Loading...
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
60#include <asm/tlbflush.h>
61#include <asm/div64.h>
62
63#include <linux/swapops.h>
64#include <linux/balloon_compaction.h>
65#include <linux/sched/sysctl.h>
66
67#include "internal.h"
68#include "swap.h"
69
70#define CREATE_TRACE_POINTS
71#include <trace/events/vmscan.h>
72
73struct scan_control {
74 /* How many pages shrink_list() should reclaim */
75 unsigned long nr_to_reclaim;
76
77 /*
78 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * are scanned.
80 */
81 nodemask_t *nodemask;
82
83 /*
84 * The memory cgroup that hit its limit and as a result is the
85 * primary target of this reclaim invocation.
86 */
87 struct mem_cgroup *target_mem_cgroup;
88
89 /*
90 * Scan pressure balancing between anon and file LRUs
91 */
92 unsigned long anon_cost;
93 unsigned long file_cost;
94
95 /* Can active folios be deactivated as part of reclaim? */
96#define DEACTIVATE_ANON 1
97#define DEACTIVATE_FILE 2
98 unsigned int may_deactivate:2;
99 unsigned int force_deactivate:1;
100 unsigned int skipped_deactivate:1;
101
102 /* Writepage batching in laptop mode; RECLAIM_WRITE */
103 unsigned int may_writepage:1;
104
105 /* Can mapped folios be reclaimed? */
106 unsigned int may_unmap:1;
107
108 /* Can folios be swapped as part of reclaim? */
109 unsigned int may_swap:1;
110
111 /* Proactive reclaim invoked by userspace through memory.reclaim */
112 unsigned int proactive:1;
113
114 /*
115 * Cgroup memory below memory.low is protected as long as we
116 * don't threaten to OOM. If any cgroup is reclaimed at
117 * reduced force or passed over entirely due to its memory.low
118 * setting (memcg_low_skipped), and nothing is reclaimed as a
119 * result, then go back for one more cycle that reclaims the protected
120 * memory (memcg_low_reclaim) to avert OOM.
121 */
122 unsigned int memcg_low_reclaim:1;
123 unsigned int memcg_low_skipped:1;
124
125 unsigned int hibernation_mode:1;
126
127 /* One of the zones is ready for compaction */
128 unsigned int compaction_ready:1;
129
130 /* There is easily reclaimable cold cache in the current node */
131 unsigned int cache_trim_mode:1;
132
133 /* The file folios on the current node are dangerously low */
134 unsigned int file_is_tiny:1;
135
136 /* Always discard instead of demoting to lower tier memory */
137 unsigned int no_demotion:1;
138
139 /* Allocation order */
140 s8 order;
141
142 /* Scan (total_size >> priority) pages at once */
143 s8 priority;
144
145 /* The highest zone to isolate folios for reclaim from */
146 s8 reclaim_idx;
147
148 /* This context's GFP mask */
149 gfp_t gfp_mask;
150
151 /* Incremented by the number of inactive pages that were scanned */
152 unsigned long nr_scanned;
153
154 /* Number of pages freed so far during a call to shrink_zones() */
155 unsigned long nr_reclaimed;
156
157 struct {
158 unsigned int dirty;
159 unsigned int unqueued_dirty;
160 unsigned int congested;
161 unsigned int writeback;
162 unsigned int immediate;
163 unsigned int file_taken;
164 unsigned int taken;
165 } nr;
166
167 /* for recording the reclaimed slab by now */
168 struct reclaim_state reclaim_state;
169};
170
171#ifdef ARCH_HAS_PREFETCHW
172#define prefetchw_prev_lru_folio(_folio, _base, _field) \
173 do { \
174 if ((_folio)->lru.prev != _base) { \
175 struct folio *prev; \
176 \
177 prev = lru_to_folio(&(_folio->lru)); \
178 prefetchw(&prev->_field); \
179 } \
180 } while (0)
181#else
182#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
183#endif
184
185/*
186 * From 0 .. 200. Higher means more swappy.
187 */
188int vm_swappiness = 60;
189
190#ifdef CONFIG_MEMCG
191
192/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
193static bool cgroup_reclaim(struct scan_control *sc)
194{
195 return sc->target_mem_cgroup;
196}
197
198/*
199 * Returns true for reclaim on the root cgroup. This is true for direct
200 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
201 */
202static bool root_reclaim(struct scan_control *sc)
203{
204 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
205}
206
207/**
208 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
209 * @sc: scan_control in question
210 *
211 * The normal page dirty throttling mechanism in balance_dirty_pages() is
212 * completely broken with the legacy memcg and direct stalling in
213 * shrink_folio_list() is used for throttling instead, which lacks all the
214 * niceties such as fairness, adaptive pausing, bandwidth proportional
215 * allocation and configurability.
216 *
217 * This function tests whether the vmscan currently in progress can assume
218 * that the normal dirty throttling mechanism is operational.
219 */
220static bool writeback_throttling_sane(struct scan_control *sc)
221{
222 if (!cgroup_reclaim(sc))
223 return true;
224#ifdef CONFIG_CGROUP_WRITEBACK
225 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
226 return true;
227#endif
228 return false;
229}
230#else
231static bool cgroup_reclaim(struct scan_control *sc)
232{
233 return false;
234}
235
236static bool root_reclaim(struct scan_control *sc)
237{
238 return true;
239}
240
241static bool writeback_throttling_sane(struct scan_control *sc)
242{
243 return true;
244}
245#endif
246
247static void set_task_reclaim_state(struct task_struct *task,
248 struct reclaim_state *rs)
249{
250 /* Check for an overwrite */
251 WARN_ON_ONCE(rs && task->reclaim_state);
252
253 /* Check for the nulling of an already-nulled member */
254 WARN_ON_ONCE(!rs && !task->reclaim_state);
255
256 task->reclaim_state = rs;
257}
258
259/*
260 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
261 * scan_control->nr_reclaimed.
262 */
263static void flush_reclaim_state(struct scan_control *sc)
264{
265 /*
266 * Currently, reclaim_state->reclaimed includes three types of pages
267 * freed outside of vmscan:
268 * (1) Slab pages.
269 * (2) Clean file pages from pruned inodes (on highmem systems).
270 * (3) XFS freed buffer pages.
271 *
272 * For all of these cases, we cannot universally link the pages to a
273 * single memcg. For example, a memcg-aware shrinker can free one object
274 * charged to the target memcg, causing an entire page to be freed.
275 * If we count the entire page as reclaimed from the memcg, we end up
276 * overestimating the reclaimed amount (potentially under-reclaiming).
277 *
278 * Only count such pages for global reclaim to prevent under-reclaiming
279 * from the target memcg; preventing unnecessary retries during memcg
280 * charging and false positives from proactive reclaim.
281 *
282 * For uncommon cases where the freed pages were actually mostly
283 * charged to the target memcg, we end up underestimating the reclaimed
284 * amount. This should be fine. The freed pages will be uncharged
285 * anyway, even if they are not counted here properly, and we will be
286 * able to make forward progress in charging (which is usually in a
287 * retry loop).
288 *
289 * We can go one step further, and report the uncharged objcg pages in
290 * memcg reclaim, to make reporting more accurate and reduce
291 * underestimation, but it's probably not worth the complexity for now.
292 */
293 if (current->reclaim_state && root_reclaim(sc)) {
294 sc->nr_reclaimed += current->reclaim_state->reclaimed;
295 current->reclaim_state->reclaimed = 0;
296 }
297}
298
299static bool can_demote(int nid, struct scan_control *sc)
300{
301 if (!numa_demotion_enabled)
302 return false;
303 if (sc && sc->no_demotion)
304 return false;
305 if (next_demotion_node(nid) == NUMA_NO_NODE)
306 return false;
307
308 return true;
309}
310
311static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
312 int nid,
313 struct scan_control *sc)
314{
315 if (memcg == NULL) {
316 /*
317 * For non-memcg reclaim, is there
318 * space in any swap device?
319 */
320 if (get_nr_swap_pages() > 0)
321 return true;
322 } else {
323 /* Is the memcg below its swap limit? */
324 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
325 return true;
326 }
327
328 /*
329 * The page can not be swapped.
330 *
331 * Can it be reclaimed from this node via demotion?
332 */
333 return can_demote(nid, sc);
334}
335
336/*
337 * This misses isolated folios which are not accounted for to save counters.
338 * As the data only determines if reclaim or compaction continues, it is
339 * not expected that isolated folios will be a dominating factor.
340 */
341unsigned long zone_reclaimable_pages(struct zone *zone)
342{
343 unsigned long nr;
344
345 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
346 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
347 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
348 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
349 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
350
351 return nr;
352}
353
354/**
355 * lruvec_lru_size - Returns the number of pages on the given LRU list.
356 * @lruvec: lru vector
357 * @lru: lru to use
358 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
359 */
360static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
361 int zone_idx)
362{
363 unsigned long size = 0;
364 int zid;
365
366 for (zid = 0; zid <= zone_idx; zid++) {
367 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
368
369 if (!managed_zone(zone))
370 continue;
371
372 if (!mem_cgroup_disabled())
373 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
374 else
375 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
376 }
377 return size;
378}
379
380static unsigned long drop_slab_node(int nid)
381{
382 unsigned long freed = 0;
383 struct mem_cgroup *memcg = NULL;
384
385 memcg = mem_cgroup_iter(NULL, NULL, NULL);
386 do {
387 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
388 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
389
390 return freed;
391}
392
393void drop_slab(void)
394{
395 int nid;
396 int shift = 0;
397 unsigned long freed;
398
399 do {
400 freed = 0;
401 for_each_online_node(nid) {
402 if (fatal_signal_pending(current))
403 return;
404
405 freed += drop_slab_node(nid);
406 }
407 } while ((freed >> shift++) > 1);
408}
409
410static int reclaimer_offset(void)
411{
412 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
413 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
414 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
415 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
416 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
417 PGSCAN_DIRECT - PGSCAN_KSWAPD);
418 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
419 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
420
421 if (current_is_kswapd())
422 return 0;
423 if (current_is_khugepaged())
424 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
425 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
426}
427
428static inline int is_page_cache_freeable(struct folio *folio)
429{
430 /*
431 * A freeable page cache folio is referenced only by the caller
432 * that isolated the folio, the page cache and optional filesystem
433 * private data at folio->private.
434 */
435 return folio_ref_count(folio) - folio_test_private(folio) ==
436 1 + folio_nr_pages(folio);
437}
438
439/*
440 * We detected a synchronous write error writing a folio out. Probably
441 * -ENOSPC. We need to propagate that into the address_space for a subsequent
442 * fsync(), msync() or close().
443 *
444 * The tricky part is that after writepage we cannot touch the mapping: nothing
445 * prevents it from being freed up. But we have a ref on the folio and once
446 * that folio is locked, the mapping is pinned.
447 *
448 * We're allowed to run sleeping folio_lock() here because we know the caller has
449 * __GFP_FS.
450 */
451static void handle_write_error(struct address_space *mapping,
452 struct folio *folio, int error)
453{
454 folio_lock(folio);
455 if (folio_mapping(folio) == mapping)
456 mapping_set_error(mapping, error);
457 folio_unlock(folio);
458}
459
460static bool skip_throttle_noprogress(pg_data_t *pgdat)
461{
462 int reclaimable = 0, write_pending = 0;
463 int i;
464
465 /*
466 * If kswapd is disabled, reschedule if necessary but do not
467 * throttle as the system is likely near OOM.
468 */
469 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
470 return true;
471
472 /*
473 * If there are a lot of dirty/writeback folios then do not
474 * throttle as throttling will occur when the folios cycle
475 * towards the end of the LRU if still under writeback.
476 */
477 for (i = 0; i < MAX_NR_ZONES; i++) {
478 struct zone *zone = pgdat->node_zones + i;
479
480 if (!managed_zone(zone))
481 continue;
482
483 reclaimable += zone_reclaimable_pages(zone);
484 write_pending += zone_page_state_snapshot(zone,
485 NR_ZONE_WRITE_PENDING);
486 }
487 if (2 * write_pending <= reclaimable)
488 return true;
489
490 return false;
491}
492
493void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
494{
495 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
496 long timeout, ret;
497 DEFINE_WAIT(wait);
498
499 /*
500 * Do not throttle user workers, kthreads other than kswapd or
501 * workqueues. They may be required for reclaim to make
502 * forward progress (e.g. journalling workqueues or kthreads).
503 */
504 if (!current_is_kswapd() &&
505 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
506 cond_resched();
507 return;
508 }
509
510 /*
511 * These figures are pulled out of thin air.
512 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
513 * parallel reclaimers which is a short-lived event so the timeout is
514 * short. Failing to make progress or waiting on writeback are
515 * potentially long-lived events so use a longer timeout. This is shaky
516 * logic as a failure to make progress could be due to anything from
517 * writeback to a slow device to excessive referenced folios at the tail
518 * of the inactive LRU.
519 */
520 switch(reason) {
521 case VMSCAN_THROTTLE_WRITEBACK:
522 timeout = HZ/10;
523
524 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
525 WRITE_ONCE(pgdat->nr_reclaim_start,
526 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
527 }
528
529 break;
530 case VMSCAN_THROTTLE_CONGESTED:
531 fallthrough;
532 case VMSCAN_THROTTLE_NOPROGRESS:
533 if (skip_throttle_noprogress(pgdat)) {
534 cond_resched();
535 return;
536 }
537
538 timeout = 1;
539
540 break;
541 case VMSCAN_THROTTLE_ISOLATED:
542 timeout = HZ/50;
543 break;
544 default:
545 WARN_ON_ONCE(1);
546 timeout = HZ;
547 break;
548 }
549
550 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
551 ret = schedule_timeout(timeout);
552 finish_wait(wqh, &wait);
553
554 if (reason == VMSCAN_THROTTLE_WRITEBACK)
555 atomic_dec(&pgdat->nr_writeback_throttled);
556
557 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
558 jiffies_to_usecs(timeout - ret),
559 reason);
560}
561
562/*
563 * Account for folios written if tasks are throttled waiting on dirty
564 * folios to clean. If enough folios have been cleaned since throttling
565 * started then wakeup the throttled tasks.
566 */
567void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
568 int nr_throttled)
569{
570 unsigned long nr_written;
571
572 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
573
574 /*
575 * This is an inaccurate read as the per-cpu deltas may not
576 * be synchronised. However, given that the system is
577 * writeback throttled, it is not worth taking the penalty
578 * of getting an accurate count. At worst, the throttle
579 * timeout guarantees forward progress.
580 */
581 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
582 READ_ONCE(pgdat->nr_reclaim_start);
583
584 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
585 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
586}
587
588/* possible outcome of pageout() */
589typedef enum {
590 /* failed to write folio out, folio is locked */
591 PAGE_KEEP,
592 /* move folio to the active list, folio is locked */
593 PAGE_ACTIVATE,
594 /* folio has been sent to the disk successfully, folio is unlocked */
595 PAGE_SUCCESS,
596 /* folio is clean and locked */
597 PAGE_CLEAN,
598} pageout_t;
599
600/*
601 * pageout is called by shrink_folio_list() for each dirty folio.
602 * Calls ->writepage().
603 */
604static pageout_t pageout(struct folio *folio, struct address_space *mapping,
605 struct swap_iocb **plug)
606{
607 /*
608 * If the folio is dirty, only perform writeback if that write
609 * will be non-blocking. To prevent this allocation from being
610 * stalled by pagecache activity. But note that there may be
611 * stalls if we need to run get_block(). We could test
612 * PagePrivate for that.
613 *
614 * If this process is currently in __generic_file_write_iter() against
615 * this folio's queue, we can perform writeback even if that
616 * will block.
617 *
618 * If the folio is swapcache, write it back even if that would
619 * block, for some throttling. This happens by accident, because
620 * swap_backing_dev_info is bust: it doesn't reflect the
621 * congestion state of the swapdevs. Easy to fix, if needed.
622 */
623 if (!is_page_cache_freeable(folio))
624 return PAGE_KEEP;
625 if (!mapping) {
626 /*
627 * Some data journaling orphaned folios can have
628 * folio->mapping == NULL while being dirty with clean buffers.
629 */
630 if (folio_test_private(folio)) {
631 if (try_to_free_buffers(folio)) {
632 folio_clear_dirty(folio);
633 pr_info("%s: orphaned folio\n", __func__);
634 return PAGE_CLEAN;
635 }
636 }
637 return PAGE_KEEP;
638 }
639 if (mapping->a_ops->writepage == NULL)
640 return PAGE_ACTIVATE;
641
642 if (folio_clear_dirty_for_io(folio)) {
643 int res;
644 struct writeback_control wbc = {
645 .sync_mode = WB_SYNC_NONE,
646 .nr_to_write = SWAP_CLUSTER_MAX,
647 .range_start = 0,
648 .range_end = LLONG_MAX,
649 .for_reclaim = 1,
650 .swap_plug = plug,
651 };
652
653 folio_set_reclaim(folio);
654 res = mapping->a_ops->writepage(&folio->page, &wbc);
655 if (res < 0)
656 handle_write_error(mapping, folio, res);
657 if (res == AOP_WRITEPAGE_ACTIVATE) {
658 folio_clear_reclaim(folio);
659 return PAGE_ACTIVATE;
660 }
661
662 if (!folio_test_writeback(folio)) {
663 /* synchronous write or broken a_ops? */
664 folio_clear_reclaim(folio);
665 }
666 trace_mm_vmscan_write_folio(folio);
667 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
668 return PAGE_SUCCESS;
669 }
670
671 return PAGE_CLEAN;
672}
673
674/*
675 * Same as remove_mapping, but if the folio is removed from the mapping, it
676 * gets returned with a refcount of 0.
677 */
678static int __remove_mapping(struct address_space *mapping, struct folio *folio,
679 bool reclaimed, struct mem_cgroup *target_memcg)
680{
681 int refcount;
682 void *shadow = NULL;
683
684 BUG_ON(!folio_test_locked(folio));
685 BUG_ON(mapping != folio_mapping(folio));
686
687 if (!folio_test_swapcache(folio))
688 spin_lock(&mapping->host->i_lock);
689 xa_lock_irq(&mapping->i_pages);
690 /*
691 * The non racy check for a busy folio.
692 *
693 * Must be careful with the order of the tests. When someone has
694 * a ref to the folio, it may be possible that they dirty it then
695 * drop the reference. So if the dirty flag is tested before the
696 * refcount here, then the following race may occur:
697 *
698 * get_user_pages(&page);
699 * [user mapping goes away]
700 * write_to(page);
701 * !folio_test_dirty(folio) [good]
702 * folio_set_dirty(folio);
703 * folio_put(folio);
704 * !refcount(folio) [good, discard it]
705 *
706 * [oops, our write_to data is lost]
707 *
708 * Reversing the order of the tests ensures such a situation cannot
709 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
710 * load is not satisfied before that of folio->_refcount.
711 *
712 * Note that if the dirty flag is always set via folio_mark_dirty,
713 * and thus under the i_pages lock, then this ordering is not required.
714 */
715 refcount = 1 + folio_nr_pages(folio);
716 if (!folio_ref_freeze(folio, refcount))
717 goto cannot_free;
718 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
719 if (unlikely(folio_test_dirty(folio))) {
720 folio_ref_unfreeze(folio, refcount);
721 goto cannot_free;
722 }
723
724 if (folio_test_swapcache(folio)) {
725 swp_entry_t swap = folio->swap;
726
727 if (reclaimed && !mapping_exiting(mapping))
728 shadow = workingset_eviction(folio, target_memcg);
729 __delete_from_swap_cache(folio, swap, shadow);
730 mem_cgroup_swapout(folio, swap);
731 xa_unlock_irq(&mapping->i_pages);
732 put_swap_folio(folio, swap);
733 } else {
734 void (*free_folio)(struct folio *);
735
736 free_folio = mapping->a_ops->free_folio;
737 /*
738 * Remember a shadow entry for reclaimed file cache in
739 * order to detect refaults, thus thrashing, later on.
740 *
741 * But don't store shadows in an address space that is
742 * already exiting. This is not just an optimization,
743 * inode reclaim needs to empty out the radix tree or
744 * the nodes are lost. Don't plant shadows behind its
745 * back.
746 *
747 * We also don't store shadows for DAX mappings because the
748 * only page cache folios found in these are zero pages
749 * covering holes, and because we don't want to mix DAX
750 * exceptional entries and shadow exceptional entries in the
751 * same address_space.
752 */
753 if (reclaimed && folio_is_file_lru(folio) &&
754 !mapping_exiting(mapping) && !dax_mapping(mapping))
755 shadow = workingset_eviction(folio, target_memcg);
756 __filemap_remove_folio(folio, shadow);
757 xa_unlock_irq(&mapping->i_pages);
758 if (mapping_shrinkable(mapping))
759 inode_add_lru(mapping->host);
760 spin_unlock(&mapping->host->i_lock);
761
762 if (free_folio)
763 free_folio(folio);
764 }
765
766 return 1;
767
768cannot_free:
769 xa_unlock_irq(&mapping->i_pages);
770 if (!folio_test_swapcache(folio))
771 spin_unlock(&mapping->host->i_lock);
772 return 0;
773}
774
775/**
776 * remove_mapping() - Attempt to remove a folio from its mapping.
777 * @mapping: The address space.
778 * @folio: The folio to remove.
779 *
780 * If the folio is dirty, under writeback or if someone else has a ref
781 * on it, removal will fail.
782 * Return: The number of pages removed from the mapping. 0 if the folio
783 * could not be removed.
784 * Context: The caller should have a single refcount on the folio and
785 * hold its lock.
786 */
787long remove_mapping(struct address_space *mapping, struct folio *folio)
788{
789 if (__remove_mapping(mapping, folio, false, NULL)) {
790 /*
791 * Unfreezing the refcount with 1 effectively
792 * drops the pagecache ref for us without requiring another
793 * atomic operation.
794 */
795 folio_ref_unfreeze(folio, 1);
796 return folio_nr_pages(folio);
797 }
798 return 0;
799}
800
801/**
802 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
803 * @folio: Folio to be returned to an LRU list.
804 *
805 * Add previously isolated @folio to appropriate LRU list.
806 * The folio may still be unevictable for other reasons.
807 *
808 * Context: lru_lock must not be held, interrupts must be enabled.
809 */
810void folio_putback_lru(struct folio *folio)
811{
812 folio_add_lru(folio);
813 folio_put(folio); /* drop ref from isolate */
814}
815
816enum folio_references {
817 FOLIOREF_RECLAIM,
818 FOLIOREF_RECLAIM_CLEAN,
819 FOLIOREF_KEEP,
820 FOLIOREF_ACTIVATE,
821};
822
823static enum folio_references folio_check_references(struct folio *folio,
824 struct scan_control *sc)
825{
826 int referenced_ptes, referenced_folio;
827 unsigned long vm_flags;
828
829 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
830 &vm_flags);
831 referenced_folio = folio_test_clear_referenced(folio);
832
833 /*
834 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
835 * Let the folio, now marked Mlocked, be moved to the unevictable list.
836 */
837 if (vm_flags & VM_LOCKED)
838 return FOLIOREF_ACTIVATE;
839
840 /* rmap lock contention: rotate */
841 if (referenced_ptes == -1)
842 return FOLIOREF_KEEP;
843
844 if (referenced_ptes) {
845 /*
846 * All mapped folios start out with page table
847 * references from the instantiating fault, so we need
848 * to look twice if a mapped file/anon folio is used more
849 * than once.
850 *
851 * Mark it and spare it for another trip around the
852 * inactive list. Another page table reference will
853 * lead to its activation.
854 *
855 * Note: the mark is set for activated folios as well
856 * so that recently deactivated but used folios are
857 * quickly recovered.
858 */
859 folio_set_referenced(folio);
860
861 if (referenced_folio || referenced_ptes > 1)
862 return FOLIOREF_ACTIVATE;
863
864 /*
865 * Activate file-backed executable folios after first usage.
866 */
867 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
868 return FOLIOREF_ACTIVATE;
869
870 return FOLIOREF_KEEP;
871 }
872
873 /* Reclaim if clean, defer dirty folios to writeback */
874 if (referenced_folio && folio_is_file_lru(folio))
875 return FOLIOREF_RECLAIM_CLEAN;
876
877 return FOLIOREF_RECLAIM;
878}
879
880/* Check if a folio is dirty or under writeback */
881static void folio_check_dirty_writeback(struct folio *folio,
882 bool *dirty, bool *writeback)
883{
884 struct address_space *mapping;
885
886 /*
887 * Anonymous folios are not handled by flushers and must be written
888 * from reclaim context. Do not stall reclaim based on them.
889 * MADV_FREE anonymous folios are put into inactive file list too.
890 * They could be mistakenly treated as file lru. So further anon
891 * test is needed.
892 */
893 if (!folio_is_file_lru(folio) ||
894 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
895 *dirty = false;
896 *writeback = false;
897 return;
898 }
899
900 /* By default assume that the folio flags are accurate */
901 *dirty = folio_test_dirty(folio);
902 *writeback = folio_test_writeback(folio);
903
904 /* Verify dirty/writeback state if the filesystem supports it */
905 if (!folio_test_private(folio))
906 return;
907
908 mapping = folio_mapping(folio);
909 if (mapping && mapping->a_ops->is_dirty_writeback)
910 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
911}
912
913static struct folio *alloc_demote_folio(struct folio *src,
914 unsigned long private)
915{
916 struct folio *dst;
917 nodemask_t *allowed_mask;
918 struct migration_target_control *mtc;
919
920 mtc = (struct migration_target_control *)private;
921
922 allowed_mask = mtc->nmask;
923 /*
924 * make sure we allocate from the target node first also trying to
925 * demote or reclaim pages from the target node via kswapd if we are
926 * low on free memory on target node. If we don't do this and if
927 * we have free memory on the slower(lower) memtier, we would start
928 * allocating pages from slower(lower) memory tiers without even forcing
929 * a demotion of cold pages from the target memtier. This can result
930 * in the kernel placing hot pages in slower(lower) memory tiers.
931 */
932 mtc->nmask = NULL;
933 mtc->gfp_mask |= __GFP_THISNODE;
934 dst = alloc_migration_target(src, (unsigned long)mtc);
935 if (dst)
936 return dst;
937
938 mtc->gfp_mask &= ~__GFP_THISNODE;
939 mtc->nmask = allowed_mask;
940
941 return alloc_migration_target(src, (unsigned long)mtc);
942}
943
944/*
945 * Take folios on @demote_folios and attempt to demote them to another node.
946 * Folios which are not demoted are left on @demote_folios.
947 */
948static unsigned int demote_folio_list(struct list_head *demote_folios,
949 struct pglist_data *pgdat)
950{
951 int target_nid = next_demotion_node(pgdat->node_id);
952 unsigned int nr_succeeded;
953 nodemask_t allowed_mask;
954
955 struct migration_target_control mtc = {
956 /*
957 * Allocate from 'node', or fail quickly and quietly.
958 * When this happens, 'page' will likely just be discarded
959 * instead of migrated.
960 */
961 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
962 __GFP_NOMEMALLOC | GFP_NOWAIT,
963 .nid = target_nid,
964 .nmask = &allowed_mask
965 };
966
967 if (list_empty(demote_folios))
968 return 0;
969
970 if (target_nid == NUMA_NO_NODE)
971 return 0;
972
973 node_get_allowed_targets(pgdat, &allowed_mask);
974
975 /* Demotion ignores all cpuset and mempolicy settings */
976 migrate_pages(demote_folios, alloc_demote_folio, NULL,
977 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
978 &nr_succeeded);
979
980 mod_node_page_state(pgdat, PGDEMOTE_KSWAPD + reclaimer_offset(),
981 nr_succeeded);
982
983 return nr_succeeded;
984}
985
986static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
987{
988 if (gfp_mask & __GFP_FS)
989 return true;
990 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
991 return false;
992 /*
993 * We can "enter_fs" for swap-cache with only __GFP_IO
994 * providing this isn't SWP_FS_OPS.
995 * ->flags can be updated non-atomicially (scan_swap_map_slots),
996 * but that will never affect SWP_FS_OPS, so the data_race
997 * is safe.
998 */
999 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1000}
1001
1002/*
1003 * shrink_folio_list() returns the number of reclaimed pages
1004 */
1005static unsigned int shrink_folio_list(struct list_head *folio_list,
1006 struct pglist_data *pgdat, struct scan_control *sc,
1007 struct reclaim_stat *stat, bool ignore_references)
1008{
1009 LIST_HEAD(ret_folios);
1010 LIST_HEAD(free_folios);
1011 LIST_HEAD(demote_folios);
1012 unsigned int nr_reclaimed = 0;
1013 unsigned int pgactivate = 0;
1014 bool do_demote_pass;
1015 struct swap_iocb *plug = NULL;
1016
1017 memset(stat, 0, sizeof(*stat));
1018 cond_resched();
1019 do_demote_pass = can_demote(pgdat->node_id, sc);
1020
1021retry:
1022 while (!list_empty(folio_list)) {
1023 struct address_space *mapping;
1024 struct folio *folio;
1025 enum folio_references references = FOLIOREF_RECLAIM;
1026 bool dirty, writeback;
1027 unsigned int nr_pages;
1028
1029 cond_resched();
1030
1031 folio = lru_to_folio(folio_list);
1032 list_del(&folio->lru);
1033
1034 if (!folio_trylock(folio))
1035 goto keep;
1036
1037 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1038
1039 nr_pages = folio_nr_pages(folio);
1040
1041 /* Account the number of base pages */
1042 sc->nr_scanned += nr_pages;
1043
1044 if (unlikely(!folio_evictable(folio)))
1045 goto activate_locked;
1046
1047 if (!sc->may_unmap && folio_mapped(folio))
1048 goto keep_locked;
1049
1050 /* folio_update_gen() tried to promote this page? */
1051 if (lru_gen_enabled() && !ignore_references &&
1052 folio_mapped(folio) && folio_test_referenced(folio))
1053 goto keep_locked;
1054
1055 /*
1056 * The number of dirty pages determines if a node is marked
1057 * reclaim_congested. kswapd will stall and start writing
1058 * folios if the tail of the LRU is all dirty unqueued folios.
1059 */
1060 folio_check_dirty_writeback(folio, &dirty, &writeback);
1061 if (dirty || writeback)
1062 stat->nr_dirty += nr_pages;
1063
1064 if (dirty && !writeback)
1065 stat->nr_unqueued_dirty += nr_pages;
1066
1067 /*
1068 * Treat this folio as congested if folios are cycling
1069 * through the LRU so quickly that the folios marked
1070 * for immediate reclaim are making it to the end of
1071 * the LRU a second time.
1072 */
1073 if (writeback && folio_test_reclaim(folio))
1074 stat->nr_congested += nr_pages;
1075
1076 /*
1077 * If a folio at the tail of the LRU is under writeback, there
1078 * are three cases to consider.
1079 *
1080 * 1) If reclaim is encountering an excessive number
1081 * of folios under writeback and this folio has both
1082 * the writeback and reclaim flags set, then it
1083 * indicates that folios are being queued for I/O but
1084 * are being recycled through the LRU before the I/O
1085 * can complete. Waiting on the folio itself risks an
1086 * indefinite stall if it is impossible to writeback
1087 * the folio due to I/O error or disconnected storage
1088 * so instead note that the LRU is being scanned too
1089 * quickly and the caller can stall after the folio
1090 * list has been processed.
1091 *
1092 * 2) Global or new memcg reclaim encounters a folio that is
1093 * not marked for immediate reclaim, or the caller does not
1094 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1095 * not to fs). In this case mark the folio for immediate
1096 * reclaim and continue scanning.
1097 *
1098 * Require may_enter_fs() because we would wait on fs, which
1099 * may not have submitted I/O yet. And the loop driver might
1100 * enter reclaim, and deadlock if it waits on a folio for
1101 * which it is needed to do the write (loop masks off
1102 * __GFP_IO|__GFP_FS for this reason); but more thought
1103 * would probably show more reasons.
1104 *
1105 * 3) Legacy memcg encounters a folio that already has the
1106 * reclaim flag set. memcg does not have any dirty folio
1107 * throttling so we could easily OOM just because too many
1108 * folios are in writeback and there is nothing else to
1109 * reclaim. Wait for the writeback to complete.
1110 *
1111 * In cases 1) and 2) we activate the folios to get them out of
1112 * the way while we continue scanning for clean folios on the
1113 * inactive list and refilling from the active list. The
1114 * observation here is that waiting for disk writes is more
1115 * expensive than potentially causing reloads down the line.
1116 * Since they're marked for immediate reclaim, they won't put
1117 * memory pressure on the cache working set any longer than it
1118 * takes to write them to disk.
1119 */
1120 if (folio_test_writeback(folio)) {
1121 /* Case 1 above */
1122 if (current_is_kswapd() &&
1123 folio_test_reclaim(folio) &&
1124 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1125 stat->nr_immediate += nr_pages;
1126 goto activate_locked;
1127
1128 /* Case 2 above */
1129 } else if (writeback_throttling_sane(sc) ||
1130 !folio_test_reclaim(folio) ||
1131 !may_enter_fs(folio, sc->gfp_mask)) {
1132 /*
1133 * This is slightly racy -
1134 * folio_end_writeback() might have
1135 * just cleared the reclaim flag, then
1136 * setting the reclaim flag here ends up
1137 * interpreted as the readahead flag - but
1138 * that does not matter enough to care.
1139 * What we do want is for this folio to
1140 * have the reclaim flag set next time
1141 * memcg reclaim reaches the tests above,
1142 * so it will then wait for writeback to
1143 * avoid OOM; and it's also appropriate
1144 * in global reclaim.
1145 */
1146 folio_set_reclaim(folio);
1147 stat->nr_writeback += nr_pages;
1148 goto activate_locked;
1149
1150 /* Case 3 above */
1151 } else {
1152 folio_unlock(folio);
1153 folio_wait_writeback(folio);
1154 /* then go back and try same folio again */
1155 list_add_tail(&folio->lru, folio_list);
1156 continue;
1157 }
1158 }
1159
1160 if (!ignore_references)
1161 references = folio_check_references(folio, sc);
1162
1163 switch (references) {
1164 case FOLIOREF_ACTIVATE:
1165 goto activate_locked;
1166 case FOLIOREF_KEEP:
1167 stat->nr_ref_keep += nr_pages;
1168 goto keep_locked;
1169 case FOLIOREF_RECLAIM:
1170 case FOLIOREF_RECLAIM_CLEAN:
1171 ; /* try to reclaim the folio below */
1172 }
1173
1174 /*
1175 * Before reclaiming the folio, try to relocate
1176 * its contents to another node.
1177 */
1178 if (do_demote_pass &&
1179 (thp_migration_supported() || !folio_test_large(folio))) {
1180 list_add(&folio->lru, &demote_folios);
1181 folio_unlock(folio);
1182 continue;
1183 }
1184
1185 /*
1186 * Anonymous process memory has backing store?
1187 * Try to allocate it some swap space here.
1188 * Lazyfree folio could be freed directly
1189 */
1190 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1191 if (!folio_test_swapcache(folio)) {
1192 if (!(sc->gfp_mask & __GFP_IO))
1193 goto keep_locked;
1194 if (folio_maybe_dma_pinned(folio))
1195 goto keep_locked;
1196 if (folio_test_large(folio)) {
1197 /* cannot split folio, skip it */
1198 if (!can_split_folio(folio, NULL))
1199 goto activate_locked;
1200 /*
1201 * Split folios without a PMD map right
1202 * away. Chances are some or all of the
1203 * tail pages can be freed without IO.
1204 */
1205 if (!folio_entire_mapcount(folio) &&
1206 split_folio_to_list(folio,
1207 folio_list))
1208 goto activate_locked;
1209 }
1210 if (!add_to_swap(folio)) {
1211 if (!folio_test_large(folio))
1212 goto activate_locked_split;
1213 /* Fallback to swap normal pages */
1214 if (split_folio_to_list(folio,
1215 folio_list))
1216 goto activate_locked;
1217#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1218 count_memcg_folio_events(folio, THP_SWPOUT_FALLBACK, 1);
1219 count_vm_event(THP_SWPOUT_FALLBACK);
1220#endif
1221 if (!add_to_swap(folio))
1222 goto activate_locked_split;
1223 }
1224 }
1225 } else if (folio_test_swapbacked(folio) &&
1226 folio_test_large(folio)) {
1227 /* Split shmem folio */
1228 if (split_folio_to_list(folio, folio_list))
1229 goto keep_locked;
1230 }
1231
1232 /*
1233 * If the folio was split above, the tail pages will make
1234 * their own pass through this function and be accounted
1235 * then.
1236 */
1237 if ((nr_pages > 1) && !folio_test_large(folio)) {
1238 sc->nr_scanned -= (nr_pages - 1);
1239 nr_pages = 1;
1240 }
1241
1242 /*
1243 * The folio is mapped into the page tables of one or more
1244 * processes. Try to unmap it here.
1245 */
1246 if (folio_mapped(folio)) {
1247 enum ttu_flags flags = TTU_BATCH_FLUSH;
1248 bool was_swapbacked = folio_test_swapbacked(folio);
1249
1250 if (folio_test_pmd_mappable(folio))
1251 flags |= TTU_SPLIT_HUGE_PMD;
1252
1253 try_to_unmap(folio, flags);
1254 if (folio_mapped(folio)) {
1255 stat->nr_unmap_fail += nr_pages;
1256 if (!was_swapbacked &&
1257 folio_test_swapbacked(folio))
1258 stat->nr_lazyfree_fail += nr_pages;
1259 goto activate_locked;
1260 }
1261 }
1262
1263 /*
1264 * Folio is unmapped now so it cannot be newly pinned anymore.
1265 * No point in trying to reclaim folio if it is pinned.
1266 * Furthermore we don't want to reclaim underlying fs metadata
1267 * if the folio is pinned and thus potentially modified by the
1268 * pinning process as that may upset the filesystem.
1269 */
1270 if (folio_maybe_dma_pinned(folio))
1271 goto activate_locked;
1272
1273 mapping = folio_mapping(folio);
1274 if (folio_test_dirty(folio)) {
1275 /*
1276 * Only kswapd can writeback filesystem folios
1277 * to avoid risk of stack overflow. But avoid
1278 * injecting inefficient single-folio I/O into
1279 * flusher writeback as much as possible: only
1280 * write folios when we've encountered many
1281 * dirty folios, and when we've already scanned
1282 * the rest of the LRU for clean folios and see
1283 * the same dirty folios again (with the reclaim
1284 * flag set).
1285 */
1286 if (folio_is_file_lru(folio) &&
1287 (!current_is_kswapd() ||
1288 !folio_test_reclaim(folio) ||
1289 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1290 /*
1291 * Immediately reclaim when written back.
1292 * Similar in principle to folio_deactivate()
1293 * except we already have the folio isolated
1294 * and know it's dirty
1295 */
1296 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1297 nr_pages);
1298 folio_set_reclaim(folio);
1299
1300 goto activate_locked;
1301 }
1302
1303 if (references == FOLIOREF_RECLAIM_CLEAN)
1304 goto keep_locked;
1305 if (!may_enter_fs(folio, sc->gfp_mask))
1306 goto keep_locked;
1307 if (!sc->may_writepage)
1308 goto keep_locked;
1309
1310 /*
1311 * Folio is dirty. Flush the TLB if a writable entry
1312 * potentially exists to avoid CPU writes after I/O
1313 * starts and then write it out here.
1314 */
1315 try_to_unmap_flush_dirty();
1316 switch (pageout(folio, mapping, &plug)) {
1317 case PAGE_KEEP:
1318 goto keep_locked;
1319 case PAGE_ACTIVATE:
1320 goto activate_locked;
1321 case PAGE_SUCCESS:
1322 stat->nr_pageout += nr_pages;
1323
1324 if (folio_test_writeback(folio))
1325 goto keep;
1326 if (folio_test_dirty(folio))
1327 goto keep;
1328
1329 /*
1330 * A synchronous write - probably a ramdisk. Go
1331 * ahead and try to reclaim the folio.
1332 */
1333 if (!folio_trylock(folio))
1334 goto keep;
1335 if (folio_test_dirty(folio) ||
1336 folio_test_writeback(folio))
1337 goto keep_locked;
1338 mapping = folio_mapping(folio);
1339 fallthrough;
1340 case PAGE_CLEAN:
1341 ; /* try to free the folio below */
1342 }
1343 }
1344
1345 /*
1346 * If the folio has buffers, try to free the buffer
1347 * mappings associated with this folio. If we succeed
1348 * we try to free the folio as well.
1349 *
1350 * We do this even if the folio is dirty.
1351 * filemap_release_folio() does not perform I/O, but it
1352 * is possible for a folio to have the dirty flag set,
1353 * but it is actually clean (all its buffers are clean).
1354 * This happens if the buffers were written out directly,
1355 * with submit_bh(). ext3 will do this, as well as
1356 * the blockdev mapping. filemap_release_folio() will
1357 * discover that cleanness and will drop the buffers
1358 * and mark the folio clean - it can be freed.
1359 *
1360 * Rarely, folios can have buffers and no ->mapping.
1361 * These are the folios which were not successfully
1362 * invalidated in truncate_cleanup_folio(). We try to
1363 * drop those buffers here and if that worked, and the
1364 * folio is no longer mapped into process address space
1365 * (refcount == 1) it can be freed. Otherwise, leave
1366 * the folio on the LRU so it is swappable.
1367 */
1368 if (folio_needs_release(folio)) {
1369 if (!filemap_release_folio(folio, sc->gfp_mask))
1370 goto activate_locked;
1371 if (!mapping && folio_ref_count(folio) == 1) {
1372 folio_unlock(folio);
1373 if (folio_put_testzero(folio))
1374 goto free_it;
1375 else {
1376 /*
1377 * rare race with speculative reference.
1378 * the speculative reference will free
1379 * this folio shortly, so we may
1380 * increment nr_reclaimed here (and
1381 * leave it off the LRU).
1382 */
1383 nr_reclaimed += nr_pages;
1384 continue;
1385 }
1386 }
1387 }
1388
1389 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1390 /* follow __remove_mapping for reference */
1391 if (!folio_ref_freeze(folio, 1))
1392 goto keep_locked;
1393 /*
1394 * The folio has only one reference left, which is
1395 * from the isolation. After the caller puts the
1396 * folio back on the lru and drops the reference, the
1397 * folio will be freed anyway. It doesn't matter
1398 * which lru it goes on. So we don't bother checking
1399 * the dirty flag here.
1400 */
1401 count_vm_events(PGLAZYFREED, nr_pages);
1402 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1403 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1404 sc->target_mem_cgroup))
1405 goto keep_locked;
1406
1407 folio_unlock(folio);
1408free_it:
1409 /*
1410 * Folio may get swapped out as a whole, need to account
1411 * all pages in it.
1412 */
1413 nr_reclaimed += nr_pages;
1414
1415 /*
1416 * Is there need to periodically free_folio_list? It would
1417 * appear not as the counts should be low
1418 */
1419 if (unlikely(folio_test_large(folio)))
1420 destroy_large_folio(folio);
1421 else
1422 list_add(&folio->lru, &free_folios);
1423 continue;
1424
1425activate_locked_split:
1426 /*
1427 * The tail pages that are failed to add into swap cache
1428 * reach here. Fixup nr_scanned and nr_pages.
1429 */
1430 if (nr_pages > 1) {
1431 sc->nr_scanned -= (nr_pages - 1);
1432 nr_pages = 1;
1433 }
1434activate_locked:
1435 /* Not a candidate for swapping, so reclaim swap space. */
1436 if (folio_test_swapcache(folio) &&
1437 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1438 folio_free_swap(folio);
1439 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1440 if (!folio_test_mlocked(folio)) {
1441 int type = folio_is_file_lru(folio);
1442 folio_set_active(folio);
1443 stat->nr_activate[type] += nr_pages;
1444 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1445 }
1446keep_locked:
1447 folio_unlock(folio);
1448keep:
1449 list_add(&folio->lru, &ret_folios);
1450 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1451 folio_test_unevictable(folio), folio);
1452 }
1453 /* 'folio_list' is always empty here */
1454
1455 /* Migrate folios selected for demotion */
1456 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
1457 /* Folios that could not be demoted are still in @demote_folios */
1458 if (!list_empty(&demote_folios)) {
1459 /* Folios which weren't demoted go back on @folio_list */
1460 list_splice_init(&demote_folios, folio_list);
1461
1462 /*
1463 * goto retry to reclaim the undemoted folios in folio_list if
1464 * desired.
1465 *
1466 * Reclaiming directly from top tier nodes is not often desired
1467 * due to it breaking the LRU ordering: in general memory
1468 * should be reclaimed from lower tier nodes and demoted from
1469 * top tier nodes.
1470 *
1471 * However, disabling reclaim from top tier nodes entirely
1472 * would cause ooms in edge scenarios where lower tier memory
1473 * is unreclaimable for whatever reason, eg memory being
1474 * mlocked or too hot to reclaim. We can disable reclaim
1475 * from top tier nodes in proactive reclaim though as that is
1476 * not real memory pressure.
1477 */
1478 if (!sc->proactive) {
1479 do_demote_pass = false;
1480 goto retry;
1481 }
1482 }
1483
1484 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1485
1486 mem_cgroup_uncharge_list(&free_folios);
1487 try_to_unmap_flush();
1488 free_unref_page_list(&free_folios);
1489
1490 list_splice(&ret_folios, folio_list);
1491 count_vm_events(PGACTIVATE, pgactivate);
1492
1493 if (plug)
1494 swap_write_unplug(plug);
1495 return nr_reclaimed;
1496}
1497
1498unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1499 struct list_head *folio_list)
1500{
1501 struct scan_control sc = {
1502 .gfp_mask = GFP_KERNEL,
1503 .may_unmap = 1,
1504 };
1505 struct reclaim_stat stat;
1506 unsigned int nr_reclaimed;
1507 struct folio *folio, *next;
1508 LIST_HEAD(clean_folios);
1509 unsigned int noreclaim_flag;
1510
1511 list_for_each_entry_safe(folio, next, folio_list, lru) {
1512 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1513 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1514 !folio_test_unevictable(folio)) {
1515 folio_clear_active(folio);
1516 list_move(&folio->lru, &clean_folios);
1517 }
1518 }
1519
1520 /*
1521 * We should be safe here since we are only dealing with file pages and
1522 * we are not kswapd and therefore cannot write dirty file pages. But
1523 * call memalloc_noreclaim_save() anyway, just in case these conditions
1524 * change in the future.
1525 */
1526 noreclaim_flag = memalloc_noreclaim_save();
1527 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1528 &stat, true);
1529 memalloc_noreclaim_restore(noreclaim_flag);
1530
1531 list_splice(&clean_folios, folio_list);
1532 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1533 -(long)nr_reclaimed);
1534 /*
1535 * Since lazyfree pages are isolated from file LRU from the beginning,
1536 * they will rotate back to anonymous LRU in the end if it failed to
1537 * discard so isolated count will be mismatched.
1538 * Compensate the isolated count for both LRU lists.
1539 */
1540 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1541 stat.nr_lazyfree_fail);
1542 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1543 -(long)stat.nr_lazyfree_fail);
1544 return nr_reclaimed;
1545}
1546
1547/*
1548 * Update LRU sizes after isolating pages. The LRU size updates must
1549 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1550 */
1551static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1552 enum lru_list lru, unsigned long *nr_zone_taken)
1553{
1554 int zid;
1555
1556 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1557 if (!nr_zone_taken[zid])
1558 continue;
1559
1560 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1561 }
1562
1563}
1564
1565#ifdef CONFIG_CMA
1566/*
1567 * It is waste of effort to scan and reclaim CMA pages if it is not available
1568 * for current allocation context. Kswapd can not be enrolled as it can not
1569 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1570 */
1571static bool skip_cma(struct folio *folio, struct scan_control *sc)
1572{
1573 return !current_is_kswapd() &&
1574 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1575 folio_migratetype(folio) == MIGRATE_CMA;
1576}
1577#else
1578static bool skip_cma(struct folio *folio, struct scan_control *sc)
1579{
1580 return false;
1581}
1582#endif
1583
1584/*
1585 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1586 *
1587 * lruvec->lru_lock is heavily contended. Some of the functions that
1588 * shrink the lists perform better by taking out a batch of pages
1589 * and working on them outside the LRU lock.
1590 *
1591 * For pagecache intensive workloads, this function is the hottest
1592 * spot in the kernel (apart from copy_*_user functions).
1593 *
1594 * Lru_lock must be held before calling this function.
1595 *
1596 * @nr_to_scan: The number of eligible pages to look through on the list.
1597 * @lruvec: The LRU vector to pull pages from.
1598 * @dst: The temp list to put pages on to.
1599 * @nr_scanned: The number of pages that were scanned.
1600 * @sc: The scan_control struct for this reclaim session
1601 * @lru: LRU list id for isolating
1602 *
1603 * returns how many pages were moved onto *@dst.
1604 */
1605static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1606 struct lruvec *lruvec, struct list_head *dst,
1607 unsigned long *nr_scanned, struct scan_control *sc,
1608 enum lru_list lru)
1609{
1610 struct list_head *src = &lruvec->lists[lru];
1611 unsigned long nr_taken = 0;
1612 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1613 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1614 unsigned long skipped = 0;
1615 unsigned long scan, total_scan, nr_pages;
1616 LIST_HEAD(folios_skipped);
1617
1618 total_scan = 0;
1619 scan = 0;
1620 while (scan < nr_to_scan && !list_empty(src)) {
1621 struct list_head *move_to = src;
1622 struct folio *folio;
1623
1624 folio = lru_to_folio(src);
1625 prefetchw_prev_lru_folio(folio, src, flags);
1626
1627 nr_pages = folio_nr_pages(folio);
1628 total_scan += nr_pages;
1629
1630 if (folio_zonenum(folio) > sc->reclaim_idx ||
1631 skip_cma(folio, sc)) {
1632 nr_skipped[folio_zonenum(folio)] += nr_pages;
1633 move_to = &folios_skipped;
1634 goto move;
1635 }
1636
1637 /*
1638 * Do not count skipped folios because that makes the function
1639 * return with no isolated folios if the LRU mostly contains
1640 * ineligible folios. This causes the VM to not reclaim any
1641 * folios, triggering a premature OOM.
1642 * Account all pages in a folio.
1643 */
1644 scan += nr_pages;
1645
1646 if (!folio_test_lru(folio))
1647 goto move;
1648 if (!sc->may_unmap && folio_mapped(folio))
1649 goto move;
1650
1651 /*
1652 * Be careful not to clear the lru flag until after we're
1653 * sure the folio is not being freed elsewhere -- the
1654 * folio release code relies on it.
1655 */
1656 if (unlikely(!folio_try_get(folio)))
1657 goto move;
1658
1659 if (!folio_test_clear_lru(folio)) {
1660 /* Another thread is already isolating this folio */
1661 folio_put(folio);
1662 goto move;
1663 }
1664
1665 nr_taken += nr_pages;
1666 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1667 move_to = dst;
1668move:
1669 list_move(&folio->lru, move_to);
1670 }
1671
1672 /*
1673 * Splice any skipped folios to the start of the LRU list. Note that
1674 * this disrupts the LRU order when reclaiming for lower zones but
1675 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1676 * scanning would soon rescan the same folios to skip and waste lots
1677 * of cpu cycles.
1678 */
1679 if (!list_empty(&folios_skipped)) {
1680 int zid;
1681
1682 list_splice(&folios_skipped, src);
1683 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1684 if (!nr_skipped[zid])
1685 continue;
1686
1687 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1688 skipped += nr_skipped[zid];
1689 }
1690 }
1691 *nr_scanned = total_scan;
1692 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1693 total_scan, skipped, nr_taken, lru);
1694 update_lru_sizes(lruvec, lru, nr_zone_taken);
1695 return nr_taken;
1696}
1697
1698/**
1699 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1700 * @folio: Folio to isolate from its LRU list.
1701 *
1702 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1703 * corresponding to whatever LRU list the folio was on.
1704 *
1705 * The folio will have its LRU flag cleared. If it was found on the
1706 * active list, it will have the Active flag set. If it was found on the
1707 * unevictable list, it will have the Unevictable flag set. These flags
1708 * may need to be cleared by the caller before letting the page go.
1709 *
1710 * Context:
1711 *
1712 * (1) Must be called with an elevated refcount on the folio. This is a
1713 * fundamental difference from isolate_lru_folios() (which is called
1714 * without a stable reference).
1715 * (2) The lru_lock must not be held.
1716 * (3) Interrupts must be enabled.
1717 *
1718 * Return: true if the folio was removed from an LRU list.
1719 * false if the folio was not on an LRU list.
1720 */
1721bool folio_isolate_lru(struct folio *folio)
1722{
1723 bool ret = false;
1724
1725 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1726
1727 if (folio_test_clear_lru(folio)) {
1728 struct lruvec *lruvec;
1729
1730 folio_get(folio);
1731 lruvec = folio_lruvec_lock_irq(folio);
1732 lruvec_del_folio(lruvec, folio);
1733 unlock_page_lruvec_irq(lruvec);
1734 ret = true;
1735 }
1736
1737 return ret;
1738}
1739
1740/*
1741 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1742 * then get rescheduled. When there are massive number of tasks doing page
1743 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1744 * the LRU list will go small and be scanned faster than necessary, leading to
1745 * unnecessary swapping, thrashing and OOM.
1746 */
1747static int too_many_isolated(struct pglist_data *pgdat, int file,
1748 struct scan_control *sc)
1749{
1750 unsigned long inactive, isolated;
1751 bool too_many;
1752
1753 if (current_is_kswapd())
1754 return 0;
1755
1756 if (!writeback_throttling_sane(sc))
1757 return 0;
1758
1759 if (file) {
1760 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1761 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1762 } else {
1763 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1764 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1765 }
1766
1767 /*
1768 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1769 * won't get blocked by normal direct-reclaimers, forming a circular
1770 * deadlock.
1771 */
1772 if (gfp_has_io_fs(sc->gfp_mask))
1773 inactive >>= 3;
1774
1775 too_many = isolated > inactive;
1776
1777 /* Wake up tasks throttled due to too_many_isolated. */
1778 if (!too_many)
1779 wake_throttle_isolated(pgdat);
1780
1781 return too_many;
1782}
1783
1784/*
1785 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1786 * On return, @list is reused as a list of folios to be freed by the caller.
1787 *
1788 * Returns the number of pages moved to the given lruvec.
1789 */
1790static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1791 struct list_head *list)
1792{
1793 int nr_pages, nr_moved = 0;
1794 LIST_HEAD(folios_to_free);
1795
1796 while (!list_empty(list)) {
1797 struct folio *folio = lru_to_folio(list);
1798
1799 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1800 list_del(&folio->lru);
1801 if (unlikely(!folio_evictable(folio))) {
1802 spin_unlock_irq(&lruvec->lru_lock);
1803 folio_putback_lru(folio);
1804 spin_lock_irq(&lruvec->lru_lock);
1805 continue;
1806 }
1807
1808 /*
1809 * The folio_set_lru needs to be kept here for list integrity.
1810 * Otherwise:
1811 * #0 move_folios_to_lru #1 release_pages
1812 * if (!folio_put_testzero())
1813 * if (folio_put_testzero())
1814 * !lru //skip lru_lock
1815 * folio_set_lru()
1816 * list_add(&folio->lru,)
1817 * list_add(&folio->lru,)
1818 */
1819 folio_set_lru(folio);
1820
1821 if (unlikely(folio_put_testzero(folio))) {
1822 __folio_clear_lru_flags(folio);
1823
1824 if (unlikely(folio_test_large(folio))) {
1825 spin_unlock_irq(&lruvec->lru_lock);
1826 destroy_large_folio(folio);
1827 spin_lock_irq(&lruvec->lru_lock);
1828 } else
1829 list_add(&folio->lru, &folios_to_free);
1830
1831 continue;
1832 }
1833
1834 /*
1835 * All pages were isolated from the same lruvec (and isolation
1836 * inhibits memcg migration).
1837 */
1838 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1839 lruvec_add_folio(lruvec, folio);
1840 nr_pages = folio_nr_pages(folio);
1841 nr_moved += nr_pages;
1842 if (folio_test_active(folio))
1843 workingset_age_nonresident(lruvec, nr_pages);
1844 }
1845
1846 /*
1847 * To save our caller's stack, now use input list for pages to free.
1848 */
1849 list_splice(&folios_to_free, list);
1850
1851 return nr_moved;
1852}
1853
1854/*
1855 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1856 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1857 * we should not throttle. Otherwise it is safe to do so.
1858 */
1859static int current_may_throttle(void)
1860{
1861 return !(current->flags & PF_LOCAL_THROTTLE);
1862}
1863
1864/*
1865 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1866 * of reclaimed pages
1867 */
1868static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1869 struct lruvec *lruvec, struct scan_control *sc,
1870 enum lru_list lru)
1871{
1872 LIST_HEAD(folio_list);
1873 unsigned long nr_scanned;
1874 unsigned int nr_reclaimed = 0;
1875 unsigned long nr_taken;
1876 struct reclaim_stat stat;
1877 bool file = is_file_lru(lru);
1878 enum vm_event_item item;
1879 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1880 bool stalled = false;
1881
1882 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1883 if (stalled)
1884 return 0;
1885
1886 /* wait a bit for the reclaimer. */
1887 stalled = true;
1888 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1889
1890 /* We are about to die and free our memory. Return now. */
1891 if (fatal_signal_pending(current))
1892 return SWAP_CLUSTER_MAX;
1893 }
1894
1895 lru_add_drain();
1896
1897 spin_lock_irq(&lruvec->lru_lock);
1898
1899 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1900 &nr_scanned, sc, lru);
1901
1902 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1903 item = PGSCAN_KSWAPD + reclaimer_offset();
1904 if (!cgroup_reclaim(sc))
1905 __count_vm_events(item, nr_scanned);
1906 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1907 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1908
1909 spin_unlock_irq(&lruvec->lru_lock);
1910
1911 if (nr_taken == 0)
1912 return 0;
1913
1914 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1915
1916 spin_lock_irq(&lruvec->lru_lock);
1917 move_folios_to_lru(lruvec, &folio_list);
1918
1919 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1920 item = PGSTEAL_KSWAPD + reclaimer_offset();
1921 if (!cgroup_reclaim(sc))
1922 __count_vm_events(item, nr_reclaimed);
1923 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1924 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1925 spin_unlock_irq(&lruvec->lru_lock);
1926
1927 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1928 mem_cgroup_uncharge_list(&folio_list);
1929 free_unref_page_list(&folio_list);
1930
1931 /*
1932 * If dirty folios are scanned that are not queued for IO, it
1933 * implies that flushers are not doing their job. This can
1934 * happen when memory pressure pushes dirty folios to the end of
1935 * the LRU before the dirty limits are breached and the dirty
1936 * data has expired. It can also happen when the proportion of
1937 * dirty folios grows not through writes but through memory
1938 * pressure reclaiming all the clean cache. And in some cases,
1939 * the flushers simply cannot keep up with the allocation
1940 * rate. Nudge the flusher threads in case they are asleep.
1941 */
1942 if (stat.nr_unqueued_dirty == nr_taken) {
1943 wakeup_flusher_threads(WB_REASON_VMSCAN);
1944 /*
1945 * For cgroupv1 dirty throttling is achieved by waking up
1946 * the kernel flusher here and later waiting on folios
1947 * which are in writeback to finish (see shrink_folio_list()).
1948 *
1949 * Flusher may not be able to issue writeback quickly
1950 * enough for cgroupv1 writeback throttling to work
1951 * on a large system.
1952 */
1953 if (!writeback_throttling_sane(sc))
1954 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1955 }
1956
1957 sc->nr.dirty += stat.nr_dirty;
1958 sc->nr.congested += stat.nr_congested;
1959 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1960 sc->nr.writeback += stat.nr_writeback;
1961 sc->nr.immediate += stat.nr_immediate;
1962 sc->nr.taken += nr_taken;
1963 if (file)
1964 sc->nr.file_taken += nr_taken;
1965
1966 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1967 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1968 return nr_reclaimed;
1969}
1970
1971/*
1972 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1973 *
1974 * We move them the other way if the folio is referenced by one or more
1975 * processes.
1976 *
1977 * If the folios are mostly unmapped, the processing is fast and it is
1978 * appropriate to hold lru_lock across the whole operation. But if
1979 * the folios are mapped, the processing is slow (folio_referenced()), so
1980 * we should drop lru_lock around each folio. It's impossible to balance
1981 * this, so instead we remove the folios from the LRU while processing them.
1982 * It is safe to rely on the active flag against the non-LRU folios in here
1983 * because nobody will play with that bit on a non-LRU folio.
1984 *
1985 * The downside is that we have to touch folio->_refcount against each folio.
1986 * But we had to alter folio->flags anyway.
1987 */
1988static void shrink_active_list(unsigned long nr_to_scan,
1989 struct lruvec *lruvec,
1990 struct scan_control *sc,
1991 enum lru_list lru)
1992{
1993 unsigned long nr_taken;
1994 unsigned long nr_scanned;
1995 unsigned long vm_flags;
1996 LIST_HEAD(l_hold); /* The folios which were snipped off */
1997 LIST_HEAD(l_active);
1998 LIST_HEAD(l_inactive);
1999 unsigned nr_deactivate, nr_activate;
2000 unsigned nr_rotated = 0;
2001 int file = is_file_lru(lru);
2002 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2003
2004 lru_add_drain();
2005
2006 spin_lock_irq(&lruvec->lru_lock);
2007
2008 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2009 &nr_scanned, sc, lru);
2010
2011 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2012
2013 if (!cgroup_reclaim(sc))
2014 __count_vm_events(PGREFILL, nr_scanned);
2015 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2016
2017 spin_unlock_irq(&lruvec->lru_lock);
2018
2019 while (!list_empty(&l_hold)) {
2020 struct folio *folio;
2021
2022 cond_resched();
2023 folio = lru_to_folio(&l_hold);
2024 list_del(&folio->lru);
2025
2026 if (unlikely(!folio_evictable(folio))) {
2027 folio_putback_lru(folio);
2028 continue;
2029 }
2030
2031 if (unlikely(buffer_heads_over_limit)) {
2032 if (folio_needs_release(folio) &&
2033 folio_trylock(folio)) {
2034 filemap_release_folio(folio, 0);
2035 folio_unlock(folio);
2036 }
2037 }
2038
2039 /* Referenced or rmap lock contention: rotate */
2040 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2041 &vm_flags) != 0) {
2042 /*
2043 * Identify referenced, file-backed active folios and
2044 * give them one more trip around the active list. So
2045 * that executable code get better chances to stay in
2046 * memory under moderate memory pressure. Anon folios
2047 * are not likely to be evicted by use-once streaming
2048 * IO, plus JVM can create lots of anon VM_EXEC folios,
2049 * so we ignore them here.
2050 */
2051 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2052 nr_rotated += folio_nr_pages(folio);
2053 list_add(&folio->lru, &l_active);
2054 continue;
2055 }
2056 }
2057
2058 folio_clear_active(folio); /* we are de-activating */
2059 folio_set_workingset(folio);
2060 list_add(&folio->lru, &l_inactive);
2061 }
2062
2063 /*
2064 * Move folios back to the lru list.
2065 */
2066 spin_lock_irq(&lruvec->lru_lock);
2067
2068 nr_activate = move_folios_to_lru(lruvec, &l_active);
2069 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2070 /* Keep all free folios in l_active list */
2071 list_splice(&l_inactive, &l_active);
2072
2073 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2074 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2075
2076 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2077 spin_unlock_irq(&lruvec->lru_lock);
2078
2079 if (nr_rotated)
2080 lru_note_cost(lruvec, file, 0, nr_rotated);
2081 mem_cgroup_uncharge_list(&l_active);
2082 free_unref_page_list(&l_active);
2083 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2084 nr_deactivate, nr_rotated, sc->priority, file);
2085}
2086
2087static unsigned int reclaim_folio_list(struct list_head *folio_list,
2088 struct pglist_data *pgdat)
2089{
2090 struct reclaim_stat dummy_stat;
2091 unsigned int nr_reclaimed;
2092 struct folio *folio;
2093 struct scan_control sc = {
2094 .gfp_mask = GFP_KERNEL,
2095 .may_writepage = 1,
2096 .may_unmap = 1,
2097 .may_swap = 1,
2098 .no_demotion = 1,
2099 };
2100
2101 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2102 while (!list_empty(folio_list)) {
2103 folio = lru_to_folio(folio_list);
2104 list_del(&folio->lru);
2105 folio_putback_lru(folio);
2106 }
2107
2108 return nr_reclaimed;
2109}
2110
2111unsigned long reclaim_pages(struct list_head *folio_list)
2112{
2113 int nid;
2114 unsigned int nr_reclaimed = 0;
2115 LIST_HEAD(node_folio_list);
2116 unsigned int noreclaim_flag;
2117
2118 if (list_empty(folio_list))
2119 return nr_reclaimed;
2120
2121 noreclaim_flag = memalloc_noreclaim_save();
2122
2123 nid = folio_nid(lru_to_folio(folio_list));
2124 do {
2125 struct folio *folio = lru_to_folio(folio_list);
2126
2127 if (nid == folio_nid(folio)) {
2128 folio_clear_active(folio);
2129 list_move(&folio->lru, &node_folio_list);
2130 continue;
2131 }
2132
2133 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2134 nid = folio_nid(lru_to_folio(folio_list));
2135 } while (!list_empty(folio_list));
2136
2137 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2138
2139 memalloc_noreclaim_restore(noreclaim_flag);
2140
2141 return nr_reclaimed;
2142}
2143
2144static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2145 struct lruvec *lruvec, struct scan_control *sc)
2146{
2147 if (is_active_lru(lru)) {
2148 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2149 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2150 else
2151 sc->skipped_deactivate = 1;
2152 return 0;
2153 }
2154
2155 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2156}
2157
2158/*
2159 * The inactive anon list should be small enough that the VM never has
2160 * to do too much work.
2161 *
2162 * The inactive file list should be small enough to leave most memory
2163 * to the established workingset on the scan-resistant active list,
2164 * but large enough to avoid thrashing the aggregate readahead window.
2165 *
2166 * Both inactive lists should also be large enough that each inactive
2167 * folio has a chance to be referenced again before it is reclaimed.
2168 *
2169 * If that fails and refaulting is observed, the inactive list grows.
2170 *
2171 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2172 * on this LRU, maintained by the pageout code. An inactive_ratio
2173 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2174 *
2175 * total target max
2176 * memory ratio inactive
2177 * -------------------------------------
2178 * 10MB 1 5MB
2179 * 100MB 1 50MB
2180 * 1GB 3 250MB
2181 * 10GB 10 0.9GB
2182 * 100GB 31 3GB
2183 * 1TB 101 10GB
2184 * 10TB 320 32GB
2185 */
2186static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2187{
2188 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2189 unsigned long inactive, active;
2190 unsigned long inactive_ratio;
2191 unsigned long gb;
2192
2193 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2194 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2195
2196 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2197 if (gb)
2198 inactive_ratio = int_sqrt(10 * gb);
2199 else
2200 inactive_ratio = 1;
2201
2202 return inactive * inactive_ratio < active;
2203}
2204
2205enum scan_balance {
2206 SCAN_EQUAL,
2207 SCAN_FRACT,
2208 SCAN_ANON,
2209 SCAN_FILE,
2210};
2211
2212static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2213{
2214 unsigned long file;
2215 struct lruvec *target_lruvec;
2216
2217 if (lru_gen_enabled())
2218 return;
2219
2220 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2221
2222 /*
2223 * Flush the memory cgroup stats, so that we read accurate per-memcg
2224 * lruvec stats for heuristics.
2225 */
2226 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2227
2228 /*
2229 * Determine the scan balance between anon and file LRUs.
2230 */
2231 spin_lock_irq(&target_lruvec->lru_lock);
2232 sc->anon_cost = target_lruvec->anon_cost;
2233 sc->file_cost = target_lruvec->file_cost;
2234 spin_unlock_irq(&target_lruvec->lru_lock);
2235
2236 /*
2237 * Target desirable inactive:active list ratios for the anon
2238 * and file LRU lists.
2239 */
2240 if (!sc->force_deactivate) {
2241 unsigned long refaults;
2242
2243 /*
2244 * When refaults are being observed, it means a new
2245 * workingset is being established. Deactivate to get
2246 * rid of any stale active pages quickly.
2247 */
2248 refaults = lruvec_page_state(target_lruvec,
2249 WORKINGSET_ACTIVATE_ANON);
2250 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2251 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2252 sc->may_deactivate |= DEACTIVATE_ANON;
2253 else
2254 sc->may_deactivate &= ~DEACTIVATE_ANON;
2255
2256 refaults = lruvec_page_state(target_lruvec,
2257 WORKINGSET_ACTIVATE_FILE);
2258 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2259 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2260 sc->may_deactivate |= DEACTIVATE_FILE;
2261 else
2262 sc->may_deactivate &= ~DEACTIVATE_FILE;
2263 } else
2264 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2265
2266 /*
2267 * If we have plenty of inactive file pages that aren't
2268 * thrashing, try to reclaim those first before touching
2269 * anonymous pages.
2270 */
2271 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2272 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2273 sc->cache_trim_mode = 1;
2274 else
2275 sc->cache_trim_mode = 0;
2276
2277 /*
2278 * Prevent the reclaimer from falling into the cache trap: as
2279 * cache pages start out inactive, every cache fault will tip
2280 * the scan balance towards the file LRU. And as the file LRU
2281 * shrinks, so does the window for rotation from references.
2282 * This means we have a runaway feedback loop where a tiny
2283 * thrashing file LRU becomes infinitely more attractive than
2284 * anon pages. Try to detect this based on file LRU size.
2285 */
2286 if (!cgroup_reclaim(sc)) {
2287 unsigned long total_high_wmark = 0;
2288 unsigned long free, anon;
2289 int z;
2290
2291 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2292 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2293 node_page_state(pgdat, NR_INACTIVE_FILE);
2294
2295 for (z = 0; z < MAX_NR_ZONES; z++) {
2296 struct zone *zone = &pgdat->node_zones[z];
2297
2298 if (!managed_zone(zone))
2299 continue;
2300
2301 total_high_wmark += high_wmark_pages(zone);
2302 }
2303
2304 /*
2305 * Consider anon: if that's low too, this isn't a
2306 * runaway file reclaim problem, but rather just
2307 * extreme pressure. Reclaim as per usual then.
2308 */
2309 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2310
2311 sc->file_is_tiny =
2312 file + free <= total_high_wmark &&
2313 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2314 anon >> sc->priority;
2315 }
2316}
2317
2318/*
2319 * Determine how aggressively the anon and file LRU lists should be
2320 * scanned.
2321 *
2322 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2323 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2324 */
2325static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2326 unsigned long *nr)
2327{
2328 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2329 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2330 unsigned long anon_cost, file_cost, total_cost;
2331 int swappiness = mem_cgroup_swappiness(memcg);
2332 u64 fraction[ANON_AND_FILE];
2333 u64 denominator = 0; /* gcc */
2334 enum scan_balance scan_balance;
2335 unsigned long ap, fp;
2336 enum lru_list lru;
2337
2338 /* If we have no swap space, do not bother scanning anon folios. */
2339 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2340 scan_balance = SCAN_FILE;
2341 goto out;
2342 }
2343
2344 /*
2345 * Global reclaim will swap to prevent OOM even with no
2346 * swappiness, but memcg users want to use this knob to
2347 * disable swapping for individual groups completely when
2348 * using the memory controller's swap limit feature would be
2349 * too expensive.
2350 */
2351 if (cgroup_reclaim(sc) && !swappiness) {
2352 scan_balance = SCAN_FILE;
2353 goto out;
2354 }
2355
2356 /*
2357 * Do not apply any pressure balancing cleverness when the
2358 * system is close to OOM, scan both anon and file equally
2359 * (unless the swappiness setting disagrees with swapping).
2360 */
2361 if (!sc->priority && swappiness) {
2362 scan_balance = SCAN_EQUAL;
2363 goto out;
2364 }
2365
2366 /*
2367 * If the system is almost out of file pages, force-scan anon.
2368 */
2369 if (sc->file_is_tiny) {
2370 scan_balance = SCAN_ANON;
2371 goto out;
2372 }
2373
2374 /*
2375 * If there is enough inactive page cache, we do not reclaim
2376 * anything from the anonymous working right now.
2377 */
2378 if (sc->cache_trim_mode) {
2379 scan_balance = SCAN_FILE;
2380 goto out;
2381 }
2382
2383 scan_balance = SCAN_FRACT;
2384 /*
2385 * Calculate the pressure balance between anon and file pages.
2386 *
2387 * The amount of pressure we put on each LRU is inversely
2388 * proportional to the cost of reclaiming each list, as
2389 * determined by the share of pages that are refaulting, times
2390 * the relative IO cost of bringing back a swapped out
2391 * anonymous page vs reloading a filesystem page (swappiness).
2392 *
2393 * Although we limit that influence to ensure no list gets
2394 * left behind completely: at least a third of the pressure is
2395 * applied, before swappiness.
2396 *
2397 * With swappiness at 100, anon and file have equal IO cost.
2398 */
2399 total_cost = sc->anon_cost + sc->file_cost;
2400 anon_cost = total_cost + sc->anon_cost;
2401 file_cost = total_cost + sc->file_cost;
2402 total_cost = anon_cost + file_cost;
2403
2404 ap = swappiness * (total_cost + 1);
2405 ap /= anon_cost + 1;
2406
2407 fp = (200 - swappiness) * (total_cost + 1);
2408 fp /= file_cost + 1;
2409
2410 fraction[0] = ap;
2411 fraction[1] = fp;
2412 denominator = ap + fp;
2413out:
2414 for_each_evictable_lru(lru) {
2415 int file = is_file_lru(lru);
2416 unsigned long lruvec_size;
2417 unsigned long low, min;
2418 unsigned long scan;
2419
2420 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2421 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2422 &min, &low);
2423
2424 if (min || low) {
2425 /*
2426 * Scale a cgroup's reclaim pressure by proportioning
2427 * its current usage to its memory.low or memory.min
2428 * setting.
2429 *
2430 * This is important, as otherwise scanning aggression
2431 * becomes extremely binary -- from nothing as we
2432 * approach the memory protection threshold, to totally
2433 * nominal as we exceed it. This results in requiring
2434 * setting extremely liberal protection thresholds. It
2435 * also means we simply get no protection at all if we
2436 * set it too low, which is not ideal.
2437 *
2438 * If there is any protection in place, we reduce scan
2439 * pressure by how much of the total memory used is
2440 * within protection thresholds.
2441 *
2442 * There is one special case: in the first reclaim pass,
2443 * we skip over all groups that are within their low
2444 * protection. If that fails to reclaim enough pages to
2445 * satisfy the reclaim goal, we come back and override
2446 * the best-effort low protection. However, we still
2447 * ideally want to honor how well-behaved groups are in
2448 * that case instead of simply punishing them all
2449 * equally. As such, we reclaim them based on how much
2450 * memory they are using, reducing the scan pressure
2451 * again by how much of the total memory used is under
2452 * hard protection.
2453 */
2454 unsigned long cgroup_size = mem_cgroup_size(memcg);
2455 unsigned long protection;
2456
2457 /* memory.low scaling, make sure we retry before OOM */
2458 if (!sc->memcg_low_reclaim && low > min) {
2459 protection = low;
2460 sc->memcg_low_skipped = 1;
2461 } else {
2462 protection = min;
2463 }
2464
2465 /* Avoid TOCTOU with earlier protection check */
2466 cgroup_size = max(cgroup_size, protection);
2467
2468 scan = lruvec_size - lruvec_size * protection /
2469 (cgroup_size + 1);
2470
2471 /*
2472 * Minimally target SWAP_CLUSTER_MAX pages to keep
2473 * reclaim moving forwards, avoiding decrementing
2474 * sc->priority further than desirable.
2475 */
2476 scan = max(scan, SWAP_CLUSTER_MAX);
2477 } else {
2478 scan = lruvec_size;
2479 }
2480
2481 scan >>= sc->priority;
2482
2483 /*
2484 * If the cgroup's already been deleted, make sure to
2485 * scrape out the remaining cache.
2486 */
2487 if (!scan && !mem_cgroup_online(memcg))
2488 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2489
2490 switch (scan_balance) {
2491 case SCAN_EQUAL:
2492 /* Scan lists relative to size */
2493 break;
2494 case SCAN_FRACT:
2495 /*
2496 * Scan types proportional to swappiness and
2497 * their relative recent reclaim efficiency.
2498 * Make sure we don't miss the last page on
2499 * the offlined memory cgroups because of a
2500 * round-off error.
2501 */
2502 scan = mem_cgroup_online(memcg) ?
2503 div64_u64(scan * fraction[file], denominator) :
2504 DIV64_U64_ROUND_UP(scan * fraction[file],
2505 denominator);
2506 break;
2507 case SCAN_FILE:
2508 case SCAN_ANON:
2509 /* Scan one type exclusively */
2510 if ((scan_balance == SCAN_FILE) != file)
2511 scan = 0;
2512 break;
2513 default:
2514 /* Look ma, no brain */
2515 BUG();
2516 }
2517
2518 nr[lru] = scan;
2519 }
2520}
2521
2522/*
2523 * Anonymous LRU management is a waste if there is
2524 * ultimately no way to reclaim the memory.
2525 */
2526static bool can_age_anon_pages(struct pglist_data *pgdat,
2527 struct scan_control *sc)
2528{
2529 /* Aging the anon LRU is valuable if swap is present: */
2530 if (total_swap_pages > 0)
2531 return true;
2532
2533 /* Also valuable if anon pages can be demoted: */
2534 return can_demote(pgdat->node_id, sc);
2535}
2536
2537#ifdef CONFIG_LRU_GEN
2538
2539#ifdef CONFIG_LRU_GEN_ENABLED
2540DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2541#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2542#else
2543DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2544#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2545#endif
2546
2547static bool should_walk_mmu(void)
2548{
2549 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2550}
2551
2552static bool should_clear_pmd_young(void)
2553{
2554 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2555}
2556
2557/******************************************************************************
2558 * shorthand helpers
2559 ******************************************************************************/
2560
2561#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2562
2563#define DEFINE_MAX_SEQ(lruvec) \
2564 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2565
2566#define DEFINE_MIN_SEQ(lruvec) \
2567 unsigned long min_seq[ANON_AND_FILE] = { \
2568 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2569 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2570 }
2571
2572#define for_each_gen_type_zone(gen, type, zone) \
2573 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2574 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2575 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2576
2577#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2578#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2579
2580static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2581{
2582 struct pglist_data *pgdat = NODE_DATA(nid);
2583
2584#ifdef CONFIG_MEMCG
2585 if (memcg) {
2586 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2587
2588 /* see the comment in mem_cgroup_lruvec() */
2589 if (!lruvec->pgdat)
2590 lruvec->pgdat = pgdat;
2591
2592 return lruvec;
2593 }
2594#endif
2595 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2596
2597 return &pgdat->__lruvec;
2598}
2599
2600static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2601{
2602 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2603 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2604
2605 if (!sc->may_swap)
2606 return 0;
2607
2608 if (!can_demote(pgdat->node_id, sc) &&
2609 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2610 return 0;
2611
2612 return mem_cgroup_swappiness(memcg);
2613}
2614
2615static int get_nr_gens(struct lruvec *lruvec, int type)
2616{
2617 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2618}
2619
2620static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2621{
2622 /* see the comment on lru_gen_folio */
2623 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2624 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2625 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2626}
2627
2628/******************************************************************************
2629 * Bloom filters
2630 ******************************************************************************/
2631
2632/*
2633 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2634 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2635 * bits in a bitmap, k is the number of hash functions and n is the number of
2636 * inserted items.
2637 *
2638 * Page table walkers use one of the two filters to reduce their search space.
2639 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2640 * aging uses the double-buffering technique to flip to the other filter each
2641 * time it produces a new generation. For non-leaf entries that have enough
2642 * leaf entries, the aging carries them over to the next generation in
2643 * walk_pmd_range(); the eviction also report them when walking the rmap
2644 * in lru_gen_look_around().
2645 *
2646 * For future optimizations:
2647 * 1. It's not necessary to keep both filters all the time. The spare one can be
2648 * freed after the RCU grace period and reallocated if needed again.
2649 * 2. And when reallocating, it's worth scaling its size according to the number
2650 * of inserted entries in the other filter, to reduce the memory overhead on
2651 * small systems and false positives on large systems.
2652 * 3. Jenkins' hash function is an alternative to Knuth's.
2653 */
2654#define BLOOM_FILTER_SHIFT 15
2655
2656static inline int filter_gen_from_seq(unsigned long seq)
2657{
2658 return seq % NR_BLOOM_FILTERS;
2659}
2660
2661static void get_item_key(void *item, int *key)
2662{
2663 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2664
2665 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2666
2667 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2668 key[1] = hash >> BLOOM_FILTER_SHIFT;
2669}
2670
2671static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2672 void *item)
2673{
2674 int key[2];
2675 unsigned long *filter;
2676 int gen = filter_gen_from_seq(seq);
2677
2678 filter = READ_ONCE(mm_state->filters[gen]);
2679 if (!filter)
2680 return true;
2681
2682 get_item_key(item, key);
2683
2684 return test_bit(key[0], filter) && test_bit(key[1], filter);
2685}
2686
2687static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2688 void *item)
2689{
2690 int key[2];
2691 unsigned long *filter;
2692 int gen = filter_gen_from_seq(seq);
2693
2694 filter = READ_ONCE(mm_state->filters[gen]);
2695 if (!filter)
2696 return;
2697
2698 get_item_key(item, key);
2699
2700 if (!test_bit(key[0], filter))
2701 set_bit(key[0], filter);
2702 if (!test_bit(key[1], filter))
2703 set_bit(key[1], filter);
2704}
2705
2706static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2707{
2708 unsigned long *filter;
2709 int gen = filter_gen_from_seq(seq);
2710
2711 filter = mm_state->filters[gen];
2712 if (filter) {
2713 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2714 return;
2715 }
2716
2717 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2718 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2719 WRITE_ONCE(mm_state->filters[gen], filter);
2720}
2721
2722/******************************************************************************
2723 * mm_struct list
2724 ******************************************************************************/
2725
2726#ifdef CONFIG_LRU_GEN_WALKS_MMU
2727
2728static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2729{
2730 static struct lru_gen_mm_list mm_list = {
2731 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2732 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2733 };
2734
2735#ifdef CONFIG_MEMCG
2736 if (memcg)
2737 return &memcg->mm_list;
2738#endif
2739 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2740
2741 return &mm_list;
2742}
2743
2744static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2745{
2746 return &lruvec->mm_state;
2747}
2748
2749static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2750{
2751 int key;
2752 struct mm_struct *mm;
2753 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2754 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2755
2756 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2757 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2758
2759 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2760 return NULL;
2761
2762 clear_bit(key, &mm->lru_gen.bitmap);
2763
2764 return mmget_not_zero(mm) ? mm : NULL;
2765}
2766
2767void lru_gen_add_mm(struct mm_struct *mm)
2768{
2769 int nid;
2770 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2771 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2772
2773 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2774#ifdef CONFIG_MEMCG
2775 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2776 mm->lru_gen.memcg = memcg;
2777#endif
2778 spin_lock(&mm_list->lock);
2779
2780 for_each_node_state(nid, N_MEMORY) {
2781 struct lruvec *lruvec = get_lruvec(memcg, nid);
2782 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2783
2784 /* the first addition since the last iteration */
2785 if (mm_state->tail == &mm_list->fifo)
2786 mm_state->tail = &mm->lru_gen.list;
2787 }
2788
2789 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2790
2791 spin_unlock(&mm_list->lock);
2792}
2793
2794void lru_gen_del_mm(struct mm_struct *mm)
2795{
2796 int nid;
2797 struct lru_gen_mm_list *mm_list;
2798 struct mem_cgroup *memcg = NULL;
2799
2800 if (list_empty(&mm->lru_gen.list))
2801 return;
2802
2803#ifdef CONFIG_MEMCG
2804 memcg = mm->lru_gen.memcg;
2805#endif
2806 mm_list = get_mm_list(memcg);
2807
2808 spin_lock(&mm_list->lock);
2809
2810 for_each_node(nid) {
2811 struct lruvec *lruvec = get_lruvec(memcg, nid);
2812 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2813
2814 /* where the current iteration continues after */
2815 if (mm_state->head == &mm->lru_gen.list)
2816 mm_state->head = mm_state->head->prev;
2817
2818 /* where the last iteration ended before */
2819 if (mm_state->tail == &mm->lru_gen.list)
2820 mm_state->tail = mm_state->tail->next;
2821 }
2822
2823 list_del_init(&mm->lru_gen.list);
2824
2825 spin_unlock(&mm_list->lock);
2826
2827#ifdef CONFIG_MEMCG
2828 mem_cgroup_put(mm->lru_gen.memcg);
2829 mm->lru_gen.memcg = NULL;
2830#endif
2831}
2832
2833#ifdef CONFIG_MEMCG
2834void lru_gen_migrate_mm(struct mm_struct *mm)
2835{
2836 struct mem_cgroup *memcg;
2837 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2838
2839 VM_WARN_ON_ONCE(task->mm != mm);
2840 lockdep_assert_held(&task->alloc_lock);
2841
2842 /* for mm_update_next_owner() */
2843 if (mem_cgroup_disabled())
2844 return;
2845
2846 /* migration can happen before addition */
2847 if (!mm->lru_gen.memcg)
2848 return;
2849
2850 rcu_read_lock();
2851 memcg = mem_cgroup_from_task(task);
2852 rcu_read_unlock();
2853 if (memcg == mm->lru_gen.memcg)
2854 return;
2855
2856 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2857
2858 lru_gen_del_mm(mm);
2859 lru_gen_add_mm(mm);
2860}
2861#endif
2862
2863#else /* !CONFIG_LRU_GEN_WALKS_MMU */
2864
2865static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2866{
2867 return NULL;
2868}
2869
2870static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2871{
2872 return NULL;
2873}
2874
2875static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2876{
2877 return NULL;
2878}
2879
2880#endif
2881
2882static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
2883{
2884 int i;
2885 int hist;
2886 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2887
2888 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2889
2890 if (walk) {
2891 hist = lru_hist_from_seq(walk->max_seq);
2892
2893 for (i = 0; i < NR_MM_STATS; i++) {
2894 WRITE_ONCE(mm_state->stats[hist][i],
2895 mm_state->stats[hist][i] + walk->mm_stats[i]);
2896 walk->mm_stats[i] = 0;
2897 }
2898 }
2899
2900 if (NR_HIST_GENS > 1 && last) {
2901 hist = lru_hist_from_seq(mm_state->seq + 1);
2902
2903 for (i = 0; i < NR_MM_STATS; i++)
2904 WRITE_ONCE(mm_state->stats[hist][i], 0);
2905 }
2906}
2907
2908static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
2909 struct mm_struct **iter)
2910{
2911 bool first = false;
2912 bool last = false;
2913 struct mm_struct *mm = NULL;
2914 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2915 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2916 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2917
2918 /*
2919 * mm_state->seq is incremented after each iteration of mm_list. There
2920 * are three interesting cases for this page table walker:
2921 * 1. It tries to start a new iteration with a stale max_seq: there is
2922 * nothing left to do.
2923 * 2. It started the next iteration: it needs to reset the Bloom filter
2924 * so that a fresh set of PTE tables can be recorded.
2925 * 3. It ended the current iteration: it needs to reset the mm stats
2926 * counters and tell its caller to increment max_seq.
2927 */
2928 spin_lock(&mm_list->lock);
2929
2930 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
2931
2932 if (walk->max_seq <= mm_state->seq)
2933 goto done;
2934
2935 if (!mm_state->head)
2936 mm_state->head = &mm_list->fifo;
2937
2938 if (mm_state->head == &mm_list->fifo)
2939 first = true;
2940
2941 do {
2942 mm_state->head = mm_state->head->next;
2943 if (mm_state->head == &mm_list->fifo) {
2944 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2945 last = true;
2946 break;
2947 }
2948
2949 /* force scan for those added after the last iteration */
2950 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2951 mm_state->tail = mm_state->head->next;
2952 walk->force_scan = true;
2953 }
2954 } while (!(mm = get_next_mm(walk)));
2955done:
2956 if (*iter || last)
2957 reset_mm_stats(lruvec, walk, last);
2958
2959 spin_unlock(&mm_list->lock);
2960
2961 if (mm && first)
2962 reset_bloom_filter(mm_state, walk->max_seq + 1);
2963
2964 if (*iter)
2965 mmput_async(*iter);
2966
2967 *iter = mm;
2968
2969 return last;
2970}
2971
2972static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
2973{
2974 bool success = false;
2975 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2976 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2977 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2978
2979 spin_lock(&mm_list->lock);
2980
2981 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
2982
2983 if (max_seq > mm_state->seq) {
2984 mm_state->head = NULL;
2985 mm_state->tail = NULL;
2986 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2987 reset_mm_stats(lruvec, NULL, true);
2988 success = true;
2989 }
2990
2991 spin_unlock(&mm_list->lock);
2992
2993 return success;
2994}
2995
2996/******************************************************************************
2997 * PID controller
2998 ******************************************************************************/
2999
3000/*
3001 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3002 *
3003 * The P term is refaulted/(evicted+protected) from a tier in the generation
3004 * currently being evicted; the I term is the exponential moving average of the
3005 * P term over the generations previously evicted, using the smoothing factor
3006 * 1/2; the D term isn't supported.
3007 *
3008 * The setpoint (SP) is always the first tier of one type; the process variable
3009 * (PV) is either any tier of the other type or any other tier of the same
3010 * type.
3011 *
3012 * The error is the difference between the SP and the PV; the correction is to
3013 * turn off protection when SP>PV or turn on protection when SP<PV.
3014 *
3015 * For future optimizations:
3016 * 1. The D term may discount the other two terms over time so that long-lived
3017 * generations can resist stale information.
3018 */
3019struct ctrl_pos {
3020 unsigned long refaulted;
3021 unsigned long total;
3022 int gain;
3023};
3024
3025static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3026 struct ctrl_pos *pos)
3027{
3028 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3029 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3030
3031 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3032 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3033 pos->total = lrugen->avg_total[type][tier] +
3034 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3035 if (tier)
3036 pos->total += lrugen->protected[hist][type][tier - 1];
3037 pos->gain = gain;
3038}
3039
3040static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3041{
3042 int hist, tier;
3043 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3044 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3045 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3046
3047 lockdep_assert_held(&lruvec->lru_lock);
3048
3049 if (!carryover && !clear)
3050 return;
3051
3052 hist = lru_hist_from_seq(seq);
3053
3054 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3055 if (carryover) {
3056 unsigned long sum;
3057
3058 sum = lrugen->avg_refaulted[type][tier] +
3059 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3060 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3061
3062 sum = lrugen->avg_total[type][tier] +
3063 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3064 if (tier)
3065 sum += lrugen->protected[hist][type][tier - 1];
3066 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3067 }
3068
3069 if (clear) {
3070 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3071 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3072 if (tier)
3073 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3074 }
3075 }
3076}
3077
3078static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3079{
3080 /*
3081 * Return true if the PV has a limited number of refaults or a lower
3082 * refaulted/total than the SP.
3083 */
3084 return pv->refaulted < MIN_LRU_BATCH ||
3085 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3086 (sp->refaulted + 1) * pv->total * pv->gain;
3087}
3088
3089/******************************************************************************
3090 * the aging
3091 ******************************************************************************/
3092
3093/* promote pages accessed through page tables */
3094static int folio_update_gen(struct folio *folio, int gen)
3095{
3096 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3097
3098 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3099 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3100
3101 do {
3102 /* lru_gen_del_folio() has isolated this page? */
3103 if (!(old_flags & LRU_GEN_MASK)) {
3104 /* for shrink_folio_list() */
3105 new_flags = old_flags | BIT(PG_referenced);
3106 continue;
3107 }
3108
3109 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3110 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3111 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3112
3113 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3114}
3115
3116/* protect pages accessed multiple times through file descriptors */
3117static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3118{
3119 int type = folio_is_file_lru(folio);
3120 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3121 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3122 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3123
3124 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3125
3126 do {
3127 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3128 /* folio_update_gen() has promoted this page? */
3129 if (new_gen >= 0 && new_gen != old_gen)
3130 return new_gen;
3131
3132 new_gen = (old_gen + 1) % MAX_NR_GENS;
3133
3134 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3135 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3136 /* for folio_end_writeback() */
3137 if (reclaiming)
3138 new_flags |= BIT(PG_reclaim);
3139 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3140
3141 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3142
3143 return new_gen;
3144}
3145
3146static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3147 int old_gen, int new_gen)
3148{
3149 int type = folio_is_file_lru(folio);
3150 int zone = folio_zonenum(folio);
3151 int delta = folio_nr_pages(folio);
3152
3153 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3154 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3155
3156 walk->batched++;
3157
3158 walk->nr_pages[old_gen][type][zone] -= delta;
3159 walk->nr_pages[new_gen][type][zone] += delta;
3160}
3161
3162static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3163{
3164 int gen, type, zone;
3165 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3166
3167 walk->batched = 0;
3168
3169 for_each_gen_type_zone(gen, type, zone) {
3170 enum lru_list lru = type * LRU_INACTIVE_FILE;
3171 int delta = walk->nr_pages[gen][type][zone];
3172
3173 if (!delta)
3174 continue;
3175
3176 walk->nr_pages[gen][type][zone] = 0;
3177 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3178 lrugen->nr_pages[gen][type][zone] + delta);
3179
3180 if (lru_gen_is_active(lruvec, gen))
3181 lru += LRU_ACTIVE;
3182 __update_lru_size(lruvec, lru, zone, delta);
3183 }
3184}
3185
3186static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3187{
3188 struct address_space *mapping;
3189 struct vm_area_struct *vma = args->vma;
3190 struct lru_gen_mm_walk *walk = args->private;
3191
3192 if (!vma_is_accessible(vma))
3193 return true;
3194
3195 if (is_vm_hugetlb_page(vma))
3196 return true;
3197
3198 if (!vma_has_recency(vma))
3199 return true;
3200
3201 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3202 return true;
3203
3204 if (vma == get_gate_vma(vma->vm_mm))
3205 return true;
3206
3207 if (vma_is_anonymous(vma))
3208 return !walk->can_swap;
3209
3210 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3211 return true;
3212
3213 mapping = vma->vm_file->f_mapping;
3214 if (mapping_unevictable(mapping))
3215 return true;
3216
3217 if (shmem_mapping(mapping))
3218 return !walk->can_swap;
3219
3220 /* to exclude special mappings like dax, etc. */
3221 return !mapping->a_ops->read_folio;
3222}
3223
3224/*
3225 * Some userspace memory allocators map many single-page VMAs. Instead of
3226 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3227 * table to reduce zigzags and improve cache performance.
3228 */
3229static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3230 unsigned long *vm_start, unsigned long *vm_end)
3231{
3232 unsigned long start = round_up(*vm_end, size);
3233 unsigned long end = (start | ~mask) + 1;
3234 VMA_ITERATOR(vmi, args->mm, start);
3235
3236 VM_WARN_ON_ONCE(mask & size);
3237 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3238
3239 for_each_vma(vmi, args->vma) {
3240 if (end && end <= args->vma->vm_start)
3241 return false;
3242
3243 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3244 continue;
3245
3246 *vm_start = max(start, args->vma->vm_start);
3247 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3248
3249 return true;
3250 }
3251
3252 return false;
3253}
3254
3255static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3256{
3257 unsigned long pfn = pte_pfn(pte);
3258
3259 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3260
3261 if (!pte_present(pte) || is_zero_pfn(pfn))
3262 return -1;
3263
3264 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3265 return -1;
3266
3267 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3268 return -1;
3269
3270 return pfn;
3271}
3272
3273static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3274{
3275 unsigned long pfn = pmd_pfn(pmd);
3276
3277 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3278
3279 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3280 return -1;
3281
3282 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3283 return -1;
3284
3285 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3286 return -1;
3287
3288 return pfn;
3289}
3290
3291static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3292 struct pglist_data *pgdat, bool can_swap)
3293{
3294 struct folio *folio;
3295
3296 /* try to avoid unnecessary memory loads */
3297 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3298 return NULL;
3299
3300 folio = pfn_folio(pfn);
3301 if (folio_nid(folio) != pgdat->node_id)
3302 return NULL;
3303
3304 if (folio_memcg_rcu(folio) != memcg)
3305 return NULL;
3306
3307 /* file VMAs can contain anon pages from COW */
3308 if (!folio_is_file_lru(folio) && !can_swap)
3309 return NULL;
3310
3311 return folio;
3312}
3313
3314static bool suitable_to_scan(int total, int young)
3315{
3316 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3317
3318 /* suitable if the average number of young PTEs per cacheline is >=1 */
3319 return young * n >= total;
3320}
3321
3322static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3323 struct mm_walk *args)
3324{
3325 int i;
3326 pte_t *pte;
3327 spinlock_t *ptl;
3328 unsigned long addr;
3329 int total = 0;
3330 int young = 0;
3331 struct lru_gen_mm_walk *walk = args->private;
3332 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3333 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3334 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3335
3336 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3337 if (!pte)
3338 return false;
3339 if (!spin_trylock(ptl)) {
3340 pte_unmap(pte);
3341 return false;
3342 }
3343
3344 arch_enter_lazy_mmu_mode();
3345restart:
3346 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3347 unsigned long pfn;
3348 struct folio *folio;
3349 pte_t ptent = ptep_get(pte + i);
3350
3351 total++;
3352 walk->mm_stats[MM_LEAF_TOTAL]++;
3353
3354 pfn = get_pte_pfn(ptent, args->vma, addr);
3355 if (pfn == -1)
3356 continue;
3357
3358 if (!pte_young(ptent)) {
3359 walk->mm_stats[MM_LEAF_OLD]++;
3360 continue;
3361 }
3362
3363 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3364 if (!folio)
3365 continue;
3366
3367 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3368 VM_WARN_ON_ONCE(true);
3369
3370 young++;
3371 walk->mm_stats[MM_LEAF_YOUNG]++;
3372
3373 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3374 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3375 !folio_test_swapcache(folio)))
3376 folio_mark_dirty(folio);
3377
3378 old_gen = folio_update_gen(folio, new_gen);
3379 if (old_gen >= 0 && old_gen != new_gen)
3380 update_batch_size(walk, folio, old_gen, new_gen);
3381 }
3382
3383 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3384 goto restart;
3385
3386 arch_leave_lazy_mmu_mode();
3387 pte_unmap_unlock(pte, ptl);
3388
3389 return suitable_to_scan(total, young);
3390}
3391
3392static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3393 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3394{
3395 int i;
3396 pmd_t *pmd;
3397 spinlock_t *ptl;
3398 struct lru_gen_mm_walk *walk = args->private;
3399 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3400 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3401 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3402
3403 VM_WARN_ON_ONCE(pud_leaf(*pud));
3404
3405 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3406 if (*first == -1) {
3407 *first = addr;
3408 bitmap_zero(bitmap, MIN_LRU_BATCH);
3409 return;
3410 }
3411
3412 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3413 if (i && i <= MIN_LRU_BATCH) {
3414 __set_bit(i - 1, bitmap);
3415 return;
3416 }
3417
3418 pmd = pmd_offset(pud, *first);
3419
3420 ptl = pmd_lockptr(args->mm, pmd);
3421 if (!spin_trylock(ptl))
3422 goto done;
3423
3424 arch_enter_lazy_mmu_mode();
3425
3426 do {
3427 unsigned long pfn;
3428 struct folio *folio;
3429
3430 /* don't round down the first address */
3431 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3432
3433 pfn = get_pmd_pfn(pmd[i], vma, addr);
3434 if (pfn == -1)
3435 goto next;
3436
3437 if (!pmd_trans_huge(pmd[i])) {
3438 if (should_clear_pmd_young())
3439 pmdp_test_and_clear_young(vma, addr, pmd + i);
3440 goto next;
3441 }
3442
3443 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3444 if (!folio)
3445 goto next;
3446
3447 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3448 goto next;
3449
3450 walk->mm_stats[MM_LEAF_YOUNG]++;
3451
3452 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3453 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3454 !folio_test_swapcache(folio)))
3455 folio_mark_dirty(folio);
3456
3457 old_gen = folio_update_gen(folio, new_gen);
3458 if (old_gen >= 0 && old_gen != new_gen)
3459 update_batch_size(walk, folio, old_gen, new_gen);
3460next:
3461 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3462 } while (i <= MIN_LRU_BATCH);
3463
3464 arch_leave_lazy_mmu_mode();
3465 spin_unlock(ptl);
3466done:
3467 *first = -1;
3468}
3469
3470static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3471 struct mm_walk *args)
3472{
3473 int i;
3474 pmd_t *pmd;
3475 unsigned long next;
3476 unsigned long addr;
3477 struct vm_area_struct *vma;
3478 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3479 unsigned long first = -1;
3480 struct lru_gen_mm_walk *walk = args->private;
3481 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3482
3483 VM_WARN_ON_ONCE(pud_leaf(*pud));
3484
3485 /*
3486 * Finish an entire PMD in two passes: the first only reaches to PTE
3487 * tables to avoid taking the PMD lock; the second, if necessary, takes
3488 * the PMD lock to clear the accessed bit in PMD entries.
3489 */
3490 pmd = pmd_offset(pud, start & PUD_MASK);
3491restart:
3492 /* walk_pte_range() may call get_next_vma() */
3493 vma = args->vma;
3494 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3495 pmd_t val = pmdp_get_lockless(pmd + i);
3496
3497 next = pmd_addr_end(addr, end);
3498
3499 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3500 walk->mm_stats[MM_LEAF_TOTAL]++;
3501 continue;
3502 }
3503
3504 if (pmd_trans_huge(val)) {
3505 unsigned long pfn = pmd_pfn(val);
3506 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3507
3508 walk->mm_stats[MM_LEAF_TOTAL]++;
3509
3510 if (!pmd_young(val)) {
3511 walk->mm_stats[MM_LEAF_OLD]++;
3512 continue;
3513 }
3514
3515 /* try to avoid unnecessary memory loads */
3516 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3517 continue;
3518
3519 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3520 continue;
3521 }
3522
3523 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3524
3525 if (should_clear_pmd_young()) {
3526 if (!pmd_young(val))
3527 continue;
3528
3529 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3530 }
3531
3532 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->max_seq, pmd + i))
3533 continue;
3534
3535 walk->mm_stats[MM_NONLEAF_FOUND]++;
3536
3537 if (!walk_pte_range(&val, addr, next, args))
3538 continue;
3539
3540 walk->mm_stats[MM_NONLEAF_ADDED]++;
3541
3542 /* carry over to the next generation */
3543 update_bloom_filter(mm_state, walk->max_seq + 1, pmd + i);
3544 }
3545
3546 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3547
3548 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3549 goto restart;
3550}
3551
3552static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3553 struct mm_walk *args)
3554{
3555 int i;
3556 pud_t *pud;
3557 unsigned long addr;
3558 unsigned long next;
3559 struct lru_gen_mm_walk *walk = args->private;
3560
3561 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3562
3563 pud = pud_offset(p4d, start & P4D_MASK);
3564restart:
3565 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3566 pud_t val = READ_ONCE(pud[i]);
3567
3568 next = pud_addr_end(addr, end);
3569
3570 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3571 continue;
3572
3573 walk_pmd_range(&val, addr, next, args);
3574
3575 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3576 end = (addr | ~PUD_MASK) + 1;
3577 goto done;
3578 }
3579 }
3580
3581 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3582 goto restart;
3583
3584 end = round_up(end, P4D_SIZE);
3585done:
3586 if (!end || !args->vma)
3587 return 1;
3588
3589 walk->next_addr = max(end, args->vma->vm_start);
3590
3591 return -EAGAIN;
3592}
3593
3594static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3595{
3596 static const struct mm_walk_ops mm_walk_ops = {
3597 .test_walk = should_skip_vma,
3598 .p4d_entry = walk_pud_range,
3599 .walk_lock = PGWALK_RDLOCK,
3600 };
3601
3602 int err;
3603 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3604
3605 walk->next_addr = FIRST_USER_ADDRESS;
3606
3607 do {
3608 DEFINE_MAX_SEQ(lruvec);
3609
3610 err = -EBUSY;
3611
3612 /* another thread might have called inc_max_seq() */
3613 if (walk->max_seq != max_seq)
3614 break;
3615
3616 /* folio_update_gen() requires stable folio_memcg() */
3617 if (!mem_cgroup_trylock_pages(memcg))
3618 break;
3619
3620 /* the caller might be holding the lock for write */
3621 if (mmap_read_trylock(mm)) {
3622 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3623
3624 mmap_read_unlock(mm);
3625 }
3626
3627 mem_cgroup_unlock_pages();
3628
3629 if (walk->batched) {
3630 spin_lock_irq(&lruvec->lru_lock);
3631 reset_batch_size(lruvec, walk);
3632 spin_unlock_irq(&lruvec->lru_lock);
3633 }
3634
3635 cond_resched();
3636 } while (err == -EAGAIN);
3637}
3638
3639static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3640{
3641 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3642
3643 if (pgdat && current_is_kswapd()) {
3644 VM_WARN_ON_ONCE(walk);
3645
3646 walk = &pgdat->mm_walk;
3647 } else if (!walk && force_alloc) {
3648 VM_WARN_ON_ONCE(current_is_kswapd());
3649
3650 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3651 }
3652
3653 current->reclaim_state->mm_walk = walk;
3654
3655 return walk;
3656}
3657
3658static void clear_mm_walk(void)
3659{
3660 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3661
3662 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3663 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3664
3665 current->reclaim_state->mm_walk = NULL;
3666
3667 if (!current_is_kswapd())
3668 kfree(walk);
3669}
3670
3671static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3672{
3673 int zone;
3674 int remaining = MAX_LRU_BATCH;
3675 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3676 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3677
3678 if (type == LRU_GEN_ANON && !can_swap)
3679 goto done;
3680
3681 /* prevent cold/hot inversion if force_scan is true */
3682 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3683 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3684
3685 while (!list_empty(head)) {
3686 struct folio *folio = lru_to_folio(head);
3687
3688 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3689 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3690 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3691 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3692
3693 new_gen = folio_inc_gen(lruvec, folio, false);
3694 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3695
3696 if (!--remaining)
3697 return false;
3698 }
3699 }
3700done:
3701 reset_ctrl_pos(lruvec, type, true);
3702 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3703
3704 return true;
3705}
3706
3707static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3708{
3709 int gen, type, zone;
3710 bool success = false;
3711 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3712 DEFINE_MIN_SEQ(lruvec);
3713
3714 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3715
3716 /* find the oldest populated generation */
3717 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3718 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3719 gen = lru_gen_from_seq(min_seq[type]);
3720
3721 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3722 if (!list_empty(&lrugen->folios[gen][type][zone]))
3723 goto next;
3724 }
3725
3726 min_seq[type]++;
3727 }
3728next:
3729 ;
3730 }
3731
3732 /* see the comment on lru_gen_folio */
3733 if (can_swap) {
3734 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3735 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3736 }
3737
3738 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3739 if (min_seq[type] == lrugen->min_seq[type])
3740 continue;
3741
3742 reset_ctrl_pos(lruvec, type, true);
3743 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3744 success = true;
3745 }
3746
3747 return success;
3748}
3749
3750static bool inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
3751 bool can_swap, bool force_scan)
3752{
3753 bool success;
3754 int prev, next;
3755 int type, zone;
3756 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3757restart:
3758 if (max_seq < READ_ONCE(lrugen->max_seq))
3759 return false;
3760
3761 spin_lock_irq(&lruvec->lru_lock);
3762
3763 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3764
3765 success = max_seq == lrugen->max_seq;
3766 if (!success)
3767 goto unlock;
3768
3769 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3770 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3771 continue;
3772
3773 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3774
3775 if (inc_min_seq(lruvec, type, can_swap))
3776 continue;
3777
3778 spin_unlock_irq(&lruvec->lru_lock);
3779 cond_resched();
3780 goto restart;
3781 }
3782
3783 /*
3784 * Update the active/inactive LRU sizes for compatibility. Both sides of
3785 * the current max_seq need to be covered, since max_seq+1 can overlap
3786 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3787 * overlap, cold/hot inversion happens.
3788 */
3789 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3790 next = lru_gen_from_seq(lrugen->max_seq + 1);
3791
3792 for (type = 0; type < ANON_AND_FILE; type++) {
3793 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3794 enum lru_list lru = type * LRU_INACTIVE_FILE;
3795 long delta = lrugen->nr_pages[prev][type][zone] -
3796 lrugen->nr_pages[next][type][zone];
3797
3798 if (!delta)
3799 continue;
3800
3801 __update_lru_size(lruvec, lru, zone, delta);
3802 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3803 }
3804 }
3805
3806 for (type = 0; type < ANON_AND_FILE; type++)
3807 reset_ctrl_pos(lruvec, type, false);
3808
3809 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3810 /* make sure preceding modifications appear */
3811 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3812unlock:
3813 spin_unlock_irq(&lruvec->lru_lock);
3814
3815 return success;
3816}
3817
3818static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
3819 struct scan_control *sc, bool can_swap, bool force_scan)
3820{
3821 bool success;
3822 struct lru_gen_mm_walk *walk;
3823 struct mm_struct *mm = NULL;
3824 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3825 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3826
3827 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
3828
3829 if (!mm_state)
3830 return inc_max_seq(lruvec, max_seq, can_swap, force_scan);
3831
3832 /* see the comment in iterate_mm_list() */
3833 if (max_seq <= READ_ONCE(mm_state->seq))
3834 return false;
3835
3836 /*
3837 * If the hardware doesn't automatically set the accessed bit, fallback
3838 * to lru_gen_look_around(), which only clears the accessed bit in a
3839 * handful of PTEs. Spreading the work out over a period of time usually
3840 * is less efficient, but it avoids bursty page faults.
3841 */
3842 if (!should_walk_mmu()) {
3843 success = iterate_mm_list_nowalk(lruvec, max_seq);
3844 goto done;
3845 }
3846
3847 walk = set_mm_walk(NULL, true);
3848 if (!walk) {
3849 success = iterate_mm_list_nowalk(lruvec, max_seq);
3850 goto done;
3851 }
3852
3853 walk->lruvec = lruvec;
3854 walk->max_seq = max_seq;
3855 walk->can_swap = can_swap;
3856 walk->force_scan = force_scan;
3857
3858 do {
3859 success = iterate_mm_list(lruvec, walk, &mm);
3860 if (mm)
3861 walk_mm(lruvec, mm, walk);
3862 } while (mm);
3863done:
3864 if (success) {
3865 success = inc_max_seq(lruvec, max_seq, can_swap, force_scan);
3866 WARN_ON_ONCE(!success);
3867 }
3868
3869 return success;
3870}
3871
3872/******************************************************************************
3873 * working set protection
3874 ******************************************************************************/
3875
3876static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3877{
3878 int gen, type, zone;
3879 unsigned long total = 0;
3880 bool can_swap = get_swappiness(lruvec, sc);
3881 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3882 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3883 DEFINE_MAX_SEQ(lruvec);
3884 DEFINE_MIN_SEQ(lruvec);
3885
3886 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3887 unsigned long seq;
3888
3889 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3890 gen = lru_gen_from_seq(seq);
3891
3892 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3893 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3894 }
3895 }
3896
3897 /* whether the size is big enough to be helpful */
3898 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3899}
3900
3901static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3902 unsigned long min_ttl)
3903{
3904 int gen;
3905 unsigned long birth;
3906 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3907 DEFINE_MIN_SEQ(lruvec);
3908
3909 /* see the comment on lru_gen_folio */
3910 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3911 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3912
3913 if (time_is_after_jiffies(birth + min_ttl))
3914 return false;
3915
3916 if (!lruvec_is_sizable(lruvec, sc))
3917 return false;
3918
3919 mem_cgroup_calculate_protection(NULL, memcg);
3920
3921 return !mem_cgroup_below_min(NULL, memcg);
3922}
3923
3924/* to protect the working set of the last N jiffies */
3925static unsigned long lru_gen_min_ttl __read_mostly;
3926
3927static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3928{
3929 struct mem_cgroup *memcg;
3930 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3931
3932 VM_WARN_ON_ONCE(!current_is_kswapd());
3933
3934 /* check the order to exclude compaction-induced reclaim */
3935 if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
3936 return;
3937
3938 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3939 do {
3940 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3941
3942 if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
3943 mem_cgroup_iter_break(NULL, memcg);
3944 return;
3945 }
3946
3947 cond_resched();
3948 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
3949
3950 /*
3951 * The main goal is to OOM kill if every generation from all memcgs is
3952 * younger than min_ttl. However, another possibility is all memcgs are
3953 * either too small or below min.
3954 */
3955 if (mutex_trylock(&oom_lock)) {
3956 struct oom_control oc = {
3957 .gfp_mask = sc->gfp_mask,
3958 };
3959
3960 out_of_memory(&oc);
3961
3962 mutex_unlock(&oom_lock);
3963 }
3964}
3965
3966/******************************************************************************
3967 * rmap/PT walk feedback
3968 ******************************************************************************/
3969
3970/*
3971 * This function exploits spatial locality when shrink_folio_list() walks the
3972 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3973 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3974 * the PTE table to the Bloom filter. This forms a feedback loop between the
3975 * eviction and the aging.
3976 */
3977void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
3978{
3979 int i;
3980 unsigned long start;
3981 unsigned long end;
3982 struct lru_gen_mm_walk *walk;
3983 int young = 0;
3984 pte_t *pte = pvmw->pte;
3985 unsigned long addr = pvmw->address;
3986 struct vm_area_struct *vma = pvmw->vma;
3987 struct folio *folio = pfn_folio(pvmw->pfn);
3988 bool can_swap = !folio_is_file_lru(folio);
3989 struct mem_cgroup *memcg = folio_memcg(folio);
3990 struct pglist_data *pgdat = folio_pgdat(folio);
3991 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3992 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3993 DEFINE_MAX_SEQ(lruvec);
3994 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3995
3996 lockdep_assert_held(pvmw->ptl);
3997 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
3998
3999 if (spin_is_contended(pvmw->ptl))
4000 return;
4001
4002 /* exclude special VMAs containing anon pages from COW */
4003 if (vma->vm_flags & VM_SPECIAL)
4004 return;
4005
4006 /* avoid taking the LRU lock under the PTL when possible */
4007 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4008
4009 start = max(addr & PMD_MASK, vma->vm_start);
4010 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4011
4012 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4013 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4014 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4015 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4016 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4017 else {
4018 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4019 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4020 }
4021 }
4022
4023 /* folio_update_gen() requires stable folio_memcg() */
4024 if (!mem_cgroup_trylock_pages(memcg))
4025 return;
4026
4027 arch_enter_lazy_mmu_mode();
4028
4029 pte -= (addr - start) / PAGE_SIZE;
4030
4031 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4032 unsigned long pfn;
4033 pte_t ptent = ptep_get(pte + i);
4034
4035 pfn = get_pte_pfn(ptent, vma, addr);
4036 if (pfn == -1)
4037 continue;
4038
4039 if (!pte_young(ptent))
4040 continue;
4041
4042 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4043 if (!folio)
4044 continue;
4045
4046 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4047 VM_WARN_ON_ONCE(true);
4048
4049 young++;
4050
4051 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4052 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4053 !folio_test_swapcache(folio)))
4054 folio_mark_dirty(folio);
4055
4056 if (walk) {
4057 old_gen = folio_update_gen(folio, new_gen);
4058 if (old_gen >= 0 && old_gen != new_gen)
4059 update_batch_size(walk, folio, old_gen, new_gen);
4060
4061 continue;
4062 }
4063
4064 old_gen = folio_lru_gen(folio);
4065 if (old_gen < 0)
4066 folio_set_referenced(folio);
4067 else if (old_gen != new_gen)
4068 folio_activate(folio);
4069 }
4070
4071 arch_leave_lazy_mmu_mode();
4072 mem_cgroup_unlock_pages();
4073
4074 /* feedback from rmap walkers to page table walkers */
4075 if (mm_state && suitable_to_scan(i, young))
4076 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4077}
4078
4079/******************************************************************************
4080 * memcg LRU
4081 ******************************************************************************/
4082
4083/* see the comment on MEMCG_NR_GENS */
4084enum {
4085 MEMCG_LRU_NOP,
4086 MEMCG_LRU_HEAD,
4087 MEMCG_LRU_TAIL,
4088 MEMCG_LRU_OLD,
4089 MEMCG_LRU_YOUNG,
4090};
4091
4092static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4093{
4094 int seg;
4095 int old, new;
4096 unsigned long flags;
4097 int bin = get_random_u32_below(MEMCG_NR_BINS);
4098 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4099
4100 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4101
4102 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4103
4104 seg = 0;
4105 new = old = lruvec->lrugen.gen;
4106
4107 /* see the comment on MEMCG_NR_GENS */
4108 if (op == MEMCG_LRU_HEAD)
4109 seg = MEMCG_LRU_HEAD;
4110 else if (op == MEMCG_LRU_TAIL)
4111 seg = MEMCG_LRU_TAIL;
4112 else if (op == MEMCG_LRU_OLD)
4113 new = get_memcg_gen(pgdat->memcg_lru.seq);
4114 else if (op == MEMCG_LRU_YOUNG)
4115 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4116 else
4117 VM_WARN_ON_ONCE(true);
4118
4119 WRITE_ONCE(lruvec->lrugen.seg, seg);
4120 WRITE_ONCE(lruvec->lrugen.gen, new);
4121
4122 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4123
4124 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4125 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4126 else
4127 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4128
4129 pgdat->memcg_lru.nr_memcgs[old]--;
4130 pgdat->memcg_lru.nr_memcgs[new]++;
4131
4132 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4133 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4134
4135 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4136}
4137
4138#ifdef CONFIG_MEMCG
4139
4140void lru_gen_online_memcg(struct mem_cgroup *memcg)
4141{
4142 int gen;
4143 int nid;
4144 int bin = get_random_u32_below(MEMCG_NR_BINS);
4145
4146 for_each_node(nid) {
4147 struct pglist_data *pgdat = NODE_DATA(nid);
4148 struct lruvec *lruvec = get_lruvec(memcg, nid);
4149
4150 spin_lock_irq(&pgdat->memcg_lru.lock);
4151
4152 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4153
4154 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4155
4156 lruvec->lrugen.gen = gen;
4157
4158 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4159 pgdat->memcg_lru.nr_memcgs[gen]++;
4160
4161 spin_unlock_irq(&pgdat->memcg_lru.lock);
4162 }
4163}
4164
4165void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4166{
4167 int nid;
4168
4169 for_each_node(nid) {
4170 struct lruvec *lruvec = get_lruvec(memcg, nid);
4171
4172 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4173 }
4174}
4175
4176void lru_gen_release_memcg(struct mem_cgroup *memcg)
4177{
4178 int gen;
4179 int nid;
4180
4181 for_each_node(nid) {
4182 struct pglist_data *pgdat = NODE_DATA(nid);
4183 struct lruvec *lruvec = get_lruvec(memcg, nid);
4184
4185 spin_lock_irq(&pgdat->memcg_lru.lock);
4186
4187 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4188 goto unlock;
4189
4190 gen = lruvec->lrugen.gen;
4191
4192 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4193 pgdat->memcg_lru.nr_memcgs[gen]--;
4194
4195 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4196 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4197unlock:
4198 spin_unlock_irq(&pgdat->memcg_lru.lock);
4199 }
4200}
4201
4202void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4203{
4204 struct lruvec *lruvec = get_lruvec(memcg, nid);
4205
4206 /* see the comment on MEMCG_NR_GENS */
4207 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4208 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4209}
4210
4211#endif /* CONFIG_MEMCG */
4212
4213/******************************************************************************
4214 * the eviction
4215 ******************************************************************************/
4216
4217static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4218 int tier_idx)
4219{
4220 bool success;
4221 int gen = folio_lru_gen(folio);
4222 int type = folio_is_file_lru(folio);
4223 int zone = folio_zonenum(folio);
4224 int delta = folio_nr_pages(folio);
4225 int refs = folio_lru_refs(folio);
4226 int tier = lru_tier_from_refs(refs);
4227 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4228
4229 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4230
4231 /* unevictable */
4232 if (!folio_evictable(folio)) {
4233 success = lru_gen_del_folio(lruvec, folio, true);
4234 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4235 folio_set_unevictable(folio);
4236 lruvec_add_folio(lruvec, folio);
4237 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4238 return true;
4239 }
4240
4241 /* dirty lazyfree */
4242 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4243 success = lru_gen_del_folio(lruvec, folio, true);
4244 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4245 folio_set_swapbacked(folio);
4246 lruvec_add_folio_tail(lruvec, folio);
4247 return true;
4248 }
4249
4250 /* promoted */
4251 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4252 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4253 return true;
4254 }
4255
4256 /* protected */
4257 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4258 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4259
4260 gen = folio_inc_gen(lruvec, folio, false);
4261 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4262
4263 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4264 lrugen->protected[hist][type][tier - 1] + delta);
4265 return true;
4266 }
4267
4268 /* ineligible */
4269 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4270 gen = folio_inc_gen(lruvec, folio, false);
4271 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4272 return true;
4273 }
4274
4275 /* waiting for writeback */
4276 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4277 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4278 gen = folio_inc_gen(lruvec, folio, true);
4279 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4280 return true;
4281 }
4282
4283 return false;
4284}
4285
4286static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4287{
4288 bool success;
4289
4290 /* swapping inhibited */
4291 if (!(sc->gfp_mask & __GFP_IO) &&
4292 (folio_test_dirty(folio) ||
4293 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4294 return false;
4295
4296 /* raced with release_pages() */
4297 if (!folio_try_get(folio))
4298 return false;
4299
4300 /* raced with another isolation */
4301 if (!folio_test_clear_lru(folio)) {
4302 folio_put(folio);
4303 return false;
4304 }
4305
4306 /* see the comment on MAX_NR_TIERS */
4307 if (!folio_test_referenced(folio))
4308 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4309
4310 /* for shrink_folio_list() */
4311 folio_clear_reclaim(folio);
4312 folio_clear_referenced(folio);
4313
4314 success = lru_gen_del_folio(lruvec, folio, true);
4315 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4316
4317 return true;
4318}
4319
4320static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4321 int type, int tier, struct list_head *list)
4322{
4323 int i;
4324 int gen;
4325 enum vm_event_item item;
4326 int sorted = 0;
4327 int scanned = 0;
4328 int isolated = 0;
4329 int skipped = 0;
4330 int remaining = MAX_LRU_BATCH;
4331 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4332 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4333
4334 VM_WARN_ON_ONCE(!list_empty(list));
4335
4336 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4337 return 0;
4338
4339 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4340
4341 for (i = MAX_NR_ZONES; i > 0; i--) {
4342 LIST_HEAD(moved);
4343 int skipped_zone = 0;
4344 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4345 struct list_head *head = &lrugen->folios[gen][type][zone];
4346
4347 while (!list_empty(head)) {
4348 struct folio *folio = lru_to_folio(head);
4349 int delta = folio_nr_pages(folio);
4350
4351 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4352 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4353 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4354 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4355
4356 scanned += delta;
4357
4358 if (sort_folio(lruvec, folio, sc, tier))
4359 sorted += delta;
4360 else if (isolate_folio(lruvec, folio, sc)) {
4361 list_add(&folio->lru, list);
4362 isolated += delta;
4363 } else {
4364 list_move(&folio->lru, &moved);
4365 skipped_zone += delta;
4366 }
4367
4368 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4369 break;
4370 }
4371
4372 if (skipped_zone) {
4373 list_splice(&moved, head);
4374 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4375 skipped += skipped_zone;
4376 }
4377
4378 if (!remaining || isolated >= MIN_LRU_BATCH)
4379 break;
4380 }
4381
4382 item = PGSCAN_KSWAPD + reclaimer_offset();
4383 if (!cgroup_reclaim(sc)) {
4384 __count_vm_events(item, isolated);
4385 __count_vm_events(PGREFILL, sorted);
4386 }
4387 __count_memcg_events(memcg, item, isolated);
4388 __count_memcg_events(memcg, PGREFILL, sorted);
4389 __count_vm_events(PGSCAN_ANON + type, isolated);
4390 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4391 scanned, skipped, isolated,
4392 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4393
4394 /*
4395 * There might not be eligible folios due to reclaim_idx. Check the
4396 * remaining to prevent livelock if it's not making progress.
4397 */
4398 return isolated || !remaining ? scanned : 0;
4399}
4400
4401static int get_tier_idx(struct lruvec *lruvec, int type)
4402{
4403 int tier;
4404 struct ctrl_pos sp, pv;
4405
4406 /*
4407 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4408 * This value is chosen because any other tier would have at least twice
4409 * as many refaults as the first tier.
4410 */
4411 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4412 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4413 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4414 if (!positive_ctrl_err(&sp, &pv))
4415 break;
4416 }
4417
4418 return tier - 1;
4419}
4420
4421static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4422{
4423 int type, tier;
4424 struct ctrl_pos sp, pv;
4425 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4426
4427 /*
4428 * Compare the first tier of anon with that of file to determine which
4429 * type to scan. Also need to compare other tiers of the selected type
4430 * with the first tier of the other type to determine the last tier (of
4431 * the selected type) to evict.
4432 */
4433 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4434 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4435 type = positive_ctrl_err(&sp, &pv);
4436
4437 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4438 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4439 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4440 if (!positive_ctrl_err(&sp, &pv))
4441 break;
4442 }
4443
4444 *tier_idx = tier - 1;
4445
4446 return type;
4447}
4448
4449static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4450 int *type_scanned, struct list_head *list)
4451{
4452 int i;
4453 int type;
4454 int scanned;
4455 int tier = -1;
4456 DEFINE_MIN_SEQ(lruvec);
4457
4458 /*
4459 * Try to make the obvious choice first. When anon and file are both
4460 * available from the same generation, interpret swappiness 1 as file
4461 * first and 200 as anon first.
4462 */
4463 if (!swappiness)
4464 type = LRU_GEN_FILE;
4465 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4466 type = LRU_GEN_ANON;
4467 else if (swappiness == 1)
4468 type = LRU_GEN_FILE;
4469 else if (swappiness == 200)
4470 type = LRU_GEN_ANON;
4471 else
4472 type = get_type_to_scan(lruvec, swappiness, &tier);
4473
4474 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4475 if (tier < 0)
4476 tier = get_tier_idx(lruvec, type);
4477
4478 scanned = scan_folios(lruvec, sc, type, tier, list);
4479 if (scanned)
4480 break;
4481
4482 type = !type;
4483 tier = -1;
4484 }
4485
4486 *type_scanned = type;
4487
4488 return scanned;
4489}
4490
4491static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4492{
4493 int type;
4494 int scanned;
4495 int reclaimed;
4496 LIST_HEAD(list);
4497 LIST_HEAD(clean);
4498 struct folio *folio;
4499 struct folio *next;
4500 enum vm_event_item item;
4501 struct reclaim_stat stat;
4502 struct lru_gen_mm_walk *walk;
4503 bool skip_retry = false;
4504 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4505 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4506
4507 spin_lock_irq(&lruvec->lru_lock);
4508
4509 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4510
4511 scanned += try_to_inc_min_seq(lruvec, swappiness);
4512
4513 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4514 scanned = 0;
4515
4516 spin_unlock_irq(&lruvec->lru_lock);
4517
4518 if (list_empty(&list))
4519 return scanned;
4520retry:
4521 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4522 sc->nr_reclaimed += reclaimed;
4523 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4524 scanned, reclaimed, &stat, sc->priority,
4525 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4526
4527 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4528 if (!folio_evictable(folio)) {
4529 list_del(&folio->lru);
4530 folio_putback_lru(folio);
4531 continue;
4532 }
4533
4534 if (folio_test_reclaim(folio) &&
4535 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4536 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4537 if (folio_test_workingset(folio))
4538 folio_set_referenced(folio);
4539 continue;
4540 }
4541
4542 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4543 folio_mapped(folio) || folio_test_locked(folio) ||
4544 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4545 /* don't add rejected folios to the oldest generation */
4546 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4547 BIT(PG_active));
4548 continue;
4549 }
4550
4551 /* retry folios that may have missed folio_rotate_reclaimable() */
4552 list_move(&folio->lru, &clean);
4553 sc->nr_scanned -= folio_nr_pages(folio);
4554 }
4555
4556 spin_lock_irq(&lruvec->lru_lock);
4557
4558 move_folios_to_lru(lruvec, &list);
4559
4560 walk = current->reclaim_state->mm_walk;
4561 if (walk && walk->batched)
4562 reset_batch_size(lruvec, walk);
4563
4564 item = PGSTEAL_KSWAPD + reclaimer_offset();
4565 if (!cgroup_reclaim(sc))
4566 __count_vm_events(item, reclaimed);
4567 __count_memcg_events(memcg, item, reclaimed);
4568 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4569
4570 spin_unlock_irq(&lruvec->lru_lock);
4571
4572 mem_cgroup_uncharge_list(&list);
4573 free_unref_page_list(&list);
4574
4575 INIT_LIST_HEAD(&list);
4576 list_splice_init(&clean, &list);
4577
4578 if (!list_empty(&list)) {
4579 skip_retry = true;
4580 goto retry;
4581 }
4582
4583 return scanned;
4584}
4585
4586static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4587 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4588{
4589 int gen, type, zone;
4590 unsigned long old = 0;
4591 unsigned long young = 0;
4592 unsigned long total = 0;
4593 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4594 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4595 DEFINE_MIN_SEQ(lruvec);
4596
4597 /* whether this lruvec is completely out of cold folios */
4598 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4599 *nr_to_scan = 0;
4600 return true;
4601 }
4602
4603 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4604 unsigned long seq;
4605
4606 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4607 unsigned long size = 0;
4608
4609 gen = lru_gen_from_seq(seq);
4610
4611 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4612 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4613
4614 total += size;
4615 if (seq == max_seq)
4616 young += size;
4617 else if (seq + MIN_NR_GENS == max_seq)
4618 old += size;
4619 }
4620 }
4621
4622 /* try to scrape all its memory if this memcg was deleted */
4623 if (!mem_cgroup_online(memcg)) {
4624 *nr_to_scan = total;
4625 return false;
4626 }
4627
4628 *nr_to_scan = total >> sc->priority;
4629
4630 /*
4631 * The aging tries to be lazy to reduce the overhead, while the eviction
4632 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4633 * ideal number of generations is MIN_NR_GENS+1.
4634 */
4635 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4636 return false;
4637
4638 /*
4639 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4640 * of the total number of pages for each generation. A reasonable range
4641 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4642 * aging cares about the upper bound of hot pages, while the eviction
4643 * cares about the lower bound of cold pages.
4644 */
4645 if (young * MIN_NR_GENS > total)
4646 return true;
4647 if (old * (MIN_NR_GENS + 2) < total)
4648 return true;
4649
4650 return false;
4651}
4652
4653/*
4654 * For future optimizations:
4655 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4656 * reclaim.
4657 */
4658static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4659{
4660 unsigned long nr_to_scan;
4661 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4662 DEFINE_MAX_SEQ(lruvec);
4663
4664 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4665 return -1;
4666
4667 if (!should_run_aging(lruvec, max_seq, sc, can_swap, &nr_to_scan))
4668 return nr_to_scan;
4669
4670 /* skip the aging path at the default priority */
4671 if (sc->priority == DEF_PRIORITY)
4672 return nr_to_scan;
4673
4674 /* skip this lruvec as it's low on cold folios */
4675 return try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false) ? -1 : 0;
4676}
4677
4678static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4679{
4680 int i;
4681 enum zone_watermarks mark;
4682
4683 /* don't abort memcg reclaim to ensure fairness */
4684 if (!root_reclaim(sc))
4685 return false;
4686
4687 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4688 return true;
4689
4690 /* check the order to exclude compaction-induced reclaim */
4691 if (!current_is_kswapd() || sc->order)
4692 return false;
4693
4694 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4695 WMARK_PROMO : WMARK_HIGH;
4696
4697 for (i = 0; i <= sc->reclaim_idx; i++) {
4698 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4699 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4700
4701 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4702 return false;
4703 }
4704
4705 /* kswapd should abort if all eligible zones are safe */
4706 return true;
4707}
4708
4709static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4710{
4711 long nr_to_scan;
4712 unsigned long scanned = 0;
4713 int swappiness = get_swappiness(lruvec, sc);
4714
4715 /* clean file folios are more likely to exist */
4716 if (swappiness && !(sc->gfp_mask & __GFP_IO))
4717 swappiness = 1;
4718
4719 while (true) {
4720 int delta;
4721
4722 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4723 if (nr_to_scan <= 0)
4724 break;
4725
4726 delta = evict_folios(lruvec, sc, swappiness);
4727 if (!delta)
4728 break;
4729
4730 scanned += delta;
4731 if (scanned >= nr_to_scan)
4732 break;
4733
4734 if (should_abort_scan(lruvec, sc))
4735 break;
4736
4737 cond_resched();
4738 }
4739
4740 /* whether this lruvec should be rotated */
4741 return nr_to_scan < 0;
4742}
4743
4744static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4745{
4746 bool success;
4747 unsigned long scanned = sc->nr_scanned;
4748 unsigned long reclaimed = sc->nr_reclaimed;
4749 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4750 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4751
4752 mem_cgroup_calculate_protection(NULL, memcg);
4753
4754 if (mem_cgroup_below_min(NULL, memcg))
4755 return MEMCG_LRU_YOUNG;
4756
4757 if (mem_cgroup_below_low(NULL, memcg)) {
4758 /* see the comment on MEMCG_NR_GENS */
4759 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4760 return MEMCG_LRU_TAIL;
4761
4762 memcg_memory_event(memcg, MEMCG_LOW);
4763 }
4764
4765 success = try_to_shrink_lruvec(lruvec, sc);
4766
4767 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4768
4769 if (!sc->proactive)
4770 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4771 sc->nr_reclaimed - reclaimed);
4772
4773 flush_reclaim_state(sc);
4774
4775 if (success && mem_cgroup_online(memcg))
4776 return MEMCG_LRU_YOUNG;
4777
4778 if (!success && lruvec_is_sizable(lruvec, sc))
4779 return 0;
4780
4781 /* one retry if offlined or too small */
4782 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4783 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4784}
4785
4786static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4787{
4788 int op;
4789 int gen;
4790 int bin;
4791 int first_bin;
4792 struct lruvec *lruvec;
4793 struct lru_gen_folio *lrugen;
4794 struct mem_cgroup *memcg;
4795 struct hlist_nulls_node *pos;
4796
4797 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4798 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4799restart:
4800 op = 0;
4801 memcg = NULL;
4802
4803 rcu_read_lock();
4804
4805 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4806 if (op) {
4807 lru_gen_rotate_memcg(lruvec, op);
4808 op = 0;
4809 }
4810
4811 mem_cgroup_put(memcg);
4812 memcg = NULL;
4813
4814 if (gen != READ_ONCE(lrugen->gen))
4815 continue;
4816
4817 lruvec = container_of(lrugen, struct lruvec, lrugen);
4818 memcg = lruvec_memcg(lruvec);
4819
4820 if (!mem_cgroup_tryget(memcg)) {
4821 lru_gen_release_memcg(memcg);
4822 memcg = NULL;
4823 continue;
4824 }
4825
4826 rcu_read_unlock();
4827
4828 op = shrink_one(lruvec, sc);
4829
4830 rcu_read_lock();
4831
4832 if (should_abort_scan(lruvec, sc))
4833 break;
4834 }
4835
4836 rcu_read_unlock();
4837
4838 if (op)
4839 lru_gen_rotate_memcg(lruvec, op);
4840
4841 mem_cgroup_put(memcg);
4842
4843 if (!is_a_nulls(pos))
4844 return;
4845
4846 /* restart if raced with lru_gen_rotate_memcg() */
4847 if (gen != get_nulls_value(pos))
4848 goto restart;
4849
4850 /* try the rest of the bins of the current generation */
4851 bin = get_memcg_bin(bin + 1);
4852 if (bin != first_bin)
4853 goto restart;
4854}
4855
4856static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4857{
4858 struct blk_plug plug;
4859
4860 VM_WARN_ON_ONCE(root_reclaim(sc));
4861 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4862
4863 lru_add_drain();
4864
4865 blk_start_plug(&plug);
4866
4867 set_mm_walk(NULL, sc->proactive);
4868
4869 if (try_to_shrink_lruvec(lruvec, sc))
4870 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4871
4872 clear_mm_walk();
4873
4874 blk_finish_plug(&plug);
4875}
4876
4877static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4878{
4879 int priority;
4880 unsigned long reclaimable;
4881 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
4882
4883 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4884 return;
4885 /*
4886 * Determine the initial priority based on
4887 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4888 * where reclaimed_to_scanned_ratio = inactive / total.
4889 */
4890 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4891 if (get_swappiness(lruvec, sc))
4892 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4893
4894 /* round down reclaimable and round up sc->nr_to_reclaim */
4895 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4896
4897 sc->priority = clamp(priority, 0, DEF_PRIORITY);
4898}
4899
4900static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4901{
4902 struct blk_plug plug;
4903 unsigned long reclaimed = sc->nr_reclaimed;
4904
4905 VM_WARN_ON_ONCE(!root_reclaim(sc));
4906
4907 /*
4908 * Unmapped clean folios are already prioritized. Scanning for more of
4909 * them is likely futile and can cause high reclaim latency when there
4910 * is a large number of memcgs.
4911 */
4912 if (!sc->may_writepage || !sc->may_unmap)
4913 goto done;
4914
4915 lru_add_drain();
4916
4917 blk_start_plug(&plug);
4918
4919 set_mm_walk(pgdat, sc->proactive);
4920
4921 set_initial_priority(pgdat, sc);
4922
4923 if (current_is_kswapd())
4924 sc->nr_reclaimed = 0;
4925
4926 if (mem_cgroup_disabled())
4927 shrink_one(&pgdat->__lruvec, sc);
4928 else
4929 shrink_many(pgdat, sc);
4930
4931 if (current_is_kswapd())
4932 sc->nr_reclaimed += reclaimed;
4933
4934 clear_mm_walk();
4935
4936 blk_finish_plug(&plug);
4937done:
4938 /* kswapd should never fail */
4939 pgdat->kswapd_failures = 0;
4940}
4941
4942/******************************************************************************
4943 * state change
4944 ******************************************************************************/
4945
4946static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4947{
4948 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4949
4950 if (lrugen->enabled) {
4951 enum lru_list lru;
4952
4953 for_each_evictable_lru(lru) {
4954 if (!list_empty(&lruvec->lists[lru]))
4955 return false;
4956 }
4957 } else {
4958 int gen, type, zone;
4959
4960 for_each_gen_type_zone(gen, type, zone) {
4961 if (!list_empty(&lrugen->folios[gen][type][zone]))
4962 return false;
4963 }
4964 }
4965
4966 return true;
4967}
4968
4969static bool fill_evictable(struct lruvec *lruvec)
4970{
4971 enum lru_list lru;
4972 int remaining = MAX_LRU_BATCH;
4973
4974 for_each_evictable_lru(lru) {
4975 int type = is_file_lru(lru);
4976 bool active = is_active_lru(lru);
4977 struct list_head *head = &lruvec->lists[lru];
4978
4979 while (!list_empty(head)) {
4980 bool success;
4981 struct folio *folio = lru_to_folio(head);
4982
4983 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4984 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
4985 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4986 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
4987
4988 lruvec_del_folio(lruvec, folio);
4989 success = lru_gen_add_folio(lruvec, folio, false);
4990 VM_WARN_ON_ONCE(!success);
4991
4992 if (!--remaining)
4993 return false;
4994 }
4995 }
4996
4997 return true;
4998}
4999
5000static bool drain_evictable(struct lruvec *lruvec)
5001{
5002 int gen, type, zone;
5003 int remaining = MAX_LRU_BATCH;
5004
5005 for_each_gen_type_zone(gen, type, zone) {
5006 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5007
5008 while (!list_empty(head)) {
5009 bool success;
5010 struct folio *folio = lru_to_folio(head);
5011
5012 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5013 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5014 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5015 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5016
5017 success = lru_gen_del_folio(lruvec, folio, false);
5018 VM_WARN_ON_ONCE(!success);
5019 lruvec_add_folio(lruvec, folio);
5020
5021 if (!--remaining)
5022 return false;
5023 }
5024 }
5025
5026 return true;
5027}
5028
5029static void lru_gen_change_state(bool enabled)
5030{
5031 static DEFINE_MUTEX(state_mutex);
5032
5033 struct mem_cgroup *memcg;
5034
5035 cgroup_lock();
5036 cpus_read_lock();
5037 get_online_mems();
5038 mutex_lock(&state_mutex);
5039
5040 if (enabled == lru_gen_enabled())
5041 goto unlock;
5042
5043 if (enabled)
5044 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5045 else
5046 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5047
5048 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5049 do {
5050 int nid;
5051
5052 for_each_node(nid) {
5053 struct lruvec *lruvec = get_lruvec(memcg, nid);
5054
5055 spin_lock_irq(&lruvec->lru_lock);
5056
5057 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5058 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5059
5060 lruvec->lrugen.enabled = enabled;
5061
5062 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5063 spin_unlock_irq(&lruvec->lru_lock);
5064 cond_resched();
5065 spin_lock_irq(&lruvec->lru_lock);
5066 }
5067
5068 spin_unlock_irq(&lruvec->lru_lock);
5069 }
5070
5071 cond_resched();
5072 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5073unlock:
5074 mutex_unlock(&state_mutex);
5075 put_online_mems();
5076 cpus_read_unlock();
5077 cgroup_unlock();
5078}
5079
5080/******************************************************************************
5081 * sysfs interface
5082 ******************************************************************************/
5083
5084static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5085{
5086 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5087}
5088
5089/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5090static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5091 const char *buf, size_t len)
5092{
5093 unsigned int msecs;
5094
5095 if (kstrtouint(buf, 0, &msecs))
5096 return -EINVAL;
5097
5098 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5099
5100 return len;
5101}
5102
5103static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5104
5105static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5106{
5107 unsigned int caps = 0;
5108
5109 if (get_cap(LRU_GEN_CORE))
5110 caps |= BIT(LRU_GEN_CORE);
5111
5112 if (should_walk_mmu())
5113 caps |= BIT(LRU_GEN_MM_WALK);
5114
5115 if (should_clear_pmd_young())
5116 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5117
5118 return sysfs_emit(buf, "0x%04x\n", caps);
5119}
5120
5121/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5122static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5123 const char *buf, size_t len)
5124{
5125 int i;
5126 unsigned int caps;
5127
5128 if (tolower(*buf) == 'n')
5129 caps = 0;
5130 else if (tolower(*buf) == 'y')
5131 caps = -1;
5132 else if (kstrtouint(buf, 0, &caps))
5133 return -EINVAL;
5134
5135 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5136 bool enabled = caps & BIT(i);
5137
5138 if (i == LRU_GEN_CORE)
5139 lru_gen_change_state(enabled);
5140 else if (enabled)
5141 static_branch_enable(&lru_gen_caps[i]);
5142 else
5143 static_branch_disable(&lru_gen_caps[i]);
5144 }
5145
5146 return len;
5147}
5148
5149static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5150
5151static struct attribute *lru_gen_attrs[] = {
5152 &lru_gen_min_ttl_attr.attr,
5153 &lru_gen_enabled_attr.attr,
5154 NULL
5155};
5156
5157static const struct attribute_group lru_gen_attr_group = {
5158 .name = "lru_gen",
5159 .attrs = lru_gen_attrs,
5160};
5161
5162/******************************************************************************
5163 * debugfs interface
5164 ******************************************************************************/
5165
5166static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5167{
5168 struct mem_cgroup *memcg;
5169 loff_t nr_to_skip = *pos;
5170
5171 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5172 if (!m->private)
5173 return ERR_PTR(-ENOMEM);
5174
5175 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5176 do {
5177 int nid;
5178
5179 for_each_node_state(nid, N_MEMORY) {
5180 if (!nr_to_skip--)
5181 return get_lruvec(memcg, nid);
5182 }
5183 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5184
5185 return NULL;
5186}
5187
5188static void lru_gen_seq_stop(struct seq_file *m, void *v)
5189{
5190 if (!IS_ERR_OR_NULL(v))
5191 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5192
5193 kvfree(m->private);
5194 m->private = NULL;
5195}
5196
5197static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5198{
5199 int nid = lruvec_pgdat(v)->node_id;
5200 struct mem_cgroup *memcg = lruvec_memcg(v);
5201
5202 ++*pos;
5203
5204 nid = next_memory_node(nid);
5205 if (nid == MAX_NUMNODES) {
5206 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5207 if (!memcg)
5208 return NULL;
5209
5210 nid = first_memory_node;
5211 }
5212
5213 return get_lruvec(memcg, nid);
5214}
5215
5216static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5217 unsigned long max_seq, unsigned long *min_seq,
5218 unsigned long seq)
5219{
5220 int i;
5221 int type, tier;
5222 int hist = lru_hist_from_seq(seq);
5223 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5224 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5225
5226 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5227 seq_printf(m, " %10d", tier);
5228 for (type = 0; type < ANON_AND_FILE; type++) {
5229 const char *s = " ";
5230 unsigned long n[3] = {};
5231
5232 if (seq == max_seq) {
5233 s = "RT ";
5234 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5235 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5236 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5237 s = "rep";
5238 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5239 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5240 if (tier)
5241 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5242 }
5243
5244 for (i = 0; i < 3; i++)
5245 seq_printf(m, " %10lu%c", n[i], s[i]);
5246 }
5247 seq_putc(m, '\n');
5248 }
5249
5250 if (!mm_state)
5251 return;
5252
5253 seq_puts(m, " ");
5254 for (i = 0; i < NR_MM_STATS; i++) {
5255 const char *s = " ";
5256 unsigned long n = 0;
5257
5258 if (seq == max_seq && NR_HIST_GENS == 1) {
5259 s = "LOYNFA";
5260 n = READ_ONCE(mm_state->stats[hist][i]);
5261 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5262 s = "loynfa";
5263 n = READ_ONCE(mm_state->stats[hist][i]);
5264 }
5265
5266 seq_printf(m, " %10lu%c", n, s[i]);
5267 }
5268 seq_putc(m, '\n');
5269}
5270
5271/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5272static int lru_gen_seq_show(struct seq_file *m, void *v)
5273{
5274 unsigned long seq;
5275 bool full = !debugfs_real_fops(m->file)->write;
5276 struct lruvec *lruvec = v;
5277 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5278 int nid = lruvec_pgdat(lruvec)->node_id;
5279 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5280 DEFINE_MAX_SEQ(lruvec);
5281 DEFINE_MIN_SEQ(lruvec);
5282
5283 if (nid == first_memory_node) {
5284 const char *path = memcg ? m->private : "";
5285
5286#ifdef CONFIG_MEMCG
5287 if (memcg)
5288 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5289#endif
5290 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5291 }
5292
5293 seq_printf(m, " node %5d\n", nid);
5294
5295 if (!full)
5296 seq = min_seq[LRU_GEN_ANON];
5297 else if (max_seq >= MAX_NR_GENS)
5298 seq = max_seq - MAX_NR_GENS + 1;
5299 else
5300 seq = 0;
5301
5302 for (; seq <= max_seq; seq++) {
5303 int type, zone;
5304 int gen = lru_gen_from_seq(seq);
5305 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5306
5307 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5308
5309 for (type = 0; type < ANON_AND_FILE; type++) {
5310 unsigned long size = 0;
5311 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5312
5313 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5314 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5315
5316 seq_printf(m, " %10lu%c", size, mark);
5317 }
5318
5319 seq_putc(m, '\n');
5320
5321 if (full)
5322 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5323 }
5324
5325 return 0;
5326}
5327
5328static const struct seq_operations lru_gen_seq_ops = {
5329 .start = lru_gen_seq_start,
5330 .stop = lru_gen_seq_stop,
5331 .next = lru_gen_seq_next,
5332 .show = lru_gen_seq_show,
5333};
5334
5335static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5336 bool can_swap, bool force_scan)
5337{
5338 DEFINE_MAX_SEQ(lruvec);
5339 DEFINE_MIN_SEQ(lruvec);
5340
5341 if (seq < max_seq)
5342 return 0;
5343
5344 if (seq > max_seq)
5345 return -EINVAL;
5346
5347 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5348 return -ERANGE;
5349
5350 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5351
5352 return 0;
5353}
5354
5355static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5356 int swappiness, unsigned long nr_to_reclaim)
5357{
5358 DEFINE_MAX_SEQ(lruvec);
5359
5360 if (seq + MIN_NR_GENS > max_seq)
5361 return -EINVAL;
5362
5363 sc->nr_reclaimed = 0;
5364
5365 while (!signal_pending(current)) {
5366 DEFINE_MIN_SEQ(lruvec);
5367
5368 if (seq < min_seq[!swappiness])
5369 return 0;
5370
5371 if (sc->nr_reclaimed >= nr_to_reclaim)
5372 return 0;
5373
5374 if (!evict_folios(lruvec, sc, swappiness))
5375 return 0;
5376
5377 cond_resched();
5378 }
5379
5380 return -EINTR;
5381}
5382
5383static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5384 struct scan_control *sc, int swappiness, unsigned long opt)
5385{
5386 struct lruvec *lruvec;
5387 int err = -EINVAL;
5388 struct mem_cgroup *memcg = NULL;
5389
5390 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5391 return -EINVAL;
5392
5393 if (!mem_cgroup_disabled()) {
5394 rcu_read_lock();
5395
5396 memcg = mem_cgroup_from_id(memcg_id);
5397 if (!mem_cgroup_tryget(memcg))
5398 memcg = NULL;
5399
5400 rcu_read_unlock();
5401
5402 if (!memcg)
5403 return -EINVAL;
5404 }
5405
5406 if (memcg_id != mem_cgroup_id(memcg))
5407 goto done;
5408
5409 lruvec = get_lruvec(memcg, nid);
5410
5411 if (swappiness < 0)
5412 swappiness = get_swappiness(lruvec, sc);
5413 else if (swappiness > 200)
5414 goto done;
5415
5416 switch (cmd) {
5417 case '+':
5418 err = run_aging(lruvec, seq, sc, swappiness, opt);
5419 break;
5420 case '-':
5421 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5422 break;
5423 }
5424done:
5425 mem_cgroup_put(memcg);
5426
5427 return err;
5428}
5429
5430/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5431static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5432 size_t len, loff_t *pos)
5433{
5434 void *buf;
5435 char *cur, *next;
5436 unsigned int flags;
5437 struct blk_plug plug;
5438 int err = -EINVAL;
5439 struct scan_control sc = {
5440 .may_writepage = true,
5441 .may_unmap = true,
5442 .may_swap = true,
5443 .reclaim_idx = MAX_NR_ZONES - 1,
5444 .gfp_mask = GFP_KERNEL,
5445 };
5446
5447 buf = kvmalloc(len + 1, GFP_KERNEL);
5448 if (!buf)
5449 return -ENOMEM;
5450
5451 if (copy_from_user(buf, src, len)) {
5452 kvfree(buf);
5453 return -EFAULT;
5454 }
5455
5456 set_task_reclaim_state(current, &sc.reclaim_state);
5457 flags = memalloc_noreclaim_save();
5458 blk_start_plug(&plug);
5459 if (!set_mm_walk(NULL, true)) {
5460 err = -ENOMEM;
5461 goto done;
5462 }
5463
5464 next = buf;
5465 next[len] = '\0';
5466
5467 while ((cur = strsep(&next, ",;\n"))) {
5468 int n;
5469 int end;
5470 char cmd;
5471 unsigned int memcg_id;
5472 unsigned int nid;
5473 unsigned long seq;
5474 unsigned int swappiness = -1;
5475 unsigned long opt = -1;
5476
5477 cur = skip_spaces(cur);
5478 if (!*cur)
5479 continue;
5480
5481 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5482 &seq, &end, &swappiness, &end, &opt, &end);
5483 if (n < 4 || cur[end]) {
5484 err = -EINVAL;
5485 break;
5486 }
5487
5488 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5489 if (err)
5490 break;
5491 }
5492done:
5493 clear_mm_walk();
5494 blk_finish_plug(&plug);
5495 memalloc_noreclaim_restore(flags);
5496 set_task_reclaim_state(current, NULL);
5497
5498 kvfree(buf);
5499
5500 return err ? : len;
5501}
5502
5503static int lru_gen_seq_open(struct inode *inode, struct file *file)
5504{
5505 return seq_open(file, &lru_gen_seq_ops);
5506}
5507
5508static const struct file_operations lru_gen_rw_fops = {
5509 .open = lru_gen_seq_open,
5510 .read = seq_read,
5511 .write = lru_gen_seq_write,
5512 .llseek = seq_lseek,
5513 .release = seq_release,
5514};
5515
5516static const struct file_operations lru_gen_ro_fops = {
5517 .open = lru_gen_seq_open,
5518 .read = seq_read,
5519 .llseek = seq_lseek,
5520 .release = seq_release,
5521};
5522
5523/******************************************************************************
5524 * initialization
5525 ******************************************************************************/
5526
5527void lru_gen_init_pgdat(struct pglist_data *pgdat)
5528{
5529 int i, j;
5530
5531 spin_lock_init(&pgdat->memcg_lru.lock);
5532
5533 for (i = 0; i < MEMCG_NR_GENS; i++) {
5534 for (j = 0; j < MEMCG_NR_BINS; j++)
5535 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5536 }
5537}
5538
5539void lru_gen_init_lruvec(struct lruvec *lruvec)
5540{
5541 int i;
5542 int gen, type, zone;
5543 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5544 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5545
5546 lrugen->max_seq = MIN_NR_GENS + 1;
5547 lrugen->enabled = lru_gen_enabled();
5548
5549 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5550 lrugen->timestamps[i] = jiffies;
5551
5552 for_each_gen_type_zone(gen, type, zone)
5553 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5554
5555 if (mm_state)
5556 mm_state->seq = MIN_NR_GENS;
5557}
5558
5559#ifdef CONFIG_MEMCG
5560
5561void lru_gen_init_memcg(struct mem_cgroup *memcg)
5562{
5563 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5564
5565 if (!mm_list)
5566 return;
5567
5568 INIT_LIST_HEAD(&mm_list->fifo);
5569 spin_lock_init(&mm_list->lock);
5570}
5571
5572void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5573{
5574 int i;
5575 int nid;
5576 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5577
5578 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5579
5580 for_each_node(nid) {
5581 struct lruvec *lruvec = get_lruvec(memcg, nid);
5582 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5583
5584 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5585 sizeof(lruvec->lrugen.nr_pages)));
5586
5587 lruvec->lrugen.list.next = LIST_POISON1;
5588
5589 if (!mm_state)
5590 continue;
5591
5592 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5593 bitmap_free(mm_state->filters[i]);
5594 mm_state->filters[i] = NULL;
5595 }
5596 }
5597}
5598
5599#endif /* CONFIG_MEMCG */
5600
5601static int __init init_lru_gen(void)
5602{
5603 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5604 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5605
5606 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5607 pr_err("lru_gen: failed to create sysfs group\n");
5608
5609 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5610 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5611
5612 return 0;
5613};
5614late_initcall(init_lru_gen);
5615
5616#else /* !CONFIG_LRU_GEN */
5617
5618static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5619{
5620 BUILD_BUG();
5621}
5622
5623static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5624{
5625 BUILD_BUG();
5626}
5627
5628static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5629{
5630 BUILD_BUG();
5631}
5632
5633#endif /* CONFIG_LRU_GEN */
5634
5635static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5636{
5637 unsigned long nr[NR_LRU_LISTS];
5638 unsigned long targets[NR_LRU_LISTS];
5639 unsigned long nr_to_scan;
5640 enum lru_list lru;
5641 unsigned long nr_reclaimed = 0;
5642 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5643 bool proportional_reclaim;
5644 struct blk_plug plug;
5645
5646 if (lru_gen_enabled() && !root_reclaim(sc)) {
5647 lru_gen_shrink_lruvec(lruvec, sc);
5648 return;
5649 }
5650
5651 get_scan_count(lruvec, sc, nr);
5652
5653 /* Record the original scan target for proportional adjustments later */
5654 memcpy(targets, nr, sizeof(nr));
5655
5656 /*
5657 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5658 * event that can occur when there is little memory pressure e.g.
5659 * multiple streaming readers/writers. Hence, we do not abort scanning
5660 * when the requested number of pages are reclaimed when scanning at
5661 * DEF_PRIORITY on the assumption that the fact we are direct
5662 * reclaiming implies that kswapd is not keeping up and it is best to
5663 * do a batch of work at once. For memcg reclaim one check is made to
5664 * abort proportional reclaim if either the file or anon lru has already
5665 * dropped to zero at the first pass.
5666 */
5667 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5668 sc->priority == DEF_PRIORITY);
5669
5670 blk_start_plug(&plug);
5671 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5672 nr[LRU_INACTIVE_FILE]) {
5673 unsigned long nr_anon, nr_file, percentage;
5674 unsigned long nr_scanned;
5675
5676 for_each_evictable_lru(lru) {
5677 if (nr[lru]) {
5678 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5679 nr[lru] -= nr_to_scan;
5680
5681 nr_reclaimed += shrink_list(lru, nr_to_scan,
5682 lruvec, sc);
5683 }
5684 }
5685
5686 cond_resched();
5687
5688 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5689 continue;
5690
5691 /*
5692 * For kswapd and memcg, reclaim at least the number of pages
5693 * requested. Ensure that the anon and file LRUs are scanned
5694 * proportionally what was requested by get_scan_count(). We
5695 * stop reclaiming one LRU and reduce the amount scanning
5696 * proportional to the original scan target.
5697 */
5698 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5699 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5700
5701 /*
5702 * It's just vindictive to attack the larger once the smaller
5703 * has gone to zero. And given the way we stop scanning the
5704 * smaller below, this makes sure that we only make one nudge
5705 * towards proportionality once we've got nr_to_reclaim.
5706 */
5707 if (!nr_file || !nr_anon)
5708 break;
5709
5710 if (nr_file > nr_anon) {
5711 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5712 targets[LRU_ACTIVE_ANON] + 1;
5713 lru = LRU_BASE;
5714 percentage = nr_anon * 100 / scan_target;
5715 } else {
5716 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5717 targets[LRU_ACTIVE_FILE] + 1;
5718 lru = LRU_FILE;
5719 percentage = nr_file * 100 / scan_target;
5720 }
5721
5722 /* Stop scanning the smaller of the LRU */
5723 nr[lru] = 0;
5724 nr[lru + LRU_ACTIVE] = 0;
5725
5726 /*
5727 * Recalculate the other LRU scan count based on its original
5728 * scan target and the percentage scanning already complete
5729 */
5730 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5731 nr_scanned = targets[lru] - nr[lru];
5732 nr[lru] = targets[lru] * (100 - percentage) / 100;
5733 nr[lru] -= min(nr[lru], nr_scanned);
5734
5735 lru += LRU_ACTIVE;
5736 nr_scanned = targets[lru] - nr[lru];
5737 nr[lru] = targets[lru] * (100 - percentage) / 100;
5738 nr[lru] -= min(nr[lru], nr_scanned);
5739 }
5740 blk_finish_plug(&plug);
5741 sc->nr_reclaimed += nr_reclaimed;
5742
5743 /*
5744 * Even if we did not try to evict anon pages at all, we want to
5745 * rebalance the anon lru active/inactive ratio.
5746 */
5747 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5748 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5749 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5750 sc, LRU_ACTIVE_ANON);
5751}
5752
5753/* Use reclaim/compaction for costly allocs or under memory pressure */
5754static bool in_reclaim_compaction(struct scan_control *sc)
5755{
5756 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5757 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5758 sc->priority < DEF_PRIORITY - 2))
5759 return true;
5760
5761 return false;
5762}
5763
5764/*
5765 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5766 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5767 * true if more pages should be reclaimed such that when the page allocator
5768 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5769 * It will give up earlier than that if there is difficulty reclaiming pages.
5770 */
5771static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5772 unsigned long nr_reclaimed,
5773 struct scan_control *sc)
5774{
5775 unsigned long pages_for_compaction;
5776 unsigned long inactive_lru_pages;
5777 int z;
5778
5779 /* If not in reclaim/compaction mode, stop */
5780 if (!in_reclaim_compaction(sc))
5781 return false;
5782
5783 /*
5784 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5785 * number of pages that were scanned. This will return to the caller
5786 * with the risk reclaim/compaction and the resulting allocation attempt
5787 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5788 * allocations through requiring that the full LRU list has been scanned
5789 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5790 * scan, but that approximation was wrong, and there were corner cases
5791 * where always a non-zero amount of pages were scanned.
5792 */
5793 if (!nr_reclaimed)
5794 return false;
5795
5796 /* If compaction would go ahead or the allocation would succeed, stop */
5797 for (z = 0; z <= sc->reclaim_idx; z++) {
5798 struct zone *zone = &pgdat->node_zones[z];
5799 if (!managed_zone(zone))
5800 continue;
5801
5802 /* Allocation can already succeed, nothing to do */
5803 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5804 sc->reclaim_idx, 0))
5805 return false;
5806
5807 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5808 return false;
5809 }
5810
5811 /*
5812 * If we have not reclaimed enough pages for compaction and the
5813 * inactive lists are large enough, continue reclaiming
5814 */
5815 pages_for_compaction = compact_gap(sc->order);
5816 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5817 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5818 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5819
5820 return inactive_lru_pages > pages_for_compaction;
5821}
5822
5823static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5824{
5825 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5826 struct mem_cgroup *memcg;
5827
5828 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
5829 do {
5830 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5831 unsigned long reclaimed;
5832 unsigned long scanned;
5833
5834 /*
5835 * This loop can become CPU-bound when target memcgs
5836 * aren't eligible for reclaim - either because they
5837 * don't have any reclaimable pages, or because their
5838 * memory is explicitly protected. Avoid soft lockups.
5839 */
5840 cond_resched();
5841
5842 mem_cgroup_calculate_protection(target_memcg, memcg);
5843
5844 if (mem_cgroup_below_min(target_memcg, memcg)) {
5845 /*
5846 * Hard protection.
5847 * If there is no reclaimable memory, OOM.
5848 */
5849 continue;
5850 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5851 /*
5852 * Soft protection.
5853 * Respect the protection only as long as
5854 * there is an unprotected supply
5855 * of reclaimable memory from other cgroups.
5856 */
5857 if (!sc->memcg_low_reclaim) {
5858 sc->memcg_low_skipped = 1;
5859 continue;
5860 }
5861 memcg_memory_event(memcg, MEMCG_LOW);
5862 }
5863
5864 reclaimed = sc->nr_reclaimed;
5865 scanned = sc->nr_scanned;
5866
5867 shrink_lruvec(lruvec, sc);
5868
5869 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5870 sc->priority);
5871
5872 /* Record the group's reclaim efficiency */
5873 if (!sc->proactive)
5874 vmpressure(sc->gfp_mask, memcg, false,
5875 sc->nr_scanned - scanned,
5876 sc->nr_reclaimed - reclaimed);
5877
5878 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5879}
5880
5881static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5882{
5883 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5884 struct lruvec *target_lruvec;
5885 bool reclaimable = false;
5886
5887 if (lru_gen_enabled() && root_reclaim(sc)) {
5888 lru_gen_shrink_node(pgdat, sc);
5889 return;
5890 }
5891
5892 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5893
5894again:
5895 memset(&sc->nr, 0, sizeof(sc->nr));
5896
5897 nr_reclaimed = sc->nr_reclaimed;
5898 nr_scanned = sc->nr_scanned;
5899
5900 prepare_scan_control(pgdat, sc);
5901
5902 shrink_node_memcgs(pgdat, sc);
5903
5904 flush_reclaim_state(sc);
5905
5906 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5907
5908 /* Record the subtree's reclaim efficiency */
5909 if (!sc->proactive)
5910 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5911 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5912
5913 if (nr_node_reclaimed)
5914 reclaimable = true;
5915
5916 if (current_is_kswapd()) {
5917 /*
5918 * If reclaim is isolating dirty pages under writeback,
5919 * it implies that the long-lived page allocation rate
5920 * is exceeding the page laundering rate. Either the
5921 * global limits are not being effective at throttling
5922 * processes due to the page distribution throughout
5923 * zones or there is heavy usage of a slow backing
5924 * device. The only option is to throttle from reclaim
5925 * context which is not ideal as there is no guarantee
5926 * the dirtying process is throttled in the same way
5927 * balance_dirty_pages() manages.
5928 *
5929 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5930 * count the number of pages under pages flagged for
5931 * immediate reclaim and stall if any are encountered
5932 * in the nr_immediate check below.
5933 */
5934 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5935 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5936
5937 /* Allow kswapd to start writing pages during reclaim.*/
5938 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5939 set_bit(PGDAT_DIRTY, &pgdat->flags);
5940
5941 /*
5942 * If kswapd scans pages marked for immediate
5943 * reclaim and under writeback (nr_immediate), it
5944 * implies that pages are cycling through the LRU
5945 * faster than they are written so forcibly stall
5946 * until some pages complete writeback.
5947 */
5948 if (sc->nr.immediate)
5949 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5950 }
5951
5952 /*
5953 * Tag a node/memcg as congested if all the dirty pages were marked
5954 * for writeback and immediate reclaim (counted in nr.congested).
5955 *
5956 * Legacy memcg will stall in page writeback so avoid forcibly
5957 * stalling in reclaim_throttle().
5958 */
5959 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
5960 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
5961 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
5962
5963 if (current_is_kswapd())
5964 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
5965 }
5966
5967 /*
5968 * Stall direct reclaim for IO completions if the lruvec is
5969 * node is congested. Allow kswapd to continue until it
5970 * starts encountering unqueued dirty pages or cycling through
5971 * the LRU too quickly.
5972 */
5973 if (!current_is_kswapd() && current_may_throttle() &&
5974 !sc->hibernation_mode &&
5975 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
5976 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
5977 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
5978
5979 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
5980 goto again;
5981
5982 /*
5983 * Kswapd gives up on balancing particular nodes after too
5984 * many failures to reclaim anything from them and goes to
5985 * sleep. On reclaim progress, reset the failure counter. A
5986 * successful direct reclaim run will revive a dormant kswapd.
5987 */
5988 if (reclaimable)
5989 pgdat->kswapd_failures = 0;
5990}
5991
5992/*
5993 * Returns true if compaction should go ahead for a costly-order request, or
5994 * the allocation would already succeed without compaction. Return false if we
5995 * should reclaim first.
5996 */
5997static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
5998{
5999 unsigned long watermark;
6000
6001 if (!gfp_compaction_allowed(sc->gfp_mask))
6002 return false;
6003
6004 /* Allocation can already succeed, nothing to do */
6005 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6006 sc->reclaim_idx, 0))
6007 return true;
6008
6009 /* Compaction cannot yet proceed. Do reclaim. */
6010 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6011 return false;
6012
6013 /*
6014 * Compaction is already possible, but it takes time to run and there
6015 * are potentially other callers using the pages just freed. So proceed
6016 * with reclaim to make a buffer of free pages available to give
6017 * compaction a reasonable chance of completing and allocating the page.
6018 * Note that we won't actually reclaim the whole buffer in one attempt
6019 * as the target watermark in should_continue_reclaim() is lower. But if
6020 * we are already above the high+gap watermark, don't reclaim at all.
6021 */
6022 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6023
6024 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6025}
6026
6027static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6028{
6029 /*
6030 * If reclaim is making progress greater than 12% efficiency then
6031 * wake all the NOPROGRESS throttled tasks.
6032 */
6033 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6034 wait_queue_head_t *wqh;
6035
6036 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6037 if (waitqueue_active(wqh))
6038 wake_up(wqh);
6039
6040 return;
6041 }
6042
6043 /*
6044 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6045 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6046 * under writeback and marked for immediate reclaim at the tail of the
6047 * LRU.
6048 */
6049 if (current_is_kswapd() || cgroup_reclaim(sc))
6050 return;
6051
6052 /* Throttle if making no progress at high prioities. */
6053 if (sc->priority == 1 && !sc->nr_reclaimed)
6054 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6055}
6056
6057/*
6058 * This is the direct reclaim path, for page-allocating processes. We only
6059 * try to reclaim pages from zones which will satisfy the caller's allocation
6060 * request.
6061 *
6062 * If a zone is deemed to be full of pinned pages then just give it a light
6063 * scan then give up on it.
6064 */
6065static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6066{
6067 struct zoneref *z;
6068 struct zone *zone;
6069 unsigned long nr_soft_reclaimed;
6070 unsigned long nr_soft_scanned;
6071 gfp_t orig_mask;
6072 pg_data_t *last_pgdat = NULL;
6073 pg_data_t *first_pgdat = NULL;
6074
6075 /*
6076 * If the number of buffer_heads in the machine exceeds the maximum
6077 * allowed level, force direct reclaim to scan the highmem zone as
6078 * highmem pages could be pinning lowmem pages storing buffer_heads
6079 */
6080 orig_mask = sc->gfp_mask;
6081 if (buffer_heads_over_limit) {
6082 sc->gfp_mask |= __GFP_HIGHMEM;
6083 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6084 }
6085
6086 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6087 sc->reclaim_idx, sc->nodemask) {
6088 /*
6089 * Take care memory controller reclaiming has small influence
6090 * to global LRU.
6091 */
6092 if (!cgroup_reclaim(sc)) {
6093 if (!cpuset_zone_allowed(zone,
6094 GFP_KERNEL | __GFP_HARDWALL))
6095 continue;
6096
6097 /*
6098 * If we already have plenty of memory free for
6099 * compaction in this zone, don't free any more.
6100 * Even though compaction is invoked for any
6101 * non-zero order, only frequent costly order
6102 * reclamation is disruptive enough to become a
6103 * noticeable problem, like transparent huge
6104 * page allocations.
6105 */
6106 if (IS_ENABLED(CONFIG_COMPACTION) &&
6107 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6108 compaction_ready(zone, sc)) {
6109 sc->compaction_ready = true;
6110 continue;
6111 }
6112
6113 /*
6114 * Shrink each node in the zonelist once. If the
6115 * zonelist is ordered by zone (not the default) then a
6116 * node may be shrunk multiple times but in that case
6117 * the user prefers lower zones being preserved.
6118 */
6119 if (zone->zone_pgdat == last_pgdat)
6120 continue;
6121
6122 /*
6123 * This steals pages from memory cgroups over softlimit
6124 * and returns the number of reclaimed pages and
6125 * scanned pages. This works for global memory pressure
6126 * and balancing, not for a memcg's limit.
6127 */
6128 nr_soft_scanned = 0;
6129 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6130 sc->order, sc->gfp_mask,
6131 &nr_soft_scanned);
6132 sc->nr_reclaimed += nr_soft_reclaimed;
6133 sc->nr_scanned += nr_soft_scanned;
6134 /* need some check for avoid more shrink_zone() */
6135 }
6136
6137 if (!first_pgdat)
6138 first_pgdat = zone->zone_pgdat;
6139
6140 /* See comment about same check for global reclaim above */
6141 if (zone->zone_pgdat == last_pgdat)
6142 continue;
6143 last_pgdat = zone->zone_pgdat;
6144 shrink_node(zone->zone_pgdat, sc);
6145 }
6146
6147 if (first_pgdat)
6148 consider_reclaim_throttle(first_pgdat, sc);
6149
6150 /*
6151 * Restore to original mask to avoid the impact on the caller if we
6152 * promoted it to __GFP_HIGHMEM.
6153 */
6154 sc->gfp_mask = orig_mask;
6155}
6156
6157static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6158{
6159 struct lruvec *target_lruvec;
6160 unsigned long refaults;
6161
6162 if (lru_gen_enabled())
6163 return;
6164
6165 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6166 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6167 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6168 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6169 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6170}
6171
6172/*
6173 * This is the main entry point to direct page reclaim.
6174 *
6175 * If a full scan of the inactive list fails to free enough memory then we
6176 * are "out of memory" and something needs to be killed.
6177 *
6178 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6179 * high - the zone may be full of dirty or under-writeback pages, which this
6180 * caller can't do much about. We kick the writeback threads and take explicit
6181 * naps in the hope that some of these pages can be written. But if the
6182 * allocating task holds filesystem locks which prevent writeout this might not
6183 * work, and the allocation attempt will fail.
6184 *
6185 * returns: 0, if no pages reclaimed
6186 * else, the number of pages reclaimed
6187 */
6188static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6189 struct scan_control *sc)
6190{
6191 int initial_priority = sc->priority;
6192 pg_data_t *last_pgdat;
6193 struct zoneref *z;
6194 struct zone *zone;
6195retry:
6196 delayacct_freepages_start();
6197
6198 if (!cgroup_reclaim(sc))
6199 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6200
6201 do {
6202 if (!sc->proactive)
6203 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6204 sc->priority);
6205 sc->nr_scanned = 0;
6206 shrink_zones(zonelist, sc);
6207
6208 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6209 break;
6210
6211 if (sc->compaction_ready)
6212 break;
6213
6214 /*
6215 * If we're getting trouble reclaiming, start doing
6216 * writepage even in laptop mode.
6217 */
6218 if (sc->priority < DEF_PRIORITY - 2)
6219 sc->may_writepage = 1;
6220 } while (--sc->priority >= 0);
6221
6222 last_pgdat = NULL;
6223 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6224 sc->nodemask) {
6225 if (zone->zone_pgdat == last_pgdat)
6226 continue;
6227 last_pgdat = zone->zone_pgdat;
6228
6229 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6230
6231 if (cgroup_reclaim(sc)) {
6232 struct lruvec *lruvec;
6233
6234 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6235 zone->zone_pgdat);
6236 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6237 }
6238 }
6239
6240 delayacct_freepages_end();
6241
6242 if (sc->nr_reclaimed)
6243 return sc->nr_reclaimed;
6244
6245 /* Aborted reclaim to try compaction? don't OOM, then */
6246 if (sc->compaction_ready)
6247 return 1;
6248
6249 /*
6250 * We make inactive:active ratio decisions based on the node's
6251 * composition of memory, but a restrictive reclaim_idx or a
6252 * memory.low cgroup setting can exempt large amounts of
6253 * memory from reclaim. Neither of which are very common, so
6254 * instead of doing costly eligibility calculations of the
6255 * entire cgroup subtree up front, we assume the estimates are
6256 * good, and retry with forcible deactivation if that fails.
6257 */
6258 if (sc->skipped_deactivate) {
6259 sc->priority = initial_priority;
6260 sc->force_deactivate = 1;
6261 sc->skipped_deactivate = 0;
6262 goto retry;
6263 }
6264
6265 /* Untapped cgroup reserves? Don't OOM, retry. */
6266 if (sc->memcg_low_skipped) {
6267 sc->priority = initial_priority;
6268 sc->force_deactivate = 0;
6269 sc->memcg_low_reclaim = 1;
6270 sc->memcg_low_skipped = 0;
6271 goto retry;
6272 }
6273
6274 return 0;
6275}
6276
6277static bool allow_direct_reclaim(pg_data_t *pgdat)
6278{
6279 struct zone *zone;
6280 unsigned long pfmemalloc_reserve = 0;
6281 unsigned long free_pages = 0;
6282 int i;
6283 bool wmark_ok;
6284
6285 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6286 return true;
6287
6288 for (i = 0; i <= ZONE_NORMAL; i++) {
6289 zone = &pgdat->node_zones[i];
6290 if (!managed_zone(zone))
6291 continue;
6292
6293 if (!zone_reclaimable_pages(zone))
6294 continue;
6295
6296 pfmemalloc_reserve += min_wmark_pages(zone);
6297 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6298 }
6299
6300 /* If there are no reserves (unexpected config) then do not throttle */
6301 if (!pfmemalloc_reserve)
6302 return true;
6303
6304 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6305
6306 /* kswapd must be awake if processes are being throttled */
6307 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6308 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6309 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6310
6311 wake_up_interruptible(&pgdat->kswapd_wait);
6312 }
6313
6314 return wmark_ok;
6315}
6316
6317/*
6318 * Throttle direct reclaimers if backing storage is backed by the network
6319 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6320 * depleted. kswapd will continue to make progress and wake the processes
6321 * when the low watermark is reached.
6322 *
6323 * Returns true if a fatal signal was delivered during throttling. If this
6324 * happens, the page allocator should not consider triggering the OOM killer.
6325 */
6326static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6327 nodemask_t *nodemask)
6328{
6329 struct zoneref *z;
6330 struct zone *zone;
6331 pg_data_t *pgdat = NULL;
6332
6333 /*
6334 * Kernel threads should not be throttled as they may be indirectly
6335 * responsible for cleaning pages necessary for reclaim to make forward
6336 * progress. kjournald for example may enter direct reclaim while
6337 * committing a transaction where throttling it could forcing other
6338 * processes to block on log_wait_commit().
6339 */
6340 if (current->flags & PF_KTHREAD)
6341 goto out;
6342
6343 /*
6344 * If a fatal signal is pending, this process should not throttle.
6345 * It should return quickly so it can exit and free its memory
6346 */
6347 if (fatal_signal_pending(current))
6348 goto out;
6349
6350 /*
6351 * Check if the pfmemalloc reserves are ok by finding the first node
6352 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6353 * GFP_KERNEL will be required for allocating network buffers when
6354 * swapping over the network so ZONE_HIGHMEM is unusable.
6355 *
6356 * Throttling is based on the first usable node and throttled processes
6357 * wait on a queue until kswapd makes progress and wakes them. There
6358 * is an affinity then between processes waking up and where reclaim
6359 * progress has been made assuming the process wakes on the same node.
6360 * More importantly, processes running on remote nodes will not compete
6361 * for remote pfmemalloc reserves and processes on different nodes
6362 * should make reasonable progress.
6363 */
6364 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6365 gfp_zone(gfp_mask), nodemask) {
6366 if (zone_idx(zone) > ZONE_NORMAL)
6367 continue;
6368
6369 /* Throttle based on the first usable node */
6370 pgdat = zone->zone_pgdat;
6371 if (allow_direct_reclaim(pgdat))
6372 goto out;
6373 break;
6374 }
6375
6376 /* If no zone was usable by the allocation flags then do not throttle */
6377 if (!pgdat)
6378 goto out;
6379
6380 /* Account for the throttling */
6381 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6382
6383 /*
6384 * If the caller cannot enter the filesystem, it's possible that it
6385 * is due to the caller holding an FS lock or performing a journal
6386 * transaction in the case of a filesystem like ext[3|4]. In this case,
6387 * it is not safe to block on pfmemalloc_wait as kswapd could be
6388 * blocked waiting on the same lock. Instead, throttle for up to a
6389 * second before continuing.
6390 */
6391 if (!(gfp_mask & __GFP_FS))
6392 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6393 allow_direct_reclaim(pgdat), HZ);
6394 else
6395 /* Throttle until kswapd wakes the process */
6396 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6397 allow_direct_reclaim(pgdat));
6398
6399 if (fatal_signal_pending(current))
6400 return true;
6401
6402out:
6403 return false;
6404}
6405
6406unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6407 gfp_t gfp_mask, nodemask_t *nodemask)
6408{
6409 unsigned long nr_reclaimed;
6410 struct scan_control sc = {
6411 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6412 .gfp_mask = current_gfp_context(gfp_mask),
6413 .reclaim_idx = gfp_zone(gfp_mask),
6414 .order = order,
6415 .nodemask = nodemask,
6416 .priority = DEF_PRIORITY,
6417 .may_writepage = !laptop_mode,
6418 .may_unmap = 1,
6419 .may_swap = 1,
6420 };
6421
6422 /*
6423 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6424 * Confirm they are large enough for max values.
6425 */
6426 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6427 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6428 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6429
6430 /*
6431 * Do not enter reclaim if fatal signal was delivered while throttled.
6432 * 1 is returned so that the page allocator does not OOM kill at this
6433 * point.
6434 */
6435 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6436 return 1;
6437
6438 set_task_reclaim_state(current, &sc.reclaim_state);
6439 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6440
6441 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6442
6443 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6444 set_task_reclaim_state(current, NULL);
6445
6446 return nr_reclaimed;
6447}
6448
6449#ifdef CONFIG_MEMCG
6450
6451/* Only used by soft limit reclaim. Do not reuse for anything else. */
6452unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6453 gfp_t gfp_mask, bool noswap,
6454 pg_data_t *pgdat,
6455 unsigned long *nr_scanned)
6456{
6457 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6458 struct scan_control sc = {
6459 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6460 .target_mem_cgroup = memcg,
6461 .may_writepage = !laptop_mode,
6462 .may_unmap = 1,
6463 .reclaim_idx = MAX_NR_ZONES - 1,
6464 .may_swap = !noswap,
6465 };
6466
6467 WARN_ON_ONCE(!current->reclaim_state);
6468
6469 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6470 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6471
6472 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6473 sc.gfp_mask);
6474
6475 /*
6476 * NOTE: Although we can get the priority field, using it
6477 * here is not a good idea, since it limits the pages we can scan.
6478 * if we don't reclaim here, the shrink_node from balance_pgdat
6479 * will pick up pages from other mem cgroup's as well. We hack
6480 * the priority and make it zero.
6481 */
6482 shrink_lruvec(lruvec, &sc);
6483
6484 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6485
6486 *nr_scanned = sc.nr_scanned;
6487
6488 return sc.nr_reclaimed;
6489}
6490
6491unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6492 unsigned long nr_pages,
6493 gfp_t gfp_mask,
6494 unsigned int reclaim_options)
6495{
6496 unsigned long nr_reclaimed;
6497 unsigned int noreclaim_flag;
6498 struct scan_control sc = {
6499 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6500 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6501 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6502 .reclaim_idx = MAX_NR_ZONES - 1,
6503 .target_mem_cgroup = memcg,
6504 .priority = DEF_PRIORITY,
6505 .may_writepage = !laptop_mode,
6506 .may_unmap = 1,
6507 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6508 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6509 };
6510 /*
6511 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6512 * equal pressure on all the nodes. This is based on the assumption that
6513 * the reclaim does not bail out early.
6514 */
6515 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6516
6517 set_task_reclaim_state(current, &sc.reclaim_state);
6518 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6519 noreclaim_flag = memalloc_noreclaim_save();
6520
6521 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6522
6523 memalloc_noreclaim_restore(noreclaim_flag);
6524 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6525 set_task_reclaim_state(current, NULL);
6526
6527 return nr_reclaimed;
6528}
6529#endif
6530
6531static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6532{
6533 struct mem_cgroup *memcg;
6534 struct lruvec *lruvec;
6535
6536 if (lru_gen_enabled()) {
6537 lru_gen_age_node(pgdat, sc);
6538 return;
6539 }
6540
6541 if (!can_age_anon_pages(pgdat, sc))
6542 return;
6543
6544 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6545 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6546 return;
6547
6548 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6549 do {
6550 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6551 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6552 sc, LRU_ACTIVE_ANON);
6553 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6554 } while (memcg);
6555}
6556
6557static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6558{
6559 int i;
6560 struct zone *zone;
6561
6562 /*
6563 * Check for watermark boosts top-down as the higher zones
6564 * are more likely to be boosted. Both watermarks and boosts
6565 * should not be checked at the same time as reclaim would
6566 * start prematurely when there is no boosting and a lower
6567 * zone is balanced.
6568 */
6569 for (i = highest_zoneidx; i >= 0; i--) {
6570 zone = pgdat->node_zones + i;
6571 if (!managed_zone(zone))
6572 continue;
6573
6574 if (zone->watermark_boost)
6575 return true;
6576 }
6577
6578 return false;
6579}
6580
6581/*
6582 * Returns true if there is an eligible zone balanced for the request order
6583 * and highest_zoneidx
6584 */
6585static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6586{
6587 int i;
6588 unsigned long mark = -1;
6589 struct zone *zone;
6590
6591 /*
6592 * Check watermarks bottom-up as lower zones are more likely to
6593 * meet watermarks.
6594 */
6595 for (i = 0; i <= highest_zoneidx; i++) {
6596 zone = pgdat->node_zones + i;
6597
6598 if (!managed_zone(zone))
6599 continue;
6600
6601 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6602 mark = wmark_pages(zone, WMARK_PROMO);
6603 else
6604 mark = high_wmark_pages(zone);
6605 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6606 return true;
6607 }
6608
6609 /*
6610 * If a node has no managed zone within highest_zoneidx, it does not
6611 * need balancing by definition. This can happen if a zone-restricted
6612 * allocation tries to wake a remote kswapd.
6613 */
6614 if (mark == -1)
6615 return true;
6616
6617 return false;
6618}
6619
6620/* Clear pgdat state for congested, dirty or under writeback. */
6621static void clear_pgdat_congested(pg_data_t *pgdat)
6622{
6623 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6624
6625 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6626 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6627 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6628 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6629}
6630
6631/*
6632 * Prepare kswapd for sleeping. This verifies that there are no processes
6633 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6634 *
6635 * Returns true if kswapd is ready to sleep
6636 */
6637static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6638 int highest_zoneidx)
6639{
6640 /*
6641 * The throttled processes are normally woken up in balance_pgdat() as
6642 * soon as allow_direct_reclaim() is true. But there is a potential
6643 * race between when kswapd checks the watermarks and a process gets
6644 * throttled. There is also a potential race if processes get
6645 * throttled, kswapd wakes, a large process exits thereby balancing the
6646 * zones, which causes kswapd to exit balance_pgdat() before reaching
6647 * the wake up checks. If kswapd is going to sleep, no process should
6648 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6649 * the wake up is premature, processes will wake kswapd and get
6650 * throttled again. The difference from wake ups in balance_pgdat() is
6651 * that here we are under prepare_to_wait().
6652 */
6653 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6654 wake_up_all(&pgdat->pfmemalloc_wait);
6655
6656 /* Hopeless node, leave it to direct reclaim */
6657 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6658 return true;
6659
6660 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6661 clear_pgdat_congested(pgdat);
6662 return true;
6663 }
6664
6665 return false;
6666}
6667
6668/*
6669 * kswapd shrinks a node of pages that are at or below the highest usable
6670 * zone that is currently unbalanced.
6671 *
6672 * Returns true if kswapd scanned at least the requested number of pages to
6673 * reclaim or if the lack of progress was due to pages under writeback.
6674 * This is used to determine if the scanning priority needs to be raised.
6675 */
6676static bool kswapd_shrink_node(pg_data_t *pgdat,
6677 struct scan_control *sc)
6678{
6679 struct zone *zone;
6680 int z;
6681
6682 /* Reclaim a number of pages proportional to the number of zones */
6683 sc->nr_to_reclaim = 0;
6684 for (z = 0; z <= sc->reclaim_idx; z++) {
6685 zone = pgdat->node_zones + z;
6686 if (!managed_zone(zone))
6687 continue;
6688
6689 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6690 }
6691
6692 /*
6693 * Historically care was taken to put equal pressure on all zones but
6694 * now pressure is applied based on node LRU order.
6695 */
6696 shrink_node(pgdat, sc);
6697
6698 /*
6699 * Fragmentation may mean that the system cannot be rebalanced for
6700 * high-order allocations. If twice the allocation size has been
6701 * reclaimed then recheck watermarks only at order-0 to prevent
6702 * excessive reclaim. Assume that a process requested a high-order
6703 * can direct reclaim/compact.
6704 */
6705 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6706 sc->order = 0;
6707
6708 return sc->nr_scanned >= sc->nr_to_reclaim;
6709}
6710
6711/* Page allocator PCP high watermark is lowered if reclaim is active. */
6712static inline void
6713update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6714{
6715 int i;
6716 struct zone *zone;
6717
6718 for (i = 0; i <= highest_zoneidx; i++) {
6719 zone = pgdat->node_zones + i;
6720
6721 if (!managed_zone(zone))
6722 continue;
6723
6724 if (active)
6725 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6726 else
6727 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6728 }
6729}
6730
6731static inline void
6732set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6733{
6734 update_reclaim_active(pgdat, highest_zoneidx, true);
6735}
6736
6737static inline void
6738clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6739{
6740 update_reclaim_active(pgdat, highest_zoneidx, false);
6741}
6742
6743/*
6744 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6745 * that are eligible for use by the caller until at least one zone is
6746 * balanced.
6747 *
6748 * Returns the order kswapd finished reclaiming at.
6749 *
6750 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6751 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6752 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6753 * or lower is eligible for reclaim until at least one usable zone is
6754 * balanced.
6755 */
6756static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6757{
6758 int i;
6759 unsigned long nr_soft_reclaimed;
6760 unsigned long nr_soft_scanned;
6761 unsigned long pflags;
6762 unsigned long nr_boost_reclaim;
6763 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6764 bool boosted;
6765 struct zone *zone;
6766 struct scan_control sc = {
6767 .gfp_mask = GFP_KERNEL,
6768 .order = order,
6769 .may_unmap = 1,
6770 };
6771
6772 set_task_reclaim_state(current, &sc.reclaim_state);
6773 psi_memstall_enter(&pflags);
6774 __fs_reclaim_acquire(_THIS_IP_);
6775
6776 count_vm_event(PAGEOUTRUN);
6777
6778 /*
6779 * Account for the reclaim boost. Note that the zone boost is left in
6780 * place so that parallel allocations that are near the watermark will
6781 * stall or direct reclaim until kswapd is finished.
6782 */
6783 nr_boost_reclaim = 0;
6784 for (i = 0; i <= highest_zoneidx; i++) {
6785 zone = pgdat->node_zones + i;
6786 if (!managed_zone(zone))
6787 continue;
6788
6789 nr_boost_reclaim += zone->watermark_boost;
6790 zone_boosts[i] = zone->watermark_boost;
6791 }
6792 boosted = nr_boost_reclaim;
6793
6794restart:
6795 set_reclaim_active(pgdat, highest_zoneidx);
6796 sc.priority = DEF_PRIORITY;
6797 do {
6798 unsigned long nr_reclaimed = sc.nr_reclaimed;
6799 bool raise_priority = true;
6800 bool balanced;
6801 bool ret;
6802
6803 sc.reclaim_idx = highest_zoneidx;
6804
6805 /*
6806 * If the number of buffer_heads exceeds the maximum allowed
6807 * then consider reclaiming from all zones. This has a dual
6808 * purpose -- on 64-bit systems it is expected that
6809 * buffer_heads are stripped during active rotation. On 32-bit
6810 * systems, highmem pages can pin lowmem memory and shrinking
6811 * buffers can relieve lowmem pressure. Reclaim may still not
6812 * go ahead if all eligible zones for the original allocation
6813 * request are balanced to avoid excessive reclaim from kswapd.
6814 */
6815 if (buffer_heads_over_limit) {
6816 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6817 zone = pgdat->node_zones + i;
6818 if (!managed_zone(zone))
6819 continue;
6820
6821 sc.reclaim_idx = i;
6822 break;
6823 }
6824 }
6825
6826 /*
6827 * If the pgdat is imbalanced then ignore boosting and preserve
6828 * the watermarks for a later time and restart. Note that the
6829 * zone watermarks will be still reset at the end of balancing
6830 * on the grounds that the normal reclaim should be enough to
6831 * re-evaluate if boosting is required when kswapd next wakes.
6832 */
6833 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6834 if (!balanced && nr_boost_reclaim) {
6835 nr_boost_reclaim = 0;
6836 goto restart;
6837 }
6838
6839 /*
6840 * If boosting is not active then only reclaim if there are no
6841 * eligible zones. Note that sc.reclaim_idx is not used as
6842 * buffer_heads_over_limit may have adjusted it.
6843 */
6844 if (!nr_boost_reclaim && balanced)
6845 goto out;
6846
6847 /* Limit the priority of boosting to avoid reclaim writeback */
6848 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6849 raise_priority = false;
6850
6851 /*
6852 * Do not writeback or swap pages for boosted reclaim. The
6853 * intent is to relieve pressure not issue sub-optimal IO
6854 * from reclaim context. If no pages are reclaimed, the
6855 * reclaim will be aborted.
6856 */
6857 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6858 sc.may_swap = !nr_boost_reclaim;
6859
6860 /*
6861 * Do some background aging, to give pages a chance to be
6862 * referenced before reclaiming. All pages are rotated
6863 * regardless of classzone as this is about consistent aging.
6864 */
6865 kswapd_age_node(pgdat, &sc);
6866
6867 /*
6868 * If we're getting trouble reclaiming, start doing writepage
6869 * even in laptop mode.
6870 */
6871 if (sc.priority < DEF_PRIORITY - 2)
6872 sc.may_writepage = 1;
6873
6874 /* Call soft limit reclaim before calling shrink_node. */
6875 sc.nr_scanned = 0;
6876 nr_soft_scanned = 0;
6877 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
6878 sc.gfp_mask, &nr_soft_scanned);
6879 sc.nr_reclaimed += nr_soft_reclaimed;
6880
6881 /*
6882 * There should be no need to raise the scanning priority if
6883 * enough pages are already being scanned that that high
6884 * watermark would be met at 100% efficiency.
6885 */
6886 if (kswapd_shrink_node(pgdat, &sc))
6887 raise_priority = false;
6888
6889 /*
6890 * If the low watermark is met there is no need for processes
6891 * to be throttled on pfmemalloc_wait as they should not be
6892 * able to safely make forward progress. Wake them
6893 */
6894 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6895 allow_direct_reclaim(pgdat))
6896 wake_up_all(&pgdat->pfmemalloc_wait);
6897
6898 /* Check if kswapd should be suspending */
6899 __fs_reclaim_release(_THIS_IP_);
6900 ret = try_to_freeze();
6901 __fs_reclaim_acquire(_THIS_IP_);
6902 if (ret || kthread_should_stop())
6903 break;
6904
6905 /*
6906 * Raise priority if scanning rate is too low or there was no
6907 * progress in reclaiming pages
6908 */
6909 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6910 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6911
6912 /*
6913 * If reclaim made no progress for a boost, stop reclaim as
6914 * IO cannot be queued and it could be an infinite loop in
6915 * extreme circumstances.
6916 */
6917 if (nr_boost_reclaim && !nr_reclaimed)
6918 break;
6919
6920 if (raise_priority || !nr_reclaimed)
6921 sc.priority--;
6922 } while (sc.priority >= 1);
6923
6924 if (!sc.nr_reclaimed)
6925 pgdat->kswapd_failures++;
6926
6927out:
6928 clear_reclaim_active(pgdat, highest_zoneidx);
6929
6930 /* If reclaim was boosted, account for the reclaim done in this pass */
6931 if (boosted) {
6932 unsigned long flags;
6933
6934 for (i = 0; i <= highest_zoneidx; i++) {
6935 if (!zone_boosts[i])
6936 continue;
6937
6938 /* Increments are under the zone lock */
6939 zone = pgdat->node_zones + i;
6940 spin_lock_irqsave(&zone->lock, flags);
6941 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6942 spin_unlock_irqrestore(&zone->lock, flags);
6943 }
6944
6945 /*
6946 * As there is now likely space, wakeup kcompact to defragment
6947 * pageblocks.
6948 */
6949 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
6950 }
6951
6952 snapshot_refaults(NULL, pgdat);
6953 __fs_reclaim_release(_THIS_IP_);
6954 psi_memstall_leave(&pflags);
6955 set_task_reclaim_state(current, NULL);
6956
6957 /*
6958 * Return the order kswapd stopped reclaiming at as
6959 * prepare_kswapd_sleep() takes it into account. If another caller
6960 * entered the allocator slow path while kswapd was awake, order will
6961 * remain at the higher level.
6962 */
6963 return sc.order;
6964}
6965
6966/*
6967 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6968 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6969 * not a valid index then either kswapd runs for first time or kswapd couldn't
6970 * sleep after previous reclaim attempt (node is still unbalanced). In that
6971 * case return the zone index of the previous kswapd reclaim cycle.
6972 */
6973static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6974 enum zone_type prev_highest_zoneidx)
6975{
6976 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
6977
6978 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
6979}
6980
6981static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
6982 unsigned int highest_zoneidx)
6983{
6984 long remaining = 0;
6985 DEFINE_WAIT(wait);
6986
6987 if (freezing(current) || kthread_should_stop())
6988 return;
6989
6990 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
6991
6992 /*
6993 * Try to sleep for a short interval. Note that kcompactd will only be
6994 * woken if it is possible to sleep for a short interval. This is
6995 * deliberate on the assumption that if reclaim cannot keep an
6996 * eligible zone balanced that it's also unlikely that compaction will
6997 * succeed.
6998 */
6999 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7000 /*
7001 * Compaction records what page blocks it recently failed to
7002 * isolate pages from and skips them in the future scanning.
7003 * When kswapd is going to sleep, it is reasonable to assume
7004 * that pages and compaction may succeed so reset the cache.
7005 */
7006 reset_isolation_suitable(pgdat);
7007
7008 /*
7009 * We have freed the memory, now we should compact it to make
7010 * allocation of the requested order possible.
7011 */
7012 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7013
7014 remaining = schedule_timeout(HZ/10);
7015
7016 /*
7017 * If woken prematurely then reset kswapd_highest_zoneidx and
7018 * order. The values will either be from a wakeup request or
7019 * the previous request that slept prematurely.
7020 */
7021 if (remaining) {
7022 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7023 kswapd_highest_zoneidx(pgdat,
7024 highest_zoneidx));
7025
7026 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7027 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7028 }
7029
7030 finish_wait(&pgdat->kswapd_wait, &wait);
7031 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7032 }
7033
7034 /*
7035 * After a short sleep, check if it was a premature sleep. If not, then
7036 * go fully to sleep until explicitly woken up.
7037 */
7038 if (!remaining &&
7039 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7040 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7041
7042 /*
7043 * vmstat counters are not perfectly accurate and the estimated
7044 * value for counters such as NR_FREE_PAGES can deviate from the
7045 * true value by nr_online_cpus * threshold. To avoid the zone
7046 * watermarks being breached while under pressure, we reduce the
7047 * per-cpu vmstat threshold while kswapd is awake and restore
7048 * them before going back to sleep.
7049 */
7050 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7051
7052 if (!kthread_should_stop())
7053 schedule();
7054
7055 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7056 } else {
7057 if (remaining)
7058 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7059 else
7060 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7061 }
7062 finish_wait(&pgdat->kswapd_wait, &wait);
7063}
7064
7065/*
7066 * The background pageout daemon, started as a kernel thread
7067 * from the init process.
7068 *
7069 * This basically trickles out pages so that we have _some_
7070 * free memory available even if there is no other activity
7071 * that frees anything up. This is needed for things like routing
7072 * etc, where we otherwise might have all activity going on in
7073 * asynchronous contexts that cannot page things out.
7074 *
7075 * If there are applications that are active memory-allocators
7076 * (most normal use), this basically shouldn't matter.
7077 */
7078static int kswapd(void *p)
7079{
7080 unsigned int alloc_order, reclaim_order;
7081 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7082 pg_data_t *pgdat = (pg_data_t *)p;
7083 struct task_struct *tsk = current;
7084 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7085
7086 if (!cpumask_empty(cpumask))
7087 set_cpus_allowed_ptr(tsk, cpumask);
7088
7089 /*
7090 * Tell the memory management that we're a "memory allocator",
7091 * and that if we need more memory we should get access to it
7092 * regardless (see "__alloc_pages()"). "kswapd" should
7093 * never get caught in the normal page freeing logic.
7094 *
7095 * (Kswapd normally doesn't need memory anyway, but sometimes
7096 * you need a small amount of memory in order to be able to
7097 * page out something else, and this flag essentially protects
7098 * us from recursively trying to free more memory as we're
7099 * trying to free the first piece of memory in the first place).
7100 */
7101 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7102 set_freezable();
7103
7104 WRITE_ONCE(pgdat->kswapd_order, 0);
7105 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7106 atomic_set(&pgdat->nr_writeback_throttled, 0);
7107 for ( ; ; ) {
7108 bool ret;
7109
7110 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7111 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7112 highest_zoneidx);
7113
7114kswapd_try_sleep:
7115 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7116 highest_zoneidx);
7117
7118 /* Read the new order and highest_zoneidx */
7119 alloc_order = READ_ONCE(pgdat->kswapd_order);
7120 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7121 highest_zoneidx);
7122 WRITE_ONCE(pgdat->kswapd_order, 0);
7123 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7124
7125 ret = try_to_freeze();
7126 if (kthread_should_stop())
7127 break;
7128
7129 /*
7130 * We can speed up thawing tasks if we don't call balance_pgdat
7131 * after returning from the refrigerator
7132 */
7133 if (ret)
7134 continue;
7135
7136 /*
7137 * Reclaim begins at the requested order but if a high-order
7138 * reclaim fails then kswapd falls back to reclaiming for
7139 * order-0. If that happens, kswapd will consider sleeping
7140 * for the order it finished reclaiming at (reclaim_order)
7141 * but kcompactd is woken to compact for the original
7142 * request (alloc_order).
7143 */
7144 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7145 alloc_order);
7146 reclaim_order = balance_pgdat(pgdat, alloc_order,
7147 highest_zoneidx);
7148 if (reclaim_order < alloc_order)
7149 goto kswapd_try_sleep;
7150 }
7151
7152 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7153
7154 return 0;
7155}
7156
7157/*
7158 * A zone is low on free memory or too fragmented for high-order memory. If
7159 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7160 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7161 * has failed or is not needed, still wake up kcompactd if only compaction is
7162 * needed.
7163 */
7164void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7165 enum zone_type highest_zoneidx)
7166{
7167 pg_data_t *pgdat;
7168 enum zone_type curr_idx;
7169
7170 if (!managed_zone(zone))
7171 return;
7172
7173 if (!cpuset_zone_allowed(zone, gfp_flags))
7174 return;
7175
7176 pgdat = zone->zone_pgdat;
7177 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7178
7179 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7180 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7181
7182 if (READ_ONCE(pgdat->kswapd_order) < order)
7183 WRITE_ONCE(pgdat->kswapd_order, order);
7184
7185 if (!waitqueue_active(&pgdat->kswapd_wait))
7186 return;
7187
7188 /* Hopeless node, leave it to direct reclaim if possible */
7189 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7190 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7191 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7192 /*
7193 * There may be plenty of free memory available, but it's too
7194 * fragmented for high-order allocations. Wake up kcompactd
7195 * and rely on compaction_suitable() to determine if it's
7196 * needed. If it fails, it will defer subsequent attempts to
7197 * ratelimit its work.
7198 */
7199 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7200 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7201 return;
7202 }
7203
7204 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7205 gfp_flags);
7206 wake_up_interruptible(&pgdat->kswapd_wait);
7207}
7208
7209#ifdef CONFIG_HIBERNATION
7210/*
7211 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7212 * freed pages.
7213 *
7214 * Rather than trying to age LRUs the aim is to preserve the overall
7215 * LRU order by reclaiming preferentially
7216 * inactive > active > active referenced > active mapped
7217 */
7218unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7219{
7220 struct scan_control sc = {
7221 .nr_to_reclaim = nr_to_reclaim,
7222 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7223 .reclaim_idx = MAX_NR_ZONES - 1,
7224 .priority = DEF_PRIORITY,
7225 .may_writepage = 1,
7226 .may_unmap = 1,
7227 .may_swap = 1,
7228 .hibernation_mode = 1,
7229 };
7230 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7231 unsigned long nr_reclaimed;
7232 unsigned int noreclaim_flag;
7233
7234 fs_reclaim_acquire(sc.gfp_mask);
7235 noreclaim_flag = memalloc_noreclaim_save();
7236 set_task_reclaim_state(current, &sc.reclaim_state);
7237
7238 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7239
7240 set_task_reclaim_state(current, NULL);
7241 memalloc_noreclaim_restore(noreclaim_flag);
7242 fs_reclaim_release(sc.gfp_mask);
7243
7244 return nr_reclaimed;
7245}
7246#endif /* CONFIG_HIBERNATION */
7247
7248/*
7249 * This kswapd start function will be called by init and node-hot-add.
7250 */
7251void __meminit kswapd_run(int nid)
7252{
7253 pg_data_t *pgdat = NODE_DATA(nid);
7254
7255 pgdat_kswapd_lock(pgdat);
7256 if (!pgdat->kswapd) {
7257 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7258 if (IS_ERR(pgdat->kswapd)) {
7259 /* failure at boot is fatal */
7260 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7261 nid, PTR_ERR(pgdat->kswapd));
7262 BUG_ON(system_state < SYSTEM_RUNNING);
7263 pgdat->kswapd = NULL;
7264 }
7265 }
7266 pgdat_kswapd_unlock(pgdat);
7267}
7268
7269/*
7270 * Called by memory hotplug when all memory in a node is offlined. Caller must
7271 * be holding mem_hotplug_begin/done().
7272 */
7273void __meminit kswapd_stop(int nid)
7274{
7275 pg_data_t *pgdat = NODE_DATA(nid);
7276 struct task_struct *kswapd;
7277
7278 pgdat_kswapd_lock(pgdat);
7279 kswapd = pgdat->kswapd;
7280 if (kswapd) {
7281 kthread_stop(kswapd);
7282 pgdat->kswapd = NULL;
7283 }
7284 pgdat_kswapd_unlock(pgdat);
7285}
7286
7287static int __init kswapd_init(void)
7288{
7289 int nid;
7290
7291 swap_setup();
7292 for_each_node_state(nid, N_MEMORY)
7293 kswapd_run(nid);
7294 return 0;
7295}
7296
7297module_init(kswapd_init)
7298
7299#ifdef CONFIG_NUMA
7300/*
7301 * Node reclaim mode
7302 *
7303 * If non-zero call node_reclaim when the number of free pages falls below
7304 * the watermarks.
7305 */
7306int node_reclaim_mode __read_mostly;
7307
7308/*
7309 * Priority for NODE_RECLAIM. This determines the fraction of pages
7310 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7311 * a zone.
7312 */
7313#define NODE_RECLAIM_PRIORITY 4
7314
7315/*
7316 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7317 * occur.
7318 */
7319int sysctl_min_unmapped_ratio = 1;
7320
7321/*
7322 * If the number of slab pages in a zone grows beyond this percentage then
7323 * slab reclaim needs to occur.
7324 */
7325int sysctl_min_slab_ratio = 5;
7326
7327static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7328{
7329 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7330 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7331 node_page_state(pgdat, NR_ACTIVE_FILE);
7332
7333 /*
7334 * It's possible for there to be more file mapped pages than
7335 * accounted for by the pages on the file LRU lists because
7336 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7337 */
7338 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7339}
7340
7341/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7342static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7343{
7344 unsigned long nr_pagecache_reclaimable;
7345 unsigned long delta = 0;
7346
7347 /*
7348 * If RECLAIM_UNMAP is set, then all file pages are considered
7349 * potentially reclaimable. Otherwise, we have to worry about
7350 * pages like swapcache and node_unmapped_file_pages() provides
7351 * a better estimate
7352 */
7353 if (node_reclaim_mode & RECLAIM_UNMAP)
7354 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7355 else
7356 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7357
7358 /* If we can't clean pages, remove dirty pages from consideration */
7359 if (!(node_reclaim_mode & RECLAIM_WRITE))
7360 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7361
7362 /* Watch for any possible underflows due to delta */
7363 if (unlikely(delta > nr_pagecache_reclaimable))
7364 delta = nr_pagecache_reclaimable;
7365
7366 return nr_pagecache_reclaimable - delta;
7367}
7368
7369/*
7370 * Try to free up some pages from this node through reclaim.
7371 */
7372static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7373{
7374 /* Minimum pages needed in order to stay on node */
7375 const unsigned long nr_pages = 1 << order;
7376 struct task_struct *p = current;
7377 unsigned int noreclaim_flag;
7378 struct scan_control sc = {
7379 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7380 .gfp_mask = current_gfp_context(gfp_mask),
7381 .order = order,
7382 .priority = NODE_RECLAIM_PRIORITY,
7383 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7384 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7385 .may_swap = 1,
7386 .reclaim_idx = gfp_zone(gfp_mask),
7387 };
7388 unsigned long pflags;
7389
7390 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7391 sc.gfp_mask);
7392
7393 cond_resched();
7394 psi_memstall_enter(&pflags);
7395 delayacct_freepages_start();
7396 fs_reclaim_acquire(sc.gfp_mask);
7397 /*
7398 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7399 */
7400 noreclaim_flag = memalloc_noreclaim_save();
7401 set_task_reclaim_state(p, &sc.reclaim_state);
7402
7403 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7404 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7405 /*
7406 * Free memory by calling shrink node with increasing
7407 * priorities until we have enough memory freed.
7408 */
7409 do {
7410 shrink_node(pgdat, &sc);
7411 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7412 }
7413
7414 set_task_reclaim_state(p, NULL);
7415 memalloc_noreclaim_restore(noreclaim_flag);
7416 fs_reclaim_release(sc.gfp_mask);
7417 psi_memstall_leave(&pflags);
7418 delayacct_freepages_end();
7419
7420 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7421
7422 return sc.nr_reclaimed >= nr_pages;
7423}
7424
7425int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7426{
7427 int ret;
7428
7429 /*
7430 * Node reclaim reclaims unmapped file backed pages and
7431 * slab pages if we are over the defined limits.
7432 *
7433 * A small portion of unmapped file backed pages is needed for
7434 * file I/O otherwise pages read by file I/O will be immediately
7435 * thrown out if the node is overallocated. So we do not reclaim
7436 * if less than a specified percentage of the node is used by
7437 * unmapped file backed pages.
7438 */
7439 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7440 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7441 pgdat->min_slab_pages)
7442 return NODE_RECLAIM_FULL;
7443
7444 /*
7445 * Do not scan if the allocation should not be delayed.
7446 */
7447 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7448 return NODE_RECLAIM_NOSCAN;
7449
7450 /*
7451 * Only run node reclaim on the local node or on nodes that do not
7452 * have associated processors. This will favor the local processor
7453 * over remote processors and spread off node memory allocations
7454 * as wide as possible.
7455 */
7456 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7457 return NODE_RECLAIM_NOSCAN;
7458
7459 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7460 return NODE_RECLAIM_NOSCAN;
7461
7462 ret = __node_reclaim(pgdat, gfp_mask, order);
7463 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7464
7465 if (!ret)
7466 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7467
7468 return ret;
7469}
7470#endif
7471
7472/**
7473 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7474 * lru list
7475 * @fbatch: Batch of lru folios to check.
7476 *
7477 * Checks folios for evictability, if an evictable folio is in the unevictable
7478 * lru list, moves it to the appropriate evictable lru list. This function
7479 * should be only used for lru folios.
7480 */
7481void check_move_unevictable_folios(struct folio_batch *fbatch)
7482{
7483 struct lruvec *lruvec = NULL;
7484 int pgscanned = 0;
7485 int pgrescued = 0;
7486 int i;
7487
7488 for (i = 0; i < fbatch->nr; i++) {
7489 struct folio *folio = fbatch->folios[i];
7490 int nr_pages = folio_nr_pages(folio);
7491
7492 pgscanned += nr_pages;
7493
7494 /* block memcg migration while the folio moves between lrus */
7495 if (!folio_test_clear_lru(folio))
7496 continue;
7497
7498 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7499 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7500 lruvec_del_folio(lruvec, folio);
7501 folio_clear_unevictable(folio);
7502 lruvec_add_folio(lruvec, folio);
7503 pgrescued += nr_pages;
7504 }
7505 folio_set_lru(folio);
7506 }
7507
7508 if (lruvec) {
7509 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7510 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7511 unlock_page_lruvec_irq(lruvec);
7512 } else if (pgscanned) {
7513 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7514 }
7515}
7516EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
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
60#include <asm/tlbflush.h>
61#include <asm/div64.h>
62
63#include <linux/swapops.h>
64#include <linux/balloon_compaction.h>
65#include <linux/sched/sysctl.h>
66
67#include "internal.h"
68#include "swap.h"
69
70#define CREATE_TRACE_POINTS
71#include <trace/events/vmscan.h>
72
73struct scan_control {
74 /* How many pages shrink_list() should reclaim */
75 unsigned long nr_to_reclaim;
76
77 /*
78 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * are scanned.
80 */
81 nodemask_t *nodemask;
82
83 /*
84 * The memory cgroup that hit its limit and as a result is the
85 * primary target of this reclaim invocation.
86 */
87 struct mem_cgroup *target_mem_cgroup;
88
89 /*
90 * Scan pressure balancing between anon and file LRUs
91 */
92 unsigned long anon_cost;
93 unsigned long file_cost;
94
95 /* Can active folios be deactivated as part of reclaim? */
96#define DEACTIVATE_ANON 1
97#define DEACTIVATE_FILE 2
98 unsigned int may_deactivate:2;
99 unsigned int force_deactivate:1;
100 unsigned int skipped_deactivate:1;
101
102 /* Writepage batching in laptop mode; RECLAIM_WRITE */
103 unsigned int may_writepage:1;
104
105 /* Can mapped folios be reclaimed? */
106 unsigned int may_unmap:1;
107
108 /* Can folios be swapped as part of reclaim? */
109 unsigned int may_swap:1;
110
111 /* Not allow cache_trim_mode to be turned on as part of reclaim? */
112 unsigned int no_cache_trim_mode:1;
113
114 /* Has cache_trim_mode failed at least once? */
115 unsigned int cache_trim_mode_failed:1;
116
117 /* Proactive reclaim invoked by userspace through memory.reclaim */
118 unsigned int proactive:1;
119
120 /*
121 * Cgroup memory below memory.low is protected as long as we
122 * don't threaten to OOM. If any cgroup is reclaimed at
123 * reduced force or passed over entirely due to its memory.low
124 * setting (memcg_low_skipped), and nothing is reclaimed as a
125 * result, then go back for one more cycle that reclaims the protected
126 * memory (memcg_low_reclaim) to avert OOM.
127 */
128 unsigned int memcg_low_reclaim:1;
129 unsigned int memcg_low_skipped:1;
130
131 unsigned int hibernation_mode:1;
132
133 /* One of the zones is ready for compaction */
134 unsigned int compaction_ready:1;
135
136 /* There is easily reclaimable cold cache in the current node */
137 unsigned int cache_trim_mode:1;
138
139 /* The file folios on the current node are dangerously low */
140 unsigned int file_is_tiny:1;
141
142 /* Always discard instead of demoting to lower tier memory */
143 unsigned int no_demotion:1;
144
145 /* Allocation order */
146 s8 order;
147
148 /* Scan (total_size >> priority) pages at once */
149 s8 priority;
150
151 /* The highest zone to isolate folios for reclaim from */
152 s8 reclaim_idx;
153
154 /* This context's GFP mask */
155 gfp_t gfp_mask;
156
157 /* Incremented by the number of inactive pages that were scanned */
158 unsigned long nr_scanned;
159
160 /* Number of pages freed so far during a call to shrink_zones() */
161 unsigned long nr_reclaimed;
162
163 struct {
164 unsigned int dirty;
165 unsigned int unqueued_dirty;
166 unsigned int congested;
167 unsigned int writeback;
168 unsigned int immediate;
169 unsigned int file_taken;
170 unsigned int taken;
171 } nr;
172
173 /* for recording the reclaimed slab by now */
174 struct reclaim_state reclaim_state;
175};
176
177#ifdef ARCH_HAS_PREFETCHW
178#define prefetchw_prev_lru_folio(_folio, _base, _field) \
179 do { \
180 if ((_folio)->lru.prev != _base) { \
181 struct folio *prev; \
182 \
183 prev = lru_to_folio(&(_folio->lru)); \
184 prefetchw(&prev->_field); \
185 } \
186 } while (0)
187#else
188#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
189#endif
190
191/*
192 * From 0 .. 200. Higher means more swappy.
193 */
194int vm_swappiness = 60;
195
196#ifdef CONFIG_MEMCG
197
198/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
199static bool cgroup_reclaim(struct scan_control *sc)
200{
201 return sc->target_mem_cgroup;
202}
203
204/*
205 * Returns true for reclaim on the root cgroup. This is true for direct
206 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
207 */
208static bool root_reclaim(struct scan_control *sc)
209{
210 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
211}
212
213/**
214 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
215 * @sc: scan_control in question
216 *
217 * The normal page dirty throttling mechanism in balance_dirty_pages() is
218 * completely broken with the legacy memcg and direct stalling in
219 * shrink_folio_list() is used for throttling instead, which lacks all the
220 * niceties such as fairness, adaptive pausing, bandwidth proportional
221 * allocation and configurability.
222 *
223 * This function tests whether the vmscan currently in progress can assume
224 * that the normal dirty throttling mechanism is operational.
225 */
226static bool writeback_throttling_sane(struct scan_control *sc)
227{
228 if (!cgroup_reclaim(sc))
229 return true;
230#ifdef CONFIG_CGROUP_WRITEBACK
231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
232 return true;
233#endif
234 return false;
235}
236#else
237static bool cgroup_reclaim(struct scan_control *sc)
238{
239 return false;
240}
241
242static bool root_reclaim(struct scan_control *sc)
243{
244 return true;
245}
246
247static bool writeback_throttling_sane(struct scan_control *sc)
248{
249 return true;
250}
251#endif
252
253static void set_task_reclaim_state(struct task_struct *task,
254 struct reclaim_state *rs)
255{
256 /* Check for an overwrite */
257 WARN_ON_ONCE(rs && task->reclaim_state);
258
259 /* Check for the nulling of an already-nulled member */
260 WARN_ON_ONCE(!rs && !task->reclaim_state);
261
262 task->reclaim_state = rs;
263}
264
265/*
266 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
267 * scan_control->nr_reclaimed.
268 */
269static void flush_reclaim_state(struct scan_control *sc)
270{
271 /*
272 * Currently, reclaim_state->reclaimed includes three types of pages
273 * freed outside of vmscan:
274 * (1) Slab pages.
275 * (2) Clean file pages from pruned inodes (on highmem systems).
276 * (3) XFS freed buffer pages.
277 *
278 * For all of these cases, we cannot universally link the pages to a
279 * single memcg. For example, a memcg-aware shrinker can free one object
280 * charged to the target memcg, causing an entire page to be freed.
281 * If we count the entire page as reclaimed from the memcg, we end up
282 * overestimating the reclaimed amount (potentially under-reclaiming).
283 *
284 * Only count such pages for global reclaim to prevent under-reclaiming
285 * from the target memcg; preventing unnecessary retries during memcg
286 * charging and false positives from proactive reclaim.
287 *
288 * For uncommon cases where the freed pages were actually mostly
289 * charged to the target memcg, we end up underestimating the reclaimed
290 * amount. This should be fine. The freed pages will be uncharged
291 * anyway, even if they are not counted here properly, and we will be
292 * able to make forward progress in charging (which is usually in a
293 * retry loop).
294 *
295 * We can go one step further, and report the uncharged objcg pages in
296 * memcg reclaim, to make reporting more accurate and reduce
297 * underestimation, but it's probably not worth the complexity for now.
298 */
299 if (current->reclaim_state && root_reclaim(sc)) {
300 sc->nr_reclaimed += current->reclaim_state->reclaimed;
301 current->reclaim_state->reclaimed = 0;
302 }
303}
304
305static bool can_demote(int nid, struct scan_control *sc)
306{
307 if (!numa_demotion_enabled)
308 return false;
309 if (sc && sc->no_demotion)
310 return false;
311 if (next_demotion_node(nid) == NUMA_NO_NODE)
312 return false;
313
314 return true;
315}
316
317static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
318 int nid,
319 struct scan_control *sc)
320{
321 if (memcg == NULL) {
322 /*
323 * For non-memcg reclaim, is there
324 * space in any swap device?
325 */
326 if (get_nr_swap_pages() > 0)
327 return true;
328 } else {
329 /* Is the memcg below its swap limit? */
330 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
331 return true;
332 }
333
334 /*
335 * The page can not be swapped.
336 *
337 * Can it be reclaimed from this node via demotion?
338 */
339 return can_demote(nid, sc);
340}
341
342/*
343 * This misses isolated folios which are not accounted for to save counters.
344 * As the data only determines if reclaim or compaction continues, it is
345 * not expected that isolated folios will be a dominating factor.
346 */
347unsigned long zone_reclaimable_pages(struct zone *zone)
348{
349 unsigned long nr;
350
351 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
352 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
353 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
354 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
355 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
356
357 return nr;
358}
359
360/**
361 * lruvec_lru_size - Returns the number of pages on the given LRU list.
362 * @lruvec: lru vector
363 * @lru: lru to use
364 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
365 */
366static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
367 int zone_idx)
368{
369 unsigned long size = 0;
370 int zid;
371
372 for (zid = 0; zid <= zone_idx; zid++) {
373 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
374
375 if (!managed_zone(zone))
376 continue;
377
378 if (!mem_cgroup_disabled())
379 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
380 else
381 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
382 }
383 return size;
384}
385
386static unsigned long drop_slab_node(int nid)
387{
388 unsigned long freed = 0;
389 struct mem_cgroup *memcg = NULL;
390
391 memcg = mem_cgroup_iter(NULL, NULL, NULL);
392 do {
393 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
394 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
395
396 return freed;
397}
398
399void drop_slab(void)
400{
401 int nid;
402 int shift = 0;
403 unsigned long freed;
404
405 do {
406 freed = 0;
407 for_each_online_node(nid) {
408 if (fatal_signal_pending(current))
409 return;
410
411 freed += drop_slab_node(nid);
412 }
413 } while ((freed >> shift++) > 1);
414}
415
416static int reclaimer_offset(void)
417{
418 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
419 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
420 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
421 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
422 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
423 PGSCAN_DIRECT - PGSCAN_KSWAPD);
424 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
425 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
426
427 if (current_is_kswapd())
428 return 0;
429 if (current_is_khugepaged())
430 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
431 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
432}
433
434static inline int is_page_cache_freeable(struct folio *folio)
435{
436 /*
437 * A freeable page cache folio is referenced only by the caller
438 * that isolated the folio, the page cache and optional filesystem
439 * private data at folio->private.
440 */
441 return folio_ref_count(folio) - folio_test_private(folio) ==
442 1 + folio_nr_pages(folio);
443}
444
445/*
446 * We detected a synchronous write error writing a folio out. Probably
447 * -ENOSPC. We need to propagate that into the address_space for a subsequent
448 * fsync(), msync() or close().
449 *
450 * The tricky part is that after writepage we cannot touch the mapping: nothing
451 * prevents it from being freed up. But we have a ref on the folio and once
452 * that folio is locked, the mapping is pinned.
453 *
454 * We're allowed to run sleeping folio_lock() here because we know the caller has
455 * __GFP_FS.
456 */
457static void handle_write_error(struct address_space *mapping,
458 struct folio *folio, int error)
459{
460 folio_lock(folio);
461 if (folio_mapping(folio) == mapping)
462 mapping_set_error(mapping, error);
463 folio_unlock(folio);
464}
465
466static bool skip_throttle_noprogress(pg_data_t *pgdat)
467{
468 int reclaimable = 0, write_pending = 0;
469 int i;
470
471 /*
472 * If kswapd is disabled, reschedule if necessary but do not
473 * throttle as the system is likely near OOM.
474 */
475 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
476 return true;
477
478 /*
479 * If there are a lot of dirty/writeback folios then do not
480 * throttle as throttling will occur when the folios cycle
481 * towards the end of the LRU if still under writeback.
482 */
483 for (i = 0; i < MAX_NR_ZONES; i++) {
484 struct zone *zone = pgdat->node_zones + i;
485
486 if (!managed_zone(zone))
487 continue;
488
489 reclaimable += zone_reclaimable_pages(zone);
490 write_pending += zone_page_state_snapshot(zone,
491 NR_ZONE_WRITE_PENDING);
492 }
493 if (2 * write_pending <= reclaimable)
494 return true;
495
496 return false;
497}
498
499void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
500{
501 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
502 long timeout, ret;
503 DEFINE_WAIT(wait);
504
505 /*
506 * Do not throttle user workers, kthreads other than kswapd or
507 * workqueues. They may be required for reclaim to make
508 * forward progress (e.g. journalling workqueues or kthreads).
509 */
510 if (!current_is_kswapd() &&
511 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
512 cond_resched();
513 return;
514 }
515
516 /*
517 * These figures are pulled out of thin air.
518 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
519 * parallel reclaimers which is a short-lived event so the timeout is
520 * short. Failing to make progress or waiting on writeback are
521 * potentially long-lived events so use a longer timeout. This is shaky
522 * logic as a failure to make progress could be due to anything from
523 * writeback to a slow device to excessive referenced folios at the tail
524 * of the inactive LRU.
525 */
526 switch(reason) {
527 case VMSCAN_THROTTLE_WRITEBACK:
528 timeout = HZ/10;
529
530 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
531 WRITE_ONCE(pgdat->nr_reclaim_start,
532 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
533 }
534
535 break;
536 case VMSCAN_THROTTLE_CONGESTED:
537 fallthrough;
538 case VMSCAN_THROTTLE_NOPROGRESS:
539 if (skip_throttle_noprogress(pgdat)) {
540 cond_resched();
541 return;
542 }
543
544 timeout = 1;
545
546 break;
547 case VMSCAN_THROTTLE_ISOLATED:
548 timeout = HZ/50;
549 break;
550 default:
551 WARN_ON_ONCE(1);
552 timeout = HZ;
553 break;
554 }
555
556 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
557 ret = schedule_timeout(timeout);
558 finish_wait(wqh, &wait);
559
560 if (reason == VMSCAN_THROTTLE_WRITEBACK)
561 atomic_dec(&pgdat->nr_writeback_throttled);
562
563 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
564 jiffies_to_usecs(timeout - ret),
565 reason);
566}
567
568/*
569 * Account for folios written if tasks are throttled waiting on dirty
570 * folios to clean. If enough folios have been cleaned since throttling
571 * started then wakeup the throttled tasks.
572 */
573void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
574 int nr_throttled)
575{
576 unsigned long nr_written;
577
578 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
579
580 /*
581 * This is an inaccurate read as the per-cpu deltas may not
582 * be synchronised. However, given that the system is
583 * writeback throttled, it is not worth taking the penalty
584 * of getting an accurate count. At worst, the throttle
585 * timeout guarantees forward progress.
586 */
587 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
588 READ_ONCE(pgdat->nr_reclaim_start);
589
590 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
591 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
592}
593
594/* possible outcome of pageout() */
595typedef enum {
596 /* failed to write folio out, folio is locked */
597 PAGE_KEEP,
598 /* move folio to the active list, folio is locked */
599 PAGE_ACTIVATE,
600 /* folio has been sent to the disk successfully, folio is unlocked */
601 PAGE_SUCCESS,
602 /* folio is clean and locked */
603 PAGE_CLEAN,
604} pageout_t;
605
606/*
607 * pageout is called by shrink_folio_list() for each dirty folio.
608 * Calls ->writepage().
609 */
610static pageout_t pageout(struct folio *folio, struct address_space *mapping,
611 struct swap_iocb **plug)
612{
613 /*
614 * If the folio is dirty, only perform writeback if that write
615 * will be non-blocking. To prevent this allocation from being
616 * stalled by pagecache activity. But note that there may be
617 * stalls if we need to run get_block(). We could test
618 * PagePrivate for that.
619 *
620 * If this process is currently in __generic_file_write_iter() against
621 * this folio's queue, we can perform writeback even if that
622 * will block.
623 *
624 * If the folio is swapcache, write it back even if that would
625 * block, for some throttling. This happens by accident, because
626 * swap_backing_dev_info is bust: it doesn't reflect the
627 * congestion state of the swapdevs. Easy to fix, if needed.
628 */
629 if (!is_page_cache_freeable(folio))
630 return PAGE_KEEP;
631 if (!mapping) {
632 /*
633 * Some data journaling orphaned folios can have
634 * folio->mapping == NULL while being dirty with clean buffers.
635 */
636 if (folio_test_private(folio)) {
637 if (try_to_free_buffers(folio)) {
638 folio_clear_dirty(folio);
639 pr_info("%s: orphaned folio\n", __func__);
640 return PAGE_CLEAN;
641 }
642 }
643 return PAGE_KEEP;
644 }
645 if (mapping->a_ops->writepage == NULL)
646 return PAGE_ACTIVATE;
647
648 if (folio_clear_dirty_for_io(folio)) {
649 int res;
650 struct writeback_control wbc = {
651 .sync_mode = WB_SYNC_NONE,
652 .nr_to_write = SWAP_CLUSTER_MAX,
653 .range_start = 0,
654 .range_end = LLONG_MAX,
655 .for_reclaim = 1,
656 .swap_plug = plug,
657 };
658
659 folio_set_reclaim(folio);
660 res = mapping->a_ops->writepage(&folio->page, &wbc);
661 if (res < 0)
662 handle_write_error(mapping, folio, res);
663 if (res == AOP_WRITEPAGE_ACTIVATE) {
664 folio_clear_reclaim(folio);
665 return PAGE_ACTIVATE;
666 }
667
668 if (!folio_test_writeback(folio)) {
669 /* synchronous write or broken a_ops? */
670 folio_clear_reclaim(folio);
671 }
672 trace_mm_vmscan_write_folio(folio);
673 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
674 return PAGE_SUCCESS;
675 }
676
677 return PAGE_CLEAN;
678}
679
680/*
681 * Same as remove_mapping, but if the folio is removed from the mapping, it
682 * gets returned with a refcount of 0.
683 */
684static int __remove_mapping(struct address_space *mapping, struct folio *folio,
685 bool reclaimed, struct mem_cgroup *target_memcg)
686{
687 int refcount;
688 void *shadow = NULL;
689
690 BUG_ON(!folio_test_locked(folio));
691 BUG_ON(mapping != folio_mapping(folio));
692
693 if (!folio_test_swapcache(folio))
694 spin_lock(&mapping->host->i_lock);
695 xa_lock_irq(&mapping->i_pages);
696 /*
697 * The non racy check for a busy folio.
698 *
699 * Must be careful with the order of the tests. When someone has
700 * a ref to the folio, it may be possible that they dirty it then
701 * drop the reference. So if the dirty flag is tested before the
702 * refcount here, then the following race may occur:
703 *
704 * get_user_pages(&page);
705 * [user mapping goes away]
706 * write_to(page);
707 * !folio_test_dirty(folio) [good]
708 * folio_set_dirty(folio);
709 * folio_put(folio);
710 * !refcount(folio) [good, discard it]
711 *
712 * [oops, our write_to data is lost]
713 *
714 * Reversing the order of the tests ensures such a situation cannot
715 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
716 * load is not satisfied before that of folio->_refcount.
717 *
718 * Note that if the dirty flag is always set via folio_mark_dirty,
719 * and thus under the i_pages lock, then this ordering is not required.
720 */
721 refcount = 1 + folio_nr_pages(folio);
722 if (!folio_ref_freeze(folio, refcount))
723 goto cannot_free;
724 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
725 if (unlikely(folio_test_dirty(folio))) {
726 folio_ref_unfreeze(folio, refcount);
727 goto cannot_free;
728 }
729
730 if (folio_test_swapcache(folio)) {
731 swp_entry_t swap = folio->swap;
732
733 if (reclaimed && !mapping_exiting(mapping))
734 shadow = workingset_eviction(folio, target_memcg);
735 __delete_from_swap_cache(folio, swap, shadow);
736 mem_cgroup_swapout(folio, swap);
737 xa_unlock_irq(&mapping->i_pages);
738 put_swap_folio(folio, swap);
739 } else {
740 void (*free_folio)(struct folio *);
741
742 free_folio = mapping->a_ops->free_folio;
743 /*
744 * Remember a shadow entry for reclaimed file cache in
745 * order to detect refaults, thus thrashing, later on.
746 *
747 * But don't store shadows in an address space that is
748 * already exiting. This is not just an optimization,
749 * inode reclaim needs to empty out the radix tree or
750 * the nodes are lost. Don't plant shadows behind its
751 * back.
752 *
753 * We also don't store shadows for DAX mappings because the
754 * only page cache folios found in these are zero pages
755 * covering holes, and because we don't want to mix DAX
756 * exceptional entries and shadow exceptional entries in the
757 * same address_space.
758 */
759 if (reclaimed && folio_is_file_lru(folio) &&
760 !mapping_exiting(mapping) && !dax_mapping(mapping))
761 shadow = workingset_eviction(folio, target_memcg);
762 __filemap_remove_folio(folio, shadow);
763 xa_unlock_irq(&mapping->i_pages);
764 if (mapping_shrinkable(mapping))
765 inode_add_lru(mapping->host);
766 spin_unlock(&mapping->host->i_lock);
767
768 if (free_folio)
769 free_folio(folio);
770 }
771
772 return 1;
773
774cannot_free:
775 xa_unlock_irq(&mapping->i_pages);
776 if (!folio_test_swapcache(folio))
777 spin_unlock(&mapping->host->i_lock);
778 return 0;
779}
780
781/**
782 * remove_mapping() - Attempt to remove a folio from its mapping.
783 * @mapping: The address space.
784 * @folio: The folio to remove.
785 *
786 * If the folio is dirty, under writeback or if someone else has a ref
787 * on it, removal will fail.
788 * Return: The number of pages removed from the mapping. 0 if the folio
789 * could not be removed.
790 * Context: The caller should have a single refcount on the folio and
791 * hold its lock.
792 */
793long remove_mapping(struct address_space *mapping, struct folio *folio)
794{
795 if (__remove_mapping(mapping, folio, false, NULL)) {
796 /*
797 * Unfreezing the refcount with 1 effectively
798 * drops the pagecache ref for us without requiring another
799 * atomic operation.
800 */
801 folio_ref_unfreeze(folio, 1);
802 return folio_nr_pages(folio);
803 }
804 return 0;
805}
806
807/**
808 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
809 * @folio: Folio to be returned to an LRU list.
810 *
811 * Add previously isolated @folio to appropriate LRU list.
812 * The folio may still be unevictable for other reasons.
813 *
814 * Context: lru_lock must not be held, interrupts must be enabled.
815 */
816void folio_putback_lru(struct folio *folio)
817{
818 folio_add_lru(folio);
819 folio_put(folio); /* drop ref from isolate */
820}
821
822enum folio_references {
823 FOLIOREF_RECLAIM,
824 FOLIOREF_RECLAIM_CLEAN,
825 FOLIOREF_KEEP,
826 FOLIOREF_ACTIVATE,
827};
828
829static enum folio_references folio_check_references(struct folio *folio,
830 struct scan_control *sc)
831{
832 int referenced_ptes, referenced_folio;
833 unsigned long vm_flags;
834
835 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
836 &vm_flags);
837 referenced_folio = folio_test_clear_referenced(folio);
838
839 /*
840 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
841 * Let the folio, now marked Mlocked, be moved to the unevictable list.
842 */
843 if (vm_flags & VM_LOCKED)
844 return FOLIOREF_ACTIVATE;
845
846 /* rmap lock contention: rotate */
847 if (referenced_ptes == -1)
848 return FOLIOREF_KEEP;
849
850 if (referenced_ptes) {
851 /*
852 * All mapped folios start out with page table
853 * references from the instantiating fault, so we need
854 * to look twice if a mapped file/anon folio is used more
855 * than once.
856 *
857 * Mark it and spare it for another trip around the
858 * inactive list. Another page table reference will
859 * lead to its activation.
860 *
861 * Note: the mark is set for activated folios as well
862 * so that recently deactivated but used folios are
863 * quickly recovered.
864 */
865 folio_set_referenced(folio);
866
867 if (referenced_folio || referenced_ptes > 1)
868 return FOLIOREF_ACTIVATE;
869
870 /*
871 * Activate file-backed executable folios after first usage.
872 */
873 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
874 return FOLIOREF_ACTIVATE;
875
876 return FOLIOREF_KEEP;
877 }
878
879 /* Reclaim if clean, defer dirty folios to writeback */
880 if (referenced_folio && folio_is_file_lru(folio))
881 return FOLIOREF_RECLAIM_CLEAN;
882
883 return FOLIOREF_RECLAIM;
884}
885
886/* Check if a folio is dirty or under writeback */
887static void folio_check_dirty_writeback(struct folio *folio,
888 bool *dirty, bool *writeback)
889{
890 struct address_space *mapping;
891
892 /*
893 * Anonymous folios are not handled by flushers and must be written
894 * from reclaim context. Do not stall reclaim based on them.
895 * MADV_FREE anonymous folios are put into inactive file list too.
896 * They could be mistakenly treated as file lru. So further anon
897 * test is needed.
898 */
899 if (!folio_is_file_lru(folio) ||
900 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
901 *dirty = false;
902 *writeback = false;
903 return;
904 }
905
906 /* By default assume that the folio flags are accurate */
907 *dirty = folio_test_dirty(folio);
908 *writeback = folio_test_writeback(folio);
909
910 /* Verify dirty/writeback state if the filesystem supports it */
911 if (!folio_test_private(folio))
912 return;
913
914 mapping = folio_mapping(folio);
915 if (mapping && mapping->a_ops->is_dirty_writeback)
916 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
917}
918
919static struct folio *alloc_demote_folio(struct folio *src,
920 unsigned long private)
921{
922 struct folio *dst;
923 nodemask_t *allowed_mask;
924 struct migration_target_control *mtc;
925
926 mtc = (struct migration_target_control *)private;
927
928 allowed_mask = mtc->nmask;
929 /*
930 * make sure we allocate from the target node first also trying to
931 * demote or reclaim pages from the target node via kswapd if we are
932 * low on free memory on target node. If we don't do this and if
933 * we have free memory on the slower(lower) memtier, we would start
934 * allocating pages from slower(lower) memory tiers without even forcing
935 * a demotion of cold pages from the target memtier. This can result
936 * in the kernel placing hot pages in slower(lower) memory tiers.
937 */
938 mtc->nmask = NULL;
939 mtc->gfp_mask |= __GFP_THISNODE;
940 dst = alloc_migration_target(src, (unsigned long)mtc);
941 if (dst)
942 return dst;
943
944 mtc->gfp_mask &= ~__GFP_THISNODE;
945 mtc->nmask = allowed_mask;
946
947 return alloc_migration_target(src, (unsigned long)mtc);
948}
949
950/*
951 * Take folios on @demote_folios and attempt to demote them to another node.
952 * Folios which are not demoted are left on @demote_folios.
953 */
954static unsigned int demote_folio_list(struct list_head *demote_folios,
955 struct pglist_data *pgdat)
956{
957 int target_nid = next_demotion_node(pgdat->node_id);
958 unsigned int nr_succeeded;
959 nodemask_t allowed_mask;
960
961 struct migration_target_control mtc = {
962 /*
963 * Allocate from 'node', or fail quickly and quietly.
964 * When this happens, 'page' will likely just be discarded
965 * instead of migrated.
966 */
967 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
968 __GFP_NOMEMALLOC | GFP_NOWAIT,
969 .nid = target_nid,
970 .nmask = &allowed_mask
971 };
972
973 if (list_empty(demote_folios))
974 return 0;
975
976 if (target_nid == NUMA_NO_NODE)
977 return 0;
978
979 node_get_allowed_targets(pgdat, &allowed_mask);
980
981 /* Demotion ignores all cpuset and mempolicy settings */
982 migrate_pages(demote_folios, alloc_demote_folio, NULL,
983 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
984 &nr_succeeded);
985
986 mod_node_page_state(pgdat, PGDEMOTE_KSWAPD + reclaimer_offset(),
987 nr_succeeded);
988
989 return nr_succeeded;
990}
991
992static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
993{
994 if (gfp_mask & __GFP_FS)
995 return true;
996 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
997 return false;
998 /*
999 * We can "enter_fs" for swap-cache with only __GFP_IO
1000 * providing this isn't SWP_FS_OPS.
1001 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1002 * but that will never affect SWP_FS_OPS, so the data_race
1003 * is safe.
1004 */
1005 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1006}
1007
1008/*
1009 * shrink_folio_list() returns the number of reclaimed pages
1010 */
1011static unsigned int shrink_folio_list(struct list_head *folio_list,
1012 struct pglist_data *pgdat, struct scan_control *sc,
1013 struct reclaim_stat *stat, bool ignore_references)
1014{
1015 struct folio_batch free_folios;
1016 LIST_HEAD(ret_folios);
1017 LIST_HEAD(demote_folios);
1018 unsigned int nr_reclaimed = 0;
1019 unsigned int pgactivate = 0;
1020 bool do_demote_pass;
1021 struct swap_iocb *plug = NULL;
1022
1023 folio_batch_init(&free_folios);
1024 memset(stat, 0, sizeof(*stat));
1025 cond_resched();
1026 do_demote_pass = can_demote(pgdat->node_id, sc);
1027
1028retry:
1029 while (!list_empty(folio_list)) {
1030 struct address_space *mapping;
1031 struct folio *folio;
1032 enum folio_references references = FOLIOREF_RECLAIM;
1033 bool dirty, writeback;
1034 unsigned int nr_pages;
1035
1036 cond_resched();
1037
1038 folio = lru_to_folio(folio_list);
1039 list_del(&folio->lru);
1040
1041 if (!folio_trylock(folio))
1042 goto keep;
1043
1044 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1045
1046 nr_pages = folio_nr_pages(folio);
1047
1048 /* Account the number of base pages */
1049 sc->nr_scanned += nr_pages;
1050
1051 if (unlikely(!folio_evictable(folio)))
1052 goto activate_locked;
1053
1054 if (!sc->may_unmap && folio_mapped(folio))
1055 goto keep_locked;
1056
1057 /* folio_update_gen() tried to promote this page? */
1058 if (lru_gen_enabled() && !ignore_references &&
1059 folio_mapped(folio) && folio_test_referenced(folio))
1060 goto keep_locked;
1061
1062 /*
1063 * The number of dirty pages determines if a node is marked
1064 * reclaim_congested. kswapd will stall and start writing
1065 * folios if the tail of the LRU is all dirty unqueued folios.
1066 */
1067 folio_check_dirty_writeback(folio, &dirty, &writeback);
1068 if (dirty || writeback)
1069 stat->nr_dirty += nr_pages;
1070
1071 if (dirty && !writeback)
1072 stat->nr_unqueued_dirty += nr_pages;
1073
1074 /*
1075 * Treat this folio as congested if folios are cycling
1076 * through the LRU so quickly that the folios marked
1077 * for immediate reclaim are making it to the end of
1078 * the LRU a second time.
1079 */
1080 if (writeback && folio_test_reclaim(folio))
1081 stat->nr_congested += nr_pages;
1082
1083 /*
1084 * If a folio at the tail of the LRU is under writeback, there
1085 * are three cases to consider.
1086 *
1087 * 1) If reclaim is encountering an excessive number
1088 * of folios under writeback and this folio has both
1089 * the writeback and reclaim flags set, then it
1090 * indicates that folios are being queued for I/O but
1091 * are being recycled through the LRU before the I/O
1092 * can complete. Waiting on the folio itself risks an
1093 * indefinite stall if it is impossible to writeback
1094 * the folio due to I/O error or disconnected storage
1095 * so instead note that the LRU is being scanned too
1096 * quickly and the caller can stall after the folio
1097 * list has been processed.
1098 *
1099 * 2) Global or new memcg reclaim encounters a folio that is
1100 * not marked for immediate reclaim, or the caller does not
1101 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1102 * not to fs). In this case mark the folio for immediate
1103 * reclaim and continue scanning.
1104 *
1105 * Require may_enter_fs() because we would wait on fs, which
1106 * may not have submitted I/O yet. And the loop driver might
1107 * enter reclaim, and deadlock if it waits on a folio for
1108 * which it is needed to do the write (loop masks off
1109 * __GFP_IO|__GFP_FS for this reason); but more thought
1110 * would probably show more reasons.
1111 *
1112 * 3) Legacy memcg encounters a folio that already has the
1113 * reclaim flag set. memcg does not have any dirty folio
1114 * throttling so we could easily OOM just because too many
1115 * folios are in writeback and there is nothing else to
1116 * reclaim. Wait for the writeback to complete.
1117 *
1118 * In cases 1) and 2) we activate the folios to get them out of
1119 * the way while we continue scanning for clean folios on the
1120 * inactive list and refilling from the active list. The
1121 * observation here is that waiting for disk writes is more
1122 * expensive than potentially causing reloads down the line.
1123 * Since they're marked for immediate reclaim, they won't put
1124 * memory pressure on the cache working set any longer than it
1125 * takes to write them to disk.
1126 */
1127 if (folio_test_writeback(folio)) {
1128 /* Case 1 above */
1129 if (current_is_kswapd() &&
1130 folio_test_reclaim(folio) &&
1131 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1132 stat->nr_immediate += nr_pages;
1133 goto activate_locked;
1134
1135 /* Case 2 above */
1136 } else if (writeback_throttling_sane(sc) ||
1137 !folio_test_reclaim(folio) ||
1138 !may_enter_fs(folio, sc->gfp_mask)) {
1139 /*
1140 * This is slightly racy -
1141 * folio_end_writeback() might have
1142 * just cleared the reclaim flag, then
1143 * setting the reclaim flag here ends up
1144 * interpreted as the readahead flag - but
1145 * that does not matter enough to care.
1146 * What we do want is for this folio to
1147 * have the reclaim flag set next time
1148 * memcg reclaim reaches the tests above,
1149 * so it will then wait for writeback to
1150 * avoid OOM; and it's also appropriate
1151 * in global reclaim.
1152 */
1153 folio_set_reclaim(folio);
1154 stat->nr_writeback += nr_pages;
1155 goto activate_locked;
1156
1157 /* Case 3 above */
1158 } else {
1159 folio_unlock(folio);
1160 folio_wait_writeback(folio);
1161 /* then go back and try same folio again */
1162 list_add_tail(&folio->lru, folio_list);
1163 continue;
1164 }
1165 }
1166
1167 if (!ignore_references)
1168 references = folio_check_references(folio, sc);
1169
1170 switch (references) {
1171 case FOLIOREF_ACTIVATE:
1172 goto activate_locked;
1173 case FOLIOREF_KEEP:
1174 stat->nr_ref_keep += nr_pages;
1175 goto keep_locked;
1176 case FOLIOREF_RECLAIM:
1177 case FOLIOREF_RECLAIM_CLEAN:
1178 ; /* try to reclaim the folio below */
1179 }
1180
1181 /*
1182 * Before reclaiming the folio, try to relocate
1183 * its contents to another node.
1184 */
1185 if (do_demote_pass &&
1186 (thp_migration_supported() || !folio_test_large(folio))) {
1187 list_add(&folio->lru, &demote_folios);
1188 folio_unlock(folio);
1189 continue;
1190 }
1191
1192 /*
1193 * Anonymous process memory has backing store?
1194 * Try to allocate it some swap space here.
1195 * Lazyfree folio could be freed directly
1196 */
1197 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1198 if (!folio_test_swapcache(folio)) {
1199 if (!(sc->gfp_mask & __GFP_IO))
1200 goto keep_locked;
1201 if (folio_maybe_dma_pinned(folio))
1202 goto keep_locked;
1203 if (folio_test_large(folio)) {
1204 /* cannot split folio, skip it */
1205 if (!can_split_folio(folio, NULL))
1206 goto activate_locked;
1207 /*
1208 * Split folios without a PMD map right
1209 * away. Chances are some or all of the
1210 * tail pages can be freed without IO.
1211 */
1212 if (!folio_entire_mapcount(folio) &&
1213 split_folio_to_list(folio,
1214 folio_list))
1215 goto activate_locked;
1216 }
1217 if (!add_to_swap(folio)) {
1218 if (!folio_test_large(folio))
1219 goto activate_locked_split;
1220 /* Fallback to swap normal pages */
1221 if (split_folio_to_list(folio,
1222 folio_list))
1223 goto activate_locked;
1224#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1225 count_memcg_folio_events(folio, THP_SWPOUT_FALLBACK, 1);
1226 count_vm_event(THP_SWPOUT_FALLBACK);
1227#endif
1228 if (!add_to_swap(folio))
1229 goto activate_locked_split;
1230 }
1231 }
1232 } else if (folio_test_swapbacked(folio) &&
1233 folio_test_large(folio)) {
1234 /* Split shmem folio */
1235 if (split_folio_to_list(folio, folio_list))
1236 goto keep_locked;
1237 }
1238
1239 /*
1240 * If the folio was split above, the tail pages will make
1241 * their own pass through this function and be accounted
1242 * then.
1243 */
1244 if ((nr_pages > 1) && !folio_test_large(folio)) {
1245 sc->nr_scanned -= (nr_pages - 1);
1246 nr_pages = 1;
1247 }
1248
1249 /*
1250 * The folio is mapped into the page tables of one or more
1251 * processes. Try to unmap it here.
1252 */
1253 if (folio_mapped(folio)) {
1254 enum ttu_flags flags = TTU_BATCH_FLUSH;
1255 bool was_swapbacked = folio_test_swapbacked(folio);
1256
1257 if (folio_test_pmd_mappable(folio))
1258 flags |= TTU_SPLIT_HUGE_PMD;
1259
1260 try_to_unmap(folio, flags);
1261 if (folio_mapped(folio)) {
1262 stat->nr_unmap_fail += nr_pages;
1263 if (!was_swapbacked &&
1264 folio_test_swapbacked(folio))
1265 stat->nr_lazyfree_fail += nr_pages;
1266 goto activate_locked;
1267 }
1268 }
1269
1270 /*
1271 * Folio is unmapped now so it cannot be newly pinned anymore.
1272 * No point in trying to reclaim folio if it is pinned.
1273 * Furthermore we don't want to reclaim underlying fs metadata
1274 * if the folio is pinned and thus potentially modified by the
1275 * pinning process as that may upset the filesystem.
1276 */
1277 if (folio_maybe_dma_pinned(folio))
1278 goto activate_locked;
1279
1280 mapping = folio_mapping(folio);
1281 if (folio_test_dirty(folio)) {
1282 /*
1283 * Only kswapd can writeback filesystem folios
1284 * to avoid risk of stack overflow. But avoid
1285 * injecting inefficient single-folio I/O into
1286 * flusher writeback as much as possible: only
1287 * write folios when we've encountered many
1288 * dirty folios, and when we've already scanned
1289 * the rest of the LRU for clean folios and see
1290 * the same dirty folios again (with the reclaim
1291 * flag set).
1292 */
1293 if (folio_is_file_lru(folio) &&
1294 (!current_is_kswapd() ||
1295 !folio_test_reclaim(folio) ||
1296 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1297 /*
1298 * Immediately reclaim when written back.
1299 * Similar in principle to folio_deactivate()
1300 * except we already have the folio isolated
1301 * and know it's dirty
1302 */
1303 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1304 nr_pages);
1305 folio_set_reclaim(folio);
1306
1307 goto activate_locked;
1308 }
1309
1310 if (references == FOLIOREF_RECLAIM_CLEAN)
1311 goto keep_locked;
1312 if (!may_enter_fs(folio, sc->gfp_mask))
1313 goto keep_locked;
1314 if (!sc->may_writepage)
1315 goto keep_locked;
1316
1317 /*
1318 * Folio is dirty. Flush the TLB if a writable entry
1319 * potentially exists to avoid CPU writes after I/O
1320 * starts and then write it out here.
1321 */
1322 try_to_unmap_flush_dirty();
1323 switch (pageout(folio, mapping, &plug)) {
1324 case PAGE_KEEP:
1325 goto keep_locked;
1326 case PAGE_ACTIVATE:
1327 goto activate_locked;
1328 case PAGE_SUCCESS:
1329 stat->nr_pageout += nr_pages;
1330
1331 if (folio_test_writeback(folio))
1332 goto keep;
1333 if (folio_test_dirty(folio))
1334 goto keep;
1335
1336 /*
1337 * A synchronous write - probably a ramdisk. Go
1338 * ahead and try to reclaim the folio.
1339 */
1340 if (!folio_trylock(folio))
1341 goto keep;
1342 if (folio_test_dirty(folio) ||
1343 folio_test_writeback(folio))
1344 goto keep_locked;
1345 mapping = folio_mapping(folio);
1346 fallthrough;
1347 case PAGE_CLEAN:
1348 ; /* try to free the folio below */
1349 }
1350 }
1351
1352 /*
1353 * If the folio has buffers, try to free the buffer
1354 * mappings associated with this folio. If we succeed
1355 * we try to free the folio as well.
1356 *
1357 * We do this even if the folio is dirty.
1358 * filemap_release_folio() does not perform I/O, but it
1359 * is possible for a folio to have the dirty flag set,
1360 * but it is actually clean (all its buffers are clean).
1361 * This happens if the buffers were written out directly,
1362 * with submit_bh(). ext3 will do this, as well as
1363 * the blockdev mapping. filemap_release_folio() will
1364 * discover that cleanness and will drop the buffers
1365 * and mark the folio clean - it can be freed.
1366 *
1367 * Rarely, folios can have buffers and no ->mapping.
1368 * These are the folios which were not successfully
1369 * invalidated in truncate_cleanup_folio(). We try to
1370 * drop those buffers here and if that worked, and the
1371 * folio is no longer mapped into process address space
1372 * (refcount == 1) it can be freed. Otherwise, leave
1373 * the folio on the LRU so it is swappable.
1374 */
1375 if (folio_needs_release(folio)) {
1376 if (!filemap_release_folio(folio, sc->gfp_mask))
1377 goto activate_locked;
1378 if (!mapping && folio_ref_count(folio) == 1) {
1379 folio_unlock(folio);
1380 if (folio_put_testzero(folio))
1381 goto free_it;
1382 else {
1383 /*
1384 * rare race with speculative reference.
1385 * the speculative reference will free
1386 * this folio shortly, so we may
1387 * increment nr_reclaimed here (and
1388 * leave it off the LRU).
1389 */
1390 nr_reclaimed += nr_pages;
1391 continue;
1392 }
1393 }
1394 }
1395
1396 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1397 /* follow __remove_mapping for reference */
1398 if (!folio_ref_freeze(folio, 1))
1399 goto keep_locked;
1400 /*
1401 * The folio has only one reference left, which is
1402 * from the isolation. After the caller puts the
1403 * folio back on the lru and drops the reference, the
1404 * folio will be freed anyway. It doesn't matter
1405 * which lru it goes on. So we don't bother checking
1406 * the dirty flag here.
1407 */
1408 count_vm_events(PGLAZYFREED, nr_pages);
1409 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1410 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1411 sc->target_mem_cgroup))
1412 goto keep_locked;
1413
1414 folio_unlock(folio);
1415free_it:
1416 /*
1417 * Folio may get swapped out as a whole, need to account
1418 * all pages in it.
1419 */
1420 nr_reclaimed += nr_pages;
1421
1422 if (folio_test_large(folio) &&
1423 folio_test_large_rmappable(folio))
1424 folio_undo_large_rmappable(folio);
1425 if (folio_batch_add(&free_folios, folio) == 0) {
1426 mem_cgroup_uncharge_folios(&free_folios);
1427 try_to_unmap_flush();
1428 free_unref_folios(&free_folios);
1429 }
1430 continue;
1431
1432activate_locked_split:
1433 /*
1434 * The tail pages that are failed to add into swap cache
1435 * reach here. Fixup nr_scanned and nr_pages.
1436 */
1437 if (nr_pages > 1) {
1438 sc->nr_scanned -= (nr_pages - 1);
1439 nr_pages = 1;
1440 }
1441activate_locked:
1442 /* Not a candidate for swapping, so reclaim swap space. */
1443 if (folio_test_swapcache(folio) &&
1444 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1445 folio_free_swap(folio);
1446 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1447 if (!folio_test_mlocked(folio)) {
1448 int type = folio_is_file_lru(folio);
1449 folio_set_active(folio);
1450 stat->nr_activate[type] += nr_pages;
1451 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1452 }
1453keep_locked:
1454 folio_unlock(folio);
1455keep:
1456 list_add(&folio->lru, &ret_folios);
1457 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1458 folio_test_unevictable(folio), folio);
1459 }
1460 /* 'folio_list' is always empty here */
1461
1462 /* Migrate folios selected for demotion */
1463 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
1464 /* Folios that could not be demoted are still in @demote_folios */
1465 if (!list_empty(&demote_folios)) {
1466 /* Folios which weren't demoted go back on @folio_list */
1467 list_splice_init(&demote_folios, folio_list);
1468
1469 /*
1470 * goto retry to reclaim the undemoted folios in folio_list if
1471 * desired.
1472 *
1473 * Reclaiming directly from top tier nodes is not often desired
1474 * due to it breaking the LRU ordering: in general memory
1475 * should be reclaimed from lower tier nodes and demoted from
1476 * top tier nodes.
1477 *
1478 * However, disabling reclaim from top tier nodes entirely
1479 * would cause ooms in edge scenarios where lower tier memory
1480 * is unreclaimable for whatever reason, eg memory being
1481 * mlocked or too hot to reclaim. We can disable reclaim
1482 * from top tier nodes in proactive reclaim though as that is
1483 * not real memory pressure.
1484 */
1485 if (!sc->proactive) {
1486 do_demote_pass = false;
1487 goto retry;
1488 }
1489 }
1490
1491 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1492
1493 mem_cgroup_uncharge_folios(&free_folios);
1494 try_to_unmap_flush();
1495 free_unref_folios(&free_folios);
1496
1497 list_splice(&ret_folios, folio_list);
1498 count_vm_events(PGACTIVATE, pgactivate);
1499
1500 if (plug)
1501 swap_write_unplug(plug);
1502 return nr_reclaimed;
1503}
1504
1505unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1506 struct list_head *folio_list)
1507{
1508 struct scan_control sc = {
1509 .gfp_mask = GFP_KERNEL,
1510 .may_unmap = 1,
1511 };
1512 struct reclaim_stat stat;
1513 unsigned int nr_reclaimed;
1514 struct folio *folio, *next;
1515 LIST_HEAD(clean_folios);
1516 unsigned int noreclaim_flag;
1517
1518 list_for_each_entry_safe(folio, next, folio_list, lru) {
1519 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1520 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1521 !folio_test_unevictable(folio)) {
1522 folio_clear_active(folio);
1523 list_move(&folio->lru, &clean_folios);
1524 }
1525 }
1526
1527 /*
1528 * We should be safe here since we are only dealing with file pages and
1529 * we are not kswapd and therefore cannot write dirty file pages. But
1530 * call memalloc_noreclaim_save() anyway, just in case these conditions
1531 * change in the future.
1532 */
1533 noreclaim_flag = memalloc_noreclaim_save();
1534 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1535 &stat, true);
1536 memalloc_noreclaim_restore(noreclaim_flag);
1537
1538 list_splice(&clean_folios, folio_list);
1539 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1540 -(long)nr_reclaimed);
1541 /*
1542 * Since lazyfree pages are isolated from file LRU from the beginning,
1543 * they will rotate back to anonymous LRU in the end if it failed to
1544 * discard so isolated count will be mismatched.
1545 * Compensate the isolated count for both LRU lists.
1546 */
1547 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1548 stat.nr_lazyfree_fail);
1549 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1550 -(long)stat.nr_lazyfree_fail);
1551 return nr_reclaimed;
1552}
1553
1554/*
1555 * Update LRU sizes after isolating pages. The LRU size updates must
1556 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1557 */
1558static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1559 enum lru_list lru, unsigned long *nr_zone_taken)
1560{
1561 int zid;
1562
1563 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1564 if (!nr_zone_taken[zid])
1565 continue;
1566
1567 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1568 }
1569
1570}
1571
1572#ifdef CONFIG_CMA
1573/*
1574 * It is waste of effort to scan and reclaim CMA pages if it is not available
1575 * for current allocation context. Kswapd can not be enrolled as it can not
1576 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1577 */
1578static bool skip_cma(struct folio *folio, struct scan_control *sc)
1579{
1580 return !current_is_kswapd() &&
1581 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1582 folio_migratetype(folio) == MIGRATE_CMA;
1583}
1584#else
1585static bool skip_cma(struct folio *folio, struct scan_control *sc)
1586{
1587 return false;
1588}
1589#endif
1590
1591/*
1592 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1593 *
1594 * lruvec->lru_lock is heavily contended. Some of the functions that
1595 * shrink the lists perform better by taking out a batch of pages
1596 * and working on them outside the LRU lock.
1597 *
1598 * For pagecache intensive workloads, this function is the hottest
1599 * spot in the kernel (apart from copy_*_user functions).
1600 *
1601 * Lru_lock must be held before calling this function.
1602 *
1603 * @nr_to_scan: The number of eligible pages to look through on the list.
1604 * @lruvec: The LRU vector to pull pages from.
1605 * @dst: The temp list to put pages on to.
1606 * @nr_scanned: The number of pages that were scanned.
1607 * @sc: The scan_control struct for this reclaim session
1608 * @lru: LRU list id for isolating
1609 *
1610 * returns how many pages were moved onto *@dst.
1611 */
1612static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1613 struct lruvec *lruvec, struct list_head *dst,
1614 unsigned long *nr_scanned, struct scan_control *sc,
1615 enum lru_list lru)
1616{
1617 struct list_head *src = &lruvec->lists[lru];
1618 unsigned long nr_taken = 0;
1619 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1620 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1621 unsigned long skipped = 0;
1622 unsigned long scan, total_scan, nr_pages;
1623 LIST_HEAD(folios_skipped);
1624
1625 total_scan = 0;
1626 scan = 0;
1627 while (scan < nr_to_scan && !list_empty(src)) {
1628 struct list_head *move_to = src;
1629 struct folio *folio;
1630
1631 folio = lru_to_folio(src);
1632 prefetchw_prev_lru_folio(folio, src, flags);
1633
1634 nr_pages = folio_nr_pages(folio);
1635 total_scan += nr_pages;
1636
1637 if (folio_zonenum(folio) > sc->reclaim_idx ||
1638 skip_cma(folio, sc)) {
1639 nr_skipped[folio_zonenum(folio)] += nr_pages;
1640 move_to = &folios_skipped;
1641 goto move;
1642 }
1643
1644 /*
1645 * Do not count skipped folios because that makes the function
1646 * return with no isolated folios if the LRU mostly contains
1647 * ineligible folios. This causes the VM to not reclaim any
1648 * folios, triggering a premature OOM.
1649 * Account all pages in a folio.
1650 */
1651 scan += nr_pages;
1652
1653 if (!folio_test_lru(folio))
1654 goto move;
1655 if (!sc->may_unmap && folio_mapped(folio))
1656 goto move;
1657
1658 /*
1659 * Be careful not to clear the lru flag until after we're
1660 * sure the folio is not being freed elsewhere -- the
1661 * folio release code relies on it.
1662 */
1663 if (unlikely(!folio_try_get(folio)))
1664 goto move;
1665
1666 if (!folio_test_clear_lru(folio)) {
1667 /* Another thread is already isolating this folio */
1668 folio_put(folio);
1669 goto move;
1670 }
1671
1672 nr_taken += nr_pages;
1673 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1674 move_to = dst;
1675move:
1676 list_move(&folio->lru, move_to);
1677 }
1678
1679 /*
1680 * Splice any skipped folios to the start of the LRU list. Note that
1681 * this disrupts the LRU order when reclaiming for lower zones but
1682 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1683 * scanning would soon rescan the same folios to skip and waste lots
1684 * of cpu cycles.
1685 */
1686 if (!list_empty(&folios_skipped)) {
1687 int zid;
1688
1689 list_splice(&folios_skipped, src);
1690 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1691 if (!nr_skipped[zid])
1692 continue;
1693
1694 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1695 skipped += nr_skipped[zid];
1696 }
1697 }
1698 *nr_scanned = total_scan;
1699 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1700 total_scan, skipped, nr_taken, lru);
1701 update_lru_sizes(lruvec, lru, nr_zone_taken);
1702 return nr_taken;
1703}
1704
1705/**
1706 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1707 * @folio: Folio to isolate from its LRU list.
1708 *
1709 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1710 * corresponding to whatever LRU list the folio was on.
1711 *
1712 * The folio will have its LRU flag cleared. If it was found on the
1713 * active list, it will have the Active flag set. If it was found on the
1714 * unevictable list, it will have the Unevictable flag set. These flags
1715 * may need to be cleared by the caller before letting the page go.
1716 *
1717 * Context:
1718 *
1719 * (1) Must be called with an elevated refcount on the folio. This is a
1720 * fundamental difference from isolate_lru_folios() (which is called
1721 * without a stable reference).
1722 * (2) The lru_lock must not be held.
1723 * (3) Interrupts must be enabled.
1724 *
1725 * Return: true if the folio was removed from an LRU list.
1726 * false if the folio was not on an LRU list.
1727 */
1728bool folio_isolate_lru(struct folio *folio)
1729{
1730 bool ret = false;
1731
1732 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1733
1734 if (folio_test_clear_lru(folio)) {
1735 struct lruvec *lruvec;
1736
1737 folio_get(folio);
1738 lruvec = folio_lruvec_lock_irq(folio);
1739 lruvec_del_folio(lruvec, folio);
1740 unlock_page_lruvec_irq(lruvec);
1741 ret = true;
1742 }
1743
1744 return ret;
1745}
1746
1747/*
1748 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1749 * then get rescheduled. When there are massive number of tasks doing page
1750 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1751 * the LRU list will go small and be scanned faster than necessary, leading to
1752 * unnecessary swapping, thrashing and OOM.
1753 */
1754static bool too_many_isolated(struct pglist_data *pgdat, int file,
1755 struct scan_control *sc)
1756{
1757 unsigned long inactive, isolated;
1758 bool too_many;
1759
1760 if (current_is_kswapd())
1761 return false;
1762
1763 if (!writeback_throttling_sane(sc))
1764 return false;
1765
1766 if (file) {
1767 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1768 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1769 } else {
1770 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1771 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1772 }
1773
1774 /*
1775 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1776 * won't get blocked by normal direct-reclaimers, forming a circular
1777 * deadlock.
1778 */
1779 if (gfp_has_io_fs(sc->gfp_mask))
1780 inactive >>= 3;
1781
1782 too_many = isolated > inactive;
1783
1784 /* Wake up tasks throttled due to too_many_isolated. */
1785 if (!too_many)
1786 wake_throttle_isolated(pgdat);
1787
1788 return too_many;
1789}
1790
1791/*
1792 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1793 *
1794 * Returns the number of pages moved to the given lruvec.
1795 */
1796static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1797 struct list_head *list)
1798{
1799 int nr_pages, nr_moved = 0;
1800 struct folio_batch free_folios;
1801
1802 folio_batch_init(&free_folios);
1803 while (!list_empty(list)) {
1804 struct folio *folio = lru_to_folio(list);
1805
1806 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1807 list_del(&folio->lru);
1808 if (unlikely(!folio_evictable(folio))) {
1809 spin_unlock_irq(&lruvec->lru_lock);
1810 folio_putback_lru(folio);
1811 spin_lock_irq(&lruvec->lru_lock);
1812 continue;
1813 }
1814
1815 /*
1816 * The folio_set_lru needs to be kept here for list integrity.
1817 * Otherwise:
1818 * #0 move_folios_to_lru #1 release_pages
1819 * if (!folio_put_testzero())
1820 * if (folio_put_testzero())
1821 * !lru //skip lru_lock
1822 * folio_set_lru()
1823 * list_add(&folio->lru,)
1824 * list_add(&folio->lru,)
1825 */
1826 folio_set_lru(folio);
1827
1828 if (unlikely(folio_put_testzero(folio))) {
1829 __folio_clear_lru_flags(folio);
1830
1831 if (folio_test_large(folio) &&
1832 folio_test_large_rmappable(folio))
1833 folio_undo_large_rmappable(folio);
1834 if (folio_batch_add(&free_folios, folio) == 0) {
1835 spin_unlock_irq(&lruvec->lru_lock);
1836 mem_cgroup_uncharge_folios(&free_folios);
1837 free_unref_folios(&free_folios);
1838 spin_lock_irq(&lruvec->lru_lock);
1839 }
1840
1841 continue;
1842 }
1843
1844 /*
1845 * All pages were isolated from the same lruvec (and isolation
1846 * inhibits memcg migration).
1847 */
1848 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1849 lruvec_add_folio(lruvec, folio);
1850 nr_pages = folio_nr_pages(folio);
1851 nr_moved += nr_pages;
1852 if (folio_test_active(folio))
1853 workingset_age_nonresident(lruvec, nr_pages);
1854 }
1855
1856 if (free_folios.nr) {
1857 spin_unlock_irq(&lruvec->lru_lock);
1858 mem_cgroup_uncharge_folios(&free_folios);
1859 free_unref_folios(&free_folios);
1860 spin_lock_irq(&lruvec->lru_lock);
1861 }
1862
1863 return nr_moved;
1864}
1865
1866/*
1867 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1868 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1869 * we should not throttle. Otherwise it is safe to do so.
1870 */
1871static int current_may_throttle(void)
1872{
1873 return !(current->flags & PF_LOCAL_THROTTLE);
1874}
1875
1876/*
1877 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1878 * of reclaimed pages
1879 */
1880static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1881 struct lruvec *lruvec, struct scan_control *sc,
1882 enum lru_list lru)
1883{
1884 LIST_HEAD(folio_list);
1885 unsigned long nr_scanned;
1886 unsigned int nr_reclaimed = 0;
1887 unsigned long nr_taken;
1888 struct reclaim_stat stat;
1889 bool file = is_file_lru(lru);
1890 enum vm_event_item item;
1891 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1892 bool stalled = false;
1893
1894 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1895 if (stalled)
1896 return 0;
1897
1898 /* wait a bit for the reclaimer. */
1899 stalled = true;
1900 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1901
1902 /* We are about to die and free our memory. Return now. */
1903 if (fatal_signal_pending(current))
1904 return SWAP_CLUSTER_MAX;
1905 }
1906
1907 lru_add_drain();
1908
1909 spin_lock_irq(&lruvec->lru_lock);
1910
1911 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1912 &nr_scanned, sc, lru);
1913
1914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1915 item = PGSCAN_KSWAPD + reclaimer_offset();
1916 if (!cgroup_reclaim(sc))
1917 __count_vm_events(item, nr_scanned);
1918 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1919 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1920
1921 spin_unlock_irq(&lruvec->lru_lock);
1922
1923 if (nr_taken == 0)
1924 return 0;
1925
1926 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1927
1928 spin_lock_irq(&lruvec->lru_lock);
1929 move_folios_to_lru(lruvec, &folio_list);
1930
1931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1932 item = PGSTEAL_KSWAPD + reclaimer_offset();
1933 if (!cgroup_reclaim(sc))
1934 __count_vm_events(item, nr_reclaimed);
1935 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1936 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1937 spin_unlock_irq(&lruvec->lru_lock);
1938
1939 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1940
1941 /*
1942 * If dirty folios are scanned that are not queued for IO, it
1943 * implies that flushers are not doing their job. This can
1944 * happen when memory pressure pushes dirty folios to the end of
1945 * the LRU before the dirty limits are breached and the dirty
1946 * data has expired. It can also happen when the proportion of
1947 * dirty folios grows not through writes but through memory
1948 * pressure reclaiming all the clean cache. And in some cases,
1949 * the flushers simply cannot keep up with the allocation
1950 * rate. Nudge the flusher threads in case they are asleep.
1951 */
1952 if (stat.nr_unqueued_dirty == nr_taken) {
1953 wakeup_flusher_threads(WB_REASON_VMSCAN);
1954 /*
1955 * For cgroupv1 dirty throttling is achieved by waking up
1956 * the kernel flusher here and later waiting on folios
1957 * which are in writeback to finish (see shrink_folio_list()).
1958 *
1959 * Flusher may not be able to issue writeback quickly
1960 * enough for cgroupv1 writeback throttling to work
1961 * on a large system.
1962 */
1963 if (!writeback_throttling_sane(sc))
1964 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1965 }
1966
1967 sc->nr.dirty += stat.nr_dirty;
1968 sc->nr.congested += stat.nr_congested;
1969 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1970 sc->nr.writeback += stat.nr_writeback;
1971 sc->nr.immediate += stat.nr_immediate;
1972 sc->nr.taken += nr_taken;
1973 if (file)
1974 sc->nr.file_taken += nr_taken;
1975
1976 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1977 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1978 return nr_reclaimed;
1979}
1980
1981/*
1982 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1983 *
1984 * We move them the other way if the folio is referenced by one or more
1985 * processes.
1986 *
1987 * If the folios are mostly unmapped, the processing is fast and it is
1988 * appropriate to hold lru_lock across the whole operation. But if
1989 * the folios are mapped, the processing is slow (folio_referenced()), so
1990 * we should drop lru_lock around each folio. It's impossible to balance
1991 * this, so instead we remove the folios from the LRU while processing them.
1992 * It is safe to rely on the active flag against the non-LRU folios in here
1993 * because nobody will play with that bit on a non-LRU folio.
1994 *
1995 * The downside is that we have to touch folio->_refcount against each folio.
1996 * But we had to alter folio->flags anyway.
1997 */
1998static void shrink_active_list(unsigned long nr_to_scan,
1999 struct lruvec *lruvec,
2000 struct scan_control *sc,
2001 enum lru_list lru)
2002{
2003 unsigned long nr_taken;
2004 unsigned long nr_scanned;
2005 unsigned long vm_flags;
2006 LIST_HEAD(l_hold); /* The folios which were snipped off */
2007 LIST_HEAD(l_active);
2008 LIST_HEAD(l_inactive);
2009 unsigned nr_deactivate, nr_activate;
2010 unsigned nr_rotated = 0;
2011 bool file = is_file_lru(lru);
2012 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2013
2014 lru_add_drain();
2015
2016 spin_lock_irq(&lruvec->lru_lock);
2017
2018 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2019 &nr_scanned, sc, lru);
2020
2021 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2022
2023 if (!cgroup_reclaim(sc))
2024 __count_vm_events(PGREFILL, nr_scanned);
2025 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2026
2027 spin_unlock_irq(&lruvec->lru_lock);
2028
2029 while (!list_empty(&l_hold)) {
2030 struct folio *folio;
2031
2032 cond_resched();
2033 folio = lru_to_folio(&l_hold);
2034 list_del(&folio->lru);
2035
2036 if (unlikely(!folio_evictable(folio))) {
2037 folio_putback_lru(folio);
2038 continue;
2039 }
2040
2041 if (unlikely(buffer_heads_over_limit)) {
2042 if (folio_needs_release(folio) &&
2043 folio_trylock(folio)) {
2044 filemap_release_folio(folio, 0);
2045 folio_unlock(folio);
2046 }
2047 }
2048
2049 /* Referenced or rmap lock contention: rotate */
2050 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2051 &vm_flags) != 0) {
2052 /*
2053 * Identify referenced, file-backed active folios and
2054 * give them one more trip around the active list. So
2055 * that executable code get better chances to stay in
2056 * memory under moderate memory pressure. Anon folios
2057 * are not likely to be evicted by use-once streaming
2058 * IO, plus JVM can create lots of anon VM_EXEC folios,
2059 * so we ignore them here.
2060 */
2061 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2062 nr_rotated += folio_nr_pages(folio);
2063 list_add(&folio->lru, &l_active);
2064 continue;
2065 }
2066 }
2067
2068 folio_clear_active(folio); /* we are de-activating */
2069 folio_set_workingset(folio);
2070 list_add(&folio->lru, &l_inactive);
2071 }
2072
2073 /*
2074 * Move folios back to the lru list.
2075 */
2076 spin_lock_irq(&lruvec->lru_lock);
2077
2078 nr_activate = move_folios_to_lru(lruvec, &l_active);
2079 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2080
2081 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2082 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2083
2084 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2085 spin_unlock_irq(&lruvec->lru_lock);
2086
2087 if (nr_rotated)
2088 lru_note_cost(lruvec, file, 0, nr_rotated);
2089 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2090 nr_deactivate, nr_rotated, sc->priority, file);
2091}
2092
2093static unsigned int reclaim_folio_list(struct list_head *folio_list,
2094 struct pglist_data *pgdat,
2095 bool ignore_references)
2096{
2097 struct reclaim_stat dummy_stat;
2098 unsigned int nr_reclaimed;
2099 struct folio *folio;
2100 struct scan_control sc = {
2101 .gfp_mask = GFP_KERNEL,
2102 .may_writepage = 1,
2103 .may_unmap = 1,
2104 .may_swap = 1,
2105 .no_demotion = 1,
2106 };
2107
2108 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, ignore_references);
2109 while (!list_empty(folio_list)) {
2110 folio = lru_to_folio(folio_list);
2111 list_del(&folio->lru);
2112 folio_putback_lru(folio);
2113 }
2114
2115 return nr_reclaimed;
2116}
2117
2118unsigned long reclaim_pages(struct list_head *folio_list, bool ignore_references)
2119{
2120 int nid;
2121 unsigned int nr_reclaimed = 0;
2122 LIST_HEAD(node_folio_list);
2123 unsigned int noreclaim_flag;
2124
2125 if (list_empty(folio_list))
2126 return nr_reclaimed;
2127
2128 noreclaim_flag = memalloc_noreclaim_save();
2129
2130 nid = folio_nid(lru_to_folio(folio_list));
2131 do {
2132 struct folio *folio = lru_to_folio(folio_list);
2133
2134 if (nid == folio_nid(folio)) {
2135 folio_clear_active(folio);
2136 list_move(&folio->lru, &node_folio_list);
2137 continue;
2138 }
2139
2140 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid),
2141 ignore_references);
2142 nid = folio_nid(lru_to_folio(folio_list));
2143 } while (!list_empty(folio_list));
2144
2145 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid), ignore_references);
2146
2147 memalloc_noreclaim_restore(noreclaim_flag);
2148
2149 return nr_reclaimed;
2150}
2151
2152static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2153 struct lruvec *lruvec, struct scan_control *sc)
2154{
2155 if (is_active_lru(lru)) {
2156 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2157 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2158 else
2159 sc->skipped_deactivate = 1;
2160 return 0;
2161 }
2162
2163 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2164}
2165
2166/*
2167 * The inactive anon list should be small enough that the VM never has
2168 * to do too much work.
2169 *
2170 * The inactive file list should be small enough to leave most memory
2171 * to the established workingset on the scan-resistant active list,
2172 * but large enough to avoid thrashing the aggregate readahead window.
2173 *
2174 * Both inactive lists should also be large enough that each inactive
2175 * folio has a chance to be referenced again before it is reclaimed.
2176 *
2177 * If that fails and refaulting is observed, the inactive list grows.
2178 *
2179 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2180 * on this LRU, maintained by the pageout code. An inactive_ratio
2181 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2182 *
2183 * total target max
2184 * memory ratio inactive
2185 * -------------------------------------
2186 * 10MB 1 5MB
2187 * 100MB 1 50MB
2188 * 1GB 3 250MB
2189 * 10GB 10 0.9GB
2190 * 100GB 31 3GB
2191 * 1TB 101 10GB
2192 * 10TB 320 32GB
2193 */
2194static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2195{
2196 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2197 unsigned long inactive, active;
2198 unsigned long inactive_ratio;
2199 unsigned long gb;
2200
2201 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2202 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2203
2204 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2205 if (gb)
2206 inactive_ratio = int_sqrt(10 * gb);
2207 else
2208 inactive_ratio = 1;
2209
2210 return inactive * inactive_ratio < active;
2211}
2212
2213enum scan_balance {
2214 SCAN_EQUAL,
2215 SCAN_FRACT,
2216 SCAN_ANON,
2217 SCAN_FILE,
2218};
2219
2220static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2221{
2222 unsigned long file;
2223 struct lruvec *target_lruvec;
2224
2225 if (lru_gen_enabled())
2226 return;
2227
2228 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2229
2230 /*
2231 * Flush the memory cgroup stats, so that we read accurate per-memcg
2232 * lruvec stats for heuristics.
2233 */
2234 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2235
2236 /*
2237 * Determine the scan balance between anon and file LRUs.
2238 */
2239 spin_lock_irq(&target_lruvec->lru_lock);
2240 sc->anon_cost = target_lruvec->anon_cost;
2241 sc->file_cost = target_lruvec->file_cost;
2242 spin_unlock_irq(&target_lruvec->lru_lock);
2243
2244 /*
2245 * Target desirable inactive:active list ratios for the anon
2246 * and file LRU lists.
2247 */
2248 if (!sc->force_deactivate) {
2249 unsigned long refaults;
2250
2251 /*
2252 * When refaults are being observed, it means a new
2253 * workingset is being established. Deactivate to get
2254 * rid of any stale active pages quickly.
2255 */
2256 refaults = lruvec_page_state(target_lruvec,
2257 WORKINGSET_ACTIVATE_ANON);
2258 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2259 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2260 sc->may_deactivate |= DEACTIVATE_ANON;
2261 else
2262 sc->may_deactivate &= ~DEACTIVATE_ANON;
2263
2264 refaults = lruvec_page_state(target_lruvec,
2265 WORKINGSET_ACTIVATE_FILE);
2266 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2267 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2268 sc->may_deactivate |= DEACTIVATE_FILE;
2269 else
2270 sc->may_deactivate &= ~DEACTIVATE_FILE;
2271 } else
2272 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2273
2274 /*
2275 * If we have plenty of inactive file pages that aren't
2276 * thrashing, try to reclaim those first before touching
2277 * anonymous pages.
2278 */
2279 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2280 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2281 !sc->no_cache_trim_mode)
2282 sc->cache_trim_mode = 1;
2283 else
2284 sc->cache_trim_mode = 0;
2285
2286 /*
2287 * Prevent the reclaimer from falling into the cache trap: as
2288 * cache pages start out inactive, every cache fault will tip
2289 * the scan balance towards the file LRU. And as the file LRU
2290 * shrinks, so does the window for rotation from references.
2291 * This means we have a runaway feedback loop where a tiny
2292 * thrashing file LRU becomes infinitely more attractive than
2293 * anon pages. Try to detect this based on file LRU size.
2294 */
2295 if (!cgroup_reclaim(sc)) {
2296 unsigned long total_high_wmark = 0;
2297 unsigned long free, anon;
2298 int z;
2299
2300 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2301 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2302 node_page_state(pgdat, NR_INACTIVE_FILE);
2303
2304 for (z = 0; z < MAX_NR_ZONES; z++) {
2305 struct zone *zone = &pgdat->node_zones[z];
2306
2307 if (!managed_zone(zone))
2308 continue;
2309
2310 total_high_wmark += high_wmark_pages(zone);
2311 }
2312
2313 /*
2314 * Consider anon: if that's low too, this isn't a
2315 * runaway file reclaim problem, but rather just
2316 * extreme pressure. Reclaim as per usual then.
2317 */
2318 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2319
2320 sc->file_is_tiny =
2321 file + free <= total_high_wmark &&
2322 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2323 anon >> sc->priority;
2324 }
2325}
2326
2327/*
2328 * Determine how aggressively the anon and file LRU lists should be
2329 * scanned.
2330 *
2331 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2332 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2333 */
2334static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2335 unsigned long *nr)
2336{
2337 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2338 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2339 unsigned long anon_cost, file_cost, total_cost;
2340 int swappiness = mem_cgroup_swappiness(memcg);
2341 u64 fraction[ANON_AND_FILE];
2342 u64 denominator = 0; /* gcc */
2343 enum scan_balance scan_balance;
2344 unsigned long ap, fp;
2345 enum lru_list lru;
2346
2347 /* If we have no swap space, do not bother scanning anon folios. */
2348 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2349 scan_balance = SCAN_FILE;
2350 goto out;
2351 }
2352
2353 /*
2354 * Global reclaim will swap to prevent OOM even with no
2355 * swappiness, but memcg users want to use this knob to
2356 * disable swapping for individual groups completely when
2357 * using the memory controller's swap limit feature would be
2358 * too expensive.
2359 */
2360 if (cgroup_reclaim(sc) && !swappiness) {
2361 scan_balance = SCAN_FILE;
2362 goto out;
2363 }
2364
2365 /*
2366 * Do not apply any pressure balancing cleverness when the
2367 * system is close to OOM, scan both anon and file equally
2368 * (unless the swappiness setting disagrees with swapping).
2369 */
2370 if (!sc->priority && swappiness) {
2371 scan_balance = SCAN_EQUAL;
2372 goto out;
2373 }
2374
2375 /*
2376 * If the system is almost out of file pages, force-scan anon.
2377 */
2378 if (sc->file_is_tiny) {
2379 scan_balance = SCAN_ANON;
2380 goto out;
2381 }
2382
2383 /*
2384 * If there is enough inactive page cache, we do not reclaim
2385 * anything from the anonymous working right now.
2386 */
2387 if (sc->cache_trim_mode) {
2388 scan_balance = SCAN_FILE;
2389 goto out;
2390 }
2391
2392 scan_balance = SCAN_FRACT;
2393 /*
2394 * Calculate the pressure balance between anon and file pages.
2395 *
2396 * The amount of pressure we put on each LRU is inversely
2397 * proportional to the cost of reclaiming each list, as
2398 * determined by the share of pages that are refaulting, times
2399 * the relative IO cost of bringing back a swapped out
2400 * anonymous page vs reloading a filesystem page (swappiness).
2401 *
2402 * Although we limit that influence to ensure no list gets
2403 * left behind completely: at least a third of the pressure is
2404 * applied, before swappiness.
2405 *
2406 * With swappiness at 100, anon and file have equal IO cost.
2407 */
2408 total_cost = sc->anon_cost + sc->file_cost;
2409 anon_cost = total_cost + sc->anon_cost;
2410 file_cost = total_cost + sc->file_cost;
2411 total_cost = anon_cost + file_cost;
2412
2413 ap = swappiness * (total_cost + 1);
2414 ap /= anon_cost + 1;
2415
2416 fp = (200 - swappiness) * (total_cost + 1);
2417 fp /= file_cost + 1;
2418
2419 fraction[0] = ap;
2420 fraction[1] = fp;
2421 denominator = ap + fp;
2422out:
2423 for_each_evictable_lru(lru) {
2424 bool file = is_file_lru(lru);
2425 unsigned long lruvec_size;
2426 unsigned long low, min;
2427 unsigned long scan;
2428
2429 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2430 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2431 &min, &low);
2432
2433 if (min || low) {
2434 /*
2435 * Scale a cgroup's reclaim pressure by proportioning
2436 * its current usage to its memory.low or memory.min
2437 * setting.
2438 *
2439 * This is important, as otherwise scanning aggression
2440 * becomes extremely binary -- from nothing as we
2441 * approach the memory protection threshold, to totally
2442 * nominal as we exceed it. This results in requiring
2443 * setting extremely liberal protection thresholds. It
2444 * also means we simply get no protection at all if we
2445 * set it too low, which is not ideal.
2446 *
2447 * If there is any protection in place, we reduce scan
2448 * pressure by how much of the total memory used is
2449 * within protection thresholds.
2450 *
2451 * There is one special case: in the first reclaim pass,
2452 * we skip over all groups that are within their low
2453 * protection. If that fails to reclaim enough pages to
2454 * satisfy the reclaim goal, we come back and override
2455 * the best-effort low protection. However, we still
2456 * ideally want to honor how well-behaved groups are in
2457 * that case instead of simply punishing them all
2458 * equally. As such, we reclaim them based on how much
2459 * memory they are using, reducing the scan pressure
2460 * again by how much of the total memory used is under
2461 * hard protection.
2462 */
2463 unsigned long cgroup_size = mem_cgroup_size(memcg);
2464 unsigned long protection;
2465
2466 /* memory.low scaling, make sure we retry before OOM */
2467 if (!sc->memcg_low_reclaim && low > min) {
2468 protection = low;
2469 sc->memcg_low_skipped = 1;
2470 } else {
2471 protection = min;
2472 }
2473
2474 /* Avoid TOCTOU with earlier protection check */
2475 cgroup_size = max(cgroup_size, protection);
2476
2477 scan = lruvec_size - lruvec_size * protection /
2478 (cgroup_size + 1);
2479
2480 /*
2481 * Minimally target SWAP_CLUSTER_MAX pages to keep
2482 * reclaim moving forwards, avoiding decrementing
2483 * sc->priority further than desirable.
2484 */
2485 scan = max(scan, SWAP_CLUSTER_MAX);
2486 } else {
2487 scan = lruvec_size;
2488 }
2489
2490 scan >>= sc->priority;
2491
2492 /*
2493 * If the cgroup's already been deleted, make sure to
2494 * scrape out the remaining cache.
2495 */
2496 if (!scan && !mem_cgroup_online(memcg))
2497 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2498
2499 switch (scan_balance) {
2500 case SCAN_EQUAL:
2501 /* Scan lists relative to size */
2502 break;
2503 case SCAN_FRACT:
2504 /*
2505 * Scan types proportional to swappiness and
2506 * their relative recent reclaim efficiency.
2507 * Make sure we don't miss the last page on
2508 * the offlined memory cgroups because of a
2509 * round-off error.
2510 */
2511 scan = mem_cgroup_online(memcg) ?
2512 div64_u64(scan * fraction[file], denominator) :
2513 DIV64_U64_ROUND_UP(scan * fraction[file],
2514 denominator);
2515 break;
2516 case SCAN_FILE:
2517 case SCAN_ANON:
2518 /* Scan one type exclusively */
2519 if ((scan_balance == SCAN_FILE) != file)
2520 scan = 0;
2521 break;
2522 default:
2523 /* Look ma, no brain */
2524 BUG();
2525 }
2526
2527 nr[lru] = scan;
2528 }
2529}
2530
2531/*
2532 * Anonymous LRU management is a waste if there is
2533 * ultimately no way to reclaim the memory.
2534 */
2535static bool can_age_anon_pages(struct pglist_data *pgdat,
2536 struct scan_control *sc)
2537{
2538 /* Aging the anon LRU is valuable if swap is present: */
2539 if (total_swap_pages > 0)
2540 return true;
2541
2542 /* Also valuable if anon pages can be demoted: */
2543 return can_demote(pgdat->node_id, sc);
2544}
2545
2546#ifdef CONFIG_LRU_GEN
2547
2548#ifdef CONFIG_LRU_GEN_ENABLED
2549DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2550#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2551#else
2552DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2553#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2554#endif
2555
2556static bool should_walk_mmu(void)
2557{
2558 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2559}
2560
2561static bool should_clear_pmd_young(void)
2562{
2563 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2564}
2565
2566/******************************************************************************
2567 * shorthand helpers
2568 ******************************************************************************/
2569
2570#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2571
2572#define DEFINE_MAX_SEQ(lruvec) \
2573 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2574
2575#define DEFINE_MIN_SEQ(lruvec) \
2576 unsigned long min_seq[ANON_AND_FILE] = { \
2577 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2578 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2579 }
2580
2581#define for_each_gen_type_zone(gen, type, zone) \
2582 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2583 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2584 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2585
2586#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2587#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2588
2589static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2590{
2591 struct pglist_data *pgdat = NODE_DATA(nid);
2592
2593#ifdef CONFIG_MEMCG
2594 if (memcg) {
2595 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2596
2597 /* see the comment in mem_cgroup_lruvec() */
2598 if (!lruvec->pgdat)
2599 lruvec->pgdat = pgdat;
2600
2601 return lruvec;
2602 }
2603#endif
2604 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2605
2606 return &pgdat->__lruvec;
2607}
2608
2609static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2610{
2611 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2612 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2613
2614 if (!sc->may_swap)
2615 return 0;
2616
2617 if (!can_demote(pgdat->node_id, sc) &&
2618 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2619 return 0;
2620
2621 return mem_cgroup_swappiness(memcg);
2622}
2623
2624static int get_nr_gens(struct lruvec *lruvec, int type)
2625{
2626 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2627}
2628
2629static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2630{
2631 /* see the comment on lru_gen_folio */
2632 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2633 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2634 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2635}
2636
2637/******************************************************************************
2638 * Bloom filters
2639 ******************************************************************************/
2640
2641/*
2642 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2643 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2644 * bits in a bitmap, k is the number of hash functions and n is the number of
2645 * inserted items.
2646 *
2647 * Page table walkers use one of the two filters to reduce their search space.
2648 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2649 * aging uses the double-buffering technique to flip to the other filter each
2650 * time it produces a new generation. For non-leaf entries that have enough
2651 * leaf entries, the aging carries them over to the next generation in
2652 * walk_pmd_range(); the eviction also report them when walking the rmap
2653 * in lru_gen_look_around().
2654 *
2655 * For future optimizations:
2656 * 1. It's not necessary to keep both filters all the time. The spare one can be
2657 * freed after the RCU grace period and reallocated if needed again.
2658 * 2. And when reallocating, it's worth scaling its size according to the number
2659 * of inserted entries in the other filter, to reduce the memory overhead on
2660 * small systems and false positives on large systems.
2661 * 3. Jenkins' hash function is an alternative to Knuth's.
2662 */
2663#define BLOOM_FILTER_SHIFT 15
2664
2665static inline int filter_gen_from_seq(unsigned long seq)
2666{
2667 return seq % NR_BLOOM_FILTERS;
2668}
2669
2670static void get_item_key(void *item, int *key)
2671{
2672 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2673
2674 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2675
2676 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2677 key[1] = hash >> BLOOM_FILTER_SHIFT;
2678}
2679
2680static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2681 void *item)
2682{
2683 int key[2];
2684 unsigned long *filter;
2685 int gen = filter_gen_from_seq(seq);
2686
2687 filter = READ_ONCE(mm_state->filters[gen]);
2688 if (!filter)
2689 return true;
2690
2691 get_item_key(item, key);
2692
2693 return test_bit(key[0], filter) && test_bit(key[1], filter);
2694}
2695
2696static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2697 void *item)
2698{
2699 int key[2];
2700 unsigned long *filter;
2701 int gen = filter_gen_from_seq(seq);
2702
2703 filter = READ_ONCE(mm_state->filters[gen]);
2704 if (!filter)
2705 return;
2706
2707 get_item_key(item, key);
2708
2709 if (!test_bit(key[0], filter))
2710 set_bit(key[0], filter);
2711 if (!test_bit(key[1], filter))
2712 set_bit(key[1], filter);
2713}
2714
2715static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2716{
2717 unsigned long *filter;
2718 int gen = filter_gen_from_seq(seq);
2719
2720 filter = mm_state->filters[gen];
2721 if (filter) {
2722 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2723 return;
2724 }
2725
2726 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2727 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2728 WRITE_ONCE(mm_state->filters[gen], filter);
2729}
2730
2731/******************************************************************************
2732 * mm_struct list
2733 ******************************************************************************/
2734
2735#ifdef CONFIG_LRU_GEN_WALKS_MMU
2736
2737static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2738{
2739 static struct lru_gen_mm_list mm_list = {
2740 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2741 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2742 };
2743
2744#ifdef CONFIG_MEMCG
2745 if (memcg)
2746 return &memcg->mm_list;
2747#endif
2748 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2749
2750 return &mm_list;
2751}
2752
2753static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2754{
2755 return &lruvec->mm_state;
2756}
2757
2758static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2759{
2760 int key;
2761 struct mm_struct *mm;
2762 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2763 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2764
2765 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2766 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2767
2768 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2769 return NULL;
2770
2771 clear_bit(key, &mm->lru_gen.bitmap);
2772
2773 return mmget_not_zero(mm) ? mm : NULL;
2774}
2775
2776void lru_gen_add_mm(struct mm_struct *mm)
2777{
2778 int nid;
2779 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2780 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2781
2782 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2783#ifdef CONFIG_MEMCG
2784 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2785 mm->lru_gen.memcg = memcg;
2786#endif
2787 spin_lock(&mm_list->lock);
2788
2789 for_each_node_state(nid, N_MEMORY) {
2790 struct lruvec *lruvec = get_lruvec(memcg, nid);
2791 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2792
2793 /* the first addition since the last iteration */
2794 if (mm_state->tail == &mm_list->fifo)
2795 mm_state->tail = &mm->lru_gen.list;
2796 }
2797
2798 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2799
2800 spin_unlock(&mm_list->lock);
2801}
2802
2803void lru_gen_del_mm(struct mm_struct *mm)
2804{
2805 int nid;
2806 struct lru_gen_mm_list *mm_list;
2807 struct mem_cgroup *memcg = NULL;
2808
2809 if (list_empty(&mm->lru_gen.list))
2810 return;
2811
2812#ifdef CONFIG_MEMCG
2813 memcg = mm->lru_gen.memcg;
2814#endif
2815 mm_list = get_mm_list(memcg);
2816
2817 spin_lock(&mm_list->lock);
2818
2819 for_each_node(nid) {
2820 struct lruvec *lruvec = get_lruvec(memcg, nid);
2821 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2822
2823 /* where the current iteration continues after */
2824 if (mm_state->head == &mm->lru_gen.list)
2825 mm_state->head = mm_state->head->prev;
2826
2827 /* where the last iteration ended before */
2828 if (mm_state->tail == &mm->lru_gen.list)
2829 mm_state->tail = mm_state->tail->next;
2830 }
2831
2832 list_del_init(&mm->lru_gen.list);
2833
2834 spin_unlock(&mm_list->lock);
2835
2836#ifdef CONFIG_MEMCG
2837 mem_cgroup_put(mm->lru_gen.memcg);
2838 mm->lru_gen.memcg = NULL;
2839#endif
2840}
2841
2842#ifdef CONFIG_MEMCG
2843void lru_gen_migrate_mm(struct mm_struct *mm)
2844{
2845 struct mem_cgroup *memcg;
2846 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2847
2848 VM_WARN_ON_ONCE(task->mm != mm);
2849 lockdep_assert_held(&task->alloc_lock);
2850
2851 /* for mm_update_next_owner() */
2852 if (mem_cgroup_disabled())
2853 return;
2854
2855 /* migration can happen before addition */
2856 if (!mm->lru_gen.memcg)
2857 return;
2858
2859 rcu_read_lock();
2860 memcg = mem_cgroup_from_task(task);
2861 rcu_read_unlock();
2862 if (memcg == mm->lru_gen.memcg)
2863 return;
2864
2865 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2866
2867 lru_gen_del_mm(mm);
2868 lru_gen_add_mm(mm);
2869}
2870#endif
2871
2872#else /* !CONFIG_LRU_GEN_WALKS_MMU */
2873
2874static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2875{
2876 return NULL;
2877}
2878
2879static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2880{
2881 return NULL;
2882}
2883
2884static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2885{
2886 return NULL;
2887}
2888
2889#endif
2890
2891static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
2892{
2893 int i;
2894 int hist;
2895 struct lruvec *lruvec = walk->lruvec;
2896 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2897
2898 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2899
2900 hist = lru_hist_from_seq(walk->seq);
2901
2902 for (i = 0; i < NR_MM_STATS; i++) {
2903 WRITE_ONCE(mm_state->stats[hist][i],
2904 mm_state->stats[hist][i] + walk->mm_stats[i]);
2905 walk->mm_stats[i] = 0;
2906 }
2907
2908 if (NR_HIST_GENS > 1 && last) {
2909 hist = lru_hist_from_seq(walk->seq + 1);
2910
2911 for (i = 0; i < NR_MM_STATS; i++)
2912 WRITE_ONCE(mm_state->stats[hist][i], 0);
2913 }
2914}
2915
2916static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
2917{
2918 bool first = false;
2919 bool last = false;
2920 struct mm_struct *mm = NULL;
2921 struct lruvec *lruvec = walk->lruvec;
2922 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2923 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2924 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2925
2926 /*
2927 * mm_state->seq is incremented after each iteration of mm_list. There
2928 * are three interesting cases for this page table walker:
2929 * 1. It tries to start a new iteration with a stale max_seq: there is
2930 * nothing left to do.
2931 * 2. It started the next iteration: it needs to reset the Bloom filter
2932 * so that a fresh set of PTE tables can be recorded.
2933 * 3. It ended the current iteration: it needs to reset the mm stats
2934 * counters and tell its caller to increment max_seq.
2935 */
2936 spin_lock(&mm_list->lock);
2937
2938 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
2939
2940 if (walk->seq <= mm_state->seq)
2941 goto done;
2942
2943 if (!mm_state->head)
2944 mm_state->head = &mm_list->fifo;
2945
2946 if (mm_state->head == &mm_list->fifo)
2947 first = true;
2948
2949 do {
2950 mm_state->head = mm_state->head->next;
2951 if (mm_state->head == &mm_list->fifo) {
2952 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2953 last = true;
2954 break;
2955 }
2956
2957 /* force scan for those added after the last iteration */
2958 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2959 mm_state->tail = mm_state->head->next;
2960 walk->force_scan = true;
2961 }
2962 } while (!(mm = get_next_mm(walk)));
2963done:
2964 if (*iter || last)
2965 reset_mm_stats(walk, last);
2966
2967 spin_unlock(&mm_list->lock);
2968
2969 if (mm && first)
2970 reset_bloom_filter(mm_state, walk->seq + 1);
2971
2972 if (*iter)
2973 mmput_async(*iter);
2974
2975 *iter = mm;
2976
2977 return last;
2978}
2979
2980static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
2981{
2982 bool success = false;
2983 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2984 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2985 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2986
2987 spin_lock(&mm_list->lock);
2988
2989 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
2990
2991 if (seq > mm_state->seq) {
2992 mm_state->head = NULL;
2993 mm_state->tail = NULL;
2994 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2995 success = true;
2996 }
2997
2998 spin_unlock(&mm_list->lock);
2999
3000 return success;
3001}
3002
3003/******************************************************************************
3004 * PID controller
3005 ******************************************************************************/
3006
3007/*
3008 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3009 *
3010 * The P term is refaulted/(evicted+protected) from a tier in the generation
3011 * currently being evicted; the I term is the exponential moving average of the
3012 * P term over the generations previously evicted, using the smoothing factor
3013 * 1/2; the D term isn't supported.
3014 *
3015 * The setpoint (SP) is always the first tier of one type; the process variable
3016 * (PV) is either any tier of the other type or any other tier of the same
3017 * type.
3018 *
3019 * The error is the difference between the SP and the PV; the correction is to
3020 * turn off protection when SP>PV or turn on protection when SP<PV.
3021 *
3022 * For future optimizations:
3023 * 1. The D term may discount the other two terms over time so that long-lived
3024 * generations can resist stale information.
3025 */
3026struct ctrl_pos {
3027 unsigned long refaulted;
3028 unsigned long total;
3029 int gain;
3030};
3031
3032static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3033 struct ctrl_pos *pos)
3034{
3035 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3036 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3037
3038 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3039 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3040 pos->total = lrugen->avg_total[type][tier] +
3041 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3042 if (tier)
3043 pos->total += lrugen->protected[hist][type][tier - 1];
3044 pos->gain = gain;
3045}
3046
3047static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3048{
3049 int hist, tier;
3050 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3051 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3052 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3053
3054 lockdep_assert_held(&lruvec->lru_lock);
3055
3056 if (!carryover && !clear)
3057 return;
3058
3059 hist = lru_hist_from_seq(seq);
3060
3061 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3062 if (carryover) {
3063 unsigned long sum;
3064
3065 sum = lrugen->avg_refaulted[type][tier] +
3066 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3067 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3068
3069 sum = lrugen->avg_total[type][tier] +
3070 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3071 if (tier)
3072 sum += lrugen->protected[hist][type][tier - 1];
3073 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3074 }
3075
3076 if (clear) {
3077 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3078 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3079 if (tier)
3080 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3081 }
3082 }
3083}
3084
3085static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3086{
3087 /*
3088 * Return true if the PV has a limited number of refaults or a lower
3089 * refaulted/total than the SP.
3090 */
3091 return pv->refaulted < MIN_LRU_BATCH ||
3092 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3093 (sp->refaulted + 1) * pv->total * pv->gain;
3094}
3095
3096/******************************************************************************
3097 * the aging
3098 ******************************************************************************/
3099
3100/* promote pages accessed through page tables */
3101static int folio_update_gen(struct folio *folio, int gen)
3102{
3103 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3104
3105 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3106 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3107
3108 do {
3109 /* lru_gen_del_folio() has isolated this page? */
3110 if (!(old_flags & LRU_GEN_MASK)) {
3111 /* for shrink_folio_list() */
3112 new_flags = old_flags | BIT(PG_referenced);
3113 continue;
3114 }
3115
3116 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3117 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3118 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3119
3120 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3121}
3122
3123/* protect pages accessed multiple times through file descriptors */
3124static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3125{
3126 int type = folio_is_file_lru(folio);
3127 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3128 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3129 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3130
3131 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3132
3133 do {
3134 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3135 /* folio_update_gen() has promoted this page? */
3136 if (new_gen >= 0 && new_gen != old_gen)
3137 return new_gen;
3138
3139 new_gen = (old_gen + 1) % MAX_NR_GENS;
3140
3141 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3142 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3143 /* for folio_end_writeback() */
3144 if (reclaiming)
3145 new_flags |= BIT(PG_reclaim);
3146 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3147
3148 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3149
3150 return new_gen;
3151}
3152
3153static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3154 int old_gen, int new_gen)
3155{
3156 int type = folio_is_file_lru(folio);
3157 int zone = folio_zonenum(folio);
3158 int delta = folio_nr_pages(folio);
3159
3160 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3161 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3162
3163 walk->batched++;
3164
3165 walk->nr_pages[old_gen][type][zone] -= delta;
3166 walk->nr_pages[new_gen][type][zone] += delta;
3167}
3168
3169static void reset_batch_size(struct lru_gen_mm_walk *walk)
3170{
3171 int gen, type, zone;
3172 struct lruvec *lruvec = walk->lruvec;
3173 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3174
3175 walk->batched = 0;
3176
3177 for_each_gen_type_zone(gen, type, zone) {
3178 enum lru_list lru = type * LRU_INACTIVE_FILE;
3179 int delta = walk->nr_pages[gen][type][zone];
3180
3181 if (!delta)
3182 continue;
3183
3184 walk->nr_pages[gen][type][zone] = 0;
3185 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3186 lrugen->nr_pages[gen][type][zone] + delta);
3187
3188 if (lru_gen_is_active(lruvec, gen))
3189 lru += LRU_ACTIVE;
3190 __update_lru_size(lruvec, lru, zone, delta);
3191 }
3192}
3193
3194static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3195{
3196 struct address_space *mapping;
3197 struct vm_area_struct *vma = args->vma;
3198 struct lru_gen_mm_walk *walk = args->private;
3199
3200 if (!vma_is_accessible(vma))
3201 return true;
3202
3203 if (is_vm_hugetlb_page(vma))
3204 return true;
3205
3206 if (!vma_has_recency(vma))
3207 return true;
3208
3209 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3210 return true;
3211
3212 if (vma == get_gate_vma(vma->vm_mm))
3213 return true;
3214
3215 if (vma_is_anonymous(vma))
3216 return !walk->can_swap;
3217
3218 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3219 return true;
3220
3221 mapping = vma->vm_file->f_mapping;
3222 if (mapping_unevictable(mapping))
3223 return true;
3224
3225 if (shmem_mapping(mapping))
3226 return !walk->can_swap;
3227
3228 /* to exclude special mappings like dax, etc. */
3229 return !mapping->a_ops->read_folio;
3230}
3231
3232/*
3233 * Some userspace memory allocators map many single-page VMAs. Instead of
3234 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3235 * table to reduce zigzags and improve cache performance.
3236 */
3237static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3238 unsigned long *vm_start, unsigned long *vm_end)
3239{
3240 unsigned long start = round_up(*vm_end, size);
3241 unsigned long end = (start | ~mask) + 1;
3242 VMA_ITERATOR(vmi, args->mm, start);
3243
3244 VM_WARN_ON_ONCE(mask & size);
3245 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3246
3247 for_each_vma(vmi, args->vma) {
3248 if (end && end <= args->vma->vm_start)
3249 return false;
3250
3251 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3252 continue;
3253
3254 *vm_start = max(start, args->vma->vm_start);
3255 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3256
3257 return true;
3258 }
3259
3260 return false;
3261}
3262
3263static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3264{
3265 unsigned long pfn = pte_pfn(pte);
3266
3267 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3268
3269 if (!pte_present(pte) || is_zero_pfn(pfn))
3270 return -1;
3271
3272 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3273 return -1;
3274
3275 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3276 return -1;
3277
3278 return pfn;
3279}
3280
3281static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3282{
3283 unsigned long pfn = pmd_pfn(pmd);
3284
3285 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3286
3287 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3288 return -1;
3289
3290 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3291 return -1;
3292
3293 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3294 return -1;
3295
3296 return pfn;
3297}
3298
3299static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3300 struct pglist_data *pgdat, bool can_swap)
3301{
3302 struct folio *folio;
3303
3304 /* try to avoid unnecessary memory loads */
3305 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3306 return NULL;
3307
3308 folio = pfn_folio(pfn);
3309 if (folio_nid(folio) != pgdat->node_id)
3310 return NULL;
3311
3312 if (folio_memcg_rcu(folio) != memcg)
3313 return NULL;
3314
3315 /* file VMAs can contain anon pages from COW */
3316 if (!folio_is_file_lru(folio) && !can_swap)
3317 return NULL;
3318
3319 return folio;
3320}
3321
3322static bool suitable_to_scan(int total, int young)
3323{
3324 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3325
3326 /* suitable if the average number of young PTEs per cacheline is >=1 */
3327 return young * n >= total;
3328}
3329
3330static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3331 struct mm_walk *args)
3332{
3333 int i;
3334 pte_t *pte;
3335 spinlock_t *ptl;
3336 unsigned long addr;
3337 int total = 0;
3338 int young = 0;
3339 struct lru_gen_mm_walk *walk = args->private;
3340 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3341 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3342 DEFINE_MAX_SEQ(walk->lruvec);
3343 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3344
3345 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3346 if (!pte)
3347 return false;
3348 if (!spin_trylock(ptl)) {
3349 pte_unmap(pte);
3350 return false;
3351 }
3352
3353 arch_enter_lazy_mmu_mode();
3354restart:
3355 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3356 unsigned long pfn;
3357 struct folio *folio;
3358 pte_t ptent = ptep_get(pte + i);
3359
3360 total++;
3361 walk->mm_stats[MM_LEAF_TOTAL]++;
3362
3363 pfn = get_pte_pfn(ptent, args->vma, addr);
3364 if (pfn == -1)
3365 continue;
3366
3367 if (!pte_young(ptent)) {
3368 walk->mm_stats[MM_LEAF_OLD]++;
3369 continue;
3370 }
3371
3372 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3373 if (!folio)
3374 continue;
3375
3376 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3377 VM_WARN_ON_ONCE(true);
3378
3379 young++;
3380 walk->mm_stats[MM_LEAF_YOUNG]++;
3381
3382 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3383 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3384 !folio_test_swapcache(folio)))
3385 folio_mark_dirty(folio);
3386
3387 old_gen = folio_update_gen(folio, new_gen);
3388 if (old_gen >= 0 && old_gen != new_gen)
3389 update_batch_size(walk, folio, old_gen, new_gen);
3390 }
3391
3392 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3393 goto restart;
3394
3395 arch_leave_lazy_mmu_mode();
3396 pte_unmap_unlock(pte, ptl);
3397
3398 return suitable_to_scan(total, young);
3399}
3400
3401static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3402 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3403{
3404 int i;
3405 pmd_t *pmd;
3406 spinlock_t *ptl;
3407 struct lru_gen_mm_walk *walk = args->private;
3408 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3409 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3410 DEFINE_MAX_SEQ(walk->lruvec);
3411 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3412
3413 VM_WARN_ON_ONCE(pud_leaf(*pud));
3414
3415 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3416 if (*first == -1) {
3417 *first = addr;
3418 bitmap_zero(bitmap, MIN_LRU_BATCH);
3419 return;
3420 }
3421
3422 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3423 if (i && i <= MIN_LRU_BATCH) {
3424 __set_bit(i - 1, bitmap);
3425 return;
3426 }
3427
3428 pmd = pmd_offset(pud, *first);
3429
3430 ptl = pmd_lockptr(args->mm, pmd);
3431 if (!spin_trylock(ptl))
3432 goto done;
3433
3434 arch_enter_lazy_mmu_mode();
3435
3436 do {
3437 unsigned long pfn;
3438 struct folio *folio;
3439
3440 /* don't round down the first address */
3441 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3442
3443 pfn = get_pmd_pfn(pmd[i], vma, addr);
3444 if (pfn == -1)
3445 goto next;
3446
3447 if (!pmd_trans_huge(pmd[i])) {
3448 if (should_clear_pmd_young())
3449 pmdp_test_and_clear_young(vma, addr, pmd + i);
3450 goto next;
3451 }
3452
3453 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3454 if (!folio)
3455 goto next;
3456
3457 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3458 goto next;
3459
3460 walk->mm_stats[MM_LEAF_YOUNG]++;
3461
3462 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3463 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3464 !folio_test_swapcache(folio)))
3465 folio_mark_dirty(folio);
3466
3467 old_gen = folio_update_gen(folio, new_gen);
3468 if (old_gen >= 0 && old_gen != new_gen)
3469 update_batch_size(walk, folio, old_gen, new_gen);
3470next:
3471 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3472 } while (i <= MIN_LRU_BATCH);
3473
3474 arch_leave_lazy_mmu_mode();
3475 spin_unlock(ptl);
3476done:
3477 *first = -1;
3478}
3479
3480static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3481 struct mm_walk *args)
3482{
3483 int i;
3484 pmd_t *pmd;
3485 unsigned long next;
3486 unsigned long addr;
3487 struct vm_area_struct *vma;
3488 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3489 unsigned long first = -1;
3490 struct lru_gen_mm_walk *walk = args->private;
3491 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3492
3493 VM_WARN_ON_ONCE(pud_leaf(*pud));
3494
3495 /*
3496 * Finish an entire PMD in two passes: the first only reaches to PTE
3497 * tables to avoid taking the PMD lock; the second, if necessary, takes
3498 * the PMD lock to clear the accessed bit in PMD entries.
3499 */
3500 pmd = pmd_offset(pud, start & PUD_MASK);
3501restart:
3502 /* walk_pte_range() may call get_next_vma() */
3503 vma = args->vma;
3504 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3505 pmd_t val = pmdp_get_lockless(pmd + i);
3506
3507 next = pmd_addr_end(addr, end);
3508
3509 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3510 walk->mm_stats[MM_LEAF_TOTAL]++;
3511 continue;
3512 }
3513
3514 if (pmd_trans_huge(val)) {
3515 unsigned long pfn = pmd_pfn(val);
3516 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3517
3518 walk->mm_stats[MM_LEAF_TOTAL]++;
3519
3520 if (!pmd_young(val)) {
3521 walk->mm_stats[MM_LEAF_OLD]++;
3522 continue;
3523 }
3524
3525 /* try to avoid unnecessary memory loads */
3526 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3527 continue;
3528
3529 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3530 continue;
3531 }
3532
3533 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3534
3535 if (should_clear_pmd_young()) {
3536 if (!pmd_young(val))
3537 continue;
3538
3539 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3540 }
3541
3542 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3543 continue;
3544
3545 walk->mm_stats[MM_NONLEAF_FOUND]++;
3546
3547 if (!walk_pte_range(&val, addr, next, args))
3548 continue;
3549
3550 walk->mm_stats[MM_NONLEAF_ADDED]++;
3551
3552 /* carry over to the next generation */
3553 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3554 }
3555
3556 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3557
3558 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3559 goto restart;
3560}
3561
3562static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3563 struct mm_walk *args)
3564{
3565 int i;
3566 pud_t *pud;
3567 unsigned long addr;
3568 unsigned long next;
3569 struct lru_gen_mm_walk *walk = args->private;
3570
3571 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3572
3573 pud = pud_offset(p4d, start & P4D_MASK);
3574restart:
3575 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3576 pud_t val = READ_ONCE(pud[i]);
3577
3578 next = pud_addr_end(addr, end);
3579
3580 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3581 continue;
3582
3583 walk_pmd_range(&val, addr, next, args);
3584
3585 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3586 end = (addr | ~PUD_MASK) + 1;
3587 goto done;
3588 }
3589 }
3590
3591 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3592 goto restart;
3593
3594 end = round_up(end, P4D_SIZE);
3595done:
3596 if (!end || !args->vma)
3597 return 1;
3598
3599 walk->next_addr = max(end, args->vma->vm_start);
3600
3601 return -EAGAIN;
3602}
3603
3604static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3605{
3606 static const struct mm_walk_ops mm_walk_ops = {
3607 .test_walk = should_skip_vma,
3608 .p4d_entry = walk_pud_range,
3609 .walk_lock = PGWALK_RDLOCK,
3610 };
3611
3612 int err;
3613 struct lruvec *lruvec = walk->lruvec;
3614 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3615
3616 walk->next_addr = FIRST_USER_ADDRESS;
3617
3618 do {
3619 DEFINE_MAX_SEQ(lruvec);
3620
3621 err = -EBUSY;
3622
3623 /* another thread might have called inc_max_seq() */
3624 if (walk->seq != max_seq)
3625 break;
3626
3627 /* folio_update_gen() requires stable folio_memcg() */
3628 if (!mem_cgroup_trylock_pages(memcg))
3629 break;
3630
3631 /* the caller might be holding the lock for write */
3632 if (mmap_read_trylock(mm)) {
3633 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3634
3635 mmap_read_unlock(mm);
3636 }
3637
3638 mem_cgroup_unlock_pages();
3639
3640 if (walk->batched) {
3641 spin_lock_irq(&lruvec->lru_lock);
3642 reset_batch_size(walk);
3643 spin_unlock_irq(&lruvec->lru_lock);
3644 }
3645
3646 cond_resched();
3647 } while (err == -EAGAIN);
3648}
3649
3650static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3651{
3652 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3653
3654 if (pgdat && current_is_kswapd()) {
3655 VM_WARN_ON_ONCE(walk);
3656
3657 walk = &pgdat->mm_walk;
3658 } else if (!walk && force_alloc) {
3659 VM_WARN_ON_ONCE(current_is_kswapd());
3660
3661 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3662 }
3663
3664 current->reclaim_state->mm_walk = walk;
3665
3666 return walk;
3667}
3668
3669static void clear_mm_walk(void)
3670{
3671 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3672
3673 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3674 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3675
3676 current->reclaim_state->mm_walk = NULL;
3677
3678 if (!current_is_kswapd())
3679 kfree(walk);
3680}
3681
3682static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3683{
3684 int zone;
3685 int remaining = MAX_LRU_BATCH;
3686 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3687 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3688
3689 if (type == LRU_GEN_ANON && !can_swap)
3690 goto done;
3691
3692 /* prevent cold/hot inversion if force_scan is true */
3693 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3694 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3695
3696 while (!list_empty(head)) {
3697 struct folio *folio = lru_to_folio(head);
3698
3699 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3700 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3701 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3702 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3703
3704 new_gen = folio_inc_gen(lruvec, folio, false);
3705 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3706
3707 if (!--remaining)
3708 return false;
3709 }
3710 }
3711done:
3712 reset_ctrl_pos(lruvec, type, true);
3713 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3714
3715 return true;
3716}
3717
3718static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3719{
3720 int gen, type, zone;
3721 bool success = false;
3722 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3723 DEFINE_MIN_SEQ(lruvec);
3724
3725 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3726
3727 /* find the oldest populated generation */
3728 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3729 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3730 gen = lru_gen_from_seq(min_seq[type]);
3731
3732 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3733 if (!list_empty(&lrugen->folios[gen][type][zone]))
3734 goto next;
3735 }
3736
3737 min_seq[type]++;
3738 }
3739next:
3740 ;
3741 }
3742
3743 /* see the comment on lru_gen_folio */
3744 if (can_swap) {
3745 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3746 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3747 }
3748
3749 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3750 if (min_seq[type] == lrugen->min_seq[type])
3751 continue;
3752
3753 reset_ctrl_pos(lruvec, type, true);
3754 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3755 success = true;
3756 }
3757
3758 return success;
3759}
3760
3761static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3762 bool can_swap, bool force_scan)
3763{
3764 bool success;
3765 int prev, next;
3766 int type, zone;
3767 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3768restart:
3769 if (seq < READ_ONCE(lrugen->max_seq))
3770 return false;
3771
3772 spin_lock_irq(&lruvec->lru_lock);
3773
3774 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3775
3776 success = seq == lrugen->max_seq;
3777 if (!success)
3778 goto unlock;
3779
3780 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3781 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3782 continue;
3783
3784 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3785
3786 if (inc_min_seq(lruvec, type, can_swap))
3787 continue;
3788
3789 spin_unlock_irq(&lruvec->lru_lock);
3790 cond_resched();
3791 goto restart;
3792 }
3793
3794 /*
3795 * Update the active/inactive LRU sizes for compatibility. Both sides of
3796 * the current max_seq need to be covered, since max_seq+1 can overlap
3797 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3798 * overlap, cold/hot inversion happens.
3799 */
3800 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3801 next = lru_gen_from_seq(lrugen->max_seq + 1);
3802
3803 for (type = 0; type < ANON_AND_FILE; type++) {
3804 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3805 enum lru_list lru = type * LRU_INACTIVE_FILE;
3806 long delta = lrugen->nr_pages[prev][type][zone] -
3807 lrugen->nr_pages[next][type][zone];
3808
3809 if (!delta)
3810 continue;
3811
3812 __update_lru_size(lruvec, lru, zone, delta);
3813 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3814 }
3815 }
3816
3817 for (type = 0; type < ANON_AND_FILE; type++)
3818 reset_ctrl_pos(lruvec, type, false);
3819
3820 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3821 /* make sure preceding modifications appear */
3822 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3823unlock:
3824 spin_unlock_irq(&lruvec->lru_lock);
3825
3826 return success;
3827}
3828
3829static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3830 bool can_swap, bool force_scan)
3831{
3832 bool success;
3833 struct lru_gen_mm_walk *walk;
3834 struct mm_struct *mm = NULL;
3835 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3836 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3837
3838 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
3839
3840 if (!mm_state)
3841 return inc_max_seq(lruvec, seq, can_swap, force_scan);
3842
3843 /* see the comment in iterate_mm_list() */
3844 if (seq <= READ_ONCE(mm_state->seq))
3845 return false;
3846
3847 /*
3848 * If the hardware doesn't automatically set the accessed bit, fallback
3849 * to lru_gen_look_around(), which only clears the accessed bit in a
3850 * handful of PTEs. Spreading the work out over a period of time usually
3851 * is less efficient, but it avoids bursty page faults.
3852 */
3853 if (!should_walk_mmu()) {
3854 success = iterate_mm_list_nowalk(lruvec, seq);
3855 goto done;
3856 }
3857
3858 walk = set_mm_walk(NULL, true);
3859 if (!walk) {
3860 success = iterate_mm_list_nowalk(lruvec, seq);
3861 goto done;
3862 }
3863
3864 walk->lruvec = lruvec;
3865 walk->seq = seq;
3866 walk->can_swap = can_swap;
3867 walk->force_scan = force_scan;
3868
3869 do {
3870 success = iterate_mm_list(walk, &mm);
3871 if (mm)
3872 walk_mm(mm, walk);
3873 } while (mm);
3874done:
3875 if (success) {
3876 success = inc_max_seq(lruvec, seq, can_swap, force_scan);
3877 WARN_ON_ONCE(!success);
3878 }
3879
3880 return success;
3881}
3882
3883/******************************************************************************
3884 * working set protection
3885 ******************************************************************************/
3886
3887static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3888{
3889 int gen, type, zone;
3890 unsigned long total = 0;
3891 bool can_swap = get_swappiness(lruvec, sc);
3892 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3893 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3894 DEFINE_MAX_SEQ(lruvec);
3895 DEFINE_MIN_SEQ(lruvec);
3896
3897 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3898 unsigned long seq;
3899
3900 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3901 gen = lru_gen_from_seq(seq);
3902
3903 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3904 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3905 }
3906 }
3907
3908 /* whether the size is big enough to be helpful */
3909 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3910}
3911
3912static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3913 unsigned long min_ttl)
3914{
3915 int gen;
3916 unsigned long birth;
3917 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3918 DEFINE_MIN_SEQ(lruvec);
3919
3920 /* see the comment on lru_gen_folio */
3921 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3922 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3923
3924 if (time_is_after_jiffies(birth + min_ttl))
3925 return false;
3926
3927 if (!lruvec_is_sizable(lruvec, sc))
3928 return false;
3929
3930 mem_cgroup_calculate_protection(NULL, memcg);
3931
3932 return !mem_cgroup_below_min(NULL, memcg);
3933}
3934
3935/* to protect the working set of the last N jiffies */
3936static unsigned long lru_gen_min_ttl __read_mostly;
3937
3938static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3939{
3940 struct mem_cgroup *memcg;
3941 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3942
3943 VM_WARN_ON_ONCE(!current_is_kswapd());
3944
3945 /* check the order to exclude compaction-induced reclaim */
3946 if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
3947 return;
3948
3949 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3950 do {
3951 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3952
3953 if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
3954 mem_cgroup_iter_break(NULL, memcg);
3955 return;
3956 }
3957
3958 cond_resched();
3959 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
3960
3961 /*
3962 * The main goal is to OOM kill if every generation from all memcgs is
3963 * younger than min_ttl. However, another possibility is all memcgs are
3964 * either too small or below min.
3965 */
3966 if (mutex_trylock(&oom_lock)) {
3967 struct oom_control oc = {
3968 .gfp_mask = sc->gfp_mask,
3969 };
3970
3971 out_of_memory(&oc);
3972
3973 mutex_unlock(&oom_lock);
3974 }
3975}
3976
3977/******************************************************************************
3978 * rmap/PT walk feedback
3979 ******************************************************************************/
3980
3981/*
3982 * This function exploits spatial locality when shrink_folio_list() walks the
3983 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3984 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3985 * the PTE table to the Bloom filter. This forms a feedback loop between the
3986 * eviction and the aging.
3987 */
3988void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
3989{
3990 int i;
3991 unsigned long start;
3992 unsigned long end;
3993 struct lru_gen_mm_walk *walk;
3994 int young = 0;
3995 pte_t *pte = pvmw->pte;
3996 unsigned long addr = pvmw->address;
3997 struct vm_area_struct *vma = pvmw->vma;
3998 struct folio *folio = pfn_folio(pvmw->pfn);
3999 bool can_swap = !folio_is_file_lru(folio);
4000 struct mem_cgroup *memcg = folio_memcg(folio);
4001 struct pglist_data *pgdat = folio_pgdat(folio);
4002 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4003 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4004 DEFINE_MAX_SEQ(lruvec);
4005 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4006
4007 lockdep_assert_held(pvmw->ptl);
4008 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4009
4010 if (spin_is_contended(pvmw->ptl))
4011 return;
4012
4013 /* exclude special VMAs containing anon pages from COW */
4014 if (vma->vm_flags & VM_SPECIAL)
4015 return;
4016
4017 /* avoid taking the LRU lock under the PTL when possible */
4018 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4019
4020 start = max(addr & PMD_MASK, vma->vm_start);
4021 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4022
4023 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4024 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4025 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4026 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4027 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4028 else {
4029 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4030 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4031 }
4032 }
4033
4034 /* folio_update_gen() requires stable folio_memcg() */
4035 if (!mem_cgroup_trylock_pages(memcg))
4036 return;
4037
4038 arch_enter_lazy_mmu_mode();
4039
4040 pte -= (addr - start) / PAGE_SIZE;
4041
4042 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4043 unsigned long pfn;
4044 pte_t ptent = ptep_get(pte + i);
4045
4046 pfn = get_pte_pfn(ptent, vma, addr);
4047 if (pfn == -1)
4048 continue;
4049
4050 if (!pte_young(ptent))
4051 continue;
4052
4053 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4054 if (!folio)
4055 continue;
4056
4057 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4058 VM_WARN_ON_ONCE(true);
4059
4060 young++;
4061
4062 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4063 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4064 !folio_test_swapcache(folio)))
4065 folio_mark_dirty(folio);
4066
4067 if (walk) {
4068 old_gen = folio_update_gen(folio, new_gen);
4069 if (old_gen >= 0 && old_gen != new_gen)
4070 update_batch_size(walk, folio, old_gen, new_gen);
4071
4072 continue;
4073 }
4074
4075 old_gen = folio_lru_gen(folio);
4076 if (old_gen < 0)
4077 folio_set_referenced(folio);
4078 else if (old_gen != new_gen)
4079 folio_activate(folio);
4080 }
4081
4082 arch_leave_lazy_mmu_mode();
4083 mem_cgroup_unlock_pages();
4084
4085 /* feedback from rmap walkers to page table walkers */
4086 if (mm_state && suitable_to_scan(i, young))
4087 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4088}
4089
4090/******************************************************************************
4091 * memcg LRU
4092 ******************************************************************************/
4093
4094/* see the comment on MEMCG_NR_GENS */
4095enum {
4096 MEMCG_LRU_NOP,
4097 MEMCG_LRU_HEAD,
4098 MEMCG_LRU_TAIL,
4099 MEMCG_LRU_OLD,
4100 MEMCG_LRU_YOUNG,
4101};
4102
4103static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4104{
4105 int seg;
4106 int old, new;
4107 unsigned long flags;
4108 int bin = get_random_u32_below(MEMCG_NR_BINS);
4109 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4110
4111 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4112
4113 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4114
4115 seg = 0;
4116 new = old = lruvec->lrugen.gen;
4117
4118 /* see the comment on MEMCG_NR_GENS */
4119 if (op == MEMCG_LRU_HEAD)
4120 seg = MEMCG_LRU_HEAD;
4121 else if (op == MEMCG_LRU_TAIL)
4122 seg = MEMCG_LRU_TAIL;
4123 else if (op == MEMCG_LRU_OLD)
4124 new = get_memcg_gen(pgdat->memcg_lru.seq);
4125 else if (op == MEMCG_LRU_YOUNG)
4126 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4127 else
4128 VM_WARN_ON_ONCE(true);
4129
4130 WRITE_ONCE(lruvec->lrugen.seg, seg);
4131 WRITE_ONCE(lruvec->lrugen.gen, new);
4132
4133 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4134
4135 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4136 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4137 else
4138 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4139
4140 pgdat->memcg_lru.nr_memcgs[old]--;
4141 pgdat->memcg_lru.nr_memcgs[new]++;
4142
4143 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4144 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4145
4146 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4147}
4148
4149#ifdef CONFIG_MEMCG
4150
4151void lru_gen_online_memcg(struct mem_cgroup *memcg)
4152{
4153 int gen;
4154 int nid;
4155 int bin = get_random_u32_below(MEMCG_NR_BINS);
4156
4157 for_each_node(nid) {
4158 struct pglist_data *pgdat = NODE_DATA(nid);
4159 struct lruvec *lruvec = get_lruvec(memcg, nid);
4160
4161 spin_lock_irq(&pgdat->memcg_lru.lock);
4162
4163 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4164
4165 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4166
4167 lruvec->lrugen.gen = gen;
4168
4169 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4170 pgdat->memcg_lru.nr_memcgs[gen]++;
4171
4172 spin_unlock_irq(&pgdat->memcg_lru.lock);
4173 }
4174}
4175
4176void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4177{
4178 int nid;
4179
4180 for_each_node(nid) {
4181 struct lruvec *lruvec = get_lruvec(memcg, nid);
4182
4183 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4184 }
4185}
4186
4187void lru_gen_release_memcg(struct mem_cgroup *memcg)
4188{
4189 int gen;
4190 int nid;
4191
4192 for_each_node(nid) {
4193 struct pglist_data *pgdat = NODE_DATA(nid);
4194 struct lruvec *lruvec = get_lruvec(memcg, nid);
4195
4196 spin_lock_irq(&pgdat->memcg_lru.lock);
4197
4198 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4199 goto unlock;
4200
4201 gen = lruvec->lrugen.gen;
4202
4203 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4204 pgdat->memcg_lru.nr_memcgs[gen]--;
4205
4206 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4207 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4208unlock:
4209 spin_unlock_irq(&pgdat->memcg_lru.lock);
4210 }
4211}
4212
4213void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4214{
4215 struct lruvec *lruvec = get_lruvec(memcg, nid);
4216
4217 /* see the comment on MEMCG_NR_GENS */
4218 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4219 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4220}
4221
4222#endif /* CONFIG_MEMCG */
4223
4224/******************************************************************************
4225 * the eviction
4226 ******************************************************************************/
4227
4228static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4229 int tier_idx)
4230{
4231 bool success;
4232 int gen = folio_lru_gen(folio);
4233 int type = folio_is_file_lru(folio);
4234 int zone = folio_zonenum(folio);
4235 int delta = folio_nr_pages(folio);
4236 int refs = folio_lru_refs(folio);
4237 int tier = lru_tier_from_refs(refs);
4238 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4239
4240 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4241
4242 /* unevictable */
4243 if (!folio_evictable(folio)) {
4244 success = lru_gen_del_folio(lruvec, folio, true);
4245 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4246 folio_set_unevictable(folio);
4247 lruvec_add_folio(lruvec, folio);
4248 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4249 return true;
4250 }
4251
4252 /* dirty lazyfree */
4253 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4254 success = lru_gen_del_folio(lruvec, folio, true);
4255 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4256 folio_set_swapbacked(folio);
4257 lruvec_add_folio_tail(lruvec, folio);
4258 return true;
4259 }
4260
4261 /* promoted */
4262 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4263 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4264 return true;
4265 }
4266
4267 /* protected */
4268 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4269 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4270
4271 gen = folio_inc_gen(lruvec, folio, false);
4272 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4273
4274 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4275 lrugen->protected[hist][type][tier - 1] + delta);
4276 return true;
4277 }
4278
4279 /* ineligible */
4280 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4281 gen = folio_inc_gen(lruvec, folio, false);
4282 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4283 return true;
4284 }
4285
4286 /* waiting for writeback */
4287 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4288 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4289 gen = folio_inc_gen(lruvec, folio, true);
4290 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4291 return true;
4292 }
4293
4294 return false;
4295}
4296
4297static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4298{
4299 bool success;
4300
4301 /* swap constrained */
4302 if (!(sc->gfp_mask & __GFP_IO) &&
4303 (folio_test_dirty(folio) ||
4304 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4305 return false;
4306
4307 /* raced with release_pages() */
4308 if (!folio_try_get(folio))
4309 return false;
4310
4311 /* raced with another isolation */
4312 if (!folio_test_clear_lru(folio)) {
4313 folio_put(folio);
4314 return false;
4315 }
4316
4317 /* see the comment on MAX_NR_TIERS */
4318 if (!folio_test_referenced(folio))
4319 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4320
4321 /* for shrink_folio_list() */
4322 folio_clear_reclaim(folio);
4323 folio_clear_referenced(folio);
4324
4325 success = lru_gen_del_folio(lruvec, folio, true);
4326 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4327
4328 return true;
4329}
4330
4331static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4332 int type, int tier, struct list_head *list)
4333{
4334 int i;
4335 int gen;
4336 enum vm_event_item item;
4337 int sorted = 0;
4338 int scanned = 0;
4339 int isolated = 0;
4340 int skipped = 0;
4341 int remaining = MAX_LRU_BATCH;
4342 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4343 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4344
4345 VM_WARN_ON_ONCE(!list_empty(list));
4346
4347 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4348 return 0;
4349
4350 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4351
4352 for (i = MAX_NR_ZONES; i > 0; i--) {
4353 LIST_HEAD(moved);
4354 int skipped_zone = 0;
4355 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4356 struct list_head *head = &lrugen->folios[gen][type][zone];
4357
4358 while (!list_empty(head)) {
4359 struct folio *folio = lru_to_folio(head);
4360 int delta = folio_nr_pages(folio);
4361
4362 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4363 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4364 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4365 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4366
4367 scanned += delta;
4368
4369 if (sort_folio(lruvec, folio, sc, tier))
4370 sorted += delta;
4371 else if (isolate_folio(lruvec, folio, sc)) {
4372 list_add(&folio->lru, list);
4373 isolated += delta;
4374 } else {
4375 list_move(&folio->lru, &moved);
4376 skipped_zone += delta;
4377 }
4378
4379 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4380 break;
4381 }
4382
4383 if (skipped_zone) {
4384 list_splice(&moved, head);
4385 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4386 skipped += skipped_zone;
4387 }
4388
4389 if (!remaining || isolated >= MIN_LRU_BATCH)
4390 break;
4391 }
4392
4393 item = PGSCAN_KSWAPD + reclaimer_offset();
4394 if (!cgroup_reclaim(sc)) {
4395 __count_vm_events(item, isolated);
4396 __count_vm_events(PGREFILL, sorted);
4397 }
4398 __count_memcg_events(memcg, item, isolated);
4399 __count_memcg_events(memcg, PGREFILL, sorted);
4400 __count_vm_events(PGSCAN_ANON + type, isolated);
4401 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4402 scanned, skipped, isolated,
4403 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4404
4405 /*
4406 * There might not be eligible folios due to reclaim_idx. Check the
4407 * remaining to prevent livelock if it's not making progress.
4408 */
4409 return isolated || !remaining ? scanned : 0;
4410}
4411
4412static int get_tier_idx(struct lruvec *lruvec, int type)
4413{
4414 int tier;
4415 struct ctrl_pos sp, pv;
4416
4417 /*
4418 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4419 * This value is chosen because any other tier would have at least twice
4420 * as many refaults as the first tier.
4421 */
4422 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4423 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4424 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4425 if (!positive_ctrl_err(&sp, &pv))
4426 break;
4427 }
4428
4429 return tier - 1;
4430}
4431
4432static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4433{
4434 int type, tier;
4435 struct ctrl_pos sp, pv;
4436 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4437
4438 /*
4439 * Compare the first tier of anon with that of file to determine which
4440 * type to scan. Also need to compare other tiers of the selected type
4441 * with the first tier of the other type to determine the last tier (of
4442 * the selected type) to evict.
4443 */
4444 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4445 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4446 type = positive_ctrl_err(&sp, &pv);
4447
4448 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4449 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4450 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4451 if (!positive_ctrl_err(&sp, &pv))
4452 break;
4453 }
4454
4455 *tier_idx = tier - 1;
4456
4457 return type;
4458}
4459
4460static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4461 int *type_scanned, struct list_head *list)
4462{
4463 int i;
4464 int type;
4465 int scanned;
4466 int tier = -1;
4467 DEFINE_MIN_SEQ(lruvec);
4468
4469 /*
4470 * Try to make the obvious choice first, and if anon and file are both
4471 * available from the same generation,
4472 * 1. Interpret swappiness 1 as file first and MAX_SWAPPINESS as anon
4473 * first.
4474 * 2. If !__GFP_IO, file first since clean pagecache is more likely to
4475 * exist than clean swapcache.
4476 */
4477 if (!swappiness)
4478 type = LRU_GEN_FILE;
4479 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4480 type = LRU_GEN_ANON;
4481 else if (swappiness == 1)
4482 type = LRU_GEN_FILE;
4483 else if (swappiness == 200)
4484 type = LRU_GEN_ANON;
4485 else if (!(sc->gfp_mask & __GFP_IO))
4486 type = LRU_GEN_FILE;
4487 else
4488 type = get_type_to_scan(lruvec, swappiness, &tier);
4489
4490 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4491 if (tier < 0)
4492 tier = get_tier_idx(lruvec, type);
4493
4494 scanned = scan_folios(lruvec, sc, type, tier, list);
4495 if (scanned)
4496 break;
4497
4498 type = !type;
4499 tier = -1;
4500 }
4501
4502 *type_scanned = type;
4503
4504 return scanned;
4505}
4506
4507static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4508{
4509 int type;
4510 int scanned;
4511 int reclaimed;
4512 LIST_HEAD(list);
4513 LIST_HEAD(clean);
4514 struct folio *folio;
4515 struct folio *next;
4516 enum vm_event_item item;
4517 struct reclaim_stat stat;
4518 struct lru_gen_mm_walk *walk;
4519 bool skip_retry = false;
4520 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4521 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4522
4523 spin_lock_irq(&lruvec->lru_lock);
4524
4525 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4526
4527 scanned += try_to_inc_min_seq(lruvec, swappiness);
4528
4529 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4530 scanned = 0;
4531
4532 spin_unlock_irq(&lruvec->lru_lock);
4533
4534 if (list_empty(&list))
4535 return scanned;
4536retry:
4537 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4538 sc->nr_reclaimed += reclaimed;
4539 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4540 scanned, reclaimed, &stat, sc->priority,
4541 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4542
4543 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4544 if (!folio_evictable(folio)) {
4545 list_del(&folio->lru);
4546 folio_putback_lru(folio);
4547 continue;
4548 }
4549
4550 if (folio_test_reclaim(folio) &&
4551 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4552 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4553 if (folio_test_workingset(folio))
4554 folio_set_referenced(folio);
4555 continue;
4556 }
4557
4558 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4559 folio_mapped(folio) || folio_test_locked(folio) ||
4560 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4561 /* don't add rejected folios to the oldest generation */
4562 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4563 BIT(PG_active));
4564 continue;
4565 }
4566
4567 /* retry folios that may have missed folio_rotate_reclaimable() */
4568 list_move(&folio->lru, &clean);
4569 sc->nr_scanned -= folio_nr_pages(folio);
4570 }
4571
4572 spin_lock_irq(&lruvec->lru_lock);
4573
4574 move_folios_to_lru(lruvec, &list);
4575
4576 walk = current->reclaim_state->mm_walk;
4577 if (walk && walk->batched) {
4578 walk->lruvec = lruvec;
4579 reset_batch_size(walk);
4580 }
4581
4582 item = PGSTEAL_KSWAPD + reclaimer_offset();
4583 if (!cgroup_reclaim(sc))
4584 __count_vm_events(item, reclaimed);
4585 __count_memcg_events(memcg, item, reclaimed);
4586 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4587
4588 spin_unlock_irq(&lruvec->lru_lock);
4589
4590 list_splice_init(&clean, &list);
4591
4592 if (!list_empty(&list)) {
4593 skip_retry = true;
4594 goto retry;
4595 }
4596
4597 return scanned;
4598}
4599
4600static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4601 bool can_swap, unsigned long *nr_to_scan)
4602{
4603 int gen, type, zone;
4604 unsigned long old = 0;
4605 unsigned long young = 0;
4606 unsigned long total = 0;
4607 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4608 DEFINE_MIN_SEQ(lruvec);
4609
4610 /* whether this lruvec is completely out of cold folios */
4611 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4612 *nr_to_scan = 0;
4613 return true;
4614 }
4615
4616 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4617 unsigned long seq;
4618
4619 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4620 unsigned long size = 0;
4621
4622 gen = lru_gen_from_seq(seq);
4623
4624 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4625 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4626
4627 total += size;
4628 if (seq == max_seq)
4629 young += size;
4630 else if (seq + MIN_NR_GENS == max_seq)
4631 old += size;
4632 }
4633 }
4634
4635 *nr_to_scan = total;
4636
4637 /*
4638 * The aging tries to be lazy to reduce the overhead, while the eviction
4639 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4640 * ideal number of generations is MIN_NR_GENS+1.
4641 */
4642 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4643 return false;
4644
4645 /*
4646 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4647 * of the total number of pages for each generation. A reasonable range
4648 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4649 * aging cares about the upper bound of hot pages, while the eviction
4650 * cares about the lower bound of cold pages.
4651 */
4652 if (young * MIN_NR_GENS > total)
4653 return true;
4654 if (old * (MIN_NR_GENS + 2) < total)
4655 return true;
4656
4657 return false;
4658}
4659
4660/*
4661 * For future optimizations:
4662 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4663 * reclaim.
4664 */
4665static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4666{
4667 bool success;
4668 unsigned long nr_to_scan;
4669 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4670 DEFINE_MAX_SEQ(lruvec);
4671
4672 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4673 return -1;
4674
4675 success = should_run_aging(lruvec, max_seq, can_swap, &nr_to_scan);
4676
4677 /* try to scrape all its memory if this memcg was deleted */
4678 if (nr_to_scan && !mem_cgroup_online(memcg))
4679 return nr_to_scan;
4680
4681 /* try to get away with not aging at the default priority */
4682 if (!success || sc->priority == DEF_PRIORITY)
4683 return nr_to_scan >> sc->priority;
4684
4685 /* stop scanning this lruvec as it's low on cold folios */
4686 return try_to_inc_max_seq(lruvec, max_seq, can_swap, false) ? -1 : 0;
4687}
4688
4689static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4690{
4691 int i;
4692 enum zone_watermarks mark;
4693
4694 /* don't abort memcg reclaim to ensure fairness */
4695 if (!root_reclaim(sc))
4696 return false;
4697
4698 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4699 return true;
4700
4701 /* check the order to exclude compaction-induced reclaim */
4702 if (!current_is_kswapd() || sc->order)
4703 return false;
4704
4705 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4706 WMARK_PROMO : WMARK_HIGH;
4707
4708 for (i = 0; i <= sc->reclaim_idx; i++) {
4709 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4710 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4711
4712 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4713 return false;
4714 }
4715
4716 /* kswapd should abort if all eligible zones are safe */
4717 return true;
4718}
4719
4720static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4721{
4722 long nr_to_scan;
4723 unsigned long scanned = 0;
4724 int swappiness = get_swappiness(lruvec, sc);
4725
4726 while (true) {
4727 int delta;
4728
4729 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4730 if (nr_to_scan <= 0)
4731 break;
4732
4733 delta = evict_folios(lruvec, sc, swappiness);
4734 if (!delta)
4735 break;
4736
4737 scanned += delta;
4738 if (scanned >= nr_to_scan)
4739 break;
4740
4741 if (should_abort_scan(lruvec, sc))
4742 break;
4743
4744 cond_resched();
4745 }
4746
4747 /* whether this lruvec should be rotated */
4748 return nr_to_scan < 0;
4749}
4750
4751static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4752{
4753 bool success;
4754 unsigned long scanned = sc->nr_scanned;
4755 unsigned long reclaimed = sc->nr_reclaimed;
4756 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4757 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4758
4759 mem_cgroup_calculate_protection(NULL, memcg);
4760
4761 if (mem_cgroup_below_min(NULL, memcg))
4762 return MEMCG_LRU_YOUNG;
4763
4764 if (mem_cgroup_below_low(NULL, memcg)) {
4765 /* see the comment on MEMCG_NR_GENS */
4766 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4767 return MEMCG_LRU_TAIL;
4768
4769 memcg_memory_event(memcg, MEMCG_LOW);
4770 }
4771
4772 success = try_to_shrink_lruvec(lruvec, sc);
4773
4774 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4775
4776 if (!sc->proactive)
4777 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4778 sc->nr_reclaimed - reclaimed);
4779
4780 flush_reclaim_state(sc);
4781
4782 if (success && mem_cgroup_online(memcg))
4783 return MEMCG_LRU_YOUNG;
4784
4785 if (!success && lruvec_is_sizable(lruvec, sc))
4786 return 0;
4787
4788 /* one retry if offlined or too small */
4789 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4790 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4791}
4792
4793static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4794{
4795 int op;
4796 int gen;
4797 int bin;
4798 int first_bin;
4799 struct lruvec *lruvec;
4800 struct lru_gen_folio *lrugen;
4801 struct mem_cgroup *memcg;
4802 struct hlist_nulls_node *pos;
4803
4804 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4805 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4806restart:
4807 op = 0;
4808 memcg = NULL;
4809
4810 rcu_read_lock();
4811
4812 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4813 if (op) {
4814 lru_gen_rotate_memcg(lruvec, op);
4815 op = 0;
4816 }
4817
4818 mem_cgroup_put(memcg);
4819 memcg = NULL;
4820
4821 if (gen != READ_ONCE(lrugen->gen))
4822 continue;
4823
4824 lruvec = container_of(lrugen, struct lruvec, lrugen);
4825 memcg = lruvec_memcg(lruvec);
4826
4827 if (!mem_cgroup_tryget(memcg)) {
4828 lru_gen_release_memcg(memcg);
4829 memcg = NULL;
4830 continue;
4831 }
4832
4833 rcu_read_unlock();
4834
4835 op = shrink_one(lruvec, sc);
4836
4837 rcu_read_lock();
4838
4839 if (should_abort_scan(lruvec, sc))
4840 break;
4841 }
4842
4843 rcu_read_unlock();
4844
4845 if (op)
4846 lru_gen_rotate_memcg(lruvec, op);
4847
4848 mem_cgroup_put(memcg);
4849
4850 if (!is_a_nulls(pos))
4851 return;
4852
4853 /* restart if raced with lru_gen_rotate_memcg() */
4854 if (gen != get_nulls_value(pos))
4855 goto restart;
4856
4857 /* try the rest of the bins of the current generation */
4858 bin = get_memcg_bin(bin + 1);
4859 if (bin != first_bin)
4860 goto restart;
4861}
4862
4863static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4864{
4865 struct blk_plug plug;
4866
4867 VM_WARN_ON_ONCE(root_reclaim(sc));
4868 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4869
4870 lru_add_drain();
4871
4872 blk_start_plug(&plug);
4873
4874 set_mm_walk(NULL, sc->proactive);
4875
4876 if (try_to_shrink_lruvec(lruvec, sc))
4877 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4878
4879 clear_mm_walk();
4880
4881 blk_finish_plug(&plug);
4882}
4883
4884static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4885{
4886 int priority;
4887 unsigned long reclaimable;
4888
4889 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4890 return;
4891 /*
4892 * Determine the initial priority based on
4893 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4894 * where reclaimed_to_scanned_ratio = inactive / total.
4895 */
4896 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4897 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
4898 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4899
4900 /* round down reclaimable and round up sc->nr_to_reclaim */
4901 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4902
4903 sc->priority = clamp(priority, 0, DEF_PRIORITY);
4904}
4905
4906static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4907{
4908 struct blk_plug plug;
4909 unsigned long reclaimed = sc->nr_reclaimed;
4910
4911 VM_WARN_ON_ONCE(!root_reclaim(sc));
4912
4913 /*
4914 * Unmapped clean folios are already prioritized. Scanning for more of
4915 * them is likely futile and can cause high reclaim latency when there
4916 * is a large number of memcgs.
4917 */
4918 if (!sc->may_writepage || !sc->may_unmap)
4919 goto done;
4920
4921 lru_add_drain();
4922
4923 blk_start_plug(&plug);
4924
4925 set_mm_walk(pgdat, sc->proactive);
4926
4927 set_initial_priority(pgdat, sc);
4928
4929 if (current_is_kswapd())
4930 sc->nr_reclaimed = 0;
4931
4932 if (mem_cgroup_disabled())
4933 shrink_one(&pgdat->__lruvec, sc);
4934 else
4935 shrink_many(pgdat, sc);
4936
4937 if (current_is_kswapd())
4938 sc->nr_reclaimed += reclaimed;
4939
4940 clear_mm_walk();
4941
4942 blk_finish_plug(&plug);
4943done:
4944 /* kswapd should never fail */
4945 pgdat->kswapd_failures = 0;
4946}
4947
4948/******************************************************************************
4949 * state change
4950 ******************************************************************************/
4951
4952static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4953{
4954 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4955
4956 if (lrugen->enabled) {
4957 enum lru_list lru;
4958
4959 for_each_evictable_lru(lru) {
4960 if (!list_empty(&lruvec->lists[lru]))
4961 return false;
4962 }
4963 } else {
4964 int gen, type, zone;
4965
4966 for_each_gen_type_zone(gen, type, zone) {
4967 if (!list_empty(&lrugen->folios[gen][type][zone]))
4968 return false;
4969 }
4970 }
4971
4972 return true;
4973}
4974
4975static bool fill_evictable(struct lruvec *lruvec)
4976{
4977 enum lru_list lru;
4978 int remaining = MAX_LRU_BATCH;
4979
4980 for_each_evictable_lru(lru) {
4981 int type = is_file_lru(lru);
4982 bool active = is_active_lru(lru);
4983 struct list_head *head = &lruvec->lists[lru];
4984
4985 while (!list_empty(head)) {
4986 bool success;
4987 struct folio *folio = lru_to_folio(head);
4988
4989 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4990 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
4991 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4992 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
4993
4994 lruvec_del_folio(lruvec, folio);
4995 success = lru_gen_add_folio(lruvec, folio, false);
4996 VM_WARN_ON_ONCE(!success);
4997
4998 if (!--remaining)
4999 return false;
5000 }
5001 }
5002
5003 return true;
5004}
5005
5006static bool drain_evictable(struct lruvec *lruvec)
5007{
5008 int gen, type, zone;
5009 int remaining = MAX_LRU_BATCH;
5010
5011 for_each_gen_type_zone(gen, type, zone) {
5012 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5013
5014 while (!list_empty(head)) {
5015 bool success;
5016 struct folio *folio = lru_to_folio(head);
5017
5018 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5019 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5020 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5021 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5022
5023 success = lru_gen_del_folio(lruvec, folio, false);
5024 VM_WARN_ON_ONCE(!success);
5025 lruvec_add_folio(lruvec, folio);
5026
5027 if (!--remaining)
5028 return false;
5029 }
5030 }
5031
5032 return true;
5033}
5034
5035static void lru_gen_change_state(bool enabled)
5036{
5037 static DEFINE_MUTEX(state_mutex);
5038
5039 struct mem_cgroup *memcg;
5040
5041 cgroup_lock();
5042 cpus_read_lock();
5043 get_online_mems();
5044 mutex_lock(&state_mutex);
5045
5046 if (enabled == lru_gen_enabled())
5047 goto unlock;
5048
5049 if (enabled)
5050 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5051 else
5052 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5053
5054 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5055 do {
5056 int nid;
5057
5058 for_each_node(nid) {
5059 struct lruvec *lruvec = get_lruvec(memcg, nid);
5060
5061 spin_lock_irq(&lruvec->lru_lock);
5062
5063 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5064 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5065
5066 lruvec->lrugen.enabled = enabled;
5067
5068 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5069 spin_unlock_irq(&lruvec->lru_lock);
5070 cond_resched();
5071 spin_lock_irq(&lruvec->lru_lock);
5072 }
5073
5074 spin_unlock_irq(&lruvec->lru_lock);
5075 }
5076
5077 cond_resched();
5078 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5079unlock:
5080 mutex_unlock(&state_mutex);
5081 put_online_mems();
5082 cpus_read_unlock();
5083 cgroup_unlock();
5084}
5085
5086/******************************************************************************
5087 * sysfs interface
5088 ******************************************************************************/
5089
5090static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5091{
5092 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5093}
5094
5095/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5096static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5097 const char *buf, size_t len)
5098{
5099 unsigned int msecs;
5100
5101 if (kstrtouint(buf, 0, &msecs))
5102 return -EINVAL;
5103
5104 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5105
5106 return len;
5107}
5108
5109static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5110
5111static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5112{
5113 unsigned int caps = 0;
5114
5115 if (get_cap(LRU_GEN_CORE))
5116 caps |= BIT(LRU_GEN_CORE);
5117
5118 if (should_walk_mmu())
5119 caps |= BIT(LRU_GEN_MM_WALK);
5120
5121 if (should_clear_pmd_young())
5122 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5123
5124 return sysfs_emit(buf, "0x%04x\n", caps);
5125}
5126
5127/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5128static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5129 const char *buf, size_t len)
5130{
5131 int i;
5132 unsigned int caps;
5133
5134 if (tolower(*buf) == 'n')
5135 caps = 0;
5136 else if (tolower(*buf) == 'y')
5137 caps = -1;
5138 else if (kstrtouint(buf, 0, &caps))
5139 return -EINVAL;
5140
5141 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5142 bool enabled = caps & BIT(i);
5143
5144 if (i == LRU_GEN_CORE)
5145 lru_gen_change_state(enabled);
5146 else if (enabled)
5147 static_branch_enable(&lru_gen_caps[i]);
5148 else
5149 static_branch_disable(&lru_gen_caps[i]);
5150 }
5151
5152 return len;
5153}
5154
5155static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5156
5157static struct attribute *lru_gen_attrs[] = {
5158 &lru_gen_min_ttl_attr.attr,
5159 &lru_gen_enabled_attr.attr,
5160 NULL
5161};
5162
5163static const struct attribute_group lru_gen_attr_group = {
5164 .name = "lru_gen",
5165 .attrs = lru_gen_attrs,
5166};
5167
5168/******************************************************************************
5169 * debugfs interface
5170 ******************************************************************************/
5171
5172static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5173{
5174 struct mem_cgroup *memcg;
5175 loff_t nr_to_skip = *pos;
5176
5177 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5178 if (!m->private)
5179 return ERR_PTR(-ENOMEM);
5180
5181 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5182 do {
5183 int nid;
5184
5185 for_each_node_state(nid, N_MEMORY) {
5186 if (!nr_to_skip--)
5187 return get_lruvec(memcg, nid);
5188 }
5189 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5190
5191 return NULL;
5192}
5193
5194static void lru_gen_seq_stop(struct seq_file *m, void *v)
5195{
5196 if (!IS_ERR_OR_NULL(v))
5197 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5198
5199 kvfree(m->private);
5200 m->private = NULL;
5201}
5202
5203static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5204{
5205 int nid = lruvec_pgdat(v)->node_id;
5206 struct mem_cgroup *memcg = lruvec_memcg(v);
5207
5208 ++*pos;
5209
5210 nid = next_memory_node(nid);
5211 if (nid == MAX_NUMNODES) {
5212 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5213 if (!memcg)
5214 return NULL;
5215
5216 nid = first_memory_node;
5217 }
5218
5219 return get_lruvec(memcg, nid);
5220}
5221
5222static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5223 unsigned long max_seq, unsigned long *min_seq,
5224 unsigned long seq)
5225{
5226 int i;
5227 int type, tier;
5228 int hist = lru_hist_from_seq(seq);
5229 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5230 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5231
5232 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5233 seq_printf(m, " %10d", tier);
5234 for (type = 0; type < ANON_AND_FILE; type++) {
5235 const char *s = " ";
5236 unsigned long n[3] = {};
5237
5238 if (seq == max_seq) {
5239 s = "RT ";
5240 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5241 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5242 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5243 s = "rep";
5244 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5245 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5246 if (tier)
5247 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5248 }
5249
5250 for (i = 0; i < 3; i++)
5251 seq_printf(m, " %10lu%c", n[i], s[i]);
5252 }
5253 seq_putc(m, '\n');
5254 }
5255
5256 if (!mm_state)
5257 return;
5258
5259 seq_puts(m, " ");
5260 for (i = 0; i < NR_MM_STATS; i++) {
5261 const char *s = " ";
5262 unsigned long n = 0;
5263
5264 if (seq == max_seq && NR_HIST_GENS == 1) {
5265 s = "LOYNFA";
5266 n = READ_ONCE(mm_state->stats[hist][i]);
5267 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5268 s = "loynfa";
5269 n = READ_ONCE(mm_state->stats[hist][i]);
5270 }
5271
5272 seq_printf(m, " %10lu%c", n, s[i]);
5273 }
5274 seq_putc(m, '\n');
5275}
5276
5277/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5278static int lru_gen_seq_show(struct seq_file *m, void *v)
5279{
5280 unsigned long seq;
5281 bool full = !debugfs_real_fops(m->file)->write;
5282 struct lruvec *lruvec = v;
5283 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5284 int nid = lruvec_pgdat(lruvec)->node_id;
5285 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5286 DEFINE_MAX_SEQ(lruvec);
5287 DEFINE_MIN_SEQ(lruvec);
5288
5289 if (nid == first_memory_node) {
5290 const char *path = memcg ? m->private : "";
5291
5292#ifdef CONFIG_MEMCG
5293 if (memcg)
5294 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5295#endif
5296 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5297 }
5298
5299 seq_printf(m, " node %5d\n", nid);
5300
5301 if (!full)
5302 seq = min_seq[LRU_GEN_ANON];
5303 else if (max_seq >= MAX_NR_GENS)
5304 seq = max_seq - MAX_NR_GENS + 1;
5305 else
5306 seq = 0;
5307
5308 for (; seq <= max_seq; seq++) {
5309 int type, zone;
5310 int gen = lru_gen_from_seq(seq);
5311 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5312
5313 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5314
5315 for (type = 0; type < ANON_AND_FILE; type++) {
5316 unsigned long size = 0;
5317 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5318
5319 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5320 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5321
5322 seq_printf(m, " %10lu%c", size, mark);
5323 }
5324
5325 seq_putc(m, '\n');
5326
5327 if (full)
5328 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5329 }
5330
5331 return 0;
5332}
5333
5334static const struct seq_operations lru_gen_seq_ops = {
5335 .start = lru_gen_seq_start,
5336 .stop = lru_gen_seq_stop,
5337 .next = lru_gen_seq_next,
5338 .show = lru_gen_seq_show,
5339};
5340
5341static int run_aging(struct lruvec *lruvec, unsigned long seq,
5342 bool can_swap, bool force_scan)
5343{
5344 DEFINE_MAX_SEQ(lruvec);
5345 DEFINE_MIN_SEQ(lruvec);
5346
5347 if (seq < max_seq)
5348 return 0;
5349
5350 if (seq > max_seq)
5351 return -EINVAL;
5352
5353 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5354 return -ERANGE;
5355
5356 try_to_inc_max_seq(lruvec, max_seq, can_swap, force_scan);
5357
5358 return 0;
5359}
5360
5361static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5362 int swappiness, unsigned long nr_to_reclaim)
5363{
5364 DEFINE_MAX_SEQ(lruvec);
5365
5366 if (seq + MIN_NR_GENS > max_seq)
5367 return -EINVAL;
5368
5369 sc->nr_reclaimed = 0;
5370
5371 while (!signal_pending(current)) {
5372 DEFINE_MIN_SEQ(lruvec);
5373
5374 if (seq < min_seq[!swappiness])
5375 return 0;
5376
5377 if (sc->nr_reclaimed >= nr_to_reclaim)
5378 return 0;
5379
5380 if (!evict_folios(lruvec, sc, swappiness))
5381 return 0;
5382
5383 cond_resched();
5384 }
5385
5386 return -EINTR;
5387}
5388
5389static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5390 struct scan_control *sc, int swappiness, unsigned long opt)
5391{
5392 struct lruvec *lruvec;
5393 int err = -EINVAL;
5394 struct mem_cgroup *memcg = NULL;
5395
5396 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5397 return -EINVAL;
5398
5399 if (!mem_cgroup_disabled()) {
5400 rcu_read_lock();
5401
5402 memcg = mem_cgroup_from_id(memcg_id);
5403 if (!mem_cgroup_tryget(memcg))
5404 memcg = NULL;
5405
5406 rcu_read_unlock();
5407
5408 if (!memcg)
5409 return -EINVAL;
5410 }
5411
5412 if (memcg_id != mem_cgroup_id(memcg))
5413 goto done;
5414
5415 lruvec = get_lruvec(memcg, nid);
5416
5417 if (swappiness < 0)
5418 swappiness = get_swappiness(lruvec, sc);
5419 else if (swappiness > 200)
5420 goto done;
5421
5422 switch (cmd) {
5423 case '+':
5424 err = run_aging(lruvec, seq, swappiness, opt);
5425 break;
5426 case '-':
5427 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5428 break;
5429 }
5430done:
5431 mem_cgroup_put(memcg);
5432
5433 return err;
5434}
5435
5436/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5437static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5438 size_t len, loff_t *pos)
5439{
5440 void *buf;
5441 char *cur, *next;
5442 unsigned int flags;
5443 struct blk_plug plug;
5444 int err = -EINVAL;
5445 struct scan_control sc = {
5446 .may_writepage = true,
5447 .may_unmap = true,
5448 .may_swap = true,
5449 .reclaim_idx = MAX_NR_ZONES - 1,
5450 .gfp_mask = GFP_KERNEL,
5451 };
5452
5453 buf = kvmalloc(len + 1, GFP_KERNEL);
5454 if (!buf)
5455 return -ENOMEM;
5456
5457 if (copy_from_user(buf, src, len)) {
5458 kvfree(buf);
5459 return -EFAULT;
5460 }
5461
5462 set_task_reclaim_state(current, &sc.reclaim_state);
5463 flags = memalloc_noreclaim_save();
5464 blk_start_plug(&plug);
5465 if (!set_mm_walk(NULL, true)) {
5466 err = -ENOMEM;
5467 goto done;
5468 }
5469
5470 next = buf;
5471 next[len] = '\0';
5472
5473 while ((cur = strsep(&next, ",;\n"))) {
5474 int n;
5475 int end;
5476 char cmd;
5477 unsigned int memcg_id;
5478 unsigned int nid;
5479 unsigned long seq;
5480 unsigned int swappiness = -1;
5481 unsigned long opt = -1;
5482
5483 cur = skip_spaces(cur);
5484 if (!*cur)
5485 continue;
5486
5487 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5488 &seq, &end, &swappiness, &end, &opt, &end);
5489 if (n < 4 || cur[end]) {
5490 err = -EINVAL;
5491 break;
5492 }
5493
5494 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5495 if (err)
5496 break;
5497 }
5498done:
5499 clear_mm_walk();
5500 blk_finish_plug(&plug);
5501 memalloc_noreclaim_restore(flags);
5502 set_task_reclaim_state(current, NULL);
5503
5504 kvfree(buf);
5505
5506 return err ? : len;
5507}
5508
5509static int lru_gen_seq_open(struct inode *inode, struct file *file)
5510{
5511 return seq_open(file, &lru_gen_seq_ops);
5512}
5513
5514static const struct file_operations lru_gen_rw_fops = {
5515 .open = lru_gen_seq_open,
5516 .read = seq_read,
5517 .write = lru_gen_seq_write,
5518 .llseek = seq_lseek,
5519 .release = seq_release,
5520};
5521
5522static const struct file_operations lru_gen_ro_fops = {
5523 .open = lru_gen_seq_open,
5524 .read = seq_read,
5525 .llseek = seq_lseek,
5526 .release = seq_release,
5527};
5528
5529/******************************************************************************
5530 * initialization
5531 ******************************************************************************/
5532
5533void lru_gen_init_pgdat(struct pglist_data *pgdat)
5534{
5535 int i, j;
5536
5537 spin_lock_init(&pgdat->memcg_lru.lock);
5538
5539 for (i = 0; i < MEMCG_NR_GENS; i++) {
5540 for (j = 0; j < MEMCG_NR_BINS; j++)
5541 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5542 }
5543}
5544
5545void lru_gen_init_lruvec(struct lruvec *lruvec)
5546{
5547 int i;
5548 int gen, type, zone;
5549 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5550 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5551
5552 lrugen->max_seq = MIN_NR_GENS + 1;
5553 lrugen->enabled = lru_gen_enabled();
5554
5555 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5556 lrugen->timestamps[i] = jiffies;
5557
5558 for_each_gen_type_zone(gen, type, zone)
5559 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5560
5561 if (mm_state)
5562 mm_state->seq = MIN_NR_GENS;
5563}
5564
5565#ifdef CONFIG_MEMCG
5566
5567void lru_gen_init_memcg(struct mem_cgroup *memcg)
5568{
5569 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5570
5571 if (!mm_list)
5572 return;
5573
5574 INIT_LIST_HEAD(&mm_list->fifo);
5575 spin_lock_init(&mm_list->lock);
5576}
5577
5578void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5579{
5580 int i;
5581 int nid;
5582 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5583
5584 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5585
5586 for_each_node(nid) {
5587 struct lruvec *lruvec = get_lruvec(memcg, nid);
5588 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5589
5590 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5591 sizeof(lruvec->lrugen.nr_pages)));
5592
5593 lruvec->lrugen.list.next = LIST_POISON1;
5594
5595 if (!mm_state)
5596 continue;
5597
5598 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5599 bitmap_free(mm_state->filters[i]);
5600 mm_state->filters[i] = NULL;
5601 }
5602 }
5603}
5604
5605#endif /* CONFIG_MEMCG */
5606
5607static int __init init_lru_gen(void)
5608{
5609 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5610 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5611
5612 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5613 pr_err("lru_gen: failed to create sysfs group\n");
5614
5615 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5616 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5617
5618 return 0;
5619};
5620late_initcall(init_lru_gen);
5621
5622#else /* !CONFIG_LRU_GEN */
5623
5624static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5625{
5626 BUILD_BUG();
5627}
5628
5629static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5630{
5631 BUILD_BUG();
5632}
5633
5634static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5635{
5636 BUILD_BUG();
5637}
5638
5639#endif /* CONFIG_LRU_GEN */
5640
5641static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5642{
5643 unsigned long nr[NR_LRU_LISTS];
5644 unsigned long targets[NR_LRU_LISTS];
5645 unsigned long nr_to_scan;
5646 enum lru_list lru;
5647 unsigned long nr_reclaimed = 0;
5648 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5649 bool proportional_reclaim;
5650 struct blk_plug plug;
5651
5652 if (lru_gen_enabled() && !root_reclaim(sc)) {
5653 lru_gen_shrink_lruvec(lruvec, sc);
5654 return;
5655 }
5656
5657 get_scan_count(lruvec, sc, nr);
5658
5659 /* Record the original scan target for proportional adjustments later */
5660 memcpy(targets, nr, sizeof(nr));
5661
5662 /*
5663 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5664 * event that can occur when there is little memory pressure e.g.
5665 * multiple streaming readers/writers. Hence, we do not abort scanning
5666 * when the requested number of pages are reclaimed when scanning at
5667 * DEF_PRIORITY on the assumption that the fact we are direct
5668 * reclaiming implies that kswapd is not keeping up and it is best to
5669 * do a batch of work at once. For memcg reclaim one check is made to
5670 * abort proportional reclaim if either the file or anon lru has already
5671 * dropped to zero at the first pass.
5672 */
5673 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5674 sc->priority == DEF_PRIORITY);
5675
5676 blk_start_plug(&plug);
5677 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5678 nr[LRU_INACTIVE_FILE]) {
5679 unsigned long nr_anon, nr_file, percentage;
5680 unsigned long nr_scanned;
5681
5682 for_each_evictable_lru(lru) {
5683 if (nr[lru]) {
5684 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5685 nr[lru] -= nr_to_scan;
5686
5687 nr_reclaimed += shrink_list(lru, nr_to_scan,
5688 lruvec, sc);
5689 }
5690 }
5691
5692 cond_resched();
5693
5694 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5695 continue;
5696
5697 /*
5698 * For kswapd and memcg, reclaim at least the number of pages
5699 * requested. Ensure that the anon and file LRUs are scanned
5700 * proportionally what was requested by get_scan_count(). We
5701 * stop reclaiming one LRU and reduce the amount scanning
5702 * proportional to the original scan target.
5703 */
5704 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5705 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5706
5707 /*
5708 * It's just vindictive to attack the larger once the smaller
5709 * has gone to zero. And given the way we stop scanning the
5710 * smaller below, this makes sure that we only make one nudge
5711 * towards proportionality once we've got nr_to_reclaim.
5712 */
5713 if (!nr_file || !nr_anon)
5714 break;
5715
5716 if (nr_file > nr_anon) {
5717 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5718 targets[LRU_ACTIVE_ANON] + 1;
5719 lru = LRU_BASE;
5720 percentage = nr_anon * 100 / scan_target;
5721 } else {
5722 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5723 targets[LRU_ACTIVE_FILE] + 1;
5724 lru = LRU_FILE;
5725 percentage = nr_file * 100 / scan_target;
5726 }
5727
5728 /* Stop scanning the smaller of the LRU */
5729 nr[lru] = 0;
5730 nr[lru + LRU_ACTIVE] = 0;
5731
5732 /*
5733 * Recalculate the other LRU scan count based on its original
5734 * scan target and the percentage scanning already complete
5735 */
5736 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5737 nr_scanned = targets[lru] - nr[lru];
5738 nr[lru] = targets[lru] * (100 - percentage) / 100;
5739 nr[lru] -= min(nr[lru], nr_scanned);
5740
5741 lru += LRU_ACTIVE;
5742 nr_scanned = targets[lru] - nr[lru];
5743 nr[lru] = targets[lru] * (100 - percentage) / 100;
5744 nr[lru] -= min(nr[lru], nr_scanned);
5745 }
5746 blk_finish_plug(&plug);
5747 sc->nr_reclaimed += nr_reclaimed;
5748
5749 /*
5750 * Even if we did not try to evict anon pages at all, we want to
5751 * rebalance the anon lru active/inactive ratio.
5752 */
5753 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5754 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5755 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5756 sc, LRU_ACTIVE_ANON);
5757}
5758
5759/* Use reclaim/compaction for costly allocs or under memory pressure */
5760static bool in_reclaim_compaction(struct scan_control *sc)
5761{
5762 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5763 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5764 sc->priority < DEF_PRIORITY - 2))
5765 return true;
5766
5767 return false;
5768}
5769
5770/*
5771 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5772 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5773 * true if more pages should be reclaimed such that when the page allocator
5774 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5775 * It will give up earlier than that if there is difficulty reclaiming pages.
5776 */
5777static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5778 unsigned long nr_reclaimed,
5779 struct scan_control *sc)
5780{
5781 unsigned long pages_for_compaction;
5782 unsigned long inactive_lru_pages;
5783 int z;
5784
5785 /* If not in reclaim/compaction mode, stop */
5786 if (!in_reclaim_compaction(sc))
5787 return false;
5788
5789 /*
5790 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5791 * number of pages that were scanned. This will return to the caller
5792 * with the risk reclaim/compaction and the resulting allocation attempt
5793 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5794 * allocations through requiring that the full LRU list has been scanned
5795 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5796 * scan, but that approximation was wrong, and there were corner cases
5797 * where always a non-zero amount of pages were scanned.
5798 */
5799 if (!nr_reclaimed)
5800 return false;
5801
5802 /* If compaction would go ahead or the allocation would succeed, stop */
5803 for (z = 0; z <= sc->reclaim_idx; z++) {
5804 struct zone *zone = &pgdat->node_zones[z];
5805 if (!managed_zone(zone))
5806 continue;
5807
5808 /* Allocation can already succeed, nothing to do */
5809 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5810 sc->reclaim_idx, 0))
5811 return false;
5812
5813 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5814 return false;
5815 }
5816
5817 /*
5818 * If we have not reclaimed enough pages for compaction and the
5819 * inactive lists are large enough, continue reclaiming
5820 */
5821 pages_for_compaction = compact_gap(sc->order);
5822 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5823 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5824 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5825
5826 return inactive_lru_pages > pages_for_compaction;
5827}
5828
5829static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5830{
5831 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5832 struct mem_cgroup *memcg;
5833
5834 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
5835 do {
5836 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5837 unsigned long reclaimed;
5838 unsigned long scanned;
5839
5840 /*
5841 * This loop can become CPU-bound when target memcgs
5842 * aren't eligible for reclaim - either because they
5843 * don't have any reclaimable pages, or because their
5844 * memory is explicitly protected. Avoid soft lockups.
5845 */
5846 cond_resched();
5847
5848 mem_cgroup_calculate_protection(target_memcg, memcg);
5849
5850 if (mem_cgroup_below_min(target_memcg, memcg)) {
5851 /*
5852 * Hard protection.
5853 * If there is no reclaimable memory, OOM.
5854 */
5855 continue;
5856 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5857 /*
5858 * Soft protection.
5859 * Respect the protection only as long as
5860 * there is an unprotected supply
5861 * of reclaimable memory from other cgroups.
5862 */
5863 if (!sc->memcg_low_reclaim) {
5864 sc->memcg_low_skipped = 1;
5865 continue;
5866 }
5867 memcg_memory_event(memcg, MEMCG_LOW);
5868 }
5869
5870 reclaimed = sc->nr_reclaimed;
5871 scanned = sc->nr_scanned;
5872
5873 shrink_lruvec(lruvec, sc);
5874
5875 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5876 sc->priority);
5877
5878 /* Record the group's reclaim efficiency */
5879 if (!sc->proactive)
5880 vmpressure(sc->gfp_mask, memcg, false,
5881 sc->nr_scanned - scanned,
5882 sc->nr_reclaimed - reclaimed);
5883
5884 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5885}
5886
5887static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5888{
5889 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5890 struct lruvec *target_lruvec;
5891 bool reclaimable = false;
5892
5893 if (lru_gen_enabled() && root_reclaim(sc)) {
5894 lru_gen_shrink_node(pgdat, sc);
5895 return;
5896 }
5897
5898 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5899
5900again:
5901 memset(&sc->nr, 0, sizeof(sc->nr));
5902
5903 nr_reclaimed = sc->nr_reclaimed;
5904 nr_scanned = sc->nr_scanned;
5905
5906 prepare_scan_control(pgdat, sc);
5907
5908 shrink_node_memcgs(pgdat, sc);
5909
5910 flush_reclaim_state(sc);
5911
5912 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5913
5914 /* Record the subtree's reclaim efficiency */
5915 if (!sc->proactive)
5916 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5917 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5918
5919 if (nr_node_reclaimed)
5920 reclaimable = true;
5921
5922 if (current_is_kswapd()) {
5923 /*
5924 * If reclaim is isolating dirty pages under writeback,
5925 * it implies that the long-lived page allocation rate
5926 * is exceeding the page laundering rate. Either the
5927 * global limits are not being effective at throttling
5928 * processes due to the page distribution throughout
5929 * zones or there is heavy usage of a slow backing
5930 * device. The only option is to throttle from reclaim
5931 * context which is not ideal as there is no guarantee
5932 * the dirtying process is throttled in the same way
5933 * balance_dirty_pages() manages.
5934 *
5935 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5936 * count the number of pages under pages flagged for
5937 * immediate reclaim and stall if any are encountered
5938 * in the nr_immediate check below.
5939 */
5940 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5941 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5942
5943 /* Allow kswapd to start writing pages during reclaim.*/
5944 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5945 set_bit(PGDAT_DIRTY, &pgdat->flags);
5946
5947 /*
5948 * If kswapd scans pages marked for immediate
5949 * reclaim and under writeback (nr_immediate), it
5950 * implies that pages are cycling through the LRU
5951 * faster than they are written so forcibly stall
5952 * until some pages complete writeback.
5953 */
5954 if (sc->nr.immediate)
5955 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5956 }
5957
5958 /*
5959 * Tag a node/memcg as congested if all the dirty pages were marked
5960 * for writeback and immediate reclaim (counted in nr.congested).
5961 *
5962 * Legacy memcg will stall in page writeback so avoid forcibly
5963 * stalling in reclaim_throttle().
5964 */
5965 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
5966 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
5967 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
5968
5969 if (current_is_kswapd())
5970 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
5971 }
5972
5973 /*
5974 * Stall direct reclaim for IO completions if the lruvec is
5975 * node is congested. Allow kswapd to continue until it
5976 * starts encountering unqueued dirty pages or cycling through
5977 * the LRU too quickly.
5978 */
5979 if (!current_is_kswapd() && current_may_throttle() &&
5980 !sc->hibernation_mode &&
5981 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
5982 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
5983 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
5984
5985 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
5986 goto again;
5987
5988 /*
5989 * Kswapd gives up on balancing particular nodes after too
5990 * many failures to reclaim anything from them and goes to
5991 * sleep. On reclaim progress, reset the failure counter. A
5992 * successful direct reclaim run will revive a dormant kswapd.
5993 */
5994 if (reclaimable)
5995 pgdat->kswapd_failures = 0;
5996 else if (sc->cache_trim_mode)
5997 sc->cache_trim_mode_failed = 1;
5998}
5999
6000/*
6001 * Returns true if compaction should go ahead for a costly-order request, or
6002 * the allocation would already succeed without compaction. Return false if we
6003 * should reclaim first.
6004 */
6005static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6006{
6007 unsigned long watermark;
6008
6009 if (!gfp_compaction_allowed(sc->gfp_mask))
6010 return false;
6011
6012 /* Allocation can already succeed, nothing to do */
6013 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6014 sc->reclaim_idx, 0))
6015 return true;
6016
6017 /* Compaction cannot yet proceed. Do reclaim. */
6018 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6019 return false;
6020
6021 /*
6022 * Compaction is already possible, but it takes time to run and there
6023 * are potentially other callers using the pages just freed. So proceed
6024 * with reclaim to make a buffer of free pages available to give
6025 * compaction a reasonable chance of completing and allocating the page.
6026 * Note that we won't actually reclaim the whole buffer in one attempt
6027 * as the target watermark in should_continue_reclaim() is lower. But if
6028 * we are already above the high+gap watermark, don't reclaim at all.
6029 */
6030 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6031
6032 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6033}
6034
6035static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6036{
6037 /*
6038 * If reclaim is making progress greater than 12% efficiency then
6039 * wake all the NOPROGRESS throttled tasks.
6040 */
6041 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6042 wait_queue_head_t *wqh;
6043
6044 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6045 if (waitqueue_active(wqh))
6046 wake_up(wqh);
6047
6048 return;
6049 }
6050
6051 /*
6052 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6053 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6054 * under writeback and marked for immediate reclaim at the tail of the
6055 * LRU.
6056 */
6057 if (current_is_kswapd() || cgroup_reclaim(sc))
6058 return;
6059
6060 /* Throttle if making no progress at high prioities. */
6061 if (sc->priority == 1 && !sc->nr_reclaimed)
6062 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6063}
6064
6065/*
6066 * This is the direct reclaim path, for page-allocating processes. We only
6067 * try to reclaim pages from zones which will satisfy the caller's allocation
6068 * request.
6069 *
6070 * If a zone is deemed to be full of pinned pages then just give it a light
6071 * scan then give up on it.
6072 */
6073static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6074{
6075 struct zoneref *z;
6076 struct zone *zone;
6077 unsigned long nr_soft_reclaimed;
6078 unsigned long nr_soft_scanned;
6079 gfp_t orig_mask;
6080 pg_data_t *last_pgdat = NULL;
6081 pg_data_t *first_pgdat = NULL;
6082
6083 /*
6084 * If the number of buffer_heads in the machine exceeds the maximum
6085 * allowed level, force direct reclaim to scan the highmem zone as
6086 * highmem pages could be pinning lowmem pages storing buffer_heads
6087 */
6088 orig_mask = sc->gfp_mask;
6089 if (buffer_heads_over_limit) {
6090 sc->gfp_mask |= __GFP_HIGHMEM;
6091 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6092 }
6093
6094 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6095 sc->reclaim_idx, sc->nodemask) {
6096 /*
6097 * Take care memory controller reclaiming has small influence
6098 * to global LRU.
6099 */
6100 if (!cgroup_reclaim(sc)) {
6101 if (!cpuset_zone_allowed(zone,
6102 GFP_KERNEL | __GFP_HARDWALL))
6103 continue;
6104
6105 /*
6106 * If we already have plenty of memory free for
6107 * compaction in this zone, don't free any more.
6108 * Even though compaction is invoked for any
6109 * non-zero order, only frequent costly order
6110 * reclamation is disruptive enough to become a
6111 * noticeable problem, like transparent huge
6112 * page allocations.
6113 */
6114 if (IS_ENABLED(CONFIG_COMPACTION) &&
6115 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6116 compaction_ready(zone, sc)) {
6117 sc->compaction_ready = true;
6118 continue;
6119 }
6120
6121 /*
6122 * Shrink each node in the zonelist once. If the
6123 * zonelist is ordered by zone (not the default) then a
6124 * node may be shrunk multiple times but in that case
6125 * the user prefers lower zones being preserved.
6126 */
6127 if (zone->zone_pgdat == last_pgdat)
6128 continue;
6129
6130 /*
6131 * This steals pages from memory cgroups over softlimit
6132 * and returns the number of reclaimed pages and
6133 * scanned pages. This works for global memory pressure
6134 * and balancing, not for a memcg's limit.
6135 */
6136 nr_soft_scanned = 0;
6137 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6138 sc->order, sc->gfp_mask,
6139 &nr_soft_scanned);
6140 sc->nr_reclaimed += nr_soft_reclaimed;
6141 sc->nr_scanned += nr_soft_scanned;
6142 /* need some check for avoid more shrink_zone() */
6143 }
6144
6145 if (!first_pgdat)
6146 first_pgdat = zone->zone_pgdat;
6147
6148 /* See comment about same check for global reclaim above */
6149 if (zone->zone_pgdat == last_pgdat)
6150 continue;
6151 last_pgdat = zone->zone_pgdat;
6152 shrink_node(zone->zone_pgdat, sc);
6153 }
6154
6155 if (first_pgdat)
6156 consider_reclaim_throttle(first_pgdat, sc);
6157
6158 /*
6159 * Restore to original mask to avoid the impact on the caller if we
6160 * promoted it to __GFP_HIGHMEM.
6161 */
6162 sc->gfp_mask = orig_mask;
6163}
6164
6165static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6166{
6167 struct lruvec *target_lruvec;
6168 unsigned long refaults;
6169
6170 if (lru_gen_enabled())
6171 return;
6172
6173 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6174 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6175 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6176 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6177 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6178}
6179
6180/*
6181 * This is the main entry point to direct page reclaim.
6182 *
6183 * If a full scan of the inactive list fails to free enough memory then we
6184 * are "out of memory" and something needs to be killed.
6185 *
6186 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6187 * high - the zone may be full of dirty or under-writeback pages, which this
6188 * caller can't do much about. We kick the writeback threads and take explicit
6189 * naps in the hope that some of these pages can be written. But if the
6190 * allocating task holds filesystem locks which prevent writeout this might not
6191 * work, and the allocation attempt will fail.
6192 *
6193 * returns: 0, if no pages reclaimed
6194 * else, the number of pages reclaimed
6195 */
6196static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6197 struct scan_control *sc)
6198{
6199 int initial_priority = sc->priority;
6200 pg_data_t *last_pgdat;
6201 struct zoneref *z;
6202 struct zone *zone;
6203retry:
6204 delayacct_freepages_start();
6205
6206 if (!cgroup_reclaim(sc))
6207 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6208
6209 do {
6210 if (!sc->proactive)
6211 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6212 sc->priority);
6213 sc->nr_scanned = 0;
6214 shrink_zones(zonelist, sc);
6215
6216 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6217 break;
6218
6219 if (sc->compaction_ready)
6220 break;
6221
6222 /*
6223 * If we're getting trouble reclaiming, start doing
6224 * writepage even in laptop mode.
6225 */
6226 if (sc->priority < DEF_PRIORITY - 2)
6227 sc->may_writepage = 1;
6228 } while (--sc->priority >= 0);
6229
6230 last_pgdat = NULL;
6231 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6232 sc->nodemask) {
6233 if (zone->zone_pgdat == last_pgdat)
6234 continue;
6235 last_pgdat = zone->zone_pgdat;
6236
6237 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6238
6239 if (cgroup_reclaim(sc)) {
6240 struct lruvec *lruvec;
6241
6242 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6243 zone->zone_pgdat);
6244 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6245 }
6246 }
6247
6248 delayacct_freepages_end();
6249
6250 if (sc->nr_reclaimed)
6251 return sc->nr_reclaimed;
6252
6253 /* Aborted reclaim to try compaction? don't OOM, then */
6254 if (sc->compaction_ready)
6255 return 1;
6256
6257 /*
6258 * We make inactive:active ratio decisions based on the node's
6259 * composition of memory, but a restrictive reclaim_idx or a
6260 * memory.low cgroup setting can exempt large amounts of
6261 * memory from reclaim. Neither of which are very common, so
6262 * instead of doing costly eligibility calculations of the
6263 * entire cgroup subtree up front, we assume the estimates are
6264 * good, and retry with forcible deactivation if that fails.
6265 */
6266 if (sc->skipped_deactivate) {
6267 sc->priority = initial_priority;
6268 sc->force_deactivate = 1;
6269 sc->skipped_deactivate = 0;
6270 goto retry;
6271 }
6272
6273 /* Untapped cgroup reserves? Don't OOM, retry. */
6274 if (sc->memcg_low_skipped) {
6275 sc->priority = initial_priority;
6276 sc->force_deactivate = 0;
6277 sc->memcg_low_reclaim = 1;
6278 sc->memcg_low_skipped = 0;
6279 goto retry;
6280 }
6281
6282 return 0;
6283}
6284
6285static bool allow_direct_reclaim(pg_data_t *pgdat)
6286{
6287 struct zone *zone;
6288 unsigned long pfmemalloc_reserve = 0;
6289 unsigned long free_pages = 0;
6290 int i;
6291 bool wmark_ok;
6292
6293 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6294 return true;
6295
6296 for (i = 0; i <= ZONE_NORMAL; i++) {
6297 zone = &pgdat->node_zones[i];
6298 if (!managed_zone(zone))
6299 continue;
6300
6301 if (!zone_reclaimable_pages(zone))
6302 continue;
6303
6304 pfmemalloc_reserve += min_wmark_pages(zone);
6305 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6306 }
6307
6308 /* If there are no reserves (unexpected config) then do not throttle */
6309 if (!pfmemalloc_reserve)
6310 return true;
6311
6312 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6313
6314 /* kswapd must be awake if processes are being throttled */
6315 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6316 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6317 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6318
6319 wake_up_interruptible(&pgdat->kswapd_wait);
6320 }
6321
6322 return wmark_ok;
6323}
6324
6325/*
6326 * Throttle direct reclaimers if backing storage is backed by the network
6327 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6328 * depleted. kswapd will continue to make progress and wake the processes
6329 * when the low watermark is reached.
6330 *
6331 * Returns true if a fatal signal was delivered during throttling. If this
6332 * happens, the page allocator should not consider triggering the OOM killer.
6333 */
6334static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6335 nodemask_t *nodemask)
6336{
6337 struct zoneref *z;
6338 struct zone *zone;
6339 pg_data_t *pgdat = NULL;
6340
6341 /*
6342 * Kernel threads should not be throttled as they may be indirectly
6343 * responsible for cleaning pages necessary for reclaim to make forward
6344 * progress. kjournald for example may enter direct reclaim while
6345 * committing a transaction where throttling it could forcing other
6346 * processes to block on log_wait_commit().
6347 */
6348 if (current->flags & PF_KTHREAD)
6349 goto out;
6350
6351 /*
6352 * If a fatal signal is pending, this process should not throttle.
6353 * It should return quickly so it can exit and free its memory
6354 */
6355 if (fatal_signal_pending(current))
6356 goto out;
6357
6358 /*
6359 * Check if the pfmemalloc reserves are ok by finding the first node
6360 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6361 * GFP_KERNEL will be required for allocating network buffers when
6362 * swapping over the network so ZONE_HIGHMEM is unusable.
6363 *
6364 * Throttling is based on the first usable node and throttled processes
6365 * wait on a queue until kswapd makes progress and wakes them. There
6366 * is an affinity then between processes waking up and where reclaim
6367 * progress has been made assuming the process wakes on the same node.
6368 * More importantly, processes running on remote nodes will not compete
6369 * for remote pfmemalloc reserves and processes on different nodes
6370 * should make reasonable progress.
6371 */
6372 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6373 gfp_zone(gfp_mask), nodemask) {
6374 if (zone_idx(zone) > ZONE_NORMAL)
6375 continue;
6376
6377 /* Throttle based on the first usable node */
6378 pgdat = zone->zone_pgdat;
6379 if (allow_direct_reclaim(pgdat))
6380 goto out;
6381 break;
6382 }
6383
6384 /* If no zone was usable by the allocation flags then do not throttle */
6385 if (!pgdat)
6386 goto out;
6387
6388 /* Account for the throttling */
6389 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6390
6391 /*
6392 * If the caller cannot enter the filesystem, it's possible that it
6393 * is due to the caller holding an FS lock or performing a journal
6394 * transaction in the case of a filesystem like ext[3|4]. In this case,
6395 * it is not safe to block on pfmemalloc_wait as kswapd could be
6396 * blocked waiting on the same lock. Instead, throttle for up to a
6397 * second before continuing.
6398 */
6399 if (!(gfp_mask & __GFP_FS))
6400 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6401 allow_direct_reclaim(pgdat), HZ);
6402 else
6403 /* Throttle until kswapd wakes the process */
6404 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6405 allow_direct_reclaim(pgdat));
6406
6407 if (fatal_signal_pending(current))
6408 return true;
6409
6410out:
6411 return false;
6412}
6413
6414unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6415 gfp_t gfp_mask, nodemask_t *nodemask)
6416{
6417 unsigned long nr_reclaimed;
6418 struct scan_control sc = {
6419 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6420 .gfp_mask = current_gfp_context(gfp_mask),
6421 .reclaim_idx = gfp_zone(gfp_mask),
6422 .order = order,
6423 .nodemask = nodemask,
6424 .priority = DEF_PRIORITY,
6425 .may_writepage = !laptop_mode,
6426 .may_unmap = 1,
6427 .may_swap = 1,
6428 };
6429
6430 /*
6431 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6432 * Confirm they are large enough for max values.
6433 */
6434 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6435 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6436 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6437
6438 /*
6439 * Do not enter reclaim if fatal signal was delivered while throttled.
6440 * 1 is returned so that the page allocator does not OOM kill at this
6441 * point.
6442 */
6443 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6444 return 1;
6445
6446 set_task_reclaim_state(current, &sc.reclaim_state);
6447 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6448
6449 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6450
6451 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6452 set_task_reclaim_state(current, NULL);
6453
6454 return nr_reclaimed;
6455}
6456
6457#ifdef CONFIG_MEMCG
6458
6459/* Only used by soft limit reclaim. Do not reuse for anything else. */
6460unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6461 gfp_t gfp_mask, bool noswap,
6462 pg_data_t *pgdat,
6463 unsigned long *nr_scanned)
6464{
6465 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6466 struct scan_control sc = {
6467 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6468 .target_mem_cgroup = memcg,
6469 .may_writepage = !laptop_mode,
6470 .may_unmap = 1,
6471 .reclaim_idx = MAX_NR_ZONES - 1,
6472 .may_swap = !noswap,
6473 };
6474
6475 WARN_ON_ONCE(!current->reclaim_state);
6476
6477 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6478 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6479
6480 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6481 sc.gfp_mask);
6482
6483 /*
6484 * NOTE: Although we can get the priority field, using it
6485 * here is not a good idea, since it limits the pages we can scan.
6486 * if we don't reclaim here, the shrink_node from balance_pgdat
6487 * will pick up pages from other mem cgroup's as well. We hack
6488 * the priority and make it zero.
6489 */
6490 shrink_lruvec(lruvec, &sc);
6491
6492 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6493
6494 *nr_scanned = sc.nr_scanned;
6495
6496 return sc.nr_reclaimed;
6497}
6498
6499unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6500 unsigned long nr_pages,
6501 gfp_t gfp_mask,
6502 unsigned int reclaim_options)
6503{
6504 unsigned long nr_reclaimed;
6505 unsigned int noreclaim_flag;
6506 struct scan_control sc = {
6507 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6508 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6509 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6510 .reclaim_idx = MAX_NR_ZONES - 1,
6511 .target_mem_cgroup = memcg,
6512 .priority = DEF_PRIORITY,
6513 .may_writepage = !laptop_mode,
6514 .may_unmap = 1,
6515 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6516 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6517 };
6518 /*
6519 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6520 * equal pressure on all the nodes. This is based on the assumption that
6521 * the reclaim does not bail out early.
6522 */
6523 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6524
6525 set_task_reclaim_state(current, &sc.reclaim_state);
6526 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6527 noreclaim_flag = memalloc_noreclaim_save();
6528
6529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6530
6531 memalloc_noreclaim_restore(noreclaim_flag);
6532 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6533 set_task_reclaim_state(current, NULL);
6534
6535 return nr_reclaimed;
6536}
6537#endif
6538
6539static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6540{
6541 struct mem_cgroup *memcg;
6542 struct lruvec *lruvec;
6543
6544 if (lru_gen_enabled()) {
6545 lru_gen_age_node(pgdat, sc);
6546 return;
6547 }
6548
6549 if (!can_age_anon_pages(pgdat, sc))
6550 return;
6551
6552 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6553 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6554 return;
6555
6556 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6557 do {
6558 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6559 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6560 sc, LRU_ACTIVE_ANON);
6561 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6562 } while (memcg);
6563}
6564
6565static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6566{
6567 int i;
6568 struct zone *zone;
6569
6570 /*
6571 * Check for watermark boosts top-down as the higher zones
6572 * are more likely to be boosted. Both watermarks and boosts
6573 * should not be checked at the same time as reclaim would
6574 * start prematurely when there is no boosting and a lower
6575 * zone is balanced.
6576 */
6577 for (i = highest_zoneidx; i >= 0; i--) {
6578 zone = pgdat->node_zones + i;
6579 if (!managed_zone(zone))
6580 continue;
6581
6582 if (zone->watermark_boost)
6583 return true;
6584 }
6585
6586 return false;
6587}
6588
6589/*
6590 * Returns true if there is an eligible zone balanced for the request order
6591 * and highest_zoneidx
6592 */
6593static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6594{
6595 int i;
6596 unsigned long mark = -1;
6597 struct zone *zone;
6598
6599 /*
6600 * Check watermarks bottom-up as lower zones are more likely to
6601 * meet watermarks.
6602 */
6603 for (i = 0; i <= highest_zoneidx; i++) {
6604 zone = pgdat->node_zones + i;
6605
6606 if (!managed_zone(zone))
6607 continue;
6608
6609 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6610 mark = wmark_pages(zone, WMARK_PROMO);
6611 else
6612 mark = high_wmark_pages(zone);
6613 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6614 return true;
6615 }
6616
6617 /*
6618 * If a node has no managed zone within highest_zoneidx, it does not
6619 * need balancing by definition. This can happen if a zone-restricted
6620 * allocation tries to wake a remote kswapd.
6621 */
6622 if (mark == -1)
6623 return true;
6624
6625 return false;
6626}
6627
6628/* Clear pgdat state for congested, dirty or under writeback. */
6629static void clear_pgdat_congested(pg_data_t *pgdat)
6630{
6631 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6632
6633 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6634 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6635 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6636 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6637}
6638
6639/*
6640 * Prepare kswapd for sleeping. This verifies that there are no processes
6641 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6642 *
6643 * Returns true if kswapd is ready to sleep
6644 */
6645static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6646 int highest_zoneidx)
6647{
6648 /*
6649 * The throttled processes are normally woken up in balance_pgdat() as
6650 * soon as allow_direct_reclaim() is true. But there is a potential
6651 * race between when kswapd checks the watermarks and a process gets
6652 * throttled. There is also a potential race if processes get
6653 * throttled, kswapd wakes, a large process exits thereby balancing the
6654 * zones, which causes kswapd to exit balance_pgdat() before reaching
6655 * the wake up checks. If kswapd is going to sleep, no process should
6656 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6657 * the wake up is premature, processes will wake kswapd and get
6658 * throttled again. The difference from wake ups in balance_pgdat() is
6659 * that here we are under prepare_to_wait().
6660 */
6661 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6662 wake_up_all(&pgdat->pfmemalloc_wait);
6663
6664 /* Hopeless node, leave it to direct reclaim */
6665 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6666 return true;
6667
6668 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6669 clear_pgdat_congested(pgdat);
6670 return true;
6671 }
6672
6673 return false;
6674}
6675
6676/*
6677 * kswapd shrinks a node of pages that are at or below the highest usable
6678 * zone that is currently unbalanced.
6679 *
6680 * Returns true if kswapd scanned at least the requested number of pages to
6681 * reclaim or if the lack of progress was due to pages under writeback.
6682 * This is used to determine if the scanning priority needs to be raised.
6683 */
6684static bool kswapd_shrink_node(pg_data_t *pgdat,
6685 struct scan_control *sc)
6686{
6687 struct zone *zone;
6688 int z;
6689
6690 /* Reclaim a number of pages proportional to the number of zones */
6691 sc->nr_to_reclaim = 0;
6692 for (z = 0; z <= sc->reclaim_idx; z++) {
6693 zone = pgdat->node_zones + z;
6694 if (!managed_zone(zone))
6695 continue;
6696
6697 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6698 }
6699
6700 /*
6701 * Historically care was taken to put equal pressure on all zones but
6702 * now pressure is applied based on node LRU order.
6703 */
6704 shrink_node(pgdat, sc);
6705
6706 /*
6707 * Fragmentation may mean that the system cannot be rebalanced for
6708 * high-order allocations. If twice the allocation size has been
6709 * reclaimed then recheck watermarks only at order-0 to prevent
6710 * excessive reclaim. Assume that a process requested a high-order
6711 * can direct reclaim/compact.
6712 */
6713 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6714 sc->order = 0;
6715
6716 return sc->nr_scanned >= sc->nr_to_reclaim;
6717}
6718
6719/* Page allocator PCP high watermark is lowered if reclaim is active. */
6720static inline void
6721update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6722{
6723 int i;
6724 struct zone *zone;
6725
6726 for (i = 0; i <= highest_zoneidx; i++) {
6727 zone = pgdat->node_zones + i;
6728
6729 if (!managed_zone(zone))
6730 continue;
6731
6732 if (active)
6733 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6734 else
6735 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6736 }
6737}
6738
6739static inline void
6740set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6741{
6742 update_reclaim_active(pgdat, highest_zoneidx, true);
6743}
6744
6745static inline void
6746clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6747{
6748 update_reclaim_active(pgdat, highest_zoneidx, false);
6749}
6750
6751/*
6752 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6753 * that are eligible for use by the caller until at least one zone is
6754 * balanced.
6755 *
6756 * Returns the order kswapd finished reclaiming at.
6757 *
6758 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6759 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6760 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6761 * or lower is eligible for reclaim until at least one usable zone is
6762 * balanced.
6763 */
6764static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6765{
6766 int i;
6767 unsigned long nr_soft_reclaimed;
6768 unsigned long nr_soft_scanned;
6769 unsigned long pflags;
6770 unsigned long nr_boost_reclaim;
6771 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6772 bool boosted;
6773 struct zone *zone;
6774 struct scan_control sc = {
6775 .gfp_mask = GFP_KERNEL,
6776 .order = order,
6777 .may_unmap = 1,
6778 };
6779
6780 set_task_reclaim_state(current, &sc.reclaim_state);
6781 psi_memstall_enter(&pflags);
6782 __fs_reclaim_acquire(_THIS_IP_);
6783
6784 count_vm_event(PAGEOUTRUN);
6785
6786 /*
6787 * Account for the reclaim boost. Note that the zone boost is left in
6788 * place so that parallel allocations that are near the watermark will
6789 * stall or direct reclaim until kswapd is finished.
6790 */
6791 nr_boost_reclaim = 0;
6792 for (i = 0; i <= highest_zoneidx; i++) {
6793 zone = pgdat->node_zones + i;
6794 if (!managed_zone(zone))
6795 continue;
6796
6797 nr_boost_reclaim += zone->watermark_boost;
6798 zone_boosts[i] = zone->watermark_boost;
6799 }
6800 boosted = nr_boost_reclaim;
6801
6802restart:
6803 set_reclaim_active(pgdat, highest_zoneidx);
6804 sc.priority = DEF_PRIORITY;
6805 do {
6806 unsigned long nr_reclaimed = sc.nr_reclaimed;
6807 bool raise_priority = true;
6808 bool balanced;
6809 bool ret;
6810 bool was_frozen;
6811
6812 sc.reclaim_idx = highest_zoneidx;
6813
6814 /*
6815 * If the number of buffer_heads exceeds the maximum allowed
6816 * then consider reclaiming from all zones. This has a dual
6817 * purpose -- on 64-bit systems it is expected that
6818 * buffer_heads are stripped during active rotation. On 32-bit
6819 * systems, highmem pages can pin lowmem memory and shrinking
6820 * buffers can relieve lowmem pressure. Reclaim may still not
6821 * go ahead if all eligible zones for the original allocation
6822 * request are balanced to avoid excessive reclaim from kswapd.
6823 */
6824 if (buffer_heads_over_limit) {
6825 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6826 zone = pgdat->node_zones + i;
6827 if (!managed_zone(zone))
6828 continue;
6829
6830 sc.reclaim_idx = i;
6831 break;
6832 }
6833 }
6834
6835 /*
6836 * If the pgdat is imbalanced then ignore boosting and preserve
6837 * the watermarks for a later time and restart. Note that the
6838 * zone watermarks will be still reset at the end of balancing
6839 * on the grounds that the normal reclaim should be enough to
6840 * re-evaluate if boosting is required when kswapd next wakes.
6841 */
6842 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6843 if (!balanced && nr_boost_reclaim) {
6844 nr_boost_reclaim = 0;
6845 goto restart;
6846 }
6847
6848 /*
6849 * If boosting is not active then only reclaim if there are no
6850 * eligible zones. Note that sc.reclaim_idx is not used as
6851 * buffer_heads_over_limit may have adjusted it.
6852 */
6853 if (!nr_boost_reclaim && balanced)
6854 goto out;
6855
6856 /* Limit the priority of boosting to avoid reclaim writeback */
6857 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6858 raise_priority = false;
6859
6860 /*
6861 * Do not writeback or swap pages for boosted reclaim. The
6862 * intent is to relieve pressure not issue sub-optimal IO
6863 * from reclaim context. If no pages are reclaimed, the
6864 * reclaim will be aborted.
6865 */
6866 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6867 sc.may_swap = !nr_boost_reclaim;
6868
6869 /*
6870 * Do some background aging, to give pages a chance to be
6871 * referenced before reclaiming. All pages are rotated
6872 * regardless of classzone as this is about consistent aging.
6873 */
6874 kswapd_age_node(pgdat, &sc);
6875
6876 /*
6877 * If we're getting trouble reclaiming, start doing writepage
6878 * even in laptop mode.
6879 */
6880 if (sc.priority < DEF_PRIORITY - 2)
6881 sc.may_writepage = 1;
6882
6883 /* Call soft limit reclaim before calling shrink_node. */
6884 sc.nr_scanned = 0;
6885 nr_soft_scanned = 0;
6886 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
6887 sc.gfp_mask, &nr_soft_scanned);
6888 sc.nr_reclaimed += nr_soft_reclaimed;
6889
6890 /*
6891 * There should be no need to raise the scanning priority if
6892 * enough pages are already being scanned that that high
6893 * watermark would be met at 100% efficiency.
6894 */
6895 if (kswapd_shrink_node(pgdat, &sc))
6896 raise_priority = false;
6897
6898 /*
6899 * If the low watermark is met there is no need for processes
6900 * to be throttled on pfmemalloc_wait as they should not be
6901 * able to safely make forward progress. Wake them
6902 */
6903 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6904 allow_direct_reclaim(pgdat))
6905 wake_up_all(&pgdat->pfmemalloc_wait);
6906
6907 /* Check if kswapd should be suspending */
6908 __fs_reclaim_release(_THIS_IP_);
6909 ret = kthread_freezable_should_stop(&was_frozen);
6910 __fs_reclaim_acquire(_THIS_IP_);
6911 if (was_frozen || ret)
6912 break;
6913
6914 /*
6915 * Raise priority if scanning rate is too low or there was no
6916 * progress in reclaiming pages
6917 */
6918 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6919 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6920
6921 /*
6922 * If reclaim made no progress for a boost, stop reclaim as
6923 * IO cannot be queued and it could be an infinite loop in
6924 * extreme circumstances.
6925 */
6926 if (nr_boost_reclaim && !nr_reclaimed)
6927 break;
6928
6929 if (raise_priority || !nr_reclaimed)
6930 sc.priority--;
6931 } while (sc.priority >= 1);
6932
6933 /*
6934 * Restart only if it went through the priority loop all the way,
6935 * but cache_trim_mode didn't work.
6936 */
6937 if (!sc.nr_reclaimed && sc.priority < 1 &&
6938 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
6939 sc.no_cache_trim_mode = 1;
6940 goto restart;
6941 }
6942
6943 if (!sc.nr_reclaimed)
6944 pgdat->kswapd_failures++;
6945
6946out:
6947 clear_reclaim_active(pgdat, highest_zoneidx);
6948
6949 /* If reclaim was boosted, account for the reclaim done in this pass */
6950 if (boosted) {
6951 unsigned long flags;
6952
6953 for (i = 0; i <= highest_zoneidx; i++) {
6954 if (!zone_boosts[i])
6955 continue;
6956
6957 /* Increments are under the zone lock */
6958 zone = pgdat->node_zones + i;
6959 spin_lock_irqsave(&zone->lock, flags);
6960 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6961 spin_unlock_irqrestore(&zone->lock, flags);
6962 }
6963
6964 /*
6965 * As there is now likely space, wakeup kcompact to defragment
6966 * pageblocks.
6967 */
6968 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
6969 }
6970
6971 snapshot_refaults(NULL, pgdat);
6972 __fs_reclaim_release(_THIS_IP_);
6973 psi_memstall_leave(&pflags);
6974 set_task_reclaim_state(current, NULL);
6975
6976 /*
6977 * Return the order kswapd stopped reclaiming at as
6978 * prepare_kswapd_sleep() takes it into account. If another caller
6979 * entered the allocator slow path while kswapd was awake, order will
6980 * remain at the higher level.
6981 */
6982 return sc.order;
6983}
6984
6985/*
6986 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6987 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6988 * not a valid index then either kswapd runs for first time or kswapd couldn't
6989 * sleep after previous reclaim attempt (node is still unbalanced). In that
6990 * case return the zone index of the previous kswapd reclaim cycle.
6991 */
6992static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6993 enum zone_type prev_highest_zoneidx)
6994{
6995 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
6996
6997 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
6998}
6999
7000static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7001 unsigned int highest_zoneidx)
7002{
7003 long remaining = 0;
7004 DEFINE_WAIT(wait);
7005
7006 if (freezing(current) || kthread_should_stop())
7007 return;
7008
7009 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7010
7011 /*
7012 * Try to sleep for a short interval. Note that kcompactd will only be
7013 * woken if it is possible to sleep for a short interval. This is
7014 * deliberate on the assumption that if reclaim cannot keep an
7015 * eligible zone balanced that it's also unlikely that compaction will
7016 * succeed.
7017 */
7018 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7019 /*
7020 * Compaction records what page blocks it recently failed to
7021 * isolate pages from and skips them in the future scanning.
7022 * When kswapd is going to sleep, it is reasonable to assume
7023 * that pages and compaction may succeed so reset the cache.
7024 */
7025 reset_isolation_suitable(pgdat);
7026
7027 /*
7028 * We have freed the memory, now we should compact it to make
7029 * allocation of the requested order possible.
7030 */
7031 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7032
7033 remaining = schedule_timeout(HZ/10);
7034
7035 /*
7036 * If woken prematurely then reset kswapd_highest_zoneidx and
7037 * order. The values will either be from a wakeup request or
7038 * the previous request that slept prematurely.
7039 */
7040 if (remaining) {
7041 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7042 kswapd_highest_zoneidx(pgdat,
7043 highest_zoneidx));
7044
7045 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7046 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7047 }
7048
7049 finish_wait(&pgdat->kswapd_wait, &wait);
7050 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7051 }
7052
7053 /*
7054 * After a short sleep, check if it was a premature sleep. If not, then
7055 * go fully to sleep until explicitly woken up.
7056 */
7057 if (!remaining &&
7058 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7059 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7060
7061 /*
7062 * vmstat counters are not perfectly accurate and the estimated
7063 * value for counters such as NR_FREE_PAGES can deviate from the
7064 * true value by nr_online_cpus * threshold. To avoid the zone
7065 * watermarks being breached while under pressure, we reduce the
7066 * per-cpu vmstat threshold while kswapd is awake and restore
7067 * them before going back to sleep.
7068 */
7069 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7070
7071 if (!kthread_should_stop())
7072 schedule();
7073
7074 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7075 } else {
7076 if (remaining)
7077 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7078 else
7079 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7080 }
7081 finish_wait(&pgdat->kswapd_wait, &wait);
7082}
7083
7084/*
7085 * The background pageout daemon, started as a kernel thread
7086 * from the init process.
7087 *
7088 * This basically trickles out pages so that we have _some_
7089 * free memory available even if there is no other activity
7090 * that frees anything up. This is needed for things like routing
7091 * etc, where we otherwise might have all activity going on in
7092 * asynchronous contexts that cannot page things out.
7093 *
7094 * If there are applications that are active memory-allocators
7095 * (most normal use), this basically shouldn't matter.
7096 */
7097static int kswapd(void *p)
7098{
7099 unsigned int alloc_order, reclaim_order;
7100 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7101 pg_data_t *pgdat = (pg_data_t *)p;
7102 struct task_struct *tsk = current;
7103 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7104
7105 if (!cpumask_empty(cpumask))
7106 set_cpus_allowed_ptr(tsk, cpumask);
7107
7108 /*
7109 * Tell the memory management that we're a "memory allocator",
7110 * and that if we need more memory we should get access to it
7111 * regardless (see "__alloc_pages()"). "kswapd" should
7112 * never get caught in the normal page freeing logic.
7113 *
7114 * (Kswapd normally doesn't need memory anyway, but sometimes
7115 * you need a small amount of memory in order to be able to
7116 * page out something else, and this flag essentially protects
7117 * us from recursively trying to free more memory as we're
7118 * trying to free the first piece of memory in the first place).
7119 */
7120 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7121 set_freezable();
7122
7123 WRITE_ONCE(pgdat->kswapd_order, 0);
7124 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7125 atomic_set(&pgdat->nr_writeback_throttled, 0);
7126 for ( ; ; ) {
7127 bool was_frozen;
7128
7129 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7130 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7131 highest_zoneidx);
7132
7133kswapd_try_sleep:
7134 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7135 highest_zoneidx);
7136
7137 /* Read the new order and highest_zoneidx */
7138 alloc_order = READ_ONCE(pgdat->kswapd_order);
7139 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7140 highest_zoneidx);
7141 WRITE_ONCE(pgdat->kswapd_order, 0);
7142 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7143
7144 if (kthread_freezable_should_stop(&was_frozen))
7145 break;
7146
7147 /*
7148 * We can speed up thawing tasks if we don't call balance_pgdat
7149 * after returning from the refrigerator
7150 */
7151 if (was_frozen)
7152 continue;
7153
7154 /*
7155 * Reclaim begins at the requested order but if a high-order
7156 * reclaim fails then kswapd falls back to reclaiming for
7157 * order-0. If that happens, kswapd will consider sleeping
7158 * for the order it finished reclaiming at (reclaim_order)
7159 * but kcompactd is woken to compact for the original
7160 * request (alloc_order).
7161 */
7162 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7163 alloc_order);
7164 reclaim_order = balance_pgdat(pgdat, alloc_order,
7165 highest_zoneidx);
7166 if (reclaim_order < alloc_order)
7167 goto kswapd_try_sleep;
7168 }
7169
7170 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7171
7172 return 0;
7173}
7174
7175/*
7176 * A zone is low on free memory or too fragmented for high-order memory. If
7177 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7178 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7179 * has failed or is not needed, still wake up kcompactd if only compaction is
7180 * needed.
7181 */
7182void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7183 enum zone_type highest_zoneidx)
7184{
7185 pg_data_t *pgdat;
7186 enum zone_type curr_idx;
7187
7188 if (!managed_zone(zone))
7189 return;
7190
7191 if (!cpuset_zone_allowed(zone, gfp_flags))
7192 return;
7193
7194 pgdat = zone->zone_pgdat;
7195 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7196
7197 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7198 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7199
7200 if (READ_ONCE(pgdat->kswapd_order) < order)
7201 WRITE_ONCE(pgdat->kswapd_order, order);
7202
7203 if (!waitqueue_active(&pgdat->kswapd_wait))
7204 return;
7205
7206 /* Hopeless node, leave it to direct reclaim if possible */
7207 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7208 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7209 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7210 /*
7211 * There may be plenty of free memory available, but it's too
7212 * fragmented for high-order allocations. Wake up kcompactd
7213 * and rely on compaction_suitable() to determine if it's
7214 * needed. If it fails, it will defer subsequent attempts to
7215 * ratelimit its work.
7216 */
7217 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7218 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7219 return;
7220 }
7221
7222 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7223 gfp_flags);
7224 wake_up_interruptible(&pgdat->kswapd_wait);
7225}
7226
7227#ifdef CONFIG_HIBERNATION
7228/*
7229 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7230 * freed pages.
7231 *
7232 * Rather than trying to age LRUs the aim is to preserve the overall
7233 * LRU order by reclaiming preferentially
7234 * inactive > active > active referenced > active mapped
7235 */
7236unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7237{
7238 struct scan_control sc = {
7239 .nr_to_reclaim = nr_to_reclaim,
7240 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7241 .reclaim_idx = MAX_NR_ZONES - 1,
7242 .priority = DEF_PRIORITY,
7243 .may_writepage = 1,
7244 .may_unmap = 1,
7245 .may_swap = 1,
7246 .hibernation_mode = 1,
7247 };
7248 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7249 unsigned long nr_reclaimed;
7250 unsigned int noreclaim_flag;
7251
7252 fs_reclaim_acquire(sc.gfp_mask);
7253 noreclaim_flag = memalloc_noreclaim_save();
7254 set_task_reclaim_state(current, &sc.reclaim_state);
7255
7256 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7257
7258 set_task_reclaim_state(current, NULL);
7259 memalloc_noreclaim_restore(noreclaim_flag);
7260 fs_reclaim_release(sc.gfp_mask);
7261
7262 return nr_reclaimed;
7263}
7264#endif /* CONFIG_HIBERNATION */
7265
7266/*
7267 * This kswapd start function will be called by init and node-hot-add.
7268 */
7269void __meminit kswapd_run(int nid)
7270{
7271 pg_data_t *pgdat = NODE_DATA(nid);
7272
7273 pgdat_kswapd_lock(pgdat);
7274 if (!pgdat->kswapd) {
7275 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7276 if (IS_ERR(pgdat->kswapd)) {
7277 /* failure at boot is fatal */
7278 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7279 nid, PTR_ERR(pgdat->kswapd));
7280 BUG_ON(system_state < SYSTEM_RUNNING);
7281 pgdat->kswapd = NULL;
7282 }
7283 }
7284 pgdat_kswapd_unlock(pgdat);
7285}
7286
7287/*
7288 * Called by memory hotplug when all memory in a node is offlined. Caller must
7289 * be holding mem_hotplug_begin/done().
7290 */
7291void __meminit kswapd_stop(int nid)
7292{
7293 pg_data_t *pgdat = NODE_DATA(nid);
7294 struct task_struct *kswapd;
7295
7296 pgdat_kswapd_lock(pgdat);
7297 kswapd = pgdat->kswapd;
7298 if (kswapd) {
7299 kthread_stop(kswapd);
7300 pgdat->kswapd = NULL;
7301 }
7302 pgdat_kswapd_unlock(pgdat);
7303}
7304
7305static int __init kswapd_init(void)
7306{
7307 int nid;
7308
7309 swap_setup();
7310 for_each_node_state(nid, N_MEMORY)
7311 kswapd_run(nid);
7312 return 0;
7313}
7314
7315module_init(kswapd_init)
7316
7317#ifdef CONFIG_NUMA
7318/*
7319 * Node reclaim mode
7320 *
7321 * If non-zero call node_reclaim when the number of free pages falls below
7322 * the watermarks.
7323 */
7324int node_reclaim_mode __read_mostly;
7325
7326/*
7327 * Priority for NODE_RECLAIM. This determines the fraction of pages
7328 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7329 * a zone.
7330 */
7331#define NODE_RECLAIM_PRIORITY 4
7332
7333/*
7334 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7335 * occur.
7336 */
7337int sysctl_min_unmapped_ratio = 1;
7338
7339/*
7340 * If the number of slab pages in a zone grows beyond this percentage then
7341 * slab reclaim needs to occur.
7342 */
7343int sysctl_min_slab_ratio = 5;
7344
7345static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7346{
7347 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7348 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7349 node_page_state(pgdat, NR_ACTIVE_FILE);
7350
7351 /*
7352 * It's possible for there to be more file mapped pages than
7353 * accounted for by the pages on the file LRU lists because
7354 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7355 */
7356 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7357}
7358
7359/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7360static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7361{
7362 unsigned long nr_pagecache_reclaimable;
7363 unsigned long delta = 0;
7364
7365 /*
7366 * If RECLAIM_UNMAP is set, then all file pages are considered
7367 * potentially reclaimable. Otherwise, we have to worry about
7368 * pages like swapcache and node_unmapped_file_pages() provides
7369 * a better estimate
7370 */
7371 if (node_reclaim_mode & RECLAIM_UNMAP)
7372 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7373 else
7374 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7375
7376 /* If we can't clean pages, remove dirty pages from consideration */
7377 if (!(node_reclaim_mode & RECLAIM_WRITE))
7378 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7379
7380 /* Watch for any possible underflows due to delta */
7381 if (unlikely(delta > nr_pagecache_reclaimable))
7382 delta = nr_pagecache_reclaimable;
7383
7384 return nr_pagecache_reclaimable - delta;
7385}
7386
7387/*
7388 * Try to free up some pages from this node through reclaim.
7389 */
7390static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7391{
7392 /* Minimum pages needed in order to stay on node */
7393 const unsigned long nr_pages = 1 << order;
7394 struct task_struct *p = current;
7395 unsigned int noreclaim_flag;
7396 struct scan_control sc = {
7397 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7398 .gfp_mask = current_gfp_context(gfp_mask),
7399 .order = order,
7400 .priority = NODE_RECLAIM_PRIORITY,
7401 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7402 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7403 .may_swap = 1,
7404 .reclaim_idx = gfp_zone(gfp_mask),
7405 };
7406 unsigned long pflags;
7407
7408 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7409 sc.gfp_mask);
7410
7411 cond_resched();
7412 psi_memstall_enter(&pflags);
7413 delayacct_freepages_start();
7414 fs_reclaim_acquire(sc.gfp_mask);
7415 /*
7416 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7417 */
7418 noreclaim_flag = memalloc_noreclaim_save();
7419 set_task_reclaim_state(p, &sc.reclaim_state);
7420
7421 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7422 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7423 /*
7424 * Free memory by calling shrink node with increasing
7425 * priorities until we have enough memory freed.
7426 */
7427 do {
7428 shrink_node(pgdat, &sc);
7429 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7430 }
7431
7432 set_task_reclaim_state(p, NULL);
7433 memalloc_noreclaim_restore(noreclaim_flag);
7434 fs_reclaim_release(sc.gfp_mask);
7435 psi_memstall_leave(&pflags);
7436 delayacct_freepages_end();
7437
7438 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7439
7440 return sc.nr_reclaimed >= nr_pages;
7441}
7442
7443int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7444{
7445 int ret;
7446
7447 /*
7448 * Node reclaim reclaims unmapped file backed pages and
7449 * slab pages if we are over the defined limits.
7450 *
7451 * A small portion of unmapped file backed pages is needed for
7452 * file I/O otherwise pages read by file I/O will be immediately
7453 * thrown out if the node is overallocated. So we do not reclaim
7454 * if less than a specified percentage of the node is used by
7455 * unmapped file backed pages.
7456 */
7457 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7458 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7459 pgdat->min_slab_pages)
7460 return NODE_RECLAIM_FULL;
7461
7462 /*
7463 * Do not scan if the allocation should not be delayed.
7464 */
7465 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7466 return NODE_RECLAIM_NOSCAN;
7467
7468 /*
7469 * Only run node reclaim on the local node or on nodes that do not
7470 * have associated processors. This will favor the local processor
7471 * over remote processors and spread off node memory allocations
7472 * as wide as possible.
7473 */
7474 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7475 return NODE_RECLAIM_NOSCAN;
7476
7477 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7478 return NODE_RECLAIM_NOSCAN;
7479
7480 ret = __node_reclaim(pgdat, gfp_mask, order);
7481 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7482
7483 if (!ret)
7484 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7485
7486 return ret;
7487}
7488#endif
7489
7490/**
7491 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7492 * lru list
7493 * @fbatch: Batch of lru folios to check.
7494 *
7495 * Checks folios for evictability, if an evictable folio is in the unevictable
7496 * lru list, moves it to the appropriate evictable lru list. This function
7497 * should be only used for lru folios.
7498 */
7499void check_move_unevictable_folios(struct folio_batch *fbatch)
7500{
7501 struct lruvec *lruvec = NULL;
7502 int pgscanned = 0;
7503 int pgrescued = 0;
7504 int i;
7505
7506 for (i = 0; i < fbatch->nr; i++) {
7507 struct folio *folio = fbatch->folios[i];
7508 int nr_pages = folio_nr_pages(folio);
7509
7510 pgscanned += nr_pages;
7511
7512 /* block memcg migration while the folio moves between lrus */
7513 if (!folio_test_clear_lru(folio))
7514 continue;
7515
7516 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7517 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7518 lruvec_del_folio(lruvec, folio);
7519 folio_clear_unevictable(folio);
7520 lruvec_add_folio(lruvec, folio);
7521 pgrescued += nr_pages;
7522 }
7523 folio_set_lru(folio);
7524 }
7525
7526 if (lruvec) {
7527 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7528 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7529 unlock_page_lruvec_irq(lruvec);
7530 } else if (pgscanned) {
7531 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7532 }
7533}
7534EXPORT_SYMBOL_GPL(check_move_unevictable_folios);