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