1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4	*
5	* Swap reorganised 29.12.95, Stephen Tweedie.
6	* kswapd added: 7.1.96 sct
7	* Removed kswapd_ctl limits, and swap out as many pages as needed
8	* to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9	* Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10	* Multiqueue VM started 5.8.00, Rik van Riel.
11	*/
12
13	#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15	#include <linux/mm.h>
16	#include <linux/sched/mm.h>
17	#include <linux/module.h>
18	#include <linux/gfp.h>
19	#include <linux/kernel_stat.h>
20	#include <linux/swap.h>
21	#include <linux/pagemap.h>
22	#include <linux/init.h>
23	#include <linux/highmem.h>
24	#include <linux/vmpressure.h>
25	#include <linux/vmstat.h>
26	#include <linux/file.h>
27	#include <linux/writeback.h>
28	#include <linux/blkdev.h>
29	#include <linux/buffer_head.h> /* for buffer_heads_over_limit */
30	#include <linux/mm_inline.h>
31	#include <linux/backing-dev.h>
32	#include <linux/rmap.h>
33	#include <linux/topology.h>
34	#include <linux/cpu.h>
35	#include <linux/cpuset.h>
36	#include <linux/compaction.h>
37	#include <linux/notifier.h>
38	#include <linux/delay.h>
39	#include <linux/kthread.h>
40	#include <linux/freezer.h>
41	#include <linux/memcontrol.h>
42	#include <linux/migrate.h>
43	#include <linux/delayacct.h>
44	#include <linux/sysctl.h>
45	#include <linux/memory-tiers.h>
46	#include <linux/oom.h>
47	#include <linux/pagevec.h>
48	#include <linux/prefetch.h>
49	#include <linux/printk.h>
50	#include <linux/dax.h>
51	#include <linux/psi.h>
52	#include <linux/pagewalk.h>
53	#include <linux/shmem_fs.h>
54	#include <linux/ctype.h>
55	#include <linux/debugfs.h>
56	#include <linux/khugepaged.h>
57	#include <linux/rculist_nulls.h>
58	#include <linux/random.h>
59	#include <linux/mmu_notifier.h>
60
61	#include <asm/tlbflush.h>
62	#include <asm/div64.h>
63
64	#include <linux/swapops.h>
65	#include <linux/balloon_compaction.h>
66	#include <linux/sched/sysctl.h>
67
68	#include "internal.h"
69	#include "swap.h"
70
71	#define CREATE_TRACE_POINTS
72	#include <trace/events/vmscan.h>
73
74	struct scan_control {
75	/* How many pages shrink_list() should reclaim */
76	unsigned long nr_to_reclaim;
77
78	/*
79	* Nodemask of nodes allowed by the caller. If NULL, all nodes
80	* are scanned.
81	*/
82	nodemask_t *nodemask;
83
84	/*
85	* The memory cgroup that hit its limit and as a result is the
86	* primary target of this reclaim invocation.
87	*/
88	struct mem_cgroup *target_mem_cgroup;
89
90	/*
91	* Scan pressure balancing between anon and file LRUs
92	*/
93	unsigned long anon_cost;
94	unsigned long file_cost;
95
96	#ifdef CONFIG_MEMCG
97	/* Swappiness value for proactive reclaim. Always use sc_swappiness()! */
98	int *proactive_swappiness;
99	#endif
100
101	/* Can active folios be deactivated as part of reclaim? */
102	#define DEACTIVATE_ANON 1
103	#define DEACTIVATE_FILE 2
104	unsigned int may_deactivate:2;
105	unsigned int force_deactivate:1;
106	unsigned int skipped_deactivate:1;
107
108	/* Writepage batching in laptop mode; RECLAIM_WRITE */
109	unsigned int may_writepage:1;
110
111	/* Can mapped folios be reclaimed? */
112	unsigned int may_unmap:1;
113
114	/* Can folios be swapped as part of reclaim? */
115	unsigned int may_swap:1;
116
117	/* Not allow cache_trim_mode to be turned on as part of reclaim? */
118	unsigned int no_cache_trim_mode:1;
119
120	/* Has cache_trim_mode failed at least once? */
121	unsigned int cache_trim_mode_failed:1;
122
123	/* Proactive reclaim invoked by userspace through memory.reclaim */
124	unsigned int proactive:1;
125
126	/*
127	* Cgroup memory below memory.low is protected as long as we
128	* don't threaten to OOM. If any cgroup is reclaimed at
129	* reduced force or passed over entirely due to its memory.low
130	* setting (memcg_low_skipped), and nothing is reclaimed as a
131	* result, then go back for one more cycle that reclaims the protected
132	* memory (memcg_low_reclaim) to avert OOM.
133	*/
134	unsigned int memcg_low_reclaim:1;
135	unsigned int memcg_low_skipped:1;
136
137	/* Shared cgroup tree walk failed, rescan the whole tree */
138	unsigned int memcg_full_walk:1;
139
140	unsigned int hibernation_mode:1;
141
142	/* One of the zones is ready for compaction */
143	unsigned int compaction_ready:1;
144
145	/* There is easily reclaimable cold cache in the current node */
146	unsigned int cache_trim_mode:1;
147
148	/* The file folios on the current node are dangerously low */
149	unsigned int file_is_tiny:1;
150
151	/* Always discard instead of demoting to lower tier memory */
152	unsigned int no_demotion:1;
153
154	/* Allocation order */
155	s8 order;
156
157	/* Scan (total_size >> priority) pages at once */
158	s8 priority;
159
160	/* The highest zone to isolate folios for reclaim from */
161	s8 reclaim_idx;
162
163	/* This context's GFP mask */
164	gfp_t gfp_mask;
165
166	/* Incremented by the number of inactive pages that were scanned */
167	unsigned long nr_scanned;
168
169	/* Number of pages freed so far during a call to shrink_zones() */
170	unsigned long nr_reclaimed;
171
172	struct {
173	unsigned int dirty;
174	unsigned int unqueued_dirty;
175	unsigned int congested;
176	unsigned int writeback;
177	unsigned int immediate;
178	unsigned int file_taken;
179	unsigned int taken;
180	} nr;
181
182	/* for recording the reclaimed slab by now */
183	struct reclaim_state reclaim_state;
184	};
185
186	#ifdef ARCH_HAS_PREFETCHW
187	#define prefetchw_prev_lru_folio(_folio, _base, _field) \
188	do { \
189	if ((_folio)->lru.prev != _base) { \
190	struct folio *prev; \
191	\
192	prev = lru_to_folio(&(_folio->lru)); \
193	prefetchw(&prev->_field); \
194	} \
195	} while (0)
196	#else
197	#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
198	#endif
199
200	/*
201	* From 0 .. MAX_SWAPPINESS. Higher means more swappy.
202	*/
203	int vm_swappiness = 60;
204
205	#ifdef CONFIG_MEMCG
206
207	/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
208	static bool cgroup_reclaim(struct scan_control *sc)
209	{
210	return sc->target_mem_cgroup;
211	}
212
213	/*
214	* Returns true for reclaim on the root cgroup. This is true for direct
215	* allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
216	*/
217	static bool root_reclaim(struct scan_control *sc)
218	{
219	return !sc->target_mem_cgroup || mem_cgroup_is_root(memcg: sc->target_mem_cgroup);
220	}
221
222	/**
223	* writeback_throttling_sane - is the usual dirty throttling mechanism available?
224	* @sc: scan_control in question
225	*
226	* The normal page dirty throttling mechanism in balance_dirty_pages() is
227	* completely broken with the legacy memcg and direct stalling in
228	* shrink_folio_list() is used for throttling instead, which lacks all the
229	* niceties such as fairness, adaptive pausing, bandwidth proportional
230	* allocation and configurability.
231	*
232	* This function tests whether the vmscan currently in progress can assume
233	* that the normal dirty throttling mechanism is operational.
234	*/
235	static bool writeback_throttling_sane(struct scan_control *sc)
236	{
237	if (!cgroup_reclaim(sc))
238	return true;
239	#ifdef CONFIG_CGROUP_WRITEBACK
240	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
241	return true;
242	#endif
243	return false;
244	}
245
246	static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
247	{
248	if (sc->proactive && sc->proactive_swappiness)
249	return *sc->proactive_swappiness;
250	return mem_cgroup_swappiness(memcg);
251	}
252	#else
253	static bool cgroup_reclaim(struct scan_control *sc)
254	{
255	return false;
256	}
257
258	static bool root_reclaim(struct scan_control *sc)
259	{
260	return true;
261	}
262
263	static bool writeback_throttling_sane(struct scan_control *sc)
264	{
265	return true;
266	}
267
268	static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
269	{
270	return READ_ONCE(vm_swappiness);
271	}
272	#endif
273
274	/* for_each_managed_zone_pgdat - helper macro to iterate over all managed zones in a pgdat up to
275	* and including the specified highidx
276	* @zone: The current zone in the iterator
277	* @pgdat: The pgdat which node_zones are being iterated
278	* @idx: The index variable
279	* @highidx: The index of the highest zone to return
280	*
281	* This macro iterates through all managed zones up to and including the specified highidx.
282	* The zone iterator enters an invalid state after macro call and must be reinitialized
283	* before it can be used again.
284	*/
285	#define for_each_managed_zone_pgdat(zone, pgdat, idx, highidx) \
286	for ((idx) = 0, (zone) = (pgdat)->node_zones; \
287	(idx) <= (highidx); \
288	(idx)++, (zone)++) \
289	if (!managed_zone(zone)) \
290	continue; \
291	else
292
293	static void set_task_reclaim_state(struct task_struct *task,
294	struct reclaim_state *rs)
295	{
296	/* Check for an overwrite */
297	WARN_ON_ONCE(rs && task->reclaim_state);
298
299	/* Check for the nulling of an already-nulled member */
300	WARN_ON_ONCE(!rs && !task->reclaim_state);
301
302	task->reclaim_state = rs;
303	}
304
305	/*
306	* flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
307	* scan_control->nr_reclaimed.
308	*/
309	static void flush_reclaim_state(struct scan_control *sc)
310	{
311	/*
312	* Currently, reclaim_state->reclaimed includes three types of pages
313	* freed outside of vmscan:
314	* (1) Slab pages.
315	* (2) Clean file pages from pruned inodes (on highmem systems).
316	* (3) XFS freed buffer pages.
317	*
318	* For all of these cases, we cannot universally link the pages to a
319	* single memcg. For example, a memcg-aware shrinker can free one object
320	* charged to the target memcg, causing an entire page to be freed.
321	* If we count the entire page as reclaimed from the memcg, we end up
322	* overestimating the reclaimed amount (potentially under-reclaiming).
323	*
324	* Only count such pages for global reclaim to prevent under-reclaiming
325	* from the target memcg; preventing unnecessary retries during memcg
326	* charging and false positives from proactive reclaim.
327	*
328	* For uncommon cases where the freed pages were actually mostly
329	* charged to the target memcg, we end up underestimating the reclaimed
330	* amount. This should be fine. The freed pages will be uncharged
331	* anyway, even if they are not counted here properly, and we will be
332	* able to make forward progress in charging (which is usually in a
333	* retry loop).
334	*
335	* We can go one step further, and report the uncharged objcg pages in
336	* memcg reclaim, to make reporting more accurate and reduce
337	* underestimation, but it's probably not worth the complexity for now.
338	*/
339	if (current->reclaim_state && root_reclaim(sc)) {
340	sc->nr_reclaimed += current->reclaim_state->reclaimed;
341	current->reclaim_state->reclaimed = 0;
342	}
343	}
344
345	static bool can_demote(int nid, struct scan_control *sc,
346	struct mem_cgroup *memcg)
347	{
348	int demotion_nid;
349
350	if (!numa_demotion_enabled)
351	return false;
352	if (sc && sc->no_demotion)
353	return false;
354
355	demotion_nid = next_demotion_node(node: nid);
356	if (demotion_nid == NUMA_NO_NODE)
357	return false;
358
359	/* If demotion node isn't in the cgroup's mems_allowed, fall back */
360	return mem_cgroup_node_allowed(memcg, nid: demotion_nid);
361	}
362
363	static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
364	int nid,
365	struct scan_control *sc)
366	{
367	if (memcg == NULL) {
368	/*
369	* For non-memcg reclaim, is there
370	* space in any swap device?
371	*/
372	if (get_nr_swap_pages() > 0)
373	return true;
374	} else {
375	/* Is the memcg below its swap limit? */
376	if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
377	return true;
378	}
379
380	/*
381	* The page can not be swapped.
382	*
383	* Can it be reclaimed from this node via demotion?
384	*/
385	return can_demote(nid, sc, memcg);
386	}
387
388	/*
389	* This misses isolated folios which are not accounted for to save counters.
390	* As the data only determines if reclaim or compaction continues, it is
391	* not expected that isolated folios will be a dominating factor.
392	*/
393	unsigned long zone_reclaimable_pages(struct zone *zone)
394	{
395	unsigned long nr;
396
397	nr = zone_page_state_snapshot(zone, item: NR_ZONE_INACTIVE_FILE) +
398	zone_page_state_snapshot(zone, item: NR_ZONE_ACTIVE_FILE);
399	if (can_reclaim_anon_pages(NULL, nid: zone_to_nid(zone), NULL))
400	nr += zone_page_state_snapshot(zone, item: NR_ZONE_INACTIVE_ANON) +
401	zone_page_state_snapshot(zone, item: NR_ZONE_ACTIVE_ANON);
402	/*
403	* If there are no reclaimable file-backed or anonymous pages,
404	* ensure zones with sufficient free pages are not skipped.
405	* This prevents zones like DMA32 from being ignored in reclaim
406	* scenarios where they can still help alleviate memory pressure.
407	*/
408	if (nr == 0)
409	nr = zone_page_state_snapshot(zone, item: NR_FREE_PAGES);
410	return nr;
411	}
412
413	/**
414	* lruvec_lru_size - Returns the number of pages on the given LRU list.
415	* @lruvec: lru vector
416	* @lru: lru to use
417	* @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
418	*/
419	static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
420	int zone_idx)
421	{
422	unsigned long size = 0;
423	int zid;
424	struct zone *zone;
425
426	for_each_managed_zone_pgdat(zone, lruvec_pgdat(lruvec), zid, zone_idx) {
427	if (!mem_cgroup_disabled())
428	size += mem_cgroup_get_zone_lru_size(lruvec, lru, zone_idx: zid);
429	else
430	size += zone_page_state(zone, item: NR_ZONE_LRU_BASE + lru);
431	}
432	return size;
433	}
434
435	static unsigned long drop_slab_node(int nid)
436	{
437	unsigned long freed = 0;
438	struct mem_cgroup *memcg = NULL;
439
440	memcg = mem_cgroup_iter(NULL, NULL, NULL);
441	do {
442	freed += shrink_slab(GFP_KERNEL, nid, memcg, priority: 0);
443	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
444
445	return freed;
446	}
447
448	void drop_slab(void)
449	{
450	int nid;
451	int shift = 0;
452	unsigned long freed;
453
454	do {
455	freed = 0;
456	for_each_online_node(nid) {
457	if (fatal_signal_pending(current))
458	return;
459
460	freed += drop_slab_node(nid);
461	}
462	} while ((freed >> shift++) > 1);
463	}
464
465	#define CHECK_RECLAIMER_OFFSET(type) \
466	do { \
467	BUILD_BUG_ON(PGSTEAL_##type - PGSTEAL_KSWAPD != \
468	PGDEMOTE_##type - PGDEMOTE_KSWAPD); \
469	BUILD_BUG_ON(PGSTEAL_##type - PGSTEAL_KSWAPD != \
470	PGSCAN_##type - PGSCAN_KSWAPD); \
471	} while (0)
472
473	static int reclaimer_offset(struct scan_control *sc)
474	{
475	CHECK_RECLAIMER_OFFSET(DIRECT);
476	CHECK_RECLAIMER_OFFSET(KHUGEPAGED);
477	CHECK_RECLAIMER_OFFSET(PROACTIVE);
478
479	if (current_is_kswapd())
480	return 0;
481	if (current_is_khugepaged())
482	return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
483	if (sc->proactive)
484	return PGSTEAL_PROACTIVE - PGSTEAL_KSWAPD;
485	return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
486	}
487
488	static inline int is_page_cache_freeable(struct folio *folio)
489	{
490	/*
491	* A freeable page cache folio is referenced only by the caller
492	* that isolated the folio, the page cache and optional filesystem
493	* private data at folio->private.
494	*/
495	return folio_ref_count(folio) - folio_test_private(folio) ==
496	1 + folio_nr_pages(folio);
497	}
498
499	/*
500	* We detected a synchronous write error writing a folio out. Probably
501	* -ENOSPC. We need to propagate that into the address_space for a subsequent
502	* fsync(), msync() or close().
503	*
504	* The tricky part is that after writepage we cannot touch the mapping: nothing
505	* prevents it from being freed up. But we have a ref on the folio and once
506	* that folio is locked, the mapping is pinned.
507	*
508	* We're allowed to run sleeping folio_lock() here because we know the caller has
509	* __GFP_FS.
510	*/
511	static void handle_write_error(struct address_space *mapping,
512	struct folio *folio, int error)
513	{
514	folio_lock(folio);
515	if (folio_mapping(folio) == mapping)
516	mapping_set_error(mapping, error);
517	folio_unlock(folio);
518	}
519
520	static bool skip_throttle_noprogress(pg_data_t *pgdat)
521	{
522	int reclaimable = 0, write_pending = 0;
523	int i;
524	struct zone *zone;
525	/*
526	* If kswapd is disabled, reschedule if necessary but do not
527	* throttle as the system is likely near OOM.
528	*/
529	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
530	return true;
531
532	/*
533	* If there are a lot of dirty/writeback folios then do not
534	* throttle as throttling will occur when the folios cycle
535	* towards the end of the LRU if still under writeback.
536	*/
537	for_each_managed_zone_pgdat(zone, pgdat, i, MAX_NR_ZONES - 1) {
538	reclaimable += zone_reclaimable_pages(zone);
539	write_pending += zone_page_state_snapshot(zone,
540	item: NR_ZONE_WRITE_PENDING);
541	}
542	if (2 * write_pending <= reclaimable)
543	return true;
544
545	return false;
546	}
547
548	void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
549	{
550	wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
551	long timeout, ret;
552	DEFINE_WAIT(wait);
553
554	/*
555	* Do not throttle user workers, kthreads other than kswapd or
556	* workqueues. They may be required for reclaim to make
557	* forward progress (e.g. journalling workqueues or kthreads).
558	*/
559	if (!current_is_kswapd() &&
560	current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
561	cond_resched();
562	return;
563	}
564
565	/*
566	* These figures are pulled out of thin air.
567	* VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
568	* parallel reclaimers which is a short-lived event so the timeout is
569	* short. Failing to make progress or waiting on writeback are
570	* potentially long-lived events so use a longer timeout. This is shaky
571	* logic as a failure to make progress could be due to anything from
572	* writeback to a slow device to excessive referenced folios at the tail
573	* of the inactive LRU.
574	*/
575	switch(reason) {
576	case VMSCAN_THROTTLE_WRITEBACK:
577	timeout = HZ/10;
578
579	if (atomic_inc_return(v: &pgdat->nr_writeback_throttled) == 1) {
580	WRITE_ONCE(pgdat->nr_reclaim_start,
581	node_page_state(pgdat, NR_THROTTLED_WRITTEN));
582	}
583
584	break;
585	case VMSCAN_THROTTLE_CONGESTED:
586	fallthrough;
587	case VMSCAN_THROTTLE_NOPROGRESS:
588	if (skip_throttle_noprogress(pgdat)) {
589	cond_resched();
590	return;
591	}
592
593	timeout = 1;
594
595	break;
596	case VMSCAN_THROTTLE_ISOLATED:
597	timeout = HZ/50;
598	break;
599	default:
600	WARN_ON_ONCE(1);
601	timeout = HZ;
602	break;
603	}
604
605	prepare_to_wait(wq_head: wqh, wq_entry: &wait, TASK_UNINTERRUPTIBLE);
606	ret = schedule_timeout(timeout);
607	finish_wait(wq_head: wqh, wq_entry: &wait);
608
609	if (reason == VMSCAN_THROTTLE_WRITEBACK)
610	atomic_dec(v: &pgdat->nr_writeback_throttled);
611
612	trace_mm_vmscan_throttled(nid: pgdat->node_id, usec_timeout: jiffies_to_usecs(j: timeout),
613	usec_delayed: jiffies_to_usecs(j: timeout - ret),
614	reason);
615	}
616
617	/*
618	* Account for folios written if tasks are throttled waiting on dirty
619	* folios to clean. If enough folios have been cleaned since throttling
620	* started then wakeup the throttled tasks.
621	*/
622	void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
623	int nr_throttled)
624	{
625	unsigned long nr_written;
626
627	node_stat_add_folio(folio, item: NR_THROTTLED_WRITTEN);
628
629	/*
630	* This is an inaccurate read as the per-cpu deltas may not
631	* be synchronised. However, given that the system is
632	* writeback throttled, it is not worth taking the penalty
633	* of getting an accurate count. At worst, the throttle
634	* timeout guarantees forward progress.
635	*/
636	nr_written = node_page_state(pgdat, item: NR_THROTTLED_WRITTEN) -
637	READ_ONCE(pgdat->nr_reclaim_start);
638
639	if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
640	wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
641	}
642
643	/* possible outcome of pageout() */
644	typedef enum {
645	/* failed to write folio out, folio is locked */
646	PAGE_KEEP,
647	/* move folio to the active list, folio is locked */
648	PAGE_ACTIVATE,
649	/* folio has been sent to the disk successfully, folio is unlocked */
650	PAGE_SUCCESS,
651	/* folio is clean and locked */
652	PAGE_CLEAN,
653	} pageout_t;
654
655	/*
656	* pageout is called by shrink_folio_list() for each dirty folio.
657	*/
658	static pageout_t pageout(struct folio *folio, struct address_space *mapping,
659	struct swap_iocb **plug, struct list_head *folio_list)
660	{
661	int (*writeout)(struct folio *, struct writeback_control *);
662
663	/*
664	* We no longer attempt to writeback filesystem folios here, other
665	* than tmpfs/shmem. That's taken care of in page-writeback.
666	* If we find a dirty filesystem folio at the end of the LRU list,
667	* typically that means the filesystem is saturating the storage
668	* with contiguous writes and telling it to write a folio here
669	* would only make the situation worse by injecting an element
670	* of random access.
671	*
672	* If the folio is swapcache, write it back even if that would
673	* block, for some throttling. This happens by accident, because
674	* swap_backing_dev_info is bust: it doesn't reflect the
675	* congestion state of the swapdevs. Easy to fix, if needed.
676	*/
677	if (!is_page_cache_freeable(folio))
678	return PAGE_KEEP;
679	if (!mapping) {
680	/*
681	* Some data journaling orphaned folios can have
682	* folio->mapping == NULL while being dirty with clean buffers.
683	*/
684	if (folio_test_private(folio)) {
685	if (try_to_free_buffers(folio)) {
686	folio_clear_dirty(folio);
687	pr_info("%s: orphaned folio\n", __func__);
688	return PAGE_CLEAN;
689	}
690	}
691	return PAGE_KEEP;
692	}
693	if (shmem_mapping(mapping))
694	writeout = shmem_writeout;
695	else if (folio_test_anon(folio))
696	writeout = swap_writeout;
697	else
698	return PAGE_ACTIVATE;
699
700	if (folio_clear_dirty_for_io(folio)) {
701	int res;
702	struct writeback_control wbc = {
703	.sync_mode = WB_SYNC_NONE,
704	.nr_to_write = SWAP_CLUSTER_MAX,
705	.range_start = 0,
706	.range_end = LLONG_MAX,
707	.for_reclaim = 1,
708	.swap_plug = plug,
709	};
710
711	/*
712	* The large shmem folio can be split if CONFIG_THP_SWAP is
713	* not enabled or contiguous swap entries are failed to
714	* allocate.
715	*/
716	if (shmem_mapping(mapping) && folio_test_large(folio))
717	wbc.list = folio_list;
718
719	folio_set_reclaim(folio);
720	res = writeout(folio, &wbc);
721	if (res < 0)
722	handle_write_error(mapping, folio, error: res);
723	if (res == AOP_WRITEPAGE_ACTIVATE) {
724	folio_clear_reclaim(folio);
725	return PAGE_ACTIVATE;
726	}
727
728	if (!folio_test_writeback(folio)) {
729	/* synchronous write? */
730	folio_clear_reclaim(folio);
731	}
732	trace_mm_vmscan_write_folio(folio);
733	node_stat_add_folio(folio, item: NR_VMSCAN_WRITE);
734	return PAGE_SUCCESS;
735	}
736
737	return PAGE_CLEAN;
738	}
739
740	/*
741	* Same as remove_mapping, but if the folio is removed from the mapping, it
742	* gets returned with a refcount of 0.
743	*/
744	static int __remove_mapping(struct address_space *mapping, struct folio *folio,
745	bool reclaimed, struct mem_cgroup *target_memcg)
746	{
747	int refcount;
748	void *shadow = NULL;
749
750	BUG_ON(!folio_test_locked(folio));
751	BUG_ON(mapping != folio_mapping(folio));
752
753	if (!folio_test_swapcache(folio))
754	spin_lock(lock: &mapping->host->i_lock);
755	xa_lock_irq(&mapping->i_pages);
756	/*
757	* The non racy check for a busy folio.
758	*
759	* Must be careful with the order of the tests. When someone has
760	* a ref to the folio, it may be possible that they dirty it then
761	* drop the reference. So if the dirty flag is tested before the
762	* refcount here, then the following race may occur:
763	*
764	* get_user_pages(&page);
765	* [user mapping goes away]
766	* write_to(page);
767	* !folio_test_dirty(folio) [good]
768	* folio_set_dirty(folio);
769	* folio_put(folio);
770	* !refcount(folio) [good, discard it]
771	*
772	* [oops, our write_to data is lost]
773	*
774	* Reversing the order of the tests ensures such a situation cannot
775	* escape unnoticed. The smp_rmb is needed to ensure the folio->flags
776	* load is not satisfied before that of folio->_refcount.
777	*
778	* Note that if the dirty flag is always set via folio_mark_dirty,
779	* and thus under the i_pages lock, then this ordering is not required.
780	*/
781	refcount = 1 + folio_nr_pages(folio);
782	if (!folio_ref_freeze(folio, count: refcount))
783	goto cannot_free;
784	/* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
785	if (unlikely(folio_test_dirty(folio))) {
786	folio_ref_unfreeze(folio, count: refcount);
787	goto cannot_free;
788	}
789
790	if (folio_test_swapcache(folio)) {
791	swp_entry_t swap = folio->swap;
792
793	if (reclaimed && !mapping_exiting(mapping))
794	shadow = workingset_eviction(folio, target_memcg);
795	__delete_from_swap_cache(folio, entry: swap, shadow);
796	memcg1_swapout(folio, entry: swap);
797	xa_unlock_irq(&mapping->i_pages);
798	put_swap_folio(folio, entry: swap);
799	} else {
800	void (*free_folio)(struct folio *);
801
802	free_folio = mapping->a_ops->free_folio;
803	/*
804	* Remember a shadow entry for reclaimed file cache in
805	* order to detect refaults, thus thrashing, later on.
806	*
807	* But don't store shadows in an address space that is
808	* already exiting. This is not just an optimization,
809	* inode reclaim needs to empty out the radix tree or
810	* the nodes are lost. Don't plant shadows behind its
811	* back.
812	*
813	* We also don't store shadows for DAX mappings because the
814	* only page cache folios found in these are zero pages
815	* covering holes, and because we don't want to mix DAX
816	* exceptional entries and shadow exceptional entries in the
817	* same address_space.
818	*/
819	if (reclaimed && folio_is_file_lru(folio) &&
820	!mapping_exiting(mapping) && !dax_mapping(mapping))
821	shadow = workingset_eviction(folio, target_memcg);
822	__filemap_remove_folio(folio, shadow);
823	xa_unlock_irq(&mapping->i_pages);
824	if (mapping_shrinkable(mapping))
825	inode_add_lru(inode: mapping->host);
826	spin_unlock(lock: &mapping->host->i_lock);
827
828	if (free_folio)
829	free_folio(folio);
830	}
831
832	return 1;
833
834	cannot_free:
835	xa_unlock_irq(&mapping->i_pages);
836	if (!folio_test_swapcache(folio))
837	spin_unlock(lock: &mapping->host->i_lock);
838	return 0;
839	}
840
841	/**
842	* remove_mapping() - Attempt to remove a folio from its mapping.
843	* @mapping: The address space.
844	* @folio: The folio to remove.
845	*
846	* If the folio is dirty, under writeback or if someone else has a ref
847	* on it, removal will fail.
848	* Return: The number of pages removed from the mapping. 0 if the folio
849	* could not be removed.
850	* Context: The caller should have a single refcount on the folio and
851	* hold its lock.
852	*/
853	long remove_mapping(struct address_space *mapping, struct folio *folio)
854	{
855	if (__remove_mapping(mapping, folio, reclaimed: false, NULL)) {
856	/*
857	* Unfreezing the refcount with 1 effectively
858	* drops the pagecache ref for us without requiring another
859	* atomic operation.
860	*/
861	folio_ref_unfreeze(folio, count: 1);
862	return folio_nr_pages(folio);
863	}
864	return 0;
865	}
866
867	/**
868	* folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
869	* @folio: Folio to be returned to an LRU list.
870	*
871	* Add previously isolated @folio to appropriate LRU list.
872	* The folio may still be unevictable for other reasons.
873	*
874	* Context: lru_lock must not be held, interrupts must be enabled.
875	*/
876	void folio_putback_lru(struct folio *folio)
877	{
878	folio_add_lru(folio);
879	folio_put(folio); /* drop ref from isolate */
880	}
881
882	enum folio_references {
883	FOLIOREF_RECLAIM,
884	FOLIOREF_RECLAIM_CLEAN,
885	FOLIOREF_KEEP,
886	FOLIOREF_ACTIVATE,
887	};
888
889	#ifdef CONFIG_LRU_GEN
890	/*
891	* Only used on a mapped folio in the eviction (rmap walk) path, where promotion
892	* needs to be done by taking the folio off the LRU list and then adding it back
893	* with PG_active set. In contrast, the aging (page table walk) path uses
894	* folio_update_gen().
895	*/
896	static bool lru_gen_set_refs(struct folio *folio)
897	{
898	/* see the comment on LRU_REFS_FLAGS */
899	if (!folio_test_referenced(folio) && !folio_test_workingset(folio)) {
900	set_mask_bits(&folio->flags, LRU_REFS_MASK, BIT(PG_referenced));
901	return false;
902	}
903
904	set_mask_bits(&folio->flags, LRU_REFS_FLAGS, BIT(PG_workingset));
905	return true;
906	}
907	#else
908	static bool lru_gen_set_refs(struct folio *folio)
909	{
910	return false;
911	}
912	#endif /* CONFIG_LRU_GEN */
913
914	static enum folio_references folio_check_references(struct folio *folio,
915	struct scan_control *sc)
916	{
917	int referenced_ptes, referenced_folio;
918	unsigned long vm_flags;
919
920	referenced_ptes = folio_referenced(folio, is_locked: 1, memcg: sc->target_mem_cgroup,
921	vm_flags: &vm_flags);
922
923	/*
924	* The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
925	* Let the folio, now marked Mlocked, be moved to the unevictable list.
926	*/
927	if (vm_flags & VM_LOCKED)
928	return FOLIOREF_ACTIVATE;
929
930	/*
931	* There are two cases to consider.
932	* 1) Rmap lock contention: rotate.
933	* 2) Skip the non-shared swapbacked folio mapped solely by
934	* the exiting or OOM-reaped process.
935	*/
936	if (referenced_ptes == -1)
937	return FOLIOREF_KEEP;
938
939	if (lru_gen_enabled()) {
940	if (!referenced_ptes)
941	return FOLIOREF_RECLAIM;
942
943	return lru_gen_set_refs(folio) ? FOLIOREF_ACTIVATE : FOLIOREF_KEEP;
944	}
945
946	referenced_folio = folio_test_clear_referenced(folio);
947
948	if (referenced_ptes) {
949	/*
950	* All mapped folios start out with page table
951	* references from the instantiating fault, so we need
952	* to look twice if a mapped file/anon folio is used more
953	* than once.
954	*
955	* Mark it and spare it for another trip around the
956	* inactive list. Another page table reference will
957	* lead to its activation.
958	*
959	* Note: the mark is set for activated folios as well
960	* so that recently deactivated but used folios are
961	* quickly recovered.
962	*/
963	folio_set_referenced(folio);
964
965	if (referenced_folio || referenced_ptes > 1)
966	return FOLIOREF_ACTIVATE;
967
968	/*
969	* Activate file-backed executable folios after first usage.
970	*/
971	if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
972	return FOLIOREF_ACTIVATE;
973
974	return FOLIOREF_KEEP;
975	}
976
977	/* Reclaim if clean, defer dirty folios to writeback */
978	if (referenced_folio && folio_is_file_lru(folio))
979	return FOLIOREF_RECLAIM_CLEAN;
980
981	return FOLIOREF_RECLAIM;
982	}
983
984	/* Check if a folio is dirty or under writeback */
985	static void folio_check_dirty_writeback(struct folio *folio,
986	bool *dirty, bool *writeback)
987	{
988	struct address_space *mapping;
989
990	/*
991	* Anonymous folios are not handled by flushers and must be written
992	* from reclaim context. Do not stall reclaim based on them.
993	* MADV_FREE anonymous folios are put into inactive file list too.
994	* They could be mistakenly treated as file lru. So further anon
995	* test is needed.
996	*/
997	if (!folio_is_file_lru(folio) ||
998	(folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
999	*dirty = false;
1000	*writeback = false;
1001	return;
1002	}
1003
1004	/* By default assume that the folio flags are accurate */
1005	*dirty = folio_test_dirty(folio);
1006	*writeback = folio_test_writeback(folio);
1007
1008	/* Verify dirty/writeback state if the filesystem supports it */
1009	if (!folio_test_private(folio))
1010	return;
1011
1012	mapping = folio_mapping(folio);
1013	if (mapping && mapping->a_ops->is_dirty_writeback)
1014	mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1015	}
1016
1017	struct folio *alloc_migrate_folio(struct folio *src, unsigned long private)
1018	{
1019	struct folio *dst;
1020	nodemask_t *allowed_mask;
1021	struct migration_target_control *mtc;
1022
1023	mtc = (struct migration_target_control *)private;
1024
1025	allowed_mask = mtc->nmask;
1026	/*
1027	* make sure we allocate from the target node first also trying to
1028	* demote or reclaim pages from the target node via kswapd if we are
1029	* low on free memory on target node. If we don't do this and if
1030	* we have free memory on the slower(lower) memtier, we would start
1031	* allocating pages from slower(lower) memory tiers without even forcing
1032	* a demotion of cold pages from the target memtier. This can result
1033	* in the kernel placing hot pages in slower(lower) memory tiers.
1034	*/
1035	mtc->nmask = NULL;
1036	mtc->gfp_mask |= __GFP_THISNODE;
1037	dst = alloc_migration_target(src, private: (unsigned long)mtc);
1038	if (dst)
1039	return dst;
1040
1041	mtc->gfp_mask &= ~__GFP_THISNODE;
1042	mtc->nmask = allowed_mask;
1043
1044	return alloc_migration_target(src, private: (unsigned long)mtc);
1045	}
1046
1047	/*
1048	* Take folios on @demote_folios and attempt to demote them to another node.
1049	* Folios which are not demoted are left on @demote_folios.
1050	*/
1051	static unsigned int demote_folio_list(struct list_head *demote_folios,
1052	struct pglist_data *pgdat)
1053	{
1054	int target_nid = next_demotion_node(node: pgdat->node_id);
1055	unsigned int nr_succeeded;
1056	nodemask_t allowed_mask;
1057
1058	struct migration_target_control mtc = {
1059	/*
1060	* Allocate from 'node', or fail quickly and quietly.
1061	* When this happens, 'page' will likely just be discarded
1062	* instead of migrated.
1063	*/
1064	.gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1065	__GFP_NOMEMALLOC | GFP_NOWAIT,
1066	.nid = target_nid,
1067	.nmask = &allowed_mask,
1068	.reason = MR_DEMOTION,
1069	};
1070
1071	if (list_empty(head: demote_folios))
1072	return 0;
1073
1074	if (target_nid == NUMA_NO_NODE)
1075	return 0;
1076
1077	node_get_allowed_targets(pgdat, targets: &allowed_mask);
1078
1079	/* Demotion ignores all cpuset and mempolicy settings */
1080	migrate_pages(l: demote_folios, new: alloc_migrate_folio, NULL,
1081	private: (unsigned long)&mtc, mode: MIGRATE_ASYNC, reason: MR_DEMOTION,
1082	ret_succeeded: &nr_succeeded);
1083
1084	return nr_succeeded;
1085	}
1086
1087	static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1088	{
1089	if (gfp_mask & __GFP_FS)
1090	return true;
1091	if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1092	return false;
1093	/*
1094	* We can "enter_fs" for swap-cache with only __GFP_IO
1095	* providing this isn't SWP_FS_OPS.
1096	* ->flags can be updated non-atomicially (scan_swap_map_slots),
1097	* but that will never affect SWP_FS_OPS, so the data_race
1098	* is safe.
1099	*/
1100	return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1101	}
1102
1103	/*
1104	* shrink_folio_list() returns the number of reclaimed pages
1105	*/
1106	static unsigned int shrink_folio_list(struct list_head *folio_list,
1107	struct pglist_data *pgdat, struct scan_control *sc,
1108	struct reclaim_stat *stat, bool ignore_references,
1109	struct mem_cgroup *memcg)
1110	{
1111	struct folio_batch free_folios;
1112	LIST_HEAD(ret_folios);
1113	LIST_HEAD(demote_folios);
1114	unsigned int nr_reclaimed = 0, nr_demoted = 0;
1115	unsigned int pgactivate = 0;
1116	bool do_demote_pass;
1117	struct swap_iocb *plug = NULL;
1118
1119	folio_batch_init(fbatch: &free_folios);
1120	memset(stat, 0, sizeof(*stat));
1121	cond_resched();
1122	do_demote_pass = can_demote(nid: pgdat->node_id, sc, memcg);
1123
1124	retry:
1125	while (!list_empty(head: folio_list)) {
1126	struct address_space *mapping;
1127	struct folio *folio;
1128	enum folio_references references = FOLIOREF_RECLAIM;
1129	bool dirty, writeback;
1130	unsigned int nr_pages;
1131
1132	cond_resched();
1133
1134	folio = lru_to_folio(head: folio_list);
1135	list_del(entry: &folio->lru);
1136
1137	if (!folio_trylock(folio))
1138	goto keep;
1139
1140	if (folio_contain_hwpoisoned_page(folio)) {
1141	unmap_poisoned_folio(folio, pfn: folio_pfn(folio), must_kill: false);
1142	folio_unlock(folio);
1143	folio_put(folio);
1144	continue;
1145	}
1146
1147	VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1148
1149	nr_pages = folio_nr_pages(folio);
1150
1151	/* Account the number of base pages */
1152	sc->nr_scanned += nr_pages;
1153
1154	if (unlikely(!folio_evictable(folio)))
1155	goto activate_locked;
1156
1157	if (!sc->may_unmap && folio_mapped(folio))
1158	goto keep_locked;
1159
1160	/*
1161	* The number of dirty pages determines if a node is marked
1162	* reclaim_congested. kswapd will stall and start writing
1163	* folios if the tail of the LRU is all dirty unqueued folios.
1164	*/
1165	folio_check_dirty_writeback(folio, dirty: &dirty, writeback: &writeback);
1166	if (dirty || writeback)
1167	stat->nr_dirty += nr_pages;
1168
1169	if (dirty && !writeback)
1170	stat->nr_unqueued_dirty += nr_pages;
1171
1172	/*
1173	* Treat this folio as congested if folios are cycling
1174	* through the LRU so quickly that the folios marked
1175	* for immediate reclaim are making it to the end of
1176	* the LRU a second time.
1177	*/
1178	if (writeback && folio_test_reclaim(folio))
1179	stat->nr_congested += nr_pages;
1180
1181	/*
1182	* If a folio at the tail of the LRU is under writeback, there
1183	* are three cases to consider.
1184	*
1185	* 1) If reclaim is encountering an excessive number
1186	* of folios under writeback and this folio has both
1187	* the writeback and reclaim flags set, then it
1188	* indicates that folios are being queued for I/O but
1189	* are being recycled through the LRU before the I/O
1190	* can complete. Waiting on the folio itself risks an
1191	* indefinite stall if it is impossible to writeback
1192	* the folio due to I/O error or disconnected storage
1193	* so instead note that the LRU is being scanned too
1194	* quickly and the caller can stall after the folio
1195	* list has been processed.
1196	*
1197	* 2) Global or new memcg reclaim encounters a folio that is
1198	* not marked for immediate reclaim, or the caller does not
1199	* have __GFP_FS (or __GFP_IO if it's simply going to swap,
1200	* not to fs), or the folio belongs to a mapping where
1201	* waiting on writeback during reclaim may lead to a deadlock.
1202	* In this case mark the folio for immediate reclaim and
1203	* continue scanning.
1204	*
1205	* Require may_enter_fs() because we would wait on fs, which
1206	* may not have submitted I/O yet. And the loop driver might
1207	* enter reclaim, and deadlock if it waits on a folio for
1208	* which it is needed to do the write (loop masks off
1209	* __GFP_IO|__GFP_FS for this reason); but more thought
1210	* would probably show more reasons.
1211	*
1212	* 3) Legacy memcg encounters a folio that already has the
1213	* reclaim flag set. memcg does not have any dirty folio
1214	* throttling so we could easily OOM just because too many
1215	* folios are in writeback and there is nothing else to
1216	* reclaim. Wait for the writeback to complete.
1217	*
1218	* In cases 1) and 2) we activate the folios to get them out of
1219	* the way while we continue scanning for clean folios on the
1220	* inactive list and refilling from the active list. The
1221	* observation here is that waiting for disk writes is more
1222	* expensive than potentially causing reloads down the line.
1223	* Since they're marked for immediate reclaim, they won't put
1224	* memory pressure on the cache working set any longer than it
1225	* takes to write them to disk.
1226	*/
1227	if (folio_test_writeback(folio)) {
1228	mapping = folio_mapping(folio);
1229
1230	/* Case 1 above */
1231	if (current_is_kswapd() &&
1232	folio_test_reclaim(folio) &&
1233	test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1234	stat->nr_immediate += nr_pages;
1235	goto activate_locked;
1236
1237	/* Case 2 above */
1238	} else if (writeback_throttling_sane(sc) ||
1239	!folio_test_reclaim(folio) ||
1240	!may_enter_fs(folio, gfp_mask: sc->gfp_mask) ||
1241	(mapping &&
1242	mapping_writeback_may_deadlock_on_reclaim(mapping))) {
1243	/*
1244	* This is slightly racy -
1245	* folio_end_writeback() might have
1246	* just cleared the reclaim flag, then
1247	* setting the reclaim flag here ends up
1248	* interpreted as the readahead flag - but
1249	* that does not matter enough to care.
1250	* What we do want is for this folio to
1251	* have the reclaim flag set next time
1252	* memcg reclaim reaches the tests above,
1253	* so it will then wait for writeback to
1254	* avoid OOM; and it's also appropriate
1255	* in global reclaim.
1256	*/
1257	folio_set_reclaim(folio);
1258	stat->nr_writeback += nr_pages;
1259	goto activate_locked;
1260
1261	/* Case 3 above */
1262	} else {
1263	folio_unlock(folio);
1264	folio_wait_writeback(folio);
1265	/* then go back and try same folio again */
1266	list_add_tail(new: &folio->lru, head: folio_list);
1267	continue;
1268	}
1269	}
1270
1271	if (!ignore_references)
1272	references = folio_check_references(folio, sc);
1273
1274	switch (references) {
1275	case FOLIOREF_ACTIVATE:
1276	goto activate_locked;
1277	case FOLIOREF_KEEP:
1278	stat->nr_ref_keep += nr_pages;
1279	goto keep_locked;
1280	case FOLIOREF_RECLAIM:
1281	case FOLIOREF_RECLAIM_CLEAN:
1282	; /* try to reclaim the folio below */
1283	}
1284
1285	/*
1286	* Before reclaiming the folio, try to relocate
1287	* its contents to another node.
1288	*/
1289	if (do_demote_pass &&
1290	(thp_migration_supported() || !folio_test_large(folio))) {
1291	list_add(new: &folio->lru, head: &demote_folios);
1292	folio_unlock(folio);
1293	continue;
1294	}
1295
1296	/*
1297	* Anonymous process memory has backing store?
1298	* Try to allocate it some swap space here.
1299	* Lazyfree folio could be freed directly
1300	*/
1301	if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1302	if (!folio_test_swapcache(folio)) {
1303	if (!(sc->gfp_mask & __GFP_IO))
1304	goto keep_locked;
1305	if (folio_maybe_dma_pinned(folio))
1306	goto keep_locked;
1307	if (folio_test_large(folio)) {
1308	/* cannot split folio, skip it */
1309	if (!can_split_folio(folio, caller_pins: 1, NULL))
1310	goto activate_locked;
1311	/*
1312	* Split partially mapped folios right away.
1313	* We can free the unmapped pages without IO.
1314	*/
1315	if (data_race(!list_empty(&folio->_deferred_list) &&
1316	folio_test_partially_mapped(folio)) &&
1317	split_folio_to_list(folio, list: folio_list))
1318	goto activate_locked;
1319	}
1320	if (folio_alloc_swap(folio, __GFP_HIGH | __GFP_NOWARN)) {
1321	int __maybe_unused order = folio_order(folio);
1322
1323	if (!folio_test_large(folio))
1324	goto activate_locked_split;
1325	/* Fallback to swap normal pages */
1326	if (split_folio_to_list(folio, list: folio_list))
1327	goto activate_locked;
1328	#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1329	if (nr_pages >= HPAGE_PMD_NR) {
1330	count_memcg_folio_events(folio,
1331	idx: THP_SWPOUT_FALLBACK, nr: 1);
1332	count_vm_event(item: THP_SWPOUT_FALLBACK);
1333	}
1334	#endif
1335	count_mthp_stat(order, item: MTHP_STAT_SWPOUT_FALLBACK);
1336	if (folio_alloc_swap(folio, __GFP_HIGH | __GFP_NOWARN))
1337	goto activate_locked_split;
1338	}
1339	/*
1340	* Normally the folio will be dirtied in unmap because its
1341	* pte should be dirty. A special case is MADV_FREE page. The
1342	* page's pte could have dirty bit cleared but the folio's
1343	* SwapBacked flag is still set because clearing the dirty bit
1344	* and SwapBacked flag has no lock protected. For such folio,
1345	* unmap will not set dirty bit for it, so folio reclaim will
1346	* not write the folio out. This can cause data corruption when
1347	* the folio is swapped in later. Always setting the dirty flag
1348	* for the folio solves the problem.
1349	*/
1350	folio_mark_dirty(folio);
1351	}
1352	}
1353
1354	/*
1355	* If the folio was split above, the tail pages will make
1356	* their own pass through this function and be accounted
1357	* then.
1358	*/
1359	if ((nr_pages > 1) && !folio_test_large(folio)) {
1360	sc->nr_scanned -= (nr_pages - 1);
1361	nr_pages = 1;
1362	}
1363
1364	/*
1365	* The folio is mapped into the page tables of one or more
1366	* processes. Try to unmap it here.
1367	*/
1368	if (folio_mapped(folio)) {
1369	enum ttu_flags flags = TTU_BATCH_FLUSH;
1370	bool was_swapbacked = folio_test_swapbacked(folio);
1371
1372	if (folio_test_pmd_mappable(folio))
1373	flags |= TTU_SPLIT_HUGE_PMD;
1374	/*
1375	* Without TTU_SYNC, try_to_unmap will only begin to
1376	* hold PTL from the first present PTE within a large
1377	* folio. Some initial PTEs might be skipped due to
1378	* races with parallel PTE writes in which PTEs can be
1379	* cleared temporarily before being written new present
1380	* values. This will lead to a large folio is still
1381	* mapped while some subpages have been partially
1382	* unmapped after try_to_unmap; TTU_SYNC helps
1383	* try_to_unmap acquire PTL from the first PTE,
1384	* eliminating the influence of temporary PTE values.
1385	*/
1386	if (folio_test_large(folio))
1387	flags |= TTU_SYNC;
1388
1389	try_to_unmap(folio, flags);
1390	if (folio_mapped(folio)) {
1391	stat->nr_unmap_fail += nr_pages;
1392	if (!was_swapbacked &&
1393	folio_test_swapbacked(folio))
1394	stat->nr_lazyfree_fail += nr_pages;
1395	goto activate_locked;
1396	}
1397	}
1398
1399	/*
1400	* Folio is unmapped now so it cannot be newly pinned anymore.
1401	* No point in trying to reclaim folio if it is pinned.
1402	* Furthermore we don't want to reclaim underlying fs metadata
1403	* if the folio is pinned and thus potentially modified by the
1404	* pinning process as that may upset the filesystem.
1405	*/
1406	if (folio_maybe_dma_pinned(folio))
1407	goto activate_locked;
1408
1409	mapping = folio_mapping(folio);
1410	if (folio_test_dirty(folio)) {
1411	/*
1412	* Only kswapd can writeback filesystem folios
1413	* to avoid risk of stack overflow. But avoid
1414	* injecting inefficient single-folio I/O into
1415	* flusher writeback as much as possible: only
1416	* write folios when we've encountered many
1417	* dirty folios, and when we've already scanned
1418	* the rest of the LRU for clean folios and see
1419	* the same dirty folios again (with the reclaim
1420	* flag set).
1421	*/
1422	if (folio_is_file_lru(folio) &&
1423	(!current_is_kswapd() ||
1424	!folio_test_reclaim(folio) ||
1425	!test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1426	/*
1427	* Immediately reclaim when written back.
1428	* Similar in principle to folio_deactivate()
1429	* except we already have the folio isolated
1430	* and know it's dirty
1431	*/
1432	node_stat_mod_folio(folio, item: NR_VMSCAN_IMMEDIATE,
1433	nr: nr_pages);
1434	folio_set_reclaim(folio);
1435
1436	goto activate_locked;
1437	}
1438
1439	if (references == FOLIOREF_RECLAIM_CLEAN)
1440	goto keep_locked;
1441	if (!may_enter_fs(folio, gfp_mask: sc->gfp_mask))
1442	goto keep_locked;
1443	if (!sc->may_writepage)
1444	goto keep_locked;
1445
1446	/*
1447	* Folio is dirty. Flush the TLB if a writable entry
1448	* potentially exists to avoid CPU writes after I/O
1449	* starts and then write it out here.
1450	*/
1451	try_to_unmap_flush_dirty();
1452	switch (pageout(folio, mapping, plug: &plug, folio_list)) {
1453	case PAGE_KEEP:
1454	goto keep_locked;
1455	case PAGE_ACTIVATE:
1456	/*
1457	* If shmem folio is split when writeback to swap,
1458	* the tail pages will make their own pass through
1459	* this function and be accounted then.
1460	*/
1461	if (nr_pages > 1 && !folio_test_large(folio)) {
1462	sc->nr_scanned -= (nr_pages - 1);
1463	nr_pages = 1;
1464	}
1465	goto activate_locked;
1466	case PAGE_SUCCESS:
1467	if (nr_pages > 1 && !folio_test_large(folio)) {
1468	sc->nr_scanned -= (nr_pages - 1);
1469	nr_pages = 1;
1470	}
1471	stat->nr_pageout += nr_pages;
1472
1473	if (folio_test_writeback(folio))
1474	goto keep;
1475	if (folio_test_dirty(folio))
1476	goto keep;
1477
1478	/*
1479	* A synchronous write - probably a ramdisk. Go
1480	* ahead and try to reclaim the folio.
1481	*/
1482	if (!folio_trylock(folio))
1483	goto keep;
1484	if (folio_test_dirty(folio) ||
1485	folio_test_writeback(folio))
1486	goto keep_locked;
1487	mapping = folio_mapping(folio);
1488	fallthrough;
1489	case PAGE_CLEAN:
1490	; /* try to free the folio below */
1491	}
1492	}
1493
1494	/*
1495	* If the folio has buffers, try to free the buffer
1496	* mappings associated with this folio. If we succeed
1497	* we try to free the folio as well.
1498	*
1499	* We do this even if the folio is dirty.
1500	* filemap_release_folio() does not perform I/O, but it
1501	* is possible for a folio to have the dirty flag set,
1502	* but it is actually clean (all its buffers are clean).
1503	* This happens if the buffers were written out directly,
1504	* with submit_bh(). ext3 will do this, as well as
1505	* the blockdev mapping. filemap_release_folio() will
1506	* discover that cleanness and will drop the buffers
1507	* and mark the folio clean - it can be freed.
1508	*
1509	* Rarely, folios can have buffers and no ->mapping.
1510	* These are the folios which were not successfully
1511	* invalidated in truncate_cleanup_folio(). We try to
1512	* drop those buffers here and if that worked, and the
1513	* folio is no longer mapped into process address space
1514	* (refcount == 1) it can be freed. Otherwise, leave
1515	* the folio on the LRU so it is swappable.
1516	*/
1517	if (folio_needs_release(folio)) {
1518	if (!filemap_release_folio(folio, gfp: sc->gfp_mask))
1519	goto activate_locked;
1520	if (!mapping && folio_ref_count(folio) == 1) {
1521	folio_unlock(folio);
1522	if (folio_put_testzero(folio))
1523	goto free_it;
1524	else {
1525	/*
1526	* rare race with speculative reference.
1527	* the speculative reference will free
1528	* this folio shortly, so we may
1529	* increment nr_reclaimed here (and
1530	* leave it off the LRU).
1531	*/
1532	nr_reclaimed += nr_pages;
1533	continue;
1534	}
1535	}
1536	}
1537
1538	if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1539	/* follow __remove_mapping for reference */
1540	if (!folio_ref_freeze(folio, count: 1))
1541	goto keep_locked;
1542	/*
1543	* The folio has only one reference left, which is
1544	* from the isolation. After the caller puts the
1545	* folio back on the lru and drops the reference, the
1546	* folio will be freed anyway. It doesn't matter
1547	* which lru it goes on. So we don't bother checking
1548	* the dirty flag here.
1549	*/
1550	count_vm_events(item: PGLAZYFREED, delta: nr_pages);
1551	count_memcg_folio_events(folio, idx: PGLAZYFREED, nr: nr_pages);
1552	} else if (!mapping || !__remove_mapping(mapping, folio, reclaimed: true,
1553	target_memcg: sc->target_mem_cgroup))
1554	goto keep_locked;
1555
1556	folio_unlock(folio);
1557	free_it:
1558	/*
1559	* Folio may get swapped out as a whole, need to account
1560	* all pages in it.
1561	*/
1562	nr_reclaimed += nr_pages;
1563
1564	folio_unqueue_deferred_split(folio);
1565	if (folio_batch_add(fbatch: &free_folios, folio) == 0) {
1566	mem_cgroup_uncharge_folios(folios: &free_folios);
1567	try_to_unmap_flush();
1568	free_unref_folios(fbatch: &free_folios);
1569	}
1570	continue;
1571
1572	activate_locked_split:
1573	/*
1574	* The tail pages that are failed to add into swap cache
1575	* reach here. Fixup nr_scanned and nr_pages.
1576	*/
1577	if (nr_pages > 1) {
1578	sc->nr_scanned -= (nr_pages - 1);
1579	nr_pages = 1;
1580	}
1581	activate_locked:
1582	/* Not a candidate for swapping, so reclaim swap space. */
1583	if (folio_test_swapcache(folio) &&
1584	(mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1585	folio_free_swap(folio);
1586	VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1587	if (!folio_test_mlocked(folio)) {
1588	int type = folio_is_file_lru(folio);
1589	folio_set_active(folio);
1590	stat->nr_activate[type] += nr_pages;
1591	count_memcg_folio_events(folio, idx: PGACTIVATE, nr: nr_pages);
1592	}
1593	keep_locked:
1594	folio_unlock(folio);
1595	keep:
1596	list_add(new: &folio->lru, head: &ret_folios);
1597	VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1598	folio_test_unevictable(folio), folio);
1599	}
1600	/* 'folio_list' is always empty here */
1601
1602	/* Migrate folios selected for demotion */
1603	nr_demoted = demote_folio_list(demote_folios: &demote_folios, pgdat);
1604	nr_reclaimed += nr_demoted;
1605	stat->nr_demoted += nr_demoted;
1606	/* Folios that could not be demoted are still in @demote_folios */
1607	if (!list_empty(head: &demote_folios)) {
1608	/* Folios which weren't demoted go back on @folio_list */
1609	list_splice_init(list: &demote_folios, head: folio_list);
1610
1611	/*
1612	* goto retry to reclaim the undemoted folios in folio_list if
1613	* desired.
1614	*
1615	* Reclaiming directly from top tier nodes is not often desired
1616	* due to it breaking the LRU ordering: in general memory
1617	* should be reclaimed from lower tier nodes and demoted from
1618	* top tier nodes.
1619	*
1620	* However, disabling reclaim from top tier nodes entirely
1621	* would cause ooms in edge scenarios where lower tier memory
1622	* is unreclaimable for whatever reason, eg memory being
1623	* mlocked or too hot to reclaim. We can disable reclaim
1624	* from top tier nodes in proactive reclaim though as that is
1625	* not real memory pressure.
1626	*/
1627	if (!sc->proactive) {
1628	do_demote_pass = false;
1629	goto retry;
1630	}
1631	}
1632
1633	pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1634
1635	mem_cgroup_uncharge_folios(folios: &free_folios);
1636	try_to_unmap_flush();
1637	free_unref_folios(fbatch: &free_folios);
1638
1639	list_splice(list: &ret_folios, head: folio_list);
1640	count_vm_events(item: PGACTIVATE, delta: pgactivate);
1641
1642	if (plug)
1643	swap_write_unplug(sio: plug);
1644	return nr_reclaimed;
1645	}
1646
1647	unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1648	struct list_head *folio_list)
1649	{
1650	struct scan_control sc = {
1651	.gfp_mask = GFP_KERNEL,
1652	.may_unmap = 1,
1653	};
1654	struct reclaim_stat stat;
1655	unsigned int nr_reclaimed;
1656	struct folio *folio, *next;
1657	LIST_HEAD(clean_folios);
1658	unsigned int noreclaim_flag;
1659
1660	list_for_each_entry_safe(folio, next, folio_list, lru) {
1661	if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1662	!folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1663	!folio_test_unevictable(folio)) {
1664	folio_clear_active(folio);
1665	list_move(list: &folio->lru, head: &clean_folios);
1666	}
1667	}
1668
1669	/*
1670	* We should be safe here since we are only dealing with file pages and
1671	* we are not kswapd and therefore cannot write dirty file pages. But
1672	* call memalloc_noreclaim_save() anyway, just in case these conditions
1673	* change in the future.
1674	*/
1675	noreclaim_flag = memalloc_noreclaim_save();
1676	nr_reclaimed = shrink_folio_list(folio_list: &clean_folios, pgdat: zone->zone_pgdat, sc: &sc,
1677	stat: &stat, ignore_references: true, NULL);
1678	memalloc_noreclaim_restore(flags: noreclaim_flag);
1679
1680	list_splice(list: &clean_folios, head: folio_list);
1681	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1682	-(long)nr_reclaimed);
1683	/*
1684	* Since lazyfree pages are isolated from file LRU from the beginning,
1685	* they will rotate back to anonymous LRU in the end if it failed to
1686	* discard so isolated count will be mismatched.
1687	* Compensate the isolated count for both LRU lists.
1688	*/
1689	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1690	stat.nr_lazyfree_fail);
1691	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1692	-(long)stat.nr_lazyfree_fail);
1693	return nr_reclaimed;
1694	}
1695
1696	/*
1697	* Update LRU sizes after isolating pages. The LRU size updates must
1698	* be complete before mem_cgroup_update_lru_size due to a sanity check.
1699	*/
1700	static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1701	enum lru_list lru, unsigned long *nr_zone_taken)
1702	{
1703	int zid;
1704
1705	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1706	if (!nr_zone_taken[zid])
1707	continue;
1708
1709	update_lru_size(lruvec, lru, zid, nr_pages: -nr_zone_taken[zid]);
1710	}
1711
1712	}
1713
1714	/*
1715	* Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1716	*
1717	* lruvec->lru_lock is heavily contended. Some of the functions that
1718	* shrink the lists perform better by taking out a batch of pages
1719	* and working on them outside the LRU lock.
1720	*
1721	* For pagecache intensive workloads, this function is the hottest
1722	* spot in the kernel (apart from copy_*_user functions).
1723	*
1724	* Lru_lock must be held before calling this function.
1725	*
1726	* @nr_to_scan: The number of eligible pages to look through on the list.
1727	* @lruvec: The LRU vector to pull pages from.
1728	* @dst: The temp list to put pages on to.
1729	* @nr_scanned: The number of pages that were scanned.
1730	* @sc: The scan_control struct for this reclaim session
1731	* @lru: LRU list id for isolating
1732	*
1733	* returns how many pages were moved onto *@dst.
1734	*/
1735	static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1736	struct lruvec *lruvec, struct list_head *dst,
1737	unsigned long *nr_scanned, struct scan_control *sc,
1738	enum lru_list lru)
1739	{
1740	struct list_head *src = &lruvec->lists[lru];
1741	unsigned long nr_taken = 0;
1742	unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1743	unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1744	unsigned long skipped = 0, total_scan = 0, scan = 0;
1745	unsigned long nr_pages;
1746	unsigned long max_nr_skipped = 0;
1747	LIST_HEAD(folios_skipped);
1748
1749	while (scan < nr_to_scan && !list_empty(head: src)) {
1750	struct list_head *move_to = src;
1751	struct folio *folio;
1752
1753	folio = lru_to_folio(head: src);
1754	prefetchw_prev_lru_folio(folio, src, flags);
1755
1756	nr_pages = folio_nr_pages(folio);
1757	total_scan += nr_pages;
1758
1759	/* Using max_nr_skipped to prevent hard LOCKUP*/
1760	if (max_nr_skipped < SWAP_CLUSTER_MAX_SKIPPED &&
1761	(folio_zonenum(folio) > sc->reclaim_idx)) {
1762	nr_skipped[folio_zonenum(folio)] += nr_pages;
1763	move_to = &folios_skipped;
1764	max_nr_skipped++;
1765	goto move;
1766	}
1767
1768	/*
1769	* Do not count skipped folios because that makes the function
1770	* return with no isolated folios if the LRU mostly contains
1771	* ineligible folios. This causes the VM to not reclaim any
1772	* folios, triggering a premature OOM.
1773	* Account all pages in a folio.
1774	*/
1775	scan += nr_pages;
1776
1777	if (!folio_test_lru(folio))
1778	goto move;
1779	if (!sc->may_unmap && folio_mapped(folio))
1780	goto move;
1781
1782	/*
1783	* Be careful not to clear the lru flag until after we're
1784	* sure the folio is not being freed elsewhere -- the
1785	* folio release code relies on it.
1786	*/
1787	if (unlikely(!folio_try_get(folio)))
1788	goto move;
1789
1790	if (!folio_test_clear_lru(folio)) {
1791	/* Another thread is already isolating this folio */
1792	folio_put(folio);
1793	goto move;
1794	}
1795
1796	nr_taken += nr_pages;
1797	nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1798	move_to = dst;
1799	move:
1800	list_move(list: &folio->lru, head: move_to);
1801	}
1802
1803	/*
1804	* Splice any skipped folios to the start of the LRU list. Note that
1805	* this disrupts the LRU order when reclaiming for lower zones but
1806	* we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1807	* scanning would soon rescan the same folios to skip and waste lots
1808	* of cpu cycles.
1809	*/
1810	if (!list_empty(head: &folios_skipped)) {
1811	int zid;
1812
1813	list_splice(list: &folios_skipped, head: src);
1814	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1815	if (!nr_skipped[zid])
1816	continue;
1817
1818	__count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1819	skipped += nr_skipped[zid];
1820	}
1821	}
1822	*nr_scanned = total_scan;
1823	trace_mm_vmscan_lru_isolate(highest_zoneidx: sc->reclaim_idx, order: sc->order, nr_requested: nr_to_scan,
1824	nr_scanned: total_scan, nr_skipped: skipped, nr_taken, lru);
1825	update_lru_sizes(lruvec, lru, nr_zone_taken);
1826	return nr_taken;
1827	}
1828
1829	/**
1830	* folio_isolate_lru() - Try to isolate a folio from its LRU list.
1831	* @folio: Folio to isolate from its LRU list.
1832	*
1833	* Isolate a @folio from an LRU list and adjust the vmstat statistic
1834	* corresponding to whatever LRU list the folio was on.
1835	*
1836	* The folio will have its LRU flag cleared. If it was found on the
1837	* active list, it will have the Active flag set. If it was found on the
1838	* unevictable list, it will have the Unevictable flag set. These flags
1839	* may need to be cleared by the caller before letting the page go.
1840	*
1841	* Context:
1842	*
1843	* (1) Must be called with an elevated refcount on the folio. This is a
1844	* fundamental difference from isolate_lru_folios() (which is called
1845	* without a stable reference).
1846	* (2) The lru_lock must not be held.
1847	* (3) Interrupts must be enabled.
1848	*
1849	* Return: true if the folio was removed from an LRU list.
1850	* false if the folio was not on an LRU list.
1851	*/
1852	bool folio_isolate_lru(struct folio *folio)
1853	{
1854	bool ret = false;
1855
1856	VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1857
1858	if (folio_test_clear_lru(folio)) {
1859	struct lruvec *lruvec;
1860
1861	folio_get(folio);
1862	lruvec = folio_lruvec_lock_irq(folio);
1863	lruvec_del_folio(lruvec, folio);
1864	unlock_page_lruvec_irq(lruvec);
1865	ret = true;
1866	}
1867
1868	return ret;
1869	}
1870
1871	/*
1872	* A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1873	* then get rescheduled. When there are massive number of tasks doing page
1874	* allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1875	* the LRU list will go small and be scanned faster than necessary, leading to
1876	* unnecessary swapping, thrashing and OOM.
1877	*/
1878	static bool too_many_isolated(struct pglist_data *pgdat, int file,
1879	struct scan_control *sc)
1880	{
1881	unsigned long inactive, isolated;
1882	bool too_many;
1883
1884	if (current_is_kswapd())
1885	return false;
1886
1887	if (!writeback_throttling_sane(sc))
1888	return false;
1889
1890	if (file) {
1891	inactive = node_page_state(pgdat, item: NR_INACTIVE_FILE);
1892	isolated = node_page_state(pgdat, item: NR_ISOLATED_FILE);
1893	} else {
1894	inactive = node_page_state(pgdat, item: NR_INACTIVE_ANON);
1895	isolated = node_page_state(pgdat, item: NR_ISOLATED_ANON);
1896	}
1897
1898	/*
1899	* GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1900	* won't get blocked by normal direct-reclaimers, forming a circular
1901	* deadlock.
1902	*/
1903	if (gfp_has_io_fs(gfp: sc->gfp_mask))
1904	inactive >>= 3;
1905
1906	too_many = isolated > inactive;
1907
1908	/* Wake up tasks throttled due to too_many_isolated. */
1909	if (!too_many)
1910	wake_throttle_isolated(pgdat);
1911
1912	return too_many;
1913	}
1914
1915	/*
1916	* move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1917	*
1918	* Returns the number of pages moved to the given lruvec.
1919	*/
1920	static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1921	struct list_head *list)
1922	{
1923	int nr_pages, nr_moved = 0;
1924	struct folio_batch free_folios;
1925
1926	folio_batch_init(fbatch: &free_folios);
1927	while (!list_empty(head: list)) {
1928	struct folio *folio = lru_to_folio(head: list);
1929
1930	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1931	list_del(entry: &folio->lru);
1932	if (unlikely(!folio_evictable(folio))) {
1933	spin_unlock_irq(lock: &lruvec->lru_lock);
1934	folio_putback_lru(folio);
1935	spin_lock_irq(lock: &lruvec->lru_lock);
1936	continue;
1937	}
1938
1939	/*
1940	* The folio_set_lru needs to be kept here for list integrity.
1941	* Otherwise:
1942	* #0 move_folios_to_lru #1 release_pages
1943	* if (!folio_put_testzero())
1944	* if (folio_put_testzero())
1945	* !lru //skip lru_lock
1946	* folio_set_lru()
1947	* list_add(&folio->lru,)
1948	* list_add(&folio->lru,)
1949	*/
1950	folio_set_lru(folio);
1951
1952	if (unlikely(folio_put_testzero(folio))) {
1953	__folio_clear_lru_flags(folio);
1954
1955	folio_unqueue_deferred_split(folio);
1956	if (folio_batch_add(fbatch: &free_folios, folio) == 0) {
1957	spin_unlock_irq(lock: &lruvec->lru_lock);
1958	mem_cgroup_uncharge_folios(folios: &free_folios);
1959	free_unref_folios(fbatch: &free_folios);
1960	spin_lock_irq(lock: &lruvec->lru_lock);
1961	}
1962
1963	continue;
1964	}
1965
1966	/*
1967	* All pages were isolated from the same lruvec (and isolation
1968	* inhibits memcg migration).
1969	*/
1970	VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1971	lruvec_add_folio(lruvec, folio);
1972	nr_pages = folio_nr_pages(folio);
1973	nr_moved += nr_pages;
1974	if (folio_test_active(folio))
1975	workingset_age_nonresident(lruvec, nr_pages);
1976	}
1977
1978	if (free_folios.nr) {
1979	spin_unlock_irq(lock: &lruvec->lru_lock);
1980	mem_cgroup_uncharge_folios(folios: &free_folios);
1981	free_unref_folios(fbatch: &free_folios);
1982	spin_lock_irq(lock: &lruvec->lru_lock);
1983	}
1984
1985	return nr_moved;
1986	}
1987
1988	/*
1989	* If a kernel thread (such as nfsd for loop-back mounts) services a backing
1990	* device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1991	* we should not throttle. Otherwise it is safe to do so.
1992	*/
1993	static int current_may_throttle(void)
1994	{
1995	return !(current->flags & PF_LOCAL_THROTTLE);
1996	}
1997
1998	/*
1999	* shrink_inactive_list() is a helper for shrink_node(). It returns the number
2000	* of reclaimed pages
2001	*/
2002	static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2003	struct lruvec *lruvec, struct scan_control *sc,
2004	enum lru_list lru)
2005	{
2006	LIST_HEAD(folio_list);
2007	unsigned long nr_scanned;
2008	unsigned int nr_reclaimed = 0;
2009	unsigned long nr_taken;
2010	struct reclaim_stat stat;
2011	bool file = is_file_lru(lru);
2012	enum vm_event_item item;
2013	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2014	bool stalled = false;
2015
2016	while (unlikely(too_many_isolated(pgdat, file, sc))) {
2017	if (stalled)
2018	return 0;
2019
2020	/* wait a bit for the reclaimer. */
2021	stalled = true;
2022	reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_ISOLATED);
2023
2024	/* We are about to die and free our memory. Return now. */
2025	if (fatal_signal_pending(current))
2026	return SWAP_CLUSTER_MAX;
2027	}
2028
2029	lru_add_drain();
2030
2031	spin_lock_irq(lock: &lruvec->lru_lock);
2032
2033	nr_taken = isolate_lru_folios(nr_to_scan, lruvec, dst: &folio_list,
2034	nr_scanned: &nr_scanned, sc, lru);
2035
2036	__mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, nr_taken);
2037	item = PGSCAN_KSWAPD + reclaimer_offset(sc);
2038	if (!cgroup_reclaim(sc))
2039	__count_vm_events(item, delta: nr_scanned);
2040	count_memcg_events(memcg: lruvec_memcg(lruvec), idx: item, count: nr_scanned);
2041	__count_vm_events(item: PGSCAN_ANON + file, delta: nr_scanned);
2042
2043	spin_unlock_irq(lock: &lruvec->lru_lock);
2044
2045	if (nr_taken == 0)
2046	return 0;
2047
2048	nr_reclaimed = shrink_folio_list(folio_list: &folio_list, pgdat, sc, stat: &stat, ignore_references: false,
2049	memcg: lruvec_memcg(lruvec));
2050
2051	spin_lock_irq(lock: &lruvec->lru_lock);
2052	move_folios_to_lru(lruvec, list: &folio_list);
2053
2054	__mod_lruvec_state(lruvec, idx: PGDEMOTE_KSWAPD + reclaimer_offset(sc),
2055	val: stat.nr_demoted);
2056	__mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, -nr_taken);
2057	item = PGSTEAL_KSWAPD + reclaimer_offset(sc);
2058	if (!cgroup_reclaim(sc))
2059	__count_vm_events(item, delta: nr_reclaimed);
2060	count_memcg_events(memcg: lruvec_memcg(lruvec), idx: item, count: nr_reclaimed);
2061	__count_vm_events(item: PGSTEAL_ANON + file, delta: nr_reclaimed);
2062	spin_unlock_irq(lock: &lruvec->lru_lock);
2063
2064	lru_note_cost(lruvec, file, nr_io: stat.nr_pageout, nr_rotated: nr_scanned - nr_reclaimed);
2065
2066	/*
2067	* If dirty folios are scanned that are not queued for IO, it
2068	* implies that flushers are not doing their job. This can
2069	* happen when memory pressure pushes dirty folios to the end of
2070	* the LRU before the dirty limits are breached and the dirty
2071	* data has expired. It can also happen when the proportion of
2072	* dirty folios grows not through writes but through memory
2073	* pressure reclaiming all the clean cache. And in some cases,
2074	* the flushers simply cannot keep up with the allocation
2075	* rate. Nudge the flusher threads in case they are asleep.
2076	*/
2077	if (stat.nr_unqueued_dirty == nr_taken) {
2078	wakeup_flusher_threads(reason: WB_REASON_VMSCAN);
2079	/*
2080	* For cgroupv1 dirty throttling is achieved by waking up
2081	* the kernel flusher here and later waiting on folios
2082	* which are in writeback to finish (see shrink_folio_list()).
2083	*
2084	* Flusher may not be able to issue writeback quickly
2085	* enough for cgroupv1 writeback throttling to work
2086	* on a large system.
2087	*/
2088	if (!writeback_throttling_sane(sc))
2089	reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_WRITEBACK);
2090	}
2091
2092	sc->nr.dirty += stat.nr_dirty;
2093	sc->nr.congested += stat.nr_congested;
2094	sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2095	sc->nr.writeback += stat.nr_writeback;
2096	sc->nr.immediate += stat.nr_immediate;
2097	sc->nr.taken += nr_taken;
2098	if (file)
2099	sc->nr.file_taken += nr_taken;
2100
2101	trace_mm_vmscan_lru_shrink_inactive(nid: pgdat->node_id,
2102	nr_scanned, nr_reclaimed, stat: &stat, priority: sc->priority, file);
2103	return nr_reclaimed;
2104	}
2105
2106	/*
2107	* shrink_active_list() moves folios from the active LRU to the inactive LRU.
2108	*
2109	* We move them the other way if the folio is referenced by one or more
2110	* processes.
2111	*
2112	* If the folios are mostly unmapped, the processing is fast and it is
2113	* appropriate to hold lru_lock across the whole operation. But if
2114	* the folios are mapped, the processing is slow (folio_referenced()), so
2115	* we should drop lru_lock around each folio. It's impossible to balance
2116	* this, so instead we remove the folios from the LRU while processing them.
2117	* It is safe to rely on the active flag against the non-LRU folios in here
2118	* because nobody will play with that bit on a non-LRU folio.
2119	*
2120	* The downside is that we have to touch folio->_refcount against each folio.
2121	* But we had to alter folio->flags anyway.
2122	*/
2123	static void shrink_active_list(unsigned long nr_to_scan,
2124	struct lruvec *lruvec,
2125	struct scan_control *sc,
2126	enum lru_list lru)
2127	{
2128	unsigned long nr_taken;
2129	unsigned long nr_scanned;
2130	unsigned long vm_flags;
2131	LIST_HEAD(l_hold); /* The folios which were snipped off */
2132	LIST_HEAD(l_active);
2133	LIST_HEAD(l_inactive);
2134	unsigned nr_deactivate, nr_activate;
2135	unsigned nr_rotated = 0;
2136	bool file = is_file_lru(lru);
2137	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2138
2139	lru_add_drain();
2140
2141	spin_lock_irq(lock: &lruvec->lru_lock);
2142
2143	nr_taken = isolate_lru_folios(nr_to_scan, lruvec, dst: &l_hold,
2144	nr_scanned: &nr_scanned, sc, lru);
2145
2146	__mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, nr_taken);
2147
2148	if (!cgroup_reclaim(sc))
2149	__count_vm_events(item: PGREFILL, delta: nr_scanned);
2150	count_memcg_events(memcg: lruvec_memcg(lruvec), idx: PGREFILL, count: nr_scanned);
2151
2152	spin_unlock_irq(lock: &lruvec->lru_lock);
2153
2154	while (!list_empty(head: &l_hold)) {
2155	struct folio *folio;
2156
2157	cond_resched();
2158	folio = lru_to_folio(head: &l_hold);
2159	list_del(entry: &folio->lru);
2160
2161	if (unlikely(!folio_evictable(folio))) {
2162	folio_putback_lru(folio);
2163	continue;
2164	}
2165
2166	if (unlikely(buffer_heads_over_limit)) {
2167	if (folio_needs_release(folio) &&
2168	folio_trylock(folio)) {
2169	filemap_release_folio(folio, gfp: 0);
2170	folio_unlock(folio);
2171	}
2172	}
2173
2174	/* Referenced or rmap lock contention: rotate */
2175	if (folio_referenced(folio, is_locked: 0, memcg: sc->target_mem_cgroup,
2176	vm_flags: &vm_flags) != 0) {
2177	/*
2178	* Identify referenced, file-backed active folios and
2179	* give them one more trip around the active list. So
2180	* that executable code get better chances to stay in
2181	* memory under moderate memory pressure. Anon folios
2182	* are not likely to be evicted by use-once streaming
2183	* IO, plus JVM can create lots of anon VM_EXEC folios,
2184	* so we ignore them here.
2185	*/
2186	if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2187	nr_rotated += folio_nr_pages(folio);
2188	list_add(new: &folio->lru, head: &l_active);
2189	continue;
2190	}
2191	}
2192
2193	folio_clear_active(folio); /* we are de-activating */
2194	folio_set_workingset(folio);
2195	list_add(new: &folio->lru, head: &l_inactive);
2196	}
2197
2198	/*
2199	* Move folios back to the lru list.
2200	*/
2201	spin_lock_irq(lock: &lruvec->lru_lock);
2202
2203	nr_activate = move_folios_to_lru(lruvec, list: &l_active);
2204	nr_deactivate = move_folios_to_lru(lruvec, list: &l_inactive);
2205
2206	__count_vm_events(item: PGDEACTIVATE, delta: nr_deactivate);
2207	count_memcg_events(memcg: lruvec_memcg(lruvec), idx: PGDEACTIVATE, count: nr_deactivate);
2208
2209	__mod_node_page_state(pgdat, item: NR_ISOLATED_ANON + file, -nr_taken);
2210	spin_unlock_irq(lock: &lruvec->lru_lock);
2211
2212	if (nr_rotated)
2213	lru_note_cost(lruvec, file, nr_io: 0, nr_rotated);
2214	trace_mm_vmscan_lru_shrink_active(nid: pgdat->node_id, nr_taken, nr_active: nr_activate,
2215	nr_deactivated: nr_deactivate, nr_referenced: nr_rotated, priority: sc->priority, file);
2216	}
2217
2218	static unsigned int reclaim_folio_list(struct list_head *folio_list,
2219	struct pglist_data *pgdat)
2220	{
2221	struct reclaim_stat stat;
2222	unsigned int nr_reclaimed;
2223	struct folio *folio;
2224	struct scan_control sc = {
2225	.gfp_mask = GFP_KERNEL,
2226	.may_writepage = 1,
2227	.may_unmap = 1,
2228	.may_swap = 1,
2229	.no_demotion = 1,
2230	};
2231
2232	nr_reclaimed = shrink_folio_list(folio_list, pgdat, sc: &sc, stat: &stat, ignore_references: true, NULL);
2233	while (!list_empty(head: folio_list)) {
2234	folio = lru_to_folio(head: folio_list);
2235	list_del(entry: &folio->lru);
2236	folio_putback_lru(folio);
2237	}
2238	trace_mm_vmscan_reclaim_pages(nid: pgdat->node_id, nr_scanned: sc.nr_scanned, nr_reclaimed, stat: &stat);
2239
2240	return nr_reclaimed;
2241	}
2242
2243	unsigned long reclaim_pages(struct list_head *folio_list)
2244	{
2245	int nid;
2246	unsigned int nr_reclaimed = 0;
2247	LIST_HEAD(node_folio_list);
2248	unsigned int noreclaim_flag;
2249
2250	if (list_empty(head: folio_list))
2251	return nr_reclaimed;
2252
2253	noreclaim_flag = memalloc_noreclaim_save();
2254
2255	nid = folio_nid(folio: lru_to_folio(head: folio_list));
2256	do {
2257	struct folio *folio = lru_to_folio(head: folio_list);
2258
2259	if (nid == folio_nid(folio)) {
2260	folio_clear_active(folio);
2261	list_move(list: &folio->lru, head: &node_folio_list);
2262	continue;
2263	}
2264
2265	nr_reclaimed += reclaim_folio_list(folio_list: &node_folio_list, NODE_DATA(nid));
2266	nid = folio_nid(folio: lru_to_folio(head: folio_list));
2267	} while (!list_empty(head: folio_list));
2268
2269	nr_reclaimed += reclaim_folio_list(folio_list: &node_folio_list, NODE_DATA(nid));
2270
2271	memalloc_noreclaim_restore(flags: noreclaim_flag);
2272
2273	return nr_reclaimed;
2274	}
2275
2276	static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277	struct lruvec *lruvec, struct scan_control *sc)
2278	{
2279	if (is_active_lru(lru)) {
2280	if (sc->may_deactivate & (1 << is_file_lru(lru)))
2281	shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282	else
2283	sc->skipped_deactivate = 1;
2284	return 0;
2285	}
2286
2287	return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2288	}
2289
2290	/*
2291	* The inactive anon list should be small enough that the VM never has
2292	* to do too much work.
2293	*
2294	* The inactive file list should be small enough to leave most memory
2295	* to the established workingset on the scan-resistant active list,
2296	* but large enough to avoid thrashing the aggregate readahead window.
2297	*
2298	* Both inactive lists should also be large enough that each inactive
2299	* folio has a chance to be referenced again before it is reclaimed.
2300	*
2301	* If that fails and refaulting is observed, the inactive list grows.
2302	*
2303	* The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2304	* on this LRU, maintained by the pageout code. An inactive_ratio
2305	* of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2306	*
2307	* total target max
2308	* memory ratio inactive
2309	* -------------------------------------
2310	* 10MB 1 5MB
2311	* 100MB 1 50MB
2312	* 1GB 3 250MB
2313	* 10GB 10 0.9GB
2314	* 100GB 31 3GB
2315	* 1TB 101 10GB
2316	* 10TB 320 32GB
2317	*/
2318	static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2319	{
2320	enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2321	unsigned long inactive, active;
2322	unsigned long inactive_ratio;
2323	unsigned long gb;
2324
2325	inactive = lruvec_page_state(lruvec, idx: NR_LRU_BASE + inactive_lru);
2326	active = lruvec_page_state(lruvec, idx: NR_LRU_BASE + active_lru);
2327
2328	gb = (inactive + active) >> (30 - PAGE_SHIFT);
2329	if (gb)
2330	inactive_ratio = int_sqrt(10 * gb);
2331	else
2332	inactive_ratio = 1;
2333
2334	return inactive * inactive_ratio < active;
2335	}
2336
2337	enum scan_balance {
2338	SCAN_EQUAL,
2339	SCAN_FRACT,
2340	SCAN_ANON,
2341	SCAN_FILE,
2342	};
2343
2344	static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2345	{
2346	unsigned long file;
2347	struct lruvec *target_lruvec;
2348
2349	if (lru_gen_enabled())
2350	return;
2351
2352	target_lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup, pgdat);
2353
2354	/*
2355	* Flush the memory cgroup stats in rate-limited way as we don't need
2356	* most accurate stats here. We may switch to regular stats flushing
2357	* in the future once it is cheap enough.
2358	*/
2359	mem_cgroup_flush_stats_ratelimited(memcg: sc->target_mem_cgroup);
2360
2361	/*
2362	* Determine the scan balance between anon and file LRUs.
2363	*/
2364	spin_lock_irq(lock: &target_lruvec->lru_lock);
2365	sc->anon_cost = target_lruvec->anon_cost;
2366	sc->file_cost = target_lruvec->file_cost;
2367	spin_unlock_irq(lock: &target_lruvec->lru_lock);
2368
2369	/*
2370	* Target desirable inactive:active list ratios for the anon
2371	* and file LRU lists.
2372	*/
2373	if (!sc->force_deactivate) {
2374	unsigned long refaults;
2375
2376	/*
2377	* When refaults are being observed, it means a new
2378	* workingset is being established. Deactivate to get
2379	* rid of any stale active pages quickly.
2380	*/
2381	refaults = lruvec_page_state(lruvec: target_lruvec,
2382	idx: WORKINGSET_ACTIVATE_ANON);
2383	if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2384	inactive_is_low(lruvec: target_lruvec, inactive_lru: LRU_INACTIVE_ANON))
2385	sc->may_deactivate |= DEACTIVATE_ANON;
2386	else
2387	sc->may_deactivate &= ~DEACTIVATE_ANON;
2388
2389	refaults = lruvec_page_state(lruvec: target_lruvec,
2390	idx: WORKINGSET_ACTIVATE_FILE);
2391	if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2392	inactive_is_low(lruvec: target_lruvec, inactive_lru: LRU_INACTIVE_FILE))
2393	sc->may_deactivate |= DEACTIVATE_FILE;
2394	else
2395	sc->may_deactivate &= ~DEACTIVATE_FILE;
2396	} else
2397	sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2398
2399	/*
2400	* If we have plenty of inactive file pages that aren't
2401	* thrashing, try to reclaim those first before touching
2402	* anonymous pages.
2403	*/
2404	file = lruvec_page_state(lruvec: target_lruvec, idx: NR_INACTIVE_FILE);
2405	if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2406	!sc->no_cache_trim_mode)
2407	sc->cache_trim_mode = 1;
2408	else
2409	sc->cache_trim_mode = 0;
2410
2411	/*
2412	* Prevent the reclaimer from falling into the cache trap: as
2413	* cache pages start out inactive, every cache fault will tip
2414	* the scan balance towards the file LRU. And as the file LRU
2415	* shrinks, so does the window for rotation from references.
2416	* This means we have a runaway feedback loop where a tiny
2417	* thrashing file LRU becomes infinitely more attractive than
2418	* anon pages. Try to detect this based on file LRU size.
2419	*/
2420	if (!cgroup_reclaim(sc)) {
2421	unsigned long total_high_wmark = 0;
2422	unsigned long free, anon;
2423	int z;
2424	struct zone *zone;
2425
2426	free = sum_zone_node_page_state(node: pgdat->node_id, item: NR_FREE_PAGES);
2427	file = node_page_state(pgdat, item: NR_ACTIVE_FILE) +
2428	node_page_state(pgdat, item: NR_INACTIVE_FILE);
2429
2430	for_each_managed_zone_pgdat(zone, pgdat, z, MAX_NR_ZONES - 1) {
2431	total_high_wmark += high_wmark_pages(z: zone);
2432	}
2433
2434	/*
2435	* Consider anon: if that's low too, this isn't a
2436	* runaway file reclaim problem, but rather just
2437	* extreme pressure. Reclaim as per usual then.
2438	*/
2439	anon = node_page_state(pgdat, item: NR_INACTIVE_ANON);
2440
2441	sc->file_is_tiny =
2442	file + free <= total_high_wmark &&
2443	!(sc->may_deactivate & DEACTIVATE_ANON) &&
2444	anon >> sc->priority;
2445	}
2446	}
2447
2448	static inline void calculate_pressure_balance(struct scan_control *sc,
2449	int swappiness, u64 *fraction, u64 *denominator)
2450	{
2451	unsigned long anon_cost, file_cost, total_cost;
2452	unsigned long ap, fp;
2453
2454	/*
2455	* Calculate the pressure balance between anon and file pages.
2456	*
2457	* The amount of pressure we put on each LRU is inversely
2458	* proportional to the cost of reclaiming each list, as
2459	* determined by the share of pages that are refaulting, times
2460	* the relative IO cost of bringing back a swapped out
2461	* anonymous page vs reloading a filesystem page (swappiness).
2462	*
2463	* Although we limit that influence to ensure no list gets
2464	* left behind completely: at least a third of the pressure is
2465	* applied, before swappiness.
2466	*
2467	* With swappiness at 100, anon and file have equal IO cost.
2468	*/
2469	total_cost = sc->anon_cost + sc->file_cost;
2470	anon_cost = total_cost + sc->anon_cost;
2471	file_cost = total_cost + sc->file_cost;
2472	total_cost = anon_cost + file_cost;
2473
2474	ap = swappiness * (total_cost + 1);
2475	ap /= anon_cost + 1;
2476
2477	fp = (MAX_SWAPPINESS - swappiness) * (total_cost + 1);
2478	fp /= file_cost + 1;
2479
2480	fraction[WORKINGSET_ANON] = ap;
2481	fraction[WORKINGSET_FILE] = fp;
2482	*denominator = ap + fp;
2483	}
2484
2485	/*
2486	* Determine how aggressively the anon and file LRU lists should be
2487	* scanned.
2488	*
2489	* nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2490	* nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2491	*/
2492	static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2493	unsigned long *nr)
2494	{
2495	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2496	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2497	int swappiness = sc_swappiness(sc, memcg);
2498	u64 fraction[ANON_AND_FILE];
2499	u64 denominator = 0; /* gcc */
2500	enum scan_balance scan_balance;
2501	enum lru_list lru;
2502
2503	/* If we have no swap space, do not bother scanning anon folios. */
2504	if (!sc->may_swap || !can_reclaim_anon_pages(memcg, nid: pgdat->node_id, sc)) {
2505	scan_balance = SCAN_FILE;
2506	goto out;
2507	}
2508
2509	/*
2510	* Global reclaim will swap to prevent OOM even with no
2511	* swappiness, but memcg users want to use this knob to
2512	* disable swapping for individual groups completely when
2513	* using the memory controller's swap limit feature would be
2514	* too expensive.
2515	*/
2516	if (cgroup_reclaim(sc) && !swappiness) {
2517	scan_balance = SCAN_FILE;
2518	goto out;
2519	}
2520
2521	/* Proactive reclaim initiated by userspace for anonymous memory only */
2522	if (swappiness == SWAPPINESS_ANON_ONLY) {
2523	WARN_ON_ONCE(!sc->proactive);
2524	scan_balance = SCAN_ANON;
2525	goto out;
2526	}
2527
2528	/*
2529	* Do not apply any pressure balancing cleverness when the
2530	* system is close to OOM, scan both anon and file equally
2531	* (unless the swappiness setting disagrees with swapping).
2532	*/
2533	if (!sc->priority && swappiness) {
2534	scan_balance = SCAN_EQUAL;
2535	goto out;
2536	}
2537
2538	/*
2539	* If the system is almost out of file pages, force-scan anon.
2540	*/
2541	if (sc->file_is_tiny) {
2542	scan_balance = SCAN_ANON;
2543	goto out;
2544	}
2545
2546	/*
2547	* If there is enough inactive page cache, we do not reclaim
2548	* anything from the anonymous working right now to make sure
2549	* a streaming file access pattern doesn't cause swapping.
2550	*/
2551	if (sc->cache_trim_mode) {
2552	scan_balance = SCAN_FILE;
2553	goto out;
2554	}
2555
2556	scan_balance = SCAN_FRACT;
2557	calculate_pressure_balance(sc, swappiness, fraction, denominator: &denominator);
2558
2559	out:
2560	for_each_evictable_lru(lru) {
2561	bool file = is_file_lru(lru);
2562	unsigned long lruvec_size;
2563	unsigned long low, min;
2564	unsigned long scan;
2565
2566	lruvec_size = lruvec_lru_size(lruvec, lru, zone_idx: sc->reclaim_idx);
2567	mem_cgroup_protection(root: sc->target_mem_cgroup, memcg,
2568	min: &min, low: &low);
2569
2570	if (min || low) {
2571	/*
2572	* Scale a cgroup's reclaim pressure by proportioning
2573	* its current usage to its memory.low or memory.min
2574	* setting.
2575	*
2576	* This is important, as otherwise scanning aggression
2577	* becomes extremely binary -- from nothing as we
2578	* approach the memory protection threshold, to totally
2579	* nominal as we exceed it. This results in requiring
2580	* setting extremely liberal protection thresholds. It
2581	* also means we simply get no protection at all if we
2582	* set it too low, which is not ideal.
2583	*
2584	* If there is any protection in place, we reduce scan
2585	* pressure by how much of the total memory used is
2586	* within protection thresholds.
2587	*
2588	* There is one special case: in the first reclaim pass,
2589	* we skip over all groups that are within their low
2590	* protection. If that fails to reclaim enough pages to
2591	* satisfy the reclaim goal, we come back and override
2592	* the best-effort low protection. However, we still
2593	* ideally want to honor how well-behaved groups are in
2594	* that case instead of simply punishing them all
2595	* equally. As such, we reclaim them based on how much
2596	* memory they are using, reducing the scan pressure
2597	* again by how much of the total memory used is under
2598	* hard protection.
2599	*/
2600	unsigned long cgroup_size = mem_cgroup_size(memcg);
2601	unsigned long protection;
2602
2603	/* memory.low scaling, make sure we retry before OOM */
2604	if (!sc->memcg_low_reclaim && low > min) {
2605	protection = low;
2606	sc->memcg_low_skipped = 1;
2607	} else {
2608	protection = min;
2609	}
2610
2611	/* Avoid TOCTOU with earlier protection check */
2612	cgroup_size = max(cgroup_size, protection);
2613
2614	scan = lruvec_size - lruvec_size * protection /
2615	(cgroup_size + 1);
2616
2617	/*
2618	* Minimally target SWAP_CLUSTER_MAX pages to keep
2619	* reclaim moving forwards, avoiding decrementing
2620	* sc->priority further than desirable.
2621	*/
2622	scan = max(scan, SWAP_CLUSTER_MAX);
2623	} else {
2624	scan = lruvec_size;
2625	}
2626
2627	scan >>= sc->priority;
2628
2629	/*
2630	* If the cgroup's already been deleted, make sure to
2631	* scrape out the remaining cache.
2632	*/
2633	if (!scan && !mem_cgroup_online(memcg))
2634	scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2635
2636	switch (scan_balance) {
2637	case SCAN_EQUAL:
2638	/* Scan lists relative to size */
2639	break;
2640	case SCAN_FRACT:
2641	/*
2642	* Scan types proportional to swappiness and
2643	* their relative recent reclaim efficiency.
2644	* Make sure we don't miss the last page on
2645	* the offlined memory cgroups because of a
2646	* round-off error.
2647	*/
2648	scan = mem_cgroup_online(memcg) ?
2649	div64_u64(dividend: scan * fraction[file], divisor: denominator) :
2650	DIV64_U64_ROUND_UP(scan * fraction[file],
2651	denominator);
2652	break;
2653	case SCAN_FILE:
2654	case SCAN_ANON:
2655	/* Scan one type exclusively */
2656	if ((scan_balance == SCAN_FILE) != file)
2657	scan = 0;
2658	break;
2659	default:
2660	/* Look ma, no brain */
2661	BUG();
2662	}
2663
2664	nr[lru] = scan;
2665	}
2666	}
2667
2668	/*
2669	* Anonymous LRU management is a waste if there is
2670	* ultimately no way to reclaim the memory.
2671	*/
2672	static bool can_age_anon_pages(struct lruvec *lruvec,
2673	struct scan_control *sc)
2674	{
2675	/* Aging the anon LRU is valuable if swap is present: */
2676	if (total_swap_pages > 0)
2677	return true;
2678
2679	/* Also valuable if anon pages can be demoted: */
2680	return can_demote(nid: lruvec_pgdat(lruvec)->node_id, sc,
2681	memcg: lruvec_memcg(lruvec));
2682	}
2683
2684	#ifdef CONFIG_LRU_GEN
2685
2686	#ifdef CONFIG_LRU_GEN_ENABLED
2687	DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2688	#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2689	#else
2690	DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2691	#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2692	#endif
2693
2694	static bool should_walk_mmu(void)
2695	{
2696	return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2697	}
2698
2699	static bool should_clear_pmd_young(void)
2700	{
2701	return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2702	}
2703
2704	/******************************************************************************
2705	* shorthand helpers
2706	******************************************************************************/
2707
2708	#define DEFINE_MAX_SEQ(lruvec) \
2709	unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2710
2711	#define DEFINE_MIN_SEQ(lruvec) \
2712	unsigned long min_seq[ANON_AND_FILE] = { \
2713	READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2714	READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2715	}
2716
2717	/* Get the min/max evictable type based on swappiness */
2718	#define min_type(swappiness) (!(swappiness))
2719	#define max_type(swappiness) ((swappiness) < SWAPPINESS_ANON_ONLY)
2720
2721	#define evictable_min_seq(min_seq, swappiness) \
2722	min((min_seq)[min_type(swappiness)], (min_seq)[max_type(swappiness)])
2723
2724	#define for_each_gen_type_zone(gen, type, zone) \
2725	for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2726	for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2727	for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2728
2729	#define for_each_evictable_type(type, swappiness) \
2730	for ((type) = min_type(swappiness); (type) <= max_type(swappiness); (type)++)
2731
2732	#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2733	#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2734
2735	static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2736	{
2737	struct pglist_data *pgdat = NODE_DATA(nid);
2738
2739	#ifdef CONFIG_MEMCG
2740	if (memcg) {
2741	struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2742
2743	/* see the comment in mem_cgroup_lruvec() */
2744	if (!lruvec->pgdat)
2745	lruvec->pgdat = pgdat;
2746
2747	return lruvec;
2748	}
2749	#endif
2750	VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2751
2752	return &pgdat->__lruvec;
2753	}
2754
2755	static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2756	{
2757	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2758	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2759
2760	if (!sc->may_swap)
2761	return 0;
2762
2763	if (!can_demote(nid: pgdat->node_id, sc, memcg) &&
2764	mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2765	return 0;
2766
2767	return sc_swappiness(sc, memcg);
2768	}
2769
2770	static int get_nr_gens(struct lruvec *lruvec, int type)
2771	{
2772	return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2773	}
2774
2775	static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2776	{
2777	int type;
2778
2779	for (type = 0; type < ANON_AND_FILE; type++) {
2780	int n = get_nr_gens(lruvec, type);
2781
2782	if (n < MIN_NR_GENS || n > MAX_NR_GENS)
2783	return false;
2784	}
2785
2786	return true;
2787	}
2788
2789	/******************************************************************************
2790	* Bloom filters
2791	******************************************************************************/
2792
2793	/*
2794	* Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2795	* n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2796	* bits in a bitmap, k is the number of hash functions and n is the number of
2797	* inserted items.
2798	*
2799	* Page table walkers use one of the two filters to reduce their search space.
2800	* To get rid of non-leaf entries that no longer have enough leaf entries, the
2801	* aging uses the double-buffering technique to flip to the other filter each
2802	* time it produces a new generation. For non-leaf entries that have enough
2803	* leaf entries, the aging carries them over to the next generation in
2804	* walk_pmd_range(); the eviction also report them when walking the rmap
2805	* in lru_gen_look_around().
2806	*
2807	* For future optimizations:
2808	* 1. It's not necessary to keep both filters all the time. The spare one can be
2809	* freed after the RCU grace period and reallocated if needed again.
2810	* 2. And when reallocating, it's worth scaling its size according to the number
2811	* of inserted entries in the other filter, to reduce the memory overhead on
2812	* small systems and false positives on large systems.
2813	* 3. Jenkins' hash function is an alternative to Knuth's.
2814	*/
2815	#define BLOOM_FILTER_SHIFT 15
2816
2817	static inline int filter_gen_from_seq(unsigned long seq)
2818	{
2819	return seq % NR_BLOOM_FILTERS;
2820	}
2821
2822	static void get_item_key(void *item, int *key)
2823	{
2824	u32 hash = hash_ptr(ptr: item, BLOOM_FILTER_SHIFT * 2);
2825
2826	BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2827
2828	key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2829	key[1] = hash >> BLOOM_FILTER_SHIFT;
2830	}
2831
2832	static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2833	void *item)
2834	{
2835	int key[2];
2836	unsigned long *filter;
2837	int gen = filter_gen_from_seq(seq);
2838
2839	filter = READ_ONCE(mm_state->filters[gen]);
2840	if (!filter)
2841	return true;
2842
2843	get_item_key(item, key);
2844
2845	return test_bit(key[0], filter) && test_bit(key[1], filter);
2846	}
2847
2848	static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2849	void *item)
2850	{
2851	int key[2];
2852	unsigned long *filter;
2853	int gen = filter_gen_from_seq(seq);
2854
2855	filter = READ_ONCE(mm_state->filters[gen]);
2856	if (!filter)
2857	return;
2858
2859	get_item_key(item, key);
2860
2861	if (!test_bit(key[0], filter))
2862	set_bit(nr: key[0], addr: filter);
2863	if (!test_bit(key[1], filter))
2864	set_bit(nr: key[1], addr: filter);
2865	}
2866
2867	static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2868	{
2869	unsigned long *filter;
2870	int gen = filter_gen_from_seq(seq);
2871
2872	filter = mm_state->filters[gen];
2873	if (filter) {
2874	bitmap_clear(map: filter, start: 0, BIT(BLOOM_FILTER_SHIFT));
2875	return;
2876	}
2877
2878	filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2879	__GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2880	WRITE_ONCE(mm_state->filters[gen], filter);
2881	}
2882
2883	/******************************************************************************
2884	* mm_struct list
2885	******************************************************************************/
2886
2887	#ifdef CONFIG_LRU_GEN_WALKS_MMU
2888
2889	static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2890	{
2891	static struct lru_gen_mm_list mm_list = {
2892	.fifo = LIST_HEAD_INIT(mm_list.fifo),
2893	.lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2894	};
2895
2896	#ifdef CONFIG_MEMCG
2897	if (memcg)
2898	return &memcg->mm_list;
2899	#endif
2900	VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2901
2902	return &mm_list;
2903	}
2904
2905	static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2906	{
2907	return &lruvec->mm_state;
2908	}
2909
2910	static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2911	{
2912	int key;
2913	struct mm_struct *mm;
2914	struct pglist_data *pgdat = lruvec_pgdat(lruvec: walk->lruvec);
2915	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec: walk->lruvec);
2916
2917	mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2918	key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2919
2920	if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2921	return NULL;
2922
2923	clear_bit(nr: key, addr: &mm->lru_gen.bitmap);
2924
2925	return mmget_not_zero(mm) ? mm : NULL;
2926	}
2927
2928	void lru_gen_add_mm(struct mm_struct *mm)
2929	{
2930	int nid;
2931	struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2932	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2933
2934	VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2935	#ifdef CONFIG_MEMCG
2936	VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2937	mm->lru_gen.memcg = memcg;
2938	#endif
2939	spin_lock(lock: &mm_list->lock);
2940
2941	for_each_node_state(nid, N_MEMORY) {
2942	struct lruvec *lruvec = get_lruvec(memcg, nid);
2943	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2944
2945	/* the first addition since the last iteration */
2946	if (mm_state->tail == &mm_list->fifo)
2947	mm_state->tail = &mm->lru_gen.list;
2948	}
2949
2950	list_add_tail(new: &mm->lru_gen.list, head: &mm_list->fifo);
2951
2952	spin_unlock(lock: &mm_list->lock);
2953	}
2954
2955	void lru_gen_del_mm(struct mm_struct *mm)
2956	{
2957	int nid;
2958	struct lru_gen_mm_list *mm_list;
2959	struct mem_cgroup *memcg = NULL;
2960
2961	if (list_empty(head: &mm->lru_gen.list))
2962	return;
2963
2964	#ifdef CONFIG_MEMCG
2965	memcg = mm->lru_gen.memcg;
2966	#endif
2967	mm_list = get_mm_list(memcg);
2968
2969	spin_lock(lock: &mm_list->lock);
2970
2971	for_each_node(nid) {
2972	struct lruvec *lruvec = get_lruvec(memcg, nid);
2973	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2974
2975	/* where the current iteration continues after */
2976	if (mm_state->head == &mm->lru_gen.list)
2977	mm_state->head = mm_state->head->prev;
2978
2979	/* where the last iteration ended before */
2980	if (mm_state->tail == &mm->lru_gen.list)
2981	mm_state->tail = mm_state->tail->next;
2982	}
2983
2984	list_del_init(entry: &mm->lru_gen.list);
2985
2986	spin_unlock(lock: &mm_list->lock);
2987
2988	#ifdef CONFIG_MEMCG
2989	mem_cgroup_put(memcg: mm->lru_gen.memcg);
2990	mm->lru_gen.memcg = NULL;
2991	#endif
2992	}
2993
2994	#ifdef CONFIG_MEMCG
2995	void lru_gen_migrate_mm(struct mm_struct *mm)
2996	{
2997	struct mem_cgroup *memcg;
2998	struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2999
3000	VM_WARN_ON_ONCE(task->mm != mm);
3001	lockdep_assert_held(&task->alloc_lock);
3002
3003	/* for mm_update_next_owner() */
3004	if (mem_cgroup_disabled())
3005	return;
3006
3007	/* migration can happen before addition */
3008	if (!mm->lru_gen.memcg)
3009	return;
3010
3011	rcu_read_lock();
3012	memcg = mem_cgroup_from_task(p: task);
3013	rcu_read_unlock();
3014	if (memcg == mm->lru_gen.memcg)
3015	return;
3016
3017	VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3018
3019	lru_gen_del_mm(mm);
3020	lru_gen_add_mm(mm);
3021	}
3022	#endif
3023
3024	#else /* !CONFIG_LRU_GEN_WALKS_MMU */
3025
3026	static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3027	{
3028	return NULL;
3029	}
3030
3031	static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
3032	{
3033	return NULL;
3034	}
3035
3036	static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
3037	{
3038	return NULL;
3039	}
3040
3041	#endif
3042
3043	static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
3044	{
3045	int i;
3046	int hist;
3047	struct lruvec *lruvec = walk->lruvec;
3048	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3049
3050	lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3051
3052	hist = lru_hist_from_seq(seq: walk->seq);
3053
3054	for (i = 0; i < NR_MM_STATS; i++) {
3055	WRITE_ONCE(mm_state->stats[hist][i],
3056	mm_state->stats[hist][i] + walk->mm_stats[i]);
3057	walk->mm_stats[i] = 0;
3058	}
3059
3060	if (NR_HIST_GENS > 1 && last) {
3061	hist = lru_hist_from_seq(seq: walk->seq + 1);
3062
3063	for (i = 0; i < NR_MM_STATS; i++)
3064	WRITE_ONCE(mm_state->stats[hist][i], 0);
3065	}
3066	}
3067
3068	static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
3069	{
3070	bool first = false;
3071	bool last = false;
3072	struct mm_struct *mm = NULL;
3073	struct lruvec *lruvec = walk->lruvec;
3074	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3075	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3076	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3077
3078	/*
3079	* mm_state->seq is incremented after each iteration of mm_list. There
3080	* are three interesting cases for this page table walker:
3081	* 1. It tries to start a new iteration with a stale max_seq: there is
3082	* nothing left to do.
3083	* 2. It started the next iteration: it needs to reset the Bloom filter
3084	* so that a fresh set of PTE tables can be recorded.
3085	* 3. It ended the current iteration: it needs to reset the mm stats
3086	* counters and tell its caller to increment max_seq.
3087	*/
3088	spin_lock(lock: &mm_list->lock);
3089
3090	VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
3091
3092	if (walk->seq <= mm_state->seq)
3093	goto done;
3094
3095	if (!mm_state->head)
3096	mm_state->head = &mm_list->fifo;
3097
3098	if (mm_state->head == &mm_list->fifo)
3099	first = true;
3100
3101	do {
3102	mm_state->head = mm_state->head->next;
3103	if (mm_state->head == &mm_list->fifo) {
3104	WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3105	last = true;
3106	break;
3107	}
3108
3109	/* force scan for those added after the last iteration */
3110	if (!mm_state->tail || mm_state->tail == mm_state->head) {
3111	mm_state->tail = mm_state->head->next;
3112	walk->force_scan = true;
3113	}
3114	} while (!(mm = get_next_mm(walk)));
3115	done:
3116	if (*iter || last)
3117	reset_mm_stats(walk, last);
3118
3119	spin_unlock(lock: &mm_list->lock);
3120
3121	if (mm && first)
3122	reset_bloom_filter(mm_state, seq: walk->seq + 1);
3123
3124	if (*iter)
3125	mmput_async(*iter);
3126
3127	*iter = mm;
3128
3129	return last;
3130	}
3131
3132	static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
3133	{
3134	bool success = false;
3135	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3136	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3137	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3138
3139	spin_lock(lock: &mm_list->lock);
3140
3141	VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
3142
3143	if (seq > mm_state->seq) {
3144	mm_state->head = NULL;
3145	mm_state->tail = NULL;
3146	WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3147	success = true;
3148	}
3149
3150	spin_unlock(lock: &mm_list->lock);
3151
3152	return success;
3153	}
3154
3155	/******************************************************************************
3156	* PID controller
3157	******************************************************************************/
3158
3159	/*
3160	* A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3161	*
3162	* The P term is refaulted/(evicted+protected) from a tier in the generation
3163	* currently being evicted; the I term is the exponential moving average of the
3164	* P term over the generations previously evicted, using the smoothing factor
3165	* 1/2; the D term isn't supported.
3166	*
3167	* The setpoint (SP) is always the first tier of one type; the process variable
3168	* (PV) is either any tier of the other type or any other tier of the same
3169	* type.
3170	*
3171	* The error is the difference between the SP and the PV; the correction is to
3172	* turn off protection when SP>PV or turn on protection when SP<PV.
3173	*
3174	* For future optimizations:
3175	* 1. The D term may discount the other two terms over time so that long-lived
3176	* generations can resist stale information.
3177	*/
3178	struct ctrl_pos {
3179	unsigned long refaulted;
3180	unsigned long total;
3181	int gain;
3182	};
3183
3184	static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3185	struct ctrl_pos *pos)
3186	{
3187	int i;
3188	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3189	int hist = lru_hist_from_seq(seq: lrugen->min_seq[type]);
3190
3191	pos->gain = gain;
3192	pos->refaulted = pos->total = 0;
3193
3194	for (i = tier % MAX_NR_TIERS; i <= min(tier, MAX_NR_TIERS - 1); i++) {
3195	pos->refaulted += lrugen->avg_refaulted[type][i] +
3196	atomic_long_read(v: &lrugen->refaulted[hist][type][i]);
3197	pos->total += lrugen->avg_total[type][i] +
3198	lrugen->protected[hist][type][i] +
3199	atomic_long_read(v: &lrugen->evicted[hist][type][i]);
3200	}
3201	}
3202
3203	static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3204	{
3205	int hist, tier;
3206	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3207	bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3208	unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3209
3210	lockdep_assert_held(&lruvec->lru_lock);
3211
3212	if (!carryover && !clear)
3213	return;
3214
3215	hist = lru_hist_from_seq(seq);
3216
3217	for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3218	if (carryover) {
3219	unsigned long sum;
3220
3221	sum = lrugen->avg_refaulted[type][tier] +
3222	atomic_long_read(v: &lrugen->refaulted[hist][type][tier]);
3223	WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3224
3225	sum = lrugen->avg_total[type][tier] +
3226	lrugen->protected[hist][type][tier] +
3227	atomic_long_read(v: &lrugen->evicted[hist][type][tier]);
3228	WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3229	}
3230
3231	if (clear) {
3232	atomic_long_set(v: &lrugen->refaulted[hist][type][tier], i: 0);
3233	atomic_long_set(v: &lrugen->evicted[hist][type][tier], i: 0);
3234	WRITE_ONCE(lrugen->protected[hist][type][tier], 0);
3235	}
3236	}
3237	}
3238
3239	static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3240	{
3241	/*
3242	* Return true if the PV has a limited number of refaults or a lower
3243	* refaulted/total than the SP.
3244	*/
3245	return pv->refaulted < MIN_LRU_BATCH ||
3246	pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3247	(sp->refaulted + 1) * pv->total * pv->gain;
3248	}
3249
3250	/******************************************************************************
3251	* the aging
3252	******************************************************************************/
3253
3254	/* promote pages accessed through page tables */
3255	static int folio_update_gen(struct folio *folio, int gen)
3256	{
3257	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3258
3259	VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3260
3261	/* see the comment on LRU_REFS_FLAGS */
3262	if (!folio_test_referenced(folio) && !folio_test_workingset(folio)) {
3263	set_mask_bits(&folio->flags, LRU_REFS_MASK, BIT(PG_referenced));
3264	return -1;
3265	}
3266
3267	do {
3268	/* lru_gen_del_folio() has isolated this page? */
3269	if (!(old_flags & LRU_GEN_MASK))
3270	return -1;
3271
3272	new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_FLAGS);
3273	new_flags |= ((gen + 1UL) << LRU_GEN_PGOFF) | BIT(PG_workingset);
3274	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3275
3276	return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3277	}
3278
3279	/* protect pages accessed multiple times through file descriptors */
3280	static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3281	{
3282	int type = folio_is_file_lru(folio);
3283	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3284	int new_gen, old_gen = lru_gen_from_seq(seq: lrugen->min_seq[type]);
3285	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3286
3287	VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3288
3289	do {
3290	new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3291	/* folio_update_gen() has promoted this page? */
3292	if (new_gen >= 0 && new_gen != old_gen)
3293	return new_gen;
3294
3295	new_gen = (old_gen + 1) % MAX_NR_GENS;
3296
3297	new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_FLAGS);
3298	new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3299	/* for folio_end_writeback() */
3300	if (reclaiming)
3301	new_flags |= BIT(PG_reclaim);
3302	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3303
3304	lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3305
3306	return new_gen;
3307	}
3308
3309	static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3310	int old_gen, int new_gen)
3311	{
3312	int type = folio_is_file_lru(folio);
3313	int zone = folio_zonenum(folio);
3314	int delta = folio_nr_pages(folio);
3315
3316	VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3317	VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3318
3319	walk->batched++;
3320
3321	walk->nr_pages[old_gen][type][zone] -= delta;
3322	walk->nr_pages[new_gen][type][zone] += delta;
3323	}
3324
3325	static void reset_batch_size(struct lru_gen_mm_walk *walk)
3326	{
3327	int gen, type, zone;
3328	struct lruvec *lruvec = walk->lruvec;
3329	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3330
3331	walk->batched = 0;
3332
3333	for_each_gen_type_zone(gen, type, zone) {
3334	enum lru_list lru = type * LRU_INACTIVE_FILE;
3335	int delta = walk->nr_pages[gen][type][zone];
3336
3337	if (!delta)
3338	continue;
3339
3340	walk->nr_pages[gen][type][zone] = 0;
3341	WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3342	lrugen->nr_pages[gen][type][zone] + delta);
3343
3344	if (lru_gen_is_active(lruvec, gen))
3345	lru += LRU_ACTIVE;
3346	__update_lru_size(lruvec, lru, zone, delta);
3347	}
3348	}
3349
3350	static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3351	{
3352	struct address_space *mapping;
3353	struct vm_area_struct *vma = args->vma;
3354	struct lru_gen_mm_walk *walk = args->private;
3355
3356	if (!vma_is_accessible(vma))
3357	return true;
3358
3359	if (is_vm_hugetlb_page(vma))
3360	return true;
3361
3362	if (!vma_has_recency(vma))
3363	return true;
3364
3365	if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3366	return true;
3367
3368	if (vma == get_gate_vma(mm: vma->vm_mm))
3369	return true;
3370
3371	if (vma_is_anonymous(vma))
3372	return !walk->swappiness;
3373
3374	if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3375	return true;
3376
3377	mapping = vma->vm_file->f_mapping;
3378	if (mapping_unevictable(mapping))
3379	return true;
3380
3381	if (shmem_mapping(mapping))
3382	return !walk->swappiness;
3383
3384	if (walk->swappiness > MAX_SWAPPINESS)
3385	return true;
3386
3387	/* to exclude special mappings like dax, etc. */
3388	return !mapping->a_ops->read_folio;
3389	}
3390
3391	/*
3392	* Some userspace memory allocators map many single-page VMAs. Instead of
3393	* returning back to the PGD table for each of such VMAs, finish an entire PMD
3394	* table to reduce zigzags and improve cache performance.
3395	*/
3396	static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3397	unsigned long *vm_start, unsigned long *vm_end)
3398	{
3399	unsigned long start = round_up(*vm_end, size);
3400	unsigned long end = (start | ~mask) + 1;
3401	VMA_ITERATOR(vmi, args->mm, start);
3402
3403	VM_WARN_ON_ONCE(mask & size);
3404	VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3405
3406	for_each_vma(vmi, args->vma) {
3407	if (end && end <= args->vma->vm_start)
3408	return false;
3409
3410	if (should_skip_vma(start: args->vma->vm_start, end: args->vma->vm_end, args))
3411	continue;
3412
3413	*vm_start = max(start, args->vma->vm_start);
3414	*vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3415
3416	return true;
3417	}
3418
3419	return false;
3420	}
3421
3422	static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr,
3423	struct pglist_data *pgdat)
3424	{
3425	unsigned long pfn = pte_pfn(pte);
3426
3427	VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3428
3429	if (!pte_present(a: pte) || is_zero_pfn(pfn))
3430	return -1;
3431
3432	if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3433	return -1;
3434
3435	if (!pte_young(pte) && !mm_has_notifiers(mm: vma->vm_mm))
3436	return -1;
3437
3438	if (WARN_ON_ONCE(!pfn_valid(pfn)))
3439	return -1;
3440
3441	if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3442	return -1;
3443
3444	return pfn;
3445	}
3446
3447	static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr,
3448	struct pglist_data *pgdat)
3449	{
3450	unsigned long pfn = pmd_pfn(pmd);
3451
3452	VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3453
3454	if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3455	return -1;
3456
3457	if (WARN_ON_ONCE(pmd_devmap(pmd)))
3458	return -1;
3459
3460	if (!pmd_young(pmd) && !mm_has_notifiers(mm: vma->vm_mm))
3461	return -1;
3462
3463	if (WARN_ON_ONCE(!pfn_valid(pfn)))
3464	return -1;
3465
3466	if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3467	return -1;
3468
3469	return pfn;
3470	}
3471
3472	static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3473	struct pglist_data *pgdat)
3474	{
3475	struct folio *folio = pfn_folio(pfn);
3476
3477	if (folio_lru_gen(folio) < 0)
3478	return NULL;
3479
3480	if (folio_nid(folio) != pgdat->node_id)
3481	return NULL;
3482
3483	if (folio_memcg(folio) != memcg)
3484	return NULL;
3485
3486	return folio;
3487	}
3488
3489	static bool suitable_to_scan(int total, int young)
3490	{
3491	int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3492
3493	/* suitable if the average number of young PTEs per cacheline is >=1 */
3494	return young * n >= total;
3495	}
3496
3497	static void walk_update_folio(struct lru_gen_mm_walk *walk, struct folio *folio,
3498	int new_gen, bool dirty)
3499	{
3500	int old_gen;
3501
3502	if (!folio)
3503	return;
3504
3505	if (dirty && !folio_test_dirty(folio) &&
3506	!(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3507	!folio_test_swapcache(folio)))
3508	folio_mark_dirty(folio);
3509
3510	if (walk) {
3511	old_gen = folio_update_gen(folio, gen: new_gen);
3512	if (old_gen >= 0 && old_gen != new_gen)
3513	update_batch_size(walk, folio, old_gen, new_gen);
3514	} else if (lru_gen_set_refs(folio)) {
3515	old_gen = folio_lru_gen(folio);
3516	if (old_gen >= 0 && old_gen != new_gen)
3517	folio_activate(folio);
3518	}
3519	}
3520
3521	static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3522	struct mm_walk *args)
3523	{
3524	int i;
3525	bool dirty;
3526	pte_t *pte;
3527	spinlock_t *ptl;
3528	unsigned long addr;
3529	int total = 0;
3530	int young = 0;
3531	struct folio *last = NULL;
3532	struct lru_gen_mm_walk *walk = args->private;
3533	struct mem_cgroup *memcg = lruvec_memcg(lruvec: walk->lruvec);
3534	struct pglist_data *pgdat = lruvec_pgdat(lruvec: walk->lruvec);
3535	DEFINE_MAX_SEQ(walk->lruvec);
3536	int gen = lru_gen_from_seq(seq: max_seq);
3537	pmd_t pmdval;
3538
3539	pte = pte_offset_map_rw_nolock(mm: args->mm, pmd, addr: start & PMD_MASK, pmdvalp: &pmdval, ptlp: &ptl);
3540	if (!pte)
3541	return false;
3542
3543	if (!spin_trylock(lock: ptl)) {
3544	pte_unmap(pte);
3545	return true;
3546	}
3547
3548	if (unlikely(!pmd_same(pmdval, pmdp_get_lockless(pmd)))) {
3549	pte_unmap_unlock(pte, ptl);
3550	return false;
3551	}
3552
3553	arch_enter_lazy_mmu_mode();
3554	restart:
3555	for (i = pte_index(address: start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3556	unsigned long pfn;
3557	struct folio *folio;
3558	pte_t ptent = ptep_get(ptep: pte + i);
3559
3560	total++;
3561	walk->mm_stats[MM_LEAF_TOTAL]++;
3562
3563	pfn = get_pte_pfn(pte: ptent, vma: args->vma, addr, pgdat);
3564	if (pfn == -1)
3565	continue;
3566
3567	folio = get_pfn_folio(pfn, memcg, pgdat);
3568	if (!folio)
3569	continue;
3570
3571	if (!ptep_clear_young_notify(args->vma, addr, pte + i))
3572	continue;
3573
3574	if (last != folio) {
3575	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
3576
3577	last = folio;
3578	dirty = false;
3579	}
3580
3581	if (pte_dirty(pte: ptent))
3582	dirty = true;
3583
3584	young++;
3585	walk->mm_stats[MM_LEAF_YOUNG]++;
3586	}
3587
3588	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
3589	last = NULL;
3590
3591	if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, vm_start: &start, vm_end: &end))
3592	goto restart;
3593
3594	arch_leave_lazy_mmu_mode();
3595	pte_unmap_unlock(pte, ptl);
3596
3597	return suitable_to_scan(total, young);
3598	}
3599
3600	static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3601	struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3602	{
3603	int i;
3604	bool dirty;
3605	pmd_t *pmd;
3606	spinlock_t *ptl;
3607	struct folio *last = NULL;
3608	struct lru_gen_mm_walk *walk = args->private;
3609	struct mem_cgroup *memcg = lruvec_memcg(lruvec: walk->lruvec);
3610	struct pglist_data *pgdat = lruvec_pgdat(lruvec: walk->lruvec);
3611	DEFINE_MAX_SEQ(walk->lruvec);
3612	int gen = lru_gen_from_seq(seq: max_seq);
3613
3614	VM_WARN_ON_ONCE(pud_leaf(*pud));
3615
3616	/* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3617	if (*first == -1) {
3618	*first = addr;
3619	bitmap_zero(dst: bitmap, MIN_LRU_BATCH);
3620	return;
3621	}
3622
3623	i = addr == -1 ? 0 : pmd_index(address: addr) - pmd_index(address: *first);
3624	if (i && i <= MIN_LRU_BATCH) {
3625	__set_bit(i - 1, bitmap);
3626	return;
3627	}
3628
3629	pmd = pmd_offset(pud, address: *first);
3630
3631	ptl = pmd_lockptr(mm: args->mm, pmd);
3632	if (!spin_trylock(lock: ptl))
3633	goto done;
3634
3635	arch_enter_lazy_mmu_mode();
3636
3637	do {
3638	unsigned long pfn;
3639	struct folio *folio;
3640
3641	/* don't round down the first address */
3642	addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3643
3644	if (!pmd_present(pmd: pmd[i]))
3645	goto next;
3646
3647	if (!pmd_trans_huge(pmd: pmd[i])) {
3648	if (!walk->force_scan && should_clear_pmd_young() &&
3649	!mm_has_notifiers(mm: args->mm))
3650	pmdp_test_and_clear_young(vma, addr, pmdp: pmd + i);
3651	goto next;
3652	}
3653
3654	pfn = get_pmd_pfn(pmd: pmd[i], vma, addr, pgdat);
3655	if (pfn == -1)
3656	goto next;
3657
3658	folio = get_pfn_folio(pfn, memcg, pgdat);
3659	if (!folio)
3660	goto next;
3661
3662	if (!pmdp_clear_young_notify(vma, addr, pmd + i))
3663	goto next;
3664
3665	if (last != folio) {
3666	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
3667
3668	last = folio;
3669	dirty = false;
3670	}
3671
3672	if (pmd_dirty(pmd: pmd[i]))
3673	dirty = true;
3674
3675	walk->mm_stats[MM_LEAF_YOUNG]++;
3676	next:
3677	i = i > MIN_LRU_BATCH ? 0 : find_next_bit(addr: bitmap, MIN_LRU_BATCH, offset: i) + 1;
3678	} while (i <= MIN_LRU_BATCH);
3679
3680	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
3681
3682	arch_leave_lazy_mmu_mode();
3683	spin_unlock(lock: ptl);
3684	done:
3685	*first = -1;
3686	}
3687
3688	static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3689	struct mm_walk *args)
3690	{
3691	int i;
3692	pmd_t *pmd;
3693	unsigned long next;
3694	unsigned long addr;
3695	struct vm_area_struct *vma;
3696	DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3697	unsigned long first = -1;
3698	struct lru_gen_mm_walk *walk = args->private;
3699	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec: walk->lruvec);
3700
3701	VM_WARN_ON_ONCE(pud_leaf(*pud));
3702
3703	/*
3704	* Finish an entire PMD in two passes: the first only reaches to PTE
3705	* tables to avoid taking the PMD lock; the second, if necessary, takes
3706	* the PMD lock to clear the accessed bit in PMD entries.
3707	*/
3708	pmd = pmd_offset(pud, address: start & PUD_MASK);
3709	restart:
3710	/* walk_pte_range() may call get_next_vma() */
3711	vma = args->vma;
3712	for (i = pmd_index(address: start), addr = start; addr != end; i++, addr = next) {
3713	pmd_t val = pmdp_get_lockless(pmdp: pmd + i);
3714
3715	next = pmd_addr_end(addr, end);
3716
3717	if (!pmd_present(pmd: val) || is_huge_zero_pmd(pmd: val)) {
3718	walk->mm_stats[MM_LEAF_TOTAL]++;
3719	continue;
3720	}
3721
3722	if (pmd_trans_huge(pmd: val)) {
3723	struct pglist_data *pgdat = lruvec_pgdat(lruvec: walk->lruvec);
3724	unsigned long pfn = get_pmd_pfn(pmd: val, vma, addr, pgdat);
3725
3726	walk->mm_stats[MM_LEAF_TOTAL]++;
3727
3728	if (pfn != -1)
3729	walk_pmd_range_locked(pud, addr, vma, args, bitmap, first: &first);
3730	continue;
3731	}
3732
3733	if (!walk->force_scan && should_clear_pmd_young() &&
3734	!mm_has_notifiers(mm: args->mm)) {
3735	if (!pmd_young(pmd: val))
3736	continue;
3737
3738	walk_pmd_range_locked(pud, addr, vma, args, bitmap, first: &first);
3739	}
3740
3741	if (!walk->force_scan && !test_bloom_filter(mm_state, seq: walk->seq, item: pmd + i))
3742	continue;
3743
3744	walk->mm_stats[MM_NONLEAF_FOUND]++;
3745
3746	if (!walk_pte_range(pmd: &val, start: addr, end: next, args))
3747	continue;
3748
3749	walk->mm_stats[MM_NONLEAF_ADDED]++;
3750
3751	/* carry over to the next generation */
3752	update_bloom_filter(mm_state, seq: walk->seq + 1, item: pmd + i);
3753	}
3754
3755	walk_pmd_range_locked(pud, addr: -1, vma, args, bitmap, first: &first);
3756
3757	if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, vm_start: &start, vm_end: &end))
3758	goto restart;
3759	}
3760
3761	static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3762	struct mm_walk *args)
3763	{
3764	int i;
3765	pud_t *pud;
3766	unsigned long addr;
3767	unsigned long next;
3768	struct lru_gen_mm_walk *walk = args->private;
3769
3770	VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3771
3772	pud = pud_offset(p4d, address: start & P4D_MASK);
3773	restart:
3774	for (i = pud_index(address: start), addr = start; addr != end; i++, addr = next) {
3775	pud_t val = READ_ONCE(pud[i]);
3776
3777	next = pud_addr_end(addr, end);
3778
3779	if (!pud_present(pud: val) || WARN_ON_ONCE(pud_leaf(val)))
3780	continue;
3781
3782	walk_pmd_range(pud: &val, start: addr, end: next, args);
3783
3784	if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3785	end = (addr | ~PUD_MASK) + 1;
3786	goto done;
3787	}
3788	}
3789
3790	if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, vm_start: &start, vm_end: &end))
3791	goto restart;
3792
3793	end = round_up(end, P4D_SIZE);
3794	done:
3795	if (!end || !args->vma)
3796	return 1;
3797
3798	walk->next_addr = max(end, args->vma->vm_start);
3799
3800	return -EAGAIN;
3801	}
3802
3803	static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3804	{
3805	static const struct mm_walk_ops mm_walk_ops = {
3806	.test_walk = should_skip_vma,
3807	.p4d_entry = walk_pud_range,
3808	.walk_lock = PGWALK_RDLOCK,
3809	};
3810	int err;
3811	struct lruvec *lruvec = walk->lruvec;
3812
3813	walk->next_addr = FIRST_USER_ADDRESS;
3814
3815	do {
3816	DEFINE_MAX_SEQ(lruvec);
3817
3818	err = -EBUSY;
3819
3820	/* another thread might have called inc_max_seq() */
3821	if (walk->seq != max_seq)
3822	break;
3823
3824	/* the caller might be holding the lock for write */
3825	if (mmap_read_trylock(mm)) {
3826	err = walk_page_range(mm, start: walk->next_addr, ULONG_MAX, ops: &mm_walk_ops, private: walk);
3827
3828	mmap_read_unlock(mm);
3829	}
3830
3831	if (walk->batched) {
3832	spin_lock_irq(lock: &lruvec->lru_lock);
3833	reset_batch_size(walk);
3834	spin_unlock_irq(lock: &lruvec->lru_lock);
3835	}
3836
3837	cond_resched();
3838	} while (err == -EAGAIN);
3839	}
3840
3841	static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3842	{
3843	struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3844
3845	if (pgdat && current_is_kswapd()) {
3846	VM_WARN_ON_ONCE(walk);
3847
3848	walk = &pgdat->mm_walk;
3849	} else if (!walk && force_alloc) {
3850	VM_WARN_ON_ONCE(current_is_kswapd());
3851
3852	walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3853	}
3854
3855	current->reclaim_state->mm_walk = walk;
3856
3857	return walk;
3858	}
3859
3860	static void clear_mm_walk(void)
3861	{
3862	struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3863
3864	VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3865	VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3866
3867	current->reclaim_state->mm_walk = NULL;
3868
3869	if (!current_is_kswapd())
3870	kfree(objp: walk);
3871	}
3872
3873	static bool inc_min_seq(struct lruvec *lruvec, int type, int swappiness)
3874	{
3875	int zone;
3876	int remaining = MAX_LRU_BATCH;
3877	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3878	int hist = lru_hist_from_seq(seq: lrugen->min_seq[type]);
3879	int new_gen, old_gen = lru_gen_from_seq(seq: lrugen->min_seq[type]);
3880
3881	/* For file type, skip the check if swappiness is anon only */
3882	if (type && (swappiness == SWAPPINESS_ANON_ONLY))
3883	goto done;
3884
3885	/* For anon type, skip the check if swappiness is zero (file only) */
3886	if (!type && !swappiness)
3887	goto done;
3888
3889	/* prevent cold/hot inversion if the type is evictable */
3890	for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3891	struct list_head *head = &lrugen->folios[old_gen][type][zone];
3892
3893	while (!list_empty(head)) {
3894	struct folio *folio = lru_to_folio(head);
3895	int refs = folio_lru_refs(folio);
3896	bool workingset = folio_test_workingset(folio);
3897
3898	VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3899	VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3900	VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3901	VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3902
3903	new_gen = folio_inc_gen(lruvec, folio, false);
3904	list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3905
3906	/* don't count the workingset being lazily promoted */
3907	if (refs + workingset != BIT(LRU_REFS_WIDTH) + 1) {
3908	int tier = lru_tier_from_refs(refs, workingset);
3909	int delta = folio_nr_pages(folio);
3910
3911	WRITE_ONCE(lrugen->protected[hist][type][tier],
3912	lrugen->protected[hist][type][tier] + delta);
3913	}
3914
3915	if (!--remaining)
3916	return false;
3917	}
3918	}
3919	done:
3920	reset_ctrl_pos(lruvec, type, carryover: true);
3921	WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3922
3923	return true;
3924	}
3925
3926	static bool try_to_inc_min_seq(struct lruvec *lruvec, int swappiness)
3927	{
3928	int gen, type, zone;
3929	bool success = false;
3930	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3931	DEFINE_MIN_SEQ(lruvec);
3932
3933	VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3934
3935	/* find the oldest populated generation */
3936	for_each_evictable_type(type, swappiness) {
3937	while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3938	gen = lru_gen_from_seq(seq: min_seq[type]);
3939
3940	for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3941	if (!list_empty(&lrugen->folios[gen][type][zone]))
3942	goto next;
3943	}
3944
3945	min_seq[type]++;
3946	}
3947	next:
3948	;
3949	}
3950
3951	/* see the comment on lru_gen_folio */
3952	if (swappiness && swappiness <= MAX_SWAPPINESS) {
3953	unsigned long seq = lrugen->max_seq - MIN_NR_GENS;
3954
3955	if (min_seq[LRU_GEN_ANON] > seq && min_seq[LRU_GEN_FILE] < seq)
3956	min_seq[LRU_GEN_ANON] = seq;
3957	else if (min_seq[LRU_GEN_FILE] > seq && min_seq[LRU_GEN_ANON] < seq)
3958	min_seq[LRU_GEN_FILE] = seq;
3959	}
3960
3961	for_each_evictable_type(type, swappiness) {
3962	if (min_seq[type] <= lrugen->min_seq[type])
3963	continue;
3964
3965	reset_ctrl_pos(lruvec, type, carryover: true);
3966	WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3967	success = true;
3968	}
3969
3970	return success;
3971	}
3972
3973	static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq, int swappiness)
3974	{
3975	bool success;
3976	int prev, next;
3977	int type, zone;
3978	struct lru_gen_folio *lrugen = &lruvec->lrugen;
3979	restart:
3980	if (seq < READ_ONCE(lrugen->max_seq))
3981	return false;
3982
3983	spin_lock_irq(lock: &lruvec->lru_lock);
3984
3985	VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3986
3987	success = seq == lrugen->max_seq;
3988	if (!success)
3989	goto unlock;
3990
3991	for (type = 0; type < ANON_AND_FILE; type++) {
3992	if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3993	continue;
3994
3995	if (inc_min_seq(lruvec, type, swappiness))
3996	continue;
3997
3998	spin_unlock_irq(lock: &lruvec->lru_lock);
3999	cond_resched();
4000	goto restart;
4001	}
4002
4003	/*
4004	* Update the active/inactive LRU sizes for compatibility. Both sides of
4005	* the current max_seq need to be covered, since max_seq+1 can overlap
4006	* with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4007	* overlap, cold/hot inversion happens.
4008	*/
4009	prev = lru_gen_from_seq(seq: lrugen->max_seq - 1);
4010	next = lru_gen_from_seq(seq: lrugen->max_seq + 1);
4011
4012	for (type = 0; type < ANON_AND_FILE; type++) {
4013	for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4014	enum lru_list lru = type * LRU_INACTIVE_FILE;
4015	long delta = lrugen->nr_pages[prev][type][zone] -
4016	lrugen->nr_pages[next][type][zone];
4017
4018	if (!delta)
4019	continue;
4020
4021	__update_lru_size(lruvec, lru, zone, delta);
4022	__update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4023	}
4024	}
4025
4026	for (type = 0; type < ANON_AND_FILE; type++)
4027	reset_ctrl_pos(lruvec, type, carryover: false);
4028
4029	WRITE_ONCE(lrugen->timestamps[next], jiffies);
4030	/* make sure preceding modifications appear */
4031	smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4032	unlock:
4033	spin_unlock_irq(lock: &lruvec->lru_lock);
4034
4035	return success;
4036	}
4037
4038	static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
4039	int swappiness, bool force_scan)
4040	{
4041	bool success;
4042	struct lru_gen_mm_walk *walk;
4043	struct mm_struct *mm = NULL;
4044	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4045	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4046
4047	VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
4048
4049	if (!mm_state)
4050	return inc_max_seq(lruvec, seq, swappiness);
4051
4052	/* see the comment in iterate_mm_list() */
4053	if (seq <= READ_ONCE(mm_state->seq))
4054	return false;
4055
4056	/*
4057	* If the hardware doesn't automatically set the accessed bit, fallback
4058	* to lru_gen_look_around(), which only clears the accessed bit in a
4059	* handful of PTEs. Spreading the work out over a period of time usually
4060	* is less efficient, but it avoids bursty page faults.
4061	*/
4062	if (!should_walk_mmu()) {
4063	success = iterate_mm_list_nowalk(lruvec, seq);
4064	goto done;
4065	}
4066
4067	walk = set_mm_walk(NULL, force_alloc: true);
4068	if (!walk) {
4069	success = iterate_mm_list_nowalk(lruvec, seq);
4070	goto done;
4071	}
4072
4073	walk->lruvec = lruvec;
4074	walk->seq = seq;
4075	walk->swappiness = swappiness;
4076	walk->force_scan = force_scan;
4077
4078	do {
4079	success = iterate_mm_list(walk, iter: &mm);
4080	if (mm)
4081	walk_mm(mm, walk);
4082	} while (mm);
4083	done:
4084	if (success) {
4085	success = inc_max_seq(lruvec, seq, swappiness);
4086	WARN_ON_ONCE(!success);
4087	}
4088
4089	return success;
4090	}
4091
4092	/******************************************************************************
4093	* working set protection
4094	******************************************************************************/
4095
4096	static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4097	{
4098	int priority;
4099	unsigned long reclaimable;
4100
4101	if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4102	return;
4103	/*
4104	* Determine the initial priority based on
4105	* (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4106	* where reclaimed_to_scanned_ratio = inactive / total.
4107	*/
4108	reclaimable = node_page_state(pgdat, item: NR_INACTIVE_FILE);
4109	if (can_reclaim_anon_pages(NULL, nid: pgdat->node_id, sc))
4110	reclaimable += node_page_state(pgdat, item: NR_INACTIVE_ANON);
4111
4112	/* round down reclaimable and round up sc->nr_to_reclaim */
4113	priority = fls_long(l: reclaimable) - 1 - fls_long(l: sc->nr_to_reclaim - 1);
4114
4115	/*
4116	* The estimation is based on LRU pages only, so cap it to prevent
4117	* overshoots of shrinker objects by large margins.
4118	*/
4119	sc->priority = clamp(priority, DEF_PRIORITY / 2, DEF_PRIORITY);
4120	}
4121
4122	static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
4123	{
4124	int gen, type, zone;
4125	unsigned long total = 0;
4126	int swappiness = get_swappiness(lruvec, sc);
4127	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4128	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4129	DEFINE_MAX_SEQ(lruvec);
4130	DEFINE_MIN_SEQ(lruvec);
4131
4132	for_each_evictable_type(type, swappiness) {
4133	unsigned long seq;
4134
4135	for (seq = min_seq[type]; seq <= max_seq; seq++) {
4136	gen = lru_gen_from_seq(seq);
4137
4138	for (zone = 0; zone < MAX_NR_ZONES; zone++)
4139	total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4140	}
4141	}
4142
4143	/* whether the size is big enough to be helpful */
4144	return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4145	}
4146
4147	static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
4148	unsigned long min_ttl)
4149	{
4150	int gen;
4151	unsigned long birth;
4152	int swappiness = get_swappiness(lruvec, sc);
4153	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4154	DEFINE_MIN_SEQ(lruvec);
4155
4156	if (mem_cgroup_below_min(NULL, memcg))
4157	return false;
4158
4159	if (!lruvec_is_sizable(lruvec, sc))
4160	return false;
4161
4162	gen = lru_gen_from_seq(evictable_min_seq(min_seq, swappiness));
4163	birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4164
4165	return time_is_before_jiffies(birth + min_ttl);
4166	}
4167
4168	/* to protect the working set of the last N jiffies */
4169	static unsigned long lru_gen_min_ttl __read_mostly;
4170
4171	static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4172	{
4173	struct mem_cgroup *memcg;
4174	unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4175	bool reclaimable = !min_ttl;
4176
4177	VM_WARN_ON_ONCE(!current_is_kswapd());
4178
4179	set_initial_priority(pgdat, sc);
4180
4181	memcg = mem_cgroup_iter(NULL, NULL, NULL);
4182	do {
4183	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4184
4185	mem_cgroup_calculate_protection(NULL, memcg);
4186
4187	if (!reclaimable)
4188	reclaimable = lruvec_is_reclaimable(lruvec, sc, min_ttl);
4189	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4190
4191	/*
4192	* The main goal is to OOM kill if every generation from all memcgs is
4193	* younger than min_ttl. However, another possibility is all memcgs are
4194	* either too small or below min.
4195	*/
4196	if (!reclaimable && mutex_trylock(&oom_lock)) {
4197	struct oom_control oc = {
4198	.gfp_mask = sc->gfp_mask,
4199	};
4200
4201	out_of_memory(oc: &oc);
4202
4203	mutex_unlock(lock: &oom_lock);
4204	}
4205	}
4206
4207	/******************************************************************************
4208	* rmap/PT walk feedback
4209	******************************************************************************/
4210
4211	/*
4212	* This function exploits spatial locality when shrink_folio_list() walks the
4213	* rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4214	* the scan was done cacheline efficiently, it adds the PMD entry pointing to
4215	* the PTE table to the Bloom filter. This forms a feedback loop between the
4216	* eviction and the aging.
4217	*/
4218	bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4219	{
4220	int i;
4221	bool dirty;
4222	unsigned long start;
4223	unsigned long end;
4224	struct lru_gen_mm_walk *walk;
4225	struct folio *last = NULL;
4226	int young = 1;
4227	pte_t *pte = pvmw->pte;
4228	unsigned long addr = pvmw->address;
4229	struct vm_area_struct *vma = pvmw->vma;
4230	struct folio *folio = pfn_folio(pfn: pvmw->pfn);
4231	struct mem_cgroup *memcg = folio_memcg(folio);
4232	struct pglist_data *pgdat = folio_pgdat(folio);
4233	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4234	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4235	DEFINE_MAX_SEQ(lruvec);
4236	int gen = lru_gen_from_seq(seq: max_seq);
4237
4238	lockdep_assert_held(pvmw->ptl);
4239	VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4240
4241	if (!ptep_clear_young_notify(vma, addr, pte))
4242	return false;
4243
4244	if (spin_is_contended(lock: pvmw->ptl))
4245	return true;
4246
4247	/* exclude special VMAs containing anon pages from COW */
4248	if (vma->vm_flags & VM_SPECIAL)
4249	return true;
4250
4251	/* avoid taking the LRU lock under the PTL when possible */
4252	walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4253
4254	start = max(addr & PMD_MASK, vma->vm_start);
4255	end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4256
4257	if (end - start == PAGE_SIZE)
4258	return true;
4259
4260	if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4261	if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4262	end = start + MIN_LRU_BATCH * PAGE_SIZE;
4263	else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4264	start = end - MIN_LRU_BATCH * PAGE_SIZE;
4265	else {
4266	start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4267	end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4268	}
4269	}
4270
4271	arch_enter_lazy_mmu_mode();
4272
4273	pte -= (addr - start) / PAGE_SIZE;
4274
4275	for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4276	unsigned long pfn;
4277	pte_t ptent = ptep_get(ptep: pte + i);
4278
4279	pfn = get_pte_pfn(pte: ptent, vma, addr, pgdat);
4280	if (pfn == -1)
4281	continue;
4282
4283	folio = get_pfn_folio(pfn, memcg, pgdat);
4284	if (!folio)
4285	continue;
4286
4287	if (!ptep_clear_young_notify(vma, addr, pte + i))
4288	continue;
4289
4290	if (last != folio) {
4291	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
4292
4293	last = folio;
4294	dirty = false;
4295	}
4296
4297	if (pte_dirty(pte: ptent))
4298	dirty = true;
4299
4300	young++;
4301	}
4302
4303	walk_update_folio(walk, folio: last, new_gen: gen, dirty);
4304
4305	arch_leave_lazy_mmu_mode();
4306
4307	/* feedback from rmap walkers to page table walkers */
4308	if (mm_state && suitable_to_scan(total: i, young))
4309	update_bloom_filter(mm_state, seq: max_seq, item: pvmw->pmd);
4310
4311	return true;
4312	}
4313
4314	/******************************************************************************
4315	* memcg LRU
4316	******************************************************************************/
4317
4318	/* see the comment on MEMCG_NR_GENS */
4319	enum {
4320	MEMCG_LRU_NOP,
4321	MEMCG_LRU_HEAD,
4322	MEMCG_LRU_TAIL,
4323	MEMCG_LRU_OLD,
4324	MEMCG_LRU_YOUNG,
4325	};
4326
4327	static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4328	{
4329	int seg;
4330	int old, new;
4331	unsigned long flags;
4332	int bin = get_random_u32_below(MEMCG_NR_BINS);
4333	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4334
4335	spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4336
4337	VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4338
4339	seg = 0;
4340	new = old = lruvec->lrugen.gen;
4341
4342	/* see the comment on MEMCG_NR_GENS */
4343	if (op == MEMCG_LRU_HEAD)
4344	seg = MEMCG_LRU_HEAD;
4345	else if (op == MEMCG_LRU_TAIL)
4346	seg = MEMCG_LRU_TAIL;
4347	else if (op == MEMCG_LRU_OLD)
4348	new = get_memcg_gen(pgdat->memcg_lru.seq);
4349	else if (op == MEMCG_LRU_YOUNG)
4350	new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4351	else
4352	VM_WARN_ON_ONCE(true);
4353
4354	WRITE_ONCE(lruvec->lrugen.seg, seg);
4355	WRITE_ONCE(lruvec->lrugen.gen, new);
4356
4357	hlist_nulls_del_rcu(n: &lruvec->lrugen.list);
4358
4359	if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4360	hlist_nulls_add_head_rcu(n: &lruvec->lrugen.list, h: &pgdat->memcg_lru.fifo[new][bin]);
4361	else
4362	hlist_nulls_add_tail_rcu(n: &lruvec->lrugen.list, h: &pgdat->memcg_lru.fifo[new][bin]);
4363
4364	pgdat->memcg_lru.nr_memcgs[old]--;
4365	pgdat->memcg_lru.nr_memcgs[new]++;
4366
4367	if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4368	WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4369
4370	spin_unlock_irqrestore(lock: &pgdat->memcg_lru.lock, flags);
4371	}
4372
4373	#ifdef CONFIG_MEMCG
4374
4375	void lru_gen_online_memcg(struct mem_cgroup *memcg)
4376	{
4377	int gen;
4378	int nid;
4379	int bin = get_random_u32_below(MEMCG_NR_BINS);
4380
4381	for_each_node(nid) {
4382	struct pglist_data *pgdat = NODE_DATA(nid);
4383	struct lruvec *lruvec = get_lruvec(memcg, nid);
4384
4385	spin_lock_irq(lock: &pgdat->memcg_lru.lock);
4386
4387	VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4388
4389	gen = get_memcg_gen(pgdat->memcg_lru.seq);
4390
4391	lruvec->lrugen.gen = gen;
4392
4393	hlist_nulls_add_tail_rcu(n: &lruvec->lrugen.list, h: &pgdat->memcg_lru.fifo[gen][bin]);
4394	pgdat->memcg_lru.nr_memcgs[gen]++;
4395
4396	spin_unlock_irq(lock: &pgdat->memcg_lru.lock);
4397	}
4398	}
4399
4400	void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4401	{
4402	int nid;
4403
4404	for_each_node(nid) {
4405	struct lruvec *lruvec = get_lruvec(memcg, nid);
4406
4407	lru_gen_rotate_memcg(lruvec, op: MEMCG_LRU_OLD);
4408	}
4409	}
4410
4411	void lru_gen_release_memcg(struct mem_cgroup *memcg)
4412	{
4413	int gen;
4414	int nid;
4415
4416	for_each_node(nid) {
4417	struct pglist_data *pgdat = NODE_DATA(nid);
4418	struct lruvec *lruvec = get_lruvec(memcg, nid);
4419
4420	spin_lock_irq(lock: &pgdat->memcg_lru.lock);
4421
4422	if (hlist_nulls_unhashed(h: &lruvec->lrugen.list))
4423	goto unlock;
4424
4425	gen = lruvec->lrugen.gen;
4426
4427	hlist_nulls_del_init_rcu(n: &lruvec->lrugen.list);
4428	pgdat->memcg_lru.nr_memcgs[gen]--;
4429
4430	if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4431	WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4432	unlock:
4433	spin_unlock_irq(lock: &pgdat->memcg_lru.lock);
4434	}
4435	}
4436
4437	void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4438	{
4439	struct lruvec *lruvec = get_lruvec(memcg, nid);
4440
4441	/* see the comment on MEMCG_NR_GENS */
4442	if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4443	lru_gen_rotate_memcg(lruvec, op: MEMCG_LRU_HEAD);
4444	}
4445
4446	#endif /* CONFIG_MEMCG */
4447
4448	/******************************************************************************
4449	* the eviction
4450	******************************************************************************/
4451
4452	static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4453	int tier_idx)
4454	{
4455	bool success;
4456	bool dirty, writeback;
4457	int gen = folio_lru_gen(folio);
4458	int type = folio_is_file_lru(folio);
4459	int zone = folio_zonenum(folio);
4460	int delta = folio_nr_pages(folio);
4461	int refs = folio_lru_refs(folio);
4462	bool workingset = folio_test_workingset(folio);
4463	int tier = lru_tier_from_refs(refs, workingset);
4464	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4465
4466	VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4467
4468	/* unevictable */
4469	if (!folio_evictable(folio)) {
4470	success = lru_gen_del_folio(lruvec, folio, reclaiming: true);
4471	VM_WARN_ON_ONCE_FOLIO(!success, folio);
4472	folio_set_unevictable(folio);
4473	lruvec_add_folio(lruvec, folio);
4474	__count_vm_events(item: UNEVICTABLE_PGCULLED, delta);
4475	return true;
4476	}
4477
4478	/* promoted */
4479	if (gen != lru_gen_from_seq(seq: lrugen->min_seq[type])) {
4480	list_move(list: &folio->lru, head: &lrugen->folios[gen][type][zone]);
4481	return true;
4482	}
4483
4484	/* protected */
4485	if (tier > tier_idx || refs + workingset == BIT(LRU_REFS_WIDTH) + 1) {
4486	gen = folio_inc_gen(lruvec, folio, reclaiming: false);
4487	list_move(list: &folio->lru, head: &lrugen->folios[gen][type][zone]);
4488
4489	/* don't count the workingset being lazily promoted */
4490	if (refs + workingset != BIT(LRU_REFS_WIDTH) + 1) {
4491	int hist = lru_hist_from_seq(seq: lrugen->min_seq[type]);
4492
4493	WRITE_ONCE(lrugen->protected[hist][type][tier],
4494	lrugen->protected[hist][type][tier] + delta);
4495	}
4496	return true;
4497	}
4498
4499	/* ineligible */
4500	if (!folio_test_lru(folio) || zone > sc->reclaim_idx) {
4501	gen = folio_inc_gen(lruvec, folio, reclaiming: false);
4502	list_move_tail(list: &folio->lru, head: &lrugen->folios[gen][type][zone]);
4503	return true;
4504	}
4505
4506	dirty = folio_test_dirty(folio);
4507	writeback = folio_test_writeback(folio);
4508	if (type == LRU_GEN_FILE && dirty) {
4509	sc->nr.file_taken += delta;
4510	if (!writeback)
4511	sc->nr.unqueued_dirty += delta;
4512	}
4513
4514	/* waiting for writeback */
4515	if (writeback || (type == LRU_GEN_FILE && dirty)) {
4516	gen = folio_inc_gen(lruvec, folio, reclaiming: true);
4517	list_move(list: &folio->lru, head: &lrugen->folios[gen][type][zone]);
4518	return true;
4519	}
4520
4521	return false;
4522	}
4523
4524	static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4525	{
4526	bool success;
4527
4528	/* swap constrained */
4529	if (!(sc->gfp_mask & __GFP_IO) &&
4530	(folio_test_dirty(folio) ||
4531	(folio_test_anon(folio) && !folio_test_swapcache(folio))))
4532	return false;
4533
4534	/* raced with release_pages() */
4535	if (!folio_try_get(folio))
4536	return false;
4537
4538	/* raced with another isolation */
4539	if (!folio_test_clear_lru(folio)) {
4540	folio_put(folio);
4541	return false;
4542	}
4543
4544	/* see the comment on LRU_REFS_FLAGS */
4545	if (!folio_test_referenced(folio))
4546	set_mask_bits(&folio->flags, LRU_REFS_MASK, 0);
4547
4548	/* for shrink_folio_list() */
4549	folio_clear_reclaim(folio);
4550
4551	success = lru_gen_del_folio(lruvec, folio, reclaiming: true);
4552	VM_WARN_ON_ONCE_FOLIO(!success, folio);
4553
4554	return true;
4555	}
4556
4557	static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4558	int type, int tier, struct list_head *list)
4559	{
4560	int i;
4561	int gen;
4562	enum vm_event_item item;
4563	int sorted = 0;
4564	int scanned = 0;
4565	int isolated = 0;
4566	int skipped = 0;
4567	int remaining = MAX_LRU_BATCH;
4568	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4569	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4570
4571	VM_WARN_ON_ONCE(!list_empty(list));
4572
4573	if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4574	return 0;
4575
4576	gen = lru_gen_from_seq(seq: lrugen->min_seq[type]);
4577
4578	for (i = MAX_NR_ZONES; i > 0; i--) {
4579	LIST_HEAD(moved);
4580	int skipped_zone = 0;
4581	int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4582	struct list_head *head = &lrugen->folios[gen][type][zone];
4583
4584	while (!list_empty(head)) {
4585	struct folio *folio = lru_to_folio(head);
4586	int delta = folio_nr_pages(folio);
4587
4588	VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4589	VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4590	VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4591	VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4592
4593	scanned += delta;
4594
4595	if (sort_folio(lruvec, folio, sc, tier_idx: tier))
4596	sorted += delta;
4597	else if (isolate_folio(lruvec, folio, sc)) {
4598	list_add(new: &folio->lru, head: list);
4599	isolated += delta;
4600	} else {
4601	list_move(list: &folio->lru, head: &moved);
4602	skipped_zone += delta;
4603	}
4604
4605	if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4606	break;
4607	}
4608
4609	if (skipped_zone) {
4610	list_splice(list: &moved, head);
4611	__count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4612	skipped += skipped_zone;
4613	}
4614
4615	if (!remaining || isolated >= MIN_LRU_BATCH)
4616	break;
4617	}
4618
4619	item = PGSCAN_KSWAPD + reclaimer_offset(sc);
4620	if (!cgroup_reclaim(sc)) {
4621	__count_vm_events(item, delta: isolated);
4622	__count_vm_events(item: PGREFILL, delta: sorted);
4623	}
4624	count_memcg_events(memcg, idx: item, count: isolated);
4625	count_memcg_events(memcg, idx: PGREFILL, count: sorted);
4626	__count_vm_events(item: PGSCAN_ANON + type, delta: isolated);
4627	trace_mm_vmscan_lru_isolate(highest_zoneidx: sc->reclaim_idx, order: sc->order, MAX_LRU_BATCH,
4628	nr_scanned: scanned, nr_skipped: skipped, nr_taken: isolated,
4629	lru: type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4630	if (type == LRU_GEN_FILE)
4631	sc->nr.file_taken += isolated;
4632	/*
4633	* There might not be eligible folios due to reclaim_idx. Check the
4634	* remaining to prevent livelock if it's not making progress.
4635	*/
4636	return isolated || !remaining ? scanned : 0;
4637	}
4638
4639	static int get_tier_idx(struct lruvec *lruvec, int type)
4640	{
4641	int tier;
4642	struct ctrl_pos sp, pv;
4643
4644	/*
4645	* To leave a margin for fluctuations, use a larger gain factor (2:3).
4646	* This value is chosen because any other tier would have at least twice
4647	* as many refaults as the first tier.
4648	*/
4649	read_ctrl_pos(lruvec, type, tier: 0, gain: 2, pos: &sp);
4650	for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4651	read_ctrl_pos(lruvec, type, tier, gain: 3, pos: &pv);
4652	if (!positive_ctrl_err(sp: &sp, pv: &pv))
4653	break;
4654	}
4655
4656	return tier - 1;
4657	}
4658
4659	static int get_type_to_scan(struct lruvec *lruvec, int swappiness)
4660	{
4661	struct ctrl_pos sp, pv;
4662
4663	if (swappiness <= MIN_SWAPPINESS + 1)
4664	return LRU_GEN_FILE;
4665
4666	if (swappiness >= MAX_SWAPPINESS)
4667	return LRU_GEN_ANON;
4668	/*
4669	* Compare the sum of all tiers of anon with that of file to determine
4670	* which type to scan.
4671	*/
4672	read_ctrl_pos(lruvec, type: LRU_GEN_ANON, MAX_NR_TIERS, gain: swappiness, pos: &sp);
4673	read_ctrl_pos(lruvec, type: LRU_GEN_FILE, MAX_NR_TIERS, MAX_SWAPPINESS - swappiness, pos: &pv);
4674
4675	return positive_ctrl_err(sp: &sp, pv: &pv);
4676	}
4677
4678	static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4679	int *type_scanned, struct list_head *list)
4680	{
4681	int i;
4682	int type = get_type_to_scan(lruvec, swappiness);
4683
4684	for_each_evictable_type(i, swappiness) {
4685	int scanned;
4686	int tier = get_tier_idx(lruvec, type);
4687
4688	*type_scanned = type;
4689
4690	scanned = scan_folios(lruvec, sc, type, tier, list);
4691	if (scanned)
4692	return scanned;
4693
4694	type = !type;
4695	}
4696
4697	return 0;
4698	}
4699
4700	static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4701	{
4702	int type;
4703	int scanned;
4704	int reclaimed;
4705	LIST_HEAD(list);
4706	LIST_HEAD(clean);
4707	struct folio *folio;
4708	struct folio *next;
4709	enum vm_event_item item;
4710	struct reclaim_stat stat;
4711	struct lru_gen_mm_walk *walk;
4712	bool skip_retry = false;
4713	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4714	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4715	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4716
4717	spin_lock_irq(lock: &lruvec->lru_lock);
4718
4719	scanned = isolate_folios(lruvec, sc, swappiness, type_scanned: &type, list: &list);
4720
4721	scanned += try_to_inc_min_seq(lruvec, swappiness);
4722
4723	if (evictable_min_seq(lrugen->min_seq, swappiness) + MIN_NR_GENS > lrugen->max_seq)
4724	scanned = 0;
4725
4726	spin_unlock_irq(lock: &lruvec->lru_lock);
4727
4728	if (list_empty(head: &list))
4729	return scanned;
4730	retry:
4731	reclaimed = shrink_folio_list(folio_list: &list, pgdat, sc, stat: &stat, ignore_references: false, memcg);
4732	sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
4733	sc->nr_reclaimed += reclaimed;
4734	trace_mm_vmscan_lru_shrink_inactive(nid: pgdat->node_id,
4735	nr_scanned: scanned, nr_reclaimed: reclaimed, stat: &stat, priority: sc->priority,
4736	file: type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4737
4738	list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4739	DEFINE_MIN_SEQ(lruvec);
4740
4741	if (!folio_evictable(folio)) {
4742	list_del(entry: &folio->lru);
4743	folio_putback_lru(folio);
4744	continue;
4745	}
4746
4747	/* retry folios that may have missed folio_rotate_reclaimable() */
4748	if (!skip_retry && !folio_test_active(folio) && !folio_mapped(folio) &&
4749	!folio_test_dirty(folio) && !folio_test_writeback(folio)) {
4750	list_move(list: &folio->lru, head: &clean);
4751	continue;
4752	}
4753
4754	/* don't add rejected folios to the oldest generation */
4755	if (lru_gen_folio_seq(lruvec, folio, reclaiming: false) == min_seq[type])
4756	set_mask_bits(&folio->flags, LRU_REFS_FLAGS, BIT(PG_active));
4757	}
4758
4759	spin_lock_irq(lock: &lruvec->lru_lock);
4760
4761	move_folios_to_lru(lruvec, list: &list);
4762
4763	walk = current->reclaim_state->mm_walk;
4764	if (walk && walk->batched) {
4765	walk->lruvec = lruvec;
4766	reset_batch_size(walk);
4767	}
4768
4769	__mod_lruvec_state(lruvec, idx: PGDEMOTE_KSWAPD + reclaimer_offset(sc),
4770	val: stat.nr_demoted);
4771
4772	item = PGSTEAL_KSWAPD + reclaimer_offset(sc);
4773	if (!cgroup_reclaim(sc))
4774	__count_vm_events(item, delta: reclaimed);
4775	count_memcg_events(memcg, idx: item, count: reclaimed);
4776	__count_vm_events(item: PGSTEAL_ANON + type, delta: reclaimed);
4777
4778	spin_unlock_irq(lock: &lruvec->lru_lock);
4779
4780	list_splice_init(list: &clean, head: &list);
4781
4782	if (!list_empty(head: &list)) {
4783	skip_retry = true;
4784	goto retry;
4785	}
4786
4787	return scanned;
4788	}
4789
4790	static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4791	int swappiness, unsigned long *nr_to_scan)
4792	{
4793	int gen, type, zone;
4794	unsigned long size = 0;
4795	struct lru_gen_folio *lrugen = &lruvec->lrugen;
4796	DEFINE_MIN_SEQ(lruvec);
4797
4798	*nr_to_scan = 0;
4799	/* have to run aging, since eviction is not possible anymore */
4800	if (evictable_min_seq(min_seq, swappiness) + MIN_NR_GENS > max_seq)
4801	return true;
4802
4803	for_each_evictable_type(type, swappiness) {
4804	unsigned long seq;
4805
4806	for (seq = min_seq[type]; seq <= max_seq; seq++) {
4807	gen = lru_gen_from_seq(seq);
4808
4809	for (zone = 0; zone < MAX_NR_ZONES; zone++)
4810	size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4811	}
4812	}
4813
4814	*nr_to_scan = size;
4815	/* better to run aging even though eviction is still possible */
4816	return evictable_min_seq(min_seq, swappiness) + MIN_NR_GENS == max_seq;
4817	}
4818
4819	/*
4820	* For future optimizations:
4821	* 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4822	* reclaim.
4823	*/
4824	static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4825	{
4826	bool success;
4827	unsigned long nr_to_scan;
4828	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4829	DEFINE_MAX_SEQ(lruvec);
4830
4831	if (mem_cgroup_below_min(target: sc->target_mem_cgroup, memcg))
4832	return -1;
4833
4834	success = should_run_aging(lruvec, max_seq, swappiness, nr_to_scan: &nr_to_scan);
4835
4836	/* try to scrape all its memory if this memcg was deleted */
4837	if (nr_to_scan && !mem_cgroup_online(memcg))
4838	return nr_to_scan;
4839
4840	/* try to get away with not aging at the default priority */
4841	if (!success || sc->priority == DEF_PRIORITY)
4842	return nr_to_scan >> sc->priority;
4843
4844	/* stop scanning this lruvec as it's low on cold folios */
4845	return try_to_inc_max_seq(lruvec, seq: max_seq, swappiness, force_scan: false) ? -1 : 0;
4846	}
4847
4848	static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4849	{
4850	int i;
4851	enum zone_watermarks mark;
4852
4853	/* don't abort memcg reclaim to ensure fairness */
4854	if (!root_reclaim(sc))
4855	return false;
4856
4857	if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4858	return true;
4859
4860	/* check the order to exclude compaction-induced reclaim */
4861	if (!current_is_kswapd() || sc->order)
4862	return false;
4863
4864	mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4865	WMARK_PROMO : WMARK_HIGH;
4866
4867	for (i = 0; i <= sc->reclaim_idx; i++) {
4868	struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4869	unsigned long size = wmark_pages(z: zone, w: mark) + MIN_LRU_BATCH;
4870
4871	if (managed_zone(zone) && !zone_watermark_ok(z: zone, order: 0, mark: size, highest_zoneidx: sc->reclaim_idx, alloc_flags: 0))
4872	return false;
4873	}
4874
4875	/* kswapd should abort if all eligible zones are safe */
4876	return true;
4877	}
4878
4879	static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4880	{
4881	long nr_to_scan;
4882	unsigned long scanned = 0;
4883	int swappiness = get_swappiness(lruvec, sc);
4884
4885	while (true) {
4886	int delta;
4887
4888	nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4889	if (nr_to_scan <= 0)
4890	break;
4891
4892	delta = evict_folios(lruvec, sc, swappiness);
4893	if (!delta)
4894	break;
4895
4896	scanned += delta;
4897	if (scanned >= nr_to_scan)
4898	break;
4899
4900	if (should_abort_scan(lruvec, sc))
4901	break;
4902
4903	cond_resched();
4904	}
4905
4906	/*
4907	* If too many file cache in the coldest generation can't be evicted
4908	* due to being dirty, wake up the flusher.
4909	*/
4910	if (sc->nr.unqueued_dirty && sc->nr.unqueued_dirty == sc->nr.file_taken)
4911	wakeup_flusher_threads(reason: WB_REASON_VMSCAN);
4912
4913	/* whether this lruvec should be rotated */
4914	return nr_to_scan < 0;
4915	}
4916
4917	static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4918	{
4919	bool success;
4920	unsigned long scanned = sc->nr_scanned;
4921	unsigned long reclaimed = sc->nr_reclaimed;
4922	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4923	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4924
4925	/* lru_gen_age_node() called mem_cgroup_calculate_protection() */
4926	if (mem_cgroup_below_min(NULL, memcg))
4927	return MEMCG_LRU_YOUNG;
4928
4929	if (mem_cgroup_below_low(NULL, memcg)) {
4930	/* see the comment on MEMCG_NR_GENS */
4931	if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4932	return MEMCG_LRU_TAIL;
4933
4934	memcg_memory_event(memcg, event: MEMCG_LOW);
4935	}
4936
4937	success = try_to_shrink_lruvec(lruvec, sc);
4938
4939	shrink_slab(gfp_mask: sc->gfp_mask, nid: pgdat->node_id, memcg, priority: sc->priority);
4940
4941	if (!sc->proactive)
4942	vmpressure(gfp: sc->gfp_mask, memcg, tree: false, scanned: sc->nr_scanned - scanned,
4943	reclaimed: sc->nr_reclaimed - reclaimed);
4944
4945	flush_reclaim_state(sc);
4946
4947	if (success && mem_cgroup_online(memcg))
4948	return MEMCG_LRU_YOUNG;
4949
4950	if (!success && lruvec_is_sizable(lruvec, sc))
4951	return 0;
4952
4953	/* one retry if offlined or too small */
4954	return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4955	MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4956	}
4957
4958	static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4959	{
4960	int op;
4961	int gen;
4962	int bin;
4963	int first_bin;
4964	struct lruvec *lruvec;
4965	struct lru_gen_folio *lrugen;
4966	struct mem_cgroup *memcg;
4967	struct hlist_nulls_node *pos;
4968
4969	gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4970	bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4971	restart:
4972	op = 0;
4973	memcg = NULL;
4974
4975	rcu_read_lock();
4976
4977	hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4978	if (op) {
4979	lru_gen_rotate_memcg(lruvec, op);
4980	op = 0;
4981	}
4982
4983	mem_cgroup_put(memcg);
4984	memcg = NULL;
4985
4986	if (gen != READ_ONCE(lrugen->gen))
4987	continue;
4988
4989	lruvec = container_of(lrugen, struct lruvec, lrugen);
4990	memcg = lruvec_memcg(lruvec);
4991
4992	if (!mem_cgroup_tryget(memcg)) {
4993	lru_gen_release_memcg(memcg);
4994	memcg = NULL;
4995	continue;
4996	}
4997
4998	rcu_read_unlock();
4999
5000	op = shrink_one(lruvec, sc);
5001
5002	rcu_read_lock();
5003
5004	if (should_abort_scan(lruvec, sc))
5005	break;
5006	}
5007
5008	rcu_read_unlock();
5009
5010	if (op)
5011	lru_gen_rotate_memcg(lruvec, op);
5012
5013	mem_cgroup_put(memcg);
5014
5015	if (!is_a_nulls(ptr: pos))
5016	return;
5017
5018	/* restart if raced with lru_gen_rotate_memcg() */
5019	if (gen != get_nulls_value(ptr: pos))
5020	goto restart;
5021
5022	/* try the rest of the bins of the current generation */
5023	bin = get_memcg_bin(bin + 1);
5024	if (bin != first_bin)
5025	goto restart;
5026	}
5027
5028	static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5029	{
5030	struct blk_plug plug;
5031
5032	VM_WARN_ON_ONCE(root_reclaim(sc));
5033	VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
5034
5035	lru_add_drain();
5036
5037	blk_start_plug(&plug);
5038
5039	set_mm_walk(NULL, force_alloc: sc->proactive);
5040
5041	if (try_to_shrink_lruvec(lruvec, sc))
5042	lru_gen_rotate_memcg(lruvec, op: MEMCG_LRU_YOUNG);
5043
5044	clear_mm_walk();
5045
5046	blk_finish_plug(&plug);
5047	}
5048
5049	static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5050	{
5051	struct blk_plug plug;
5052	unsigned long reclaimed = sc->nr_reclaimed;
5053
5054	VM_WARN_ON_ONCE(!root_reclaim(sc));
5055
5056	/*
5057	* Unmapped clean folios are already prioritized. Scanning for more of
5058	* them is likely futile and can cause high reclaim latency when there
5059	* is a large number of memcgs.
5060	*/
5061	if (!sc->may_writepage || !sc->may_unmap)
5062	goto done;
5063
5064	lru_add_drain();
5065
5066	blk_start_plug(&plug);
5067
5068	set_mm_walk(pgdat, force_alloc: sc->proactive);
5069
5070	set_initial_priority(pgdat, sc);
5071
5072	if (current_is_kswapd())
5073	sc->nr_reclaimed = 0;
5074
5075	if (mem_cgroup_disabled())
5076	shrink_one(lruvec: &pgdat->__lruvec, sc);
5077	else
5078	shrink_many(pgdat, sc);
5079
5080	if (current_is_kswapd())
5081	sc->nr_reclaimed += reclaimed;
5082
5083	clear_mm_walk();
5084
5085	blk_finish_plug(&plug);
5086	done:
5087	if (sc->nr_reclaimed > reclaimed)
5088	pgdat->kswapd_failures = 0;
5089	}
5090
5091	/******************************************************************************
5092	* state change
5093	******************************************************************************/
5094
5095	static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5096	{
5097	struct lru_gen_folio *lrugen = &lruvec->lrugen;
5098
5099	if (lrugen->enabled) {
5100	enum lru_list lru;
5101
5102	for_each_evictable_lru(lru) {
5103	if (!list_empty(head: &lruvec->lists[lru]))
5104	return false;
5105	}
5106	} else {
5107	int gen, type, zone;
5108
5109	for_each_gen_type_zone(gen, type, zone) {
5110	if (!list_empty(&lrugen->folios[gen][type][zone]))
5111	return false;
5112	}
5113	}
5114
5115	return true;
5116	}
5117
5118	static bool fill_evictable(struct lruvec *lruvec)
5119	{
5120	enum lru_list lru;
5121	int remaining = MAX_LRU_BATCH;
5122
5123	for_each_evictable_lru(lru) {
5124	int type = is_file_lru(lru);
5125	bool active = is_active_lru(lru);
5126	struct list_head *head = &lruvec->lists[lru];
5127
5128	while (!list_empty(head)) {
5129	bool success;
5130	struct folio *folio = lru_to_folio(head);
5131
5132	VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5133	VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5134	VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5135	VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5136
5137	lruvec_del_folio(lruvec, folio);
5138	success = lru_gen_add_folio(lruvec, folio, reclaiming: false);
5139	VM_WARN_ON_ONCE(!success);
5140
5141	if (!--remaining)
5142	return false;
5143	}
5144	}
5145
5146	return true;
5147	}
5148
5149	static bool drain_evictable(struct lruvec *lruvec)
5150	{
5151	int gen, type, zone;
5152	int remaining = MAX_LRU_BATCH;
5153
5154	for_each_gen_type_zone(gen, type, zone) {
5155	struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5156
5157	while (!list_empty(head)) {
5158	bool success;
5159	struct folio *folio = lru_to_folio(head);
5160
5161	VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5162	VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5163	VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5164	VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5165
5166	success = lru_gen_del_folio(lruvec, folio, false);
5167	VM_WARN_ON_ONCE(!success);
5168	lruvec_add_folio(lruvec, folio);
5169
5170	if (!--remaining)
5171	return false;
5172	}
5173	}
5174
5175	return true;
5176	}
5177
5178	static void lru_gen_change_state(bool enabled)
5179	{
5180	static DEFINE_MUTEX(state_mutex);
5181
5182	struct mem_cgroup *memcg;
5183
5184	cgroup_lock();
5185	cpus_read_lock();
5186	get_online_mems();
5187	mutex_lock(&state_mutex);
5188
5189	if (enabled == lru_gen_enabled())
5190	goto unlock;
5191
5192	if (enabled)
5193	static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5194	else
5195	static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5196
5197	memcg = mem_cgroup_iter(NULL, NULL, NULL);
5198	do {
5199	int nid;
5200
5201	for_each_node(nid) {
5202	struct lruvec *lruvec = get_lruvec(memcg, nid);
5203
5204	spin_lock_irq(lock: &lruvec->lru_lock);
5205
5206	VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5207	VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5208
5209	lruvec->lrugen.enabled = enabled;
5210
5211	while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5212	spin_unlock_irq(lock: &lruvec->lru_lock);
5213	cond_resched();
5214	spin_lock_irq(lock: &lruvec->lru_lock);
5215	}
5216
5217	spin_unlock_irq(lock: &lruvec->lru_lock);
5218	}
5219
5220	cond_resched();
5221	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5222	unlock:
5223	mutex_unlock(lock: &state_mutex);
5224	put_online_mems();
5225	cpus_read_unlock();
5226	cgroup_unlock();
5227	}
5228
5229	/******************************************************************************
5230	* sysfs interface
5231	******************************************************************************/
5232
5233	static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5234	{
5235	return sysfs_emit(buf, fmt: "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5236	}
5237
5238	/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5239	static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5240	const char *buf, size_t len)
5241	{
5242	unsigned int msecs;
5243
5244	if (kstrtouint(s: buf, base: 0, res: &msecs))
5245	return -EINVAL;
5246
5247	WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5248
5249	return len;
5250	}
5251
5252	static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5253
5254	static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5255	{
5256	unsigned int caps = 0;
5257
5258	if (get_cap(LRU_GEN_CORE))
5259	caps |= BIT(LRU_GEN_CORE);
5260
5261	if (should_walk_mmu())
5262	caps |= BIT(LRU_GEN_MM_WALK);
5263
5264	if (should_clear_pmd_young())
5265	caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5266
5267	return sysfs_emit(buf, fmt: "0x%04x\n", caps);
5268	}
5269
5270	/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5271	static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5272	const char *buf, size_t len)
5273	{
5274	int i;
5275	unsigned int caps;
5276
5277	if (tolower(*buf) == 'n')
5278	caps = 0;
5279	else if (tolower(*buf) == 'y')
5280	caps = -1;
5281	else if (kstrtouint(s: buf, base: 0, res: &caps))
5282	return -EINVAL;
5283
5284	for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5285	bool enabled = caps & BIT(i);
5286
5287	if (i == LRU_GEN_CORE)
5288	lru_gen_change_state(enabled);
5289	else if (enabled)
5290	static_branch_enable(&lru_gen_caps[i]);
5291	else
5292	static_branch_disable(&lru_gen_caps[i]);
5293	}
5294
5295	return len;
5296	}
5297
5298	static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5299
5300	static struct attribute *lru_gen_attrs[] = {
5301	&lru_gen_min_ttl_attr.attr,
5302	&lru_gen_enabled_attr.attr,
5303	NULL
5304	};
5305
5306	static const struct attribute_group lru_gen_attr_group = {
5307	.name = "lru_gen",
5308	.attrs = lru_gen_attrs,
5309	};
5310
5311	/******************************************************************************
5312	* debugfs interface
5313	******************************************************************************/
5314
5315	static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5316	{
5317	struct mem_cgroup *memcg;
5318	loff_t nr_to_skip = *pos;
5319
5320	m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5321	if (!m->private)
5322	return ERR_PTR(error: -ENOMEM);
5323
5324	memcg = mem_cgroup_iter(NULL, NULL, NULL);
5325	do {
5326	int nid;
5327
5328	for_each_node_state(nid, N_MEMORY) {
5329	if (!nr_to_skip--)
5330	return get_lruvec(memcg, nid);
5331	}
5332	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5333
5334	return NULL;
5335	}
5336
5337	static void lru_gen_seq_stop(struct seq_file *m, void *v)
5338	{
5339	if (!IS_ERR_OR_NULL(ptr: v))
5340	mem_cgroup_iter_break(NULL, lruvec_memcg(lruvec: v));
5341
5342	kvfree(addr: m->private);
5343	m->private = NULL;
5344	}
5345
5346	static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5347	{
5348	int nid = lruvec_pgdat(lruvec: v)->node_id;
5349	struct mem_cgroup *memcg = lruvec_memcg(lruvec: v);
5350
5351	++*pos;
5352
5353	nid = next_memory_node(nid);
5354	if (nid == MAX_NUMNODES) {
5355	memcg = mem_cgroup_iter(NULL, memcg, NULL);
5356	if (!memcg)
5357	return NULL;
5358
5359	nid = first_memory_node;
5360	}
5361
5362	return get_lruvec(memcg, nid);
5363	}
5364
5365	static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5366	unsigned long max_seq, unsigned long *min_seq,
5367	unsigned long seq)
5368	{
5369	int i;
5370	int type, tier;
5371	int hist = lru_hist_from_seq(seq);
5372	struct lru_gen_folio *lrugen = &lruvec->lrugen;
5373	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5374
5375	for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5376	seq_printf(m, fmt: " %10d", tier);
5377	for (type = 0; type < ANON_AND_FILE; type++) {
5378	const char *s = "xxx";
5379	unsigned long n[3] = {};
5380
5381	if (seq == max_seq) {
5382	s = "RTx";
5383	n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5384	n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5385	} else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5386	s = "rep";
5387	n[0] = atomic_long_read(v: &lrugen->refaulted[hist][type][tier]);
5388	n[1] = atomic_long_read(v: &lrugen->evicted[hist][type][tier]);
5389	n[2] = READ_ONCE(lrugen->protected[hist][type][tier]);
5390	}
5391
5392	for (i = 0; i < 3; i++)
5393	seq_printf(m, fmt: " %10lu%c", n[i], s[i]);
5394	}
5395	seq_putc(m, c: '\n');
5396	}
5397
5398	if (!mm_state)
5399	return;
5400
5401	seq_puts(m, s: " ");
5402	for (i = 0; i < NR_MM_STATS; i++) {
5403	const char *s = "xxxx";
5404	unsigned long n = 0;
5405
5406	if (seq == max_seq && NR_HIST_GENS == 1) {
5407	s = "TYFA";
5408	n = READ_ONCE(mm_state->stats[hist][i]);
5409	} else if (seq != max_seq && NR_HIST_GENS > 1) {
5410	s = "tyfa";
5411	n = READ_ONCE(mm_state->stats[hist][i]);
5412	}
5413
5414	seq_printf(m, fmt: " %10lu%c", n, s[i]);
5415	}
5416	seq_putc(m, c: '\n');
5417	}
5418
5419	/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5420	static int lru_gen_seq_show(struct seq_file *m, void *v)
5421	{
5422	unsigned long seq;
5423	bool full = !debugfs_real_fops(filp: m->file)->write;
5424	struct lruvec *lruvec = v;
5425	struct lru_gen_folio *lrugen = &lruvec->lrugen;
5426	int nid = lruvec_pgdat(lruvec)->node_id;
5427	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5428	DEFINE_MAX_SEQ(lruvec);
5429	DEFINE_MIN_SEQ(lruvec);
5430
5431	if (nid == first_memory_node) {
5432	const char *path = memcg ? m->private : "";
5433
5434	#ifdef CONFIG_MEMCG
5435	if (memcg)
5436	cgroup_path(cgrp: memcg->css.cgroup, buf: m->private, PATH_MAX);
5437	#endif
5438	seq_printf(m, fmt: "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5439	}
5440
5441	seq_printf(m, fmt: " node %5d\n", nid);
5442
5443	if (!full)
5444	seq = evictable_min_seq(min_seq, MAX_SWAPPINESS / 2);
5445	else if (max_seq >= MAX_NR_GENS)
5446	seq = max_seq - MAX_NR_GENS + 1;
5447	else
5448	seq = 0;
5449
5450	for (; seq <= max_seq; seq++) {
5451	int type, zone;
5452	int gen = lru_gen_from_seq(seq);
5453	unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5454
5455	seq_printf(m, fmt: " %10lu %10u", seq, jiffies_to_msecs(j: jiffies - birth));
5456
5457	for (type = 0; type < ANON_AND_FILE; type++) {
5458	unsigned long size = 0;
5459	char mark = full && seq < min_seq[type] ? 'x' : ' ';
5460
5461	for (zone = 0; zone < MAX_NR_ZONES; zone++)
5462	size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5463
5464	seq_printf(m, fmt: " %10lu%c", size, mark);
5465	}
5466
5467	seq_putc(m, c: '\n');
5468
5469	if (full)
5470	lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5471	}
5472
5473	return 0;
5474	}
5475
5476	static const struct seq_operations lru_gen_seq_ops = {
5477	.start = lru_gen_seq_start,
5478	.stop = lru_gen_seq_stop,
5479	.next = lru_gen_seq_next,
5480	.show = lru_gen_seq_show,
5481	};
5482
5483	static int run_aging(struct lruvec *lruvec, unsigned long seq,
5484	int swappiness, bool force_scan)
5485	{
5486	DEFINE_MAX_SEQ(lruvec);
5487
5488	if (seq > max_seq)
5489	return -EINVAL;
5490
5491	return try_to_inc_max_seq(lruvec, seq: max_seq, swappiness, force_scan) ? 0 : -EEXIST;
5492	}
5493
5494	static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5495	int swappiness, unsigned long nr_to_reclaim)
5496	{
5497	DEFINE_MAX_SEQ(lruvec);
5498
5499	if (seq + MIN_NR_GENS > max_seq)
5500	return -EINVAL;
5501
5502	sc->nr_reclaimed = 0;
5503
5504	while (!signal_pending(current)) {
5505	DEFINE_MIN_SEQ(lruvec);
5506
5507	if (seq < evictable_min_seq(min_seq, swappiness))
5508	return 0;
5509
5510	if (sc->nr_reclaimed >= nr_to_reclaim)
5511	return 0;
5512
5513	if (!evict_folios(lruvec, sc, swappiness))
5514	return 0;
5515
5516	cond_resched();
5517	}
5518
5519	return -EINTR;
5520	}
5521
5522	static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5523	struct scan_control *sc, int swappiness, unsigned long opt)
5524	{
5525	struct lruvec *lruvec;
5526	int err = -EINVAL;
5527	struct mem_cgroup *memcg = NULL;
5528
5529	if (nid < 0 || nid >= MAX_NUMNODES || !node_state(node: nid, state: N_MEMORY))
5530	return -EINVAL;
5531
5532	if (!mem_cgroup_disabled()) {
5533	rcu_read_lock();
5534
5535	memcg = mem_cgroup_from_id(id: memcg_id);
5536	if (!mem_cgroup_tryget(memcg))
5537	memcg = NULL;
5538
5539	rcu_read_unlock();
5540
5541	if (!memcg)
5542	return -EINVAL;
5543	}
5544
5545	if (memcg_id != mem_cgroup_id(memcg))
5546	goto done;
5547
5548	lruvec = get_lruvec(memcg, nid);
5549
5550	if (swappiness < MIN_SWAPPINESS)
5551	swappiness = get_swappiness(lruvec, sc);
5552	else if (swappiness > SWAPPINESS_ANON_ONLY)
5553	goto done;
5554
5555	switch (cmd) {
5556	case '+':
5557	err = run_aging(lruvec, seq, swappiness, force_scan: opt);
5558	break;
5559	case '-':
5560	err = run_eviction(lruvec, seq, sc, swappiness, nr_to_reclaim: opt);
5561	break;
5562	}
5563	done:
5564	mem_cgroup_put(memcg);
5565
5566	return err;
5567	}
5568
5569	/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5570	static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5571	size_t len, loff_t *pos)
5572	{
5573	void *buf;
5574	char *cur, *next;
5575	unsigned int flags;
5576	struct blk_plug plug;
5577	int err = -EINVAL;
5578	struct scan_control sc = {
5579	.may_writepage = true,
5580	.may_unmap = true,
5581	.may_swap = true,
5582	.reclaim_idx = MAX_NR_ZONES - 1,
5583	.gfp_mask = GFP_KERNEL,
5584	};
5585
5586	buf = kvmalloc(len + 1, GFP_KERNEL);
5587	if (!buf)
5588	return -ENOMEM;
5589
5590	if (copy_from_user(to: buf, from: src, n: len)) {
5591	kvfree(addr: buf);
5592	return -EFAULT;
5593	}
5594
5595	set_task_reclaim_state(current, rs: &sc.reclaim_state);
5596	flags = memalloc_noreclaim_save();
5597	blk_start_plug(&plug);
5598	if (!set_mm_walk(NULL, force_alloc: true)) {
5599	err = -ENOMEM;
5600	goto done;
5601	}
5602
5603	next = buf;
5604	next[len] = '\0';
5605
5606	while ((cur = strsep(&next, ",;\n"))) {
5607	int n;
5608	int end;
5609	char cmd, swap_string[5];
5610	unsigned int memcg_id;
5611	unsigned int nid;
5612	unsigned long seq;
5613	unsigned int swappiness;
5614	unsigned long opt = -1;
5615
5616	cur = skip_spaces(cur);
5617	if (!*cur)
5618	continue;
5619
5620	n = sscanf(cur, "%c %u %u %lu %n %4s %n %lu %n", &cmd, &memcg_id, &nid,
5621	&seq, &end, swap_string, &end, &opt, &end);
5622	if (n < 4 || cur[end]) {
5623	err = -EINVAL;
5624	break;
5625	}
5626
5627	if (n == 4) {
5628	swappiness = -1;
5629	} else if (!strcmp("max", swap_string)) {
5630	/* set by userspace for anonymous memory only */
5631	swappiness = SWAPPINESS_ANON_ONLY;
5632	} else {
5633	err = kstrtouint(s: swap_string, base: 0, res: &swappiness);
5634	if (err)
5635	break;
5636	}
5637
5638	err = run_cmd(cmd, memcg_id, nid, seq, sc: &sc, swappiness, opt);
5639	if (err)
5640	break;
5641	}
5642	done:
5643	clear_mm_walk();
5644	blk_finish_plug(&plug);
5645	memalloc_noreclaim_restore(flags);
5646	set_task_reclaim_state(current, NULL);
5647
5648	kvfree(addr: buf);
5649
5650	return err ? : len;
5651	}
5652
5653	static int lru_gen_seq_open(struct inode *inode, struct file *file)
5654	{
5655	return seq_open(file, &lru_gen_seq_ops);
5656	}
5657
5658	static const struct file_operations lru_gen_rw_fops = {
5659	.open = lru_gen_seq_open,
5660	.read = seq_read,
5661	.write = lru_gen_seq_write,
5662	.llseek = seq_lseek,
5663	.release = seq_release,
5664	};
5665
5666	static const struct file_operations lru_gen_ro_fops = {
5667	.open = lru_gen_seq_open,
5668	.read = seq_read,
5669	.llseek = seq_lseek,
5670	.release = seq_release,
5671	};
5672
5673	/******************************************************************************
5674	* initialization
5675	******************************************************************************/
5676
5677	void lru_gen_init_pgdat(struct pglist_data *pgdat)
5678	{
5679	int i, j;
5680
5681	spin_lock_init(&pgdat->memcg_lru.lock);
5682
5683	for (i = 0; i < MEMCG_NR_GENS; i++) {
5684	for (j = 0; j < MEMCG_NR_BINS; j++)
5685	INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5686	}
5687	}
5688
5689	void lru_gen_init_lruvec(struct lruvec *lruvec)
5690	{
5691	int i;
5692	int gen, type, zone;
5693	struct lru_gen_folio *lrugen = &lruvec->lrugen;
5694	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5695
5696	lrugen->max_seq = MIN_NR_GENS + 1;
5697	lrugen->enabled = lru_gen_enabled();
5698
5699	for (i = 0; i <= MIN_NR_GENS + 1; i++)
5700	lrugen->timestamps[i] = jiffies;
5701
5702	for_each_gen_type_zone(gen, type, zone)
5703	INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5704
5705	if (mm_state)
5706	mm_state->seq = MIN_NR_GENS;
5707	}
5708
5709	#ifdef CONFIG_MEMCG
5710
5711	void lru_gen_init_memcg(struct mem_cgroup *memcg)
5712	{
5713	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5714
5715	if (!mm_list)
5716	return;
5717
5718	INIT_LIST_HEAD(list: &mm_list->fifo);
5719	spin_lock_init(&mm_list->lock);
5720	}
5721
5722	void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5723	{
5724	int i;
5725	int nid;
5726	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5727
5728	VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5729
5730	for_each_node(nid) {
5731	struct lruvec *lruvec = get_lruvec(memcg, nid);
5732	struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5733
5734	VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5735	sizeof(lruvec->lrugen.nr_pages)));
5736
5737	lruvec->lrugen.list.next = LIST_POISON1;
5738
5739	if (!mm_state)
5740	continue;
5741
5742	for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5743	bitmap_free(bitmap: mm_state->filters[i]);
5744	mm_state->filters[i] = NULL;
5745	}
5746	}
5747	}
5748
5749	#endif /* CONFIG_MEMCG */
5750
5751	static int __init init_lru_gen(void)
5752	{
5753	BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5754	BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5755
5756	if (sysfs_create_group(kobj: mm_kobj, grp: &lru_gen_attr_group))
5757	pr_err("lru_gen: failed to create sysfs group\n");
5758
5759	debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5760	debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5761
5762	return 0;
5763	};
5764	late_initcall(init_lru_gen);
5765
5766	#else /* !CONFIG_LRU_GEN */
5767
5768	static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5769	{
5770	BUILD_BUG();
5771	}
5772
5773	static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5774	{
5775	BUILD_BUG();
5776	}
5777
5778	static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5779	{
5780	BUILD_BUG();
5781	}
5782
5783	#endif /* CONFIG_LRU_GEN */
5784
5785	static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5786	{
5787	unsigned long nr[NR_LRU_LISTS];
5788	unsigned long targets[NR_LRU_LISTS];
5789	unsigned long nr_to_scan;
5790	enum lru_list lru;
5791	unsigned long nr_reclaimed = 0;
5792	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5793	bool proportional_reclaim;
5794	struct blk_plug plug;
5795
5796	if (lru_gen_enabled() && !root_reclaim(sc)) {
5797	lru_gen_shrink_lruvec(lruvec, sc);
5798	return;
5799	}
5800
5801	get_scan_count(lruvec, sc, nr);
5802
5803	/* Record the original scan target for proportional adjustments later */
5804	memcpy(targets, nr, sizeof(nr));
5805
5806	/*
5807	* Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5808	* event that can occur when there is little memory pressure e.g.
5809	* multiple streaming readers/writers. Hence, we do not abort scanning
5810	* when the requested number of pages are reclaimed when scanning at
5811	* DEF_PRIORITY on the assumption that the fact we are direct
5812	* reclaiming implies that kswapd is not keeping up and it is best to
5813	* do a batch of work at once. For memcg reclaim one check is made to
5814	* abort proportional reclaim if either the file or anon lru has already
5815	* dropped to zero at the first pass.
5816	*/
5817	proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5818	sc->priority == DEF_PRIORITY);
5819
5820	blk_start_plug(&plug);
5821	while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5822	nr[LRU_INACTIVE_FILE]) {
5823	unsigned long nr_anon, nr_file, percentage;
5824	unsigned long nr_scanned;
5825
5826	for_each_evictable_lru(lru) {
5827	if (nr[lru]) {
5828	nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5829	nr[lru] -= nr_to_scan;
5830
5831	nr_reclaimed += shrink_list(lru, nr_to_scan,
5832	lruvec, sc);
5833	}
5834	}
5835
5836	cond_resched();
5837
5838	if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5839	continue;
5840
5841	/*
5842	* For kswapd and memcg, reclaim at least the number of pages
5843	* requested. Ensure that the anon and file LRUs are scanned
5844	* proportionally what was requested by get_scan_count(). We
5845	* stop reclaiming one LRU and reduce the amount scanning
5846	* proportional to the original scan target.
5847	*/
5848	nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5849	nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5850
5851	/*
5852	* It's just vindictive to attack the larger once the smaller
5853	* has gone to zero. And given the way we stop scanning the
5854	* smaller below, this makes sure that we only make one nudge
5855	* towards proportionality once we've got nr_to_reclaim.
5856	*/
5857	if (!nr_file || !nr_anon)
5858	break;
5859
5860	if (nr_file > nr_anon) {
5861	unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5862	targets[LRU_ACTIVE_ANON] + 1;
5863	lru = LRU_BASE;
5864	percentage = nr_anon * 100 / scan_target;
5865	} else {
5866	unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5867	targets[LRU_ACTIVE_FILE] + 1;
5868	lru = LRU_FILE;
5869	percentage = nr_file * 100 / scan_target;
5870	}
5871
5872	/* Stop scanning the smaller of the LRU */
5873	nr[lru] = 0;
5874	nr[lru + LRU_ACTIVE] = 0;
5875
5876	/*
5877	* Recalculate the other LRU scan count based on its original
5878	* scan target and the percentage scanning already complete
5879	*/
5880	lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5881	nr_scanned = targets[lru] - nr[lru];
5882	nr[lru] = targets[lru] * (100 - percentage) / 100;
5883	nr[lru] -= min(nr[lru], nr_scanned);
5884
5885	lru += LRU_ACTIVE;
5886	nr_scanned = targets[lru] - nr[lru];
5887	nr[lru] = targets[lru] * (100 - percentage) / 100;
5888	nr[lru] -= min(nr[lru], nr_scanned);
5889	}
5890	blk_finish_plug(&plug);
5891	sc->nr_reclaimed += nr_reclaimed;
5892
5893	/*
5894	* Even if we did not try to evict anon pages at all, we want to
5895	* rebalance the anon lru active/inactive ratio.
5896	*/
5897	if (can_age_anon_pages(lruvec, sc) &&
5898	inactive_is_low(lruvec, inactive_lru: LRU_INACTIVE_ANON))
5899	shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5900	sc, lru: LRU_ACTIVE_ANON);
5901	}
5902
5903	/* Use reclaim/compaction for costly allocs or under memory pressure */
5904	static bool in_reclaim_compaction(struct scan_control *sc)
5905	{
5906	if (gfp_compaction_allowed(gfp_mask: sc->gfp_mask) && sc->order &&
5907	(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5908	sc->priority < DEF_PRIORITY - 2))
5909	return true;
5910
5911	return false;
5912	}
5913
5914	/*
5915	* Reclaim/compaction is used for high-order allocation requests. It reclaims
5916	* order-0 pages before compacting the zone. should_continue_reclaim() returns
5917	* true if more pages should be reclaimed such that when the page allocator
5918	* calls try_to_compact_pages() that it will have enough free pages to succeed.
5919	* It will give up earlier than that if there is difficulty reclaiming pages.
5920	*/
5921	static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5922	unsigned long nr_reclaimed,
5923	struct scan_control *sc)
5924	{
5925	unsigned long pages_for_compaction;
5926	unsigned long inactive_lru_pages;
5927	int z;
5928	struct zone *zone;
5929
5930	/* If not in reclaim/compaction mode, stop */
5931	if (!in_reclaim_compaction(sc))
5932	return false;
5933
5934	/*
5935	* Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5936	* number of pages that were scanned. This will return to the caller
5937	* with the risk reclaim/compaction and the resulting allocation attempt
5938	* fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5939	* allocations through requiring that the full LRU list has been scanned
5940	* first, by assuming that zero delta of sc->nr_scanned means full LRU
5941	* scan, but that approximation was wrong, and there were corner cases
5942	* where always a non-zero amount of pages were scanned.
5943	*/
5944	if (!nr_reclaimed)
5945	return false;
5946
5947	/* If compaction would go ahead or the allocation would succeed, stop */
5948	for_each_managed_zone_pgdat(zone, pgdat, z, sc->reclaim_idx) {
5949	unsigned long watermark = min_wmark_pages(z: zone);
5950
5951	/* Allocation can already succeed, nothing to do */
5952	if (zone_watermark_ok(z: zone, order: sc->order, mark: watermark,
5953	highest_zoneidx: sc->reclaim_idx, alloc_flags: 0))
5954	return false;
5955
5956	if (compaction_suitable(zone, order: sc->order, watermark,
5957	highest_zoneidx: sc->reclaim_idx))
5958	return false;
5959	}
5960
5961	/*
5962	* If we have not reclaimed enough pages for compaction and the
5963	* inactive lists are large enough, continue reclaiming
5964	*/
5965	pages_for_compaction = compact_gap(order: sc->order);
5966	inactive_lru_pages = node_page_state(pgdat, item: NR_INACTIVE_FILE);
5967	if (can_reclaim_anon_pages(NULL, nid: pgdat->node_id, sc))
5968	inactive_lru_pages += node_page_state(pgdat, item: NR_INACTIVE_ANON);
5969
5970	return inactive_lru_pages > pages_for_compaction;
5971	}
5972
5973	static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5974	{
5975	struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5976	struct mem_cgroup_reclaim_cookie reclaim = {
5977	.pgdat = pgdat,
5978	};
5979	struct mem_cgroup_reclaim_cookie *partial = &reclaim;
5980	struct mem_cgroup *memcg;
5981
5982	/*
5983	* In most cases, direct reclaimers can do partial walks
5984	* through the cgroup tree, using an iterator state that
5985	* persists across invocations. This strikes a balance between
5986	* fairness and allocation latency.
5987	*
5988	* For kswapd, reliable forward progress is more important
5989	* than a quick return to idle. Always do full walks.
5990	*/
5991	if (current_is_kswapd() || sc->memcg_full_walk)
5992	partial = NULL;
5993
5994	memcg = mem_cgroup_iter(target_memcg, NULL, partial);
5995	do {
5996	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5997	unsigned long reclaimed;
5998	unsigned long scanned;
5999
6000	/*
6001	* This loop can become CPU-bound when target memcgs
6002	* aren't eligible for reclaim - either because they
6003	* don't have any reclaimable pages, or because their
6004	* memory is explicitly protected. Avoid soft lockups.
6005	*/
6006	cond_resched();
6007
6008	mem_cgroup_calculate_protection(root: target_memcg, memcg);
6009
6010	if (mem_cgroup_below_min(target: target_memcg, memcg)) {
6011	/*
6012	* Hard protection.
6013	* If there is no reclaimable memory, OOM.
6014	*/
6015	continue;
6016	} else if (mem_cgroup_below_low(target: target_memcg, memcg)) {
6017	/*
6018	* Soft protection.
6019	* Respect the protection only as long as
6020	* there is an unprotected supply
6021	* of reclaimable memory from other cgroups.
6022	*/
6023	if (!sc->memcg_low_reclaim) {
6024	sc->memcg_low_skipped = 1;
6025	continue;
6026	}
6027	memcg_memory_event(memcg, event: MEMCG_LOW);
6028	}
6029
6030	reclaimed = sc->nr_reclaimed;
6031	scanned = sc->nr_scanned;
6032
6033	shrink_lruvec(lruvec, sc);
6034
6035	shrink_slab(gfp_mask: sc->gfp_mask, nid: pgdat->node_id, memcg,
6036	priority: sc->priority);
6037
6038	/* Record the group's reclaim efficiency */
6039	if (!sc->proactive)
6040	vmpressure(gfp: sc->gfp_mask, memcg, tree: false,
6041	scanned: sc->nr_scanned - scanned,
6042	reclaimed: sc->nr_reclaimed - reclaimed);
6043
6044	/* If partial walks are allowed, bail once goal is reached */
6045	if (partial && sc->nr_reclaimed >= sc->nr_to_reclaim) {
6046	mem_cgroup_iter_break(target_memcg, memcg);
6047	break;
6048	}
6049	} while ((memcg = mem_cgroup_iter(target_memcg, memcg, partial)));
6050	}
6051
6052	static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6053	{
6054	unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
6055	struct lruvec *target_lruvec;
6056	bool reclaimable = false;
6057
6058	if (lru_gen_enabled() && root_reclaim(sc)) {
6059	memset(&sc->nr, 0, sizeof(sc->nr));
6060	lru_gen_shrink_node(pgdat, sc);
6061	return;
6062	}
6063
6064	target_lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup, pgdat);
6065
6066	again:
6067	memset(&sc->nr, 0, sizeof(sc->nr));
6068
6069	nr_reclaimed = sc->nr_reclaimed;
6070	nr_scanned = sc->nr_scanned;
6071
6072	prepare_scan_control(pgdat, sc);
6073
6074	shrink_node_memcgs(pgdat, sc);
6075
6076	flush_reclaim_state(sc);
6077
6078	nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
6079
6080	/* Record the subtree's reclaim efficiency */
6081	if (!sc->proactive)
6082	vmpressure(gfp: sc->gfp_mask, memcg: sc->target_mem_cgroup, tree: true,
6083	scanned: sc->nr_scanned - nr_scanned, reclaimed: nr_node_reclaimed);
6084
6085	if (nr_node_reclaimed)
6086	reclaimable = true;
6087
6088	if (current_is_kswapd()) {
6089	/*
6090	* If reclaim is isolating dirty pages under writeback,
6091	* it implies that the long-lived page allocation rate
6092	* is exceeding the page laundering rate. Either the
6093	* global limits are not being effective at throttling
6094	* processes due to the page distribution throughout
6095	* zones or there is heavy usage of a slow backing
6096	* device. The only option is to throttle from reclaim
6097	* context which is not ideal as there is no guarantee
6098	* the dirtying process is throttled in the same way
6099	* balance_dirty_pages() manages.
6100	*
6101	* Once a node is flagged PGDAT_WRITEBACK, kswapd will
6102	* count the number of pages under pages flagged for
6103	* immediate reclaim and stall if any are encountered
6104	* in the nr_immediate check below.
6105	*/
6106	if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6107	set_bit(nr: PGDAT_WRITEBACK, addr: &pgdat->flags);
6108
6109	/* Allow kswapd to start writing pages during reclaim.*/
6110	if (sc->nr.unqueued_dirty &&
6111	sc->nr.unqueued_dirty == sc->nr.file_taken)
6112	set_bit(nr: PGDAT_DIRTY, addr: &pgdat->flags);
6113
6114	/*
6115	* If kswapd scans pages marked for immediate
6116	* reclaim and under writeback (nr_immediate), it
6117	* implies that pages are cycling through the LRU
6118	* faster than they are written so forcibly stall
6119	* until some pages complete writeback.
6120	*/
6121	if (sc->nr.immediate)
6122	reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_WRITEBACK);
6123	}
6124
6125	/*
6126	* Tag a node/memcg as congested if all the dirty pages were marked
6127	* for writeback and immediate reclaim (counted in nr.congested).
6128	*
6129	* Legacy memcg will stall in page writeback so avoid forcibly
6130	* stalling in reclaim_throttle().
6131	*/
6132	if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
6133	if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
6134	set_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &target_lruvec->flags);
6135
6136	if (current_is_kswapd())
6137	set_bit(nr: LRUVEC_NODE_CONGESTED, addr: &target_lruvec->flags);
6138	}
6139
6140	/*
6141	* Stall direct reclaim for IO completions if the lruvec is
6142	* node is congested. Allow kswapd to continue until it
6143	* starts encountering unqueued dirty pages or cycling through
6144	* the LRU too quickly.
6145	*/
6146	if (!current_is_kswapd() && current_may_throttle() &&
6147	!sc->hibernation_mode &&
6148	(test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
6149	test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
6150	reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_CONGESTED);
6151
6152	if (should_continue_reclaim(pgdat, nr_reclaimed: nr_node_reclaimed, sc))
6153	goto again;
6154
6155	/*
6156	* Kswapd gives up on balancing particular nodes after too
6157	* many failures to reclaim anything from them and goes to
6158	* sleep. On reclaim progress, reset the failure counter. A
6159	* successful direct reclaim run will revive a dormant kswapd.
6160	*/
6161	if (reclaimable)
6162	pgdat->kswapd_failures = 0;
6163	else if (sc->cache_trim_mode)
6164	sc->cache_trim_mode_failed = 1;
6165	}
6166
6167	/*
6168	* Returns true if compaction should go ahead for a costly-order request, or
6169	* the allocation would already succeed without compaction. Return false if we
6170	* should reclaim first.
6171	*/
6172	static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6173	{
6174	unsigned long watermark;
6175
6176	if (!gfp_compaction_allowed(gfp_mask: sc->gfp_mask))
6177	return false;
6178
6179	/* Allocation can already succeed, nothing to do */
6180	if (zone_watermark_ok(z: zone, order: sc->order, mark: min_wmark_pages(z: zone),
6181	highest_zoneidx: sc->reclaim_idx, alloc_flags: 0))
6182	return true;
6183
6184	/*
6185	* Direct reclaim usually targets the min watermark, but compaction
6186	* takes time to run and there are potentially other callers using the
6187	* pages just freed. So target a higher buffer to give compaction a
6188	* reasonable chance of completing and allocating the pages.
6189	*
6190	* Note that we won't actually reclaim the whole buffer in one attempt
6191	* as the target watermark in should_continue_reclaim() is lower. But if
6192	* we are already above the high+gap watermark, don't reclaim at all.
6193	*/
6194	watermark = high_wmark_pages(z: zone);
6195	if (compaction_suitable(zone, order: sc->order, watermark, highest_zoneidx: sc->reclaim_idx))
6196	return true;
6197
6198	return false;
6199	}
6200
6201	static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6202	{
6203	/*
6204	* If reclaim is making progress greater than 12% efficiency then
6205	* wake all the NOPROGRESS throttled tasks.
6206	*/
6207	if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6208	wait_queue_head_t *wqh;
6209
6210	wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6211	if (waitqueue_active(wq_head: wqh))
6212	wake_up(wqh);
6213
6214	return;
6215	}
6216
6217	/*
6218	* Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6219	* throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6220	* under writeback and marked for immediate reclaim at the tail of the
6221	* LRU.
6222	*/
6223	if (current_is_kswapd() || cgroup_reclaim(sc))
6224	return;
6225
6226	/* Throttle if making no progress at high prioities. */
6227	if (sc->priority == 1 && !sc->nr_reclaimed)
6228	reclaim_throttle(pgdat, reason: VMSCAN_THROTTLE_NOPROGRESS);
6229	}
6230
6231	/*
6232	* This is the direct reclaim path, for page-allocating processes. We only
6233	* try to reclaim pages from zones which will satisfy the caller's allocation
6234	* request.
6235	*
6236	* If a zone is deemed to be full of pinned pages then just give it a light
6237	* scan then give up on it.
6238	*/
6239	static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6240	{
6241	struct zoneref *z;
6242	struct zone *zone;
6243	unsigned long nr_soft_reclaimed;
6244	unsigned long nr_soft_scanned;
6245	gfp_t orig_mask;
6246	pg_data_t *last_pgdat = NULL;
6247	pg_data_t *first_pgdat = NULL;
6248
6249	/*
6250	* If the number of buffer_heads in the machine exceeds the maximum
6251	* allowed level, force direct reclaim to scan the highmem zone as
6252	* highmem pages could be pinning lowmem pages storing buffer_heads
6253	*/
6254	orig_mask = sc->gfp_mask;
6255	if (buffer_heads_over_limit) {
6256	sc->gfp_mask |= __GFP_HIGHMEM;
6257	sc->reclaim_idx = gfp_zone(flags: sc->gfp_mask);
6258	}
6259
6260	for_each_zone_zonelist_nodemask(zone, z, zonelist,
6261	sc->reclaim_idx, sc->nodemask) {
6262	/*
6263	* Take care memory controller reclaiming has small influence
6264	* to global LRU.
6265	*/
6266	if (!cgroup_reclaim(sc)) {
6267	if (!cpuset_zone_allowed(z: zone,
6268	GFP_KERNEL | __GFP_HARDWALL))
6269	continue;
6270
6271	/*
6272	* If we already have plenty of memory free for
6273	* compaction in this zone, don't free any more.
6274	* Even though compaction is invoked for any
6275	* non-zero order, only frequent costly order
6276	* reclamation is disruptive enough to become a
6277	* noticeable problem, like transparent huge
6278	* page allocations.
6279	*/
6280	if (IS_ENABLED(CONFIG_COMPACTION) &&
6281	sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6282	compaction_ready(zone, sc)) {
6283	sc->compaction_ready = true;
6284	continue;
6285	}
6286
6287	/*
6288	* Shrink each node in the zonelist once. If the
6289	* zonelist is ordered by zone (not the default) then a
6290	* node may be shrunk multiple times but in that case
6291	* the user prefers lower zones being preserved.
6292	*/
6293	if (zone->zone_pgdat == last_pgdat)
6294	continue;
6295
6296	/*
6297	* This steals pages from memory cgroups over softlimit
6298	* and returns the number of reclaimed pages and
6299	* scanned pages. This works for global memory pressure
6300	* and balancing, not for a memcg's limit.
6301	*/
6302	nr_soft_scanned = 0;
6303	nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat: zone->zone_pgdat,
6304	order: sc->order, gfp_mask: sc->gfp_mask,
6305	total_scanned: &nr_soft_scanned);
6306	sc->nr_reclaimed += nr_soft_reclaimed;
6307	sc->nr_scanned += nr_soft_scanned;
6308	/* need some check for avoid more shrink_zone() */
6309	}
6310
6311	if (!first_pgdat)
6312	first_pgdat = zone->zone_pgdat;
6313
6314	/* See comment about same check for global reclaim above */
6315	if (zone->zone_pgdat == last_pgdat)
6316	continue;
6317	last_pgdat = zone->zone_pgdat;
6318	shrink_node(pgdat: zone->zone_pgdat, sc);
6319	}
6320
6321	if (first_pgdat)
6322	consider_reclaim_throttle(pgdat: first_pgdat, sc);
6323
6324	/*
6325	* Restore to original mask to avoid the impact on the caller if we
6326	* promoted it to __GFP_HIGHMEM.
6327	*/
6328	sc->gfp_mask = orig_mask;
6329	}
6330
6331	static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6332	{
6333	struct lruvec *target_lruvec;
6334	unsigned long refaults;
6335
6336	if (lru_gen_enabled())
6337	return;
6338
6339	target_lruvec = mem_cgroup_lruvec(memcg: target_memcg, pgdat);
6340	refaults = lruvec_page_state(lruvec: target_lruvec, idx: WORKINGSET_ACTIVATE_ANON);
6341	target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6342	refaults = lruvec_page_state(lruvec: target_lruvec, idx: WORKINGSET_ACTIVATE_FILE);
6343	target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6344	}
6345
6346	/*
6347	* This is the main entry point to direct page reclaim.
6348	*
6349	* If a full scan of the inactive list fails to free enough memory then we
6350	* are "out of memory" and something needs to be killed.
6351	*
6352	* If the caller is !__GFP_FS then the probability of a failure is reasonably
6353	* high - the zone may be full of dirty or under-writeback pages, which this
6354	* caller can't do much about. We kick the writeback threads and take explicit
6355	* naps in the hope that some of these pages can be written. But if the
6356	* allocating task holds filesystem locks which prevent writeout this might not
6357	* work, and the allocation attempt will fail.
6358	*
6359	* returns: 0, if no pages reclaimed
6360	* else, the number of pages reclaimed
6361	*/
6362	static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6363	struct scan_control *sc)
6364	{
6365	int initial_priority = sc->priority;
6366	pg_data_t *last_pgdat;
6367	struct zoneref *z;
6368	struct zone *zone;
6369	retry:
6370	delayacct_freepages_start();
6371
6372	if (!cgroup_reclaim(sc))
6373	__count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6374
6375	do {
6376	if (!sc->proactive)
6377	vmpressure_prio(gfp: sc->gfp_mask, memcg: sc->target_mem_cgroup,
6378	prio: sc->priority);
6379	sc->nr_scanned = 0;
6380	shrink_zones(zonelist, sc);
6381
6382	if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6383	break;
6384
6385	if (sc->compaction_ready)
6386	break;
6387
6388	/*
6389	* If we're getting trouble reclaiming, start doing
6390	* writepage even in laptop mode.
6391	*/
6392	if (sc->priority < DEF_PRIORITY - 2)
6393	sc->may_writepage = 1;
6394	} while (--sc->priority >= 0);
6395
6396	last_pgdat = NULL;
6397	for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6398	sc->nodemask) {
6399	if (zone->zone_pgdat == last_pgdat)
6400	continue;
6401	last_pgdat = zone->zone_pgdat;
6402
6403	snapshot_refaults(target_memcg: sc->target_mem_cgroup, pgdat: zone->zone_pgdat);
6404
6405	if (cgroup_reclaim(sc)) {
6406	struct lruvec *lruvec;
6407
6408	lruvec = mem_cgroup_lruvec(memcg: sc->target_mem_cgroup,
6409	pgdat: zone->zone_pgdat);
6410	clear_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &lruvec->flags);
6411	}
6412	}
6413
6414	delayacct_freepages_end();
6415
6416	if (sc->nr_reclaimed)
6417	return sc->nr_reclaimed;
6418
6419	/* Aborted reclaim to try compaction? don't OOM, then */
6420	if (sc->compaction_ready)
6421	return 1;
6422
6423	/*
6424	* In most cases, direct reclaimers can do partial walks
6425	* through the cgroup tree to meet the reclaim goal while
6426	* keeping latency low. Since the iterator state is shared
6427	* among all direct reclaim invocations (to retain fairness
6428	* among cgroups), though, high concurrency can result in
6429	* individual threads not seeing enough cgroups to make
6430	* meaningful forward progress. Avoid false OOMs in this case.
6431	*/
6432	if (!sc->memcg_full_walk) {
6433	sc->priority = initial_priority;
6434	sc->memcg_full_walk = 1;
6435	goto retry;
6436	}
6437
6438	/*
6439	* We make inactive:active ratio decisions based on the node's
6440	* composition of memory, but a restrictive reclaim_idx or a
6441	* memory.low cgroup setting can exempt large amounts of
6442	* memory from reclaim. Neither of which are very common, so
6443	* instead of doing costly eligibility calculations of the
6444	* entire cgroup subtree up front, we assume the estimates are
6445	* good, and retry with forcible deactivation if that fails.
6446	*/
6447	if (sc->skipped_deactivate) {
6448	sc->priority = initial_priority;
6449	sc->force_deactivate = 1;
6450	sc->skipped_deactivate = 0;
6451	goto retry;
6452	}
6453
6454	/* Untapped cgroup reserves? Don't OOM, retry. */
6455	if (sc->memcg_low_skipped) {
6456	sc->priority = initial_priority;
6457	sc->force_deactivate = 0;
6458	sc->memcg_low_reclaim = 1;
6459	sc->memcg_low_skipped = 0;
6460	goto retry;
6461	}
6462
6463	return 0;
6464	}
6465
6466	static bool allow_direct_reclaim(pg_data_t *pgdat)
6467	{
6468	struct zone *zone;
6469	unsigned long pfmemalloc_reserve = 0;
6470	unsigned long free_pages = 0;
6471	int i;
6472	bool wmark_ok;
6473
6474	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6475	return true;
6476
6477	for_each_managed_zone_pgdat(zone, pgdat, i, ZONE_NORMAL) {
6478	if (!zone_reclaimable_pages(zone))
6479	continue;
6480
6481	pfmemalloc_reserve += min_wmark_pages(z: zone);
6482	free_pages += zone_page_state_snapshot(zone, item: NR_FREE_PAGES);
6483	}
6484
6485	/* If there are no reserves (unexpected config) then do not throttle */
6486	if (!pfmemalloc_reserve)
6487	return true;
6488
6489	wmark_ok = free_pages > pfmemalloc_reserve / 2;
6490
6491	/* kswapd must be awake if processes are being throttled */
6492	if (!wmark_ok && waitqueue_active(wq_head: &pgdat->kswapd_wait)) {
6493	if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6494	WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6495
6496	wake_up_interruptible(&pgdat->kswapd_wait);
6497	}
6498
6499	return wmark_ok;
6500	}
6501
6502	/*
6503	* Throttle direct reclaimers if backing storage is backed by the network
6504	* and the PFMEMALLOC reserve for the preferred node is getting dangerously
6505	* depleted. kswapd will continue to make progress and wake the processes
6506	* when the low watermark is reached.
6507	*
6508	* Returns true if a fatal signal was delivered during throttling. If this
6509	* happens, the page allocator should not consider triggering the OOM killer.
6510	*/
6511	static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6512	nodemask_t *nodemask)
6513	{
6514	struct zoneref *z;
6515	struct zone *zone;
6516	pg_data_t *pgdat = NULL;
6517
6518	/*
6519	* Kernel threads should not be throttled as they may be indirectly
6520	* responsible for cleaning pages necessary for reclaim to make forward
6521	* progress. kjournald for example may enter direct reclaim while
6522	* committing a transaction where throttling it could forcing other
6523	* processes to block on log_wait_commit().
6524	*/
6525	if (current->flags & PF_KTHREAD)
6526	goto out;
6527
6528	/*
6529	* If a fatal signal is pending, this process should not throttle.
6530	* It should return quickly so it can exit and free its memory
6531	*/
6532	if (fatal_signal_pending(current))
6533	goto out;
6534
6535	/*
6536	* Check if the pfmemalloc reserves are ok by finding the first node
6537	* with a usable ZONE_NORMAL or lower zone. The expectation is that
6538	* GFP_KERNEL will be required for allocating network buffers when
6539	* swapping over the network so ZONE_HIGHMEM is unusable.
6540	*
6541	* Throttling is based on the first usable node and throttled processes
6542	* wait on a queue until kswapd makes progress and wakes them. There
6543	* is an affinity then between processes waking up and where reclaim
6544	* progress has been made assuming the process wakes on the same node.
6545	* More importantly, processes running on remote nodes will not compete
6546	* for remote pfmemalloc reserves and processes on different nodes
6547	* should make reasonable progress.
6548	*/
6549	for_each_zone_zonelist_nodemask(zone, z, zonelist,
6550	gfp_zone(gfp_mask), nodemask) {
6551	if (zone_idx(zone) > ZONE_NORMAL)
6552	continue;
6553
6554	/* Throttle based on the first usable node */
6555	pgdat = zone->zone_pgdat;
6556	if (allow_direct_reclaim(pgdat))
6557	goto out;
6558	break;
6559	}
6560
6561	/* If no zone was usable by the allocation flags then do not throttle */
6562	if (!pgdat)
6563	goto out;
6564
6565	/* Account for the throttling */
6566	count_vm_event(item: PGSCAN_DIRECT_THROTTLE);
6567
6568	/*
6569	* If the caller cannot enter the filesystem, it's possible that it
6570	* is due to the caller holding an FS lock or performing a journal
6571	* transaction in the case of a filesystem like ext[3|4]. In this case,
6572	* it is not safe to block on pfmemalloc_wait as kswapd could be
6573	* blocked waiting on the same lock. Instead, throttle for up to a
6574	* second before continuing.
6575	*/
6576	if (!(gfp_mask & __GFP_FS))
6577	wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6578	allow_direct_reclaim(pgdat), HZ);
6579	else
6580	/* Throttle until kswapd wakes the process */
6581	wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6582	allow_direct_reclaim(pgdat));
6583
6584	if (fatal_signal_pending(current))
6585	return true;
6586
6587	out:
6588	return false;
6589	}
6590
6591	unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6592	gfp_t gfp_mask, nodemask_t *nodemask)
6593	{
6594	unsigned long nr_reclaimed;
6595	struct scan_control sc = {
6596	.nr_to_reclaim = SWAP_CLUSTER_MAX,
6597	.gfp_mask = current_gfp_context(flags: gfp_mask),
6598	.reclaim_idx = gfp_zone(flags: gfp_mask),
6599	.order = order,
6600	.nodemask = nodemask,
6601	.priority = DEF_PRIORITY,
6602	.may_writepage = !laptop_mode,
6603	.may_unmap = 1,
6604	.may_swap = 1,
6605	};
6606
6607	/*
6608	* scan_control uses s8 fields for order, priority, and reclaim_idx.
6609	* Confirm they are large enough for max values.
6610	*/
6611	BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6612	BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6613	BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6614
6615	/*
6616	* Do not enter reclaim if fatal signal was delivered while throttled.
6617	* 1 is returned so that the page allocator does not OOM kill at this
6618	* point.
6619	*/
6620	if (throttle_direct_reclaim(gfp_mask: sc.gfp_mask, zonelist, nodemask))
6621	return 1;
6622
6623	set_task_reclaim_state(current, rs: &sc.reclaim_state);
6624	trace_mm_vmscan_direct_reclaim_begin(order, gfp_flags: sc.gfp_mask);
6625
6626	nr_reclaimed = do_try_to_free_pages(zonelist, sc: &sc);
6627
6628	trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6629	set_task_reclaim_state(current, NULL);
6630
6631	return nr_reclaimed;
6632	}
6633
6634	#ifdef CONFIG_MEMCG
6635
6636	/* Only used by soft limit reclaim. Do not reuse for anything else. */
6637	unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6638	gfp_t gfp_mask, bool noswap,
6639	pg_data_t *pgdat,
6640	unsigned long *nr_scanned)
6641	{
6642	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6643	struct scan_control sc = {
6644	.nr_to_reclaim = SWAP_CLUSTER_MAX,
6645	.target_mem_cgroup = memcg,
6646	.may_writepage = !laptop_mode,
6647	.may_unmap = 1,
6648	.reclaim_idx = MAX_NR_ZONES - 1,
6649	.may_swap = !noswap,
6650	};
6651
6652	WARN_ON_ONCE(!current->reclaim_state);
6653
6654	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6655	(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6656
6657	trace_mm_vmscan_memcg_softlimit_reclaim_begin(order: sc.order,
6658	gfp_flags: sc.gfp_mask);
6659
6660	/*
6661	* NOTE: Although we can get the priority field, using it
6662	* here is not a good idea, since it limits the pages we can scan.
6663	* if we don't reclaim here, the shrink_node from balance_pgdat
6664	* will pick up pages from other mem cgroup's as well. We hack
6665	* the priority and make it zero.
6666	*/
6667	shrink_lruvec(lruvec, sc: &sc);
6668
6669	trace_mm_vmscan_memcg_softlimit_reclaim_end(nr_reclaimed: sc.nr_reclaimed);
6670
6671	*nr_scanned = sc.nr_scanned;
6672
6673	return sc.nr_reclaimed;
6674	}
6675
6676	unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6677	unsigned long nr_pages,
6678	gfp_t gfp_mask,
6679	unsigned int reclaim_options,
6680	int *swappiness)
6681	{
6682	unsigned long nr_reclaimed;
6683	unsigned int noreclaim_flag;
6684	struct scan_control sc = {
6685	.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6686	.proactive_swappiness = swappiness,
6687	.gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6688	(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6689	.reclaim_idx = MAX_NR_ZONES - 1,
6690	.target_mem_cgroup = memcg,
6691	.priority = DEF_PRIORITY,
6692	.may_writepage = !laptop_mode,
6693	.may_unmap = 1,
6694	.may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6695	.proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6696	};
6697	/*
6698	* Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6699	* equal pressure on all the nodes. This is based on the assumption that
6700	* the reclaim does not bail out early.
6701	*/
6702	struct zonelist *zonelist = node_zonelist(nid: numa_node_id(), flags: sc.gfp_mask);
6703
6704	set_task_reclaim_state(current, rs: &sc.reclaim_state);
6705	trace_mm_vmscan_memcg_reclaim_begin(order: 0, gfp_flags: sc.gfp_mask);
6706	noreclaim_flag = memalloc_noreclaim_save();
6707
6708	nr_reclaimed = do_try_to_free_pages(zonelist, sc: &sc);
6709
6710	memalloc_noreclaim_restore(flags: noreclaim_flag);
6711	trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6712	set_task_reclaim_state(current, NULL);
6713
6714	return nr_reclaimed;
6715	}
6716	#endif
6717
6718	static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6719	{
6720	struct mem_cgroup *memcg;
6721	struct lruvec *lruvec;
6722
6723	if (lru_gen_enabled()) {
6724	lru_gen_age_node(pgdat, sc);
6725	return;
6726	}
6727
6728	lruvec = mem_cgroup_lruvec(NULL, pgdat);
6729	if (!can_age_anon_pages(lruvec, sc))
6730	return;
6731
6732	if (!inactive_is_low(lruvec, inactive_lru: LRU_INACTIVE_ANON))
6733	return;
6734
6735	memcg = mem_cgroup_iter(NULL, NULL, NULL);
6736	do {
6737	lruvec = mem_cgroup_lruvec(memcg, pgdat);
6738	shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6739	sc, lru: LRU_ACTIVE_ANON);
6740	memcg = mem_cgroup_iter(NULL, memcg, NULL);
6741	} while (memcg);
6742	}
6743
6744	static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6745	{
6746	int i;
6747	struct zone *zone;
6748
6749	/*
6750	* Check for watermark boosts top-down as the higher zones
6751	* are more likely to be boosted. Both watermarks and boosts
6752	* should not be checked at the same time as reclaim would
6753	* start prematurely when there is no boosting and a lower
6754	* zone is balanced.
6755	*/
6756	for (i = highest_zoneidx; i >= 0; i--) {
6757	zone = pgdat->node_zones + i;
6758	if (!managed_zone(zone))
6759	continue;
6760
6761	if (zone->watermark_boost)
6762	return true;
6763	}
6764
6765	return false;
6766	}
6767
6768	/*
6769	* Returns true if there is an eligible zone balanced for the request order
6770	* and highest_zoneidx
6771	*/
6772	static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6773	{
6774	int i;
6775	unsigned long mark = -1;
6776	struct zone *zone;
6777
6778	/*
6779	* Check watermarks bottom-up as lower zones are more likely to
6780	* meet watermarks.
6781	*/
6782	for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
6783	enum zone_stat_item item;
6784	unsigned long free_pages;
6785
6786	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6787	mark = promo_wmark_pages(z: zone);
6788	else
6789	mark = high_wmark_pages(z: zone);
6790
6791	/*
6792	* In defrag_mode, watermarks must be met in whole
6793	* blocks to avoid polluting allocator fallbacks.
6794	*
6795	* However, kswapd usually cannot accomplish this on
6796	* its own and needs kcompactd support. Once it's
6797	* reclaimed a compaction gap, and kswapd_shrink_node
6798	* has dropped order, simply ensure there are enough
6799	* base pages for compaction, wake kcompactd & sleep.
6800	*/
6801	if (defrag_mode && order)
6802	item = NR_FREE_PAGES_BLOCKS;
6803	else
6804	item = NR_FREE_PAGES;
6805
6806	/*
6807	* When there is a high number of CPUs in the system,
6808	* the cumulative error from the vmstat per-cpu cache
6809	* can blur the line between the watermarks. In that
6810	* case, be safe and get an accurate snapshot.
6811	*
6812	* TODO: NR_FREE_PAGES_BLOCKS moves in steps of
6813	* pageblock_nr_pages, while the vmstat pcp threshold
6814	* is limited to 125. On many configurations that
6815	* counter won't actually be per-cpu cached. But keep
6816	* things simple for now; revisit when somebody cares.
6817	*/
6818	free_pages = zone_page_state(zone, item);
6819	if (zone->percpu_drift_mark && free_pages < zone->percpu_drift_mark)
6820	free_pages = zone_page_state_snapshot(zone, item);
6821
6822	if (__zone_watermark_ok(z: zone, order, mark, highest_zoneidx,
6823	alloc_flags: 0, free_pages))
6824	return true;
6825	}
6826
6827	/*
6828	* If a node has no managed zone within highest_zoneidx, it does not
6829	* need balancing by definition. This can happen if a zone-restricted
6830	* allocation tries to wake a remote kswapd.
6831	*/
6832	if (mark == -1)
6833	return true;
6834
6835	return false;
6836	}
6837
6838	/* Clear pgdat state for congested, dirty or under writeback. */
6839	static void clear_pgdat_congested(pg_data_t *pgdat)
6840	{
6841	struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6842
6843	clear_bit(nr: LRUVEC_NODE_CONGESTED, addr: &lruvec->flags);
6844	clear_bit(nr: LRUVEC_CGROUP_CONGESTED, addr: &lruvec->flags);
6845	clear_bit(nr: PGDAT_DIRTY, addr: &pgdat->flags);
6846	clear_bit(nr: PGDAT_WRITEBACK, addr: &pgdat->flags);
6847	}
6848
6849	/*
6850	* Prepare kswapd for sleeping. This verifies that there are no processes
6851	* waiting in throttle_direct_reclaim() and that watermarks have been met.
6852	*
6853	* Returns true if kswapd is ready to sleep
6854	*/
6855	static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6856	int highest_zoneidx)
6857	{
6858	/*
6859	* The throttled processes are normally woken up in balance_pgdat() as
6860	* soon as allow_direct_reclaim() is true. But there is a potential
6861	* race between when kswapd checks the watermarks and a process gets
6862	* throttled. There is also a potential race if processes get
6863	* throttled, kswapd wakes, a large process exits thereby balancing the
6864	* zones, which causes kswapd to exit balance_pgdat() before reaching
6865	* the wake up checks. If kswapd is going to sleep, no process should
6866	* be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6867	* the wake up is premature, processes will wake kswapd and get
6868	* throttled again. The difference from wake ups in balance_pgdat() is
6869	* that here we are under prepare_to_wait().
6870	*/
6871	if (waitqueue_active(wq_head: &pgdat->pfmemalloc_wait))
6872	wake_up_all(&pgdat->pfmemalloc_wait);
6873
6874	/* Hopeless node, leave it to direct reclaim */
6875	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6876	return true;
6877
6878	if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6879	clear_pgdat_congested(pgdat);
6880	return true;
6881	}
6882
6883	return false;
6884	}
6885
6886	/*
6887	* kswapd shrinks a node of pages that are at or below the highest usable
6888	* zone that is currently unbalanced.
6889	*
6890	* Returns true if kswapd scanned at least the requested number of pages to
6891	* reclaim or if the lack of progress was due to pages under writeback.
6892	* This is used to determine if the scanning priority needs to be raised.
6893	*/
6894	static bool kswapd_shrink_node(pg_data_t *pgdat,
6895	struct scan_control *sc)
6896	{
6897	struct zone *zone;
6898	int z;
6899	unsigned long nr_reclaimed = sc->nr_reclaimed;
6900
6901	/* Reclaim a number of pages proportional to the number of zones */
6902	sc->nr_to_reclaim = 0;
6903	for_each_managed_zone_pgdat(zone, pgdat, z, sc->reclaim_idx) {
6904	sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6905	}
6906
6907	/*
6908	* Historically care was taken to put equal pressure on all zones but
6909	* now pressure is applied based on node LRU order.
6910	*/
6911	shrink_node(pgdat, sc);
6912
6913	/*
6914	* Fragmentation may mean that the system cannot be rebalanced for
6915	* high-order allocations. If twice the allocation size has been
6916	* reclaimed then recheck watermarks only at order-0 to prevent
6917	* excessive reclaim. Assume that a process requested a high-order
6918	* can direct reclaim/compact.
6919	*/
6920	if (sc->order && sc->nr_reclaimed >= compact_gap(order: sc->order))
6921	sc->order = 0;
6922
6923	/* account for progress from mm_account_reclaimed_pages() */
6924	return max(sc->nr_scanned, sc->nr_reclaimed - nr_reclaimed) >= sc->nr_to_reclaim;
6925	}
6926
6927	/* Page allocator PCP high watermark is lowered if reclaim is active. */
6928	static inline void
6929	update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6930	{
6931	int i;
6932	struct zone *zone;
6933
6934	for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
6935	if (active)
6936	set_bit(nr: ZONE_RECLAIM_ACTIVE, addr: &zone->flags);
6937	else
6938	clear_bit(nr: ZONE_RECLAIM_ACTIVE, addr: &zone->flags);
6939	}
6940	}
6941
6942	static inline void
6943	set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6944	{
6945	update_reclaim_active(pgdat, highest_zoneidx, active: true);
6946	}
6947
6948	static inline void
6949	clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6950	{
6951	update_reclaim_active(pgdat, highest_zoneidx, active: false);
6952	}
6953
6954	/*
6955	* For kswapd, balance_pgdat() will reclaim pages across a node from zones
6956	* that are eligible for use by the caller until at least one zone is
6957	* balanced.
6958	*
6959	* Returns the order kswapd finished reclaiming at.
6960	*
6961	* kswapd scans the zones in the highmem->normal->dma direction. It skips
6962	* zones which have free_pages > high_wmark_pages(zone), but once a zone is
6963	* found to have free_pages <= high_wmark_pages(zone), any page in that zone
6964	* or lower is eligible for reclaim until at least one usable zone is
6965	* balanced.
6966	*/
6967	static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6968	{
6969	int i;
6970	unsigned long nr_soft_reclaimed;
6971	unsigned long nr_soft_scanned;
6972	unsigned long pflags;
6973	unsigned long nr_boost_reclaim;
6974	unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6975	bool boosted;
6976	struct zone *zone;
6977	struct scan_control sc = {
6978	.gfp_mask = GFP_KERNEL,
6979	.order = order,
6980	.may_unmap = 1,
6981	};
6982
6983	set_task_reclaim_state(current, rs: &sc.reclaim_state);
6984	psi_memstall_enter(flags: &pflags);
6985	__fs_reclaim_acquire(_THIS_IP_);
6986
6987	count_vm_event(item: PAGEOUTRUN);
6988
6989	/*
6990	* Account for the reclaim boost. Note that the zone boost is left in
6991	* place so that parallel allocations that are near the watermark will
6992	* stall or direct reclaim until kswapd is finished.
6993	*/
6994	nr_boost_reclaim = 0;
6995	for_each_managed_zone_pgdat(zone, pgdat, i, highest_zoneidx) {
6996	nr_boost_reclaim += zone->watermark_boost;
6997	zone_boosts[i] = zone->watermark_boost;
6998	}
6999	boosted = nr_boost_reclaim;
7000
7001	restart:
7002	set_reclaim_active(pgdat, highest_zoneidx);
7003	sc.priority = DEF_PRIORITY;
7004	do {
7005	unsigned long nr_reclaimed = sc.nr_reclaimed;
7006	bool raise_priority = true;
7007	bool balanced;
7008	bool ret;
7009	bool was_frozen;
7010
7011	sc.reclaim_idx = highest_zoneidx;
7012
7013	/*
7014	* If the number of buffer_heads exceeds the maximum allowed
7015	* then consider reclaiming from all zones. This has a dual
7016	* purpose -- on 64-bit systems it is expected that
7017	* buffer_heads are stripped during active rotation. On 32-bit
7018	* systems, highmem pages can pin lowmem memory and shrinking
7019	* buffers can relieve lowmem pressure. Reclaim may still not
7020	* go ahead if all eligible zones for the original allocation
7021	* request are balanced to avoid excessive reclaim from kswapd.
7022	*/
7023	if (buffer_heads_over_limit) {
7024	for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7025	zone = pgdat->node_zones + i;
7026	if (!managed_zone(zone))
7027	continue;
7028
7029	sc.reclaim_idx = i;
7030	break;
7031	}
7032	}
7033
7034	/*
7035	* If the pgdat is imbalanced then ignore boosting and preserve
7036	* the watermarks for a later time and restart. Note that the
7037	* zone watermarks will be still reset at the end of balancing
7038	* on the grounds that the normal reclaim should be enough to
7039	* re-evaluate if boosting is required when kswapd next wakes.
7040	*/
7041	balanced = pgdat_balanced(pgdat, order: sc.order, highest_zoneidx);
7042	if (!balanced && nr_boost_reclaim) {
7043	nr_boost_reclaim = 0;
7044	goto restart;
7045	}
7046
7047	/*
7048	* If boosting is not active then only reclaim if there are no
7049	* eligible zones. Note that sc.reclaim_idx is not used as
7050	* buffer_heads_over_limit may have adjusted it.
7051	*/
7052	if (!nr_boost_reclaim && balanced)
7053	goto out;
7054
7055	/* Limit the priority of boosting to avoid reclaim writeback */
7056	if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7057	raise_priority = false;
7058
7059	/*
7060	* Do not writeback or swap pages for boosted reclaim. The
7061	* intent is to relieve pressure not issue sub-optimal IO
7062	* from reclaim context. If no pages are reclaimed, the
7063	* reclaim will be aborted.
7064	*/
7065	sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7066	sc.may_swap = !nr_boost_reclaim;
7067
7068	/*
7069	* Do some background aging, to give pages a chance to be
7070	* referenced before reclaiming. All pages are rotated
7071	* regardless of classzone as this is about consistent aging.
7072	*/
7073	kswapd_age_node(pgdat, sc: &sc);
7074
7075	/*
7076	* If we're getting trouble reclaiming, start doing writepage
7077	* even in laptop mode.
7078	*/
7079	if (sc.priority < DEF_PRIORITY - 2)
7080	sc.may_writepage = 1;
7081
7082	/* Call soft limit reclaim before calling shrink_node. */
7083	sc.nr_scanned = 0;
7084	nr_soft_scanned = 0;
7085	nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat, order: sc.order,
7086	gfp_mask: sc.gfp_mask, total_scanned: &nr_soft_scanned);
7087	sc.nr_reclaimed += nr_soft_reclaimed;
7088
7089	/*
7090	* There should be no need to raise the scanning priority if
7091	* enough pages are already being scanned that that high
7092	* watermark would be met at 100% efficiency.
7093	*/
7094	if (kswapd_shrink_node(pgdat, sc: &sc))
7095	raise_priority = false;
7096
7097	/*
7098	* If the low watermark is met there is no need for processes
7099	* to be throttled on pfmemalloc_wait as they should not be
7100	* able to safely make forward progress. Wake them
7101	*/
7102	if (waitqueue_active(wq_head: &pgdat->pfmemalloc_wait) &&
7103	allow_direct_reclaim(pgdat))
7104	wake_up_all(&pgdat->pfmemalloc_wait);
7105
7106	/* Check if kswapd should be suspending */
7107	__fs_reclaim_release(_THIS_IP_);
7108	ret = kthread_freezable_should_stop(was_frozen: &was_frozen);
7109	__fs_reclaim_acquire(_THIS_IP_);
7110	if (was_frozen || ret)
7111	break;
7112
7113	/*
7114	* Raise priority if scanning rate is too low or there was no
7115	* progress in reclaiming pages
7116	*/
7117	nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7118	nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7119
7120	/*
7121	* If reclaim made no progress for a boost, stop reclaim as
7122	* IO cannot be queued and it could be an infinite loop in
7123	* extreme circumstances.
7124	*/
7125	if (nr_boost_reclaim && !nr_reclaimed)
7126	break;
7127
7128	if (raise_priority || !nr_reclaimed)
7129	sc.priority--;
7130	} while (sc.priority >= 1);
7131
7132	/*
7133	* Restart only if it went through the priority loop all the way,
7134	* but cache_trim_mode didn't work.
7135	*/
7136	if (!sc.nr_reclaimed && sc.priority < 1 &&
7137	!sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
7138	sc.no_cache_trim_mode = 1;
7139	goto restart;
7140	}
7141
7142	if (!sc.nr_reclaimed)
7143	pgdat->kswapd_failures++;
7144
7145	out:
7146	clear_reclaim_active(pgdat, highest_zoneidx);
7147
7148	/* If reclaim was boosted, account for the reclaim done in this pass */
7149	if (boosted) {
7150	unsigned long flags;
7151
7152	for (i = 0; i <= highest_zoneidx; i++) {
7153	if (!zone_boosts[i])
7154	continue;
7155
7156	/* Increments are under the zone lock */
7157	zone = pgdat->node_zones + i;
7158	spin_lock_irqsave(&zone->lock, flags);
7159	zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7160	spin_unlock_irqrestore(lock: &zone->lock, flags);
7161	}
7162
7163	/*
7164	* As there is now likely space, wakeup kcompact to defragment
7165	* pageblocks.
7166	*/
7167	wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7168	}
7169
7170	snapshot_refaults(NULL, pgdat);
7171	__fs_reclaim_release(_THIS_IP_);
7172	psi_memstall_leave(flags: &pflags);
7173	set_task_reclaim_state(current, NULL);
7174
7175	/*
7176	* Return the order kswapd stopped reclaiming at as
7177	* prepare_kswapd_sleep() takes it into account. If another caller
7178	* entered the allocator slow path while kswapd was awake, order will
7179	* remain at the higher level.
7180	*/
7181	return sc.order;
7182	}
7183
7184	/*
7185	* The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7186	* be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7187	* not a valid index then either kswapd runs for first time or kswapd couldn't
7188	* sleep after previous reclaim attempt (node is still unbalanced). In that
7189	* case return the zone index of the previous kswapd reclaim cycle.
7190	*/
7191	static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7192	enum zone_type prev_highest_zoneidx)
7193	{
7194	enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7195
7196	return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7197	}
7198
7199	static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7200	unsigned int highest_zoneidx)
7201	{
7202	long remaining = 0;
7203	DEFINE_WAIT(wait);
7204
7205	if (freezing(current) || kthread_should_stop())
7206	return;
7207
7208	prepare_to_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait, TASK_INTERRUPTIBLE);
7209
7210	/*
7211	* Try to sleep for a short interval. Note that kcompactd will only be
7212	* woken if it is possible to sleep for a short interval. This is
7213	* deliberate on the assumption that if reclaim cannot keep an
7214	* eligible zone balanced that it's also unlikely that compaction will
7215	* succeed.
7216	*/
7217	if (prepare_kswapd_sleep(pgdat, order: reclaim_order, highest_zoneidx)) {
7218	/*
7219	* Compaction records what page blocks it recently failed to
7220	* isolate pages from and skips them in the future scanning.
7221	* When kswapd is going to sleep, it is reasonable to assume
7222	* that pages and compaction may succeed so reset the cache.
7223	*/
7224	reset_isolation_suitable(pgdat);
7225
7226	/*
7227	* We have freed the memory, now we should compact it to make
7228	* allocation of the requested order possible.
7229	*/
7230	wakeup_kcompactd(pgdat, order: alloc_order, highest_zoneidx);
7231
7232	remaining = schedule_timeout(HZ/10);
7233
7234	/*
7235	* If woken prematurely then reset kswapd_highest_zoneidx and
7236	* order. The values will either be from a wakeup request or
7237	* the previous request that slept prematurely.
7238	*/
7239	if (remaining) {
7240	WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7241	kswapd_highest_zoneidx(pgdat,
7242	highest_zoneidx));
7243
7244	if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7245	WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7246	}
7247
7248	finish_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait);
7249	prepare_to_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait, TASK_INTERRUPTIBLE);
7250	}
7251
7252	/*
7253	* After a short sleep, check if it was a premature sleep. If not, then
7254	* go fully to sleep until explicitly woken up.
7255	*/
7256	if (!remaining &&
7257	prepare_kswapd_sleep(pgdat, order: reclaim_order, highest_zoneidx)) {
7258	trace_mm_vmscan_kswapd_sleep(nid: pgdat->node_id);
7259
7260	/*
7261	* vmstat counters are not perfectly accurate and the estimated
7262	* value for counters such as NR_FREE_PAGES can deviate from the
7263	* true value by nr_online_cpus * threshold. To avoid the zone
7264	* watermarks being breached while under pressure, we reduce the
7265	* per-cpu vmstat threshold while kswapd is awake and restore
7266	* them before going back to sleep.
7267	*/
7268	set_pgdat_percpu_threshold(pgdat, calculate_pressure: calculate_normal_threshold);
7269
7270	if (!kthread_should_stop())
7271	schedule();
7272
7273	set_pgdat_percpu_threshold(pgdat, calculate_pressure: calculate_pressure_threshold);
7274	} else {
7275	if (remaining)
7276	count_vm_event(item: KSWAPD_LOW_WMARK_HIT_QUICKLY);
7277	else
7278	count_vm_event(item: KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7279	}
7280	finish_wait(wq_head: &pgdat->kswapd_wait, wq_entry: &wait);
7281	}
7282
7283	/*
7284	* The background pageout daemon, started as a kernel thread
7285	* from the init process.
7286	*
7287	* This basically trickles out pages so that we have _some_
7288	* free memory available even if there is no other activity
7289	* that frees anything up. This is needed for things like routing
7290	* etc, where we otherwise might have all activity going on in
7291	* asynchronous contexts that cannot page things out.
7292	*
7293	* If there are applications that are active memory-allocators
7294	* (most normal use), this basically shouldn't matter.
7295	*/
7296	static int kswapd(void *p)
7297	{
7298	unsigned int alloc_order, reclaim_order;
7299	unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7300	pg_data_t *pgdat = (pg_data_t *)p;
7301	struct task_struct *tsk = current;
7302
7303	/*
7304	* Tell the memory management that we're a "memory allocator",
7305	* and that if we need more memory we should get access to it
7306	* regardless (see "__alloc_pages()"). "kswapd" should
7307	* never get caught in the normal page freeing logic.
7308	*
7309	* (Kswapd normally doesn't need memory anyway, but sometimes
7310	* you need a small amount of memory in order to be able to
7311	* page out something else, and this flag essentially protects
7312	* us from recursively trying to free more memory as we're
7313	* trying to free the first piece of memory in the first place).
7314	*/
7315	tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7316	set_freezable();
7317
7318	WRITE_ONCE(pgdat->kswapd_order, 0);
7319	WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7320	atomic_set(v: &pgdat->nr_writeback_throttled, i: 0);
7321	for ( ; ; ) {
7322	bool was_frozen;
7323
7324	alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7325	highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7326	prev_highest_zoneidx: highest_zoneidx);
7327
7328	kswapd_try_sleep:
7329	kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7330	highest_zoneidx);
7331
7332	/* Read the new order and highest_zoneidx */
7333	alloc_order = READ_ONCE(pgdat->kswapd_order);
7334	highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7335	prev_highest_zoneidx: highest_zoneidx);
7336	WRITE_ONCE(pgdat->kswapd_order, 0);
7337	WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7338
7339	if (kthread_freezable_should_stop(was_frozen: &was_frozen))
7340	break;
7341
7342	/*
7343	* We can speed up thawing tasks if we don't call balance_pgdat
7344	* after returning from the refrigerator
7345	*/
7346	if (was_frozen)
7347	continue;
7348
7349	/*
7350	* Reclaim begins at the requested order but if a high-order
7351	* reclaim fails then kswapd falls back to reclaiming for
7352	* order-0. If that happens, kswapd will consider sleeping
7353	* for the order it finished reclaiming at (reclaim_order)
7354	* but kcompactd is woken to compact for the original
7355	* request (alloc_order).
7356	*/
7357	trace_mm_vmscan_kswapd_wake(nid: pgdat->node_id, zid: highest_zoneidx,
7358	order: alloc_order);
7359	reclaim_order = balance_pgdat(pgdat, order: alloc_order,
7360	highest_zoneidx);
7361	if (reclaim_order < alloc_order)
7362	goto kswapd_try_sleep;
7363	}
7364
7365	tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7366
7367	return 0;
7368	}
7369
7370	/*
7371	* A zone is low on free memory or too fragmented for high-order memory. If
7372	* kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7373	* pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7374	* has failed or is not needed, still wake up kcompactd if only compaction is
7375	* needed.
7376	*/
7377	void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7378	enum zone_type highest_zoneidx)
7379	{
7380	pg_data_t *pgdat;
7381	enum zone_type curr_idx;
7382
7383	if (!managed_zone(zone))
7384	return;
7385
7386	if (!cpuset_zone_allowed(z: zone, gfp_mask: gfp_flags))
7387	return;
7388
7389	pgdat = zone->zone_pgdat;
7390	curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7391
7392	if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7393	WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7394
7395	if (READ_ONCE(pgdat->kswapd_order) < order)
7396	WRITE_ONCE(pgdat->kswapd_order, order);
7397
7398	if (!waitqueue_active(wq_head: &pgdat->kswapd_wait))
7399	return;
7400
7401	/* Hopeless node, leave it to direct reclaim if possible */
7402	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7403	(pgdat_balanced(pgdat, order, highest_zoneidx) &&
7404	!pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7405	/*
7406	* There may be plenty of free memory available, but it's too
7407	* fragmented for high-order allocations. Wake up kcompactd
7408	* and rely on compaction_suitable() to determine if it's
7409	* needed. If it fails, it will defer subsequent attempts to
7410	* ratelimit its work.
7411	*/
7412	if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7413	wakeup_kcompactd(pgdat, order, highest_zoneidx);
7414	return;
7415	}
7416
7417	trace_mm_vmscan_wakeup_kswapd(nid: pgdat->node_id, zid: highest_zoneidx, order,
7418	gfp_flags);
7419	wake_up_interruptible(&pgdat->kswapd_wait);
7420	}
7421
7422	#ifdef CONFIG_HIBERNATION
7423	/*
7424	* Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7425	* freed pages.
7426	*
7427	* Rather than trying to age LRUs the aim is to preserve the overall
7428	* LRU order by reclaiming preferentially
7429	* inactive > active > active referenced > active mapped
7430	*/
7431	unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7432	{
7433	struct scan_control sc = {
7434	.nr_to_reclaim = nr_to_reclaim,
7435	.gfp_mask = GFP_HIGHUSER_MOVABLE,
7436	.reclaim_idx = MAX_NR_ZONES - 1,
7437	.priority = DEF_PRIORITY,
7438	.may_writepage = 1,
7439	.may_unmap = 1,
7440	.may_swap = 1,
7441	.hibernation_mode = 1,
7442	};
7443	struct zonelist *zonelist = node_zonelist(nid: numa_node_id(), flags: sc.gfp_mask);
7444	unsigned long nr_reclaimed;
7445	unsigned int noreclaim_flag;
7446
7447	fs_reclaim_acquire(gfp_mask: sc.gfp_mask);
7448	noreclaim_flag = memalloc_noreclaim_save();
7449	set_task_reclaim_state(current, rs: &sc.reclaim_state);
7450
7451	nr_reclaimed = do_try_to_free_pages(zonelist, sc: &sc);
7452
7453	set_task_reclaim_state(current, NULL);
7454	memalloc_noreclaim_restore(flags: noreclaim_flag);
7455	fs_reclaim_release(gfp_mask: sc.gfp_mask);
7456
7457	return nr_reclaimed;
7458	}
7459	#endif /* CONFIG_HIBERNATION */
7460
7461	/*
7462	* This kswapd start function will be called by init and node-hot-add.
7463	*/
7464	void __meminit kswapd_run(int nid)
7465	{
7466	pg_data_t *pgdat = NODE_DATA(nid);
7467
7468	pgdat_kswapd_lock(pgdat);
7469	if (!pgdat->kswapd) {
7470	pgdat->kswapd = kthread_create_on_node(threadfn: kswapd, data: pgdat, node: nid, namefmt: "kswapd%d", nid);
7471	if (IS_ERR(ptr: pgdat->kswapd)) {
7472	/* failure at boot is fatal */
7473	pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7474	nid, PTR_ERR(pgdat->kswapd));
7475	BUG_ON(system_state < SYSTEM_RUNNING);
7476	pgdat->kswapd = NULL;
7477	} else {
7478	wake_up_process(tsk: pgdat->kswapd);
7479	}
7480	}
7481	pgdat_kswapd_unlock(pgdat);
7482	}
7483
7484	/*
7485	* Called by memory hotplug when all memory in a node is offlined. Caller must
7486	* be holding mem_hotplug_begin/done().
7487	*/
7488	void __meminit kswapd_stop(int nid)
7489	{
7490	pg_data_t *pgdat = NODE_DATA(nid);
7491	struct task_struct *kswapd;
7492
7493	pgdat_kswapd_lock(pgdat);
7494	kswapd = pgdat->kswapd;
7495	if (kswapd) {
7496	kthread_stop(k: kswapd);
7497	pgdat->kswapd = NULL;
7498	}
7499	pgdat_kswapd_unlock(pgdat);
7500	}
7501
7502	static const struct ctl_table vmscan_sysctl_table[] = {
7503	{
7504	.procname = "swappiness",
7505	.data = &vm_swappiness,
7506	.maxlen = sizeof(vm_swappiness),
7507	.mode = 0644,
7508	.proc_handler = proc_dointvec_minmax,
7509	.extra1 = SYSCTL_ZERO,
7510	.extra2 = SYSCTL_TWO_HUNDRED,
7511	},
7512	#ifdef CONFIG_NUMA
7513	{
7514	.procname = "zone_reclaim_mode",
7515	.data = &node_reclaim_mode,
7516	.maxlen = sizeof(node_reclaim_mode),
7517	.mode = 0644,
7518	.proc_handler = proc_dointvec_minmax,
7519	.extra1 = SYSCTL_ZERO,
7520	}
7521	#endif
7522	};
7523
7524	static int __init kswapd_init(void)
7525	{
7526	int nid;
7527
7528	swap_setup();
7529	for_each_node_state(nid, N_MEMORY)
7530	kswapd_run(nid);
7531	register_sysctl_init("vm", vmscan_sysctl_table);
7532	return 0;
7533	}
7534
7535	module_init(kswapd_init)
7536
7537	#ifdef CONFIG_NUMA
7538	/*
7539	* Node reclaim mode
7540	*
7541	* If non-zero call node_reclaim when the number of free pages falls below
7542	* the watermarks.
7543	*/
7544	int node_reclaim_mode __read_mostly;
7545
7546	/*
7547	* Priority for NODE_RECLAIM. This determines the fraction of pages
7548	* of a node considered for each zone_reclaim. 4 scans 1/16th of
7549	* a zone.
7550	*/
7551	#define NODE_RECLAIM_PRIORITY 4
7552
7553	/*
7554	* Percentage of pages in a zone that must be unmapped for node_reclaim to
7555	* occur.
7556	*/
7557	int sysctl_min_unmapped_ratio = 1;
7558
7559	/*
7560	* If the number of slab pages in a zone grows beyond this percentage then
7561	* slab reclaim needs to occur.
7562	*/
7563	int sysctl_min_slab_ratio = 5;
7564
7565	static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7566	{
7567	unsigned long file_mapped = node_page_state(pgdat, item: NR_FILE_MAPPED);
7568	unsigned long file_lru = node_page_state(pgdat, item: NR_INACTIVE_FILE) +
7569	node_page_state(pgdat, item: NR_ACTIVE_FILE);
7570
7571	/*
7572	* It's possible for there to be more file mapped pages than
7573	* accounted for by the pages on the file LRU lists because
7574	* tmpfs pages accounted for as ANON can also be FILE_MAPPED
7575	*/
7576	return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7577	}
7578
7579	/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7580	static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7581	{
7582	unsigned long nr_pagecache_reclaimable;
7583	unsigned long delta = 0;
7584
7585	/*
7586	* If RECLAIM_UNMAP is set, then all file pages are considered
7587	* potentially reclaimable. Otherwise, we have to worry about
7588	* pages like swapcache and node_unmapped_file_pages() provides
7589	* a better estimate
7590	*/
7591	if (node_reclaim_mode & RECLAIM_UNMAP)
7592	nr_pagecache_reclaimable = node_page_state(pgdat, item: NR_FILE_PAGES);
7593	else
7594	nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7595
7596	/* If we can't clean pages, remove dirty pages from consideration */
7597	if (!(node_reclaim_mode & RECLAIM_WRITE))
7598	delta += node_page_state(pgdat, item: NR_FILE_DIRTY);
7599
7600	/* Watch for any possible underflows due to delta */
7601	if (unlikely(delta > nr_pagecache_reclaimable))
7602	delta = nr_pagecache_reclaimable;
7603
7604	return nr_pagecache_reclaimable - delta;
7605	}
7606
7607	/*
7608	* Try to free up some pages from this node through reclaim.
7609	*/
7610	static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7611	{
7612	/* Minimum pages needed in order to stay on node */
7613	const unsigned long nr_pages = 1 << order;
7614	struct task_struct *p = current;
7615	unsigned int noreclaim_flag;
7616	struct scan_control sc = {
7617	.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7618	.gfp_mask = current_gfp_context(flags: gfp_mask),
7619	.order = order,
7620	.priority = NODE_RECLAIM_PRIORITY,
7621	.may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7622	.may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7623	.may_swap = 1,
7624	.reclaim_idx = gfp_zone(flags: gfp_mask),
7625	};
7626	unsigned long pflags;
7627
7628	trace_mm_vmscan_node_reclaim_begin(nid: pgdat->node_id, order,
7629	gfp_flags: sc.gfp_mask);
7630
7631	cond_resched();
7632	psi_memstall_enter(flags: &pflags);
7633	delayacct_freepages_start();
7634	fs_reclaim_acquire(gfp_mask: sc.gfp_mask);
7635	/*
7636	* We need to be able to allocate from the reserves for RECLAIM_UNMAP
7637	*/
7638	noreclaim_flag = memalloc_noreclaim_save();
7639	set_task_reclaim_state(task: p, rs: &sc.reclaim_state);
7640
7641	if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7642	node_page_state_pages(pgdat, item: NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7643	/*
7644	* Free memory by calling shrink node with increasing
7645	* priorities until we have enough memory freed.
7646	*/
7647	do {
7648	shrink_node(pgdat, sc: &sc);
7649	} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7650	}
7651
7652	set_task_reclaim_state(task: p, NULL);
7653	memalloc_noreclaim_restore(flags: noreclaim_flag);
7654	fs_reclaim_release(gfp_mask: sc.gfp_mask);
7655	psi_memstall_leave(flags: &pflags);
7656	delayacct_freepages_end();
7657
7658	trace_mm_vmscan_node_reclaim_end(nr_reclaimed: sc.nr_reclaimed);
7659
7660	return sc.nr_reclaimed >= nr_pages;
7661	}
7662
7663	int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7664	{
7665	int ret;
7666
7667	/*
7668	* Node reclaim reclaims unmapped file backed pages and
7669	* slab pages if we are over the defined limits.
7670	*
7671	* A small portion of unmapped file backed pages is needed for
7672	* file I/O otherwise pages read by file I/O will be immediately
7673	* thrown out if the node is overallocated. So we do not reclaim
7674	* if less than a specified percentage of the node is used by
7675	* unmapped file backed pages.
7676	*/
7677	if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7678	node_page_state_pages(pgdat, item: NR_SLAB_RECLAIMABLE_B) <=
7679	pgdat->min_slab_pages)
7680	return NODE_RECLAIM_FULL;
7681
7682	/*
7683	* Do not scan if the allocation should not be delayed.
7684	*/
7685	if (!gfpflags_allow_blocking(gfp_flags: gfp_mask) || (current->flags & PF_MEMALLOC))
7686	return NODE_RECLAIM_NOSCAN;
7687
7688	/*
7689	* Only run node reclaim on the local node or on nodes that do not
7690	* have associated processors. This will favor the local processor
7691	* over remote processors and spread off node memory allocations
7692	* as wide as possible.
7693	*/
7694	if (node_state(node: pgdat->node_id, state: N_CPU) && pgdat->node_id != numa_node_id())
7695	return NODE_RECLAIM_NOSCAN;
7696
7697	if (test_and_set_bit_lock(nr: PGDAT_RECLAIM_LOCKED, addr: &pgdat->flags))
7698	return NODE_RECLAIM_NOSCAN;
7699
7700	ret = __node_reclaim(pgdat, gfp_mask, order);
7701	clear_bit_unlock(nr: PGDAT_RECLAIM_LOCKED, addr: &pgdat->flags);
7702
7703	if (ret)
7704	count_vm_event(item: PGSCAN_ZONE_RECLAIM_SUCCESS);
7705	else
7706	count_vm_event(item: PGSCAN_ZONE_RECLAIM_FAILED);
7707
7708	return ret;
7709	}
7710	#endif
7711
7712	/**
7713	* check_move_unevictable_folios - Move evictable folios to appropriate zone
7714	* lru list
7715	* @fbatch: Batch of lru folios to check.
7716	*
7717	* Checks folios for evictability, if an evictable folio is in the unevictable
7718	* lru list, moves it to the appropriate evictable lru list. This function
7719	* should be only used for lru folios.
7720	*/
7721	void check_move_unevictable_folios(struct folio_batch *fbatch)
7722	{
7723	struct lruvec *lruvec = NULL;
7724	int pgscanned = 0;
7725	int pgrescued = 0;
7726	int i;
7727
7728	for (i = 0; i < fbatch->nr; i++) {
7729	struct folio *folio = fbatch->folios[i];
7730	int nr_pages = folio_nr_pages(folio);
7731
7732	pgscanned += nr_pages;
7733
7734	/* block memcg migration while the folio moves between lrus */
7735	if (!folio_test_clear_lru(folio))
7736	continue;
7737
7738	lruvec = folio_lruvec_relock_irq(folio, locked_lruvec: lruvec);
7739	if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7740	lruvec_del_folio(lruvec, folio);
7741	folio_clear_unevictable(folio);
7742	lruvec_add_folio(lruvec, folio);
7743	pgrescued += nr_pages;
7744	}
7745	folio_set_lru(folio);
7746	}
7747
7748	if (lruvec) {
7749	__count_vm_events(item: UNEVICTABLE_PGRESCUED, delta: pgrescued);
7750	__count_vm_events(item: UNEVICTABLE_PGSCANNED, delta: pgscanned);
7751	unlock_page_lruvec_irq(lruvec);
7752	} else if (pgscanned) {
7753	count_vm_events(item: UNEVICTABLE_PGSCANNED, delta: pgscanned);
7754	}
7755	}
7756	EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
7757