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 #ifdef CONFIG_CMA
1608 /*
1609 * It is waste of effort to scan and reclaim CMA pages if it is not available
1610 * for current allocation context. Kswapd can not be enrolled as it can not
1611 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1612 */
1613 static bool skip_cma(struct folio *folio, struct scan_control *sc)
1614 {
1615 return !current_is_kswapd() &&
1616 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1617 folio_migratetype(folio) == MIGRATE_CMA;
1618 }
1619 #else
1620 static bool skip_cma(struct folio *folio, struct scan_control *sc)
1621 {
1622 return false;
1623 }
1624 #endif
1625
1626 /*
1627 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1628 *
1629 * lruvec->lru_lock is heavily contended. Some of the functions that
1630 * shrink the lists perform better by taking out a batch of pages
1631 * and working on them outside the LRU lock.
1632 *
1633 * For pagecache intensive workloads, this function is the hottest
1634 * spot in the kernel (apart from copy_*_user functions).
1635 *
1636 * Lru_lock must be held before calling this function.
1637 *
1638 * @nr_to_scan: The number of eligible pages to look through on the list.
1639 * @lruvec: The LRU vector to pull pages from.
1640 * @dst: The temp list to put pages on to.
1641 * @nr_scanned: The number of pages that were scanned.
1642 * @sc: The scan_control struct for this reclaim session
1643 * @lru: LRU list id for isolating
1644 *
1645 * returns how many pages were moved onto *@dst.
1646 */
1647 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1648 struct lruvec *lruvec, struct list_head *dst,
1649 unsigned long *nr_scanned, struct scan_control *sc,
1650 enum lru_list lru)
1651 {
1652 struct list_head *src = &lruvec->lists[lru];
1653 unsigned long nr_taken = 0;
1654 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1655 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1656 unsigned long skipped = 0;
1657 unsigned long scan, total_scan, nr_pages;
1658 LIST_HEAD(folios_skipped);
1659
1660 total_scan = 0;
1661 scan = 0;
1662 while (scan < nr_to_scan && !list_empty(src)) {
1663 struct list_head *move_to = src;
1664 struct folio *folio;
1665
1666 folio = lru_to_folio(src);
1667 prefetchw_prev_lru_folio(folio, src, flags);
1668
1669 nr_pages = folio_nr_pages(folio);
1670 total_scan += nr_pages;
1671
1672 if (folio_zonenum(folio) > sc->reclaim_idx ||
1673 skip_cma(folio, sc)) {
1674 nr_skipped[folio_zonenum(folio)] += nr_pages;
1675 move_to = &folios_skipped;
1676 goto move;
1677 }
1678
1679 /*
1680 * Do not count skipped folios because that makes the function
1681 * return with no isolated folios if the LRU mostly contains
1682 * ineligible folios. This causes the VM to not reclaim any
1683 * folios, triggering a premature OOM.
1684 * Account all pages in a folio.
1685 */
1686 scan += nr_pages;
1687
1688 if (!folio_test_lru(folio))
1689 goto move;
1690 if (!sc->may_unmap && folio_mapped(folio))
1691 goto move;
1692
1693 /*
1694 * Be careful not to clear the lru flag until after we're
1695 * sure the folio is not being freed elsewhere -- the
1696 * folio release code relies on it.
1697 */
1698 if (unlikely(!folio_try_get(folio)))
1699 goto move;
1700
1701 if (!folio_test_clear_lru(folio)) {
1702 /* Another thread is already isolating this folio */
1703 folio_put(folio);
1704 goto move;
1705 }
1706
1707 nr_taken += nr_pages;
1708 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1709 move_to = dst;
1710 move:
1711 list_move(&folio->lru, move_to);
1712 }
1713
1714 /*
1715 * Splice any skipped folios to the start of the LRU list. Note that
1716 * this disrupts the LRU order when reclaiming for lower zones but
1717 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1718 * scanning would soon rescan the same folios to skip and waste lots
1719 * of cpu cycles.
1720 */
1721 if (!list_empty(&folios_skipped)) {
1722 int zid;
1723
1724 list_splice(&folios_skipped, src);
1725 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1726 if (!nr_skipped[zid])
1727 continue;
1728
1729 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1730 skipped += nr_skipped[zid];
1731 }
1732 }
1733 *nr_scanned = total_scan;
1734 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1735 total_scan, skipped, nr_taken, lru);
1736 update_lru_sizes(lruvec, lru, nr_zone_taken);
1737 return nr_taken;
1738 }
1739
1740 /**
1741 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1742 * @folio: Folio to isolate from its LRU list.
1743 *
1744 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1745 * corresponding to whatever LRU list the folio was on.
1746 *
1747 * The folio will have its LRU flag cleared. If it was found on the
1748 * active list, it will have the Active flag set. If it was found on the
1749 * unevictable list, it will have the Unevictable flag set. These flags
1750 * may need to be cleared by the caller before letting the page go.
1751 *
1752 * Context:
1753 *
1754 * (1) Must be called with an elevated refcount on the folio. This is a
1755 * fundamental difference from isolate_lru_folios() (which is called
1756 * without a stable reference).
1757 * (2) The lru_lock must not be held.
1758 * (3) Interrupts must be enabled.
1759 *
1760 * Return: true if the folio was removed from an LRU list.
1761 * false if the folio was not on an LRU list.
1762 */
1763 bool folio_isolate_lru(struct folio *folio)
1764 {
1765 bool ret = false;
1766
1767 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1768
1769 if (folio_test_clear_lru(folio)) {
1770 struct lruvec *lruvec;
1771
1772 folio_get(folio);
1773 lruvec = folio_lruvec_lock_irq(folio);
1774 lruvec_del_folio(lruvec, folio);
1775 unlock_page_lruvec_irq(lruvec);
1776 ret = true;
1777 }
1778
1779 return ret;
1780 }
1781
1782 /*
1783 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1784 * then get rescheduled. When there are massive number of tasks doing page
1785 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1786 * the LRU list will go small and be scanned faster than necessary, leading to
1787 * unnecessary swapping, thrashing and OOM.
1788 */
1789 static bool too_many_isolated(struct pglist_data *pgdat, int file,
1790 struct scan_control *sc)
1791 {
1792 unsigned long inactive, isolated;
1793 bool too_many;
1794
1795 if (current_is_kswapd())
1796 return false;
1797
1798 if (!writeback_throttling_sane(sc))
1799 return false;
1800
1801 if (file) {
1802 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1803 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1804 } else {
1805 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1806 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1807 }
1808
1809 /*
1810 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1811 * won't get blocked by normal direct-reclaimers, forming a circular
1812 * deadlock.
1813 */
1814 if (gfp_has_io_fs(sc->gfp_mask))
1815 inactive >>= 3;
1816
1817 too_many = isolated > inactive;
1818
1819 /* Wake up tasks throttled due to too_many_isolated. */
1820 if (!too_many)
1821 wake_throttle_isolated(pgdat);
1822
1823 return too_many;
1824 }
1825
1826 /*
1827 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1828 *
1829 * Returns the number of pages moved to the given lruvec.
1830 */
1831 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1832 struct list_head *list)
1833 {
1834 int nr_pages, nr_moved = 0;
1835 struct folio_batch free_folios;
1836
1837 folio_batch_init(&free_folios);
1838 while (!list_empty(list)) {
1839 struct folio *folio = lru_to_folio(list);
1840
1841 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1842 list_del(&folio->lru);
1843 if (unlikely(!folio_evictable(folio))) {
1844 spin_unlock_irq(&lruvec->lru_lock);
1845 folio_putback_lru(folio);
1846 spin_lock_irq(&lruvec->lru_lock);
1847 continue;
1848 }
1849
1850 /*
1851 * The folio_set_lru needs to be kept here for list integrity.
1852 * Otherwise:
1853 * #0 move_folios_to_lru #1 release_pages
1854 * if (!folio_put_testzero())
1855 * if (folio_put_testzero())
1856 * !lru //skip lru_lock
1857 * folio_set_lru()
1858 * list_add(&folio->lru,)
1859 * list_add(&folio->lru,)
1860 */
1861 folio_set_lru(folio);
1862
1863 if (unlikely(folio_put_testzero(folio))) {
1864 __folio_clear_lru_flags(folio);
1865
1866 folio_undo_large_rmappable(folio);
1867 if (folio_batch_add(&free_folios, folio) == 0) {
1868 spin_unlock_irq(&lruvec->lru_lock);
1869 mem_cgroup_uncharge_folios(&free_folios);
1870 free_unref_folios(&free_folios);
1871 spin_lock_irq(&lruvec->lru_lock);
1872 }
1873
1874 continue;
1875 }
1876
1877 /*
1878 * All pages were isolated from the same lruvec (and isolation
1879 * inhibits memcg migration).
1880 */
1881 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1882 lruvec_add_folio(lruvec, folio);
1883 nr_pages = folio_nr_pages(folio);
1884 nr_moved += nr_pages;
1885 if (folio_test_active(folio))
1886 workingset_age_nonresident(lruvec, nr_pages);
1887 }
1888
1889 if (free_folios.nr) {
1890 spin_unlock_irq(&lruvec->lru_lock);
1891 mem_cgroup_uncharge_folios(&free_folios);
1892 free_unref_folios(&free_folios);
1893 spin_lock_irq(&lruvec->lru_lock);
1894 }
1895
1896 return nr_moved;
1897 }
1898
1899 /*
1900 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1901 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1902 * we should not throttle. Otherwise it is safe to do so.
1903 */
1904 static int current_may_throttle(void)
1905 {
1906 return !(current->flags & PF_LOCAL_THROTTLE);
1907 }
1908
1909 /*
1910 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1911 * of reclaimed pages
1912 */
1913 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1914 struct lruvec *lruvec, struct scan_control *sc,
1915 enum lru_list lru)
1916 {
1917 LIST_HEAD(folio_list);
1918 unsigned long nr_scanned;
1919 unsigned int nr_reclaimed = 0;
1920 unsigned long nr_taken;
1921 struct reclaim_stat stat;
1922 bool file = is_file_lru(lru);
1923 enum vm_event_item item;
1924 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1925 bool stalled = false;
1926
1927 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1928 if (stalled)
1929 return 0;
1930
1931 /* wait a bit for the reclaimer. */
1932 stalled = true;
1933 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1934
1935 /* We are about to die and free our memory. Return now. */
1936 if (fatal_signal_pending(current))
1937 return SWAP_CLUSTER_MAX;
1938 }
1939
1940 lru_add_drain();
1941
1942 spin_lock_irq(&lruvec->lru_lock);
1943
1944 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1945 &nr_scanned, sc, lru);
1946
1947 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1948 item = PGSCAN_KSWAPD + reclaimer_offset();
1949 if (!cgroup_reclaim(sc))
1950 __count_vm_events(item, nr_scanned);
1951 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1952 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1953
1954 spin_unlock_irq(&lruvec->lru_lock);
1955
1956 if (nr_taken == 0)
1957 return 0;
1958
1959 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1960
1961 spin_lock_irq(&lruvec->lru_lock);
1962 move_folios_to_lru(lruvec, &folio_list);
1963
1964 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1965 item = PGSTEAL_KSWAPD + reclaimer_offset();
1966 if (!cgroup_reclaim(sc))
1967 __count_vm_events(item, nr_reclaimed);
1968 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1969 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1970 spin_unlock_irq(&lruvec->lru_lock);
1971
1972 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1973
1974 /*
1975 * If dirty folios are scanned that are not queued for IO, it
1976 * implies that flushers are not doing their job. This can
1977 * happen when memory pressure pushes dirty folios to the end of
1978 * the LRU before the dirty limits are breached and the dirty
1979 * data has expired. It can also happen when the proportion of
1980 * dirty folios grows not through writes but through memory
1981 * pressure reclaiming all the clean cache. And in some cases,
1982 * the flushers simply cannot keep up with the allocation
1983 * rate. Nudge the flusher threads in case they are asleep.
1984 */
1985 if (stat.nr_unqueued_dirty == nr_taken) {
1986 wakeup_flusher_threads(WB_REASON_VMSCAN);
1987 /*
1988 * For cgroupv1 dirty throttling is achieved by waking up
1989 * the kernel flusher here and later waiting on folios
1990 * which are in writeback to finish (see shrink_folio_list()).
1991 *
1992 * Flusher may not be able to issue writeback quickly
1993 * enough for cgroupv1 writeback throttling to work
1994 * on a large system.
1995 */
1996 if (!writeback_throttling_sane(sc))
1997 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1998 }
1999
2000 sc->nr.dirty += stat.nr_dirty;
2001 sc->nr.congested += stat.nr_congested;
2002 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2003 sc->nr.writeback += stat.nr_writeback;
2004 sc->nr.immediate += stat.nr_immediate;
2005 sc->nr.taken += nr_taken;
2006 if (file)
2007 sc->nr.file_taken += nr_taken;
2008
2009 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2010 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2011 return nr_reclaimed;
2012 }
2013
2014 /*
2015 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2016 *
2017 * We move them the other way if the folio is referenced by one or more
2018 * processes.
2019 *
2020 * If the folios are mostly unmapped, the processing is fast and it is
2021 * appropriate to hold lru_lock across the whole operation. But if
2022 * the folios are mapped, the processing is slow (folio_referenced()), so
2023 * we should drop lru_lock around each folio. It's impossible to balance
2024 * this, so instead we remove the folios from the LRU while processing them.
2025 * It is safe to rely on the active flag against the non-LRU folios in here
2026 * because nobody will play with that bit on a non-LRU folio.
2027 *
2028 * The downside is that we have to touch folio->_refcount against each folio.
2029 * But we had to alter folio->flags anyway.
2030 */
2031 static void shrink_active_list(unsigned long nr_to_scan,
2032 struct lruvec *lruvec,
2033 struct scan_control *sc,
2034 enum lru_list lru)
2035 {
2036 unsigned long nr_taken;
2037 unsigned long nr_scanned;
2038 unsigned long vm_flags;
2039 LIST_HEAD(l_hold); /* The folios which were snipped off */
2040 LIST_HEAD(l_active);
2041 LIST_HEAD(l_inactive);
2042 unsigned nr_deactivate, nr_activate;
2043 unsigned nr_rotated = 0;
2044 bool file = is_file_lru(lru);
2045 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2046
2047 lru_add_drain();
2048
2049 spin_lock_irq(&lruvec->lru_lock);
2050
2051 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2052 &nr_scanned, sc, lru);
2053
2054 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2055
2056 if (!cgroup_reclaim(sc))
2057 __count_vm_events(PGREFILL, nr_scanned);
2058 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2059
2060 spin_unlock_irq(&lruvec->lru_lock);
2061
2062 while (!list_empty(&l_hold)) {
2063 struct folio *folio;
2064
2065 cond_resched();
2066 folio = lru_to_folio(&l_hold);
2067 list_del(&folio->lru);
2068
2069 if (unlikely(!folio_evictable(folio))) {
2070 folio_putback_lru(folio);
2071 continue;
2072 }
2073
2074 if (unlikely(buffer_heads_over_limit)) {
2075 if (folio_needs_release(folio) &&
2076 folio_trylock(folio)) {
2077 filemap_release_folio(folio, 0);
2078 folio_unlock(folio);
2079 }
2080 }
2081
2082 /* Referenced or rmap lock contention: rotate */
2083 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2084 &vm_flags) != 0) {
2085 /*
2086 * Identify referenced, file-backed active folios and
2087 * give them one more trip around the active list. So
2088 * that executable code get better chances to stay in
2089 * memory under moderate memory pressure. Anon folios
2090 * are not likely to be evicted by use-once streaming
2091 * IO, plus JVM can create lots of anon VM_EXEC folios,
2092 * so we ignore them here.
2093 */
2094 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2095 nr_rotated += folio_nr_pages(folio);
2096 list_add(&folio->lru, &l_active);
2097 continue;
2098 }
2099 }
2100
2101 folio_clear_active(folio); /* we are de-activating */
2102 folio_set_workingset(folio);
2103 list_add(&folio->lru, &l_inactive);
2104 }
2105
2106 /*
2107 * Move folios back to the lru list.
2108 */
2109 spin_lock_irq(&lruvec->lru_lock);
2110
2111 nr_activate = move_folios_to_lru(lruvec, &l_active);
2112 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2113
2114 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2115 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2116
2117 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2118 spin_unlock_irq(&lruvec->lru_lock);
2119
2120 if (nr_rotated)
2121 lru_note_cost(lruvec, file, 0, nr_rotated);
2122 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2123 nr_deactivate, nr_rotated, sc->priority, file);
2124 }
2125
2126 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2127 struct pglist_data *pgdat)
2128 {
2129 struct reclaim_stat dummy_stat;
2130 unsigned int nr_reclaimed;
2131 struct folio *folio;
2132 struct scan_control sc = {
2133 .gfp_mask = GFP_KERNEL,
2134 .may_writepage = 1,
2135 .may_unmap = 1,
2136 .may_swap = 1,
2137 .no_demotion = 1,
2138 };
2139
2140 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, true);
2141 while (!list_empty(folio_list)) {
2142 folio = lru_to_folio(folio_list);
2143 list_del(&folio->lru);
2144 folio_putback_lru(folio);
2145 }
2146
2147 return nr_reclaimed;
2148 }
2149
2150 unsigned long reclaim_pages(struct list_head *folio_list)
2151 {
2152 int nid;
2153 unsigned int nr_reclaimed = 0;
2154 LIST_HEAD(node_folio_list);
2155 unsigned int noreclaim_flag;
2156
2157 if (list_empty(folio_list))
2158 return nr_reclaimed;
2159
2160 noreclaim_flag = memalloc_noreclaim_save();
2161
2162 nid = folio_nid(lru_to_folio(folio_list));
2163 do {
2164 struct folio *folio = lru_to_folio(folio_list);
2165
2166 if (nid == folio_nid(folio)) {
2167 folio_clear_active(folio);
2168 list_move(&folio->lru, &node_folio_list);
2169 continue;
2170 }
2171
2172 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2173 nid = folio_nid(lru_to_folio(folio_list));
2174 } while (!list_empty(folio_list));
2175
2176 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2177
2178 memalloc_noreclaim_restore(noreclaim_flag);
2179
2180 return nr_reclaimed;
2181 }
2182
2183 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2184 struct lruvec *lruvec, struct scan_control *sc)
2185 {
2186 if (is_active_lru(lru)) {
2187 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2188 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2189 else
2190 sc->skipped_deactivate = 1;
2191 return 0;
2192 }
2193
2194 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2195 }
2196
2197 /*
2198 * The inactive anon list should be small enough that the VM never has
2199 * to do too much work.
2200 *
2201 * The inactive file list should be small enough to leave most memory
2202 * to the established workingset on the scan-resistant active list,
2203 * but large enough to avoid thrashing the aggregate readahead window.
2204 *
2205 * Both inactive lists should also be large enough that each inactive
2206 * folio has a chance to be referenced again before it is reclaimed.
2207 *
2208 * If that fails and refaulting is observed, the inactive list grows.
2209 *
2210 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2211 * on this LRU, maintained by the pageout code. An inactive_ratio
2212 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2213 *
2214 * total target max
2215 * memory ratio inactive
2216 * -------------------------------------
2217 * 10MB 1 5MB
2218 * 100MB 1 50MB
2219 * 1GB 3 250MB
2220 * 10GB 10 0.9GB
2221 * 100GB 31 3GB
2222 * 1TB 101 10GB
2223 * 10TB 320 32GB
2224 */
2225 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2226 {
2227 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2228 unsigned long inactive, active;
2229 unsigned long inactive_ratio;
2230 unsigned long gb;
2231
2232 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2233 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2234
2235 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2236 if (gb)
2237 inactive_ratio = int_sqrt(10 * gb);
2238 else
2239 inactive_ratio = 1;
2240
2241 return inactive * inactive_ratio < active;
2242 }
2243
2244 enum scan_balance {
2245 SCAN_EQUAL,
2246 SCAN_FRACT,
2247 SCAN_ANON,
2248 SCAN_FILE,
2249 };
2250
2251 static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2252 {
2253 unsigned long file;
2254 struct lruvec *target_lruvec;
2255
2256 if (lru_gen_enabled())
2257 return;
2258
2259 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2260
2261 /*
2262 * Flush the memory cgroup stats, so that we read accurate per-memcg
2263 * lruvec stats for heuristics.
2264 */
2265 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2266
2267 /*
2268 * Determine the scan balance between anon and file LRUs.
2269 */
2270 spin_lock_irq(&target_lruvec->lru_lock);
2271 sc->anon_cost = target_lruvec->anon_cost;
2272 sc->file_cost = target_lruvec->file_cost;
2273 spin_unlock_irq(&target_lruvec->lru_lock);
2274
2275 /*
2276 * Target desirable inactive:active list ratios for the anon
2277 * and file LRU lists.
2278 */
2279 if (!sc->force_deactivate) {
2280 unsigned long refaults;
2281
2282 /*
2283 * When refaults are being observed, it means a new
2284 * workingset is being established. Deactivate to get
2285 * rid of any stale active pages quickly.
2286 */
2287 refaults = lruvec_page_state(target_lruvec,
2288 WORKINGSET_ACTIVATE_ANON);
2289 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2290 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2291 sc->may_deactivate |= DEACTIVATE_ANON;
2292 else
2293 sc->may_deactivate &= ~DEACTIVATE_ANON;
2294
2295 refaults = lruvec_page_state(target_lruvec,
2296 WORKINGSET_ACTIVATE_FILE);
2297 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2298 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2299 sc->may_deactivate |= DEACTIVATE_FILE;
2300 else
2301 sc->may_deactivate &= ~DEACTIVATE_FILE;
2302 } else
2303 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2304
2305 /*
2306 * If we have plenty of inactive file pages that aren't
2307 * thrashing, try to reclaim those first before touching
2308 * anonymous pages.
2309 */
2310 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2311 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2312 !sc->no_cache_trim_mode)
2313 sc->cache_trim_mode = 1;
2314 else
2315 sc->cache_trim_mode = 0;
2316
2317 /*
2318 * Prevent the reclaimer from falling into the cache trap: as
2319 * cache pages start out inactive, every cache fault will tip
2320 * the scan balance towards the file LRU. And as the file LRU
2321 * shrinks, so does the window for rotation from references.
2322 * This means we have a runaway feedback loop where a tiny
2323 * thrashing file LRU becomes infinitely more attractive than
2324 * anon pages. Try to detect this based on file LRU size.
2325 */
2326 if (!cgroup_reclaim(sc)) {
2327 unsigned long total_high_wmark = 0;
2328 unsigned long free, anon;
2329 int z;
2330
2331 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2332 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2333 node_page_state(pgdat, NR_INACTIVE_FILE);
2334
2335 for (z = 0; z < MAX_NR_ZONES; z++) {
2336 struct zone *zone = &pgdat->node_zones[z];
2337
2338 if (!managed_zone(zone))
2339 continue;
2340
2341 total_high_wmark += high_wmark_pages(zone);
2342 }
2343
2344 /*
2345 * Consider anon: if that's low too, this isn't a
2346 * runaway file reclaim problem, but rather just
2347 * extreme pressure. Reclaim as per usual then.
2348 */
2349 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2350
2351 sc->file_is_tiny =
2352 file + free <= total_high_wmark &&
2353 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2354 anon >> sc->priority;
2355 }
2356 }
2357
2358 /*
2359 * Determine how aggressively the anon and file LRU lists should be
2360 * scanned.
2361 *
2362 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2363 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2364 */
2365 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2366 unsigned long *nr)
2367 {
2368 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2369 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2370 unsigned long anon_cost, file_cost, total_cost;
2371 int swappiness = sc_swappiness(sc, memcg);
2372 u64 fraction[ANON_AND_FILE];
2373 u64 denominator = 0; /* gcc */
2374 enum scan_balance scan_balance;
2375 unsigned long ap, fp;
2376 enum lru_list lru;
2377
2378 /* If we have no swap space, do not bother scanning anon folios. */
2379 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2380 scan_balance = SCAN_FILE;
2381 goto out;
2382 }
2383
2384 /*
2385 * Global reclaim will swap to prevent OOM even with no
2386 * swappiness, but memcg users want to use this knob to
2387 * disable swapping for individual groups completely when
2388 * using the memory controller's swap limit feature would be
2389 * too expensive.
2390 */
2391 if (cgroup_reclaim(sc) && !swappiness) {
2392 scan_balance = SCAN_FILE;
2393 goto out;
2394 }
2395
2396 /*
2397 * Do not apply any pressure balancing cleverness when the
2398 * system is close to OOM, scan both anon and file equally
2399 * (unless the swappiness setting disagrees with swapping).
2400 */
2401 if (!sc->priority && swappiness) {
2402 scan_balance = SCAN_EQUAL;
2403 goto out;
2404 }
2405
2406 /*
2407 * If the system is almost out of file pages, force-scan anon.
2408 */
2409 if (sc->file_is_tiny) {
2410 scan_balance = SCAN_ANON;
2411 goto out;
2412 }
2413
2414 /*
2415 * If there is enough inactive page cache, we do not reclaim
2416 * anything from the anonymous working right now.
2417 */
2418 if (sc->cache_trim_mode) {
2419 scan_balance = SCAN_FILE;
2420 goto out;
2421 }
2422
2423 scan_balance = SCAN_FRACT;
2424 /*
2425 * Calculate the pressure balance between anon and file pages.
2426 *
2427 * The amount of pressure we put on each LRU is inversely
2428 * proportional to the cost of reclaiming each list, as
2429 * determined by the share of pages that are refaulting, times
2430 * the relative IO cost of bringing back a swapped out
2431 * anonymous page vs reloading a filesystem page (swappiness).
2432 *
2433 * Although we limit that influence to ensure no list gets
2434 * left behind completely: at least a third of the pressure is
2435 * applied, before swappiness.
2436 *
2437 * With swappiness at 100, anon and file have equal IO cost.
2438 */
2439 total_cost = sc->anon_cost + sc->file_cost;
2440 anon_cost = total_cost + sc->anon_cost;
2441 file_cost = total_cost + sc->file_cost;
2442 total_cost = anon_cost + file_cost;
2443
2444 ap = swappiness * (total_cost + 1);
2445 ap /= anon_cost + 1;
2446
2447 fp = (MAX_SWAPPINESS - swappiness) * (total_cost + 1);
2448 fp /= file_cost + 1;
2449
2450 fraction[0] = ap;
2451 fraction[1] = fp;
2452 denominator = ap + fp;
2453 out:
2454 for_each_evictable_lru(lru) {
2455 bool file = is_file_lru(lru);
2456 unsigned long lruvec_size;
2457 unsigned long low, min;
2458 unsigned long scan;
2459
2460 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2461 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2462 &min, &low);
2463
2464 if (min || low) {
2465 /*
2466 * Scale a cgroup's reclaim pressure by proportioning
2467 * its current usage to its memory.low or memory.min
2468 * setting.
2469 *
2470 * This is important, as otherwise scanning aggression
2471 * becomes extremely binary -- from nothing as we
2472 * approach the memory protection threshold, to totally
2473 * nominal as we exceed it. This results in requiring
2474 * setting extremely liberal protection thresholds. It
2475 * also means we simply get no protection at all if we
2476 * set it too low, which is not ideal.
2477 *
2478 * If there is any protection in place, we reduce scan
2479 * pressure by how much of the total memory used is
2480 * within protection thresholds.
2481 *
2482 * There is one special case: in the first reclaim pass,
2483 * we skip over all groups that are within their low
2484 * protection. If that fails to reclaim enough pages to
2485 * satisfy the reclaim goal, we come back and override
2486 * the best-effort low protection. However, we still
2487 * ideally want to honor how well-behaved groups are in
2488 * that case instead of simply punishing them all
2489 * equally. As such, we reclaim them based on how much
2490 * memory they are using, reducing the scan pressure
2491 * again by how much of the total memory used is under
2492 * hard protection.
2493 */
2494 unsigned long cgroup_size = mem_cgroup_size(memcg);
2495 unsigned long protection;
2496
2497 /* memory.low scaling, make sure we retry before OOM */
2498 if (!sc->memcg_low_reclaim && low > min) {
2499 protection = low;
2500 sc->memcg_low_skipped = 1;
2501 } else {
2502 protection = min;
2503 }
2504
2505 /* Avoid TOCTOU with earlier protection check */
2506 cgroup_size = max(cgroup_size, protection);
2507
2508 scan = lruvec_size - lruvec_size * protection /
2509 (cgroup_size + 1);
2510
2511 /*
2512 * Minimally target SWAP_CLUSTER_MAX pages to keep
2513 * reclaim moving forwards, avoiding decrementing
2514 * sc->priority further than desirable.
2515 */
2516 scan = max(scan, SWAP_CLUSTER_MAX);
2517 } else {
2518 scan = lruvec_size;
2519 }
2520
2521 scan >>= sc->priority;
2522
2523 /*
2524 * If the cgroup's already been deleted, make sure to
2525 * scrape out the remaining cache.
2526 */
2527 if (!scan && !mem_cgroup_online(memcg))
2528 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2529
2530 switch (scan_balance) {
2531 case SCAN_EQUAL:
2532 /* Scan lists relative to size */
2533 break;
2534 case SCAN_FRACT:
2535 /*
2536 * Scan types proportional to swappiness and
2537 * their relative recent reclaim efficiency.
2538 * Make sure we don't miss the last page on
2539 * the offlined memory cgroups because of a
2540 * round-off error.
2541 */
2542 scan = mem_cgroup_online(memcg) ?
2543 div64_u64(scan * fraction[file], denominator) :
2544 DIV64_U64_ROUND_UP(scan * fraction[file],
2545 denominator);
2546 break;
2547 case SCAN_FILE:
2548 case SCAN_ANON:
2549 /* Scan one type exclusively */
2550 if ((scan_balance == SCAN_FILE) != file)
2551 scan = 0;
2552 break;
2553 default:
2554 /* Look ma, no brain */
2555 BUG();
2556 }
2557
2558 nr[lru] = scan;
2559 }
2560 }
2561
2562 /*
2563 * Anonymous LRU management is a waste if there is
2564 * ultimately no way to reclaim the memory.
2565 */
2566 static bool can_age_anon_pages(struct pglist_data *pgdat,
2567 struct scan_control *sc)
2568 {
2569 /* Aging the anon LRU is valuable if swap is present: */
2570 if (total_swap_pages > 0)
2571 return true;
2572
2573 /* Also valuable if anon pages can be demoted: */
2574 return can_demote(pgdat->node_id, sc);
2575 }
2576
2577 #ifdef CONFIG_LRU_GEN
2578
2579 #ifdef CONFIG_LRU_GEN_ENABLED
2580 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2581 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2582 #else
2583 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2584 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2585 #endif
2586
2587 static bool should_walk_mmu(void)
2588 {
2589 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2590 }
2591
2592 static bool should_clear_pmd_young(void)
2593 {
2594 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2595 }
2596
2597 /******************************************************************************
2598 * shorthand helpers
2599 ******************************************************************************/
2600
2601 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2602
2603 #define DEFINE_MAX_SEQ(lruvec) \
2604 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2605
2606 #define DEFINE_MIN_SEQ(lruvec) \
2607 unsigned long min_seq[ANON_AND_FILE] = { \
2608 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2609 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2610 }
2611
2612 #define for_each_gen_type_zone(gen, type, zone) \
2613 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2614 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2615 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2616
2617 #define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2618 #define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2619
2620 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2621 {
2622 struct pglist_data *pgdat = NODE_DATA(nid);
2623
2624 #ifdef CONFIG_MEMCG
2625 if (memcg) {
2626 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2627
2628 /* see the comment in mem_cgroup_lruvec() */
2629 if (!lruvec->pgdat)
2630 lruvec->pgdat = pgdat;
2631
2632 return lruvec;
2633 }
2634 #endif
2635 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2636
2637 return &pgdat->__lruvec;
2638 }
2639
2640 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2641 {
2642 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2643 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2644
2645 if (!sc->may_swap)
2646 return 0;
2647
2648 if (!can_demote(pgdat->node_id, sc) &&
2649 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2650 return 0;
2651
2652 return sc_swappiness(sc, memcg);
2653 }
2654
2655 static int get_nr_gens(struct lruvec *lruvec, int type)
2656 {
2657 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2658 }
2659
2660 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2661 {
2662 /* see the comment on lru_gen_folio */
2663 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2664 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2665 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2666 }
2667
2668 /******************************************************************************
2669 * Bloom filters
2670 ******************************************************************************/
2671
2672 /*
2673 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2674 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2675 * bits in a bitmap, k is the number of hash functions and n is the number of
2676 * inserted items.
2677 *
2678 * Page table walkers use one of the two filters to reduce their search space.
2679 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2680 * aging uses the double-buffering technique to flip to the other filter each
2681 * time it produces a new generation. For non-leaf entries that have enough
2682 * leaf entries, the aging carries them over to the next generation in
2683 * walk_pmd_range(); the eviction also report them when walking the rmap
2684 * in lru_gen_look_around().
2685 *
2686 * For future optimizations:
2687 * 1. It's not necessary to keep both filters all the time. The spare one can be
2688 * freed after the RCU grace period and reallocated if needed again.
2689 * 2. And when reallocating, it's worth scaling its size according to the number
2690 * of inserted entries in the other filter, to reduce the memory overhead on
2691 * small systems and false positives on large systems.
2692 * 3. Jenkins' hash function is an alternative to Knuth's.
2693 */
2694 #define BLOOM_FILTER_SHIFT 15
2695
2696 static inline int filter_gen_from_seq(unsigned long seq)
2697 {
2698 return seq % NR_BLOOM_FILTERS;
2699 }
2700
2701 static void get_item_key(void *item, int *key)
2702 {
2703 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2704
2705 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2706
2707 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2708 key[1] = hash >> BLOOM_FILTER_SHIFT;
2709 }
2710
2711 static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2712 void *item)
2713 {
2714 int key[2];
2715 unsigned long *filter;
2716 int gen = filter_gen_from_seq(seq);
2717
2718 filter = READ_ONCE(mm_state->filters[gen]);
2719 if (!filter)
2720 return true;
2721
2722 get_item_key(item, key);
2723
2724 return test_bit(key[0], filter) && test_bit(key[1], filter);
2725 }
2726
2727 static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2728 void *item)
2729 {
2730 int key[2];
2731 unsigned long *filter;
2732 int gen = filter_gen_from_seq(seq);
2733
2734 filter = READ_ONCE(mm_state->filters[gen]);
2735 if (!filter)
2736 return;
2737
2738 get_item_key(item, key);
2739
2740 if (!test_bit(key[0], filter))
2741 set_bit(key[0], filter);
2742 if (!test_bit(key[1], filter))
2743 set_bit(key[1], filter);
2744 }
2745
2746 static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2747 {
2748 unsigned long *filter;
2749 int gen = filter_gen_from_seq(seq);
2750
2751 filter = mm_state->filters[gen];
2752 if (filter) {
2753 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2754 return;
2755 }
2756
2757 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2758 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2759 WRITE_ONCE(mm_state->filters[gen], filter);
2760 }
2761
2762 /******************************************************************************
2763 * mm_struct list
2764 ******************************************************************************/
2765
2766 #ifdef CONFIG_LRU_GEN_WALKS_MMU
2767
2768 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2769 {
2770 static struct lru_gen_mm_list mm_list = {
2771 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2772 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2773 };
2774
2775 #ifdef CONFIG_MEMCG
2776 if (memcg)
2777 return &memcg->mm_list;
2778 #endif
2779 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2780
2781 return &mm_list;
2782 }
2783
2784 static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2785 {
2786 return &lruvec->mm_state;
2787 }
2788
2789 static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2790 {
2791 int key;
2792 struct mm_struct *mm;
2793 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2794 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2795
2796 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2797 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2798
2799 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2800 return NULL;
2801
2802 clear_bit(key, &mm->lru_gen.bitmap);
2803
2804 return mmget_not_zero(mm) ? mm : NULL;
2805 }
2806
2807 void lru_gen_add_mm(struct mm_struct *mm)
2808 {
2809 int nid;
2810 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2811 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2812
2813 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2814 #ifdef CONFIG_MEMCG
2815 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2816 mm->lru_gen.memcg = memcg;
2817 #endif
2818 spin_lock(&mm_list->lock);
2819
2820 for_each_node_state(nid, N_MEMORY) {
2821 struct lruvec *lruvec = get_lruvec(memcg, nid);
2822 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2823
2824 /* the first addition since the last iteration */
2825 if (mm_state->tail == &mm_list->fifo)
2826 mm_state->tail = &mm->lru_gen.list;
2827 }
2828
2829 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2830
2831 spin_unlock(&mm_list->lock);
2832 }
2833
2834 void lru_gen_del_mm(struct mm_struct *mm)
2835 {
2836 int nid;
2837 struct lru_gen_mm_list *mm_list;
2838 struct mem_cgroup *memcg = NULL;
2839
2840 if (list_empty(&mm->lru_gen.list))
2841 return;
2842
2843 #ifdef CONFIG_MEMCG
2844 memcg = mm->lru_gen.memcg;
2845 #endif
2846 mm_list = get_mm_list(memcg);
2847
2848 spin_lock(&mm_list->lock);
2849
2850 for_each_node(nid) {
2851 struct lruvec *lruvec = get_lruvec(memcg, nid);
2852 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2853
2854 /* where the current iteration continues after */
2855 if (mm_state->head == &mm->lru_gen.list)
2856 mm_state->head = mm_state->head->prev;
2857
2858 /* where the last iteration ended before */
2859 if (mm_state->tail == &mm->lru_gen.list)
2860 mm_state->tail = mm_state->tail->next;
2861 }
2862
2863 list_del_init(&mm->lru_gen.list);
2864
2865 spin_unlock(&mm_list->lock);
2866
2867 #ifdef CONFIG_MEMCG
2868 mem_cgroup_put(mm->lru_gen.memcg);
2869 mm->lru_gen.memcg = NULL;
2870 #endif
2871 }
2872
2873 #ifdef CONFIG_MEMCG
2874 void lru_gen_migrate_mm(struct mm_struct *mm)
2875 {
2876 struct mem_cgroup *memcg;
2877 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2878
2879 VM_WARN_ON_ONCE(task->mm != mm);
2880 lockdep_assert_held(&task->alloc_lock);
2881
2882 /* for mm_update_next_owner() */
2883 if (mem_cgroup_disabled())
2884 return;
2885
2886 /* migration can happen before addition */
2887 if (!mm->lru_gen.memcg)
2888 return;
2889
2890 rcu_read_lock();
2891 memcg = mem_cgroup_from_task(task);
2892 rcu_read_unlock();
2893 if (memcg == mm->lru_gen.memcg)
2894 return;
2895
2896 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2897
2898 lru_gen_del_mm(mm);
2899 lru_gen_add_mm(mm);
2900 }
2901 #endif
2902
2903 #else /* !CONFIG_LRU_GEN_WALKS_MMU */
2904
2905 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2906 {
2907 return NULL;
2908 }
2909
2910 static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2911 {
2912 return NULL;
2913 }
2914
2915 static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2916 {
2917 return NULL;
2918 }
2919
2920 #endif
2921
2922 static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
2923 {
2924 int i;
2925 int hist;
2926 struct lruvec *lruvec = walk->lruvec;
2927 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2928
2929 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2930
2931 hist = lru_hist_from_seq(walk->seq);
2932
2933 for (i = 0; i < NR_MM_STATS; i++) {
2934 WRITE_ONCE(mm_state->stats[hist][i],
2935 mm_state->stats[hist][i] + walk->mm_stats[i]);
2936 walk->mm_stats[i] = 0;
2937 }
2938
2939 if (NR_HIST_GENS > 1 && last) {
2940 hist = lru_hist_from_seq(walk->seq + 1);
2941
2942 for (i = 0; i < NR_MM_STATS; i++)
2943 WRITE_ONCE(mm_state->stats[hist][i], 0);
2944 }
2945 }
2946
2947 static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
2948 {
2949 bool first = false;
2950 bool last = false;
2951 struct mm_struct *mm = NULL;
2952 struct lruvec *lruvec = walk->lruvec;
2953 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2954 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2955 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2956
2957 /*
2958 * mm_state->seq is incremented after each iteration of mm_list. There
2959 * are three interesting cases for this page table walker:
2960 * 1. It tries to start a new iteration with a stale max_seq: there is
2961 * nothing left to do.
2962 * 2. It started the next iteration: it needs to reset the Bloom filter
2963 * so that a fresh set of PTE tables can be recorded.
2964 * 3. It ended the current iteration: it needs to reset the mm stats
2965 * counters and tell its caller to increment max_seq.
2966 */
2967 spin_lock(&mm_list->lock);
2968
2969 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
2970
2971 if (walk->seq <= mm_state->seq)
2972 goto done;
2973
2974 if (!mm_state->head)
2975 mm_state->head = &mm_list->fifo;
2976
2977 if (mm_state->head == &mm_list->fifo)
2978 first = true;
2979
2980 do {
2981 mm_state->head = mm_state->head->next;
2982 if (mm_state->head == &mm_list->fifo) {
2983 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2984 last = true;
2985 break;
2986 }
2987
2988 /* force scan for those added after the last iteration */
2989 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2990 mm_state->tail = mm_state->head->next;
2991 walk->force_scan = true;
2992 }
2993 } while (!(mm = get_next_mm(walk)));
2994 done:
2995 if (*iter || last)
2996 reset_mm_stats(walk, last);
2997
2998 spin_unlock(&mm_list->lock);
2999
3000 if (mm && first)
3001 reset_bloom_filter(mm_state, walk->seq + 1);
3002
3003 if (*iter)
3004 mmput_async(*iter);
3005
3006 *iter = mm;
3007
3008 return last;
3009 }
3010
3011 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
3012 {
3013 bool success = false;
3014 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3015 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3016 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3017
3018 spin_lock(&mm_list->lock);
3019
3020 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
3021
3022 if (seq > mm_state->seq) {
3023 mm_state->head = NULL;
3024 mm_state->tail = NULL;
3025 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3026 success = true;
3027 }
3028
3029 spin_unlock(&mm_list->lock);
3030
3031 return success;
3032 }
3033
3034 /******************************************************************************
3035 * PID controller
3036 ******************************************************************************/
3037
3038 /*
3039 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3040 *
3041 * The P term is refaulted/(evicted+protected) from a tier in the generation
3042 * currently being evicted; the I term is the exponential moving average of the
3043 * P term over the generations previously evicted, using the smoothing factor
3044 * 1/2; the D term isn't supported.
3045 *
3046 * The setpoint (SP) is always the first tier of one type; the process variable
3047 * (PV) is either any tier of the other type or any other tier of the same
3048 * type.
3049 *
3050 * The error is the difference between the SP and the PV; the correction is to
3051 * turn off protection when SP>PV or turn on protection when SP<PV.
3052 *
3053 * For future optimizations:
3054 * 1. The D term may discount the other two terms over time so that long-lived
3055 * generations can resist stale information.
3056 */
3057 struct ctrl_pos {
3058 unsigned long refaulted;
3059 unsigned long total;
3060 int gain;
3061 };
3062
3063 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3064 struct ctrl_pos *pos)
3065 {
3066 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3067 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3068
3069 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3070 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3071 pos->total = lrugen->avg_total[type][tier] +
3072 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3073 if (tier)
3074 pos->total += lrugen->protected[hist][type][tier - 1];
3075 pos->gain = gain;
3076 }
3077
3078 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3079 {
3080 int hist, tier;
3081 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3082 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3083 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3084
3085 lockdep_assert_held(&lruvec->lru_lock);
3086
3087 if (!carryover && !clear)
3088 return;
3089
3090 hist = lru_hist_from_seq(seq);
3091
3092 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3093 if (carryover) {
3094 unsigned long sum;
3095
3096 sum = lrugen->avg_refaulted[type][tier] +
3097 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3098 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3099
3100 sum = lrugen->avg_total[type][tier] +
3101 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3102 if (tier)
3103 sum += lrugen->protected[hist][type][tier - 1];
3104 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3105 }
3106
3107 if (clear) {
3108 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3109 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3110 if (tier)
3111 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3112 }
3113 }
3114 }
3115
3116 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3117 {
3118 /*
3119 * Return true if the PV has a limited number of refaults or a lower
3120 * refaulted/total than the SP.
3121 */
3122 return pv->refaulted < MIN_LRU_BATCH ||
3123 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3124 (sp->refaulted + 1) * pv->total * pv->gain;
3125 }
3126
3127 /******************************************************************************
3128 * the aging
3129 ******************************************************************************/
3130
3131 /* promote pages accessed through page tables */
3132 static int folio_update_gen(struct folio *folio, int gen)
3133 {
3134 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3135
3136 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3137 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3138
3139 do {
3140 /* lru_gen_del_folio() has isolated this page? */
3141 if (!(old_flags & LRU_GEN_MASK)) {
3142 /* for shrink_folio_list() */
3143 new_flags = old_flags | BIT(PG_referenced);
3144 continue;
3145 }
3146
3147 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3148 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3149 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3150
3151 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3152 }
3153
3154 /* protect pages accessed multiple times through file descriptors */
3155 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3156 {
3157 int type = folio_is_file_lru(folio);
3158 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3159 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3160 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3161
3162 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3163
3164 do {
3165 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3166 /* folio_update_gen() has promoted this page? */
3167 if (new_gen >= 0 && new_gen != old_gen)
3168 return new_gen;
3169
3170 new_gen = (old_gen + 1) % MAX_NR_GENS;
3171
3172 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3173 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3174 /* for folio_end_writeback() */
3175 if (reclaiming)
3176 new_flags |= BIT(PG_reclaim);
3177 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3178
3179 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3180
3181 return new_gen;
3182 }
3183
3184 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3185 int old_gen, int new_gen)
3186 {
3187 int type = folio_is_file_lru(folio);
3188 int zone = folio_zonenum(folio);
3189 int delta = folio_nr_pages(folio);
3190
3191 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3192 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3193
3194 walk->batched++;
3195
3196 walk->nr_pages[old_gen][type][zone] -= delta;
3197 walk->nr_pages[new_gen][type][zone] += delta;
3198 }
3199
3200 static void reset_batch_size(struct lru_gen_mm_walk *walk)
3201 {
3202 int gen, type, zone;
3203 struct lruvec *lruvec = walk->lruvec;
3204 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3205
3206 walk->batched = 0;
3207
3208 for_each_gen_type_zone(gen, type, zone) {
3209 enum lru_list lru = type * LRU_INACTIVE_FILE;
3210 int delta = walk->nr_pages[gen][type][zone];
3211
3212 if (!delta)
3213 continue;
3214
3215 walk->nr_pages[gen][type][zone] = 0;
3216 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3217 lrugen->nr_pages[gen][type][zone] + delta);
3218
3219 if (lru_gen_is_active(lruvec, gen))
3220 lru += LRU_ACTIVE;
3221 __update_lru_size(lruvec, lru, zone, delta);
3222 }
3223 }
3224
3225 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3226 {
3227 struct address_space *mapping;
3228 struct vm_area_struct *vma = args->vma;
3229 struct lru_gen_mm_walk *walk = args->private;
3230
3231 if (!vma_is_accessible(vma))
3232 return true;
3233
3234 if (is_vm_hugetlb_page(vma))
3235 return true;
3236
3237 if (!vma_has_recency(vma))
3238 return true;
3239
3240 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3241 return true;
3242
3243 if (vma == get_gate_vma(vma->vm_mm))
3244 return true;
3245
3246 if (vma_is_anonymous(vma))
3247 return !walk->can_swap;
3248
3249 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3250 return true;
3251
3252 mapping = vma->vm_file->f_mapping;
3253 if (mapping_unevictable(mapping))
3254 return true;
3255
3256 if (shmem_mapping(mapping))
3257 return !walk->can_swap;
3258
3259 /* to exclude special mappings like dax, etc. */
3260 return !mapping->a_ops->read_folio;
3261 }
3262
3263 /*
3264 * Some userspace memory allocators map many single-page VMAs. Instead of
3265 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3266 * table to reduce zigzags and improve cache performance.
3267 */
3268 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3269 unsigned long *vm_start, unsigned long *vm_end)
3270 {
3271 unsigned long start = round_up(*vm_end, size);
3272 unsigned long end = (start | ~mask) + 1;
3273 VMA_ITERATOR(vmi, args->mm, start);
3274
3275 VM_WARN_ON_ONCE(mask & size);
3276 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3277
3278 for_each_vma(vmi, args->vma) {
3279 if (end && end <= args->vma->vm_start)
3280 return false;
3281
3282 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3283 continue;
3284
3285 *vm_start = max(start, args->vma->vm_start);
3286 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3287
3288 return true;
3289 }
3290
3291 return false;
3292 }
3293
3294 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3295 {
3296 unsigned long pfn = pte_pfn(pte);
3297
3298 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3299
3300 if (!pte_present(pte) || is_zero_pfn(pfn))
3301 return -1;
3302
3303 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3304 return -1;
3305
3306 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3307 return -1;
3308
3309 return pfn;
3310 }
3311
3312 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3313 {
3314 unsigned long pfn = pmd_pfn(pmd);
3315
3316 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3317
3318 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3319 return -1;
3320
3321 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3322 return -1;
3323
3324 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3325 return -1;
3326
3327 return pfn;
3328 }
3329
3330 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3331 struct pglist_data *pgdat, bool can_swap)
3332 {
3333 struct folio *folio;
3334
3335 /* try to avoid unnecessary memory loads */
3336 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3337 return NULL;
3338
3339 folio = pfn_folio(pfn);
3340 if (folio_nid(folio) != pgdat->node_id)
3341 return NULL;
3342
3343 if (folio_memcg_rcu(folio) != memcg)
3344 return NULL;
3345
3346 /* file VMAs can contain anon pages from COW */
3347 if (!folio_is_file_lru(folio) && !can_swap)
3348 return NULL;
3349
3350 return folio;
3351 }
3352
3353 static bool suitable_to_scan(int total, int young)
3354 {
3355 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3356
3357 /* suitable if the average number of young PTEs per cacheline is >=1 */
3358 return young * n >= total;
3359 }
3360
3361 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3362 struct mm_walk *args)
3363 {
3364 int i;
3365 pte_t *pte;
3366 spinlock_t *ptl;
3367 unsigned long addr;
3368 int total = 0;
3369 int young = 0;
3370 struct lru_gen_mm_walk *walk = args->private;
3371 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3372 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3373 DEFINE_MAX_SEQ(walk->lruvec);
3374 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3375
3376 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3377 if (!pte)
3378 return false;
3379 if (!spin_trylock(ptl)) {
3380 pte_unmap(pte);
3381 return false;
3382 }
3383
3384 arch_enter_lazy_mmu_mode();
3385 restart:
3386 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3387 unsigned long pfn;
3388 struct folio *folio;
3389 pte_t ptent = ptep_get(pte + i);
3390
3391 total++;
3392 walk->mm_stats[MM_LEAF_TOTAL]++;
3393
3394 pfn = get_pte_pfn(ptent, args->vma, addr);
3395 if (pfn == -1)
3396 continue;
3397
3398 if (!pte_young(ptent)) {
3399 walk->mm_stats[MM_LEAF_OLD]++;
3400 continue;
3401 }
3402
3403 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3404 if (!folio)
3405 continue;
3406
3407 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3408 VM_WARN_ON_ONCE(true);
3409
3410 young++;
3411 walk->mm_stats[MM_LEAF_YOUNG]++;
3412
3413 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3414 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3415 !folio_test_swapcache(folio)))
3416 folio_mark_dirty(folio);
3417
3418 old_gen = folio_update_gen(folio, new_gen);
3419 if (old_gen >= 0 && old_gen != new_gen)
3420 update_batch_size(walk, folio, old_gen, new_gen);
3421 }
3422
3423 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3424 goto restart;
3425
3426 arch_leave_lazy_mmu_mode();
3427 pte_unmap_unlock(pte, ptl);
3428
3429 return suitable_to_scan(total, young);
3430 }
3431
3432 static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3433 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3434 {
3435 int i;
3436 pmd_t *pmd;
3437 spinlock_t *ptl;
3438 struct lru_gen_mm_walk *walk = args->private;
3439 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3440 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3441 DEFINE_MAX_SEQ(walk->lruvec);
3442 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3443
3444 VM_WARN_ON_ONCE(pud_leaf(*pud));
3445
3446 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3447 if (*first == -1) {
3448 *first = addr;
3449 bitmap_zero(bitmap, MIN_LRU_BATCH);
3450 return;
3451 }
3452
3453 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3454 if (i && i <= MIN_LRU_BATCH) {
3455 __set_bit(i - 1, bitmap);
3456 return;
3457 }
3458
3459 pmd = pmd_offset(pud, *first);
3460
3461 ptl = pmd_lockptr(args->mm, pmd);
3462 if (!spin_trylock(ptl))
3463 goto done;
3464
3465 arch_enter_lazy_mmu_mode();
3466
3467 do {
3468 unsigned long pfn;
3469 struct folio *folio;
3470
3471 /* don't round down the first address */
3472 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3473
3474 pfn = get_pmd_pfn(pmd[i], vma, addr);
3475 if (pfn == -1)
3476 goto next;
3477
3478 if (!pmd_trans_huge(pmd[i])) {
3479 if (should_clear_pmd_young())
3480 pmdp_test_and_clear_young(vma, addr, pmd + i);
3481 goto next;
3482 }
3483
3484 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3485 if (!folio)
3486 goto next;
3487
3488 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3489 goto next;
3490
3491 walk->mm_stats[MM_LEAF_YOUNG]++;
3492
3493 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3494 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3495 !folio_test_swapcache(folio)))
3496 folio_mark_dirty(folio);
3497
3498 old_gen = folio_update_gen(folio, new_gen);
3499 if (old_gen >= 0 && old_gen != new_gen)
3500 update_batch_size(walk, folio, old_gen, new_gen);
3501 next:
3502 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3503 } while (i <= MIN_LRU_BATCH);
3504
3505 arch_leave_lazy_mmu_mode();
3506 spin_unlock(ptl);
3507 done:
3508 *first = -1;
3509 }
3510
3511 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3512 struct mm_walk *args)
3513 {
3514 int i;
3515 pmd_t *pmd;
3516 unsigned long next;
3517 unsigned long addr;
3518 struct vm_area_struct *vma;
3519 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3520 unsigned long first = -1;
3521 struct lru_gen_mm_walk *walk = args->private;
3522 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3523
3524 VM_WARN_ON_ONCE(pud_leaf(*pud));
3525
3526 /*
3527 * Finish an entire PMD in two passes: the first only reaches to PTE
3528 * tables to avoid taking the PMD lock; the second, if necessary, takes
3529 * the PMD lock to clear the accessed bit in PMD entries.
3530 */
3531 pmd = pmd_offset(pud, start & PUD_MASK);
3532 restart:
3533 /* walk_pte_range() may call get_next_vma() */
3534 vma = args->vma;
3535 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3536 pmd_t val = pmdp_get_lockless(pmd + i);
3537
3538 next = pmd_addr_end(addr, end);
3539
3540 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3541 walk->mm_stats[MM_LEAF_TOTAL]++;
3542 continue;
3543 }
3544
3545 if (pmd_trans_huge(val)) {
3546 unsigned long pfn = pmd_pfn(val);
3547 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3548
3549 walk->mm_stats[MM_LEAF_TOTAL]++;
3550
3551 if (!pmd_young(val)) {
3552 walk->mm_stats[MM_LEAF_OLD]++;
3553 continue;
3554 }
3555
3556 /* try to avoid unnecessary memory loads */
3557 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3558 continue;
3559
3560 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3561 continue;
3562 }
3563
3564 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3565
3566 if (should_clear_pmd_young()) {
3567 if (!pmd_young(val))
3568 continue;
3569
3570 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3571 }
3572
3573 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3574 continue;
3575
3576 walk->mm_stats[MM_NONLEAF_FOUND]++;
3577
3578 if (!walk_pte_range(&val, addr, next, args))
3579 continue;
3580
3581 walk->mm_stats[MM_NONLEAF_ADDED]++;
3582
3583 /* carry over to the next generation */
3584 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3585 }
3586
3587 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3588
3589 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3590 goto restart;
3591 }
3592
3593 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3594 struct mm_walk *args)
3595 {
3596 int i;
3597 pud_t *pud;
3598 unsigned long addr;
3599 unsigned long next;
3600 struct lru_gen_mm_walk *walk = args->private;
3601
3602 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3603
3604 pud = pud_offset(p4d, start & P4D_MASK);
3605 restart:
3606 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3607 pud_t val = READ_ONCE(pud[i]);
3608
3609 next = pud_addr_end(addr, end);
3610
3611 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3612 continue;
3613
3614 walk_pmd_range(&val, addr, next, args);
3615
3616 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3617 end = (addr | ~PUD_MASK) + 1;
3618 goto done;
3619 }
3620 }
3621
3622 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3623 goto restart;
3624
3625 end = round_up(end, P4D_SIZE);
3626 done:
3627 if (!end || !args->vma)
3628 return 1;
3629
3630 walk->next_addr = max(end, args->vma->vm_start);
3631
3632 return -EAGAIN;
3633 }
3634
3635 static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3636 {
3637 static const struct mm_walk_ops mm_walk_ops = {
3638 .test_walk = should_skip_vma,
3639 .p4d_entry = walk_pud_range,
3640 .walk_lock = PGWALK_RDLOCK,
3641 };
3642
3643 int err;
3644 struct lruvec *lruvec = walk->lruvec;
3645 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3646
3647 walk->next_addr = FIRST_USER_ADDRESS;
3648
3649 do {
3650 DEFINE_MAX_SEQ(lruvec);
3651
3652 err = -EBUSY;
3653
3654 /* another thread might have called inc_max_seq() */
3655 if (walk->seq != max_seq)
3656 break;
3657
3658 /* folio_update_gen() requires stable folio_memcg() */
3659 if (!mem_cgroup_trylock_pages(memcg))
3660 break;
3661
3662 /* the caller might be holding the lock for write */
3663 if (mmap_read_trylock(mm)) {
3664 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3665
3666 mmap_read_unlock(mm);
3667 }
3668
3669 mem_cgroup_unlock_pages();
3670
3671 if (walk->batched) {
3672 spin_lock_irq(&lruvec->lru_lock);
3673 reset_batch_size(walk);
3674 spin_unlock_irq(&lruvec->lru_lock);
3675 }
3676
3677 cond_resched();
3678 } while (err == -EAGAIN);
3679 }
3680
3681 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3682 {
3683 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3684
3685 if (pgdat && current_is_kswapd()) {
3686 VM_WARN_ON_ONCE(walk);
3687
3688 walk = &pgdat->mm_walk;
3689 } else if (!walk && force_alloc) {
3690 VM_WARN_ON_ONCE(current_is_kswapd());
3691
3692 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3693 }
3694
3695 current->reclaim_state->mm_walk = walk;
3696
3697 return walk;
3698 }
3699
3700 static void clear_mm_walk(void)
3701 {
3702 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3703
3704 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3705 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3706
3707 current->reclaim_state->mm_walk = NULL;
3708
3709 if (!current_is_kswapd())
3710 kfree(walk);
3711 }
3712
3713 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3714 {
3715 int zone;
3716 int remaining = MAX_LRU_BATCH;
3717 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3718 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3719
3720 if (type == LRU_GEN_ANON && !can_swap)
3721 goto done;
3722
3723 /* prevent cold/hot inversion if force_scan is true */
3724 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3725 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3726
3727 while (!list_empty(head)) {
3728 struct folio *folio = lru_to_folio(head);
3729
3730 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3731 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3732 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3733 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3734
3735 new_gen = folio_inc_gen(lruvec, folio, false);
3736 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3737
3738 if (!--remaining)
3739 return false;
3740 }
3741 }
3742 done:
3743 reset_ctrl_pos(lruvec, type, true);
3744 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3745
3746 return true;
3747 }
3748
3749 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3750 {
3751 int gen, type, zone;
3752 bool success = false;
3753 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3754 DEFINE_MIN_SEQ(lruvec);
3755
3756 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3757
3758 /* find the oldest populated generation */
3759 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3760 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3761 gen = lru_gen_from_seq(min_seq[type]);
3762
3763 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3764 if (!list_empty(&lrugen->folios[gen][type][zone]))
3765 goto next;
3766 }
3767
3768 min_seq[type]++;
3769 }
3770 next:
3771 ;
3772 }
3773
3774 /* see the comment on lru_gen_folio */
3775 if (can_swap) {
3776 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3777 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3778 }
3779
3780 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3781 if (min_seq[type] == lrugen->min_seq[type])
3782 continue;
3783
3784 reset_ctrl_pos(lruvec, type, true);
3785 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3786 success = true;
3787 }
3788
3789 return success;
3790 }
3791
3792 static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3793 bool can_swap, bool force_scan)
3794 {
3795 bool success;
3796 int prev, next;
3797 int type, zone;
3798 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3799 restart:
3800 if (seq < READ_ONCE(lrugen->max_seq))
3801 return false;
3802
3803 spin_lock_irq(&lruvec->lru_lock);
3804
3805 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3806
3807 success = seq == lrugen->max_seq;
3808 if (!success)
3809 goto unlock;
3810
3811 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3812 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3813 continue;
3814
3815 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3816
3817 if (inc_min_seq(lruvec, type, can_swap))
3818 continue;
3819
3820 spin_unlock_irq(&lruvec->lru_lock);
3821 cond_resched();
3822 goto restart;
3823 }
3824
3825 /*
3826 * Update the active/inactive LRU sizes for compatibility. Both sides of
3827 * the current max_seq need to be covered, since max_seq+1 can overlap
3828 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3829 * overlap, cold/hot inversion happens.
3830 */
3831 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3832 next = lru_gen_from_seq(lrugen->max_seq + 1);
3833
3834 for (type = 0; type < ANON_AND_FILE; type++) {
3835 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3836 enum lru_list lru = type * LRU_INACTIVE_FILE;
3837 long delta = lrugen->nr_pages[prev][type][zone] -
3838 lrugen->nr_pages[next][type][zone];
3839
3840 if (!delta)
3841 continue;
3842
3843 __update_lru_size(lruvec, lru, zone, delta);
3844 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3845 }
3846 }
3847
3848 for (type = 0; type < ANON_AND_FILE; type++)
3849 reset_ctrl_pos(lruvec, type, false);
3850
3851 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3852 /* make sure preceding modifications appear */
3853 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3854 unlock:
3855 spin_unlock_irq(&lruvec->lru_lock);
3856
3857 return success;
3858 }
3859
3860 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3861 bool can_swap, bool force_scan)
3862 {
3863 bool success;
3864 struct lru_gen_mm_walk *walk;
3865 struct mm_struct *mm = NULL;
3866 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3867 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3868
3869 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
3870
3871 if (!mm_state)
3872 return inc_max_seq(lruvec, seq, can_swap, force_scan);
3873
3874 /* see the comment in iterate_mm_list() */
3875 if (seq <= READ_ONCE(mm_state->seq))
3876 return false;
3877
3878 /*
3879 * If the hardware doesn't automatically set the accessed bit, fallback
3880 * to lru_gen_look_around(), which only clears the accessed bit in a
3881 * handful of PTEs. Spreading the work out over a period of time usually
3882 * is less efficient, but it avoids bursty page faults.
3883 */
3884 if (!should_walk_mmu()) {
3885 success = iterate_mm_list_nowalk(lruvec, seq);
3886 goto done;
3887 }
3888
3889 walk = set_mm_walk(NULL, true);
3890 if (!walk) {
3891 success = iterate_mm_list_nowalk(lruvec, seq);
3892 goto done;
3893 }
3894
3895 walk->lruvec = lruvec;
3896 walk->seq = seq;
3897 walk->can_swap = can_swap;
3898 walk->force_scan = force_scan;
3899
3900 do {
3901 success = iterate_mm_list(walk, &mm);
3902 if (mm)
3903 walk_mm(mm, walk);
3904 } while (mm);
3905 done:
3906 if (success) {
3907 success = inc_max_seq(lruvec, seq, can_swap, force_scan);
3908 WARN_ON_ONCE(!success);
3909 }
3910
3911 return success;
3912 }
3913
3914 /******************************************************************************
3915 * working set protection
3916 ******************************************************************************/
3917
3918 static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
3919 {
3920 int priority;
3921 unsigned long reclaimable;
3922
3923 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
3924 return;
3925 /*
3926 * Determine the initial priority based on
3927 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
3928 * where reclaimed_to_scanned_ratio = inactive / total.
3929 */
3930 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
3931 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
3932 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
3933
3934 /* round down reclaimable and round up sc->nr_to_reclaim */
3935 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
3936
3937 /*
3938 * The estimation is based on LRU pages only, so cap it to prevent
3939 * overshoots of shrinker objects by large margins.
3940 */
3941 sc->priority = clamp(priority, DEF_PRIORITY / 2, DEF_PRIORITY);
3942 }
3943
3944 static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3945 {
3946 int gen, type, zone;
3947 unsigned long total = 0;
3948 bool can_swap = get_swappiness(lruvec, sc);
3949 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3950 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3951 DEFINE_MAX_SEQ(lruvec);
3952 DEFINE_MIN_SEQ(lruvec);
3953
3954 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3955 unsigned long seq;
3956
3957 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3958 gen = lru_gen_from_seq(seq);
3959
3960 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3961 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3962 }
3963 }
3964
3965 /* whether the size is big enough to be helpful */
3966 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3967 }
3968
3969 static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3970 unsigned long min_ttl)
3971 {
3972 int gen;
3973 unsigned long birth;
3974 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3975 DEFINE_MIN_SEQ(lruvec);
3976
3977 if (mem_cgroup_below_min(NULL, memcg))
3978 return false;
3979
3980 if (!lruvec_is_sizable(lruvec, sc))
3981 return false;
3982
3983 /* see the comment on lru_gen_folio */
3984 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3985 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3986
3987 return time_is_before_jiffies(birth + min_ttl);
3988 }
3989
3990 /* to protect the working set of the last N jiffies */
3991 static unsigned long lru_gen_min_ttl __read_mostly;
3992
3993 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3994 {
3995 struct mem_cgroup *memcg;
3996 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3997 bool reclaimable = !min_ttl;
3998
3999 VM_WARN_ON_ONCE(!current_is_kswapd());
4000
4001 set_initial_priority(pgdat, sc);
4002
4003 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4004 do {
4005 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4006
4007 mem_cgroup_calculate_protection(NULL, memcg);
4008
4009 if (!reclaimable)
4010 reclaimable = lruvec_is_reclaimable(lruvec, sc, min_ttl);
4011 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4012
4013 /*
4014 * The main goal is to OOM kill if every generation from all memcgs is
4015 * younger than min_ttl. However, another possibility is all memcgs are
4016 * either too small or below min.
4017 */
4018 if (!reclaimable && mutex_trylock(&oom_lock)) {
4019 struct oom_control oc = {
4020 .gfp_mask = sc->gfp_mask,
4021 };
4022
4023 out_of_memory(&oc);
4024
4025 mutex_unlock(&oom_lock);
4026 }
4027 }
4028
4029 /******************************************************************************
4030 * rmap/PT walk feedback
4031 ******************************************************************************/
4032
4033 /*
4034 * This function exploits spatial locality when shrink_folio_list() walks the
4035 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4036 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4037 * the PTE table to the Bloom filter. This forms a feedback loop between the
4038 * eviction and the aging.
4039 */
4040 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4041 {
4042 int i;
4043 unsigned long start;
4044 unsigned long end;
4045 struct lru_gen_mm_walk *walk;
4046 int young = 0;
4047 pte_t *pte = pvmw->pte;
4048 unsigned long addr = pvmw->address;
4049 struct vm_area_struct *vma = pvmw->vma;
4050 struct folio *folio = pfn_folio(pvmw->pfn);
4051 bool can_swap = !folio_is_file_lru(folio);
4052 struct mem_cgroup *memcg = folio_memcg(folio);
4053 struct pglist_data *pgdat = folio_pgdat(folio);
4054 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4055 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4056 DEFINE_MAX_SEQ(lruvec);
4057 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4058
4059 lockdep_assert_held(pvmw->ptl);
4060 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4061
4062 if (spin_is_contended(pvmw->ptl))
4063 return;
4064
4065 /* exclude special VMAs containing anon pages from COW */
4066 if (vma->vm_flags & VM_SPECIAL)
4067 return;
4068
4069 /* avoid taking the LRU lock under the PTL when possible */
4070 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4071
4072 start = max(addr & PMD_MASK, vma->vm_start);
4073 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4074
4075 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4076 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4077 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4078 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4079 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4080 else {
4081 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4082 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4083 }
4084 }
4085
4086 /* folio_update_gen() requires stable folio_memcg() */
4087 if (!mem_cgroup_trylock_pages(memcg))
4088 return;
4089
4090 arch_enter_lazy_mmu_mode();
4091
4092 pte -= (addr - start) / PAGE_SIZE;
4093
4094 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4095 unsigned long pfn;
4096 pte_t ptent = ptep_get(pte + i);
4097
4098 pfn = get_pte_pfn(ptent, vma, addr);
4099 if (pfn == -1)
4100 continue;
4101
4102 if (!pte_young(ptent))
4103 continue;
4104
4105 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4106 if (!folio)
4107 continue;
4108
4109 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4110 VM_WARN_ON_ONCE(true);
4111
4112 young++;
4113
4114 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4115 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4116 !folio_test_swapcache(folio)))
4117 folio_mark_dirty(folio);
4118
4119 if (walk) {
4120 old_gen = folio_update_gen(folio, new_gen);
4121 if (old_gen >= 0 && old_gen != new_gen)
4122 update_batch_size(walk, folio, old_gen, new_gen);
4123
4124 continue;
4125 }
4126
4127 old_gen = folio_lru_gen(folio);
4128 if (old_gen < 0)
4129 folio_set_referenced(folio);
4130 else if (old_gen != new_gen)
4131 folio_activate(folio);
4132 }
4133
4134 arch_leave_lazy_mmu_mode();
4135 mem_cgroup_unlock_pages();
4136
4137 /* feedback from rmap walkers to page table walkers */
4138 if (mm_state && suitable_to_scan(i, young))
4139 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4140 }
4141
4142 /******************************************************************************
4143 * memcg LRU
4144 ******************************************************************************/
4145
4146 /* see the comment on MEMCG_NR_GENS */
4147 enum {
4148 MEMCG_LRU_NOP,
4149 MEMCG_LRU_HEAD,
4150 MEMCG_LRU_TAIL,
4151 MEMCG_LRU_OLD,
4152 MEMCG_LRU_YOUNG,
4153 };
4154
4155 static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4156 {
4157 int seg;
4158 int old, new;
4159 unsigned long flags;
4160 int bin = get_random_u32_below(MEMCG_NR_BINS);
4161 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4162
4163 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4164
4165 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4166
4167 seg = 0;
4168 new = old = lruvec->lrugen.gen;
4169
4170 /* see the comment on MEMCG_NR_GENS */
4171 if (op == MEMCG_LRU_HEAD)
4172 seg = MEMCG_LRU_HEAD;
4173 else if (op == MEMCG_LRU_TAIL)
4174 seg = MEMCG_LRU_TAIL;
4175 else if (op == MEMCG_LRU_OLD)
4176 new = get_memcg_gen(pgdat->memcg_lru.seq);
4177 else if (op == MEMCG_LRU_YOUNG)
4178 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4179 else
4180 VM_WARN_ON_ONCE(true);
4181
4182 WRITE_ONCE(lruvec->lrugen.seg, seg);
4183 WRITE_ONCE(lruvec->lrugen.gen, new);
4184
4185 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4186
4187 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4188 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4189 else
4190 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4191
4192 pgdat->memcg_lru.nr_memcgs[old]--;
4193 pgdat->memcg_lru.nr_memcgs[new]++;
4194
4195 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4196 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4197
4198 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4199 }
4200
4201 #ifdef CONFIG_MEMCG
4202
4203 void lru_gen_online_memcg(struct mem_cgroup *memcg)
4204 {
4205 int gen;
4206 int nid;
4207 int bin = get_random_u32_below(MEMCG_NR_BINS);
4208
4209 for_each_node(nid) {
4210 struct pglist_data *pgdat = NODE_DATA(nid);
4211 struct lruvec *lruvec = get_lruvec(memcg, nid);
4212
4213 spin_lock_irq(&pgdat->memcg_lru.lock);
4214
4215 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4216
4217 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4218
4219 lruvec->lrugen.gen = gen;
4220
4221 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4222 pgdat->memcg_lru.nr_memcgs[gen]++;
4223
4224 spin_unlock_irq(&pgdat->memcg_lru.lock);
4225 }
4226 }
4227
4228 void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4229 {
4230 int nid;
4231
4232 for_each_node(nid) {
4233 struct lruvec *lruvec = get_lruvec(memcg, nid);
4234
4235 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4236 }
4237 }
4238
4239 void lru_gen_release_memcg(struct mem_cgroup *memcg)
4240 {
4241 int gen;
4242 int nid;
4243
4244 for_each_node(nid) {
4245 struct pglist_data *pgdat = NODE_DATA(nid);
4246 struct lruvec *lruvec = get_lruvec(memcg, nid);
4247
4248 spin_lock_irq(&pgdat->memcg_lru.lock);
4249
4250 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4251 goto unlock;
4252
4253 gen = lruvec->lrugen.gen;
4254
4255 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4256 pgdat->memcg_lru.nr_memcgs[gen]--;
4257
4258 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4259 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4260 unlock:
4261 spin_unlock_irq(&pgdat->memcg_lru.lock);
4262 }
4263 }
4264
4265 void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4266 {
4267 struct lruvec *lruvec = get_lruvec(memcg, nid);
4268
4269 /* see the comment on MEMCG_NR_GENS */
4270 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4271 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4272 }
4273
4274 #endif /* CONFIG_MEMCG */
4275
4276 /******************************************************************************
4277 * the eviction
4278 ******************************************************************************/
4279
4280 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4281 int tier_idx)
4282 {
4283 bool success;
4284 int gen = folio_lru_gen(folio);
4285 int type = folio_is_file_lru(folio);
4286 int zone = folio_zonenum(folio);
4287 int delta = folio_nr_pages(folio);
4288 int refs = folio_lru_refs(folio);
4289 int tier = lru_tier_from_refs(refs);
4290 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4291
4292 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4293
4294 /* unevictable */
4295 if (!folio_evictable(folio)) {
4296 success = lru_gen_del_folio(lruvec, folio, true);
4297 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4298 folio_set_unevictable(folio);
4299 lruvec_add_folio(lruvec, folio);
4300 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4301 return true;
4302 }
4303
4304 /* promoted */
4305 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4306 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4307 return true;
4308 }
4309
4310 /* protected */
4311 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4312 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4313
4314 gen = folio_inc_gen(lruvec, folio, false);
4315 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4316
4317 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4318 lrugen->protected[hist][type][tier - 1] + delta);
4319 return true;
4320 }
4321
4322 /* ineligible */
4323 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4324 gen = folio_inc_gen(lruvec, folio, false);
4325 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4326 return true;
4327 }
4328
4329 /* waiting for writeback */
4330 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4331 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4332 gen = folio_inc_gen(lruvec, folio, true);
4333 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4334 return true;
4335 }
4336
4337 return false;
4338 }
4339
4340 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4341 {
4342 bool success;
4343
4344 /* swap constrained */
4345 if (!(sc->gfp_mask & __GFP_IO) &&
4346 (folio_test_dirty(folio) ||
4347 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4348 return false;
4349
4350 /* raced with release_pages() */
4351 if (!folio_try_get(folio))
4352 return false;
4353
4354 /* raced with another isolation */
4355 if (!folio_test_clear_lru(folio)) {
4356 folio_put(folio);
4357 return false;
4358 }
4359
4360 /* see the comment on MAX_NR_TIERS */
4361 if (!folio_test_referenced(folio))
4362 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4363
4364 /* for shrink_folio_list() */
4365 folio_clear_reclaim(folio);
4366 folio_clear_referenced(folio);
4367
4368 success = lru_gen_del_folio(lruvec, folio, true);
4369 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4370
4371 return true;
4372 }
4373
4374 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4375 int type, int tier, struct list_head *list)
4376 {
4377 int i;
4378 int gen;
4379 enum vm_event_item item;
4380 int sorted = 0;
4381 int scanned = 0;
4382 int isolated = 0;
4383 int skipped = 0;
4384 int remaining = MAX_LRU_BATCH;
4385 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4386 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4387
4388 VM_WARN_ON_ONCE(!list_empty(list));
4389
4390 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4391 return 0;
4392
4393 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4394
4395 for (i = MAX_NR_ZONES; i > 0; i--) {
4396 LIST_HEAD(moved);
4397 int skipped_zone = 0;
4398 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4399 struct list_head *head = &lrugen->folios[gen][type][zone];
4400
4401 while (!list_empty(head)) {
4402 struct folio *folio = lru_to_folio(head);
4403 int delta = folio_nr_pages(folio);
4404
4405 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4406 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4407 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4408 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4409
4410 scanned += delta;
4411
4412 if (sort_folio(lruvec, folio, sc, tier))
4413 sorted += delta;
4414 else if (isolate_folio(lruvec, folio, sc)) {
4415 list_add(&folio->lru, list);
4416 isolated += delta;
4417 } else {
4418 list_move(&folio->lru, &moved);
4419 skipped_zone += delta;
4420 }
4421
4422 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4423 break;
4424 }
4425
4426 if (skipped_zone) {
4427 list_splice(&moved, head);
4428 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4429 skipped += skipped_zone;
4430 }
4431
4432 if (!remaining || isolated >= MIN_LRU_BATCH)
4433 break;
4434 }
4435
4436 item = PGSCAN_KSWAPD + reclaimer_offset();
4437 if (!cgroup_reclaim(sc)) {
4438 __count_vm_events(item, isolated);
4439 __count_vm_events(PGREFILL, sorted);
4440 }
4441 __count_memcg_events(memcg, item, isolated);
4442 __count_memcg_events(memcg, PGREFILL, sorted);
4443 __count_vm_events(PGSCAN_ANON + type, isolated);
4444 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4445 scanned, skipped, isolated,
4446 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4447
4448 /*
4449 * There might not be eligible folios due to reclaim_idx. Check the
4450 * remaining to prevent livelock if it's not making progress.
4451 */
4452 return isolated || !remaining ? scanned : 0;
4453 }
4454
4455 static int get_tier_idx(struct lruvec *lruvec, int type)
4456 {
4457 int tier;
4458 struct ctrl_pos sp, pv;
4459
4460 /*
4461 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4462 * This value is chosen because any other tier would have at least twice
4463 * as many refaults as the first tier.
4464 */
4465 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4466 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4467 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4468 if (!positive_ctrl_err(&sp, &pv))
4469 break;
4470 }
4471
4472 return tier - 1;
4473 }
4474
4475 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4476 {
4477 int type, tier;
4478 struct ctrl_pos sp, pv;
4479 int gain[ANON_AND_FILE] = { swappiness, MAX_SWAPPINESS - swappiness };
4480
4481 /*
4482 * Compare the first tier of anon with that of file to determine which
4483 * type to scan. Also need to compare other tiers of the selected type
4484 * with the first tier of the other type to determine the last tier (of
4485 * the selected type) to evict.
4486 */
4487 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4488 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4489 type = positive_ctrl_err(&sp, &pv);
4490
4491 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4492 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4493 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4494 if (!positive_ctrl_err(&sp, &pv))
4495 break;
4496 }
4497
4498 *tier_idx = tier - 1;
4499
4500 return type;
4501 }
4502
4503 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4504 int *type_scanned, struct list_head *list)
4505 {
4506 int i;
4507 int type;
4508 int scanned;
4509 int tier = -1;
4510 DEFINE_MIN_SEQ(lruvec);
4511
4512 /*
4513 * Try to make the obvious choice first, and if anon and file are both
4514 * available from the same generation,
4515 * 1. Interpret swappiness 1 as file first and MAX_SWAPPINESS as anon
4516 * first.
4517 * 2. If !__GFP_IO, file first since clean pagecache is more likely to
4518 * exist than clean swapcache.
4519 */
4520 if (!swappiness)
4521 type = LRU_GEN_FILE;
4522 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4523 type = LRU_GEN_ANON;
4524 else if (swappiness == 1)
4525 type = LRU_GEN_FILE;
4526 else if (swappiness == MAX_SWAPPINESS)
4527 type = LRU_GEN_ANON;
4528 else if (!(sc->gfp_mask & __GFP_IO))
4529 type = LRU_GEN_FILE;
4530 else
4531 type = get_type_to_scan(lruvec, swappiness, &tier);
4532
4533 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4534 if (tier < 0)
4535 tier = get_tier_idx(lruvec, type);
4536
4537 scanned = scan_folios(lruvec, sc, type, tier, list);
4538 if (scanned)
4539 break;
4540
4541 type = !type;
4542 tier = -1;
4543 }
4544
4545 *type_scanned = type;
4546
4547 return scanned;
4548 }
4549
4550 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4551 {
4552 int type;
4553 int scanned;
4554 int reclaimed;
4555 LIST_HEAD(list);
4556 LIST_HEAD(clean);
4557 struct folio *folio;
4558 struct folio *next;
4559 enum vm_event_item item;
4560 struct reclaim_stat stat;
4561 struct lru_gen_mm_walk *walk;
4562 bool skip_retry = false;
4563 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4564 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4565
4566 spin_lock_irq(&lruvec->lru_lock);
4567
4568 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4569
4570 scanned += try_to_inc_min_seq(lruvec, swappiness);
4571
4572 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4573 scanned = 0;
4574
4575 spin_unlock_irq(&lruvec->lru_lock);
4576
4577 if (list_empty(&list))
4578 return scanned;
4579 retry:
4580 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4581 sc->nr_reclaimed += reclaimed;
4582 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4583 scanned, reclaimed, &stat, sc->priority,
4584 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4585
4586 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4587 if (!folio_evictable(folio)) {
4588 list_del(&folio->lru);
4589 folio_putback_lru(folio);
4590 continue;
4591 }
4592
4593 if (folio_test_reclaim(folio) &&
4594 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4595 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4596 if (folio_test_workingset(folio))
4597 folio_set_referenced(folio);
4598 continue;
4599 }
4600
4601 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4602 folio_mapped(folio) || folio_test_locked(folio) ||
4603 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4604 /* don't add rejected folios to the oldest generation */
4605 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4606 BIT(PG_active));
4607 continue;
4608 }
4609
4610 /* retry folios that may have missed folio_rotate_reclaimable() */
4611 list_move(&folio->lru, &clean);
4612 }
4613
4614 spin_lock_irq(&lruvec->lru_lock);
4615
4616 move_folios_to_lru(lruvec, &list);
4617
4618 walk = current->reclaim_state->mm_walk;
4619 if (walk && walk->batched) {
4620 walk->lruvec = lruvec;
4621 reset_batch_size(walk);
4622 }
4623
4624 item = PGSTEAL_KSWAPD + reclaimer_offset();
4625 if (!cgroup_reclaim(sc))
4626 __count_vm_events(item, reclaimed);
4627 __count_memcg_events(memcg, item, reclaimed);
4628 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4629
4630 spin_unlock_irq(&lruvec->lru_lock);
4631
4632 list_splice_init(&clean, &list);
4633
4634 if (!list_empty(&list)) {
4635 skip_retry = true;
4636 goto retry;
4637 }
4638
4639 return scanned;
4640 }
4641
4642 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4643 bool can_swap, unsigned long *nr_to_scan)
4644 {
4645 int gen, type, zone;
4646 unsigned long old = 0;
4647 unsigned long young = 0;
4648 unsigned long total = 0;
4649 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4650 DEFINE_MIN_SEQ(lruvec);
4651
4652 /* whether this lruvec is completely out of cold folios */
4653 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4654 *nr_to_scan = 0;
4655 return true;
4656 }
4657
4658 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4659 unsigned long seq;
4660
4661 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4662 unsigned long size = 0;
4663
4664 gen = lru_gen_from_seq(seq);
4665
4666 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4667 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4668
4669 total += size;
4670 if (seq == max_seq)
4671 young += size;
4672 else if (seq + MIN_NR_GENS == max_seq)
4673 old += size;
4674 }
4675 }
4676
4677 *nr_to_scan = total;
4678
4679 /*
4680 * The aging tries to be lazy to reduce the overhead, while the eviction
4681 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4682 * ideal number of generations is MIN_NR_GENS+1.
4683 */
4684 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4685 return false;
4686
4687 /*
4688 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4689 * of the total number of pages for each generation. A reasonable range
4690 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4691 * aging cares about the upper bound of hot pages, while the eviction
4692 * cares about the lower bound of cold pages.
4693 */
4694 if (young * MIN_NR_GENS > total)
4695 return true;
4696 if (old * (MIN_NR_GENS + 2) < total)
4697 return true;
4698
4699 return false;
4700 }
4701
4702 /*
4703 * For future optimizations:
4704 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4705 * reclaim.
4706 */
4707 static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4708 {
4709 bool success;
4710 unsigned long nr_to_scan;
4711 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4712 DEFINE_MAX_SEQ(lruvec);
4713
4714 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4715 return -1;
4716
4717 success = should_run_aging(lruvec, max_seq, can_swap, &nr_to_scan);
4718
4719 /* try to scrape all its memory if this memcg was deleted */
4720 if (nr_to_scan && !mem_cgroup_online(memcg))
4721 return nr_to_scan;
4722
4723 /* try to get away with not aging at the default priority */
4724 if (!success || sc->priority == DEF_PRIORITY)
4725 return nr_to_scan >> sc->priority;
4726
4727 /* stop scanning this lruvec as it's low on cold folios */
4728 return try_to_inc_max_seq(lruvec, max_seq, can_swap, false) ? -1 : 0;
4729 }
4730
4731 static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4732 {
4733 int i;
4734 enum zone_watermarks mark;
4735
4736 /* don't abort memcg reclaim to ensure fairness */
4737 if (!root_reclaim(sc))
4738 return false;
4739
4740 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4741 return true;
4742
4743 /* check the order to exclude compaction-induced reclaim */
4744 if (!current_is_kswapd() || sc->order)
4745 return false;
4746
4747 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4748 WMARK_PROMO : WMARK_HIGH;
4749
4750 for (i = 0; i <= sc->reclaim_idx; i++) {
4751 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4752 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4753
4754 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4755 return false;
4756 }
4757
4758 /* kswapd should abort if all eligible zones are safe */
4759 return true;
4760 }
4761
4762 static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4763 {
4764 long nr_to_scan;
4765 unsigned long scanned = 0;
4766 int swappiness = get_swappiness(lruvec, sc);
4767
4768 while (true) {
4769 int delta;
4770
4771 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4772 if (nr_to_scan <= 0)
4773 break;
4774
4775 delta = evict_folios(lruvec, sc, swappiness);
4776 if (!delta)
4777 break;
4778
4779 scanned += delta;
4780 if (scanned >= nr_to_scan)
4781 break;
4782
4783 if (should_abort_scan(lruvec, sc))
4784 break;
4785
4786 cond_resched();
4787 }
4788
4789 /* whether this lruvec should be rotated */
4790 return nr_to_scan < 0;
4791 }
4792
4793 static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4794 {
4795 bool success;
4796 unsigned long scanned = sc->nr_scanned;
4797 unsigned long reclaimed = sc->nr_reclaimed;
4798 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4799 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4800
4801 /* lru_gen_age_node() called mem_cgroup_calculate_protection() */
4802 if (mem_cgroup_below_min(NULL, memcg))
4803 return MEMCG_LRU_YOUNG;
4804
4805 if (mem_cgroup_below_low(NULL, memcg)) {
4806 /* see the comment on MEMCG_NR_GENS */
4807 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4808 return MEMCG_LRU_TAIL;
4809
4810 memcg_memory_event(memcg, MEMCG_LOW);
4811 }
4812
4813 success = try_to_shrink_lruvec(lruvec, sc);
4814
4815 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4816
4817 if (!sc->proactive)
4818 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4819 sc->nr_reclaimed - reclaimed);
4820
4821 flush_reclaim_state(sc);
4822
4823 if (success && mem_cgroup_online(memcg))
4824 return MEMCG_LRU_YOUNG;
4825
4826 if (!success && lruvec_is_sizable(lruvec, sc))
4827 return 0;
4828
4829 /* one retry if offlined or too small */
4830 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4831 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4832 }
4833
4834 static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4835 {
4836 int op;
4837 int gen;
4838 int bin;
4839 int first_bin;
4840 struct lruvec *lruvec;
4841 struct lru_gen_folio *lrugen;
4842 struct mem_cgroup *memcg;
4843 struct hlist_nulls_node *pos;
4844
4845 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4846 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4847 restart:
4848 op = 0;
4849 memcg = NULL;
4850
4851 rcu_read_lock();
4852
4853 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4854 if (op) {
4855 lru_gen_rotate_memcg(lruvec, op);
4856 op = 0;
4857 }
4858
4859 mem_cgroup_put(memcg);
4860 memcg = NULL;
4861
4862 if (gen != READ_ONCE(lrugen->gen))
4863 continue;
4864
4865 lruvec = container_of(lrugen, struct lruvec, lrugen);
4866 memcg = lruvec_memcg(lruvec);
4867
4868 if (!mem_cgroup_tryget(memcg)) {
4869 lru_gen_release_memcg(memcg);
4870 memcg = NULL;
4871 continue;
4872 }
4873
4874 rcu_read_unlock();
4875
4876 op = shrink_one(lruvec, sc);
4877
4878 rcu_read_lock();
4879
4880 if (should_abort_scan(lruvec, sc))
4881 break;
4882 }
4883
4884 rcu_read_unlock();
4885
4886 if (op)
4887 lru_gen_rotate_memcg(lruvec, op);
4888
4889 mem_cgroup_put(memcg);
4890
4891 if (!is_a_nulls(pos))
4892 return;
4893
4894 /* restart if raced with lru_gen_rotate_memcg() */
4895 if (gen != get_nulls_value(pos))
4896 goto restart;
4897
4898 /* try the rest of the bins of the current generation */
4899 bin = get_memcg_bin(bin + 1);
4900 if (bin != first_bin)
4901 goto restart;
4902 }
4903
4904 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4905 {
4906 struct blk_plug plug;
4907
4908 VM_WARN_ON_ONCE(root_reclaim(sc));
4909 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4910
4911 lru_add_drain();
4912
4913 blk_start_plug(&plug);
4914
4915 set_mm_walk(NULL, sc->proactive);
4916
4917 if (try_to_shrink_lruvec(lruvec, sc))
4918 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4919
4920 clear_mm_walk();
4921
4922 blk_finish_plug(&plug);
4923 }
4924
4925 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4926 {
4927 struct blk_plug plug;
4928 unsigned long reclaimed = sc->nr_reclaimed;
4929
4930 VM_WARN_ON_ONCE(!root_reclaim(sc));
4931
4932 /*
4933 * Unmapped clean folios are already prioritized. Scanning for more of
4934 * them is likely futile and can cause high reclaim latency when there
4935 * is a large number of memcgs.
4936 */
4937 if (!sc->may_writepage || !sc->may_unmap)
4938 goto done;
4939
4940 lru_add_drain();
4941
4942 blk_start_plug(&plug);
4943
4944 set_mm_walk(pgdat, sc->proactive);
4945
4946 set_initial_priority(pgdat, sc);
4947
4948 if (current_is_kswapd())
4949 sc->nr_reclaimed = 0;
4950
4951 if (mem_cgroup_disabled())
4952 shrink_one(&pgdat->__lruvec, sc);
4953 else
4954 shrink_many(pgdat, sc);
4955
4956 if (current_is_kswapd())
4957 sc->nr_reclaimed += reclaimed;
4958
4959 clear_mm_walk();
4960
4961 blk_finish_plug(&plug);
4962 done:
4963 /* kswapd should never fail */
4964 pgdat->kswapd_failures = 0;
4965 }
4966
4967 /******************************************************************************
4968 * state change
4969 ******************************************************************************/
4970
4971 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4972 {
4973 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4974
4975 if (lrugen->enabled) {
4976 enum lru_list lru;
4977
4978 for_each_evictable_lru(lru) {
4979 if (!list_empty(&lruvec->lists[lru]))
4980 return false;
4981 }
4982 } else {
4983 int gen, type, zone;
4984
4985 for_each_gen_type_zone(gen, type, zone) {
4986 if (!list_empty(&lrugen->folios[gen][type][zone]))
4987 return false;
4988 }
4989 }
4990
4991 return true;
4992 }
4993
4994 static bool fill_evictable(struct lruvec *lruvec)
4995 {
4996 enum lru_list lru;
4997 int remaining = MAX_LRU_BATCH;
4998
4999 for_each_evictable_lru(lru) {
5000 int type = is_file_lru(lru);
5001 bool active = is_active_lru(lru);
5002 struct list_head *head = &lruvec->lists[lru];
5003
5004 while (!list_empty(head)) {
5005 bool success;
5006 struct folio *folio = lru_to_folio(head);
5007
5008 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5009 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5010 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5011 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5012
5013 lruvec_del_folio(lruvec, folio);
5014 success = lru_gen_add_folio(lruvec, folio, false);
5015 VM_WARN_ON_ONCE(!success);
5016
5017 if (!--remaining)
5018 return false;
5019 }
5020 }
5021
5022 return true;
5023 }
5024
5025 static bool drain_evictable(struct lruvec *lruvec)
5026 {
5027 int gen, type, zone;
5028 int remaining = MAX_LRU_BATCH;
5029
5030 for_each_gen_type_zone(gen, type, zone) {
5031 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5032
5033 while (!list_empty(head)) {
5034 bool success;
5035 struct folio *folio = lru_to_folio(head);
5036
5037 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5038 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5039 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5040 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5041
5042 success = lru_gen_del_folio(lruvec, folio, false);
5043 VM_WARN_ON_ONCE(!success);
5044 lruvec_add_folio(lruvec, folio);
5045
5046 if (!--remaining)
5047 return false;
5048 }
5049 }
5050
5051 return true;
5052 }
5053
5054 static void lru_gen_change_state(bool enabled)
5055 {
5056 static DEFINE_MUTEX(state_mutex);
5057
5058 struct mem_cgroup *memcg;
5059
5060 cgroup_lock();
5061 cpus_read_lock();
5062 get_online_mems();
5063 mutex_lock(&state_mutex);
5064
5065 if (enabled == lru_gen_enabled())
5066 goto unlock;
5067
5068 if (enabled)
5069 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5070 else
5071 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5072
5073 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5074 do {
5075 int nid;
5076
5077 for_each_node(nid) {
5078 struct lruvec *lruvec = get_lruvec(memcg, nid);
5079
5080 spin_lock_irq(&lruvec->lru_lock);
5081
5082 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5083 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5084
5085 lruvec->lrugen.enabled = enabled;
5086
5087 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5088 spin_unlock_irq(&lruvec->lru_lock);
5089 cond_resched();
5090 spin_lock_irq(&lruvec->lru_lock);
5091 }
5092
5093 spin_unlock_irq(&lruvec->lru_lock);
5094 }
5095
5096 cond_resched();
5097 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5098 unlock:
5099 mutex_unlock(&state_mutex);
5100 put_online_mems();
5101 cpus_read_unlock();
5102 cgroup_unlock();
5103 }
5104
5105 /******************************************************************************
5106 * sysfs interface
5107 ******************************************************************************/
5108
5109 static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5110 {
5111 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5112 }
5113
5114 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5115 static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5116 const char *buf, size_t len)
5117 {
5118 unsigned int msecs;
5119
5120 if (kstrtouint(buf, 0, &msecs))
5121 return -EINVAL;
5122
5123 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5124
5125 return len;
5126 }
5127
5128 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5129
5130 static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5131 {
5132 unsigned int caps = 0;
5133
5134 if (get_cap(LRU_GEN_CORE))
5135 caps |= BIT(LRU_GEN_CORE);
5136
5137 if (should_walk_mmu())
5138 caps |= BIT(LRU_GEN_MM_WALK);
5139
5140 if (should_clear_pmd_young())
5141 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5142
5143 return sysfs_emit(buf, "0x%04x\n", caps);
5144 }
5145
5146 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5147 static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5148 const char *buf, size_t len)
5149 {
5150 int i;
5151 unsigned int caps;
5152
5153 if (tolower(*buf) == 'n')
5154 caps = 0;
5155 else if (tolower(*buf) == 'y')
5156 caps = -1;
5157 else if (kstrtouint(buf, 0, &caps))
5158 return -EINVAL;
5159
5160 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5161 bool enabled = caps & BIT(i);
5162
5163 if (i == LRU_GEN_CORE)
5164 lru_gen_change_state(enabled);
5165 else if (enabled)
5166 static_branch_enable(&lru_gen_caps[i]);
5167 else
5168 static_branch_disable(&lru_gen_caps[i]);
5169 }
5170
5171 return len;
5172 }
5173
5174 static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5175
5176 static struct attribute *lru_gen_attrs[] = {
5177 &lru_gen_min_ttl_attr.attr,
5178 &lru_gen_enabled_attr.attr,
5179 NULL
5180 };
5181
5182 static const struct attribute_group lru_gen_attr_group = {
5183 .name = "lru_gen",
5184 .attrs = lru_gen_attrs,
5185 };
5186
5187 /******************************************************************************
5188 * debugfs interface
5189 ******************************************************************************/
5190
5191 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5192 {
5193 struct mem_cgroup *memcg;
5194 loff_t nr_to_skip = *pos;
5195
5196 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5197 if (!m->private)
5198 return ERR_PTR(-ENOMEM);
5199
5200 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5201 do {
5202 int nid;
5203
5204 for_each_node_state(nid, N_MEMORY) {
5205 if (!nr_to_skip--)
5206 return get_lruvec(memcg, nid);
5207 }
5208 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5209
5210 return NULL;
5211 }
5212
5213 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5214 {
5215 if (!IS_ERR_OR_NULL(v))
5216 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5217
5218 kvfree(m->private);
5219 m->private = NULL;
5220 }
5221
5222 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5223 {
5224 int nid = lruvec_pgdat(v)->node_id;
5225 struct mem_cgroup *memcg = lruvec_memcg(v);
5226
5227 ++*pos;
5228
5229 nid = next_memory_node(nid);
5230 if (nid == MAX_NUMNODES) {
5231 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5232 if (!memcg)
5233 return NULL;
5234
5235 nid = first_memory_node;
5236 }
5237
5238 return get_lruvec(memcg, nid);
5239 }
5240
5241 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5242 unsigned long max_seq, unsigned long *min_seq,
5243 unsigned long seq)
5244 {
5245 int i;
5246 int type, tier;
5247 int hist = lru_hist_from_seq(seq);
5248 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5249 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5250
5251 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5252 seq_printf(m, " %10d", tier);
5253 for (type = 0; type < ANON_AND_FILE; type++) {
5254 const char *s = " ";
5255 unsigned long n[3] = {};
5256
5257 if (seq == max_seq) {
5258 s = "RT ";
5259 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5260 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5261 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5262 s = "rep";
5263 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5264 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5265 if (tier)
5266 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5267 }
5268
5269 for (i = 0; i < 3; i++)
5270 seq_printf(m, " %10lu%c", n[i], s[i]);
5271 }
5272 seq_putc(m, '\n');
5273 }
5274
5275 if (!mm_state)
5276 return;
5277
5278 seq_puts(m, " ");
5279 for (i = 0; i < NR_MM_STATS; i++) {
5280 const char *s = " ";
5281 unsigned long n = 0;
5282
5283 if (seq == max_seq && NR_HIST_GENS == 1) {
5284 s = "LOYNFA";
5285 n = READ_ONCE(mm_state->stats[hist][i]);
5286 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5287 s = "loynfa";
5288 n = READ_ONCE(mm_state->stats[hist][i]);
5289 }
5290
5291 seq_printf(m, " %10lu%c", n, s[i]);
5292 }
5293 seq_putc(m, '\n');
5294 }
5295
5296 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5297 static int lru_gen_seq_show(struct seq_file *m, void *v)
5298 {
5299 unsigned long seq;
5300 bool full = !debugfs_real_fops(m->file)->write;
5301 struct lruvec *lruvec = v;
5302 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5303 int nid = lruvec_pgdat(lruvec)->node_id;
5304 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5305 DEFINE_MAX_SEQ(lruvec);
5306 DEFINE_MIN_SEQ(lruvec);
5307
5308 if (nid == first_memory_node) {
5309 const char *path = memcg ? m->private : "";
5310
5311 #ifdef CONFIG_MEMCG
5312 if (memcg)
5313 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5314 #endif
5315 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5316 }
5317
5318 seq_printf(m, " node %5d\n", nid);
5319
5320 if (!full)
5321 seq = min_seq[LRU_GEN_ANON];
5322 else if (max_seq >= MAX_NR_GENS)
5323 seq = max_seq - MAX_NR_GENS + 1;
5324 else
5325 seq = 0;
5326
5327 for (; seq <= max_seq; seq++) {
5328 int type, zone;
5329 int gen = lru_gen_from_seq(seq);
5330 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5331
5332 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5333
5334 for (type = 0; type < ANON_AND_FILE; type++) {
5335 unsigned long size = 0;
5336 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5337
5338 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5339 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5340
5341 seq_printf(m, " %10lu%c", size, mark);
5342 }
5343
5344 seq_putc(m, '\n');
5345
5346 if (full)
5347 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5348 }
5349
5350 return 0;
5351 }
5352
5353 static const struct seq_operations lru_gen_seq_ops = {
5354 .start = lru_gen_seq_start,
5355 .stop = lru_gen_seq_stop,
5356 .next = lru_gen_seq_next,
5357 .show = lru_gen_seq_show,
5358 };
5359
5360 static int run_aging(struct lruvec *lruvec, unsigned long seq,
5361 bool can_swap, bool force_scan)
5362 {
5363 DEFINE_MAX_SEQ(lruvec);
5364 DEFINE_MIN_SEQ(lruvec);
5365
5366 if (seq < max_seq)
5367 return 0;
5368
5369 if (seq > max_seq)
5370 return -EINVAL;
5371
5372 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5373 return -ERANGE;
5374
5375 try_to_inc_max_seq(lruvec, max_seq, can_swap, force_scan);
5376
5377 return 0;
5378 }
5379
5380 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5381 int swappiness, unsigned long nr_to_reclaim)
5382 {
5383 DEFINE_MAX_SEQ(lruvec);
5384
5385 if (seq + MIN_NR_GENS > max_seq)
5386 return -EINVAL;
5387
5388 sc->nr_reclaimed = 0;
5389
5390 while (!signal_pending(current)) {
5391 DEFINE_MIN_SEQ(lruvec);
5392
5393 if (seq < min_seq[!swappiness])
5394 return 0;
5395
5396 if (sc->nr_reclaimed >= nr_to_reclaim)
5397 return 0;
5398
5399 if (!evict_folios(lruvec, sc, swappiness))
5400 return 0;
5401
5402 cond_resched();
5403 }
5404
5405 return -EINTR;
5406 }
5407
5408 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5409 struct scan_control *sc, int swappiness, unsigned long opt)
5410 {
5411 struct lruvec *lruvec;
5412 int err = -EINVAL;
5413 struct mem_cgroup *memcg = NULL;
5414
5415 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5416 return -EINVAL;
5417
5418 if (!mem_cgroup_disabled()) {
5419 rcu_read_lock();
5420
5421 memcg = mem_cgroup_from_id(memcg_id);
5422 if (!mem_cgroup_tryget(memcg))
5423 memcg = NULL;
5424
5425 rcu_read_unlock();
5426
5427 if (!memcg)
5428 return -EINVAL;
5429 }
5430
5431 if (memcg_id != mem_cgroup_id(memcg))
5432 goto done;
5433
5434 lruvec = get_lruvec(memcg, nid);
5435
5436 if (swappiness < MIN_SWAPPINESS)
5437 swappiness = get_swappiness(lruvec, sc);
5438 else if (swappiness > MAX_SWAPPINESS)
5439 goto done;
5440
5441 switch (cmd) {
5442 case '+':
5443 err = run_aging(lruvec, seq, swappiness, opt);
5444 break;
5445 case '-':
5446 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5447 break;
5448 }
5449 done:
5450 mem_cgroup_put(memcg);
5451
5452 return err;
5453 }
5454
5455 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5456 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5457 size_t len, loff_t *pos)
5458 {
5459 void *buf;
5460 char *cur, *next;
5461 unsigned int flags;
5462 struct blk_plug plug;
5463 int err = -EINVAL;
5464 struct scan_control sc = {
5465 .may_writepage = true,
5466 .may_unmap = true,
5467 .may_swap = true,
5468 .reclaim_idx = MAX_NR_ZONES - 1,
5469 .gfp_mask = GFP_KERNEL,
5470 };
5471
5472 buf = kvmalloc(len + 1, GFP_KERNEL);
5473 if (!buf)
5474 return -ENOMEM;
5475
5476 if (copy_from_user(buf, src, len)) {
5477 kvfree(buf);
5478 return -EFAULT;
5479 }
5480
5481 set_task_reclaim_state(current, &sc.reclaim_state);
5482 flags = memalloc_noreclaim_save();
5483 blk_start_plug(&plug);
5484 if (!set_mm_walk(NULL, true)) {
5485 err = -ENOMEM;
5486 goto done;
5487 }
5488
5489 next = buf;
5490 next[len] = '\0';
5491
5492 while ((cur = strsep(&next, ",;\n"))) {
5493 int n;
5494 int end;
5495 char cmd;
5496 unsigned int memcg_id;
5497 unsigned int nid;
5498 unsigned long seq;
5499 unsigned int swappiness = -1;
5500 unsigned long opt = -1;
5501
5502 cur = skip_spaces(cur);
5503 if (!*cur)
5504 continue;
5505
5506 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5507 &seq, &end, &swappiness, &end, &opt, &end);
5508 if (n < 4 || cur[end]) {
5509 err = -EINVAL;
5510 break;
5511 }
5512
5513 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5514 if (err)
5515 break;
5516 }
5517 done:
5518 clear_mm_walk();
5519 blk_finish_plug(&plug);
5520 memalloc_noreclaim_restore(flags);
5521 set_task_reclaim_state(current, NULL);
5522
5523 kvfree(buf);
5524
5525 return err ? : len;
5526 }
5527
5528 static int lru_gen_seq_open(struct inode *inode, struct file *file)
5529 {
5530 return seq_open(file, &lru_gen_seq_ops);
5531 }
5532
5533 static const struct file_operations lru_gen_rw_fops = {
5534 .open = lru_gen_seq_open,
5535 .read = seq_read,
5536 .write = lru_gen_seq_write,
5537 .llseek = seq_lseek,
5538 .release = seq_release,
5539 };
5540
5541 static const struct file_operations lru_gen_ro_fops = {
5542 .open = lru_gen_seq_open,
5543 .read = seq_read,
5544 .llseek = seq_lseek,
5545 .release = seq_release,
5546 };
5547
5548 /******************************************************************************
5549 * initialization
5550 ******************************************************************************/
5551
5552 void lru_gen_init_pgdat(struct pglist_data *pgdat)
5553 {
5554 int i, j;
5555
5556 spin_lock_init(&pgdat->memcg_lru.lock);
5557
5558 for (i = 0; i < MEMCG_NR_GENS; i++) {
5559 for (j = 0; j < MEMCG_NR_BINS; j++)
5560 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5561 }
5562 }
5563
5564 void lru_gen_init_lruvec(struct lruvec *lruvec)
5565 {
5566 int i;
5567 int gen, type, zone;
5568 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5569 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5570
5571 lrugen->max_seq = MIN_NR_GENS + 1;
5572 lrugen->enabled = lru_gen_enabled();
5573
5574 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5575 lrugen->timestamps[i] = jiffies;
5576
5577 for_each_gen_type_zone(gen, type, zone)
5578 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5579
5580 if (mm_state)
5581 mm_state->seq = MIN_NR_GENS;
5582 }
5583
5584 #ifdef CONFIG_MEMCG
5585
5586 void lru_gen_init_memcg(struct mem_cgroup *memcg)
5587 {
5588 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5589
5590 if (!mm_list)
5591 return;
5592
5593 INIT_LIST_HEAD(&mm_list->fifo);
5594 spin_lock_init(&mm_list->lock);
5595 }
5596
5597 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5598 {
5599 int i;
5600 int nid;
5601 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5602
5603 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5604
5605 for_each_node(nid) {
5606 struct lruvec *lruvec = get_lruvec(memcg, nid);
5607 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5608
5609 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5610 sizeof(lruvec->lrugen.nr_pages)));
5611
5612 lruvec->lrugen.list.next = LIST_POISON1;
5613
5614 if (!mm_state)
5615 continue;
5616
5617 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5618 bitmap_free(mm_state->filters[i]);
5619 mm_state->filters[i] = NULL;
5620 }
5621 }
5622 }
5623
5624 #endif /* CONFIG_MEMCG */
5625
5626 static int __init init_lru_gen(void)
5627 {
5628 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5629 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5630
5631 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5632 pr_err("lru_gen: failed to create sysfs group\n");
5633
5634 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5635 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5636
5637 return 0;
5638 };
5639 late_initcall(init_lru_gen);
5640
5641 #else /* !CONFIG_LRU_GEN */
5642
5643 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5644 {
5645 BUILD_BUG();
5646 }
5647
5648 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5649 {
5650 BUILD_BUG();
5651 }
5652
5653 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5654 {
5655 BUILD_BUG();
5656 }
5657
5658 #endif /* CONFIG_LRU_GEN */
5659
5660 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5661 {
5662 unsigned long nr[NR_LRU_LISTS];
5663 unsigned long targets[NR_LRU_LISTS];
5664 unsigned long nr_to_scan;
5665 enum lru_list lru;
5666 unsigned long nr_reclaimed = 0;
5667 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5668 bool proportional_reclaim;
5669 struct blk_plug plug;
5670
5671 if (lru_gen_enabled() && !root_reclaim(sc)) {
5672 lru_gen_shrink_lruvec(lruvec, sc);
5673 return;
5674 }
5675
5676 get_scan_count(lruvec, sc, nr);
5677
5678 /* Record the original scan target for proportional adjustments later */
5679 memcpy(targets, nr, sizeof(nr));
5680
5681 /*
5682 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5683 * event that can occur when there is little memory pressure e.g.
5684 * multiple streaming readers/writers. Hence, we do not abort scanning
5685 * when the requested number of pages are reclaimed when scanning at
5686 * DEF_PRIORITY on the assumption that the fact we are direct
5687 * reclaiming implies that kswapd is not keeping up and it is best to
5688 * do a batch of work at once. For memcg reclaim one check is made to
5689 * abort proportional reclaim if either the file or anon lru has already
5690 * dropped to zero at the first pass.
5691 */
5692 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5693 sc->priority == DEF_PRIORITY);
5694
5695 blk_start_plug(&plug);
5696 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5697 nr[LRU_INACTIVE_FILE]) {
5698 unsigned long nr_anon, nr_file, percentage;
5699 unsigned long nr_scanned;
5700
5701 for_each_evictable_lru(lru) {
5702 if (nr[lru]) {
5703 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5704 nr[lru] -= nr_to_scan;
5705
5706 nr_reclaimed += shrink_list(lru, nr_to_scan,
5707 lruvec, sc);
5708 }
5709 }
5710
5711 cond_resched();
5712
5713 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5714 continue;
5715
5716 /*
5717 * For kswapd and memcg, reclaim at least the number of pages
5718 * requested. Ensure that the anon and file LRUs are scanned
5719 * proportionally what was requested by get_scan_count(). We
5720 * stop reclaiming one LRU and reduce the amount scanning
5721 * proportional to the original scan target.
5722 */
5723 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5724 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5725
5726 /*
5727 * It's just vindictive to attack the larger once the smaller
5728 * has gone to zero. And given the way we stop scanning the
5729 * smaller below, this makes sure that we only make one nudge
5730 * towards proportionality once we've got nr_to_reclaim.
5731 */
5732 if (!nr_file || !nr_anon)
5733 break;
5734
5735 if (nr_file > nr_anon) {
5736 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5737 targets[LRU_ACTIVE_ANON] + 1;
5738 lru = LRU_BASE;
5739 percentage = nr_anon * 100 / scan_target;
5740 } else {
5741 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5742 targets[LRU_ACTIVE_FILE] + 1;
5743 lru = LRU_FILE;
5744 percentage = nr_file * 100 / scan_target;
5745 }
5746
5747 /* Stop scanning the smaller of the LRU */
5748 nr[lru] = 0;
5749 nr[lru + LRU_ACTIVE] = 0;
5750
5751 /*
5752 * Recalculate the other LRU scan count based on its original
5753 * scan target and the percentage scanning already complete
5754 */
5755 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5756 nr_scanned = targets[lru] - nr[lru];
5757 nr[lru] = targets[lru] * (100 - percentage) / 100;
5758 nr[lru] -= min(nr[lru], nr_scanned);
5759
5760 lru += LRU_ACTIVE;
5761 nr_scanned = targets[lru] - nr[lru];
5762 nr[lru] = targets[lru] * (100 - percentage) / 100;
5763 nr[lru] -= min(nr[lru], nr_scanned);
5764 }
5765 blk_finish_plug(&plug);
5766 sc->nr_reclaimed += nr_reclaimed;
5767
5768 /*
5769 * Even if we did not try to evict anon pages at all, we want to
5770 * rebalance the anon lru active/inactive ratio.
5771 */
5772 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5773 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5774 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5775 sc, LRU_ACTIVE_ANON);
5776 }
5777
5778 /* Use reclaim/compaction for costly allocs or under memory pressure */
5779 static bool in_reclaim_compaction(struct scan_control *sc)
5780 {
5781 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5782 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5783 sc->priority < DEF_PRIORITY - 2))
5784 return true;
5785
5786 return false;
5787 }
5788
5789 /*
5790 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5791 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5792 * true if more pages should be reclaimed such that when the page allocator
5793 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5794 * It will give up earlier than that if there is difficulty reclaiming pages.
5795 */
5796 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5797 unsigned long nr_reclaimed,
5798 struct scan_control *sc)
5799 {
5800 unsigned long pages_for_compaction;
5801 unsigned long inactive_lru_pages;
5802 int z;
5803
5804 /* If not in reclaim/compaction mode, stop */
5805 if (!in_reclaim_compaction(sc))
5806 return false;
5807
5808 /*
5809 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5810 * number of pages that were scanned. This will return to the caller
5811 * with the risk reclaim/compaction and the resulting allocation attempt
5812 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5813 * allocations through requiring that the full LRU list has been scanned
5814 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5815 * scan, but that approximation was wrong, and there were corner cases
5816 * where always a non-zero amount of pages were scanned.
5817 */
5818 if (!nr_reclaimed)
5819 return false;
5820
5821 /* If compaction would go ahead or the allocation would succeed, stop */
5822 for (z = 0; z <= sc->reclaim_idx; z++) {
5823 struct zone *zone = &pgdat->node_zones[z];
5824 if (!managed_zone(zone))
5825 continue;
5826
5827 /* Allocation can already succeed, nothing to do */
5828 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5829 sc->reclaim_idx, 0))
5830 return false;
5831
5832 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5833 return false;
5834 }
5835
5836 /*
5837 * If we have not reclaimed enough pages for compaction and the
5838 * inactive lists are large enough, continue reclaiming
5839 */
5840 pages_for_compaction = compact_gap(sc->order);
5841 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5842 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5843 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5844
5845 return inactive_lru_pages > pages_for_compaction;
5846 }
5847
5848 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5849 {
5850 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5851 struct mem_cgroup_reclaim_cookie reclaim = {
5852 .pgdat = pgdat,
5853 };
5854 struct mem_cgroup_reclaim_cookie *partial = &reclaim;
5855 struct mem_cgroup *memcg;
5856
5857 /*
5858 * In most cases, direct reclaimers can do partial walks
5859 * through the cgroup tree, using an iterator state that
5860 * persists across invocations. This strikes a balance between
5861 * fairness and allocation latency.
5862 *
5863 * For kswapd, reliable forward progress is more important
5864 * than a quick return to idle. Always do full walks.
5865 */
5866 if (current_is_kswapd() || sc->memcg_full_walk)
5867 partial = NULL;
5868
5869 memcg = mem_cgroup_iter(target_memcg, NULL, partial);
5870 do {
5871 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5872 unsigned long reclaimed;
5873 unsigned long scanned;
5874
5875 /*
5876 * This loop can become CPU-bound when target memcgs
5877 * aren't eligible for reclaim - either because they
5878 * don't have any reclaimable pages, or because their
5879 * memory is explicitly protected. Avoid soft lockups.
5880 */
5881 cond_resched();
5882
5883 mem_cgroup_calculate_protection(target_memcg, memcg);
5884
5885 if (mem_cgroup_below_min(target_memcg, memcg)) {
5886 /*
5887 * Hard protection.
5888 * If there is no reclaimable memory, OOM.
5889 */
5890 continue;
5891 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5892 /*
5893 * Soft protection.
5894 * Respect the protection only as long as
5895 * there is an unprotected supply
5896 * of reclaimable memory from other cgroups.
5897 */
5898 if (!sc->memcg_low_reclaim) {
5899 sc->memcg_low_skipped = 1;
5900 continue;
5901 }
5902 memcg_memory_event(memcg, MEMCG_LOW);
5903 }
5904
5905 reclaimed = sc->nr_reclaimed;
5906 scanned = sc->nr_scanned;
5907
5908 shrink_lruvec(lruvec, sc);
5909
5910 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5911 sc->priority);
5912
5913 /* Record the group's reclaim efficiency */
5914 if (!sc->proactive)
5915 vmpressure(sc->gfp_mask, memcg, false,
5916 sc->nr_scanned - scanned,
5917 sc->nr_reclaimed - reclaimed);
5918
5919 /* If partial walks are allowed, bail once goal is reached */
5920 if (partial && sc->nr_reclaimed >= sc->nr_to_reclaim) {
5921 mem_cgroup_iter_break(target_memcg, memcg);
5922 break;
5923 }
5924 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, partial)));
5925 }
5926
5927 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5928 {
5929 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5930 struct lruvec *target_lruvec;
5931 bool reclaimable = false;
5932
5933 if (lru_gen_enabled() && root_reclaim(sc)) {
5934 lru_gen_shrink_node(pgdat, sc);
5935 return;
5936 }
5937
5938 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5939
5940 again:
5941 memset(&sc->nr, 0, sizeof(sc->nr));
5942
5943 nr_reclaimed = sc->nr_reclaimed;
5944 nr_scanned = sc->nr_scanned;
5945
5946 prepare_scan_control(pgdat, sc);
5947
5948 shrink_node_memcgs(pgdat, sc);
5949
5950 flush_reclaim_state(sc);
5951
5952 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5953
5954 /* Record the subtree's reclaim efficiency */
5955 if (!sc->proactive)
5956 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5957 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5958
5959 if (nr_node_reclaimed)
5960 reclaimable = true;
5961
5962 if (current_is_kswapd()) {
5963 /*
5964 * If reclaim is isolating dirty pages under writeback,
5965 * it implies that the long-lived page allocation rate
5966 * is exceeding the page laundering rate. Either the
5967 * global limits are not being effective at throttling
5968 * processes due to the page distribution throughout
5969 * zones or there is heavy usage of a slow backing
5970 * device. The only option is to throttle from reclaim
5971 * context which is not ideal as there is no guarantee
5972 * the dirtying process is throttled in the same way
5973 * balance_dirty_pages() manages.
5974 *
5975 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5976 * count the number of pages under pages flagged for
5977 * immediate reclaim and stall if any are encountered
5978 * in the nr_immediate check below.
5979 */
5980 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5981 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5982
5983 /* Allow kswapd to start writing pages during reclaim.*/
5984 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5985 set_bit(PGDAT_DIRTY, &pgdat->flags);
5986
5987 /*
5988 * If kswapd scans pages marked for immediate
5989 * reclaim and under writeback (nr_immediate), it
5990 * implies that pages are cycling through the LRU
5991 * faster than they are written so forcibly stall
5992 * until some pages complete writeback.
5993 */
5994 if (sc->nr.immediate)
5995 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5996 }
5997
5998 /*
5999 * Tag a node/memcg as congested if all the dirty pages were marked
6000 * for writeback and immediate reclaim (counted in nr.congested).
6001 *
6002 * Legacy memcg will stall in page writeback so avoid forcibly
6003 * stalling in reclaim_throttle().
6004 */
6005 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
6006 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
6007 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
6008
6009 if (current_is_kswapd())
6010 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
6011 }
6012
6013 /*
6014 * Stall direct reclaim for IO completions if the lruvec is
6015 * node is congested. Allow kswapd to continue until it
6016 * starts encountering unqueued dirty pages or cycling through
6017 * the LRU too quickly.
6018 */
6019 if (!current_is_kswapd() && current_may_throttle() &&
6020 !sc->hibernation_mode &&
6021 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
6022 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
6023 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6024
6025 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
6026 goto again;
6027
6028 /*
6029 * Kswapd gives up on balancing particular nodes after too
6030 * many failures to reclaim anything from them and goes to
6031 * sleep. On reclaim progress, reset the failure counter. A
6032 * successful direct reclaim run will revive a dormant kswapd.
6033 */
6034 if (reclaimable)
6035 pgdat->kswapd_failures = 0;
6036 else if (sc->cache_trim_mode)
6037 sc->cache_trim_mode_failed = 1;
6038 }
6039
6040 /*
6041 * Returns true if compaction should go ahead for a costly-order request, or
6042 * the allocation would already succeed without compaction. Return false if we
6043 * should reclaim first.
6044 */
6045 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6046 {
6047 unsigned long watermark;
6048
6049 if (!gfp_compaction_allowed(sc->gfp_mask))
6050 return false;
6051
6052 /* Allocation can already succeed, nothing to do */
6053 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6054 sc->reclaim_idx, 0))
6055 return true;
6056
6057 /* Compaction cannot yet proceed. Do reclaim. */
6058 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6059 return false;
6060
6061 /*
6062 * Compaction is already possible, but it takes time to run and there
6063 * are potentially other callers using the pages just freed. So proceed
6064 * with reclaim to make a buffer of free pages available to give
6065 * compaction a reasonable chance of completing and allocating the page.
6066 * Note that we won't actually reclaim the whole buffer in one attempt
6067 * as the target watermark in should_continue_reclaim() is lower. But if
6068 * we are already above the high+gap watermark, don't reclaim at all.
6069 */
6070 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6071
6072 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6073 }
6074
6075 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6076 {
6077 /*
6078 * If reclaim is making progress greater than 12% efficiency then
6079 * wake all the NOPROGRESS throttled tasks.
6080 */
6081 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6082 wait_queue_head_t *wqh;
6083
6084 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6085 if (waitqueue_active(wqh))
6086 wake_up(wqh);
6087
6088 return;
6089 }
6090
6091 /*
6092 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6093 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6094 * under writeback and marked for immediate reclaim at the tail of the
6095 * LRU.
6096 */
6097 if (current_is_kswapd() || cgroup_reclaim(sc))
6098 return;
6099
6100 /* Throttle if making no progress at high prioities. */
6101 if (sc->priority == 1 && !sc->nr_reclaimed)
6102 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6103 }
6104
6105 /*
6106 * This is the direct reclaim path, for page-allocating processes. We only
6107 * try to reclaim pages from zones which will satisfy the caller's allocation
6108 * request.
6109 *
6110 * If a zone is deemed to be full of pinned pages then just give it a light
6111 * scan then give up on it.
6112 */
6113 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6114 {
6115 struct zoneref *z;
6116 struct zone *zone;
6117 unsigned long nr_soft_reclaimed;
6118 unsigned long nr_soft_scanned;
6119 gfp_t orig_mask;
6120 pg_data_t *last_pgdat = NULL;
6121 pg_data_t *first_pgdat = NULL;
6122
6123 /*
6124 * If the number of buffer_heads in the machine exceeds the maximum
6125 * allowed level, force direct reclaim to scan the highmem zone as
6126 * highmem pages could be pinning lowmem pages storing buffer_heads
6127 */
6128 orig_mask = sc->gfp_mask;
6129 if (buffer_heads_over_limit) {
6130 sc->gfp_mask |= __GFP_HIGHMEM;
6131 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6132 }
6133
6134 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6135 sc->reclaim_idx, sc->nodemask) {
6136 /*
6137 * Take care memory controller reclaiming has small influence
6138 * to global LRU.
6139 */
6140 if (!cgroup_reclaim(sc)) {
6141 if (!cpuset_zone_allowed(zone,
6142 GFP_KERNEL | __GFP_HARDWALL))
6143 continue;
6144
6145 /*
6146 * If we already have plenty of memory free for
6147 * compaction in this zone, don't free any more.
6148 * Even though compaction is invoked for any
6149 * non-zero order, only frequent costly order
6150 * reclamation is disruptive enough to become a
6151 * noticeable problem, like transparent huge
6152 * page allocations.
6153 */
6154 if (IS_ENABLED(CONFIG_COMPACTION) &&
6155 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6156 compaction_ready(zone, sc)) {
6157 sc->compaction_ready = true;
6158 continue;
6159 }
6160
6161 /*
6162 * Shrink each node in the zonelist once. If the
6163 * zonelist is ordered by zone (not the default) then a
6164 * node may be shrunk multiple times but in that case
6165 * the user prefers lower zones being preserved.
6166 */
6167 if (zone->zone_pgdat == last_pgdat)
6168 continue;
6169
6170 /*
6171 * This steals pages from memory cgroups over softlimit
6172 * and returns the number of reclaimed pages and
6173 * scanned pages. This works for global memory pressure
6174 * and balancing, not for a memcg's limit.
6175 */
6176 nr_soft_scanned = 0;
6177 nr_soft_reclaimed = memcg1_soft_limit_reclaim(zone->zone_pgdat,
6178 sc->order, sc->gfp_mask,
6179 &nr_soft_scanned);
6180 sc->nr_reclaimed += nr_soft_reclaimed;
6181 sc->nr_scanned += nr_soft_scanned;
6182 /* need some check for avoid more shrink_zone() */
6183 }
6184
6185 if (!first_pgdat)
6186 first_pgdat = zone->zone_pgdat;
6187
6188 /* See comment about same check for global reclaim above */
6189 if (zone->zone_pgdat == last_pgdat)
6190 continue;
6191 last_pgdat = zone->zone_pgdat;
6192 shrink_node(zone->zone_pgdat, sc);
6193 }
6194
6195 if (first_pgdat)
6196 consider_reclaim_throttle(first_pgdat, sc);
6197
6198 /*
6199 * Restore to original mask to avoid the impact on the caller if we
6200 * promoted it to __GFP_HIGHMEM.
6201 */
6202 sc->gfp_mask = orig_mask;
6203 }
6204
6205 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6206 {
6207 struct lruvec *target_lruvec;
6208 unsigned long refaults;
6209
6210 if (lru_gen_enabled())
6211 return;
6212
6213 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6214 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6215 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6216 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6217 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6218 }
6219
6220 /*
6221 * This is the main entry point to direct page reclaim.
6222 *
6223 * If a full scan of the inactive list fails to free enough memory then we
6224 * are "out of memory" and something needs to be killed.
6225 *
6226 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6227 * high - the zone may be full of dirty or under-writeback pages, which this
6228 * caller can't do much about. We kick the writeback threads and take explicit
6229 * naps in the hope that some of these pages can be written. But if the
6230 * allocating task holds filesystem locks which prevent writeout this might not
6231 * work, and the allocation attempt will fail.
6232 *
6233 * returns: 0, if no pages reclaimed
6234 * else, the number of pages reclaimed
6235 */
6236 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6237 struct scan_control *sc)
6238 {
6239 int initial_priority = sc->priority;
6240 pg_data_t *last_pgdat;
6241 struct zoneref *z;
6242 struct zone *zone;
6243 retry:
6244 delayacct_freepages_start();
6245
6246 if (!cgroup_reclaim(sc))
6247 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6248
6249 do {
6250 if (!sc->proactive)
6251 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6252 sc->priority);
6253 sc->nr_scanned = 0;
6254 shrink_zones(zonelist, sc);
6255
6256 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6257 break;
6258
6259 if (sc->compaction_ready)
6260 break;
6261
6262 /*
6263 * If we're getting trouble reclaiming, start doing
6264 * writepage even in laptop mode.
6265 */
6266 if (sc->priority < DEF_PRIORITY - 2)
6267 sc->may_writepage = 1;
6268 } while (--sc->priority >= 0);
6269
6270 last_pgdat = NULL;
6271 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6272 sc->nodemask) {
6273 if (zone->zone_pgdat == last_pgdat)
6274 continue;
6275 last_pgdat = zone->zone_pgdat;
6276
6277 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6278
6279 if (cgroup_reclaim(sc)) {
6280 struct lruvec *lruvec;
6281
6282 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6283 zone->zone_pgdat);
6284 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6285 }
6286 }
6287
6288 delayacct_freepages_end();
6289
6290 if (sc->nr_reclaimed)
6291 return sc->nr_reclaimed;
6292
6293 /* Aborted reclaim to try compaction? don't OOM, then */
6294 if (sc->compaction_ready)
6295 return 1;
6296
6297 /*
6298 * In most cases, direct reclaimers can do partial walks
6299 * through the cgroup tree to meet the reclaim goal while
6300 * keeping latency low. Since the iterator state is shared
6301 * among all direct reclaim invocations (to retain fairness
6302 * among cgroups), though, high concurrency can result in
6303 * individual threads not seeing enough cgroups to make
6304 * meaningful forward progress. Avoid false OOMs in this case.
6305 */
6306 if (!sc->memcg_full_walk) {
6307 sc->priority = initial_priority;
6308 sc->memcg_full_walk = 1;
6309 goto retry;
6310 }
6311
6312 /*
6313 * We make inactive:active ratio decisions based on the node's
6314 * composition of memory, but a restrictive reclaim_idx or a
6315 * memory.low cgroup setting can exempt large amounts of
6316 * memory from reclaim. Neither of which are very common, so
6317 * instead of doing costly eligibility calculations of the
6318 * entire cgroup subtree up front, we assume the estimates are
6319 * good, and retry with forcible deactivation if that fails.
6320 */
6321 if (sc->skipped_deactivate) {
6322 sc->priority = initial_priority;
6323 sc->force_deactivate = 1;
6324 sc->skipped_deactivate = 0;
6325 goto retry;
6326 }
6327
6328 /* Untapped cgroup reserves? Don't OOM, retry. */
6329 if (sc->memcg_low_skipped) {
6330 sc->priority = initial_priority;
6331 sc->force_deactivate = 0;
6332 sc->memcg_low_reclaim = 1;
6333 sc->memcg_low_skipped = 0;
6334 goto retry;
6335 }
6336
6337 return 0;
6338 }
6339
6340 static bool allow_direct_reclaim(pg_data_t *pgdat)
6341 {
6342 struct zone *zone;
6343 unsigned long pfmemalloc_reserve = 0;
6344 unsigned long free_pages = 0;
6345 int i;
6346 bool wmark_ok;
6347
6348 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6349 return true;
6350
6351 for (i = 0; i <= ZONE_NORMAL; i++) {
6352 zone = &pgdat->node_zones[i];
6353 if (!managed_zone(zone))
6354 continue;
6355
6356 if (!zone_reclaimable_pages(zone))
6357 continue;
6358
6359 pfmemalloc_reserve += min_wmark_pages(zone);
6360 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6361 }
6362
6363 /* If there are no reserves (unexpected config) then do not throttle */
6364 if (!pfmemalloc_reserve)
6365 return true;
6366
6367 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6368
6369 /* kswapd must be awake if processes are being throttled */
6370 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6371 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6372 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6373
6374 wake_up_interruptible(&pgdat->kswapd_wait);
6375 }
6376
6377 return wmark_ok;
6378 }
6379
6380 /*
6381 * Throttle direct reclaimers if backing storage is backed by the network
6382 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6383 * depleted. kswapd will continue to make progress and wake the processes
6384 * when the low watermark is reached.
6385 *
6386 * Returns true if a fatal signal was delivered during throttling. If this
6387 * happens, the page allocator should not consider triggering the OOM killer.
6388 */
6389 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6390 nodemask_t *nodemask)
6391 {
6392 struct zoneref *z;
6393 struct zone *zone;
6394 pg_data_t *pgdat = NULL;
6395
6396 /*
6397 * Kernel threads should not be throttled as they may be indirectly
6398 * responsible for cleaning pages necessary for reclaim to make forward
6399 * progress. kjournald for example may enter direct reclaim while
6400 * committing a transaction where throttling it could forcing other
6401 * processes to block on log_wait_commit().
6402 */
6403 if (current->flags & PF_KTHREAD)
6404 goto out;
6405
6406 /*
6407 * If a fatal signal is pending, this process should not throttle.
6408 * It should return quickly so it can exit and free its memory
6409 */
6410 if (fatal_signal_pending(current))
6411 goto out;
6412
6413 /*
6414 * Check if the pfmemalloc reserves are ok by finding the first node
6415 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6416 * GFP_KERNEL will be required for allocating network buffers when
6417 * swapping over the network so ZONE_HIGHMEM is unusable.
6418 *
6419 * Throttling is based on the first usable node and throttled processes
6420 * wait on a queue until kswapd makes progress and wakes them. There
6421 * is an affinity then between processes waking up and where reclaim
6422 * progress has been made assuming the process wakes on the same node.
6423 * More importantly, processes running on remote nodes will not compete
6424 * for remote pfmemalloc reserves and processes on different nodes
6425 * should make reasonable progress.
6426 */
6427 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6428 gfp_zone(gfp_mask), nodemask) {
6429 if (zone_idx(zone) > ZONE_NORMAL)
6430 continue;
6431
6432 /* Throttle based on the first usable node */
6433 pgdat = zone->zone_pgdat;
6434 if (allow_direct_reclaim(pgdat))
6435 goto out;
6436 break;
6437 }
6438
6439 /* If no zone was usable by the allocation flags then do not throttle */
6440 if (!pgdat)
6441 goto out;
6442
6443 /* Account for the throttling */
6444 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6445
6446 /*
6447 * If the caller cannot enter the filesystem, it's possible that it
6448 * is due to the caller holding an FS lock or performing a journal
6449 * transaction in the case of a filesystem like ext[3|4]. In this case,
6450 * it is not safe to block on pfmemalloc_wait as kswapd could be
6451 * blocked waiting on the same lock. Instead, throttle for up to a
6452 * second before continuing.
6453 */
6454 if (!(gfp_mask & __GFP_FS))
6455 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6456 allow_direct_reclaim(pgdat), HZ);
6457 else
6458 /* Throttle until kswapd wakes the process */
6459 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6460 allow_direct_reclaim(pgdat));
6461
6462 if (fatal_signal_pending(current))
6463 return true;
6464
6465 out:
6466 return false;
6467 }
6468
6469 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6470 gfp_t gfp_mask, nodemask_t *nodemask)
6471 {
6472 unsigned long nr_reclaimed;
6473 struct scan_control sc = {
6474 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6475 .gfp_mask = current_gfp_context(gfp_mask),
6476 .reclaim_idx = gfp_zone(gfp_mask),
6477 .order = order,
6478 .nodemask = nodemask,
6479 .priority = DEF_PRIORITY,
6480 .may_writepage = !laptop_mode,
6481 .may_unmap = 1,
6482 .may_swap = 1,
6483 };
6484
6485 /*
6486 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6487 * Confirm they are large enough for max values.
6488 */
6489 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6490 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6491 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6492
6493 /*
6494 * Do not enter reclaim if fatal signal was delivered while throttled.
6495 * 1 is returned so that the page allocator does not OOM kill at this
6496 * point.
6497 */
6498 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6499 return 1;
6500
6501 set_task_reclaim_state(current, &sc.reclaim_state);
6502 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6503
6504 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6505
6506 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6507 set_task_reclaim_state(current, NULL);
6508
6509 return nr_reclaimed;
6510 }
6511
6512 #ifdef CONFIG_MEMCG
6513
6514 /* Only used by soft limit reclaim. Do not reuse for anything else. */
6515 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6516 gfp_t gfp_mask, bool noswap,
6517 pg_data_t *pgdat,
6518 unsigned long *nr_scanned)
6519 {
6520 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6521 struct scan_control sc = {
6522 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6523 .target_mem_cgroup = memcg,
6524 .may_writepage = !laptop_mode,
6525 .may_unmap = 1,
6526 .reclaim_idx = MAX_NR_ZONES - 1,
6527 .may_swap = !noswap,
6528 };
6529
6530 WARN_ON_ONCE(!current->reclaim_state);
6531
6532 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6533 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6534
6535 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6536 sc.gfp_mask);
6537
6538 /*
6539 * NOTE: Although we can get the priority field, using it
6540 * here is not a good idea, since it limits the pages we can scan.
6541 * if we don't reclaim here, the shrink_node from balance_pgdat
6542 * will pick up pages from other mem cgroup's as well. We hack
6543 * the priority and make it zero.
6544 */
6545 shrink_lruvec(lruvec, &sc);
6546
6547 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6548
6549 *nr_scanned = sc.nr_scanned;
6550
6551 return sc.nr_reclaimed;
6552 }
6553
6554 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6555 unsigned long nr_pages,
6556 gfp_t gfp_mask,
6557 unsigned int reclaim_options,
6558 int *swappiness)
6559 {
6560 unsigned long nr_reclaimed;
6561 unsigned int noreclaim_flag;
6562 struct scan_control sc = {
6563 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6564 .proactive_swappiness = swappiness,
6565 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6566 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6567 .reclaim_idx = MAX_NR_ZONES - 1,
6568 .target_mem_cgroup = memcg,
6569 .priority = DEF_PRIORITY,
6570 .may_writepage = !laptop_mode,
6571 .may_unmap = 1,
6572 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6573 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6574 };
6575 /*
6576 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6577 * equal pressure on all the nodes. This is based on the assumption that
6578 * the reclaim does not bail out early.
6579 */
6580 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6581
6582 set_task_reclaim_state(current, &sc.reclaim_state);
6583 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6584 noreclaim_flag = memalloc_noreclaim_save();
6585
6586 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6587
6588 memalloc_noreclaim_restore(noreclaim_flag);
6589 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6590 set_task_reclaim_state(current, NULL);
6591
6592 return nr_reclaimed;
6593 }
6594 #endif
6595
6596 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6597 {
6598 struct mem_cgroup *memcg;
6599 struct lruvec *lruvec;
6600
6601 if (lru_gen_enabled()) {
6602 lru_gen_age_node(pgdat, sc);
6603 return;
6604 }
6605
6606 if (!can_age_anon_pages(pgdat, sc))
6607 return;
6608
6609 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6610 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6611 return;
6612
6613 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6614 do {
6615 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6616 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6617 sc, LRU_ACTIVE_ANON);
6618 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6619 } while (memcg);
6620 }
6621
6622 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6623 {
6624 int i;
6625 struct zone *zone;
6626
6627 /*
6628 * Check for watermark boosts top-down as the higher zones
6629 * are more likely to be boosted. Both watermarks and boosts
6630 * should not be checked at the same time as reclaim would
6631 * start prematurely when there is no boosting and a lower
6632 * zone is balanced.
6633 */
6634 for (i = highest_zoneidx; i >= 0; i--) {
6635 zone = pgdat->node_zones + i;
6636 if (!managed_zone(zone))
6637 continue;
6638
6639 if (zone->watermark_boost)
6640 return true;
6641 }
6642
6643 return false;
6644 }
6645
6646 /*
6647 * Returns true if there is an eligible zone balanced for the request order
6648 * and highest_zoneidx
6649 */
6650 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6651 {
6652 int i;
6653 unsigned long mark = -1;
6654 struct zone *zone;
6655
6656 /*
6657 * Check watermarks bottom-up as lower zones are more likely to
6658 * meet watermarks.
6659 */
6660 for (i = 0; i <= highest_zoneidx; i++) {
6661 zone = pgdat->node_zones + i;
6662
6663 if (!managed_zone(zone))
6664 continue;
6665
6666 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6667 mark = wmark_pages(zone, WMARK_PROMO);
6668 else
6669 mark = high_wmark_pages(zone);
6670 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6671 return true;
6672 }
6673
6674 /*
6675 * If a node has no managed zone within highest_zoneidx, it does not
6676 * need balancing by definition. This can happen if a zone-restricted
6677 * allocation tries to wake a remote kswapd.
6678 */
6679 if (mark == -1)
6680 return true;
6681
6682 return false;
6683 }
6684
6685 /* Clear pgdat state for congested, dirty or under writeback. */
6686 static void clear_pgdat_congested(pg_data_t *pgdat)
6687 {
6688 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6689
6690 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6691 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6692 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6693 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6694 }
6695
6696 /*
6697 * Prepare kswapd for sleeping. This verifies that there are no processes
6698 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6699 *
6700 * Returns true if kswapd is ready to sleep
6701 */
6702 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6703 int highest_zoneidx)
6704 {
6705 /*
6706 * The throttled processes are normally woken up in balance_pgdat() as
6707 * soon as allow_direct_reclaim() is true. But there is a potential
6708 * race between when kswapd checks the watermarks and a process gets
6709 * throttled. There is also a potential race if processes get
6710 * throttled, kswapd wakes, a large process exits thereby balancing the
6711 * zones, which causes kswapd to exit balance_pgdat() before reaching
6712 * the wake up checks. If kswapd is going to sleep, no process should
6713 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6714 * the wake up is premature, processes will wake kswapd and get
6715 * throttled again. The difference from wake ups in balance_pgdat() is
6716 * that here we are under prepare_to_wait().
6717 */
6718 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6719 wake_up_all(&pgdat->pfmemalloc_wait);
6720
6721 /* Hopeless node, leave it to direct reclaim */
6722 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6723 return true;
6724
6725 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6726 clear_pgdat_congested(pgdat);
6727 return true;
6728 }
6729
6730 return false;
6731 }
6732
6733 /*
6734 * kswapd shrinks a node of pages that are at or below the highest usable
6735 * zone that is currently unbalanced.
6736 *
6737 * Returns true if kswapd scanned at least the requested number of pages to
6738 * reclaim or if the lack of progress was due to pages under writeback.
6739 * This is used to determine if the scanning priority needs to be raised.
6740 */
6741 static bool kswapd_shrink_node(pg_data_t *pgdat,
6742 struct scan_control *sc)
6743 {
6744 struct zone *zone;
6745 int z;
6746 unsigned long nr_reclaimed = sc->nr_reclaimed;
6747
6748 /* Reclaim a number of pages proportional to the number of zones */
6749 sc->nr_to_reclaim = 0;
6750 for (z = 0; z <= sc->reclaim_idx; z++) {
6751 zone = pgdat->node_zones + z;
6752 if (!managed_zone(zone))
6753 continue;
6754
6755 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6756 }
6757
6758 /*
6759 * Historically care was taken to put equal pressure on all zones but
6760 * now pressure is applied based on node LRU order.
6761 */
6762 shrink_node(pgdat, sc);
6763
6764 /*
6765 * Fragmentation may mean that the system cannot be rebalanced for
6766 * high-order allocations. If twice the allocation size has been
6767 * reclaimed then recheck watermarks only at order-0 to prevent
6768 * excessive reclaim. Assume that a process requested a high-order
6769 * can direct reclaim/compact.
6770 */
6771 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6772 sc->order = 0;
6773
6774 /* account for progress from mm_account_reclaimed_pages() */
6775 return max(sc->nr_scanned, sc->nr_reclaimed - nr_reclaimed) >= sc->nr_to_reclaim;
6776 }
6777
6778 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6779 static inline void
6780 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6781 {
6782 int i;
6783 struct zone *zone;
6784
6785 for (i = 0; i <= highest_zoneidx; i++) {
6786 zone = pgdat->node_zones + i;
6787
6788 if (!managed_zone(zone))
6789 continue;
6790
6791 if (active)
6792 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6793 else
6794 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6795 }
6796 }
6797
6798 static inline void
6799 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6800 {
6801 update_reclaim_active(pgdat, highest_zoneidx, true);
6802 }
6803
6804 static inline void
6805 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6806 {
6807 update_reclaim_active(pgdat, highest_zoneidx, false);
6808 }
6809
6810 /*
6811 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6812 * that are eligible for use by the caller until at least one zone is
6813 * balanced.
6814 *
6815 * Returns the order kswapd finished reclaiming at.
6816 *
6817 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6818 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6819 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6820 * or lower is eligible for reclaim until at least one usable zone is
6821 * balanced.
6822 */
6823 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6824 {
6825 int i;
6826 unsigned long nr_soft_reclaimed;
6827 unsigned long nr_soft_scanned;
6828 unsigned long pflags;
6829 unsigned long nr_boost_reclaim;
6830 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6831 bool boosted;
6832 struct zone *zone;
6833 struct scan_control sc = {
6834 .gfp_mask = GFP_KERNEL,
6835 .order = order,
6836 .may_unmap = 1,
6837 };
6838
6839 set_task_reclaim_state(current, &sc.reclaim_state);
6840 psi_memstall_enter(&pflags);
6841 __fs_reclaim_acquire(_THIS_IP_);
6842
6843 count_vm_event(PAGEOUTRUN);
6844
6845 /*
6846 * Account for the reclaim boost. Note that the zone boost is left in
6847 * place so that parallel allocations that are near the watermark will
6848 * stall or direct reclaim until kswapd is finished.
6849 */
6850 nr_boost_reclaim = 0;
6851 for (i = 0; i <= highest_zoneidx; i++) {
6852 zone = pgdat->node_zones + i;
6853 if (!managed_zone(zone))
6854 continue;
6855
6856 nr_boost_reclaim += zone->watermark_boost;
6857 zone_boosts[i] = zone->watermark_boost;
6858 }
6859 boosted = nr_boost_reclaim;
6860
6861 restart:
6862 set_reclaim_active(pgdat, highest_zoneidx);
6863 sc.priority = DEF_PRIORITY;
6864 do {
6865 unsigned long nr_reclaimed = sc.nr_reclaimed;
6866 bool raise_priority = true;
6867 bool balanced;
6868 bool ret;
6869 bool was_frozen;
6870
6871 sc.reclaim_idx = highest_zoneidx;
6872
6873 /*
6874 * If the number of buffer_heads exceeds the maximum allowed
6875 * then consider reclaiming from all zones. This has a dual
6876 * purpose -- on 64-bit systems it is expected that
6877 * buffer_heads are stripped during active rotation. On 32-bit
6878 * systems, highmem pages can pin lowmem memory and shrinking
6879 * buffers can relieve lowmem pressure. Reclaim may still not
6880 * go ahead if all eligible zones for the original allocation
6881 * request are balanced to avoid excessive reclaim from kswapd.
6882 */
6883 if (buffer_heads_over_limit) {
6884 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6885 zone = pgdat->node_zones + i;
6886 if (!managed_zone(zone))
6887 continue;
6888
6889 sc.reclaim_idx = i;
6890 break;
6891 }
6892 }
6893
6894 /*
6895 * If the pgdat is imbalanced then ignore boosting and preserve
6896 * the watermarks for a later time and restart. Note that the
6897 * zone watermarks will be still reset at the end of balancing
6898 * on the grounds that the normal reclaim should be enough to
6899 * re-evaluate if boosting is required when kswapd next wakes.
6900 */
6901 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6902 if (!balanced && nr_boost_reclaim) {
6903 nr_boost_reclaim = 0;
6904 goto restart;
6905 }
6906
6907 /*
6908 * If boosting is not active then only reclaim if there are no
6909 * eligible zones. Note that sc.reclaim_idx is not used as
6910 * buffer_heads_over_limit may have adjusted it.
6911 */
6912 if (!nr_boost_reclaim && balanced)
6913 goto out;
6914
6915 /* Limit the priority of boosting to avoid reclaim writeback */
6916 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6917 raise_priority = false;
6918
6919 /*
6920 * Do not writeback or swap pages for boosted reclaim. The
6921 * intent is to relieve pressure not issue sub-optimal IO
6922 * from reclaim context. If no pages are reclaimed, the
6923 * reclaim will be aborted.
6924 */
6925 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6926 sc.may_swap = !nr_boost_reclaim;
6927
6928 /*
6929 * Do some background aging, to give pages a chance to be
6930 * referenced before reclaiming. All pages are rotated
6931 * regardless of classzone as this is about consistent aging.
6932 */
6933 kswapd_age_node(pgdat, &sc);
6934
6935 /*
6936 * If we're getting trouble reclaiming, start doing writepage
6937 * even in laptop mode.
6938 */
6939 if (sc.priority < DEF_PRIORITY - 2)
6940 sc.may_writepage = 1;
6941
6942 /* Call soft limit reclaim before calling shrink_node. */
6943 sc.nr_scanned = 0;
6944 nr_soft_scanned = 0;
6945 nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat, sc.order,
6946 sc.gfp_mask, &nr_soft_scanned);
6947 sc.nr_reclaimed += nr_soft_reclaimed;
6948
6949 /*
6950 * There should be no need to raise the scanning priority if
6951 * enough pages are already being scanned that that high
6952 * watermark would be met at 100% efficiency.
6953 */
6954 if (kswapd_shrink_node(pgdat, &sc))
6955 raise_priority = false;
6956
6957 /*
6958 * If the low watermark is met there is no need for processes
6959 * to be throttled on pfmemalloc_wait as they should not be
6960 * able to safely make forward progress. Wake them
6961 */
6962 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6963 allow_direct_reclaim(pgdat))
6964 wake_up_all(&pgdat->pfmemalloc_wait);
6965
6966 /* Check if kswapd should be suspending */
6967 __fs_reclaim_release(_THIS_IP_);
6968 ret = kthread_freezable_should_stop(&was_frozen);
6969 __fs_reclaim_acquire(_THIS_IP_);
6970 if (was_frozen || ret)
6971 break;
6972
6973 /*
6974 * Raise priority if scanning rate is too low or there was no
6975 * progress in reclaiming pages
6976 */
6977 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6978 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6979
6980 /*
6981 * If reclaim made no progress for a boost, stop reclaim as
6982 * IO cannot be queued and it could be an infinite loop in
6983 * extreme circumstances.
6984 */
6985 if (nr_boost_reclaim && !nr_reclaimed)
6986 break;
6987
6988 if (raise_priority || !nr_reclaimed)
6989 sc.priority--;
6990 } while (sc.priority >= 1);
6991
6992 /*
6993 * Restart only if it went through the priority loop all the way,
6994 * but cache_trim_mode didn't work.
6995 */
6996 if (!sc.nr_reclaimed && sc.priority < 1 &&
6997 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
6998 sc.no_cache_trim_mode = 1;
6999 goto restart;
7000 }
7001
7002 if (!sc.nr_reclaimed)
7003 pgdat->kswapd_failures++;
7004
7005 out:
7006 clear_reclaim_active(pgdat, highest_zoneidx);
7007
7008 /* If reclaim was boosted, account for the reclaim done in this pass */
7009 if (boosted) {
7010 unsigned long flags;
7011
7012 for (i = 0; i <= highest_zoneidx; i++) {
7013 if (!zone_boosts[i])
7014 continue;
7015
7016 /* Increments are under the zone lock */
7017 zone = pgdat->node_zones + i;
7018 spin_lock_irqsave(&zone->lock, flags);
7019 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7020 spin_unlock_irqrestore(&zone->lock, flags);
7021 }
7022
7023 /*
7024 * As there is now likely space, wakeup kcompact to defragment
7025 * pageblocks.
7026 */
7027 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7028 }
7029
7030 snapshot_refaults(NULL, pgdat);
7031 __fs_reclaim_release(_THIS_IP_);
7032 psi_memstall_leave(&pflags);
7033 set_task_reclaim_state(current, NULL);
7034
7035 /*
7036 * Return the order kswapd stopped reclaiming at as
7037 * prepare_kswapd_sleep() takes it into account. If another caller
7038 * entered the allocator slow path while kswapd was awake, order will
7039 * remain at the higher level.
7040 */
7041 return sc.order;
7042 }
7043
7044 /*
7045 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7046 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7047 * not a valid index then either kswapd runs for first time or kswapd couldn't
7048 * sleep after previous reclaim attempt (node is still unbalanced). In that
7049 * case return the zone index of the previous kswapd reclaim cycle.
7050 */
7051 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7052 enum zone_type prev_highest_zoneidx)
7053 {
7054 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7055
7056 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7057 }
7058
7059 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7060 unsigned int highest_zoneidx)
7061 {
7062 long remaining = 0;
7063 DEFINE_WAIT(wait);
7064
7065 if (freezing(current) || kthread_should_stop())
7066 return;
7067
7068 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7069
7070 /*
7071 * Try to sleep for a short interval. Note that kcompactd will only be
7072 * woken if it is possible to sleep for a short interval. This is
7073 * deliberate on the assumption that if reclaim cannot keep an
7074 * eligible zone balanced that it's also unlikely that compaction will
7075 * succeed.
7076 */
7077 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7078 /*
7079 * Compaction records what page blocks it recently failed to
7080 * isolate pages from and skips them in the future scanning.
7081 * When kswapd is going to sleep, it is reasonable to assume
7082 * that pages and compaction may succeed so reset the cache.
7083 */
7084 reset_isolation_suitable(pgdat);
7085
7086 /*
7087 * We have freed the memory, now we should compact it to make
7088 * allocation of the requested order possible.
7089 */
7090 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7091
7092 remaining = schedule_timeout(HZ/10);
7093
7094 /*
7095 * If woken prematurely then reset kswapd_highest_zoneidx and
7096 * order. The values will either be from a wakeup request or
7097 * the previous request that slept prematurely.
7098 */
7099 if (remaining) {
7100 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7101 kswapd_highest_zoneidx(pgdat,
7102 highest_zoneidx));
7103
7104 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7105 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7106 }
7107
7108 finish_wait(&pgdat->kswapd_wait, &wait);
7109 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7110 }
7111
7112 /*
7113 * After a short sleep, check if it was a premature sleep. If not, then
7114 * go fully to sleep until explicitly woken up.
7115 */
7116 if (!remaining &&
7117 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7118 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7119
7120 /*
7121 * vmstat counters are not perfectly accurate and the estimated
7122 * value for counters such as NR_FREE_PAGES can deviate from the
7123 * true value by nr_online_cpus * threshold. To avoid the zone
7124 * watermarks being breached while under pressure, we reduce the
7125 * per-cpu vmstat threshold while kswapd is awake and restore
7126 * them before going back to sleep.
7127 */
7128 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7129
7130 if (!kthread_should_stop())
7131 schedule();
7132
7133 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7134 } else {
7135 if (remaining)
7136 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7137 else
7138 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7139 }
7140 finish_wait(&pgdat->kswapd_wait, &wait);
7141 }
7142
7143 /*
7144 * The background pageout daemon, started as a kernel thread
7145 * from the init process.
7146 *
7147 * This basically trickles out pages so that we have _some_
7148 * free memory available even if there is no other activity
7149 * that frees anything up. This is needed for things like routing
7150 * etc, where we otherwise might have all activity going on in
7151 * asynchronous contexts that cannot page things out.
7152 *
7153 * If there are applications that are active memory-allocators
7154 * (most normal use), this basically shouldn't matter.
7155 */
7156 static int kswapd(void *p)
7157 {
7158 unsigned int alloc_order, reclaim_order;
7159 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7160 pg_data_t *pgdat = (pg_data_t *)p;
7161 struct task_struct *tsk = current;
7162 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7163
7164 if (!cpumask_empty(cpumask))
7165 set_cpus_allowed_ptr(tsk, cpumask);
7166
7167 /*
7168 * Tell the memory management that we're a "memory allocator",
7169 * and that if we need more memory we should get access to it
7170 * regardless (see "__alloc_pages()"). "kswapd" should
7171 * never get caught in the normal page freeing logic.
7172 *
7173 * (Kswapd normally doesn't need memory anyway, but sometimes
7174 * you need a small amount of memory in order to be able to
7175 * page out something else, and this flag essentially protects
7176 * us from recursively trying to free more memory as we're
7177 * trying to free the first piece of memory in the first place).
7178 */
7179 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7180 set_freezable();
7181
7182 WRITE_ONCE(pgdat->kswapd_order, 0);
7183 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7184 atomic_set(&pgdat->nr_writeback_throttled, 0);
7185 for ( ; ; ) {
7186 bool was_frozen;
7187
7188 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7189 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7190 highest_zoneidx);
7191
7192 kswapd_try_sleep:
7193 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7194 highest_zoneidx);
7195
7196 /* Read the new order and highest_zoneidx */
7197 alloc_order = READ_ONCE(pgdat->kswapd_order);
7198 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7199 highest_zoneidx);
7200 WRITE_ONCE(pgdat->kswapd_order, 0);
7201 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7202
7203 if (kthread_freezable_should_stop(&was_frozen))
7204 break;
7205
7206 /*
7207 * We can speed up thawing tasks if we don't call balance_pgdat
7208 * after returning from the refrigerator
7209 */
7210 if (was_frozen)
7211 continue;
7212
7213 /*
7214 * Reclaim begins at the requested order but if a high-order
7215 * reclaim fails then kswapd falls back to reclaiming for
7216 * order-0. If that happens, kswapd will consider sleeping
7217 * for the order it finished reclaiming at (reclaim_order)
7218 * but kcompactd is woken to compact for the original
7219 * request (alloc_order).
7220 */
7221 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7222 alloc_order);
7223 reclaim_order = balance_pgdat(pgdat, alloc_order,
7224 highest_zoneidx);
7225 if (reclaim_order < alloc_order)
7226 goto kswapd_try_sleep;
7227 }
7228
7229 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7230
7231 return 0;
7232 }
7233
7234 /*
7235 * A zone is low on free memory or too fragmented for high-order memory. If
7236 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7237 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7238 * has failed or is not needed, still wake up kcompactd if only compaction is
7239 * needed.
7240 */
7241 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7242 enum zone_type highest_zoneidx)
7243 {
7244 pg_data_t *pgdat;
7245 enum zone_type curr_idx;
7246
7247 if (!managed_zone(zone))
7248 return;
7249
7250 if (!cpuset_zone_allowed(zone, gfp_flags))
7251 return;
7252
7253 pgdat = zone->zone_pgdat;
7254 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7255
7256 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7257 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7258
7259 if (READ_ONCE(pgdat->kswapd_order) < order)
7260 WRITE_ONCE(pgdat->kswapd_order, order);
7261
7262 if (!waitqueue_active(&pgdat->kswapd_wait))
7263 return;
7264
7265 /* Hopeless node, leave it to direct reclaim if possible */
7266 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7267 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7268 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7269 /*
7270 * There may be plenty of free memory available, but it's too
7271 * fragmented for high-order allocations. Wake up kcompactd
7272 * and rely on compaction_suitable() to determine if it's
7273 * needed. If it fails, it will defer subsequent attempts to
7274 * ratelimit its work.
7275 */
7276 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7277 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7278 return;
7279 }
7280
7281 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7282 gfp_flags);
7283 wake_up_interruptible(&pgdat->kswapd_wait);
7284 }
7285
7286 #ifdef CONFIG_HIBERNATION
7287 /*
7288 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7289 * freed pages.
7290 *
7291 * Rather than trying to age LRUs the aim is to preserve the overall
7292 * LRU order by reclaiming preferentially
7293 * inactive > active > active referenced > active mapped
7294 */
7295 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7296 {
7297 struct scan_control sc = {
7298 .nr_to_reclaim = nr_to_reclaim,
7299 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7300 .reclaim_idx = MAX_NR_ZONES - 1,
7301 .priority = DEF_PRIORITY,
7302 .may_writepage = 1,
7303 .may_unmap = 1,
7304 .may_swap = 1,
7305 .hibernation_mode = 1,
7306 };
7307 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7308 unsigned long nr_reclaimed;
7309 unsigned int noreclaim_flag;
7310
7311 fs_reclaim_acquire(sc.gfp_mask);
7312 noreclaim_flag = memalloc_noreclaim_save();
7313 set_task_reclaim_state(current, &sc.reclaim_state);
7314
7315 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7316
7317 set_task_reclaim_state(current, NULL);
7318 memalloc_noreclaim_restore(noreclaim_flag);
7319 fs_reclaim_release(sc.gfp_mask);
7320
7321 return nr_reclaimed;
7322 }
7323 #endif /* CONFIG_HIBERNATION */
7324
7325 /*
7326 * This kswapd start function will be called by init and node-hot-add.
7327 */
7328 void __meminit kswapd_run(int nid)
7329 {
7330 pg_data_t *pgdat = NODE_DATA(nid);
7331
7332 pgdat_kswapd_lock(pgdat);
7333 if (!pgdat->kswapd) {
7334 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7335 if (IS_ERR(pgdat->kswapd)) {
7336 /* failure at boot is fatal */
7337 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7338 nid, PTR_ERR(pgdat->kswapd));
7339 BUG_ON(system_state < SYSTEM_RUNNING);
7340 pgdat->kswapd = NULL;
7341 }
7342 }
7343 pgdat_kswapd_unlock(pgdat);
7344 }
7345
7346 /*
7347 * Called by memory hotplug when all memory in a node is offlined. Caller must
7348 * be holding mem_hotplug_begin/done().
7349 */
7350 void __meminit kswapd_stop(int nid)
7351 {
7352 pg_data_t *pgdat = NODE_DATA(nid);
7353 struct task_struct *kswapd;
7354
7355 pgdat_kswapd_lock(pgdat);
7356 kswapd = pgdat->kswapd;
7357 if (kswapd) {
7358 kthread_stop(kswapd);
7359 pgdat->kswapd = NULL;
7360 }
7361 pgdat_kswapd_unlock(pgdat);
7362 }
7363
7364 static int __init kswapd_init(void)
7365 {
7366 int nid;
7367
7368 swap_setup();
7369 for_each_node_state(nid, N_MEMORY)
7370 kswapd_run(nid);
7371 return 0;
7372 }
7373
7374 module_init(kswapd_init)
7375
7376 #ifdef CONFIG_NUMA
7377 /*
7378 * Node reclaim mode
7379 *
7380 * If non-zero call node_reclaim when the number of free pages falls below
7381 * the watermarks.
7382 */
7383 int node_reclaim_mode __read_mostly;
7384
7385 /*
7386 * Priority for NODE_RECLAIM. This determines the fraction of pages
7387 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7388 * a zone.
7389 */
7390 #define NODE_RECLAIM_PRIORITY 4
7391
7392 /*
7393 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7394 * occur.
7395 */
7396 int sysctl_min_unmapped_ratio = 1;
7397
7398 /*
7399 * If the number of slab pages in a zone grows beyond this percentage then
7400 * slab reclaim needs to occur.
7401 */
7402 int sysctl_min_slab_ratio = 5;
7403
7404 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7405 {
7406 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7407 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7408 node_page_state(pgdat, NR_ACTIVE_FILE);
7409
7410 /*
7411 * It's possible for there to be more file mapped pages than
7412 * accounted for by the pages on the file LRU lists because
7413 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7414 */
7415 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7416 }
7417
7418 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
7419 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7420 {
7421 unsigned long nr_pagecache_reclaimable;
7422 unsigned long delta = 0;
7423
7424 /*
7425 * If RECLAIM_UNMAP is set, then all file pages are considered
7426 * potentially reclaimable. Otherwise, we have to worry about
7427 * pages like swapcache and node_unmapped_file_pages() provides
7428 * a better estimate
7429 */
7430 if (node_reclaim_mode & RECLAIM_UNMAP)
7431 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7432 else
7433 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7434
7435 /* If we can't clean pages, remove dirty pages from consideration */
7436 if (!(node_reclaim_mode & RECLAIM_WRITE))
7437 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7438
7439 /* Watch for any possible underflows due to delta */
7440 if (unlikely(delta > nr_pagecache_reclaimable))
7441 delta = nr_pagecache_reclaimable;
7442
7443 return nr_pagecache_reclaimable - delta;
7444 }
7445
7446 /*
7447 * Try to free up some pages from this node through reclaim.
7448 */
7449 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7450 {
7451 /* Minimum pages needed in order to stay on node */
7452 const unsigned long nr_pages = 1 << order;
7453 struct task_struct *p = current;
7454 unsigned int noreclaim_flag;
7455 struct scan_control sc = {
7456 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7457 .gfp_mask = current_gfp_context(gfp_mask),
7458 .order = order,
7459 .priority = NODE_RECLAIM_PRIORITY,
7460 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7461 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7462 .may_swap = 1,
7463 .reclaim_idx = gfp_zone(gfp_mask),
7464 };
7465 unsigned long pflags;
7466
7467 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7468 sc.gfp_mask);
7469
7470 cond_resched();
7471 psi_memstall_enter(&pflags);
7472 delayacct_freepages_start();
7473 fs_reclaim_acquire(sc.gfp_mask);
7474 /*
7475 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7476 */
7477 noreclaim_flag = memalloc_noreclaim_save();
7478 set_task_reclaim_state(p, &sc.reclaim_state);
7479
7480 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7481 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7482 /*
7483 * Free memory by calling shrink node with increasing
7484 * priorities until we have enough memory freed.
7485 */
7486 do {
7487 shrink_node(pgdat, &sc);
7488 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7489 }
7490
7491 set_task_reclaim_state(p, NULL);
7492 memalloc_noreclaim_restore(noreclaim_flag);
7493 fs_reclaim_release(sc.gfp_mask);
7494 psi_memstall_leave(&pflags);
7495 delayacct_freepages_end();
7496
7497 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7498
7499 return sc.nr_reclaimed >= nr_pages;
7500 }
7501
7502 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7503 {
7504 int ret;
7505
7506 /*
7507 * Node reclaim reclaims unmapped file backed pages and
7508 * slab pages if we are over the defined limits.
7509 *
7510 * A small portion of unmapped file backed pages is needed for
7511 * file I/O otherwise pages read by file I/O will be immediately
7512 * thrown out if the node is overallocated. So we do not reclaim
7513 * if less than a specified percentage of the node is used by
7514 * unmapped file backed pages.
7515 */
7516 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7517 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7518 pgdat->min_slab_pages)
7519 return NODE_RECLAIM_FULL;
7520
7521 /*
7522 * Do not scan if the allocation should not be delayed.
7523 */
7524 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7525 return NODE_RECLAIM_NOSCAN;
7526
7527 /*
7528 * Only run node reclaim on the local node or on nodes that do not
7529 * have associated processors. This will favor the local processor
7530 * over remote processors and spread off node memory allocations
7531 * as wide as possible.
7532 */
7533 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7534 return NODE_RECLAIM_NOSCAN;
7535
7536 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7537 return NODE_RECLAIM_NOSCAN;
7538
7539 ret = __node_reclaim(pgdat, gfp_mask, order);
7540 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7541
7542 if (!ret)
7543 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7544
7545 return ret;
7546 }
7547 #endif
7548
7549 /**
7550 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7551 * lru list
7552 * @fbatch: Batch of lru folios to check.
7553 *
7554 * Checks folios for evictability, if an evictable folio is in the unevictable
7555 * lru list, moves it to the appropriate evictable lru list. This function
7556 * should be only used for lru folios.
7557 */
7558 void check_move_unevictable_folios(struct folio_batch *fbatch)
7559 {
7560 struct lruvec *lruvec = NULL;
7561 int pgscanned = 0;
7562 int pgrescued = 0;
7563 int i;
7564
7565 for (i = 0; i < fbatch->nr; i++) {
7566 struct folio *folio = fbatch->folios[i];
7567 int nr_pages = folio_nr_pages(folio);
7568
7569 pgscanned += nr_pages;
7570
7571 /* block memcg migration while the folio moves between lrus */
7572 if (!folio_test_clear_lru(folio))
7573 continue;
7574
7575 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7576 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7577 lruvec_del_folio(lruvec, folio);
7578 folio_clear_unevictable(folio);
7579 lruvec_add_folio(lruvec, folio);
7580 pgrescued += nr_pages;
7581 }
7582 folio_set_lru(folio);
7583 }
7584
7585 if (lruvec) {
7586 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7587 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7588 unlock_page_lruvec_irq(lruvec);
7589 } else if (pgscanned) {
7590 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7591 }
7592 }
7593 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
7594
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.