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