets[NR_LRU_LISTS];
5810 unsigned long nr_to_scan;
5811 enum lru_list lru;
5812 unsigned long nr_reclaimed = 0;
5813 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5814 bool proportional_reclaim;
5815 struct blk_plug plug;
5816
5817 if (lru_gen_enabled() && !root_reclaim(sc)) {
5818 lru_gen_shrink_lruvec(lruvec, sc);
5819 return;
5820 }
5821
5822 get_scan_count(lruvec, sc, nr);
5823
5824 /* Record the original scan target for proportional adjustments later */
5825 memcpy(to: targets, from: nr, len: sizeof(nr));
5826
5827 /*
5828 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5829 * event that can occur when there is little memory pressure e.g.
5830 * multiple streaming readers/writers. Hence, we do not abort scanning
5831 * when the requested number of pages are reclaimed when scanning at
5832 * DEF_PRIORITY on the assumption that the fact we are direct
5833 * reclaiming implies that kswapd is not keeping up and it is best to
5834 * do a batch of work at once. For memcg reclaim one check is made to
5835 * abort proportional reclaim if either the file or anon lru has already
5836 * dropped to zero at the first pass.
5837 */
5838 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5839 sc->priority == DEF_PRIORITY);
5840
5841 blk_start_plug(&plug);
5842 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5843 nr[LRU_INACTIVE_FILE]) {
5844 unsigned long nr_anon, nr_file, percentage;
5845 unsigned long nr_scanned;
5846
5847 for_each_evictable_lru(lru) {
5848 if (nr[lru]) {
5849 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5850 nr[lru] -= nr_to_scan;
5851
5852 nr_reclaimed += shrink_list(lru, nr_to_scan,
5853 lruvec, sc);
5854 }
5855 }
5856
5857 cond_resched();
5858
5859 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5860 continue;
5861
5862 /*
5863 * For kswapd and memcg, reclaim at least the number of pages
5864 * requested. Ensure that the anon and file LRUs are scanned
5865 * proportionally what was requested by get_scan_count(). We
5866 * stop reclaiming one LRU and reduce the amount scanning
5867 * proportional to the original scan target.
5868 */
5869 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5870 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5871
5872 /*
5873 * It's just vindictive to attack the larger once the smaller
5874 * has gone to zero. And given the way we stop scanning the
5875 * smaller below, this makes sure that we only make one nudge
5876 * towards proportionality once we've got nr_to_reclaim.
5877 */
5878 if (!nr_file || !nr_anon)
5879 break;
5880
5881 if (nr_file > nr_anon) {
5882 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5883 targets[LRU_ACTIVE_ANON] + 1;
5884 lru = LRU_BASE;
5885 percentage = nr_anon * 100 / scan_target;
5886 } else {
5887 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5888 targets[LRU_ACTIVE_FILE] + 1;
5889 lru = LRU_FILE;
5890 percentage = nr_file * 100 / scan_target;
5891 }
5892
5893 /* Stop scanning the smaller of the LRU */
5894 nr[lru] = 0;
5895 nr[lru + LRU_ACTIVE] = 0;
5896
5897 /*
5898 * Recalculate the other LRU scan count based on its original
5899 * scan target and the percentage scanning already complete
5900 */
5901 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5902 nr_scanned = targets[lru] - nr[lru];
5903 nr[lru] = targets[lru] * (100 - percentage) / 100;
5904 nr[lru] -= min(nr[lru], nr_scanned);
5905
5906 lru += LRU_ACTIVE;
5907 nr_scanned = targets[lru] - nr[lru];
5908 nr[lru] = targets[lru] * (100 - percentage) / 100;
5909 nr[lru] -= min(nr[lru], nr_scanned);
5910 }
5911 blk_finish_plug(&plug);
5912 sc->nr_reclaimed += nr_reclaimed;
5913
5914 /*
5915 * Even if we did not try to evict anon pages at all, we want to
5916 * rebalance the anon lru active/inactive ratio.
5917 */
5918 if (can_age_anon_pages(lruvec, sc) &&
5919 inactive_is_low(lruvec, inactive_lru: LRU_INACTIVE_ANON))
5920 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5921 sc, lru: LRU_ACTIVE_ANON);
5922}
5923
5924/* Use reclaim/compaction for costly allocs or under memory pressure */
5925static bool in_reclaim_compaction(struct scan_control *sc)
5926{
5927 if (gfp_compaction_allowed(gfp_mask: sc->gfp_mask) && sc->order &&
5928 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5929 sc->priority < DEF_PRIORITY - 2))
5930 return true;
5931
5932 return false;
5933}
5934
5935/*
5936 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5937 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5938 * true if more pages should be reclaimed such that when the page allocator
5939 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5940 * It will give up earlier than that if there is difficulty reclaiming pages.
5941 */
5942static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5943 unsigned long nr_reclaimed,
5944 struct scan_control *sc)
5945{
5946 unsigned long pages_for_compaction;
5947 unsigned long inactive_lru_pages;
5948 int z;
5949 struct zone *zone;
5950
5951 /* If not in reclaim/compaction mode, stop */
5952 if (!in_reclaim_compaction(sc))
5953 return false;
5954
5955 /*
5956 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5957 * number of pages that were scanned. This will return to the caller
5958 * with the risk reclaim/compaction and the resulting allocation attempt
5959 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5960 * allocations through requiring that the full LRU list has been scanned
5961 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5962 * scan, but that approximation was wrong, and there were corner cases
5963 * where always a non-zero amount of pages were scanned.
5964 */
5965 if (!nr_reclaimed)
5966 return false;
5967
5968 /* If compaction would go ahead or the allocation would succeed, stop */
5969 for_each_managed_zone_pgdat(zone, pgdat, z, sc->reclaim_idx) {
5970 unsigned long watermark = min_wmark_pages(z: zone);
5971
5972 /* Allocation can already succeed, nothing to do */
5973 if (zone_watermark_ok(z: zone, order: sc->order, mark: watermark,
5974 highest_zoneidx: sc->reclaim_idx, alloc_flags: 0))
5975 return false;
5976
5977 if (compaction_suitable(zone, order: sc->order, watermark,
5978 highest_zoneidx: sc->reclaim_idx))
5979 return false;
5980 }
5981
5982 /*
5983 * If we have not reclaimed enough pages for compaction and the
5984 * inactive lists are large enough, continue reclaiming
5985 */
5986 pages_for_compaction = compact_gap(order: sc->order);
5987 inactive_lru_pages = node_page_state(pgdat, item: NR_INACTIVE_FILE);
5988 if (can_reclaim_anon_pages(NULL, nid: pgdat->node_id, sc))
5989 inactive_lru_pages += node_page_state(pgdat, item: NR_INACTIVE_ANON);
5990
5991 return inactive_lru_pages > pages_for_compaction;
5992}
5993
5994static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5995{
5996 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5997 struct mem_cgroup_reclaim_cookie reclaim = {
5998 .pgdat = pgdat,
5999 };
6000 struct mem_cgroup_reclaim_cookie *partial = &reclaim;
6001 struct mem_cgroup *memcg;
6002
6003 /*
6004 * In most cases, direct reclaimers can do partial walks
6005 * through the cgroup tree, using an iterator state that
6006 * persists across invocations. This strikes a balance between
6007 * fairness and allocation latency.
6008 *
6009 * For kswapd, reliable forward progress is more important
6010 * than a quick return to idle. Always do full walks.
6011 */
6012 if (current_is_kswapd() || sc->memcg_full_walk)
6013 partial = NULL;
6014
6015 memcg = mem_cgroup_iter(root: target_memcg, NULL, reclaim: partial);
6016 do {
6017 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6018 unsigned long reclaimed;
6019 unsigned long scanned;
6020
6021 /*
6022 * This loop can become CPU-bound when target memcgs
6023 * aren't eligible for reclaim - either because they
6024 * don't have any reclaimable pages, or because their
6025 * memory is explicitly protected. Avoid soft lockups.
6026 */
6027 cond_resched();
6028
6029 mem_cgroup_calculate_protection(root: target_memcg, memcg);
6030
6031 if (mem_cgroup_below_min(target: target_memcg, memcg)) {
6032 /*
6033 * Hard protection.
6034 * If there is no reclaimable memory, OOM.
6035 */
6036 continue;
6037 } else if (mem_cgroup_below_low(target: target_memcg, memcg)) {
6038 /*
6039 * Soft protection.
6040 * Respect the protection only as long as
6041 * there is an unprotected supply
6042 * of reclaimable memory from other cgroups.
6043 */
6044 if (!sc->memcg_low_reclaim) {
6045 sc->memcg_low_skipped = 1;
6046 continue;
6047 }
6048 memcg_memory_event(memcg, event: MEMCG_LOW);
6049 }
6050
6051 reclaimed = sc->nr_reclaimed;
6052 scanned = sc->nr_scanned;
6053
6054 shrink_lruvec(lruvec, sc);
6055
6056 shrink_slab(gfp_mask: sc->gfp_mask, nid: pgdat->node_id, memcg,
6057 priority: sc->priority);
6058
6059 /* Record the group's reclaim efficiency */
6060 if (!sc->proactive)
6061 vmpressure(gfp: sc->gfp_mask, memcg, tree: false,
6062 scanned: sc->nr_scanned - scanned,
6063 reclaimed: sc->nr_reclaimed - reclaimed);
6064
6065 /* If partial walks are allowed, bail once goal is reached */
6066 if (partial && sc->nr_reclaimed >= sc->nr_to_reclaim) {
6067 mem_cgroup_iter_break(root: target_memcg, prev: memcg);
6068 break;
6069 }
6070 } while ((memcg = mem_cgroup_iter(root: target_memcg, prev: memcg, reclaim: partial)));
6071}
6072
6073static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6074{
6075 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
6076 struct lruvec *target_lruvec;
6077 bool reclaimable = false;
6078
6079 if (lru_gen_enabled() && root_reclaim(sc)) {
6080 memset(s: &sc->nr, c: 0, n: sizeof(sc->nr));
6081 lru_gen_shrink_node(pgdat, sc);
6082 return;
6083 }
6084
6085 target_lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup, pgdat);
6086
6087again:
6088 memset(s: &sc->nr, c: 0, n: sizeof(sc->nr));
6089
6090 nr_reclaimed = sc->nr_reclaimed;
6091 nr_scanned = sc->nr_scanned;
6092
6093 prepare_scan_control(pgdat, sc);
6094
6095 shrink_node_memcgs(pgdat, sc);
6096
6097 flush_reclaim_state(sc);
6098
6099 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
6100
6101 /* Record the subtree's reclaim efficiency */
6102 if (!sc->proactive)
6103 vmpressure(gfp: sc->gfp_mask, memcg: sc->target_mem_cgroup, tree: true,
6104 scanned: sc->nr_scanned - nr_scanned, reclaimed: nr_node_reclaimed);
6105
6106 if (nr_node_reclaimed)
6107 reclaimable = true;
6108
6109 if (current_is_kswapd()) {
6110 /*
6111 * If reclaim is isolating dirty pages under writeback,
6112 * it implies that the long-lived page allocation rate
6113 * is exceeding the page laundering rate. Either the
6114 * global limits are not being effective at throttling
6115 * processes due to the page distribution throughout
6116 * zones or there is heavy usage of a slow backing
6117 * device. The only option is to throttle from reclaim
6118 * context which is not ideal as there is no guarantee
6119 * the dirtying process is throttled in the same way
6120 * balance_dirty_pages() manages.
6121 *
6122 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6123 * count the number of pages under pages flagged for
6124 * immediate reclaim and stall if any are encountered
6125 * in the nr_immediate check below.
6126 */
6127 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6128 set_bit(nr: PGDAT_WRITEBACK, addr: &pgdat->flags);
6129
6130 /* Allow kswapd to start writing pages during reclaim.*/
6131 if (sc->nr.unqueued_dirty &&
6132 sc->nr.unqueued_dirty == sc->nr.file_taken)
6133 set_bit(nr: PGDAT_DIRTY, addr: &pgdat->flags);
6134
6135 /*
6136 * If kswapd scans pages marked for immediate
6137 * reclaim and under writeback (nr_immediate), it
6138 * implies that pages are cycling through the LRU
6139 * faster than they are written so forcibly stall
6140 * until some pages complete writeback.
6141 */
6142 if (sc->nr.immediate)
6143 reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_WRITEBACK);
6144 }
6145
6146 /*
6147 * Tag a node/memcg as congested if all the dirty pages were marked
6148 * for writeback and immediate reclaim (counted in nr.congested).
6149 *
6150 * Legacy memcg will stall in page writeback so avoid forcibly
6151 * stalling in reclaim_throttle().
6152 */
6153 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
6154 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
6155 set_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &target_lruvec->flags);
6156
6157 if (current_is_kswapd())
6158 set_bit(nr: LRUVEC_NODE_CONGESTED, addr: &target_lruvec->flags);
6159 }
6160
6161 /*
6162 * Stall direct reclaim for IO completions if the lruvec is
6163 * node is congested. Allow kswapd to continue until it
6164 * starts encountering unqueued dirty pages or cycling through
6165 * the LRU too quickly.
6166 */
6167 if (!current_is_kswapd() && current_may_throttle() &&
6168 !sc->hibernation_mode &&
6169 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
6170 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
6171 reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_CONGESTED);
6172
6173 if (should_continue_reclaim(pgdat, nr_reclaimed: nr_node_reclaimed, sc))
6174 goto again;
6175
6176 /*
6177 * Kswapd gives up on balancing particular nodes after too
6178 * many failures to reclaim anything from them and goes to
6179 * sleep. On reclaim progress, reset the failure counter. A
6180 * successful direct reclaim run will revive a dormant kswapd.
6181 */
6182 if (reclaimable)
6183 atomic_set(v: &pgdat->kswapd_failures, i: 0);
6184 else if (sc->cache_trim_mode)
6185 sc->cache_trim_mode_failed = 1;
6186}
6187
6188/*
6189 * Returns true if compaction should go ahead for a costly-order request, or
6190 * the allocation would already succeed without compaction. Return false if we
6191 * should reclaim first.
6192 */
6193static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6194{
6195 unsigned long watermark;
6196
6197 if (!gfp_compaction_allowed(gfp_mask: sc->gfp_mask))
6198 return false;
6199
6200 /* Allocation can already succeed, nothing to do */
6201 if (zone_watermark_ok(z: zone, order: sc->order, mark: min_wmark_pages(z: zone),
6202 highest_zoneidx: sc->reclaim_idx, alloc_flags: 0))
6203 return true;
6204
6205 /*
6206 * Direct reclaim usually targets the min watermark, but compaction
6207 * takes time to run and there are potentially other callers using the
6208 * pages just freed. So target a higher buffer to give compaction a
6209 * reasonable chance of completing and allocating the pages.
6210 *
6211 * Note that we won't actually reclaim the whole buffer in one attempt
6212 * as the target watermark in should_continue_reclaim() is lower. But if
6213 * we are already above the high+gap watermark, don't reclaim at all.
6214 */
6215 watermark = high_wmark_pages(z: zone);
6216 if (compaction_suitable(zone, order: sc->order, watermark, highest_zoneidx: sc->reclaim_idx))
6217 return true;
6218
6219 return false;
6220}
6221
6222static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6223{
6224 /*
6225 * If reclaim is making progress greater than 12% efficiency then
6226 * wake all the NOPROGRESS throttled tasks.
6227 */
6228 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6229 wait_queue_head_t *wqh;
6230
6231 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6232 if (waitqueue_active(wq_head: wqh))
6233 wake_up(wqh);
6234
6235 return;
6236 }
6237
6238 /*
6239 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6240 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6241 * under writeback and marked for immediate reclaim at the tail of the
6242 * LRU.
6243 */
6244 if (current_is_kswapd() || cgroup_reclaim(sc))
6245 return;
6246
6247 /* Throttle if making no progress at high prioities. */
6248 if (sc->priority == 1 && !sc->nr_reclaimed)
6249 reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_NOPROGRESS);
6250}
6251
6252/*
6253 * This is the direct reclaim path, for page-allocating processes. We only
6254 * try to reclaim pages from zones which will satisfy the caller's allocation
6255 * request.
6256 *
6257 * If a zone is deemed to be full of pinned pages then just give it a light
6258 * scan then give up on it.
6259 */
6260static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6261{
6262 struct zoneref *z;
6263 struct zone *zone;
6264 unsigned long nr_soft_reclaimed;
6265 unsigned long nr_soft_scanned;
6266 gfp_t orig_mask;
6267 pg_data_t *last_pgdat = NULL;
6268 pg_data_t *first_pgdat = NULL;
6269
6270 /*
6271 * If the number of buffer_heads in the machine exceeds the maximum
6272 * allowed level, force direct reclaim to scan the highmem zone as
6273 * highmem pages could be pinning lowmem pages storing buffer_heads
6274 */
6275 orig_mask = sc->gfp_mask;
6276 if (buffer_heads_over_limit) {
6277 sc->gfp_mask |= __GFP_HIGHMEM;
6278 sc->reclaim_idx = gfp_zone(flags: sc->gfp_mask);
6279 }
6280
6281 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6282 sc->reclaim_idx, sc->nodemask) {
6283 /*
6284 * Take care memory controller reclaiming has small influence
6285 * to global LRU.
6286 */
6287 if (!cgroup_reclaim(sc)) {
6288 if (!cpuset_zone_allowed(z: zone,
6289 GFP_KERNEL | __GFP_HARDWALL))
6290 continue;
6291
6292 /*
6293 * If we already have plenty of memory free for
6294 * compaction in this zone, don't free any more.
6295 * Even though compaction is invoked for any
6296 * non-zero order, only frequent costly order
6297 * reclamation is disruptive enough to become a
6298 * noticeable problem, like transparent huge
6299 * page allocations.
6300 */
6301 if (IS_ENABLED(CONFIG_COMPACTION) &&
6302 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6303 compaction_ready(zone, sc)) {
6304 sc->compaction_ready = true;
6305 continue;
6306 }
6307
6308 /*
6309 * Shrink each node in the zonelist once. If the
6310 * zonelist is ordered by zone (not the default) then a
6311 * node may be shrunk multiple times but in that case
6312 * the user prefers lower zones being preserved.
6313 */
6314 if (zone->zone_pgdat == last_pgdat)
6315 continue;
6316
6317 /*
6318 * This steals pages from memory cgroups over softlimit
6319 * and returns the number of reclaimed pages and
6320 * scanned pages. This works for global memory pressure
6321 * and balancing, not for a memcg's limit.
6322 */
6323 nr_soft_scanned = 0;
6324 nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat: zone->zone_pgdat,
6325 order: sc->order, gfp_mask: sc->gfp_mask,
6326 total_scanned: &nr_soft_scanned);
6327 sc->nr_reclaimed += nr_soft_reclaimed;
6328 sc->nr_scanned += nr_soft_scanned;
6329 /* need some check for avoid more shrink_zone() */
6330 }
6331
6332 if (!first_pgdat)
6333 first_pgdat = zone->zone_pgdat;
6334
6335 /* See comment about same check for global reclaim above */
6336 if (zone->zone_pgdat == last_pgdat)
6337 continue;
6338 last_pgdat = zone->zone_pgdat;
6339 shrink_node(pgdat: zone->zone_pgdat, sc);
6340 }
6341
6342 if (first_pgdat)
6343 consider_reclaim_throttle(pgdat: first_pgdat, sc);
6344
6345 /*
6346 * Restore to original mask to avoid the impact on the caller if we
6347 * promoted it to __GFP_HIGHMEM.
6348 */
6349 sc->gfp_mask = orig_mask;
6350}
6351
6352static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6353{
6354 struct lruvec *target_lruvec;
6355 unsigned long refaults;
6356
6357 if (lru_gen_enabled())
6358 return;
6359
6360 target_lruvec = mem_cgroup_lruvec(memcg: target_memcg, pgdat);
6361 refaults = lruvec_page_state(lruvec: target_lruvec, idx: WORKINGSET_ACTIVATE_ANON);
6362 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6363 refaults = lruvec_page_state(lruvec: target_lruvec, idx: WORKINGSET_ACTIVATE_FILE);
6364 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6365}
6366
6367/*
6368 * This is the main entry point to direct page reclaim.
6369 *
6370 * If a full scan of the inactive list fails to free enough memory then we
6371 * are "out of memory" and something needs to be killed.
6372 *
6373 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6374 * high - the zone may be full of dirty or under-writeback pages, which this
6375 * caller can't do much about. We kick the writeback threads and take explicit
6376 * naps in the hope that some of these pages can be written. But if the
6377 * allocating task holds filesystem locks which prevent writeout this might not
6378 * work, and the allocation attempt will fail.
6379 *
6380 * returns: 0, if no pages reclaimed
6381 * else, the number of pages reclaimed
6382 */
6383static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6384 struct scan_control *sc)
6385{
6386 int initial_priority = sc->priority;
6387 pg_data_t *last_pgdat;
6388 struct zoneref *z;
6389 struct zone *zone;
6390retry:
6391 delayacct_freepages_start();
6392
6393 if (!cgroup_reclaim(sc))
6394 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6395
6396 do {
6397 if (!sc->proactive)
6398 vmpressure_prio(gfp: sc->gfp_mask, memcg: sc->target_mem_cgroup,
6399 prio: sc->priority);
6400 sc->nr_scanned = 0;
6401 shrink_zones(zonelist, sc);
6402
6403 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6404 break;
6405
6406 if (sc->compaction_ready)
6407 break;
6408
6409 /*
6410 * If we're getting trouble reclaiming, start doing
6411 * writepage even in laptop mode.
6412 */
6413 if (sc->priority < DEF_PRIORITY - 2)
6414 sc->may_writepage = 1;
6415 } while (--sc->priority >= 0);
6416
6417 last_pgdat = NULL;
6418 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6419 sc->nodemask) {
6420 if (zone->zone_pgdat == last_pgdat)
6421 continue;
6422 last_pgdat = zone->zone_pgdat;
6423
6424 snapshot_refaults(target_memcg: sc->target_mem_cgroup, pgdat: zone->zone_pgdat);
6425
6426 if (cgroup_reclaim(sc)) {
6427 struct lruvec *lruvec;
6428
6429 lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup,
6430 pgdat: zone->zone_pgdat);
6431 clear_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &lruvec->flags);
6432 }
6433 }
6434
6435 delayacct_freepages_end();
6436
6437 if (sc->nr_reclaimed)
6438 return sc->nr_reclaimed;
6439
6440 /* Aborted reclaim to try compaction? don't OOM, then */
6441 if (sc->compaction_ready)
6442 return 1;
6443
6444 /*
6445 * In most cases, direct reclaimers can do partial walks
6446 * through the cgroup tree to meet the reclaim goal while
6447 * keeping latency low. Since the iterator state is shared
6448 * among all direct reclaim invocations (to retain fairness
6449 * among cgroups), though, high concurrency can result in
6450 * individual threads not seeing enough cgroups to make
6451 * meaningful forward progress. Avoid false OOMs in this case.
6452 */
6453 if (!sc->memcg_full_walk) {
6454 sc->priority = initial_priority;
6455 sc->memcg_full_walk = 1;
6456 goto retry;
6457 }
6458
6459 /*
6460 * We make inactive:active ratio decisions based on the node's
6461 * composition of memory, but a restrictive reclaim_idx or a
6462 * memory.low cgroup setting can exempt large amounts of
6463 * memory from reclaim. Neither of which are very common, so
6464 * instead of doing costly eligibility calculations of the
6465 * entire cgroup subtree up front, we assume the estimates are
6466 * good, and retry with forcible deactivation if that fails.
6467 */
6468 if (sc->skipped_deactivate) {
6469 sc->priority = initial_priority;
6470 sc->force_deactivate = 1;
6471 sc->skipped_deactivate = 0;
6472 goto retry;
6473 }
6474
6475 /* Untapped cgroup reserves? Don't OOM, retry. */
6476 if (sc->memcg_low_skipped) {
6477 sc->priority = initial_priority;
6478 sc->force_deactivate = 0;
6479 sc->memcg_low_reclaim = 1;
6480 sc->memcg_low_skipped = 0;
6481 goto retry;
6482 }
6483
6484 return 0;
6485}
6486
6487static bool allow_direct_reclaim(pg_data_t *pgdat)
6488{
6489 struct zone *zone;
6490 unsigned long pfmemalloc_reserve = 0;
6491 unsigned long free_pages = 0;
6492 int i;
6493 bool wmark_ok;
6494
6495 if (atomic_read(v: &pgdat->kswapd_failures) >= MAX_RECLAIM_RETRIES)
6496 return true;
6497
6498 for_each_managed_zone_pgdat(zone, pgdat, i, ZONE_NORMAL) {
6499 if (!zone_reclaimable_pages(zone) && zone_page_state_snapshot(zone, item: NR_FREE_PAGES))
6500 continue;
6501
6502 pfmemalloc_reserve += min_wmark_pages(z: zone);
6503 free_pages += zone_page_state_snapshot(zone, item: NR_FREE_PAGES);
6504 }
6505
6506 /* If there are no reserves (unexpected config) then do not throttle */
6507 if (!pfmemalloc_reserve)
6508 return true;
6509
6510 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6511
6512 /* kswapd must be awake if processes are being throttled */
6513 if (!wmark_ok && waitqueue_active(wq_head: &pgdat->kswapd_wait)) {
6514 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6515 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6516
6517 wake_up_interruptible(&pgdat->kswapd_wait);
6518 }
6519
6520 return wmark_ok;
6521}
6522
6523/*
6524 * Throttle direct reclaimers if backing storage is backed by the network
6525 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6526 * depleted. kswapd will continue to make progress and wake the processes
6527 * when the low watermark is reached.
6528 *
6529 * Returns true if a fatal signal was delivered during throttling. If this
6530 * happens, the page allocator should not consider triggering the OOM killer.
6531 */
6532static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6533 nodemask_t *nodemask)
6534{
6535 struct zoneref *z;
6536 struct zone *zone;
6537 pg_data_t *pgdat = NULL;
6538
6539 /*
6540 * Kernel threads should not be throttled as they may be indirectly
6541 * responsible for cleaning pages necessary for reclaim to make forward
6542 * progress. kjournald for example may enter direct reclaim while
6543 * committing a transaction where throttling it could forcing other
6544 * processes to block on log_wait_commit().
6545 */
6546 if (current->flags & PF_KTHREAD)
6547 goto out;
6548
6549 /*
6550 * If a fatal signal is pending, this process should not throttle.
6551 * It should return quickly so it can exit and free its memory
6552 */
6553 if (fatal_signal_pending(current))
6554 goto out;
6555
6556 /*
6557 * Check if the pfmemalloc reserves are ok by finding the first node
6558 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6559 * GFP_KERNEL will be required for allocating network buffers when
6560 * swapping over the network so ZONE_HIGHMEM is unusable.
6561 *
6562 * Throttling is based on the first usable node and throttled processes
6563 * wait on a queue until kswapd makes progress and wakes them. There
6564 * is an affinity then between processes waking up and where reclaim
6565 * progress has been made assuming the process wakes on the same node.
6566 * More importantly, processes running on remote nodes will not compete
6567 * for remote pfmemalloc reserves and processes on different nodes
6568 * should make reasonable progress.
6569 */
6570 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6571 gfp_zone(gfp_mask), nodemask) {
6572 if (zone_idx(zone) > ZONE_NORMAL)
6573 continue;
6574
6575 /* Throttle based on the first usable node */
6576 pgdat = zone->zone_pgdat;
6577 if (allow_direct_reclaim(pgdat))
6578 goto out;
6579 break;
6580 }
6581
6582 /* If no zone was usable by the allocation flags then do not throttle */
6583 if (!pgdat)
6584 goto out;
6585
6586 /* Account for the throttling */
6587 count_vm_event(item: PGSCAN_DIRECT_THROTTLE);
6588
6589 /*
6590 * If the caller cannot enter the filesystem, it's possible that it
6591 * is due to the caller holding an FS lock or performing a journal
6592 * transaction in the case of a filesystem like ext[3|4]. In this case,
6593 * it is not safe to block on pfmemalloc_wait as kswapd could be
6594 * blocked waiting on the same lock. Instead, throttle for up to a
6595 * second before continuing.
6596 */
6597 if (!(gfp_mask & __GFP_FS))
6598 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6599 allow_direct_reclaim(pgdat), HZ);
6600 else
6601 /* Throttle until kswapd wakes the process */
6602 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6603 allow_direct_reclaim(pgdat));
6604
6605 if (fatal_signal_pending(current))
6606 return true;
6607
6608out:
6609 return false;
6610}
6611
6612unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6613 gfp_t gfp_mask, nodemask_t *nodemask)
6614{
6615 unsigned long nr_reclaimed;
6616 struct scan_control sc = {
6617 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6618 .gfp_mask = current_gfp_context(flags: gfp_mask),
6619 .reclaim_idx = gfp_zone(flags: gfp_mask),
6620 .order = order,
6621 .nodemask = nodemask,
6622 .priority = DEF_PRIORITY,
6623 .may_writepage = !laptop_mode,
6624 .may_unmap = 1,
6625 .may_swap = 1,
6626 };
6627
6628 /*
6629 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6630 * Confirm they are large enough for max values.
6631 */
6632 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6633 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6634 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6635
6636 /*
6637 * Do not enter reclaim if fatal signal was delivered while throttled.
6638 * 1 is returned so that the page allocator does not OOM kill at this
6639 * point.
6640 */
6641 if (throttle_direct_reclaim(gfp_mask: sc.gfp_mask, zonelist, nodemask))
6642 return 1;
6643
6644 set_task_reclaim_state(current, rs: &sc.reclaim_state);
6645 trace_mm_vmscan_direct_reclaim_begin(order, gfp_flags: sc.gfp_mask);
6646
6647 nr_reclaimed = do_try_to_free_pages(zonelist, sc: &sc);
6648
6649 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6650 set_task_reclaim_state(current, NULL);
6651
6652 return nr_reclaimed;
6653}
6654
6655#ifdef CONFIG_MEMCG
6656
6657/* Only used by soft limit reclaim. Do not reuse for anything else. */
6658unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6659 gfp_t gfp_mask, bool noswap,
6660 pg_data_t *pgdat,
6661 unsigned long *nr_scanned)
6662{
6663 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6664 struct scan_control sc = {
6665 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6666 .target_mem_cgroup = memcg,
6667 .may_writepage = !laptop_mode,
6668 .may_unmap = 1,
6669 .reclaim_idx = MAX_NR_ZONES - 1,
6670 .may_swap = !noswap,
6671 };
6672
6673 WARN_ON_ONCE(!current->reclaim_state);
6674
6675 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6676 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6677
6678 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6679 sc.gfp_mask);
6680
6681 /*
6682 * NOTE: Although we can get the priority field, using it
6683 * here is not a good idea, since it limits the pages we can scan.
6684 * if we don't reclaim here, the shrink_node from balance_pgdat
6685 * will pick up pages from other mem cgroup's as well. We hack
6686 * the priority and make it zero.
6687 */
6688 shrink_lruvec(lruvec, &sc);
6689
6690 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6691
6692 *nr_scanned = sc.nr_scanned;
6693
6694 return sc.nr_reclaimed;
6695}
6696
6697unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6698 unsigned long nr_pages,
6699 gfp_t gfp_mask,
6700 unsigned int reclaim_options,
6701 int *swappiness)
6702{
6703 unsigned long nr_reclaimed;
6704 unsigned int noreclaim_flag;
6705 struct scan_control sc = {
6706 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6707 .proactive_swappiness = swappiness,
6708 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6709 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6710 .reclaim_idx = MAX_NR_ZONES - 1,
6711 .target_mem_cgroup = memcg,
6712 .priority = DEF_PRIORITY,
6713 .may_writepage = !laptop_mode,
6714 .may_unmap = 1,
6715 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6716 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6717 };
6718 /*
6719 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6720 * equal pressure on all the nodes. This is based on the assumption that
6721 * the reclaim does not bail out early.
6722 */
6723 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6724
6725 set_task_reclaim_state(current, &sc.reclaim_state);
6726 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6727 noreclaim_flag = memalloc_noreclaim_save();
6728
6729 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6730
6731 memalloc_noreclaim_restore(noreclaim_flag);
6732 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6733 set_task_reclaim_state(current, NULL);
6734
6735 return nr_reclaimed;
6736}
6737#else
6738unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6739 unsigned long nr_pages,
6740 gfp_t gfp_mask,
6741 unsigned int reclaim_options,
6742 int *swappiness)
6743{
6744 return 0;
6745}
6746#endif
6747
6748static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6749{
6750 struct mem_cgroup *memcg;
6751 struct lruvec *lruvec;
6752
6753 if (lru_gen_enabled()) {
6754 lru_gen_age_node(pgdat, sc);
6755 return;
6756 }
6757
6758 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6759 if (!can_age_anon_pages(lruvec, sc))
6760 return;
6761
6762 if (!inactive_is_low(lruvec, inactive_lru: LRU_INACTIVE_ANON))
6763 return;
6764
6765 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6766 do {
6767 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6768 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6769 sc, lru: LRU_ACTIVE_ANON);
6770 memcg = mem_cgroup_iter(NULL, prev: memcg, NULL);
6771 } while (memcg);
6772}
6773
6774static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6775{
6776 int i;
6777 struct zone *zone;
6778
6779 /*
6780 * Check for watermark boosts top-down as the higher zones
6781 * are more likely to be boosted. Both watermarks and boosts
6782 * should not be checked at the same time as reclaim would
6783 * start prematurely when there is no boosting and a lower
6784 * zone is balanced.
6785 */
6786 for (i = highest_zoneidx; i >= 0; i--) {
6787 zone = pgdat->node_zones + i;
6788 if (!managed_zone(zone))
6789 continue;
6790
6791 if (zone->watermark_boost)
6792 return true;
6793 }
6794
6795 return false;
6796}
6797
6798/*
6799 * Returns true if there is an eligible zone balanced for the request order
6800 * and highest_zoneidx
6801 */
6802static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6803{
6804 int i;
6805 unsigned long mark = -1;
6806 struct zone *zone;
6807
6808 /*
6809 * Check watermarks bottom-up as lower zones are more likely to
6810 * meet watermarks.
6811 */
6812 for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
6813 enum zone_stat_item item;
6814 unsigned long free_pages;
6815
6816 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6817 mark = promo_wmark_pages(z: zone);
6818 else
6819 mark = high_wmark_pages(z: zone);
6820
6821 /*
6822 * In defrag_mode, watermarks must be met in whole
6823 * blocks to avoid polluting allocator fallbacks.
6824 *
6825 * However, kswapd usually cannot accomplish this on
6826 * its own and needs kcompactd support. Once it's
6827 * reclaimed a compaction gap, and kswapd_shrink_node
6828 * has dropped order, simply ensure there are enough
6829 * base pages for compaction, wake kcompactd & sleep.
6830 */
6831 if (defrag_mode && order)
6832 item = NR_FREE_PAGES_BLOCKS;
6833 else
6834 item = NR_FREE_PAGES;
6835
6836 /*
6837 * When there is a high number of CPUs in the system,
6838 * the cumulative error from the vmstat per-cpu cache
6839 * can blur the line between the watermarks. In that
6840 * case, be safe and get an accurate snapshot.
6841 *
6842 * TODO: NR_FREE_PAGES_BLOCKS moves in steps of
6843 * pageblock_nr_pages, while the vmstat pcp threshold
6844 * is limited to 125. On many configurations that
6845 * counter won't actually be per-cpu cached. But keep
6846 * things simple for now; revisit when somebody cares.
6847 */
6848 free_pages = zone_page_state(zone, item);
6849 if (zone->percpu_drift_mark && free_pages < zone->percpu_drift_mark)
6850 free_pages = zone_page_state_snapshot(zone, item);
6851
6852 if (__zone_watermark_ok(z: zone, order, mark, highest_zoneidx,
6853 alloc_flags: 0, free_pages))
6854 return true;
6855 }
6856
6857 /*
6858 * If a node has no managed zone within highest_zoneidx, it does not
6859 * need balancing by definition. This can happen if a zone-restricted
6860 * allocation tries to wake a remote kswapd.
6861 */
6862 if (mark == -1)
6863 return true;
6864
6865 return false;
6866}
6867
6868/* Clear pgdat state for congested, dirty or under writeback. */
6869static void clear_pgdat_congested(pg_data_t *pgdat)
6870{
6871 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6872
6873 clear_bit(nr: LRUVEC_NODE_CONGESTED, addr: &lruvec->flags);
6874 clear_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &lruvec->flags);
6875 clear_bit(nr: PGDAT_DIRTY, addr: &pgdat->flags);
6876 clear_bit(nr: PGDAT_WRITEBACK, addr: &pgdat->flags);
6877}
6878
6879/*
6880 * Prepare kswapd for sleeping. This verifies that there are no processes
6881 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6882 *
6883 * Returns true if kswapd is ready to sleep
6884 */
6885static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6886 int highest_zoneidx)
6887{
6888 /*
6889 * The throttled processes are normally woken up in balance_pgdat() as
6890 * soon as allow_direct_reclaim() is true. But there is a potential
6891 * race between when kswapd checks the watermarks and a process gets
6892 * throttled. There is also a potential race if processes get
6893 * throttled, kswapd wakes, a large process exits thereby balancing the
6894 * zones, which causes kswapd to exit balance_pgdat() before reaching
6895 * the wake up checks. If kswapd is going to sleep, no process should
6896 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6897 * the wake up is premature, processes will wake kswapd and get
6898 * throttled again. The difference from wake ups in balance_pgdat() is
6899 * that here we are under prepare_to_wait().
6900 */
6901 if (waitqueue_active(wq_head: &pgdat->pfmemalloc_wait))
6902 wake_up_all(&pgdat->pfmemalloc_wait);
6903
6904 /* Hopeless node, leave it to direct reclaim */
6905 if (atomic_read(v: &pgdat->kswapd_failures) >= MAX_RECLAIM_RETRIES)
6906 return true;
6907
6908 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6909 clear_pgdat_congested(pgdat);
6910 return true;
6911 }
6912
6913 return false;
6914}
6915
6916/*
6917 * kswapd shrinks a node of pages that are at or below the highest usable
6918 * zone that is currently unbalanced.
6919 *
6920 * Returns true if kswapd scanned at least the requested number of pages to
6921 * reclaim or if the lack of progress was due to pages under writeback.
6922 * This is used to determine if the scanning priority needs to be raised.
6923 */
6924static bool kswapd_shrink_node(pg_data_t *pgdat,
6925 struct scan_control *sc)
6926{
6927 struct zone *zone;
6928 int z;
6929 unsigned long nr_reclaimed = sc->nr_reclaimed;
6930
6931 /* Reclaim a number of pages proportional to the number of zones */
6932 sc->nr_to_reclaim = 0;
6933 for_each_managed_zone_pgdat(zone, pgdat, z, sc->reclaim_idx) {
6934 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6935 }
6936
6937 /*
6938 * Historically care was taken to put equal pressure on all zones but
6939 * now pressure is applied based on node LRU order.
6940 */
6941 shrink_node(pgdat, sc);
6942
6943 /*
6944 * Fragmentation may mean that the system cannot be rebalanced for
6945 * high-order allocations. If twice the allocation size has been
6946 * reclaimed then recheck watermarks only at order-0 to prevent
6947 * excessive reclaim. Assume that a process requested a high-order
6948 * can direct reclaim/compact.
6949 */
6950 if (sc->order && sc->nr_reclaimed >= compact_gap(order: sc->order))
6951 sc->order = 0;
6952
6953 /* account for progress from mm_account_reclaimed_pages() */
6954 return max(sc->nr_scanned, sc->nr_reclaimed - nr_reclaimed) >= sc->nr_to_reclaim;
6955}
6956
6957/* Page allocator PCP high watermark is lowered if reclaim is active. */
6958static inline void
6959update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6960{
6961 int i;
6962 struct zone *zone;
6963
6964 for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
6965 if (active)
6966 set_bit(nr: ZONE_RECLAIM_ACTIVE, addr: &zone->flags);
6967 else
6968 clear_bit(nr: ZONE_RECLAIM_ACTIVE, addr: &zone->flags);
6969 }
6970}
6971
6972static inline void
6973set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6974{
6975 update_reclaim_active(pgdat, highest_zoneidx, active: true);
6976}
6977
6978static inline void
6979clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6980{
6981 update_reclaim_active(pgdat, highest_zoneidx, active: false);
6982}
6983
6984/*
6985 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6986 * that are eligible for use by the caller until at least one zone is
6987 * balanced.
6988 *
6989 * Returns the order kswapd finished reclaiming at.
6990 *
6991 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6992 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6993 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6994 * or lower is eligible for reclaim until at least one usable zone is
6995 * balanced.
6996 */
6997static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6998{
6999 int i;
7000 unsigned long nr_soft_reclaimed;
7001 unsigned long nr_soft_scanned;
7002 unsigned long pflags;
7003 unsigned long nr_boost_reclaim;
7004 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
7005 bool boosted;
7006 struct zone *zone;
7007 struct scan_control sc = {
7008 .gfp_mask = GFP_KERNEL,
7009 .order = order,
7010 .may_unmap = 1,
7011 };
7012
7013 set_task_reclaim_state(current, rs: &sc.reclaim_state);
7014 psi_memstall_enter(flags: &pflags);
7015 __fs_reclaim_acquire(_THIS_IP_);
7016
7017 count_vm_event(item: PAGEOUTRUN);
7018
7019 /*
7020 * Account for the reclaim boost. Note that the zone boost is left in
7021 * place so that parallel allocations that are near the watermark will
7022 * stall or direct reclaim until kswapd is finished.
7023 */
7024 nr_boost_reclaim = 0;
7025 for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
7026 nr_boost_reclaim += zone->watermark_boost;
7027 zone_boosts[i] = zone->watermark_boost;
7028 }
7029 boosted = nr_boost_reclaim;
7030
7031restart:
7032 set_reclaim_active(pgdat, highest_zoneidx);
7033 sc.priority = DEF_PRIORITY;
7034 do {
7035 unsigned long nr_reclaimed = sc.nr_reclaimed;
7036 bool raise_priority = true;
7037 bool balanced;
7038 bool ret;
7039 bool was_frozen;
7040
7041 sc.reclaim_idx = highest_zoneidx;
7042
7043 /*
7044 * If the number of buffer_heads exceeds the maximum allowed
7045 * then consider reclaiming from all zones. This has a dual
7046 * purpose -- on 64-bit systems it is expected that
7047 * buffer_heads are stripped during active rotation. On 32-bit
7048 * systems, highmem pages can pin lowmem memory and shrinking
7049 * buffers can relieve lowmem pressure. Reclaim may still not
7050 * go ahead if all eligible zones for the original allocation
7051 * request are balanced to avoid excessive reclaim from kswapd.
7052 */
7053 if (buffer_heads_over_limit) {
7054 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7055 zone = pgdat->node_zones + i;
7056 if (!managed_zone(zone))
7057 continue;
7058
7059 sc.reclaim_idx = i;
7060 break;
7061 }
7062 }
7063
7064 /*
7065 * If the pgdat is imbalanced then ignore boosting and preserve
7066 * the watermarks for a later time and restart. Note that the
7067 * zone watermarks will be still reset at the end of balancing
7068 * on the grounds that the normal reclaim should be enough to
7069 * re-evaluate if boosting is required when kswapd next wakes.
7070 */
7071 balanced = pgdat_balanced(pgdat, order: sc.order, highest_zoneidx);
7072 if (!balanced && nr_boost_reclaim) {
7073 nr_boost_reclaim = 0;
7074 goto restart;
7075 }
7076
7077 /*
7078 * If boosting is not active then only reclaim if there are no
7079 * eligible zones. Note that sc.reclaim_idx is not used as
7080 * buffer_heads_over_limit may have adjusted it.
7081 */
7082 if (!nr_boost_reclaim && balanced)
7083 goto out;
7084
7085 /* Limit the priority of boosting to avoid reclaim writeback */
7086 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7087 raise_priority = false;
7088
7089 /*
7090 * Do not writeback or swap pages for boosted reclaim. The
7091 * intent is to relieve pressure not issue sub-optimal IO
7092 * from reclaim context. If no pages are reclaimed, the
7093 * reclaim will be aborted.
7094 */
7095 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7096 sc.may_swap = !nr_boost_reclaim;
7097
7098 /*
7099 * Do some background aging, to give pages a chance to be
7100 * referenced before reclaiming. All pages are rotated
7101 * regardless of classzone as this is about consistent aging.
7102 */
7103 kswapd_age_node(pgdat, sc: &sc);
7104
7105 /*
7106 * If we're getting trouble reclaiming, start doing writepage
7107 * even in laptop mode.
7108 */
7109 if (sc.priority < DEF_PRIORITY - 2)
7110 sc.may_writepage = 1;
7111
7112 /* Call soft limit reclaim before calling shrink_node. */
7113 sc.nr_scanned = 0;
7114 nr_soft_scanned = 0;
7115 nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat, order: sc.order,
7116 gfp_mask: sc.gfp_mask, total_scanned: &nr_soft_scanned);
7117 sc.nr_reclaimed += nr_soft_reclaimed;
7118
7119 /*
7120 * There should be no need to raise the scanning priority if
7121 * enough pages are already being scanned that that high
7122 * watermark would be met at 100% efficiency.
7123 */
7124 if (kswapd_shrink_node(pgdat, sc: &sc))
7125 raise_priority = false;
7126
7127 /*
7128 * If the low watermark is met there is no need for processes
7129 * to be throttled on pfmemalloc_wait as they should not be
7130 * able to safely make forward progress. Wake them
7131 */
7132 if (waitqueue_active(wq_head: &pgdat->pfmemalloc_wait) &&
7133 allow_direct_reclaim(pgdat))
7134 wake_up_all(&pgdat->pfmemalloc_wait);
7135
7136 /* Check if kswapd should be suspending */
7137 __fs_reclaim_release(_THIS_IP_);
7138 ret = kthread_freezable_should_stop(was_frozen: &was_frozen);
7139 __fs_reclaim_acquire(_THIS_IP_);
7140 if (was_frozen || ret)
7141 break;
7142
7143 /*
7144 * Raise priority if scanning rate is too low or there was no
7145 * progress in reclaiming pages
7146 */
7147 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7148 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7149
7150 /*
7151 * If reclaim made no progress for a boost, stop reclaim as
7152 * IO cannot be queued and it could be an infinite loop in
7153 * extreme circumstances.
7154 */
7155 if (nr_boost_reclaim && !nr_reclaimed)
7156 break;
7157
7158 if (raise_priority || !nr_reclaimed)
7159 sc.priority--;
7160 } while (sc.priority >= 1);
7161
7162 /*
7163 * Restart only if it went through the priority loop all the way,
7164 * but cache_trim_mode didn't work.
7165 */
7166 if (!sc.nr_reclaimed && sc.priority < 1 &&
7167 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
7168 sc.no_cache_trim_mode = 1;
7169 goto restart;
7170 }
7171
7172 if (!sc.nr_reclaimed)
7173 atomic_inc(v: &pgdat->kswapd_failures);
7174
7175out:
7176 clear_reclaim_active(pgdat, highest_zoneidx);
7177
7178 /* If reclaim was boosted, account for the reclaim done in this pass */
7179 if (boosted) {
7180 unsigned long flags;
7181
7182 for (i = 0; i <= highest_zoneidx; i++) {
7183 if (!zone_boosts[i])
7184 continue;
7185
7186 /* Increments are under the zone lock */
7187 zone = pgdat->node_zones + i;
7188 spin_lock_irqsave(&zone->lock, flags);
7189 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7190 spin_unlock_irqrestore(lock: &zone->lock, flags);
7191 }
7192
7193 /*
7194 * As there is now likely space, wakeup kcompact to defragment
7195 * pageblocks.
7196 */
7197 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7198 }
7199
7200 snapshot_refaults(NULL, pgdat);
7201 __fs_reclaim_release(_THIS_IP_);
7202 psi_memstall_leave(flags: &pflags);
7203 set_task_reclaim_state(current, NULL);
7204
7205 /*
7206 * Return the order kswapd stopped reclaiming at as
7207 * prepare_kswapd_sleep() takes it into account. If another caller
7208 * entered the allocator slow path while kswapd was awake, order will
7209 * remain at the higher level.
7210 */
7211 return sc.order;
7212}
7213
7214/*
7215 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7216 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7217 * not a valid index then either kswapd runs for first time or kswapd couldn't
7218 * sleep after previous reclaim attempt (node is still unbalanced). In that
7219 * case return the zone index of the previous kswapd reclaim cycle.
7220 */
7221static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7222 enum zone_type prev_highest_zoneidx)
7223{
7224 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7225
7226 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7227}
7228
7229static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7230 unsigned int highest_zoneidx)
7231{
7232 long remaining = 0;
7233 DEFINE_WAIT(wait);
7234
7235 if (freezing(current) || kthread_should_stop())
7236 return;
7237
7238 prepare_to_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait, TASK_INTERRUPTIBLE);
7239
7240 /*
7241 * Try to sleep for a short interval. Note that kcompactd will only be
7242 * woken if it is possible to sleep for a short interval. This is
7243 * deliberate on the assumption that if reclaim cannot keep an
7244 * eligible zone balanced that it's also unlikely that compaction will
7245 * succeed.
7246 */
7247 if (prepare_kswapd_sleep(pgdat, order: reclaim_order, highest_zoneidx)) {
7248 /*
7249 * Compaction records what page blocks it recently failed to
7250 * isolate pages from and skips them in the future scanning.
7251 * When kswapd is going to sleep, it is reasonable to assume
7252 * that pages and compaction may succeed so reset the cache.
7253 */
7254 reset_isolation_suitable(pgdat);
7255
7256 /*
7257 * We have freed the memory, now we should compact it to make
7258 * allocation of the requested order possible.
7259 */
7260 wakeup_kcompactd(pgdat, order: alloc_order, highest_zoneidx);
7261
7262 remaining = schedule_timeout(HZ/10);
7263
7264 /*
7265 * If woken prematurely then reset kswapd_highest_zoneidx and
7266 * order. The values will either be from a wakeup request or
7267 * the previous request that slept prematurely.
7268 */
7269 if (remaining) {
7270 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7271 kswapd_highest_zoneidx(pgdat,
7272 highest_zoneidx));
7273
7274 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7275 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7276 }
7277
7278 finish_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait);
7279 prepare_to_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait, TASK_INTERRUPTIBLE);
7280 }
7281
7282 /*
7283 * After a short sleep, check if it was a premature sleep. If not, then
7284 * go fully to sleep until explicitly woken up.
7285 */
7286 if (!remaining &&
7287 prepare_kswapd_sleep(pgdat, order: reclaim_order, highest_zoneidx)) {
7288 trace_mm_vmscan_kswapd_sleep(nid: pgdat->node_id);
7289
7290 /*
7291 * vmstat counters are not perfectly accurate and the estimated
7292 * value for counters such as NR_FREE_PAGES can deviate from the
7293 * true value by nr_online_cpus * threshold. To avoid the zone
7294 * watermarks being breached while under pressure, we reduce the
7295 * per-cpu vmstat threshold while kswapd is awake and restore
7296 * them before going back to sleep.
7297 */
7298 set_pgdat_percpu_threshold(pgdat, calculate_pressure: calculate_normal_threshold);
7299
7300 if (!kthread_should_stop())
7301 schedule();
7302
7303 set_pgdat_percpu_threshold(pgdat, calculate_pressure: calculate_pressure_threshold);
7304 } else {
7305 if (remaining)
7306 count_vm_event(item: KSWAPD_LOW_WMARK_HIT_QUICKLY);
7307 else
7308 count_vm_event(item: KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7309 }
7310 finish_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait);
7311}
7312
7313/*
7314 * The background pageout daemon, started as a kernel thread
7315 * from the init process.
7316 *
7317 * This basically trickles out pages so that we have _some_
7318 * free memory available even if there is no other activity
7319 * that frees anything up. This is needed for things like routing
7320 * etc, where we otherwise might have all activity going on in
7321 * asynchronous contexts that cannot page things out.
7322 *
7323 * If there are applications that are active memory-allocators
7324 * (most normal use), this basically shouldn't matter.
7325 */
7326static int kswapd(void *p)
7327{
7328 unsigned int alloc_order, reclaim_order;
7329 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7330 pg_data_t *pgdat = (pg_data_t *)p;
7331 struct task_struct *tsk = current;
7332
7333 /*
7334 * Tell the memory management that we're a "memory allocator",
7335 * and that if we need more memory we should get access to it
7336 * regardless (see "__alloc_pages()"). "kswapd" should
7337 * never get caught in the normal page freeing logic.
7338 *
7339 * (Kswapd normally doesn't need memory anyway, but sometimes
7340 * you need a small amount of memory in order to be able to
7341 * page out something else, and this flag essentially protects
7342 * us from recursively trying to free more memory as we're
7343 * trying to free the first piece of memory in the first place).
7344 */
7345 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7346 set_freezable();
7347
7348 WRITE_ONCE(pgdat->kswapd_order, 0);
7349 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7350 atomic_set(v: &pgdat->nr_writeback_throttled, i: 0);
7351 for ( ; ; ) {
7352 bool was_frozen;
7353
7354 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7355 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7356 prev_highest_zoneidx: highest_zoneidx);
7357
7358kswapd_try_sleep:
7359 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7360 highest_zoneidx);
7361
7362 /* Read the new order and highest_zoneidx */
7363 alloc_order = READ_ONCE(pgdat->kswapd_order);
7364 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7365 prev_highest_zoneidx: highest_zoneidx);
7366 WRITE_ONCE(pgdat->kswapd_order, 0);
7367 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7368
7369 if (kthread_freezable_should_stop(was_frozen: &was_frozen))
7370 break;
7371
7372 /*
7373 * We can speed up thawing tasks if we don't call balance_pgdat
7374 * after returning from the refrigerator
7375 */
7376 if (was_frozen)
7377 continue;
7378
7379 /*
7380 * Reclaim begins at the requested order but if a high-order
7381 * reclaim fails then kswapd falls back to reclaiming for
7382 * order-0. If that happens, kswapd will consider sleeping
7383 * for the order it finished reclaiming at (reclaim_order)
7384 * but kcompactd is woken to compact for the original
7385 * request (alloc_order).
7386 */
7387 trace_mm_vmscan_kswapd_wake(nid: pgdat->node_id, zid: highest_zoneidx,
7388 order: alloc_order);
7389 reclaim_order = balance_pgdat(pgdat, order: alloc_order,
7390 highest_zoneidx);
7391 if (reclaim_order < alloc_order)
7392 goto kswapd_try_sleep;
7393 }
7394
7395 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7396
7397 return 0;
7398}
7399
7400/*
7401 * A zone is low on free memory or too fragmented for high-order memory. If
7402 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7403 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7404 * has failed or is not needed, still wake up kcompactd if only compaction is
7405 * needed.
7406 */
7407void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7408 enum zone_type highest_zoneidx)
7409{
7410 pg_data_t *pgdat;
7411 enum zone_type curr_idx;
7412
7413 if (!managed_zone(zone))
7414 return;
7415
7416 if (!cpuset_zone_allowed(z: zone, gfp_mask: gfp_flags))
7417 return;
7418
7419 pgdat = zone->zone_pgdat;
7420 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7421
7422 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7423 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7424
7425 if (READ_ONCE(pgdat->kswapd_order) < order)
7426 WRITE_ONCE(pgdat->kswapd_order, order);
7427
7428 if (!waitqueue_active(wq_head: &pgdat->kswapd_wait))
7429 return;
7430
7431 /* Hopeless node, leave it to direct reclaim if possible */
7432 if (atomic_read(v: &pgdat->kswapd_failures) >= MAX_RECLAIM_RETRIES ||
7433 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7434 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7435 /*
7436 * There may be plenty of free memory available, but it's too
7437 * fragmented for high-order allocations. Wake up kcompactd
7438 * and rely on compaction_suitable() to determine if it's
7439 * needed. If it fails, it will defer subsequent attempts to
7440 * ratelimit its work.
7441 */
7442 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7443 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7444 return;
7445 }
7446
7447 trace_mm_vmscan_wakeup_kswapd(nid: pgdat->node_id, zid: highest_zoneidx, order,
7448 gfp_flags);
7449 wake_up_interruptible(&pgdat->kswapd_wait);
7450}
7451
7452#ifdef CONFIG_HIBERNATION
7453/*
7454 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7455 * freed pages.
7456 *
7457 * Rather than trying to age LRUs the aim is to preserve the overall
7458 * LRU order by reclaiming preferentially
7459 * inactive > active > active referenced > active mapped
7460 */
7461unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7462{
7463 struct scan_control sc = {
7464 .nr_to_reclaim = nr_to_reclaim,
7465 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7466 .reclaim_idx = MAX_NR_ZONES - 1,
7467 .priority = DEF_PRIORITY,
7468 .may_writepage = 1,
7469 .may_unmap = 1,
7470 .may_swap = 1,
7471 .hibernation_mode = 1,
7472 };
7473 struct zonelist *zonelist = node_zonelist(nid: numa_node_id(), flags: sc.gfp_mask);
7474 unsigned long nr_reclaimed;
7475 unsigned int noreclaim_flag;
7476
7477 fs_reclaim_acquire(gfp_mask: sc.gfp_mask);
7478 noreclaim_flag = memalloc_noreclaim_save();
7479 set_task_reclaim_state(current, rs: &sc.reclaim_state);
7480
7481 nr_reclaimed = do_try_to_free_pages(zonelist, sc: &sc);
7482
7483 set_task_reclaim_state(current, NULL);
7484 memalloc_noreclaim_restore(flags: noreclaim_flag);
7485 fs_reclaim_release(gfp_mask: sc.gfp_mask);
7486
7487 return nr_reclaimed;
7488}
7489#endif /* CONFIG_HIBERNATION */
7490
7491/*
7492 * This kswapd start function will be called by init and node-hot-add.
7493 */
7494void __meminit kswapd_run(int nid)
7495{
7496 pg_data_t *pgdat = NODE_DATA(nid);
7497
7498 pgdat_kswapd_lock(pgdat);
7499 if (!pgdat->kswapd) {
7500 pgdat->kswapd = kthread_create_on_node(threadfn: kswapd, data: pgdat, node: nid, namefmt: "kswapd%d", nid);
7501 if (IS_ERR(ptr: pgdat->kswapd)) {
7502 /* failure at boot is fatal */
7503 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7504 nid, PTR_ERR(pgdat->kswapd));
7505 BUG_ON(system_state < SYSTEM_RUNNING);
7506 pgdat->kswapd = NULL;
7507 } else {
7508 wake_up_process(tsk: pgdat->kswapd);
7509 }
7510 }
7511 pgdat_kswapd_unlock(pgdat);
7512}
7513
7514/*
7515 * Called by memory hotplug when all memory in a node is offlined. Caller must
7516 * be holding mem_hotplug_begin/done().
7517 */
7518void __meminit kswapd_stop(int nid)
7519{
7520 pg_data_t *pgdat = NODE_DATA(nid);
7521 struct task_struct *kswapd;
7522
7523 pgdat_kswapd_lock(pgdat);
7524 kswapd = pgdat->kswapd;
7525 if (kswapd) {
7526 kthread_stop(k: kswapd);
7527 pgdat->kswapd = NULL;
7528 }
7529 pgdat_kswapd_unlock(pgdat);
7530}
7531
7532static const struct ctl_table vmscan_sysctl_table[] = {
7533 {
7534 .procname = "swappiness",
7535 .data = &vm_swappiness,
7536 .maxlen = sizeof(vm_swappiness),
7537 .mode = 0644,
7538 .proc_handler = proc_dointvec_minmax,
7539 .extra1 = SYSCTL_ZERO,
7540 .extra2 = SYSCTL_TWO_HUNDRED,
7541 },
7542#ifdef CONFIG_NUMA
7543 {
7544 .procname = "zone_reclaim_mode",
7545 .data = &node_reclaim_mode,
7546 .maxlen = sizeof(node_reclaim_mode),
7547 .mode = 0644,
7548 .proc_handler = proc_dointvec_minmax,
7549 .extra1 = SYSCTL_ZERO,
7550 }
7551#endif
7552};
7553
7554static int __init kswapd_init(void)
7555{
7556 int nid;
7557
7558 swap_setup();
7559 for_each_node_state(nid, N_MEMORY)
7560 kswapd_run(nid);
7561 register_sysctl_init("vm", vmscan_sysctl_table);
7562 return 0;
7563}
7564
7565module_init(kswapd_init)
7566
7567#ifdef CONFIG_NUMA
7568/*
7569 * Node reclaim mode
7570 *
7571 * If non-zero call node_reclaim when the number of free pages falls below
7572 * the watermarks.
7573 */
7574int node_reclaim_mode __read_mostly;
7575
7576/*
7577 * Priority for NODE_RECLAIM. This determines the fraction of pages
7578 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7579 * a zone.
7580 */
7581#define NODE_RECLAIM_PRIORITY 4
7582
7583/*
7584 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7585 * occur.
7586 */
7587int sysctl_min_unmapped_ratio = 1;
7588
7589/*
7590 * If the number of slab pages in a zone grows beyond this percentage then
7591 * slab reclaim needs to occur.
7592 */
7593int sysctl_min_slab_ratio = 5;
7594
7595static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7596{
7597 unsigned long file_mapped = node_page_state(pgdat, item: NR_FILE_MAPPED);
7598 unsigned long file_lru = node_page_state(pgdat, item: NR_INACTIVE_FILE) +
7599 node_page_state(pgdat, item: NR_ACTIVE_FILE);
7600
7601 /*
7602 * It's possible for there to be more file mapped pages than
7603 * accounted for by the pages on the file LRU lists because
7604 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7605 */
7606 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7607}
7608
7609/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7610static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7611{
7612 unsigned long nr_pagecache_reclaimable;
7613 unsigned long delta = 0;
7614
7615 /*
7616 * If RECLAIM_UNMAP is set, then all file pages are considered
7617 * potentially reclaimable. Otherwise, we have to worry about
7618 * pages like swapcache and node_unmapped_file_pages() provides
7619 * a better estimate
7620 */
7621 if (node_reclaim_mode & RECLAIM_UNMAP)
7622 nr_pagecache_reclaimable = node_page_state(pgdat, item: NR_FILE_PAGES);
7623 else
7624 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7625
7626 /* If we can't clean pages, remove dirty pages from consideration */
7627 if (!(node_reclaim_mode & RECLAIM_WRITE))
7628 delta += node_page_state(pgdat, item: NR_FILE_DIRTY);
7629
7630 /* Watch for any possible underflows due to delta */
7631 if (unlikely(delta > nr_pagecache_reclaimable))
7632 delta = nr_pagecache_reclaimable;
7633
7634 return nr_pagecache_reclaimable - delta;
7635}
7636
7637/*
7638 * Try to free up some pages from this node through reclaim.
7639 */
7640static unsigned long __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask,
7641 unsigned long nr_pages,
7642 struct scan_control *sc)
7643{
7644 struct task_struct *p = current;
7645 unsigned int noreclaim_flag;
7646 unsigned long pflags;
7647
7648 trace_mm_vmscan_node_reclaim_begin(nid: pgdat->node_id, order: sc->order,
7649 gfp_flags: sc->gfp_mask);
7650
7651 cond_resched();
7652 psi_memstall_enter(flags: &pflags);
7653 delayacct_freepages_start();
7654 fs_reclaim_acquire(gfp_mask: sc->gfp_mask);
7655 /*
7656 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7657 */
7658 noreclaim_flag = memalloc_noreclaim_save();
7659 set_task_reclaim_state(task: p, rs: &sc->reclaim_state);
7660
7661 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7662 node_page_state_pages(pgdat, item: NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7663 /*
7664 * Free memory by calling shrink node with increasing
7665 * priorities until we have enough memory freed.
7666 */
7667 do {
7668 shrink_node(pgdat, sc);
7669 } while (sc->nr_reclaimed < nr_pages && --sc->priority >= 0);
7670 }
7671
7672 set_task_reclaim_state(task: p, NULL);
7673 memalloc_noreclaim_restore(flags: noreclaim_flag);
7674 fs_reclaim_release(gfp_mask: sc->gfp_mask);
7675 delayacct_freepages_end();
7676 psi_memstall_leave(flags: &pflags);
7677
7678 trace_mm_vmscan_node_reclaim_end(nr_reclaimed: sc->nr_reclaimed);
7679
7680 return sc->nr_reclaimed;
7681}
7682
7683int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7684{
7685 int ret;
7686 /* Minimum pages needed in order to stay on node */
7687 const unsigned long nr_pages = 1 << order;
7688 struct scan_control sc = {
7689 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7690 .gfp_mask = current_gfp_context(flags: gfp_mask),
7691 .order = order,
7692 .priority = NODE_RECLAIM_PRIORITY,
7693 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7694 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7695 .may_swap = 1,
7696 .reclaim_idx = gfp_zone(flags: gfp_mask),
7697 };
7698
7699 /*
7700 * Node reclaim reclaims unmapped file backed pages and
7701 * slab pages if we are over the defined limits.
7702 *
7703 * A small portion of unmapped file backed pages is needed for
7704 * file I/O otherwise pages read by file I/O will be immediately
7705 * thrown out if the node is overallocated. So we do not reclaim
7706 * if less than a specified percentage of the node is used by
7707 * unmapped file backed pages.
7708 */
7709 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7710 node_page_state_pages(pgdat, item: NR_SLAB_RECLAIMABLE_B) <=
7711 pgdat->min_slab_pages)
7712 return NODE_RECLAIM_FULL;
7713
7714 /*
7715 * Do not scan if the allocation should not be delayed.
7716 */
7717 if (!gfpflags_allow_blocking(gfp_flags: gfp_mask) || (current->flags & PF_MEMALLOC))
7718 return NODE_RECLAIM_NOSCAN;
7719
7720 /*
7721 * Only run node reclaim on the local node or on nodes that do not
7722 * have associated processors. This will favor the local processor
7723 * over remote processors and spread off node memory allocations
7724 * as wide as possible.
7725 */
7726 if (node_state(node: pgdat->node_id, state: N_CPU) && pgdat->node_id != numa_node_id())
7727 return NODE_RECLAIM_NOSCAN;
7728
7729 if (test_and_set_bit_lock(nr: PGDAT_RECLAIM_LOCKED, addr: &pgdat->flags))
7730 return NODE_RECLAIM_NOSCAN;
7731
7732 ret = __node_reclaim(pgdat, gfp_mask, nr_pages, sc: &sc) >= nr_pages;
7733 clear_bit_unlock(nr: PGDAT_RECLAIM_LOCKED, addr: &pgdat->flags);
7734
7735 if (ret)
7736 count_vm_event(item: PGSCAN_ZONE_RECLAIM_SUCCESS);
7737 else
7738 count_vm_event(item: PGSCAN_ZONE_RECLAIM_FAILED);
7739
7740 return ret;
7741}
7742
7743enum {
7744 MEMORY_RECLAIM_SWAPPINESS = 0,
7745 MEMORY_RECLAIM_SWAPPINESS_MAX,
7746 MEMORY_RECLAIM_NULL,
7747};
7748static const match_table_t tokens = {
7749 { MEMORY_RECLAIM_SWAPPINESS, "swappiness=%d"},
7750 { .token: MEMORY_RECLAIM_SWAPPINESS_MAX, .pattern: "swappiness=max"},
7751 { .token: MEMORY_RECLAIM_NULL, NULL },
7752};
7753
7754int user_proactive_reclaim(char *buf,
7755 struct mem_cgroup *memcg, pg_data_t *pgdat)
7756{
7757 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
7758 unsigned long nr_to_reclaim, nr_reclaimed = 0;
7759 int swappiness = -1;
7760 char *old_buf, *start;
7761 substring_t args[MAX_OPT_ARGS];
7762 gfp_t gfp_mask = GFP_KERNEL;
7763
7764 if (!buf || (!memcg && !pgdat) || (memcg && pgdat))
7765 return -EINVAL;
7766
7767 buf = strstrip(str: buf);
7768
7769 old_buf = buf;
7770 nr_to_reclaim = memparse(ptr: buf, retptr: &buf) / PAGE_SIZE;
7771 if (buf == old_buf)
7772 return -EINVAL;
7773
7774 buf = strstrip(str: buf);
7775
7776 while ((start = strsep(&buf, " ")) != NULL) {
7777 if (!strlen(start))
7778 continue;
7779 switch (match_token(start, table: tokens, args)) {
7780 case MEMORY_RECLAIM_SWAPPINESS:
7781 if (match_int(&args[0], result: &swappiness))
7782 return -EINVAL;
7783 if (swappiness < MIN_SWAPPINESS ||
7784 swappiness > MAX_SWAPPINESS)
7785 return -EINVAL;
7786 break;
7787 case MEMORY_RECLAIM_SWAPPINESS_MAX:
7788 swappiness = SWAPPINESS_ANON_ONLY;
7789 break;
7790 default:
7791 return -EINVAL;
7792 }
7793 }
7794
7795 while (nr_reclaimed < nr_to_reclaim) {
7796 /* Will converge on zero, but reclaim enforces a minimum */
7797 unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
7798 unsigned long reclaimed;
7799
7800 if (signal_pending(current))
7801 return -EINTR;
7802
7803 /*
7804 * This is the final attempt, drain percpu lru caches in the
7805 * hope of introducing more evictable pages.
7806 */
7807 if (!nr_retries)
7808 lru_add_drain_all();
7809
7810 if (memcg) {
7811 unsigned int reclaim_options;
7812
7813 reclaim_options = MEMCG_RECLAIM_MAY_SWAP |
7814 MEMCG_RECLAIM_PROACTIVE;
7815 reclaimed = try_to_free_mem_cgroup_pages(memcg,
7816 nr_pages: batch_size, gfp_mask,
7817 reclaim_options,
7818 swappiness: swappiness == -1 ? NULL : &swappiness);
7819 } else {
7820 struct scan_control sc = {
7821 .gfp_mask = current_gfp_context(flags: gfp_mask),
7822 .reclaim_idx = gfp_zone(flags: gfp_mask),
7823 .proactive_swappiness = swappiness == -1 ? NULL : &swappiness,
7824 .priority = DEF_PRIORITY,
7825 .may_writepage = !laptop_mode,
7826 .nr_to_reclaim = max(batch_size, SWAP_CLUSTER_MAX),
7827 .may_unmap = 1,
7828 .may_swap = 1,
7829 .proactive = 1,
7830 };
7831
7832 if (test_and_set_bit_lock(nr: PGDAT_RECLAIM_LOCKED,
7833 addr: &pgdat->flags))
7834 return -EBUSY;
7835
7836 reclaimed = __node_reclaim(pgdat, gfp_mask,
7837 nr_pages: batch_size, sc: &sc);
7838 clear_bit_unlock(nr: PGDAT_RECLAIM_LOCKED, addr: &pgdat->flags);
7839 }
7840
7841 if (!reclaimed && !nr_retries--)
7842 return -EAGAIN;
7843
7844 nr_reclaimed += reclaimed;
7845 }
7846
7847 return 0;
7848}
7849
7850#endif
7851
7852/**
7853 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7854 * lru list
7855 * @fbatch: Batch of lru folios to check.
7856 *
7857 * Checks folios for evictability, if an evictable folio is in the unevictable
7858 * lru list, moves it to the appropriate evictable lru list. This function
7859 * should be only used for lru folios.
7860 */
7861void check_move_unevictable_folios(struct folio_batch *fbatch)
7862{
7863 struct lruvec *lruvec = NULL;
7864 int pgscanned = 0;
7865 int pgrescued = 0;
7866 int i;
7867
7868 for (i = 0; i < fbatch->nr; i++) {
7869 struct folio *folio = fbatch->folios[i];
7870 int nr_pages = folio_nr_pages(folio);
7871
7872 pgscanned += nr_pages;
7873
7874 /* block memcg migration while the folio moves between lrus */
7875 if (!folio_test_clear_lru(folio))
7876 continue;
7877
7878 lruvec = folio_lruvec_relock_irq(folio, locked_lruvec: lruvec);
7879 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7880 lruvec_del_folio(lruvec, folio);
7881 folio_clear_unevictable(folio);
7882 lruvec_add_folio(lruvec, folio);
7883 pgrescued += nr_pages;
7884 }
7885 folio_set_lru(folio);
7886 }
7887
7888 if (lruvec) {
7889 __count_vm_events(item: UNEVICTABLE_PGRESCUED, delta: pgrescued);
7890 __count_vm_events(item: UNEVICTABLE_PGSCANNED, delta: pgscanned);
7891 unlock_page_lruvec_irq(lruvec);
7892 } else if (pgscanned) {
7893 count_vm_events(item: UNEVICTABLE_PGSCANNED, delta: pgscanned);
7894 }
7895}
7896EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
7897
7898#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
7899static ssize_t reclaim_store(struct device *dev,
7900 struct device_attribute *attr,
7901 const char *buf, size_t count)
7902{
7903 int ret, nid = dev->id;
7904
7905 ret = user_proactive_reclaim(buf: (char *)buf, NULL, NODE_DATA(nid));
7906 return ret ? -EAGAIN : count;
7907}
7908
7909static DEVICE_ATTR_WO(reclaim);
7910int reclaim_register_node(struct node *node)
7911{
7912 return device_create_file(device: &node->dev, entry: &dev_attr_reclaim);
7913}
7914
7915void reclaim_unregister_node(struct node *node)
7916{
7917 return device_remove_file(dev: &node->dev, attr: &dev_attr_reclaim);
7918}
7919#endif
7920