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