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