Loading...
1/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16#include <linux/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 try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
37#include <linux/compaction.h>
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
40#include <linux/delay.h>
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43#include <linux/memcontrol.h>
44#include <linux/delayacct.h>
45#include <linux/sysctl.h>
46#include <linux/oom.h>
47#include <linux/prefetch.h>
48#include <linux/printk.h>
49#include <linux/dax.h>
50
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
55#include <linux/balloon_compaction.h>
56
57#include "internal.h"
58
59#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
62struct scan_control {
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
66 /* This context's GFP mask */
67 gfp_t gfp_mask;
68
69 /* Allocation order */
70 int order;
71
72 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
77
78 /*
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
83
84 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
87 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
98 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
108};
109
110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
151#ifdef CONFIG_MEMCG
152static bool global_reclaim(struct scan_control *sc)
153{
154 return !sc->target_mem_cgroup;
155}
156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
178 return true;
179#endif
180 return false;
181}
182#else
183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
192#endif
193
194static unsigned long zone_reclaimable_pages(struct zone *zone)
195{
196 unsigned long nr;
197
198 nr = zone_page_state_snapshot(zone, NR_ACTIVE_FILE) +
199 zone_page_state_snapshot(zone, NR_INACTIVE_FILE) +
200 zone_page_state_snapshot(zone, NR_ISOLATED_FILE);
201
202 if (get_nr_swap_pages() > 0)
203 nr += zone_page_state_snapshot(zone, NR_ACTIVE_ANON) +
204 zone_page_state_snapshot(zone, NR_INACTIVE_ANON) +
205 zone_page_state_snapshot(zone, NR_ISOLATED_ANON);
206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
212 return zone_page_state_snapshot(zone, NR_PAGES_SCANNED) <
213 zone_reclaimable_pages(zone) * 6;
214}
215
216unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
217{
218 if (!mem_cgroup_disabled())
219 return mem_cgroup_get_lru_size(lruvec, lru);
220
221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
222}
223
224/*
225 * Add a shrinker callback to be called from the vm.
226 */
227int register_shrinker(struct shrinker *shrinker)
228{
229 size_t size = sizeof(*shrinker->nr_deferred);
230
231 if (shrinker->flags & SHRINKER_NUMA_AWARE)
232 size *= nr_node_ids;
233
234 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
235 if (!shrinker->nr_deferred)
236 return -ENOMEM;
237
238 down_write(&shrinker_rwsem);
239 list_add_tail(&shrinker->list, &shrinker_list);
240 up_write(&shrinker_rwsem);
241 return 0;
242}
243EXPORT_SYMBOL(register_shrinker);
244
245/*
246 * Remove one
247 */
248void unregister_shrinker(struct shrinker *shrinker)
249{
250 down_write(&shrinker_rwsem);
251 list_del(&shrinker->list);
252 up_write(&shrinker_rwsem);
253 kfree(shrinker->nr_deferred);
254}
255EXPORT_SYMBOL(unregister_shrinker);
256
257#define SHRINK_BATCH 128
258
259static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
260 struct shrinker *shrinker,
261 unsigned long nr_scanned,
262 unsigned long nr_eligible)
263{
264 unsigned long freed = 0;
265 unsigned long long delta;
266 long total_scan;
267 long freeable;
268 long nr;
269 long new_nr;
270 int nid = shrinkctl->nid;
271 long batch_size = shrinker->batch ? shrinker->batch
272 : SHRINK_BATCH;
273
274 freeable = shrinker->count_objects(shrinker, shrinkctl);
275 if (freeable == 0)
276 return 0;
277
278 /*
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
282 */
283 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
284
285 total_scan = nr;
286 delta = (4 * nr_scanned) / shrinker->seeks;
287 delta *= freeable;
288 do_div(delta, nr_eligible + 1);
289 total_scan += delta;
290 if (total_scan < 0) {
291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
292 shrinker->scan_objects, total_scan);
293 total_scan = freeable;
294 }
295
296 /*
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
302 * freeable. This is bad for sustaining a working set in
303 * memory.
304 *
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
307 */
308 if (delta < freeable / 4)
309 total_scan = min(total_scan, freeable / 2);
310
311 /*
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
314 * freeable entries.
315 */
316 if (total_scan > freeable * 2)
317 total_scan = freeable * 2;
318
319 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
320 nr_scanned, nr_eligible,
321 freeable, delta, total_scan);
322
323 /*
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
330 * batch_size.
331 *
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
334 * than the total number of objects on slab (freeable), we must be
335 * scanning at high prio and therefore should try to reclaim as much as
336 * possible.
337 */
338 while (total_scan >= batch_size ||
339 total_scan >= freeable) {
340 unsigned long ret;
341 unsigned long nr_to_scan = min(batch_size, total_scan);
342
343 shrinkctl->nr_to_scan = nr_to_scan;
344 ret = shrinker->scan_objects(shrinker, shrinkctl);
345 if (ret == SHRINK_STOP)
346 break;
347 freed += ret;
348
349 count_vm_events(SLABS_SCANNED, nr_to_scan);
350 total_scan -= nr_to_scan;
351
352 cond_resched();
353 }
354
355 /*
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
359 */
360 if (total_scan > 0)
361 new_nr = atomic_long_add_return(total_scan,
362 &shrinker->nr_deferred[nid]);
363 else
364 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
365
366 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
367 return freed;
368}
369
370/**
371 * shrink_slab - shrink slab caches
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
374 * @memcg: memory cgroup whose slab caches to target
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
377 *
378 * Call the shrink functions to age shrinkable caches.
379 *
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
382 *
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
387 *
388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
395 *
396 * Returns the number of reclaimed slab objects.
397 */
398static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
399 struct mem_cgroup *memcg,
400 unsigned long nr_scanned,
401 unsigned long nr_eligible)
402{
403 struct shrinker *shrinker;
404 unsigned long freed = 0;
405
406 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
407 return 0;
408
409 if (nr_scanned == 0)
410 nr_scanned = SWAP_CLUSTER_MAX;
411
412 if (!down_read_trylock(&shrinker_rwsem)) {
413 /*
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
417 * time.
418 */
419 freed = 1;
420 goto out;
421 }
422
423 list_for_each_entry(shrinker, &shrinker_list, list) {
424 struct shrink_control sc = {
425 .gfp_mask = gfp_mask,
426 .nid = nid,
427 .memcg = memcg,
428 };
429
430 /*
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
434 */
435 if (memcg_kmem_enabled() &&
436 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
437 continue;
438
439 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
440 sc.nid = 0;
441
442 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
443 }
444
445 up_read(&shrinker_rwsem);
446out:
447 cond_resched();
448 return freed;
449}
450
451void drop_slab_node(int nid)
452{
453 unsigned long freed;
454
455 do {
456 struct mem_cgroup *memcg = NULL;
457
458 freed = 0;
459 do {
460 freed += shrink_slab(GFP_KERNEL, nid, memcg,
461 1000, 1000);
462 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
463 } while (freed > 10);
464}
465
466void drop_slab(void)
467{
468 int nid;
469
470 for_each_online_node(nid)
471 drop_slab_node(nid);
472}
473
474static inline int is_page_cache_freeable(struct page *page)
475{
476 /*
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
480 */
481 return page_count(page) - page_has_private(page) == 2;
482}
483
484static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
485{
486 if (current->flags & PF_SWAPWRITE)
487 return 1;
488 if (!inode_write_congested(inode))
489 return 1;
490 if (inode_to_bdi(inode) == current->backing_dev_info)
491 return 1;
492 return 0;
493}
494
495/*
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
499 *
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
503 *
504 * We're allowed to run sleeping lock_page() here because we know the caller has
505 * __GFP_FS.
506 */
507static void handle_write_error(struct address_space *mapping,
508 struct page *page, int error)
509{
510 lock_page(page);
511 if (page_mapping(page) == mapping)
512 mapping_set_error(mapping, error);
513 unlock_page(page);
514}
515
516/* possible outcome of pageout() */
517typedef enum {
518 /* failed to write page out, page is locked */
519 PAGE_KEEP,
520 /* move page to the active list, page is locked */
521 PAGE_ACTIVATE,
522 /* page has been sent to the disk successfully, page is unlocked */
523 PAGE_SUCCESS,
524 /* page is clean and locked */
525 PAGE_CLEAN,
526} pageout_t;
527
528/*
529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
531 */
532static pageout_t pageout(struct page *page, struct address_space *mapping,
533 struct scan_control *sc)
534{
535 /*
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
541 *
542 * If this process is currently in __generic_file_write_iter() against
543 * this page's queue, we can perform writeback even if that
544 * will block.
545 *
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
550 */
551 if (!is_page_cache_freeable(page))
552 return PAGE_KEEP;
553 if (!mapping) {
554 /*
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
557 */
558 if (page_has_private(page)) {
559 if (try_to_free_buffers(page)) {
560 ClearPageDirty(page);
561 pr_info("%s: orphaned page\n", __func__);
562 return PAGE_CLEAN;
563 }
564 }
565 return PAGE_KEEP;
566 }
567 if (mapping->a_ops->writepage == NULL)
568 return PAGE_ACTIVATE;
569 if (!may_write_to_inode(mapping->host, sc))
570 return PAGE_KEEP;
571
572 if (clear_page_dirty_for_io(page)) {
573 int res;
574 struct writeback_control wbc = {
575 .sync_mode = WB_SYNC_NONE,
576 .nr_to_write = SWAP_CLUSTER_MAX,
577 .range_start = 0,
578 .range_end = LLONG_MAX,
579 .for_reclaim = 1,
580 };
581
582 SetPageReclaim(page);
583 res = mapping->a_ops->writepage(page, &wbc);
584 if (res < 0)
585 handle_write_error(mapping, page, res);
586 if (res == AOP_WRITEPAGE_ACTIVATE) {
587 ClearPageReclaim(page);
588 return PAGE_ACTIVATE;
589 }
590
591 if (!PageWriteback(page)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page);
594 }
595 trace_mm_vmscan_writepage(page);
596 inc_zone_page_state(page, NR_VMSCAN_WRITE);
597 return PAGE_SUCCESS;
598 }
599
600 return PAGE_CLEAN;
601}
602
603/*
604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
606 */
607static int __remove_mapping(struct address_space *mapping, struct page *page,
608 bool reclaimed)
609{
610 unsigned long flags;
611
612 BUG_ON(!PageLocked(page));
613 BUG_ON(mapping != page_mapping(page));
614
615 spin_lock_irqsave(&mapping->tree_lock, flags);
616 /*
617 * The non racy check for a busy page.
618 *
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
623 *
624 * get_user_pages(&page);
625 * [user mapping goes away]
626 * write_to(page);
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
629 * put_page(page);
630 * !page_count(page) [good, discard it]
631 *
632 * [oops, our write_to data is lost]
633 *
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
636 * load is not satisfied before that of page->_count.
637 *
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
640 */
641 if (!page_ref_freeze(page, 2))
642 goto cannot_free;
643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page))) {
645 page_ref_unfreeze(page, 2);
646 goto cannot_free;
647 }
648
649 if (PageSwapCache(page)) {
650 swp_entry_t swap = { .val = page_private(page) };
651 mem_cgroup_swapout(page, swap);
652 __delete_from_swap_cache(page);
653 spin_unlock_irqrestore(&mapping->tree_lock, flags);
654 swapcache_free(swap);
655 } else {
656 void (*freepage)(struct page *);
657 void *shadow = NULL;
658
659 freepage = mapping->a_ops->freepage;
660 /*
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
663 *
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
668 * back.
669 *
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
674 * same page_tree.
675 */
676 if (reclaimed && page_is_file_cache(page) &&
677 !mapping_exiting(mapping) && !dax_mapping(mapping))
678 shadow = workingset_eviction(mapping, page);
679 __delete_from_page_cache(page, shadow);
680 spin_unlock_irqrestore(&mapping->tree_lock, flags);
681
682 if (freepage != NULL)
683 freepage(page);
684 }
685
686 return 1;
687
688cannot_free:
689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
690 return 0;
691}
692
693/*
694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
697 * this page.
698 */
699int remove_mapping(struct address_space *mapping, struct page *page)
700{
701 if (__remove_mapping(mapping, page, false)) {
702 /*
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
705 * atomic operation.
706 */
707 page_ref_unfreeze(page, 1);
708 return 1;
709 }
710 return 0;
711}
712
713/**
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
716 *
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
719 *
720 * lru_lock must not be held, interrupts must be enabled.
721 */
722void putback_lru_page(struct page *page)
723{
724 bool is_unevictable;
725 int was_unevictable = PageUnevictable(page);
726
727 VM_BUG_ON_PAGE(PageLRU(page), page);
728
729redo:
730 ClearPageUnevictable(page);
731
732 if (page_evictable(page)) {
733 /*
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
738 */
739 is_unevictable = false;
740 lru_cache_add(page);
741 } else {
742 /*
743 * Put unevictable pages directly on zone's unevictable
744 * list.
745 */
746 is_unevictable = true;
747 add_page_to_unevictable_list(page);
748 /*
749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
752 * isolation/check_move_unevictable_pages,
753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
754 * the page back to the evictable list.
755 *
756 * The other side is TestClearPageMlocked() or shmem_lock().
757 */
758 smp_mb();
759 }
760
761 /*
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
765 */
766 if (is_unevictable && page_evictable(page)) {
767 if (!isolate_lru_page(page)) {
768 put_page(page);
769 goto redo;
770 }
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
774 */
775 }
776
777 if (was_unevictable && !is_unevictable)
778 count_vm_event(UNEVICTABLE_PGRESCUED);
779 else if (!was_unevictable && is_unevictable)
780 count_vm_event(UNEVICTABLE_PGCULLED);
781
782 put_page(page); /* drop ref from isolate */
783}
784
785enum page_references {
786 PAGEREF_RECLAIM,
787 PAGEREF_RECLAIM_CLEAN,
788 PAGEREF_KEEP,
789 PAGEREF_ACTIVATE,
790};
791
792static enum page_references page_check_references(struct page *page,
793 struct scan_control *sc)
794{
795 int referenced_ptes, referenced_page;
796 unsigned long vm_flags;
797
798 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
799 &vm_flags);
800 referenced_page = TestClearPageReferenced(page);
801
802 /*
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
805 */
806 if (vm_flags & VM_LOCKED)
807 return PAGEREF_RECLAIM;
808
809 if (referenced_ptes) {
810 if (PageSwapBacked(page))
811 return PAGEREF_ACTIVATE;
812 /*
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
816 * than once.
817 *
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
821 *
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
824 * quickly recovered.
825 */
826 SetPageReferenced(page);
827
828 if (referenced_page || referenced_ptes > 1)
829 return PAGEREF_ACTIVATE;
830
831 /*
832 * Activate file-backed executable pages after first usage.
833 */
834 if (vm_flags & VM_EXEC)
835 return PAGEREF_ACTIVATE;
836
837 return PAGEREF_KEEP;
838 }
839
840 /* Reclaim if clean, defer dirty pages to writeback */
841 if (referenced_page && !PageSwapBacked(page))
842 return PAGEREF_RECLAIM_CLEAN;
843
844 return PAGEREF_RECLAIM;
845}
846
847/* Check if a page is dirty or under writeback */
848static void page_check_dirty_writeback(struct page *page,
849 bool *dirty, bool *writeback)
850{
851 struct address_space *mapping;
852
853 /*
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
856 */
857 if (!page_is_file_cache(page)) {
858 *dirty = false;
859 *writeback = false;
860 return;
861 }
862
863 /* By default assume that the page flags are accurate */
864 *dirty = PageDirty(page);
865 *writeback = PageWriteback(page);
866
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page))
869 return;
870
871 mapping = page_mapping(page);
872 if (mapping && mapping->a_ops->is_dirty_writeback)
873 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
874}
875
876/*
877 * shrink_page_list() returns the number of reclaimed pages
878 */
879static unsigned long shrink_page_list(struct list_head *page_list,
880 struct zone *zone,
881 struct scan_control *sc,
882 enum ttu_flags ttu_flags,
883 unsigned long *ret_nr_dirty,
884 unsigned long *ret_nr_unqueued_dirty,
885 unsigned long *ret_nr_congested,
886 unsigned long *ret_nr_writeback,
887 unsigned long *ret_nr_immediate,
888 bool force_reclaim)
889{
890 LIST_HEAD(ret_pages);
891 LIST_HEAD(free_pages);
892 int pgactivate = 0;
893 unsigned long nr_unqueued_dirty = 0;
894 unsigned long nr_dirty = 0;
895 unsigned long nr_congested = 0;
896 unsigned long nr_reclaimed = 0;
897 unsigned long nr_writeback = 0;
898 unsigned long nr_immediate = 0;
899
900 cond_resched();
901
902 while (!list_empty(page_list)) {
903 struct address_space *mapping;
904 struct page *page;
905 int may_enter_fs;
906 enum page_references references = PAGEREF_RECLAIM_CLEAN;
907 bool dirty, writeback;
908 bool lazyfree = false;
909 int ret = SWAP_SUCCESS;
910
911 cond_resched();
912
913 page = lru_to_page(page_list);
914 list_del(&page->lru);
915
916 if (!trylock_page(page))
917 goto keep;
918
919 VM_BUG_ON_PAGE(PageActive(page), page);
920 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
921
922 sc->nr_scanned++;
923
924 if (unlikely(!page_evictable(page)))
925 goto cull_mlocked;
926
927 if (!sc->may_unmap && page_mapped(page))
928 goto keep_locked;
929
930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page) || PageSwapCache(page))
932 sc->nr_scanned++;
933
934 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
935 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
936
937 /*
938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
942 */
943 page_check_dirty_writeback(page, &dirty, &writeback);
944 if (dirty || writeback)
945 nr_dirty++;
946
947 if (dirty && !writeback)
948 nr_unqueued_dirty++;
949
950 /*
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
955 */
956 mapping = page_mapping(page);
957 if (((dirty || writeback) && mapping &&
958 inode_write_congested(mapping->host)) ||
959 (writeback && PageReclaim(page)))
960 nr_congested++;
961
962 /*
963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
965 *
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
975 *
976 * 2) Global or new memcg reclaim encounters a page that is
977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
980 * reclaim and continue scanning.
981 *
982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
988 *
989 * 3) Legacy memcg encounters a page that is already marked
990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
994 */
995 if (PageWriteback(page)) {
996 /* Case 1 above */
997 if (current_is_kswapd() &&
998 PageReclaim(page) &&
999 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1000 nr_immediate++;
1001 goto keep_locked;
1002
1003 /* Case 2 above */
1004 } else if (sane_reclaim(sc) ||
1005 !PageReclaim(page) || !may_enter_fs) {
1006 /*
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1016 */
1017 SetPageReclaim(page);
1018 nr_writeback++;
1019 goto keep_locked;
1020
1021 /* Case 3 above */
1022 } else {
1023 unlock_page(page);
1024 wait_on_page_writeback(page);
1025 /* then go back and try same page again */
1026 list_add_tail(&page->lru, page_list);
1027 continue;
1028 }
1029 }
1030
1031 if (!force_reclaim)
1032 references = page_check_references(page, sc);
1033
1034 switch (references) {
1035 case PAGEREF_ACTIVATE:
1036 goto activate_locked;
1037 case PAGEREF_KEEP:
1038 goto keep_locked;
1039 case PAGEREF_RECLAIM:
1040 case PAGEREF_RECLAIM_CLEAN:
1041 ; /* try to reclaim the page below */
1042 }
1043
1044 /*
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1047 */
1048 if (PageAnon(page) && !PageSwapCache(page)) {
1049 if (!(sc->gfp_mask & __GFP_IO))
1050 goto keep_locked;
1051 if (!add_to_swap(page, page_list))
1052 goto activate_locked;
1053 lazyfree = true;
1054 may_enter_fs = 1;
1055
1056 /* Adding to swap updated mapping */
1057 mapping = page_mapping(page);
1058 }
1059
1060 /*
1061 * The page is mapped into the page tables of one or more
1062 * processes. Try to unmap it here.
1063 */
1064 if (page_mapped(page) && mapping) {
1065 switch (ret = try_to_unmap(page, lazyfree ?
1066 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1067 (ttu_flags | TTU_BATCH_FLUSH))) {
1068 case SWAP_FAIL:
1069 goto activate_locked;
1070 case SWAP_AGAIN:
1071 goto keep_locked;
1072 case SWAP_MLOCK:
1073 goto cull_mlocked;
1074 case SWAP_LZFREE:
1075 goto lazyfree;
1076 case SWAP_SUCCESS:
1077 ; /* try to free the page below */
1078 }
1079 }
1080
1081 if (PageDirty(page)) {
1082 /*
1083 * Only kswapd can writeback filesystem pages to
1084 * avoid risk of stack overflow but only writeback
1085 * if many dirty pages have been encountered.
1086 */
1087 if (page_is_file_cache(page) &&
1088 (!current_is_kswapd() ||
1089 !test_bit(ZONE_DIRTY, &zone->flags))) {
1090 /*
1091 * Immediately reclaim when written back.
1092 * Similar in principal to deactivate_page()
1093 * except we already have the page isolated
1094 * and know it's dirty
1095 */
1096 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1097 SetPageReclaim(page);
1098
1099 goto keep_locked;
1100 }
1101
1102 if (references == PAGEREF_RECLAIM_CLEAN)
1103 goto keep_locked;
1104 if (!may_enter_fs)
1105 goto keep_locked;
1106 if (!sc->may_writepage)
1107 goto keep_locked;
1108
1109 /*
1110 * Page is dirty. Flush the TLB if a writable entry
1111 * potentially exists to avoid CPU writes after IO
1112 * starts and then write it out here.
1113 */
1114 try_to_unmap_flush_dirty();
1115 switch (pageout(page, mapping, sc)) {
1116 case PAGE_KEEP:
1117 goto keep_locked;
1118 case PAGE_ACTIVATE:
1119 goto activate_locked;
1120 case PAGE_SUCCESS:
1121 if (PageWriteback(page))
1122 goto keep;
1123 if (PageDirty(page))
1124 goto keep;
1125
1126 /*
1127 * A synchronous write - probably a ramdisk. Go
1128 * ahead and try to reclaim the page.
1129 */
1130 if (!trylock_page(page))
1131 goto keep;
1132 if (PageDirty(page) || PageWriteback(page))
1133 goto keep_locked;
1134 mapping = page_mapping(page);
1135 case PAGE_CLEAN:
1136 ; /* try to free the page below */
1137 }
1138 }
1139
1140 /*
1141 * If the page has buffers, try to free the buffer mappings
1142 * associated with this page. If we succeed we try to free
1143 * the page as well.
1144 *
1145 * We do this even if the page is PageDirty().
1146 * try_to_release_page() does not perform I/O, but it is
1147 * possible for a page to have PageDirty set, but it is actually
1148 * clean (all its buffers are clean). This happens if the
1149 * buffers were written out directly, with submit_bh(). ext3
1150 * will do this, as well as the blockdev mapping.
1151 * try_to_release_page() will discover that cleanness and will
1152 * drop the buffers and mark the page clean - it can be freed.
1153 *
1154 * Rarely, pages can have buffers and no ->mapping. These are
1155 * the pages which were not successfully invalidated in
1156 * truncate_complete_page(). We try to drop those buffers here
1157 * and if that worked, and the page is no longer mapped into
1158 * process address space (page_count == 1) it can be freed.
1159 * Otherwise, leave the page on the LRU so it is swappable.
1160 */
1161 if (page_has_private(page)) {
1162 if (!try_to_release_page(page, sc->gfp_mask))
1163 goto activate_locked;
1164 if (!mapping && page_count(page) == 1) {
1165 unlock_page(page);
1166 if (put_page_testzero(page))
1167 goto free_it;
1168 else {
1169 /*
1170 * rare race with speculative reference.
1171 * the speculative reference will free
1172 * this page shortly, so we may
1173 * increment nr_reclaimed here (and
1174 * leave it off the LRU).
1175 */
1176 nr_reclaimed++;
1177 continue;
1178 }
1179 }
1180 }
1181
1182lazyfree:
1183 if (!mapping || !__remove_mapping(mapping, page, true))
1184 goto keep_locked;
1185
1186 /*
1187 * At this point, we have no other references and there is
1188 * no way to pick any more up (removed from LRU, removed
1189 * from pagecache). Can use non-atomic bitops now (and
1190 * we obviously don't have to worry about waking up a process
1191 * waiting on the page lock, because there are no references.
1192 */
1193 __ClearPageLocked(page);
1194free_it:
1195 if (ret == SWAP_LZFREE)
1196 count_vm_event(PGLAZYFREED);
1197
1198 nr_reclaimed++;
1199
1200 /*
1201 * Is there need to periodically free_page_list? It would
1202 * appear not as the counts should be low
1203 */
1204 list_add(&page->lru, &free_pages);
1205 continue;
1206
1207cull_mlocked:
1208 if (PageSwapCache(page))
1209 try_to_free_swap(page);
1210 unlock_page(page);
1211 list_add(&page->lru, &ret_pages);
1212 continue;
1213
1214activate_locked:
1215 /* Not a candidate for swapping, so reclaim swap space. */
1216 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
1217 try_to_free_swap(page);
1218 VM_BUG_ON_PAGE(PageActive(page), page);
1219 SetPageActive(page);
1220 pgactivate++;
1221keep_locked:
1222 unlock_page(page);
1223keep:
1224 list_add(&page->lru, &ret_pages);
1225 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1226 }
1227
1228 mem_cgroup_uncharge_list(&free_pages);
1229 try_to_unmap_flush();
1230 free_hot_cold_page_list(&free_pages, true);
1231
1232 list_splice(&ret_pages, page_list);
1233 count_vm_events(PGACTIVATE, pgactivate);
1234
1235 *ret_nr_dirty += nr_dirty;
1236 *ret_nr_congested += nr_congested;
1237 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1238 *ret_nr_writeback += nr_writeback;
1239 *ret_nr_immediate += nr_immediate;
1240 return nr_reclaimed;
1241}
1242
1243unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1244 struct list_head *page_list)
1245{
1246 struct scan_control sc = {
1247 .gfp_mask = GFP_KERNEL,
1248 .priority = DEF_PRIORITY,
1249 .may_unmap = 1,
1250 };
1251 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1252 struct page *page, *next;
1253 LIST_HEAD(clean_pages);
1254
1255 list_for_each_entry_safe(page, next, page_list, lru) {
1256 if (page_is_file_cache(page) && !PageDirty(page) &&
1257 !isolated_balloon_page(page)) {
1258 ClearPageActive(page);
1259 list_move(&page->lru, &clean_pages);
1260 }
1261 }
1262
1263 ret = shrink_page_list(&clean_pages, zone, &sc,
1264 TTU_UNMAP|TTU_IGNORE_ACCESS,
1265 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1266 list_splice(&clean_pages, page_list);
1267 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1268 return ret;
1269}
1270
1271/*
1272 * Attempt to remove the specified page from its LRU. Only take this page
1273 * if it is of the appropriate PageActive status. Pages which are being
1274 * freed elsewhere are also ignored.
1275 *
1276 * page: page to consider
1277 * mode: one of the LRU isolation modes defined above
1278 *
1279 * returns 0 on success, -ve errno on failure.
1280 */
1281int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1282{
1283 int ret = -EINVAL;
1284
1285 /* Only take pages on the LRU. */
1286 if (!PageLRU(page))
1287 return ret;
1288
1289 /* Compaction should not handle unevictable pages but CMA can do so */
1290 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1291 return ret;
1292
1293 ret = -EBUSY;
1294
1295 /*
1296 * To minimise LRU disruption, the caller can indicate that it only
1297 * wants to isolate pages it will be able to operate on without
1298 * blocking - clean pages for the most part.
1299 *
1300 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1301 * is used by reclaim when it is cannot write to backing storage
1302 *
1303 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1304 * that it is possible to migrate without blocking
1305 */
1306 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1307 /* All the caller can do on PageWriteback is block */
1308 if (PageWriteback(page))
1309 return ret;
1310
1311 if (PageDirty(page)) {
1312 struct address_space *mapping;
1313
1314 /* ISOLATE_CLEAN means only clean pages */
1315 if (mode & ISOLATE_CLEAN)
1316 return ret;
1317
1318 /*
1319 * Only pages without mappings or that have a
1320 * ->migratepage callback are possible to migrate
1321 * without blocking
1322 */
1323 mapping = page_mapping(page);
1324 if (mapping && !mapping->a_ops->migratepage)
1325 return ret;
1326 }
1327 }
1328
1329 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 return ret;
1331
1332 if (likely(get_page_unless_zero(page))) {
1333 /*
1334 * Be careful not to clear PageLRU until after we're
1335 * sure the page is not being freed elsewhere -- the
1336 * page release code relies on it.
1337 */
1338 ClearPageLRU(page);
1339 ret = 0;
1340 }
1341
1342 return ret;
1343}
1344
1345/*
1346 * zone->lru_lock is heavily contended. Some of the functions that
1347 * shrink the lists perform better by taking out a batch of pages
1348 * and working on them outside the LRU lock.
1349 *
1350 * For pagecache intensive workloads, this function is the hottest
1351 * spot in the kernel (apart from copy_*_user functions).
1352 *
1353 * Appropriate locks must be held before calling this function.
1354 *
1355 * @nr_to_scan: The number of pages to look through on the list.
1356 * @lruvec: The LRU vector to pull pages from.
1357 * @dst: The temp list to put pages on to.
1358 * @nr_scanned: The number of pages that were scanned.
1359 * @sc: The scan_control struct for this reclaim session
1360 * @mode: One of the LRU isolation modes
1361 * @lru: LRU list id for isolating
1362 *
1363 * returns how many pages were moved onto *@dst.
1364 */
1365static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1366 struct lruvec *lruvec, struct list_head *dst,
1367 unsigned long *nr_scanned, struct scan_control *sc,
1368 isolate_mode_t mode, enum lru_list lru)
1369{
1370 struct list_head *src = &lruvec->lists[lru];
1371 unsigned long nr_taken = 0;
1372 unsigned long scan;
1373
1374 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1375 !list_empty(src); scan++) {
1376 struct page *page;
1377 int nr_pages;
1378
1379 page = lru_to_page(src);
1380 prefetchw_prev_lru_page(page, src, flags);
1381
1382 VM_BUG_ON_PAGE(!PageLRU(page), page);
1383
1384 switch (__isolate_lru_page(page, mode)) {
1385 case 0:
1386 nr_pages = hpage_nr_pages(page);
1387 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1388 list_move(&page->lru, dst);
1389 nr_taken += nr_pages;
1390 break;
1391
1392 case -EBUSY:
1393 /* else it is being freed elsewhere */
1394 list_move(&page->lru, src);
1395 continue;
1396
1397 default:
1398 BUG();
1399 }
1400 }
1401
1402 *nr_scanned = scan;
1403 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1404 nr_taken, mode, is_file_lru(lru));
1405 return nr_taken;
1406}
1407
1408/**
1409 * isolate_lru_page - tries to isolate a page from its LRU list
1410 * @page: page to isolate from its LRU list
1411 *
1412 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1413 * vmstat statistic corresponding to whatever LRU list the page was on.
1414 *
1415 * Returns 0 if the page was removed from an LRU list.
1416 * Returns -EBUSY if the page was not on an LRU list.
1417 *
1418 * The returned page will have PageLRU() cleared. If it was found on
1419 * the active list, it will have PageActive set. If it was found on
1420 * the unevictable list, it will have the PageUnevictable bit set. That flag
1421 * may need to be cleared by the caller before letting the page go.
1422 *
1423 * The vmstat statistic corresponding to the list on which the page was
1424 * found will be decremented.
1425 *
1426 * Restrictions:
1427 * (1) Must be called with an elevated refcount on the page. This is a
1428 * fundamentnal difference from isolate_lru_pages (which is called
1429 * without a stable reference).
1430 * (2) the lru_lock must not be held.
1431 * (3) interrupts must be enabled.
1432 */
1433int isolate_lru_page(struct page *page)
1434{
1435 int ret = -EBUSY;
1436
1437 VM_BUG_ON_PAGE(!page_count(page), page);
1438 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1439
1440 if (PageLRU(page)) {
1441 struct zone *zone = page_zone(page);
1442 struct lruvec *lruvec;
1443
1444 spin_lock_irq(&zone->lru_lock);
1445 lruvec = mem_cgroup_page_lruvec(page, zone);
1446 if (PageLRU(page)) {
1447 int lru = page_lru(page);
1448 get_page(page);
1449 ClearPageLRU(page);
1450 del_page_from_lru_list(page, lruvec, lru);
1451 ret = 0;
1452 }
1453 spin_unlock_irq(&zone->lru_lock);
1454 }
1455 return ret;
1456}
1457
1458/*
1459 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1460 * then get resheduled. When there are massive number of tasks doing page
1461 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1462 * the LRU list will go small and be scanned faster than necessary, leading to
1463 * unnecessary swapping, thrashing and OOM.
1464 */
1465static int too_many_isolated(struct zone *zone, int file,
1466 struct scan_control *sc)
1467{
1468 unsigned long inactive, isolated;
1469
1470 if (current_is_kswapd())
1471 return 0;
1472
1473 if (!sane_reclaim(sc))
1474 return 0;
1475
1476 if (file) {
1477 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1478 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1479 } else {
1480 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1481 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1482 }
1483
1484 /*
1485 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1486 * won't get blocked by normal direct-reclaimers, forming a circular
1487 * deadlock.
1488 */
1489 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1490 inactive >>= 3;
1491
1492 return isolated > inactive;
1493}
1494
1495static noinline_for_stack void
1496putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1497{
1498 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1499 struct zone *zone = lruvec_zone(lruvec);
1500 LIST_HEAD(pages_to_free);
1501
1502 /*
1503 * Put back any unfreeable pages.
1504 */
1505 while (!list_empty(page_list)) {
1506 struct page *page = lru_to_page(page_list);
1507 int lru;
1508
1509 VM_BUG_ON_PAGE(PageLRU(page), page);
1510 list_del(&page->lru);
1511 if (unlikely(!page_evictable(page))) {
1512 spin_unlock_irq(&zone->lru_lock);
1513 putback_lru_page(page);
1514 spin_lock_irq(&zone->lru_lock);
1515 continue;
1516 }
1517
1518 lruvec = mem_cgroup_page_lruvec(page, zone);
1519
1520 SetPageLRU(page);
1521 lru = page_lru(page);
1522 add_page_to_lru_list(page, lruvec, lru);
1523
1524 if (is_active_lru(lru)) {
1525 int file = is_file_lru(lru);
1526 int numpages = hpage_nr_pages(page);
1527 reclaim_stat->recent_rotated[file] += numpages;
1528 }
1529 if (put_page_testzero(page)) {
1530 __ClearPageLRU(page);
1531 __ClearPageActive(page);
1532 del_page_from_lru_list(page, lruvec, lru);
1533
1534 if (unlikely(PageCompound(page))) {
1535 spin_unlock_irq(&zone->lru_lock);
1536 mem_cgroup_uncharge(page);
1537 (*get_compound_page_dtor(page))(page);
1538 spin_lock_irq(&zone->lru_lock);
1539 } else
1540 list_add(&page->lru, &pages_to_free);
1541 }
1542 }
1543
1544 /*
1545 * To save our caller's stack, now use input list for pages to free.
1546 */
1547 list_splice(&pages_to_free, page_list);
1548}
1549
1550/*
1551 * If a kernel thread (such as nfsd for loop-back mounts) services
1552 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1553 * In that case we should only throttle if the backing device it is
1554 * writing to is congested. In other cases it is safe to throttle.
1555 */
1556static int current_may_throttle(void)
1557{
1558 return !(current->flags & PF_LESS_THROTTLE) ||
1559 current->backing_dev_info == NULL ||
1560 bdi_write_congested(current->backing_dev_info);
1561}
1562
1563/*
1564 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1565 * of reclaimed pages
1566 */
1567static noinline_for_stack unsigned long
1568shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1569 struct scan_control *sc, enum lru_list lru)
1570{
1571 LIST_HEAD(page_list);
1572 unsigned long nr_scanned;
1573 unsigned long nr_reclaimed = 0;
1574 unsigned long nr_taken;
1575 unsigned long nr_dirty = 0;
1576 unsigned long nr_congested = 0;
1577 unsigned long nr_unqueued_dirty = 0;
1578 unsigned long nr_writeback = 0;
1579 unsigned long nr_immediate = 0;
1580 isolate_mode_t isolate_mode = 0;
1581 int file = is_file_lru(lru);
1582 struct zone *zone = lruvec_zone(lruvec);
1583 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1584
1585 while (unlikely(too_many_isolated(zone, file, sc))) {
1586 congestion_wait(BLK_RW_ASYNC, HZ/10);
1587
1588 /* We are about to die and free our memory. Return now. */
1589 if (fatal_signal_pending(current))
1590 return SWAP_CLUSTER_MAX;
1591 }
1592
1593 lru_add_drain();
1594
1595 if (!sc->may_unmap)
1596 isolate_mode |= ISOLATE_UNMAPPED;
1597 if (!sc->may_writepage)
1598 isolate_mode |= ISOLATE_CLEAN;
1599
1600 spin_lock_irq(&zone->lru_lock);
1601
1602 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1603 &nr_scanned, sc, isolate_mode, lru);
1604
1605 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1606 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1607
1608 if (global_reclaim(sc)) {
1609 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1610 if (current_is_kswapd())
1611 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1612 else
1613 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1614 }
1615 spin_unlock_irq(&zone->lru_lock);
1616
1617 if (nr_taken == 0)
1618 return 0;
1619
1620 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1621 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1622 &nr_writeback, &nr_immediate,
1623 false);
1624
1625 spin_lock_irq(&zone->lru_lock);
1626
1627 reclaim_stat->recent_scanned[file] += nr_taken;
1628
1629 if (global_reclaim(sc)) {
1630 if (current_is_kswapd())
1631 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1632 nr_reclaimed);
1633 else
1634 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1635 nr_reclaimed);
1636 }
1637
1638 putback_inactive_pages(lruvec, &page_list);
1639
1640 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1641
1642 spin_unlock_irq(&zone->lru_lock);
1643
1644 mem_cgroup_uncharge_list(&page_list);
1645 free_hot_cold_page_list(&page_list, true);
1646
1647 /*
1648 * If reclaim is isolating dirty pages under writeback, it implies
1649 * that the long-lived page allocation rate is exceeding the page
1650 * laundering rate. Either the global limits are not being effective
1651 * at throttling processes due to the page distribution throughout
1652 * zones or there is heavy usage of a slow backing device. The
1653 * only option is to throttle from reclaim context which is not ideal
1654 * as there is no guarantee the dirtying process is throttled in the
1655 * same way balance_dirty_pages() manages.
1656 *
1657 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1658 * of pages under pages flagged for immediate reclaim and stall if any
1659 * are encountered in the nr_immediate check below.
1660 */
1661 if (nr_writeback && nr_writeback == nr_taken)
1662 set_bit(ZONE_WRITEBACK, &zone->flags);
1663
1664 /*
1665 * Legacy memcg will stall in page writeback so avoid forcibly
1666 * stalling here.
1667 */
1668 if (sane_reclaim(sc)) {
1669 /*
1670 * Tag a zone as congested if all the dirty pages scanned were
1671 * backed by a congested BDI and wait_iff_congested will stall.
1672 */
1673 if (nr_dirty && nr_dirty == nr_congested)
1674 set_bit(ZONE_CONGESTED, &zone->flags);
1675
1676 /*
1677 * If dirty pages are scanned that are not queued for IO, it
1678 * implies that flushers are not keeping up. In this case, flag
1679 * the zone ZONE_DIRTY and kswapd will start writing pages from
1680 * reclaim context.
1681 */
1682 if (nr_unqueued_dirty == nr_taken)
1683 set_bit(ZONE_DIRTY, &zone->flags);
1684
1685 /*
1686 * If kswapd scans pages marked marked for immediate
1687 * reclaim and under writeback (nr_immediate), it implies
1688 * that pages are cycling through the LRU faster than
1689 * they are written so also forcibly stall.
1690 */
1691 if (nr_immediate && current_may_throttle())
1692 congestion_wait(BLK_RW_ASYNC, HZ/10);
1693 }
1694
1695 /*
1696 * Stall direct reclaim for IO completions if underlying BDIs or zone
1697 * is congested. Allow kswapd to continue until it starts encountering
1698 * unqueued dirty pages or cycling through the LRU too quickly.
1699 */
1700 if (!sc->hibernation_mode && !current_is_kswapd() &&
1701 current_may_throttle())
1702 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1703
1704 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1705 sc->priority, file);
1706 return nr_reclaimed;
1707}
1708
1709/*
1710 * This moves pages from the active list to the inactive list.
1711 *
1712 * We move them the other way if the page is referenced by one or more
1713 * processes, from rmap.
1714 *
1715 * If the pages are mostly unmapped, the processing is fast and it is
1716 * appropriate to hold zone->lru_lock across the whole operation. But if
1717 * the pages are mapped, the processing is slow (page_referenced()) so we
1718 * should drop zone->lru_lock around each page. It's impossible to balance
1719 * this, so instead we remove the pages from the LRU while processing them.
1720 * It is safe to rely on PG_active against the non-LRU pages in here because
1721 * nobody will play with that bit on a non-LRU page.
1722 *
1723 * The downside is that we have to touch page->_count against each page.
1724 * But we had to alter page->flags anyway.
1725 */
1726
1727static void move_active_pages_to_lru(struct lruvec *lruvec,
1728 struct list_head *list,
1729 struct list_head *pages_to_free,
1730 enum lru_list lru)
1731{
1732 struct zone *zone = lruvec_zone(lruvec);
1733 unsigned long pgmoved = 0;
1734 struct page *page;
1735 int nr_pages;
1736
1737 while (!list_empty(list)) {
1738 page = lru_to_page(list);
1739 lruvec = mem_cgroup_page_lruvec(page, zone);
1740
1741 VM_BUG_ON_PAGE(PageLRU(page), page);
1742 SetPageLRU(page);
1743
1744 nr_pages = hpage_nr_pages(page);
1745 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1746 list_move(&page->lru, &lruvec->lists[lru]);
1747 pgmoved += nr_pages;
1748
1749 if (put_page_testzero(page)) {
1750 __ClearPageLRU(page);
1751 __ClearPageActive(page);
1752 del_page_from_lru_list(page, lruvec, lru);
1753
1754 if (unlikely(PageCompound(page))) {
1755 spin_unlock_irq(&zone->lru_lock);
1756 mem_cgroup_uncharge(page);
1757 (*get_compound_page_dtor(page))(page);
1758 spin_lock_irq(&zone->lru_lock);
1759 } else
1760 list_add(&page->lru, pages_to_free);
1761 }
1762 }
1763 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1764 if (!is_active_lru(lru))
1765 __count_vm_events(PGDEACTIVATE, pgmoved);
1766}
1767
1768static void shrink_active_list(unsigned long nr_to_scan,
1769 struct lruvec *lruvec,
1770 struct scan_control *sc,
1771 enum lru_list lru)
1772{
1773 unsigned long nr_taken;
1774 unsigned long nr_scanned;
1775 unsigned long vm_flags;
1776 LIST_HEAD(l_hold); /* The pages which were snipped off */
1777 LIST_HEAD(l_active);
1778 LIST_HEAD(l_inactive);
1779 struct page *page;
1780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1781 unsigned long nr_rotated = 0;
1782 isolate_mode_t isolate_mode = 0;
1783 int file = is_file_lru(lru);
1784 struct zone *zone = lruvec_zone(lruvec);
1785
1786 lru_add_drain();
1787
1788 if (!sc->may_unmap)
1789 isolate_mode |= ISOLATE_UNMAPPED;
1790 if (!sc->may_writepage)
1791 isolate_mode |= ISOLATE_CLEAN;
1792
1793 spin_lock_irq(&zone->lru_lock);
1794
1795 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1796 &nr_scanned, sc, isolate_mode, lru);
1797 if (global_reclaim(sc))
1798 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1799
1800 reclaim_stat->recent_scanned[file] += nr_taken;
1801
1802 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1803 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1804 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1805 spin_unlock_irq(&zone->lru_lock);
1806
1807 while (!list_empty(&l_hold)) {
1808 cond_resched();
1809 page = lru_to_page(&l_hold);
1810 list_del(&page->lru);
1811
1812 if (unlikely(!page_evictable(page))) {
1813 putback_lru_page(page);
1814 continue;
1815 }
1816
1817 if (unlikely(buffer_heads_over_limit)) {
1818 if (page_has_private(page) && trylock_page(page)) {
1819 if (page_has_private(page))
1820 try_to_release_page(page, 0);
1821 unlock_page(page);
1822 }
1823 }
1824
1825 if (page_referenced(page, 0, sc->target_mem_cgroup,
1826 &vm_flags)) {
1827 nr_rotated += hpage_nr_pages(page);
1828 /*
1829 * Identify referenced, file-backed active pages and
1830 * give them one more trip around the active list. So
1831 * that executable code get better chances to stay in
1832 * memory under moderate memory pressure. Anon pages
1833 * are not likely to be evicted by use-once streaming
1834 * IO, plus JVM can create lots of anon VM_EXEC pages,
1835 * so we ignore them here.
1836 */
1837 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1838 list_add(&page->lru, &l_active);
1839 continue;
1840 }
1841 }
1842
1843 ClearPageActive(page); /* we are de-activating */
1844 list_add(&page->lru, &l_inactive);
1845 }
1846
1847 /*
1848 * Move pages back to the lru list.
1849 */
1850 spin_lock_irq(&zone->lru_lock);
1851 /*
1852 * Count referenced pages from currently used mappings as rotated,
1853 * even though only some of them are actually re-activated. This
1854 * helps balance scan pressure between file and anonymous pages in
1855 * get_scan_count.
1856 */
1857 reclaim_stat->recent_rotated[file] += nr_rotated;
1858
1859 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1860 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1861 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1862 spin_unlock_irq(&zone->lru_lock);
1863
1864 mem_cgroup_uncharge_list(&l_hold);
1865 free_hot_cold_page_list(&l_hold, true);
1866}
1867
1868#ifdef CONFIG_SWAP
1869static bool inactive_anon_is_low_global(struct zone *zone)
1870{
1871 unsigned long active, inactive;
1872
1873 active = zone_page_state(zone, NR_ACTIVE_ANON);
1874 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1875
1876 return inactive * zone->inactive_ratio < active;
1877}
1878
1879/**
1880 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1881 * @lruvec: LRU vector to check
1882 *
1883 * Returns true if the zone does not have enough inactive anon pages,
1884 * meaning some active anon pages need to be deactivated.
1885 */
1886static bool inactive_anon_is_low(struct lruvec *lruvec)
1887{
1888 /*
1889 * If we don't have swap space, anonymous page deactivation
1890 * is pointless.
1891 */
1892 if (!total_swap_pages)
1893 return false;
1894
1895 if (!mem_cgroup_disabled())
1896 return mem_cgroup_inactive_anon_is_low(lruvec);
1897
1898 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1899}
1900#else
1901static inline bool inactive_anon_is_low(struct lruvec *lruvec)
1902{
1903 return false;
1904}
1905#endif
1906
1907/**
1908 * inactive_file_is_low - check if file pages need to be deactivated
1909 * @lruvec: LRU vector to check
1910 *
1911 * When the system is doing streaming IO, memory pressure here
1912 * ensures that active file pages get deactivated, until more
1913 * than half of the file pages are on the inactive list.
1914 *
1915 * Once we get to that situation, protect the system's working
1916 * set from being evicted by disabling active file page aging.
1917 *
1918 * This uses a different ratio than the anonymous pages, because
1919 * the page cache uses a use-once replacement algorithm.
1920 */
1921static bool inactive_file_is_low(struct lruvec *lruvec)
1922{
1923 unsigned long inactive;
1924 unsigned long active;
1925
1926 inactive = lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
1927 active = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
1928
1929 return active > inactive;
1930}
1931
1932static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1933{
1934 if (is_file_lru(lru))
1935 return inactive_file_is_low(lruvec);
1936 else
1937 return inactive_anon_is_low(lruvec);
1938}
1939
1940static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1941 struct lruvec *lruvec, struct scan_control *sc)
1942{
1943 if (is_active_lru(lru)) {
1944 if (inactive_list_is_low(lruvec, lru))
1945 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1946 return 0;
1947 }
1948
1949 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1950}
1951
1952enum scan_balance {
1953 SCAN_EQUAL,
1954 SCAN_FRACT,
1955 SCAN_ANON,
1956 SCAN_FILE,
1957};
1958
1959/*
1960 * Determine how aggressively the anon and file LRU lists should be
1961 * scanned. The relative value of each set of LRU lists is determined
1962 * by looking at the fraction of the pages scanned we did rotate back
1963 * onto the active list instead of evict.
1964 *
1965 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1966 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1967 */
1968static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
1969 struct scan_control *sc, unsigned long *nr,
1970 unsigned long *lru_pages)
1971{
1972 int swappiness = mem_cgroup_swappiness(memcg);
1973 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1974 u64 fraction[2];
1975 u64 denominator = 0; /* gcc */
1976 struct zone *zone = lruvec_zone(lruvec);
1977 unsigned long anon_prio, file_prio;
1978 enum scan_balance scan_balance;
1979 unsigned long anon, file;
1980 bool force_scan = false;
1981 unsigned long ap, fp;
1982 enum lru_list lru;
1983 bool some_scanned;
1984 int pass;
1985
1986 /*
1987 * If the zone or memcg is small, nr[l] can be 0. This
1988 * results in no scanning on this priority and a potential
1989 * priority drop. Global direct reclaim can go to the next
1990 * zone and tends to have no problems. Global kswapd is for
1991 * zone balancing and it needs to scan a minimum amount. When
1992 * reclaiming for a memcg, a priority drop can cause high
1993 * latencies, so it's better to scan a minimum amount there as
1994 * well.
1995 */
1996 if (current_is_kswapd()) {
1997 if (!zone_reclaimable(zone))
1998 force_scan = true;
1999 if (!mem_cgroup_online(memcg))
2000 force_scan = true;
2001 }
2002 if (!global_reclaim(sc))
2003 force_scan = true;
2004
2005 /* If we have no swap space, do not bother scanning anon pages. */
2006 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2007 scan_balance = SCAN_FILE;
2008 goto out;
2009 }
2010
2011 /*
2012 * Global reclaim will swap to prevent OOM even with no
2013 * swappiness, but memcg users want to use this knob to
2014 * disable swapping for individual groups completely when
2015 * using the memory controller's swap limit feature would be
2016 * too expensive.
2017 */
2018 if (!global_reclaim(sc) && !swappiness) {
2019 scan_balance = SCAN_FILE;
2020 goto out;
2021 }
2022
2023 /*
2024 * Do not apply any pressure balancing cleverness when the
2025 * system is close to OOM, scan both anon and file equally
2026 * (unless the swappiness setting disagrees with swapping).
2027 */
2028 if (!sc->priority && swappiness) {
2029 scan_balance = SCAN_EQUAL;
2030 goto out;
2031 }
2032
2033 /*
2034 * Prevent the reclaimer from falling into the cache trap: as
2035 * cache pages start out inactive, every cache fault will tip
2036 * the scan balance towards the file LRU. And as the file LRU
2037 * shrinks, so does the window for rotation from references.
2038 * This means we have a runaway feedback loop where a tiny
2039 * thrashing file LRU becomes infinitely more attractive than
2040 * anon pages. Try to detect this based on file LRU size.
2041 */
2042 if (global_reclaim(sc)) {
2043 unsigned long zonefile;
2044 unsigned long zonefree;
2045
2046 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2047 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2048 zone_page_state(zone, NR_INACTIVE_FILE);
2049
2050 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2051 scan_balance = SCAN_ANON;
2052 goto out;
2053 }
2054 }
2055
2056 /*
2057 * If there is enough inactive page cache, i.e. if the size of the
2058 * inactive list is greater than that of the active list *and* the
2059 * inactive list actually has some pages to scan on this priority, we
2060 * do not reclaim anything from the anonymous working set right now.
2061 * Without the second condition we could end up never scanning an
2062 * lruvec even if it has plenty of old anonymous pages unless the
2063 * system is under heavy pressure.
2064 */
2065 if (!inactive_file_is_low(lruvec) &&
2066 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
2067 scan_balance = SCAN_FILE;
2068 goto out;
2069 }
2070
2071 scan_balance = SCAN_FRACT;
2072
2073 /*
2074 * With swappiness at 100, anonymous and file have the same priority.
2075 * This scanning priority is essentially the inverse of IO cost.
2076 */
2077 anon_prio = swappiness;
2078 file_prio = 200 - anon_prio;
2079
2080 /*
2081 * OK, so we have swap space and a fair amount of page cache
2082 * pages. We use the recently rotated / recently scanned
2083 * ratios to determine how valuable each cache is.
2084 *
2085 * Because workloads change over time (and to avoid overflow)
2086 * we keep these statistics as a floating average, which ends
2087 * up weighing recent references more than old ones.
2088 *
2089 * anon in [0], file in [1]
2090 */
2091
2092 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2093 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2094 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2095 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2096
2097 spin_lock_irq(&zone->lru_lock);
2098 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2099 reclaim_stat->recent_scanned[0] /= 2;
2100 reclaim_stat->recent_rotated[0] /= 2;
2101 }
2102
2103 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2104 reclaim_stat->recent_scanned[1] /= 2;
2105 reclaim_stat->recent_rotated[1] /= 2;
2106 }
2107
2108 /*
2109 * The amount of pressure on anon vs file pages is inversely
2110 * proportional to the fraction of recently scanned pages on
2111 * each list that were recently referenced and in active use.
2112 */
2113 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2114 ap /= reclaim_stat->recent_rotated[0] + 1;
2115
2116 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2117 fp /= reclaim_stat->recent_rotated[1] + 1;
2118 spin_unlock_irq(&zone->lru_lock);
2119
2120 fraction[0] = ap;
2121 fraction[1] = fp;
2122 denominator = ap + fp + 1;
2123out:
2124 some_scanned = false;
2125 /* Only use force_scan on second pass. */
2126 for (pass = 0; !some_scanned && pass < 2; pass++) {
2127 *lru_pages = 0;
2128 for_each_evictable_lru(lru) {
2129 int file = is_file_lru(lru);
2130 unsigned long size;
2131 unsigned long scan;
2132
2133 size = lruvec_lru_size(lruvec, lru);
2134 scan = size >> sc->priority;
2135
2136 if (!scan && pass && force_scan)
2137 scan = min(size, SWAP_CLUSTER_MAX);
2138
2139 switch (scan_balance) {
2140 case SCAN_EQUAL:
2141 /* Scan lists relative to size */
2142 break;
2143 case SCAN_FRACT:
2144 /*
2145 * Scan types proportional to swappiness and
2146 * their relative recent reclaim efficiency.
2147 */
2148 scan = div64_u64(scan * fraction[file],
2149 denominator);
2150 break;
2151 case SCAN_FILE:
2152 case SCAN_ANON:
2153 /* Scan one type exclusively */
2154 if ((scan_balance == SCAN_FILE) != file) {
2155 size = 0;
2156 scan = 0;
2157 }
2158 break;
2159 default:
2160 /* Look ma, no brain */
2161 BUG();
2162 }
2163
2164 *lru_pages += size;
2165 nr[lru] = scan;
2166
2167 /*
2168 * Skip the second pass and don't force_scan,
2169 * if we found something to scan.
2170 */
2171 some_scanned |= !!scan;
2172 }
2173 }
2174}
2175
2176#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2177static void init_tlb_ubc(void)
2178{
2179 /*
2180 * This deliberately does not clear the cpumask as it's expensive
2181 * and unnecessary. If there happens to be data in there then the
2182 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2183 * then will be cleared.
2184 */
2185 current->tlb_ubc.flush_required = false;
2186}
2187#else
2188static inline void init_tlb_ubc(void)
2189{
2190}
2191#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2192
2193/*
2194 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2195 */
2196static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *lru_pages)
2198{
2199 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2200 unsigned long nr[NR_LRU_LISTS];
2201 unsigned long targets[NR_LRU_LISTS];
2202 unsigned long nr_to_scan;
2203 enum lru_list lru;
2204 unsigned long nr_reclaimed = 0;
2205 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2206 struct blk_plug plug;
2207 bool scan_adjusted;
2208
2209 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2210
2211 /* Record the original scan target for proportional adjustments later */
2212 memcpy(targets, nr, sizeof(nr));
2213
2214 /*
2215 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2216 * event that can occur when there is little memory pressure e.g.
2217 * multiple streaming readers/writers. Hence, we do not abort scanning
2218 * when the requested number of pages are reclaimed when scanning at
2219 * DEF_PRIORITY on the assumption that the fact we are direct
2220 * reclaiming implies that kswapd is not keeping up and it is best to
2221 * do a batch of work at once. For memcg reclaim one check is made to
2222 * abort proportional reclaim if either the file or anon lru has already
2223 * dropped to zero at the first pass.
2224 */
2225 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2226 sc->priority == DEF_PRIORITY);
2227
2228 init_tlb_ubc();
2229
2230 blk_start_plug(&plug);
2231 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2232 nr[LRU_INACTIVE_FILE]) {
2233 unsigned long nr_anon, nr_file, percentage;
2234 unsigned long nr_scanned;
2235
2236 for_each_evictable_lru(lru) {
2237 if (nr[lru]) {
2238 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2239 nr[lru] -= nr_to_scan;
2240
2241 nr_reclaimed += shrink_list(lru, nr_to_scan,
2242 lruvec, sc);
2243 }
2244 }
2245
2246 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2247 continue;
2248
2249 /*
2250 * For kswapd and memcg, reclaim at least the number of pages
2251 * requested. Ensure that the anon and file LRUs are scanned
2252 * proportionally what was requested by get_scan_count(). We
2253 * stop reclaiming one LRU and reduce the amount scanning
2254 * proportional to the original scan target.
2255 */
2256 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2257 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2258
2259 /*
2260 * It's just vindictive to attack the larger once the smaller
2261 * has gone to zero. And given the way we stop scanning the
2262 * smaller below, this makes sure that we only make one nudge
2263 * towards proportionality once we've got nr_to_reclaim.
2264 */
2265 if (!nr_file || !nr_anon)
2266 break;
2267
2268 if (nr_file > nr_anon) {
2269 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2270 targets[LRU_ACTIVE_ANON] + 1;
2271 lru = LRU_BASE;
2272 percentage = nr_anon * 100 / scan_target;
2273 } else {
2274 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2275 targets[LRU_ACTIVE_FILE] + 1;
2276 lru = LRU_FILE;
2277 percentage = nr_file * 100 / scan_target;
2278 }
2279
2280 /* Stop scanning the smaller of the LRU */
2281 nr[lru] = 0;
2282 nr[lru + LRU_ACTIVE] = 0;
2283
2284 /*
2285 * Recalculate the other LRU scan count based on its original
2286 * scan target and the percentage scanning already complete
2287 */
2288 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2289 nr_scanned = targets[lru] - nr[lru];
2290 nr[lru] = targets[lru] * (100 - percentage) / 100;
2291 nr[lru] -= min(nr[lru], nr_scanned);
2292
2293 lru += LRU_ACTIVE;
2294 nr_scanned = targets[lru] - nr[lru];
2295 nr[lru] = targets[lru] * (100 - percentage) / 100;
2296 nr[lru] -= min(nr[lru], nr_scanned);
2297
2298 scan_adjusted = true;
2299 }
2300 blk_finish_plug(&plug);
2301 sc->nr_reclaimed += nr_reclaimed;
2302
2303 /*
2304 * Even if we did not try to evict anon pages at all, we want to
2305 * rebalance the anon lru active/inactive ratio.
2306 */
2307 if (inactive_anon_is_low(lruvec))
2308 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2309 sc, LRU_ACTIVE_ANON);
2310
2311 throttle_vm_writeout(sc->gfp_mask);
2312}
2313
2314/* Use reclaim/compaction for costly allocs or under memory pressure */
2315static bool in_reclaim_compaction(struct scan_control *sc)
2316{
2317 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2318 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2319 sc->priority < DEF_PRIORITY - 2))
2320 return true;
2321
2322 return false;
2323}
2324
2325/*
2326 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2327 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2328 * true if more pages should be reclaimed such that when the page allocator
2329 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2330 * It will give up earlier than that if there is difficulty reclaiming pages.
2331 */
2332static inline bool should_continue_reclaim(struct zone *zone,
2333 unsigned long nr_reclaimed,
2334 unsigned long nr_scanned,
2335 struct scan_control *sc)
2336{
2337 unsigned long pages_for_compaction;
2338 unsigned long inactive_lru_pages;
2339
2340 /* If not in reclaim/compaction mode, stop */
2341 if (!in_reclaim_compaction(sc))
2342 return false;
2343
2344 /* Consider stopping depending on scan and reclaim activity */
2345 if (sc->gfp_mask & __GFP_REPEAT) {
2346 /*
2347 * For __GFP_REPEAT allocations, stop reclaiming if the
2348 * full LRU list has been scanned and we are still failing
2349 * to reclaim pages. This full LRU scan is potentially
2350 * expensive but a __GFP_REPEAT caller really wants to succeed
2351 */
2352 if (!nr_reclaimed && !nr_scanned)
2353 return false;
2354 } else {
2355 /*
2356 * For non-__GFP_REPEAT allocations which can presumably
2357 * fail without consequence, stop if we failed to reclaim
2358 * any pages from the last SWAP_CLUSTER_MAX number of
2359 * pages that were scanned. This will return to the
2360 * caller faster at the risk reclaim/compaction and
2361 * the resulting allocation attempt fails
2362 */
2363 if (!nr_reclaimed)
2364 return false;
2365 }
2366
2367 /*
2368 * If we have not reclaimed enough pages for compaction and the
2369 * inactive lists are large enough, continue reclaiming
2370 */
2371 pages_for_compaction = (2UL << sc->order);
2372 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2373 if (get_nr_swap_pages() > 0)
2374 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2375 if (sc->nr_reclaimed < pages_for_compaction &&
2376 inactive_lru_pages > pages_for_compaction)
2377 return true;
2378
2379 /* If compaction would go ahead or the allocation would succeed, stop */
2380 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2381 case COMPACT_PARTIAL:
2382 case COMPACT_CONTINUE:
2383 return false;
2384 default:
2385 return true;
2386 }
2387}
2388
2389static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2390 bool is_classzone)
2391{
2392 struct reclaim_state *reclaim_state = current->reclaim_state;
2393 unsigned long nr_reclaimed, nr_scanned;
2394 bool reclaimable = false;
2395
2396 do {
2397 struct mem_cgroup *root = sc->target_mem_cgroup;
2398 struct mem_cgroup_reclaim_cookie reclaim = {
2399 .zone = zone,
2400 .priority = sc->priority,
2401 };
2402 unsigned long zone_lru_pages = 0;
2403 struct mem_cgroup *memcg;
2404
2405 nr_reclaimed = sc->nr_reclaimed;
2406 nr_scanned = sc->nr_scanned;
2407
2408 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2409 do {
2410 unsigned long lru_pages;
2411 unsigned long reclaimed;
2412 unsigned long scanned;
2413
2414 if (mem_cgroup_low(root, memcg)) {
2415 if (!sc->may_thrash)
2416 continue;
2417 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2418 }
2419
2420 reclaimed = sc->nr_reclaimed;
2421 scanned = sc->nr_scanned;
2422
2423 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
2424 zone_lru_pages += lru_pages;
2425
2426 if (memcg && is_classzone)
2427 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2428 memcg, sc->nr_scanned - scanned,
2429 lru_pages);
2430
2431 /* Record the group's reclaim efficiency */
2432 vmpressure(sc->gfp_mask, memcg, false,
2433 sc->nr_scanned - scanned,
2434 sc->nr_reclaimed - reclaimed);
2435
2436 /*
2437 * Direct reclaim and kswapd have to scan all memory
2438 * cgroups to fulfill the overall scan target for the
2439 * zone.
2440 *
2441 * Limit reclaim, on the other hand, only cares about
2442 * nr_to_reclaim pages to be reclaimed and it will
2443 * retry with decreasing priority if one round over the
2444 * whole hierarchy is not sufficient.
2445 */
2446 if (!global_reclaim(sc) &&
2447 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2448 mem_cgroup_iter_break(root, memcg);
2449 break;
2450 }
2451 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2452
2453 /*
2454 * Shrink the slab caches in the same proportion that
2455 * the eligible LRU pages were scanned.
2456 */
2457 if (global_reclaim(sc) && is_classzone)
2458 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2459 sc->nr_scanned - nr_scanned,
2460 zone_lru_pages);
2461
2462 if (reclaim_state) {
2463 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2464 reclaim_state->reclaimed_slab = 0;
2465 }
2466
2467 /* Record the subtree's reclaim efficiency */
2468 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2469 sc->nr_scanned - nr_scanned,
2470 sc->nr_reclaimed - nr_reclaimed);
2471
2472 if (sc->nr_reclaimed - nr_reclaimed)
2473 reclaimable = true;
2474
2475 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2476 sc->nr_scanned - nr_scanned, sc));
2477
2478 return reclaimable;
2479}
2480
2481/*
2482 * Returns true if compaction should go ahead for a high-order request, or
2483 * the high-order allocation would succeed without compaction.
2484 */
2485static inline bool compaction_ready(struct zone *zone, int order)
2486{
2487 unsigned long balance_gap, watermark;
2488 bool watermark_ok;
2489
2490 /*
2491 * Compaction takes time to run and there are potentially other
2492 * callers using the pages just freed. Continue reclaiming until
2493 * there is a buffer of free pages available to give compaction
2494 * a reasonable chance of completing and allocating the page
2495 */
2496 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2497 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2498 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2499 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
2500
2501 /*
2502 * If compaction is deferred, reclaim up to a point where
2503 * compaction will have a chance of success when re-enabled
2504 */
2505 if (compaction_deferred(zone, order))
2506 return watermark_ok;
2507
2508 /*
2509 * If compaction is not ready to start and allocation is not likely
2510 * to succeed without it, then keep reclaiming.
2511 */
2512 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2513 return false;
2514
2515 return watermark_ok;
2516}
2517
2518/*
2519 * This is the direct reclaim path, for page-allocating processes. We only
2520 * try to reclaim pages from zones which will satisfy the caller's allocation
2521 * request.
2522 *
2523 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2524 * Because:
2525 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2526 * allocation or
2527 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2528 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2529 * zone defense algorithm.
2530 *
2531 * If a zone is deemed to be full of pinned pages then just give it a light
2532 * scan then give up on it.
2533 *
2534 * Returns true if a zone was reclaimable.
2535 */
2536static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2537{
2538 struct zoneref *z;
2539 struct zone *zone;
2540 unsigned long nr_soft_reclaimed;
2541 unsigned long nr_soft_scanned;
2542 gfp_t orig_mask;
2543 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2544 bool reclaimable = false;
2545
2546 /*
2547 * If the number of buffer_heads in the machine exceeds the maximum
2548 * allowed level, force direct reclaim to scan the highmem zone as
2549 * highmem pages could be pinning lowmem pages storing buffer_heads
2550 */
2551 orig_mask = sc->gfp_mask;
2552 if (buffer_heads_over_limit)
2553 sc->gfp_mask |= __GFP_HIGHMEM;
2554
2555 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2556 gfp_zone(sc->gfp_mask), sc->nodemask) {
2557 enum zone_type classzone_idx;
2558
2559 if (!populated_zone(zone))
2560 continue;
2561
2562 classzone_idx = requested_highidx;
2563 while (!populated_zone(zone->zone_pgdat->node_zones +
2564 classzone_idx))
2565 classzone_idx--;
2566
2567 /*
2568 * Take care memory controller reclaiming has small influence
2569 * to global LRU.
2570 */
2571 if (global_reclaim(sc)) {
2572 if (!cpuset_zone_allowed(zone,
2573 GFP_KERNEL | __GFP_HARDWALL))
2574 continue;
2575
2576 if (sc->priority != DEF_PRIORITY &&
2577 !zone_reclaimable(zone))
2578 continue; /* Let kswapd poll it */
2579
2580 /*
2581 * If we already have plenty of memory free for
2582 * compaction in this zone, don't free any more.
2583 * Even though compaction is invoked for any
2584 * non-zero order, only frequent costly order
2585 * reclamation is disruptive enough to become a
2586 * noticeable problem, like transparent huge
2587 * page allocations.
2588 */
2589 if (IS_ENABLED(CONFIG_COMPACTION) &&
2590 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2591 zonelist_zone_idx(z) <= requested_highidx &&
2592 compaction_ready(zone, sc->order)) {
2593 sc->compaction_ready = true;
2594 continue;
2595 }
2596
2597 /*
2598 * This steals pages from memory cgroups over softlimit
2599 * and returns the number of reclaimed pages and
2600 * scanned pages. This works for global memory pressure
2601 * and balancing, not for a memcg's limit.
2602 */
2603 nr_soft_scanned = 0;
2604 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2605 sc->order, sc->gfp_mask,
2606 &nr_soft_scanned);
2607 sc->nr_reclaimed += nr_soft_reclaimed;
2608 sc->nr_scanned += nr_soft_scanned;
2609 if (nr_soft_reclaimed)
2610 reclaimable = true;
2611 /* need some check for avoid more shrink_zone() */
2612 }
2613
2614 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2615 reclaimable = true;
2616
2617 if (global_reclaim(sc) &&
2618 !reclaimable && zone_reclaimable(zone))
2619 reclaimable = true;
2620 }
2621
2622 /*
2623 * Restore to original mask to avoid the impact on the caller if we
2624 * promoted it to __GFP_HIGHMEM.
2625 */
2626 sc->gfp_mask = orig_mask;
2627
2628 return reclaimable;
2629}
2630
2631/*
2632 * This is the main entry point to direct page reclaim.
2633 *
2634 * If a full scan of the inactive list fails to free enough memory then we
2635 * are "out of memory" and something needs to be killed.
2636 *
2637 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2638 * high - the zone may be full of dirty or under-writeback pages, which this
2639 * caller can't do much about. We kick the writeback threads and take explicit
2640 * naps in the hope that some of these pages can be written. But if the
2641 * allocating task holds filesystem locks which prevent writeout this might not
2642 * work, and the allocation attempt will fail.
2643 *
2644 * returns: 0, if no pages reclaimed
2645 * else, the number of pages reclaimed
2646 */
2647static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2648 struct scan_control *sc)
2649{
2650 int initial_priority = sc->priority;
2651 unsigned long total_scanned = 0;
2652 unsigned long writeback_threshold;
2653 bool zones_reclaimable;
2654retry:
2655 delayacct_freepages_start();
2656
2657 if (global_reclaim(sc))
2658 count_vm_event(ALLOCSTALL);
2659
2660 do {
2661 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2662 sc->priority);
2663 sc->nr_scanned = 0;
2664 zones_reclaimable = shrink_zones(zonelist, sc);
2665
2666 total_scanned += sc->nr_scanned;
2667 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2668 break;
2669
2670 if (sc->compaction_ready)
2671 break;
2672
2673 /*
2674 * If we're getting trouble reclaiming, start doing
2675 * writepage even in laptop mode.
2676 */
2677 if (sc->priority < DEF_PRIORITY - 2)
2678 sc->may_writepage = 1;
2679
2680 /*
2681 * Try to write back as many pages as we just scanned. This
2682 * tends to cause slow streaming writers to write data to the
2683 * disk smoothly, at the dirtying rate, which is nice. But
2684 * that's undesirable in laptop mode, where we *want* lumpy
2685 * writeout. So in laptop mode, write out the whole world.
2686 */
2687 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2688 if (total_scanned > writeback_threshold) {
2689 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2690 WB_REASON_TRY_TO_FREE_PAGES);
2691 sc->may_writepage = 1;
2692 }
2693 } while (--sc->priority >= 0);
2694
2695 delayacct_freepages_end();
2696
2697 if (sc->nr_reclaimed)
2698 return sc->nr_reclaimed;
2699
2700 /* Aborted reclaim to try compaction? don't OOM, then */
2701 if (sc->compaction_ready)
2702 return 1;
2703
2704 /* Untapped cgroup reserves? Don't OOM, retry. */
2705 if (!sc->may_thrash) {
2706 sc->priority = initial_priority;
2707 sc->may_thrash = 1;
2708 goto retry;
2709 }
2710
2711 /* Any of the zones still reclaimable? Don't OOM. */
2712 if (zones_reclaimable)
2713 return 1;
2714
2715 return 0;
2716}
2717
2718static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2719{
2720 struct zone *zone;
2721 unsigned long pfmemalloc_reserve = 0;
2722 unsigned long free_pages = 0;
2723 int i;
2724 bool wmark_ok;
2725
2726 for (i = 0; i <= ZONE_NORMAL; i++) {
2727 zone = &pgdat->node_zones[i];
2728 if (!populated_zone(zone) ||
2729 zone_reclaimable_pages(zone) == 0)
2730 continue;
2731
2732 pfmemalloc_reserve += min_wmark_pages(zone);
2733 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2734 }
2735
2736 /* If there are no reserves (unexpected config) then do not throttle */
2737 if (!pfmemalloc_reserve)
2738 return true;
2739
2740 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2741
2742 /* kswapd must be awake if processes are being throttled */
2743 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2744 pgdat->classzone_idx = min(pgdat->classzone_idx,
2745 (enum zone_type)ZONE_NORMAL);
2746 wake_up_interruptible(&pgdat->kswapd_wait);
2747 }
2748
2749 return wmark_ok;
2750}
2751
2752/*
2753 * Throttle direct reclaimers if backing storage is backed by the network
2754 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2755 * depleted. kswapd will continue to make progress and wake the processes
2756 * when the low watermark is reached.
2757 *
2758 * Returns true if a fatal signal was delivered during throttling. If this
2759 * happens, the page allocator should not consider triggering the OOM killer.
2760 */
2761static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2762 nodemask_t *nodemask)
2763{
2764 struct zoneref *z;
2765 struct zone *zone;
2766 pg_data_t *pgdat = NULL;
2767
2768 /*
2769 * Kernel threads should not be throttled as they may be indirectly
2770 * responsible for cleaning pages necessary for reclaim to make forward
2771 * progress. kjournald for example may enter direct reclaim while
2772 * committing a transaction where throttling it could forcing other
2773 * processes to block on log_wait_commit().
2774 */
2775 if (current->flags & PF_KTHREAD)
2776 goto out;
2777
2778 /*
2779 * If a fatal signal is pending, this process should not throttle.
2780 * It should return quickly so it can exit and free its memory
2781 */
2782 if (fatal_signal_pending(current))
2783 goto out;
2784
2785 /*
2786 * Check if the pfmemalloc reserves are ok by finding the first node
2787 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2788 * GFP_KERNEL will be required for allocating network buffers when
2789 * swapping over the network so ZONE_HIGHMEM is unusable.
2790 *
2791 * Throttling is based on the first usable node and throttled processes
2792 * wait on a queue until kswapd makes progress and wakes them. There
2793 * is an affinity then between processes waking up and where reclaim
2794 * progress has been made assuming the process wakes on the same node.
2795 * More importantly, processes running on remote nodes will not compete
2796 * for remote pfmemalloc reserves and processes on different nodes
2797 * should make reasonable progress.
2798 */
2799 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2800 gfp_zone(gfp_mask), nodemask) {
2801 if (zone_idx(zone) > ZONE_NORMAL)
2802 continue;
2803
2804 /* Throttle based on the first usable node */
2805 pgdat = zone->zone_pgdat;
2806 if (pfmemalloc_watermark_ok(pgdat))
2807 goto out;
2808 break;
2809 }
2810
2811 /* If no zone was usable by the allocation flags then do not throttle */
2812 if (!pgdat)
2813 goto out;
2814
2815 /* Account for the throttling */
2816 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2817
2818 /*
2819 * If the caller cannot enter the filesystem, it's possible that it
2820 * is due to the caller holding an FS lock or performing a journal
2821 * transaction in the case of a filesystem like ext[3|4]. In this case,
2822 * it is not safe to block on pfmemalloc_wait as kswapd could be
2823 * blocked waiting on the same lock. Instead, throttle for up to a
2824 * second before continuing.
2825 */
2826 if (!(gfp_mask & __GFP_FS)) {
2827 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2828 pfmemalloc_watermark_ok(pgdat), HZ);
2829
2830 goto check_pending;
2831 }
2832
2833 /* Throttle until kswapd wakes the process */
2834 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2835 pfmemalloc_watermark_ok(pgdat));
2836
2837check_pending:
2838 if (fatal_signal_pending(current))
2839 return true;
2840
2841out:
2842 return false;
2843}
2844
2845unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2846 gfp_t gfp_mask, nodemask_t *nodemask)
2847{
2848 unsigned long nr_reclaimed;
2849 struct scan_control sc = {
2850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2851 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2852 .order = order,
2853 .nodemask = nodemask,
2854 .priority = DEF_PRIORITY,
2855 .may_writepage = !laptop_mode,
2856 .may_unmap = 1,
2857 .may_swap = 1,
2858 };
2859
2860 /*
2861 * Do not enter reclaim if fatal signal was delivered while throttled.
2862 * 1 is returned so that the page allocator does not OOM kill at this
2863 * point.
2864 */
2865 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2866 return 1;
2867
2868 trace_mm_vmscan_direct_reclaim_begin(order,
2869 sc.may_writepage,
2870 gfp_mask);
2871
2872 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2873
2874 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2875
2876 return nr_reclaimed;
2877}
2878
2879#ifdef CONFIG_MEMCG
2880
2881unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2882 gfp_t gfp_mask, bool noswap,
2883 struct zone *zone,
2884 unsigned long *nr_scanned)
2885{
2886 struct scan_control sc = {
2887 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2888 .target_mem_cgroup = memcg,
2889 .may_writepage = !laptop_mode,
2890 .may_unmap = 1,
2891 .may_swap = !noswap,
2892 };
2893 unsigned long lru_pages;
2894
2895 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2896 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2897
2898 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2899 sc.may_writepage,
2900 sc.gfp_mask);
2901
2902 /*
2903 * NOTE: Although we can get the priority field, using it
2904 * here is not a good idea, since it limits the pages we can scan.
2905 * if we don't reclaim here, the shrink_zone from balance_pgdat
2906 * will pick up pages from other mem cgroup's as well. We hack
2907 * the priority and make it zero.
2908 */
2909 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
2910
2911 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2912
2913 *nr_scanned = sc.nr_scanned;
2914 return sc.nr_reclaimed;
2915}
2916
2917unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2918 unsigned long nr_pages,
2919 gfp_t gfp_mask,
2920 bool may_swap)
2921{
2922 struct zonelist *zonelist;
2923 unsigned long nr_reclaimed;
2924 int nid;
2925 struct scan_control sc = {
2926 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2927 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2928 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2929 .target_mem_cgroup = memcg,
2930 .priority = DEF_PRIORITY,
2931 .may_writepage = !laptop_mode,
2932 .may_unmap = 1,
2933 .may_swap = may_swap,
2934 };
2935
2936 /*
2937 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2938 * take care of from where we get pages. So the node where we start the
2939 * scan does not need to be the current node.
2940 */
2941 nid = mem_cgroup_select_victim_node(memcg);
2942
2943 zonelist = NODE_DATA(nid)->node_zonelists;
2944
2945 trace_mm_vmscan_memcg_reclaim_begin(0,
2946 sc.may_writepage,
2947 sc.gfp_mask);
2948
2949 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2950
2951 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2952
2953 return nr_reclaimed;
2954}
2955#endif
2956
2957static void age_active_anon(struct zone *zone, struct scan_control *sc)
2958{
2959 struct mem_cgroup *memcg;
2960
2961 if (!total_swap_pages)
2962 return;
2963
2964 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2965 do {
2966 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2967
2968 if (inactive_anon_is_low(lruvec))
2969 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2970 sc, LRU_ACTIVE_ANON);
2971
2972 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2973 } while (memcg);
2974}
2975
2976static bool zone_balanced(struct zone *zone, int order, bool highorder,
2977 unsigned long balance_gap, int classzone_idx)
2978{
2979 unsigned long mark = high_wmark_pages(zone) + balance_gap;
2980
2981 /*
2982 * When checking from pgdat_balanced(), kswapd should stop and sleep
2983 * when it reaches the high order-0 watermark and let kcompactd take
2984 * over. Other callers such as wakeup_kswapd() want to determine the
2985 * true high-order watermark.
2986 */
2987 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
2988 mark += (1UL << order);
2989 order = 0;
2990 }
2991
2992 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
2993}
2994
2995/*
2996 * pgdat_balanced() is used when checking if a node is balanced.
2997 *
2998 * For order-0, all zones must be balanced!
2999 *
3000 * For high-order allocations only zones that meet watermarks and are in a
3001 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3002 * total of balanced pages must be at least 25% of the zones allowed by
3003 * classzone_idx for the node to be considered balanced. Forcing all zones to
3004 * be balanced for high orders can cause excessive reclaim when there are
3005 * imbalanced zones.
3006 * The choice of 25% is due to
3007 * o a 16M DMA zone that is balanced will not balance a zone on any
3008 * reasonable sized machine
3009 * o On all other machines, the top zone must be at least a reasonable
3010 * percentage of the middle zones. For example, on 32-bit x86, highmem
3011 * would need to be at least 256M for it to be balance a whole node.
3012 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3013 * to balance a node on its own. These seemed like reasonable ratios.
3014 */
3015static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3016{
3017 unsigned long managed_pages = 0;
3018 unsigned long balanced_pages = 0;
3019 int i;
3020
3021 /* Check the watermark levels */
3022 for (i = 0; i <= classzone_idx; i++) {
3023 struct zone *zone = pgdat->node_zones + i;
3024
3025 if (!populated_zone(zone))
3026 continue;
3027
3028 managed_pages += zone->managed_pages;
3029
3030 /*
3031 * A special case here:
3032 *
3033 * balance_pgdat() skips over all_unreclaimable after
3034 * DEF_PRIORITY. Effectively, it considers them balanced so
3035 * they must be considered balanced here as well!
3036 */
3037 if (!zone_reclaimable(zone)) {
3038 balanced_pages += zone->managed_pages;
3039 continue;
3040 }
3041
3042 if (zone_balanced(zone, order, false, 0, i))
3043 balanced_pages += zone->managed_pages;
3044 else if (!order)
3045 return false;
3046 }
3047
3048 if (order)
3049 return balanced_pages >= (managed_pages >> 2);
3050 else
3051 return true;
3052}
3053
3054/*
3055 * Prepare kswapd for sleeping. This verifies that there are no processes
3056 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3057 *
3058 * Returns true if kswapd is ready to sleep
3059 */
3060static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3061 int classzone_idx)
3062{
3063 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3064 if (remaining)
3065 return false;
3066
3067 /*
3068 * The throttled processes are normally woken up in balance_pgdat() as
3069 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3070 * race between when kswapd checks the watermarks and a process gets
3071 * throttled. There is also a potential race if processes get
3072 * throttled, kswapd wakes, a large process exits thereby balancing the
3073 * zones, which causes kswapd to exit balance_pgdat() before reaching
3074 * the wake up checks. If kswapd is going to sleep, no process should
3075 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3076 * the wake up is premature, processes will wake kswapd and get
3077 * throttled again. The difference from wake ups in balance_pgdat() is
3078 * that here we are under prepare_to_wait().
3079 */
3080 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3081 wake_up_all(&pgdat->pfmemalloc_wait);
3082
3083 return pgdat_balanced(pgdat, order, classzone_idx);
3084}
3085
3086/*
3087 * kswapd shrinks the zone by the number of pages required to reach
3088 * the high watermark.
3089 *
3090 * Returns true if kswapd scanned at least the requested number of pages to
3091 * reclaim or if the lack of progress was due to pages under writeback.
3092 * This is used to determine if the scanning priority needs to be raised.
3093 */
3094static bool kswapd_shrink_zone(struct zone *zone,
3095 int classzone_idx,
3096 struct scan_control *sc)
3097{
3098 unsigned long balance_gap;
3099 bool lowmem_pressure;
3100
3101 /* Reclaim above the high watermark. */
3102 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3103
3104 /*
3105 * We put equal pressure on every zone, unless one zone has way too
3106 * many pages free already. The "too many pages" is defined as the
3107 * high wmark plus a "gap" where the gap is either the low
3108 * watermark or 1% of the zone, whichever is smaller.
3109 */
3110 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3111 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3112
3113 /*
3114 * If there is no low memory pressure or the zone is balanced then no
3115 * reclaim is necessary
3116 */
3117 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3118 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
3119 balance_gap, classzone_idx))
3120 return true;
3121
3122 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3123
3124 clear_bit(ZONE_WRITEBACK, &zone->flags);
3125
3126 /*
3127 * If a zone reaches its high watermark, consider it to be no longer
3128 * congested. It's possible there are dirty pages backed by congested
3129 * BDIs but as pressure is relieved, speculatively avoid congestion
3130 * waits.
3131 */
3132 if (zone_reclaimable(zone) &&
3133 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
3134 clear_bit(ZONE_CONGESTED, &zone->flags);
3135 clear_bit(ZONE_DIRTY, &zone->flags);
3136 }
3137
3138 return sc->nr_scanned >= sc->nr_to_reclaim;
3139}
3140
3141/*
3142 * For kswapd, balance_pgdat() will work across all this node's zones until
3143 * they are all at high_wmark_pages(zone).
3144 *
3145 * Returns the highest zone idx kswapd was reclaiming at
3146 *
3147 * There is special handling here for zones which are full of pinned pages.
3148 * This can happen if the pages are all mlocked, or if they are all used by
3149 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3150 * What we do is to detect the case where all pages in the zone have been
3151 * scanned twice and there has been zero successful reclaim. Mark the zone as
3152 * dead and from now on, only perform a short scan. Basically we're polling
3153 * the zone for when the problem goes away.
3154 *
3155 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3156 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3157 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3158 * lower zones regardless of the number of free pages in the lower zones. This
3159 * interoperates with the page allocator fallback scheme to ensure that aging
3160 * of pages is balanced across the zones.
3161 */
3162static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3163{
3164 int i;
3165 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3166 unsigned long nr_soft_reclaimed;
3167 unsigned long nr_soft_scanned;
3168 struct scan_control sc = {
3169 .gfp_mask = GFP_KERNEL,
3170 .order = order,
3171 .priority = DEF_PRIORITY,
3172 .may_writepage = !laptop_mode,
3173 .may_unmap = 1,
3174 .may_swap = 1,
3175 };
3176 count_vm_event(PAGEOUTRUN);
3177
3178 do {
3179 bool raise_priority = true;
3180
3181 sc.nr_reclaimed = 0;
3182
3183 /*
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3186 */
3187 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188 struct zone *zone = pgdat->node_zones + i;
3189
3190 if (!populated_zone(zone))
3191 continue;
3192
3193 if (sc.priority != DEF_PRIORITY &&
3194 !zone_reclaimable(zone))
3195 continue;
3196
3197 /*
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3200 */
3201 age_active_anon(zone, &sc);
3202
3203 /*
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3208 */
3209 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210 end_zone = i;
3211 break;
3212 }
3213
3214 if (!zone_balanced(zone, order, false, 0, 0)) {
3215 end_zone = i;
3216 break;
3217 } else {
3218 /*
3219 * If balanced, clear the dirty and congested
3220 * flags
3221 */
3222 clear_bit(ZONE_CONGESTED, &zone->flags);
3223 clear_bit(ZONE_DIRTY, &zone->flags);
3224 }
3225 }
3226
3227 if (i < 0)
3228 goto out;
3229
3230 /*
3231 * If we're getting trouble reclaiming, start doing writepage
3232 * even in laptop mode.
3233 */
3234 if (sc.priority < DEF_PRIORITY - 2)
3235 sc.may_writepage = 1;
3236
3237 /*
3238 * Now scan the zone in the dma->highmem direction, stopping
3239 * at the last zone which needs scanning.
3240 *
3241 * We do this because the page allocator works in the opposite
3242 * direction. This prevents the page allocator from allocating
3243 * pages behind kswapd's direction of progress, which would
3244 * cause too much scanning of the lower zones.
3245 */
3246 for (i = 0; i <= end_zone; i++) {
3247 struct zone *zone = pgdat->node_zones + i;
3248
3249 if (!populated_zone(zone))
3250 continue;
3251
3252 if (sc.priority != DEF_PRIORITY &&
3253 !zone_reclaimable(zone))
3254 continue;
3255
3256 sc.nr_scanned = 0;
3257
3258 nr_soft_scanned = 0;
3259 /*
3260 * Call soft limit reclaim before calling shrink_zone.
3261 */
3262 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3263 order, sc.gfp_mask,
3264 &nr_soft_scanned);
3265 sc.nr_reclaimed += nr_soft_reclaimed;
3266
3267 /*
3268 * There should be no need to raise the scanning
3269 * priority if enough pages are already being scanned
3270 * that that high watermark would be met at 100%
3271 * efficiency.
3272 */
3273 if (kswapd_shrink_zone(zone, end_zone, &sc))
3274 raise_priority = false;
3275 }
3276
3277 /*
3278 * If the low watermark is met there is no need for processes
3279 * to be throttled on pfmemalloc_wait as they should not be
3280 * able to safely make forward progress. Wake them
3281 */
3282 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3283 pfmemalloc_watermark_ok(pgdat))
3284 wake_up_all(&pgdat->pfmemalloc_wait);
3285
3286 /* Check if kswapd should be suspending */
3287 if (try_to_freeze() || kthread_should_stop())
3288 break;
3289
3290 /*
3291 * Raise priority if scanning rate is too low or there was no
3292 * progress in reclaiming pages
3293 */
3294 if (raise_priority || !sc.nr_reclaimed)
3295 sc.priority--;
3296 } while (sc.priority >= 1 &&
3297 !pgdat_balanced(pgdat, order, classzone_idx));
3298
3299out:
3300 /*
3301 * Return the highest zone idx we were reclaiming at so
3302 * prepare_kswapd_sleep() makes the same decisions as here.
3303 */
3304 return end_zone;
3305}
3306
3307static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3308 int classzone_idx, int balanced_classzone_idx)
3309{
3310 long remaining = 0;
3311 DEFINE_WAIT(wait);
3312
3313 if (freezing(current) || kthread_should_stop())
3314 return;
3315
3316 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3317
3318 /* Try to sleep for a short interval */
3319 if (prepare_kswapd_sleep(pgdat, order, remaining,
3320 balanced_classzone_idx)) {
3321 /*
3322 * Compaction records what page blocks it recently failed to
3323 * isolate pages from and skips them in the future scanning.
3324 * When kswapd is going to sleep, it is reasonable to assume
3325 * that pages and compaction may succeed so reset the cache.
3326 */
3327 reset_isolation_suitable(pgdat);
3328
3329 /*
3330 * We have freed the memory, now we should compact it to make
3331 * allocation of the requested order possible.
3332 */
3333 wakeup_kcompactd(pgdat, order, classzone_idx);
3334
3335 remaining = schedule_timeout(HZ/10);
3336 finish_wait(&pgdat->kswapd_wait, &wait);
3337 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3338 }
3339
3340 /*
3341 * After a short sleep, check if it was a premature sleep. If not, then
3342 * go fully to sleep until explicitly woken up.
3343 */
3344 if (prepare_kswapd_sleep(pgdat, order, remaining,
3345 balanced_classzone_idx)) {
3346 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3347
3348 /*
3349 * vmstat counters are not perfectly accurate and the estimated
3350 * value for counters such as NR_FREE_PAGES can deviate from the
3351 * true value by nr_online_cpus * threshold. To avoid the zone
3352 * watermarks being breached while under pressure, we reduce the
3353 * per-cpu vmstat threshold while kswapd is awake and restore
3354 * them before going back to sleep.
3355 */
3356 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3357
3358 if (!kthread_should_stop())
3359 schedule();
3360
3361 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3362 } else {
3363 if (remaining)
3364 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3365 else
3366 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3367 }
3368 finish_wait(&pgdat->kswapd_wait, &wait);
3369}
3370
3371/*
3372 * The background pageout daemon, started as a kernel thread
3373 * from the init process.
3374 *
3375 * This basically trickles out pages so that we have _some_
3376 * free memory available even if there is no other activity
3377 * that frees anything up. This is needed for things like routing
3378 * etc, where we otherwise might have all activity going on in
3379 * asynchronous contexts that cannot page things out.
3380 *
3381 * If there are applications that are active memory-allocators
3382 * (most normal use), this basically shouldn't matter.
3383 */
3384static int kswapd(void *p)
3385{
3386 unsigned long order, new_order;
3387 int classzone_idx, new_classzone_idx;
3388 int balanced_classzone_idx;
3389 pg_data_t *pgdat = (pg_data_t*)p;
3390 struct task_struct *tsk = current;
3391
3392 struct reclaim_state reclaim_state = {
3393 .reclaimed_slab = 0,
3394 };
3395 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3396
3397 lockdep_set_current_reclaim_state(GFP_KERNEL);
3398
3399 if (!cpumask_empty(cpumask))
3400 set_cpus_allowed_ptr(tsk, cpumask);
3401 current->reclaim_state = &reclaim_state;
3402
3403 /*
3404 * Tell the memory management that we're a "memory allocator",
3405 * and that if we need more memory we should get access to it
3406 * regardless (see "__alloc_pages()"). "kswapd" should
3407 * never get caught in the normal page freeing logic.
3408 *
3409 * (Kswapd normally doesn't need memory anyway, but sometimes
3410 * you need a small amount of memory in order to be able to
3411 * page out something else, and this flag essentially protects
3412 * us from recursively trying to free more memory as we're
3413 * trying to free the first piece of memory in the first place).
3414 */
3415 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3416 set_freezable();
3417
3418 order = new_order = 0;
3419 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3420 balanced_classzone_idx = classzone_idx;
3421 for ( ; ; ) {
3422 bool ret;
3423
3424 /*
3425 * While we were reclaiming, there might have been another
3426 * wakeup, so check the values.
3427 */
3428 new_order = pgdat->kswapd_max_order;
3429 new_classzone_idx = pgdat->classzone_idx;
3430 pgdat->kswapd_max_order = 0;
3431 pgdat->classzone_idx = pgdat->nr_zones - 1;
3432
3433 if (order < new_order || classzone_idx > new_classzone_idx) {
3434 /*
3435 * Don't sleep if someone wants a larger 'order'
3436 * allocation or has tigher zone constraints
3437 */
3438 order = new_order;
3439 classzone_idx = new_classzone_idx;
3440 } else {
3441 kswapd_try_to_sleep(pgdat, order, classzone_idx,
3442 balanced_classzone_idx);
3443 order = pgdat->kswapd_max_order;
3444 classzone_idx = pgdat->classzone_idx;
3445 new_order = order;
3446 new_classzone_idx = classzone_idx;
3447 pgdat->kswapd_max_order = 0;
3448 pgdat->classzone_idx = pgdat->nr_zones - 1;
3449 }
3450
3451 ret = try_to_freeze();
3452 if (kthread_should_stop())
3453 break;
3454
3455 /*
3456 * We can speed up thawing tasks if we don't call balance_pgdat
3457 * after returning from the refrigerator
3458 */
3459 if (!ret) {
3460 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3461 balanced_classzone_idx = balance_pgdat(pgdat, order,
3462 classzone_idx);
3463 }
3464 }
3465
3466 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3467 current->reclaim_state = NULL;
3468 lockdep_clear_current_reclaim_state();
3469
3470 return 0;
3471}
3472
3473/*
3474 * A zone is low on free memory, so wake its kswapd task to service it.
3475 */
3476void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3477{
3478 pg_data_t *pgdat;
3479
3480 if (!populated_zone(zone))
3481 return;
3482
3483 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3484 return;
3485 pgdat = zone->zone_pgdat;
3486 if (pgdat->kswapd_max_order < order) {
3487 pgdat->kswapd_max_order = order;
3488 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3489 }
3490 if (!waitqueue_active(&pgdat->kswapd_wait))
3491 return;
3492 if (zone_balanced(zone, order, true, 0, 0))
3493 return;
3494
3495 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3496 wake_up_interruptible(&pgdat->kswapd_wait);
3497}
3498
3499#ifdef CONFIG_HIBERNATION
3500/*
3501 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3502 * freed pages.
3503 *
3504 * Rather than trying to age LRUs the aim is to preserve the overall
3505 * LRU order by reclaiming preferentially
3506 * inactive > active > active referenced > active mapped
3507 */
3508unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3509{
3510 struct reclaim_state reclaim_state;
3511 struct scan_control sc = {
3512 .nr_to_reclaim = nr_to_reclaim,
3513 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3514 .priority = DEF_PRIORITY,
3515 .may_writepage = 1,
3516 .may_unmap = 1,
3517 .may_swap = 1,
3518 .hibernation_mode = 1,
3519 };
3520 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3521 struct task_struct *p = current;
3522 unsigned long nr_reclaimed;
3523
3524 p->flags |= PF_MEMALLOC;
3525 lockdep_set_current_reclaim_state(sc.gfp_mask);
3526 reclaim_state.reclaimed_slab = 0;
3527 p->reclaim_state = &reclaim_state;
3528
3529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3530
3531 p->reclaim_state = NULL;
3532 lockdep_clear_current_reclaim_state();
3533 p->flags &= ~PF_MEMALLOC;
3534
3535 return nr_reclaimed;
3536}
3537#endif /* CONFIG_HIBERNATION */
3538
3539/* It's optimal to keep kswapds on the same CPUs as their memory, but
3540 not required for correctness. So if the last cpu in a node goes
3541 away, we get changed to run anywhere: as the first one comes back,
3542 restore their cpu bindings. */
3543static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3544 void *hcpu)
3545{
3546 int nid;
3547
3548 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3549 for_each_node_state(nid, N_MEMORY) {
3550 pg_data_t *pgdat = NODE_DATA(nid);
3551 const struct cpumask *mask;
3552
3553 mask = cpumask_of_node(pgdat->node_id);
3554
3555 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3556 /* One of our CPUs online: restore mask */
3557 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3558 }
3559 }
3560 return NOTIFY_OK;
3561}
3562
3563/*
3564 * This kswapd start function will be called by init and node-hot-add.
3565 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3566 */
3567int kswapd_run(int nid)
3568{
3569 pg_data_t *pgdat = NODE_DATA(nid);
3570 int ret = 0;
3571
3572 if (pgdat->kswapd)
3573 return 0;
3574
3575 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3576 if (IS_ERR(pgdat->kswapd)) {
3577 /* failure at boot is fatal */
3578 BUG_ON(system_state == SYSTEM_BOOTING);
3579 pr_err("Failed to start kswapd on node %d\n", nid);
3580 ret = PTR_ERR(pgdat->kswapd);
3581 pgdat->kswapd = NULL;
3582 }
3583 return ret;
3584}
3585
3586/*
3587 * Called by memory hotplug when all memory in a node is offlined. Caller must
3588 * hold mem_hotplug_begin/end().
3589 */
3590void kswapd_stop(int nid)
3591{
3592 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3593
3594 if (kswapd) {
3595 kthread_stop(kswapd);
3596 NODE_DATA(nid)->kswapd = NULL;
3597 }
3598}
3599
3600static int __init kswapd_init(void)
3601{
3602 int nid;
3603
3604 swap_setup();
3605 for_each_node_state(nid, N_MEMORY)
3606 kswapd_run(nid);
3607 hotcpu_notifier(cpu_callback, 0);
3608 return 0;
3609}
3610
3611module_init(kswapd_init)
3612
3613#ifdef CONFIG_NUMA
3614/*
3615 * Zone reclaim mode
3616 *
3617 * If non-zero call zone_reclaim when the number of free pages falls below
3618 * the watermarks.
3619 */
3620int zone_reclaim_mode __read_mostly;
3621
3622#define RECLAIM_OFF 0
3623#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3624#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3625#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
3626
3627/*
3628 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3629 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3630 * a zone.
3631 */
3632#define ZONE_RECLAIM_PRIORITY 4
3633
3634/*
3635 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3636 * occur.
3637 */
3638int sysctl_min_unmapped_ratio = 1;
3639
3640/*
3641 * If the number of slab pages in a zone grows beyond this percentage then
3642 * slab reclaim needs to occur.
3643 */
3644int sysctl_min_slab_ratio = 5;
3645
3646static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3647{
3648 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3649 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3650 zone_page_state(zone, NR_ACTIVE_FILE);
3651
3652 /*
3653 * It's possible for there to be more file mapped pages than
3654 * accounted for by the pages on the file LRU lists because
3655 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3656 */
3657 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3658}
3659
3660/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3661static unsigned long zone_pagecache_reclaimable(struct zone *zone)
3662{
3663 unsigned long nr_pagecache_reclaimable;
3664 unsigned long delta = 0;
3665
3666 /*
3667 * If RECLAIM_UNMAP is set, then all file pages are considered
3668 * potentially reclaimable. Otherwise, we have to worry about
3669 * pages like swapcache and zone_unmapped_file_pages() provides
3670 * a better estimate
3671 */
3672 if (zone_reclaim_mode & RECLAIM_UNMAP)
3673 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3674 else
3675 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3676
3677 /* If we can't clean pages, remove dirty pages from consideration */
3678 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3679 delta += zone_page_state(zone, NR_FILE_DIRTY);
3680
3681 /* Watch for any possible underflows due to delta */
3682 if (unlikely(delta > nr_pagecache_reclaimable))
3683 delta = nr_pagecache_reclaimable;
3684
3685 return nr_pagecache_reclaimable - delta;
3686}
3687
3688/*
3689 * Try to free up some pages from this zone through reclaim.
3690 */
3691static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3692{
3693 /* Minimum pages needed in order to stay on node */
3694 const unsigned long nr_pages = 1 << order;
3695 struct task_struct *p = current;
3696 struct reclaim_state reclaim_state;
3697 struct scan_control sc = {
3698 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3699 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3700 .order = order,
3701 .priority = ZONE_RECLAIM_PRIORITY,
3702 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3703 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
3704 .may_swap = 1,
3705 };
3706
3707 cond_resched();
3708 /*
3709 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3710 * and we also need to be able to write out pages for RECLAIM_WRITE
3711 * and RECLAIM_UNMAP.
3712 */
3713 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3714 lockdep_set_current_reclaim_state(gfp_mask);
3715 reclaim_state.reclaimed_slab = 0;
3716 p->reclaim_state = &reclaim_state;
3717
3718 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3719 /*
3720 * Free memory by calling shrink zone with increasing
3721 * priorities until we have enough memory freed.
3722 */
3723 do {
3724 shrink_zone(zone, &sc, true);
3725 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3726 }
3727
3728 p->reclaim_state = NULL;
3729 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3730 lockdep_clear_current_reclaim_state();
3731 return sc.nr_reclaimed >= nr_pages;
3732}
3733
3734int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3735{
3736 int node_id;
3737 int ret;
3738
3739 /*
3740 * Zone reclaim reclaims unmapped file backed pages and
3741 * slab pages if we are over the defined limits.
3742 *
3743 * A small portion of unmapped file backed pages is needed for
3744 * file I/O otherwise pages read by file I/O will be immediately
3745 * thrown out if the zone is overallocated. So we do not reclaim
3746 * if less than a specified percentage of the zone is used by
3747 * unmapped file backed pages.
3748 */
3749 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3750 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3751 return ZONE_RECLAIM_FULL;
3752
3753 if (!zone_reclaimable(zone))
3754 return ZONE_RECLAIM_FULL;
3755
3756 /*
3757 * Do not scan if the allocation should not be delayed.
3758 */
3759 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3760 return ZONE_RECLAIM_NOSCAN;
3761
3762 /*
3763 * Only run zone reclaim on the local zone or on zones that do not
3764 * have associated processors. This will favor the local processor
3765 * over remote processors and spread off node memory allocations
3766 * as wide as possible.
3767 */
3768 node_id = zone_to_nid(zone);
3769 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3770 return ZONE_RECLAIM_NOSCAN;
3771
3772 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3773 return ZONE_RECLAIM_NOSCAN;
3774
3775 ret = __zone_reclaim(zone, gfp_mask, order);
3776 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3777
3778 if (!ret)
3779 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3780
3781 return ret;
3782}
3783#endif
3784
3785/*
3786 * page_evictable - test whether a page is evictable
3787 * @page: the page to test
3788 *
3789 * Test whether page is evictable--i.e., should be placed on active/inactive
3790 * lists vs unevictable list.
3791 *
3792 * Reasons page might not be evictable:
3793 * (1) page's mapping marked unevictable
3794 * (2) page is part of an mlocked VMA
3795 *
3796 */
3797int page_evictable(struct page *page)
3798{
3799 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3800}
3801
3802#ifdef CONFIG_SHMEM
3803/**
3804 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3805 * @pages: array of pages to check
3806 * @nr_pages: number of pages to check
3807 *
3808 * Checks pages for evictability and moves them to the appropriate lru list.
3809 *
3810 * This function is only used for SysV IPC SHM_UNLOCK.
3811 */
3812void check_move_unevictable_pages(struct page **pages, int nr_pages)
3813{
3814 struct lruvec *lruvec;
3815 struct zone *zone = NULL;
3816 int pgscanned = 0;
3817 int pgrescued = 0;
3818 int i;
3819
3820 for (i = 0; i < nr_pages; i++) {
3821 struct page *page = pages[i];
3822 struct zone *pagezone;
3823
3824 pgscanned++;
3825 pagezone = page_zone(page);
3826 if (pagezone != zone) {
3827 if (zone)
3828 spin_unlock_irq(&zone->lru_lock);
3829 zone = pagezone;
3830 spin_lock_irq(&zone->lru_lock);
3831 }
3832 lruvec = mem_cgroup_page_lruvec(page, zone);
3833
3834 if (!PageLRU(page) || !PageUnevictable(page))
3835 continue;
3836
3837 if (page_evictable(page)) {
3838 enum lru_list lru = page_lru_base_type(page);
3839
3840 VM_BUG_ON_PAGE(PageActive(page), page);
3841 ClearPageUnevictable(page);
3842 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3843 add_page_to_lru_list(page, lruvec, lru);
3844 pgrescued++;
3845 }
3846 }
3847
3848 if (zone) {
3849 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3850 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3851 spin_unlock_irq(&zone->lru_lock);
3852 }
3853}
3854#endif /* CONFIG_SHMEM */
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
73struct scan_control {
74 /* How many pages shrink_list() should reclaim */
75 unsigned long nr_to_reclaim;
76
77 /*
78 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * are scanned.
80 */
81 nodemask_t *nodemask;
82
83 /*
84 * The memory cgroup that hit its limit and as a result is the
85 * primary target of this reclaim invocation.
86 */
87 struct mem_cgroup *target_mem_cgroup;
88
89 /*
90 * Scan pressure balancing between anon and file LRUs
91 */
92 unsigned long anon_cost;
93 unsigned long file_cost;
94
95 /* Can active folios be deactivated as part of reclaim? */
96#define DEACTIVATE_ANON 1
97#define DEACTIVATE_FILE 2
98 unsigned int may_deactivate:2;
99 unsigned int force_deactivate:1;
100 unsigned int skipped_deactivate:1;
101
102 /* Writepage batching in laptop mode; RECLAIM_WRITE */
103 unsigned int may_writepage:1;
104
105 /* Can mapped folios be reclaimed? */
106 unsigned int may_unmap:1;
107
108 /* Can folios be swapped as part of reclaim? */
109 unsigned int may_swap:1;
110
111 /* Proactive reclaim invoked by userspace through memory.reclaim */
112 unsigned int proactive:1;
113
114 /*
115 * Cgroup memory below memory.low is protected as long as we
116 * don't threaten to OOM. If any cgroup is reclaimed at
117 * reduced force or passed over entirely due to its memory.low
118 * setting (memcg_low_skipped), and nothing is reclaimed as a
119 * result, then go back for one more cycle that reclaims the protected
120 * memory (memcg_low_reclaim) to avert OOM.
121 */
122 unsigned int memcg_low_reclaim:1;
123 unsigned int memcg_low_skipped:1;
124
125 unsigned int hibernation_mode:1;
126
127 /* One of the zones is ready for compaction */
128 unsigned int compaction_ready:1;
129
130 /* There is easily reclaimable cold cache in the current node */
131 unsigned int cache_trim_mode:1;
132
133 /* The file folios on the current node are dangerously low */
134 unsigned int file_is_tiny:1;
135
136 /* Always discard instead of demoting to lower tier memory */
137 unsigned int no_demotion:1;
138
139 /* Allocation order */
140 s8 order;
141
142 /* Scan (total_size >> priority) pages at once */
143 s8 priority;
144
145 /* The highest zone to isolate folios for reclaim from */
146 s8 reclaim_idx;
147
148 /* This context's GFP mask */
149 gfp_t gfp_mask;
150
151 /* Incremented by the number of inactive pages that were scanned */
152 unsigned long nr_scanned;
153
154 /* Number of pages freed so far during a call to shrink_zones() */
155 unsigned long nr_reclaimed;
156
157 struct {
158 unsigned int dirty;
159 unsigned int unqueued_dirty;
160 unsigned int congested;
161 unsigned int writeback;
162 unsigned int immediate;
163 unsigned int file_taken;
164 unsigned int taken;
165 } nr;
166
167 /* for recording the reclaimed slab by now */
168 struct reclaim_state reclaim_state;
169};
170
171#ifdef ARCH_HAS_PREFETCHW
172#define prefetchw_prev_lru_folio(_folio, _base, _field) \
173 do { \
174 if ((_folio)->lru.prev != _base) { \
175 struct folio *prev; \
176 \
177 prev = lru_to_folio(&(_folio->lru)); \
178 prefetchw(&prev->_field); \
179 } \
180 } while (0)
181#else
182#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
183#endif
184
185/*
186 * From 0 .. 200. Higher means more swappy.
187 */
188int vm_swappiness = 60;
189
190#ifdef CONFIG_MEMCG
191
192/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
193static bool cgroup_reclaim(struct scan_control *sc)
194{
195 return sc->target_mem_cgroup;
196}
197
198/*
199 * Returns true for reclaim on the root cgroup. This is true for direct
200 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
201 */
202static bool root_reclaim(struct scan_control *sc)
203{
204 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
205}
206
207/**
208 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
209 * @sc: scan_control in question
210 *
211 * The normal page dirty throttling mechanism in balance_dirty_pages() is
212 * completely broken with the legacy memcg and direct stalling in
213 * shrink_folio_list() is used for throttling instead, which lacks all the
214 * niceties such as fairness, adaptive pausing, bandwidth proportional
215 * allocation and configurability.
216 *
217 * This function tests whether the vmscan currently in progress can assume
218 * that the normal dirty throttling mechanism is operational.
219 */
220static bool writeback_throttling_sane(struct scan_control *sc)
221{
222 if (!cgroup_reclaim(sc))
223 return true;
224#ifdef CONFIG_CGROUP_WRITEBACK
225 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
226 return true;
227#endif
228 return false;
229}
230#else
231static bool cgroup_reclaim(struct scan_control *sc)
232{
233 return false;
234}
235
236static bool root_reclaim(struct scan_control *sc)
237{
238 return true;
239}
240
241static bool writeback_throttling_sane(struct scan_control *sc)
242{
243 return true;
244}
245#endif
246
247static void set_task_reclaim_state(struct task_struct *task,
248 struct reclaim_state *rs)
249{
250 /* Check for an overwrite */
251 WARN_ON_ONCE(rs && task->reclaim_state);
252
253 /* Check for the nulling of an already-nulled member */
254 WARN_ON_ONCE(!rs && !task->reclaim_state);
255
256 task->reclaim_state = rs;
257}
258
259/*
260 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
261 * scan_control->nr_reclaimed.
262 */
263static void flush_reclaim_state(struct scan_control *sc)
264{
265 /*
266 * Currently, reclaim_state->reclaimed includes three types of pages
267 * freed outside of vmscan:
268 * (1) Slab pages.
269 * (2) Clean file pages from pruned inodes (on highmem systems).
270 * (3) XFS freed buffer pages.
271 *
272 * For all of these cases, we cannot universally link the pages to a
273 * single memcg. For example, a memcg-aware shrinker can free one object
274 * charged to the target memcg, causing an entire page to be freed.
275 * If we count the entire page as reclaimed from the memcg, we end up
276 * overestimating the reclaimed amount (potentially under-reclaiming).
277 *
278 * Only count such pages for global reclaim to prevent under-reclaiming
279 * from the target memcg; preventing unnecessary retries during memcg
280 * charging and false positives from proactive reclaim.
281 *
282 * For uncommon cases where the freed pages were actually mostly
283 * charged to the target memcg, we end up underestimating the reclaimed
284 * amount. This should be fine. The freed pages will be uncharged
285 * anyway, even if they are not counted here properly, and we will be
286 * able to make forward progress in charging (which is usually in a
287 * retry loop).
288 *
289 * We can go one step further, and report the uncharged objcg pages in
290 * memcg reclaim, to make reporting more accurate and reduce
291 * underestimation, but it's probably not worth the complexity for now.
292 */
293 if (current->reclaim_state && root_reclaim(sc)) {
294 sc->nr_reclaimed += current->reclaim_state->reclaimed;
295 current->reclaim_state->reclaimed = 0;
296 }
297}
298
299static bool can_demote(int nid, struct scan_control *sc)
300{
301 if (!numa_demotion_enabled)
302 return false;
303 if (sc && sc->no_demotion)
304 return false;
305 if (next_demotion_node(nid) == NUMA_NO_NODE)
306 return false;
307
308 return true;
309}
310
311static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
312 int nid,
313 struct scan_control *sc)
314{
315 if (memcg == NULL) {
316 /*
317 * For non-memcg reclaim, is there
318 * space in any swap device?
319 */
320 if (get_nr_swap_pages() > 0)
321 return true;
322 } else {
323 /* Is the memcg below its swap limit? */
324 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
325 return true;
326 }
327
328 /*
329 * The page can not be swapped.
330 *
331 * Can it be reclaimed from this node via demotion?
332 */
333 return can_demote(nid, sc);
334}
335
336/*
337 * This misses isolated folios which are not accounted for to save counters.
338 * As the data only determines if reclaim or compaction continues, it is
339 * not expected that isolated folios will be a dominating factor.
340 */
341unsigned long zone_reclaimable_pages(struct zone *zone)
342{
343 unsigned long nr;
344
345 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
346 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
347 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
348 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
349 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
350
351 return nr;
352}
353
354/**
355 * lruvec_lru_size - Returns the number of pages on the given LRU list.
356 * @lruvec: lru vector
357 * @lru: lru to use
358 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
359 */
360static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
361 int zone_idx)
362{
363 unsigned long size = 0;
364 int zid;
365
366 for (zid = 0; zid <= zone_idx; zid++) {
367 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
368
369 if (!managed_zone(zone))
370 continue;
371
372 if (!mem_cgroup_disabled())
373 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
374 else
375 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
376 }
377 return size;
378}
379
380static unsigned long drop_slab_node(int nid)
381{
382 unsigned long freed = 0;
383 struct mem_cgroup *memcg = NULL;
384
385 memcg = mem_cgroup_iter(NULL, NULL, NULL);
386 do {
387 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
388 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
389
390 return freed;
391}
392
393void drop_slab(void)
394{
395 int nid;
396 int shift = 0;
397 unsigned long freed;
398
399 do {
400 freed = 0;
401 for_each_online_node(nid) {
402 if (fatal_signal_pending(current))
403 return;
404
405 freed += drop_slab_node(nid);
406 }
407 } while ((freed >> shift++) > 1);
408}
409
410static int reclaimer_offset(void)
411{
412 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
413 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
414 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
415 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
416 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
417 PGSCAN_DIRECT - PGSCAN_KSWAPD);
418 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
419 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
420
421 if (current_is_kswapd())
422 return 0;
423 if (current_is_khugepaged())
424 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
425 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
426}
427
428static inline int is_page_cache_freeable(struct folio *folio)
429{
430 /*
431 * A freeable page cache folio is referenced only by the caller
432 * that isolated the folio, the page cache and optional filesystem
433 * private data at folio->private.
434 */
435 return folio_ref_count(folio) - folio_test_private(folio) ==
436 1 + folio_nr_pages(folio);
437}
438
439/*
440 * We detected a synchronous write error writing a folio out. Probably
441 * -ENOSPC. We need to propagate that into the address_space for a subsequent
442 * fsync(), msync() or close().
443 *
444 * The tricky part is that after writepage we cannot touch the mapping: nothing
445 * prevents it from being freed up. But we have a ref on the folio and once
446 * that folio is locked, the mapping is pinned.
447 *
448 * We're allowed to run sleeping folio_lock() here because we know the caller has
449 * __GFP_FS.
450 */
451static void handle_write_error(struct address_space *mapping,
452 struct folio *folio, int error)
453{
454 folio_lock(folio);
455 if (folio_mapping(folio) == mapping)
456 mapping_set_error(mapping, error);
457 folio_unlock(folio);
458}
459
460static bool skip_throttle_noprogress(pg_data_t *pgdat)
461{
462 int reclaimable = 0, write_pending = 0;
463 int i;
464
465 /*
466 * If kswapd is disabled, reschedule if necessary but do not
467 * throttle as the system is likely near OOM.
468 */
469 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
470 return true;
471
472 /*
473 * If there are a lot of dirty/writeback folios then do not
474 * throttle as throttling will occur when the folios cycle
475 * towards the end of the LRU if still under writeback.
476 */
477 for (i = 0; i < MAX_NR_ZONES; i++) {
478 struct zone *zone = pgdat->node_zones + i;
479
480 if (!managed_zone(zone))
481 continue;
482
483 reclaimable += zone_reclaimable_pages(zone);
484 write_pending += zone_page_state_snapshot(zone,
485 NR_ZONE_WRITE_PENDING);
486 }
487 if (2 * write_pending <= reclaimable)
488 return true;
489
490 return false;
491}
492
493void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
494{
495 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
496 long timeout, ret;
497 DEFINE_WAIT(wait);
498
499 /*
500 * Do not throttle user workers, kthreads other than kswapd or
501 * workqueues. They may be required for reclaim to make
502 * forward progress (e.g. journalling workqueues or kthreads).
503 */
504 if (!current_is_kswapd() &&
505 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
506 cond_resched();
507 return;
508 }
509
510 /*
511 * These figures are pulled out of thin air.
512 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
513 * parallel reclaimers which is a short-lived event so the timeout is
514 * short. Failing to make progress or waiting on writeback are
515 * potentially long-lived events so use a longer timeout. This is shaky
516 * logic as a failure to make progress could be due to anything from
517 * writeback to a slow device to excessive referenced folios at the tail
518 * of the inactive LRU.
519 */
520 switch(reason) {
521 case VMSCAN_THROTTLE_WRITEBACK:
522 timeout = HZ/10;
523
524 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
525 WRITE_ONCE(pgdat->nr_reclaim_start,
526 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
527 }
528
529 break;
530 case VMSCAN_THROTTLE_CONGESTED:
531 fallthrough;
532 case VMSCAN_THROTTLE_NOPROGRESS:
533 if (skip_throttle_noprogress(pgdat)) {
534 cond_resched();
535 return;
536 }
537
538 timeout = 1;
539
540 break;
541 case VMSCAN_THROTTLE_ISOLATED:
542 timeout = HZ/50;
543 break;
544 default:
545 WARN_ON_ONCE(1);
546 timeout = HZ;
547 break;
548 }
549
550 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
551 ret = schedule_timeout(timeout);
552 finish_wait(wqh, &wait);
553
554 if (reason == VMSCAN_THROTTLE_WRITEBACK)
555 atomic_dec(&pgdat->nr_writeback_throttled);
556
557 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
558 jiffies_to_usecs(timeout - ret),
559 reason);
560}
561
562/*
563 * Account for folios written if tasks are throttled waiting on dirty
564 * folios to clean. If enough folios have been cleaned since throttling
565 * started then wakeup the throttled tasks.
566 */
567void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
568 int nr_throttled)
569{
570 unsigned long nr_written;
571
572 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
573
574 /*
575 * This is an inaccurate read as the per-cpu deltas may not
576 * be synchronised. However, given that the system is
577 * writeback throttled, it is not worth taking the penalty
578 * of getting an accurate count. At worst, the throttle
579 * timeout guarantees forward progress.
580 */
581 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
582 READ_ONCE(pgdat->nr_reclaim_start);
583
584 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
585 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
586}
587
588/* possible outcome of pageout() */
589typedef enum {
590 /* failed to write folio out, folio is locked */
591 PAGE_KEEP,
592 /* move folio to the active list, folio is locked */
593 PAGE_ACTIVATE,
594 /* folio has been sent to the disk successfully, folio is unlocked */
595 PAGE_SUCCESS,
596 /* folio is clean and locked */
597 PAGE_CLEAN,
598} pageout_t;
599
600/*
601 * pageout is called by shrink_folio_list() for each dirty folio.
602 * Calls ->writepage().
603 */
604static pageout_t pageout(struct folio *folio, struct address_space *mapping,
605 struct swap_iocb **plug)
606{
607 /*
608 * If the folio is dirty, only perform writeback if that write
609 * will be non-blocking. To prevent this allocation from being
610 * stalled by pagecache activity. But note that there may be
611 * stalls if we need to run get_block(). We could test
612 * PagePrivate for that.
613 *
614 * If this process is currently in __generic_file_write_iter() against
615 * this folio's queue, we can perform writeback even if that
616 * will block.
617 *
618 * If the folio is swapcache, write it back even if that would
619 * block, for some throttling. This happens by accident, because
620 * swap_backing_dev_info is bust: it doesn't reflect the
621 * congestion state of the swapdevs. Easy to fix, if needed.
622 */
623 if (!is_page_cache_freeable(folio))
624 return PAGE_KEEP;
625 if (!mapping) {
626 /*
627 * Some data journaling orphaned folios can have
628 * folio->mapping == NULL while being dirty with clean buffers.
629 */
630 if (folio_test_private(folio)) {
631 if (try_to_free_buffers(folio)) {
632 folio_clear_dirty(folio);
633 pr_info("%s: orphaned folio\n", __func__);
634 return PAGE_CLEAN;
635 }
636 }
637 return PAGE_KEEP;
638 }
639 if (mapping->a_ops->writepage == NULL)
640 return PAGE_ACTIVATE;
641
642 if (folio_clear_dirty_for_io(folio)) {
643 int res;
644 struct writeback_control wbc = {
645 .sync_mode = WB_SYNC_NONE,
646 .nr_to_write = SWAP_CLUSTER_MAX,
647 .range_start = 0,
648 .range_end = LLONG_MAX,
649 .for_reclaim = 1,
650 .swap_plug = plug,
651 };
652
653 folio_set_reclaim(folio);
654 res = mapping->a_ops->writepage(&folio->page, &wbc);
655 if (res < 0)
656 handle_write_error(mapping, folio, res);
657 if (res == AOP_WRITEPAGE_ACTIVATE) {
658 folio_clear_reclaim(folio);
659 return PAGE_ACTIVATE;
660 }
661
662 if (!folio_test_writeback(folio)) {
663 /* synchronous write or broken a_ops? */
664 folio_clear_reclaim(folio);
665 }
666 trace_mm_vmscan_write_folio(folio);
667 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
668 return PAGE_SUCCESS;
669 }
670
671 return PAGE_CLEAN;
672}
673
674/*
675 * Same as remove_mapping, but if the folio is removed from the mapping, it
676 * gets returned with a refcount of 0.
677 */
678static int __remove_mapping(struct address_space *mapping, struct folio *folio,
679 bool reclaimed, struct mem_cgroup *target_memcg)
680{
681 int refcount;
682 void *shadow = NULL;
683
684 BUG_ON(!folio_test_locked(folio));
685 BUG_ON(mapping != folio_mapping(folio));
686
687 if (!folio_test_swapcache(folio))
688 spin_lock(&mapping->host->i_lock);
689 xa_lock_irq(&mapping->i_pages);
690 /*
691 * The non racy check for a busy folio.
692 *
693 * Must be careful with the order of the tests. When someone has
694 * a ref to the folio, it may be possible that they dirty it then
695 * drop the reference. So if the dirty flag is tested before the
696 * refcount here, then the following race may occur:
697 *
698 * get_user_pages(&page);
699 * [user mapping goes away]
700 * write_to(page);
701 * !folio_test_dirty(folio) [good]
702 * folio_set_dirty(folio);
703 * folio_put(folio);
704 * !refcount(folio) [good, discard it]
705 *
706 * [oops, our write_to data is lost]
707 *
708 * Reversing the order of the tests ensures such a situation cannot
709 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
710 * load is not satisfied before that of folio->_refcount.
711 *
712 * Note that if the dirty flag is always set via folio_mark_dirty,
713 * and thus under the i_pages lock, then this ordering is not required.
714 */
715 refcount = 1 + folio_nr_pages(folio);
716 if (!folio_ref_freeze(folio, refcount))
717 goto cannot_free;
718 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
719 if (unlikely(folio_test_dirty(folio))) {
720 folio_ref_unfreeze(folio, refcount);
721 goto cannot_free;
722 }
723
724 if (folio_test_swapcache(folio)) {
725 swp_entry_t swap = folio->swap;
726
727 if (reclaimed && !mapping_exiting(mapping))
728 shadow = workingset_eviction(folio, target_memcg);
729 __delete_from_swap_cache(folio, swap, shadow);
730 mem_cgroup_swapout(folio, swap);
731 xa_unlock_irq(&mapping->i_pages);
732 put_swap_folio(folio, swap);
733 } else {
734 void (*free_folio)(struct folio *);
735
736 free_folio = mapping->a_ops->free_folio;
737 /*
738 * Remember a shadow entry for reclaimed file cache in
739 * order to detect refaults, thus thrashing, later on.
740 *
741 * But don't store shadows in an address space that is
742 * already exiting. This is not just an optimization,
743 * inode reclaim needs to empty out the radix tree or
744 * the nodes are lost. Don't plant shadows behind its
745 * back.
746 *
747 * We also don't store shadows for DAX mappings because the
748 * only page cache folios found in these are zero pages
749 * covering holes, and because we don't want to mix DAX
750 * exceptional entries and shadow exceptional entries in the
751 * same address_space.
752 */
753 if (reclaimed && folio_is_file_lru(folio) &&
754 !mapping_exiting(mapping) && !dax_mapping(mapping))
755 shadow = workingset_eviction(folio, target_memcg);
756 __filemap_remove_folio(folio, shadow);
757 xa_unlock_irq(&mapping->i_pages);
758 if (mapping_shrinkable(mapping))
759 inode_add_lru(mapping->host);
760 spin_unlock(&mapping->host->i_lock);
761
762 if (free_folio)
763 free_folio(folio);
764 }
765
766 return 1;
767
768cannot_free:
769 xa_unlock_irq(&mapping->i_pages);
770 if (!folio_test_swapcache(folio))
771 spin_unlock(&mapping->host->i_lock);
772 return 0;
773}
774
775/**
776 * remove_mapping() - Attempt to remove a folio from its mapping.
777 * @mapping: The address space.
778 * @folio: The folio to remove.
779 *
780 * If the folio is dirty, under writeback or if someone else has a ref
781 * on it, removal will fail.
782 * Return: The number of pages removed from the mapping. 0 if the folio
783 * could not be removed.
784 * Context: The caller should have a single refcount on the folio and
785 * hold its lock.
786 */
787long remove_mapping(struct address_space *mapping, struct folio *folio)
788{
789 if (__remove_mapping(mapping, folio, false, NULL)) {
790 /*
791 * Unfreezing the refcount with 1 effectively
792 * drops the pagecache ref for us without requiring another
793 * atomic operation.
794 */
795 folio_ref_unfreeze(folio, 1);
796 return folio_nr_pages(folio);
797 }
798 return 0;
799}
800
801/**
802 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
803 * @folio: Folio to be returned to an LRU list.
804 *
805 * Add previously isolated @folio to appropriate LRU list.
806 * The folio may still be unevictable for other reasons.
807 *
808 * Context: lru_lock must not be held, interrupts must be enabled.
809 */
810void folio_putback_lru(struct folio *folio)
811{
812 folio_add_lru(folio);
813 folio_put(folio); /* drop ref from isolate */
814}
815
816enum folio_references {
817 FOLIOREF_RECLAIM,
818 FOLIOREF_RECLAIM_CLEAN,
819 FOLIOREF_KEEP,
820 FOLIOREF_ACTIVATE,
821};
822
823static enum folio_references folio_check_references(struct folio *folio,
824 struct scan_control *sc)
825{
826 int referenced_ptes, referenced_folio;
827 unsigned long vm_flags;
828
829 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
830 &vm_flags);
831 referenced_folio = folio_test_clear_referenced(folio);
832
833 /*
834 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
835 * Let the folio, now marked Mlocked, be moved to the unevictable list.
836 */
837 if (vm_flags & VM_LOCKED)
838 return FOLIOREF_ACTIVATE;
839
840 /* rmap lock contention: rotate */
841 if (referenced_ptes == -1)
842 return FOLIOREF_KEEP;
843
844 if (referenced_ptes) {
845 /*
846 * All mapped folios start out with page table
847 * references from the instantiating fault, so we need
848 * to look twice if a mapped file/anon folio is used more
849 * than once.
850 *
851 * Mark it and spare it for another trip around the
852 * inactive list. Another page table reference will
853 * lead to its activation.
854 *
855 * Note: the mark is set for activated folios as well
856 * so that recently deactivated but used folios are
857 * quickly recovered.
858 */
859 folio_set_referenced(folio);
860
861 if (referenced_folio || referenced_ptes > 1)
862 return FOLIOREF_ACTIVATE;
863
864 /*
865 * Activate file-backed executable folios after first usage.
866 */
867 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
868 return FOLIOREF_ACTIVATE;
869
870 return FOLIOREF_KEEP;
871 }
872
873 /* Reclaim if clean, defer dirty folios to writeback */
874 if (referenced_folio && folio_is_file_lru(folio))
875 return FOLIOREF_RECLAIM_CLEAN;
876
877 return FOLIOREF_RECLAIM;
878}
879
880/* Check if a folio is dirty or under writeback */
881static void folio_check_dirty_writeback(struct folio *folio,
882 bool *dirty, bool *writeback)
883{
884 struct address_space *mapping;
885
886 /*
887 * Anonymous folios are not handled by flushers and must be written
888 * from reclaim context. Do not stall reclaim based on them.
889 * MADV_FREE anonymous folios are put into inactive file list too.
890 * They could be mistakenly treated as file lru. So further anon
891 * test is needed.
892 */
893 if (!folio_is_file_lru(folio) ||
894 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
895 *dirty = false;
896 *writeback = false;
897 return;
898 }
899
900 /* By default assume that the folio flags are accurate */
901 *dirty = folio_test_dirty(folio);
902 *writeback = folio_test_writeback(folio);
903
904 /* Verify dirty/writeback state if the filesystem supports it */
905 if (!folio_test_private(folio))
906 return;
907
908 mapping = folio_mapping(folio);
909 if (mapping && mapping->a_ops->is_dirty_writeback)
910 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
911}
912
913static struct folio *alloc_demote_folio(struct folio *src,
914 unsigned long private)
915{
916 struct folio *dst;
917 nodemask_t *allowed_mask;
918 struct migration_target_control *mtc;
919
920 mtc = (struct migration_target_control *)private;
921
922 allowed_mask = mtc->nmask;
923 /*
924 * make sure we allocate from the target node first also trying to
925 * demote or reclaim pages from the target node via kswapd if we are
926 * low on free memory on target node. If we don't do this and if
927 * we have free memory on the slower(lower) memtier, we would start
928 * allocating pages from slower(lower) memory tiers without even forcing
929 * a demotion of cold pages from the target memtier. This can result
930 * in the kernel placing hot pages in slower(lower) memory tiers.
931 */
932 mtc->nmask = NULL;
933 mtc->gfp_mask |= __GFP_THISNODE;
934 dst = alloc_migration_target(src, (unsigned long)mtc);
935 if (dst)
936 return dst;
937
938 mtc->gfp_mask &= ~__GFP_THISNODE;
939 mtc->nmask = allowed_mask;
940
941 return alloc_migration_target(src, (unsigned long)mtc);
942}
943
944/*
945 * Take folios on @demote_folios and attempt to demote them to another node.
946 * Folios which are not demoted are left on @demote_folios.
947 */
948static unsigned int demote_folio_list(struct list_head *demote_folios,
949 struct pglist_data *pgdat)
950{
951 int target_nid = next_demotion_node(pgdat->node_id);
952 unsigned int nr_succeeded;
953 nodemask_t allowed_mask;
954
955 struct migration_target_control mtc = {
956 /*
957 * Allocate from 'node', or fail quickly and quietly.
958 * When this happens, 'page' will likely just be discarded
959 * instead of migrated.
960 */
961 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
962 __GFP_NOMEMALLOC | GFP_NOWAIT,
963 .nid = target_nid,
964 .nmask = &allowed_mask
965 };
966
967 if (list_empty(demote_folios))
968 return 0;
969
970 if (target_nid == NUMA_NO_NODE)
971 return 0;
972
973 node_get_allowed_targets(pgdat, &allowed_mask);
974
975 /* Demotion ignores all cpuset and mempolicy settings */
976 migrate_pages(demote_folios, alloc_demote_folio, NULL,
977 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
978 &nr_succeeded);
979
980 mod_node_page_state(pgdat, PGDEMOTE_KSWAPD + reclaimer_offset(),
981 nr_succeeded);
982
983 return nr_succeeded;
984}
985
986static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
987{
988 if (gfp_mask & __GFP_FS)
989 return true;
990 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
991 return false;
992 /*
993 * We can "enter_fs" for swap-cache with only __GFP_IO
994 * providing this isn't SWP_FS_OPS.
995 * ->flags can be updated non-atomicially (scan_swap_map_slots),
996 * but that will never affect SWP_FS_OPS, so the data_race
997 * is safe.
998 */
999 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1000}
1001
1002/*
1003 * shrink_folio_list() returns the number of reclaimed pages
1004 */
1005static unsigned int shrink_folio_list(struct list_head *folio_list,
1006 struct pglist_data *pgdat, struct scan_control *sc,
1007 struct reclaim_stat *stat, bool ignore_references)
1008{
1009 LIST_HEAD(ret_folios);
1010 LIST_HEAD(free_folios);
1011 LIST_HEAD(demote_folios);
1012 unsigned int nr_reclaimed = 0;
1013 unsigned int pgactivate = 0;
1014 bool do_demote_pass;
1015 struct swap_iocb *plug = NULL;
1016
1017 memset(stat, 0, sizeof(*stat));
1018 cond_resched();
1019 do_demote_pass = can_demote(pgdat->node_id, sc);
1020
1021retry:
1022 while (!list_empty(folio_list)) {
1023 struct address_space *mapping;
1024 struct folio *folio;
1025 enum folio_references references = FOLIOREF_RECLAIM;
1026 bool dirty, writeback;
1027 unsigned int nr_pages;
1028
1029 cond_resched();
1030
1031 folio = lru_to_folio(folio_list);
1032 list_del(&folio->lru);
1033
1034 if (!folio_trylock(folio))
1035 goto keep;
1036
1037 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1038
1039 nr_pages = folio_nr_pages(folio);
1040
1041 /* Account the number of base pages */
1042 sc->nr_scanned += nr_pages;
1043
1044 if (unlikely(!folio_evictable(folio)))
1045 goto activate_locked;
1046
1047 if (!sc->may_unmap && folio_mapped(folio))
1048 goto keep_locked;
1049
1050 /* folio_update_gen() tried to promote this page? */
1051 if (lru_gen_enabled() && !ignore_references &&
1052 folio_mapped(folio) && folio_test_referenced(folio))
1053 goto keep_locked;
1054
1055 /*
1056 * The number of dirty pages determines if a node is marked
1057 * reclaim_congested. kswapd will stall and start writing
1058 * folios if the tail of the LRU is all dirty unqueued folios.
1059 */
1060 folio_check_dirty_writeback(folio, &dirty, &writeback);
1061 if (dirty || writeback)
1062 stat->nr_dirty += nr_pages;
1063
1064 if (dirty && !writeback)
1065 stat->nr_unqueued_dirty += nr_pages;
1066
1067 /*
1068 * Treat this folio as congested if folios are cycling
1069 * through the LRU so quickly that the folios marked
1070 * for immediate reclaim are making it to the end of
1071 * the LRU a second time.
1072 */
1073 if (writeback && folio_test_reclaim(folio))
1074 stat->nr_congested += nr_pages;
1075
1076 /*
1077 * If a folio at the tail of the LRU is under writeback, there
1078 * are three cases to consider.
1079 *
1080 * 1) If reclaim is encountering an excessive number
1081 * of folios under writeback and this folio has both
1082 * the writeback and reclaim flags set, then it
1083 * indicates that folios are being queued for I/O but
1084 * are being recycled through the LRU before the I/O
1085 * can complete. Waiting on the folio itself risks an
1086 * indefinite stall if it is impossible to writeback
1087 * the folio due to I/O error or disconnected storage
1088 * so instead note that the LRU is being scanned too
1089 * quickly and the caller can stall after the folio
1090 * list has been processed.
1091 *
1092 * 2) Global or new memcg reclaim encounters a folio that is
1093 * not marked for immediate reclaim, or the caller does not
1094 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1095 * not to fs). In this case mark the folio for immediate
1096 * reclaim and continue scanning.
1097 *
1098 * Require may_enter_fs() because we would wait on fs, which
1099 * may not have submitted I/O yet. And the loop driver might
1100 * enter reclaim, and deadlock if it waits on a folio for
1101 * which it is needed to do the write (loop masks off
1102 * __GFP_IO|__GFP_FS for this reason); but more thought
1103 * would probably show more reasons.
1104 *
1105 * 3) Legacy memcg encounters a folio that already has the
1106 * reclaim flag set. memcg does not have any dirty folio
1107 * throttling so we could easily OOM just because too many
1108 * folios are in writeback and there is nothing else to
1109 * reclaim. Wait for the writeback to complete.
1110 *
1111 * In cases 1) and 2) we activate the folios to get them out of
1112 * the way while we continue scanning for clean folios on the
1113 * inactive list and refilling from the active list. The
1114 * observation here is that waiting for disk writes is more
1115 * expensive than potentially causing reloads down the line.
1116 * Since they're marked for immediate reclaim, they won't put
1117 * memory pressure on the cache working set any longer than it
1118 * takes to write them to disk.
1119 */
1120 if (folio_test_writeback(folio)) {
1121 /* Case 1 above */
1122 if (current_is_kswapd() &&
1123 folio_test_reclaim(folio) &&
1124 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1125 stat->nr_immediate += nr_pages;
1126 goto activate_locked;
1127
1128 /* Case 2 above */
1129 } else if (writeback_throttling_sane(sc) ||
1130 !folio_test_reclaim(folio) ||
1131 !may_enter_fs(folio, sc->gfp_mask)) {
1132 /*
1133 * This is slightly racy -
1134 * folio_end_writeback() might have
1135 * just cleared the reclaim flag, then
1136 * setting the reclaim flag here ends up
1137 * interpreted as the readahead flag - but
1138 * that does not matter enough to care.
1139 * What we do want is for this folio to
1140 * have the reclaim flag set next time
1141 * memcg reclaim reaches the tests above,
1142 * so it will then wait for writeback to
1143 * avoid OOM; and it's also appropriate
1144 * in global reclaim.
1145 */
1146 folio_set_reclaim(folio);
1147 stat->nr_writeback += nr_pages;
1148 goto activate_locked;
1149
1150 /* Case 3 above */
1151 } else {
1152 folio_unlock(folio);
1153 folio_wait_writeback(folio);
1154 /* then go back and try same folio again */
1155 list_add_tail(&folio->lru, folio_list);
1156 continue;
1157 }
1158 }
1159
1160 if (!ignore_references)
1161 references = folio_check_references(folio, sc);
1162
1163 switch (references) {
1164 case FOLIOREF_ACTIVATE:
1165 goto activate_locked;
1166 case FOLIOREF_KEEP:
1167 stat->nr_ref_keep += nr_pages;
1168 goto keep_locked;
1169 case FOLIOREF_RECLAIM:
1170 case FOLIOREF_RECLAIM_CLEAN:
1171 ; /* try to reclaim the folio below */
1172 }
1173
1174 /*
1175 * Before reclaiming the folio, try to relocate
1176 * its contents to another node.
1177 */
1178 if (do_demote_pass &&
1179 (thp_migration_supported() || !folio_test_large(folio))) {
1180 list_add(&folio->lru, &demote_folios);
1181 folio_unlock(folio);
1182 continue;
1183 }
1184
1185 /*
1186 * Anonymous process memory has backing store?
1187 * Try to allocate it some swap space here.
1188 * Lazyfree folio could be freed directly
1189 */
1190 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1191 if (!folio_test_swapcache(folio)) {
1192 if (!(sc->gfp_mask & __GFP_IO))
1193 goto keep_locked;
1194 if (folio_maybe_dma_pinned(folio))
1195 goto keep_locked;
1196 if (folio_test_large(folio)) {
1197 /* cannot split folio, skip it */
1198 if (!can_split_folio(folio, NULL))
1199 goto activate_locked;
1200 /*
1201 * Split folios without a PMD map right
1202 * away. Chances are some or all of the
1203 * tail pages can be freed without IO.
1204 */
1205 if (!folio_entire_mapcount(folio) &&
1206 split_folio_to_list(folio,
1207 folio_list))
1208 goto activate_locked;
1209 }
1210 if (!add_to_swap(folio)) {
1211 if (!folio_test_large(folio))
1212 goto activate_locked_split;
1213 /* Fallback to swap normal pages */
1214 if (split_folio_to_list(folio,
1215 folio_list))
1216 goto activate_locked;
1217#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1218 count_memcg_folio_events(folio, THP_SWPOUT_FALLBACK, 1);
1219 count_vm_event(THP_SWPOUT_FALLBACK);
1220#endif
1221 if (!add_to_swap(folio))
1222 goto activate_locked_split;
1223 }
1224 }
1225 } else if (folio_test_swapbacked(folio) &&
1226 folio_test_large(folio)) {
1227 /* Split shmem folio */
1228 if (split_folio_to_list(folio, folio_list))
1229 goto keep_locked;
1230 }
1231
1232 /*
1233 * If the folio was split above, the tail pages will make
1234 * their own pass through this function and be accounted
1235 * then.
1236 */
1237 if ((nr_pages > 1) && !folio_test_large(folio)) {
1238 sc->nr_scanned -= (nr_pages - 1);
1239 nr_pages = 1;
1240 }
1241
1242 /*
1243 * The folio is mapped into the page tables of one or more
1244 * processes. Try to unmap it here.
1245 */
1246 if (folio_mapped(folio)) {
1247 enum ttu_flags flags = TTU_BATCH_FLUSH;
1248 bool was_swapbacked = folio_test_swapbacked(folio);
1249
1250 if (folio_test_pmd_mappable(folio))
1251 flags |= TTU_SPLIT_HUGE_PMD;
1252
1253 try_to_unmap(folio, flags);
1254 if (folio_mapped(folio)) {
1255 stat->nr_unmap_fail += nr_pages;
1256 if (!was_swapbacked &&
1257 folio_test_swapbacked(folio))
1258 stat->nr_lazyfree_fail += nr_pages;
1259 goto activate_locked;
1260 }
1261 }
1262
1263 /*
1264 * Folio is unmapped now so it cannot be newly pinned anymore.
1265 * No point in trying to reclaim folio if it is pinned.
1266 * Furthermore we don't want to reclaim underlying fs metadata
1267 * if the folio is pinned and thus potentially modified by the
1268 * pinning process as that may upset the filesystem.
1269 */
1270 if (folio_maybe_dma_pinned(folio))
1271 goto activate_locked;
1272
1273 mapping = folio_mapping(folio);
1274 if (folio_test_dirty(folio)) {
1275 /*
1276 * Only kswapd can writeback filesystem folios
1277 * to avoid risk of stack overflow. But avoid
1278 * injecting inefficient single-folio I/O into
1279 * flusher writeback as much as possible: only
1280 * write folios when we've encountered many
1281 * dirty folios, and when we've already scanned
1282 * the rest of the LRU for clean folios and see
1283 * the same dirty folios again (with the reclaim
1284 * flag set).
1285 */
1286 if (folio_is_file_lru(folio) &&
1287 (!current_is_kswapd() ||
1288 !folio_test_reclaim(folio) ||
1289 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1290 /*
1291 * Immediately reclaim when written back.
1292 * Similar in principle to folio_deactivate()
1293 * except we already have the folio isolated
1294 * and know it's dirty
1295 */
1296 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1297 nr_pages);
1298 folio_set_reclaim(folio);
1299
1300 goto activate_locked;
1301 }
1302
1303 if (references == FOLIOREF_RECLAIM_CLEAN)
1304 goto keep_locked;
1305 if (!may_enter_fs(folio, sc->gfp_mask))
1306 goto keep_locked;
1307 if (!sc->may_writepage)
1308 goto keep_locked;
1309
1310 /*
1311 * Folio is dirty. Flush the TLB if a writable entry
1312 * potentially exists to avoid CPU writes after I/O
1313 * starts and then write it out here.
1314 */
1315 try_to_unmap_flush_dirty();
1316 switch (pageout(folio, mapping, &plug)) {
1317 case PAGE_KEEP:
1318 goto keep_locked;
1319 case PAGE_ACTIVATE:
1320 goto activate_locked;
1321 case PAGE_SUCCESS:
1322 stat->nr_pageout += nr_pages;
1323
1324 if (folio_test_writeback(folio))
1325 goto keep;
1326 if (folio_test_dirty(folio))
1327 goto keep;
1328
1329 /*
1330 * A synchronous write - probably a ramdisk. Go
1331 * ahead and try to reclaim the folio.
1332 */
1333 if (!folio_trylock(folio))
1334 goto keep;
1335 if (folio_test_dirty(folio) ||
1336 folio_test_writeback(folio))
1337 goto keep_locked;
1338 mapping = folio_mapping(folio);
1339 fallthrough;
1340 case PAGE_CLEAN:
1341 ; /* try to free the folio below */
1342 }
1343 }
1344
1345 /*
1346 * If the folio has buffers, try to free the buffer
1347 * mappings associated with this folio. If we succeed
1348 * we try to free the folio as well.
1349 *
1350 * We do this even if the folio is dirty.
1351 * filemap_release_folio() does not perform I/O, but it
1352 * is possible for a folio to have the dirty flag set,
1353 * but it is actually clean (all its buffers are clean).
1354 * This happens if the buffers were written out directly,
1355 * with submit_bh(). ext3 will do this, as well as
1356 * the blockdev mapping. filemap_release_folio() will
1357 * discover that cleanness and will drop the buffers
1358 * and mark the folio clean - it can be freed.
1359 *
1360 * Rarely, folios can have buffers and no ->mapping.
1361 * These are the folios which were not successfully
1362 * invalidated in truncate_cleanup_folio(). We try to
1363 * drop those buffers here and if that worked, and the
1364 * folio is no longer mapped into process address space
1365 * (refcount == 1) it can be freed. Otherwise, leave
1366 * the folio on the LRU so it is swappable.
1367 */
1368 if (folio_needs_release(folio)) {
1369 if (!filemap_release_folio(folio, sc->gfp_mask))
1370 goto activate_locked;
1371 if (!mapping && folio_ref_count(folio) == 1) {
1372 folio_unlock(folio);
1373 if (folio_put_testzero(folio))
1374 goto free_it;
1375 else {
1376 /*
1377 * rare race with speculative reference.
1378 * the speculative reference will free
1379 * this folio shortly, so we may
1380 * increment nr_reclaimed here (and
1381 * leave it off the LRU).
1382 */
1383 nr_reclaimed += nr_pages;
1384 continue;
1385 }
1386 }
1387 }
1388
1389 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1390 /* follow __remove_mapping for reference */
1391 if (!folio_ref_freeze(folio, 1))
1392 goto keep_locked;
1393 /*
1394 * The folio has only one reference left, which is
1395 * from the isolation. After the caller puts the
1396 * folio back on the lru and drops the reference, the
1397 * folio will be freed anyway. It doesn't matter
1398 * which lru it goes on. So we don't bother checking
1399 * the dirty flag here.
1400 */
1401 count_vm_events(PGLAZYFREED, nr_pages);
1402 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1403 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1404 sc->target_mem_cgroup))
1405 goto keep_locked;
1406
1407 folio_unlock(folio);
1408free_it:
1409 /*
1410 * Folio may get swapped out as a whole, need to account
1411 * all pages in it.
1412 */
1413 nr_reclaimed += nr_pages;
1414
1415 /*
1416 * Is there need to periodically free_folio_list? It would
1417 * appear not as the counts should be low
1418 */
1419 if (unlikely(folio_test_large(folio)))
1420 destroy_large_folio(folio);
1421 else
1422 list_add(&folio->lru, &free_folios);
1423 continue;
1424
1425activate_locked_split:
1426 /*
1427 * The tail pages that are failed to add into swap cache
1428 * reach here. Fixup nr_scanned and nr_pages.
1429 */
1430 if (nr_pages > 1) {
1431 sc->nr_scanned -= (nr_pages - 1);
1432 nr_pages = 1;
1433 }
1434activate_locked:
1435 /* Not a candidate for swapping, so reclaim swap space. */
1436 if (folio_test_swapcache(folio) &&
1437 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1438 folio_free_swap(folio);
1439 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1440 if (!folio_test_mlocked(folio)) {
1441 int type = folio_is_file_lru(folio);
1442 folio_set_active(folio);
1443 stat->nr_activate[type] += nr_pages;
1444 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1445 }
1446keep_locked:
1447 folio_unlock(folio);
1448keep:
1449 list_add(&folio->lru, &ret_folios);
1450 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1451 folio_test_unevictable(folio), folio);
1452 }
1453 /* 'folio_list' is always empty here */
1454
1455 /* Migrate folios selected for demotion */
1456 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
1457 /* Folios that could not be demoted are still in @demote_folios */
1458 if (!list_empty(&demote_folios)) {
1459 /* Folios which weren't demoted go back on @folio_list */
1460 list_splice_init(&demote_folios, folio_list);
1461
1462 /*
1463 * goto retry to reclaim the undemoted folios in folio_list if
1464 * desired.
1465 *
1466 * Reclaiming directly from top tier nodes is not often desired
1467 * due to it breaking the LRU ordering: in general memory
1468 * should be reclaimed from lower tier nodes and demoted from
1469 * top tier nodes.
1470 *
1471 * However, disabling reclaim from top tier nodes entirely
1472 * would cause ooms in edge scenarios where lower tier memory
1473 * is unreclaimable for whatever reason, eg memory being
1474 * mlocked or too hot to reclaim. We can disable reclaim
1475 * from top tier nodes in proactive reclaim though as that is
1476 * not real memory pressure.
1477 */
1478 if (!sc->proactive) {
1479 do_demote_pass = false;
1480 goto retry;
1481 }
1482 }
1483
1484 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1485
1486 mem_cgroup_uncharge_list(&free_folios);
1487 try_to_unmap_flush();
1488 free_unref_page_list(&free_folios);
1489
1490 list_splice(&ret_folios, folio_list);
1491 count_vm_events(PGACTIVATE, pgactivate);
1492
1493 if (plug)
1494 swap_write_unplug(plug);
1495 return nr_reclaimed;
1496}
1497
1498unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1499 struct list_head *folio_list)
1500{
1501 struct scan_control sc = {
1502 .gfp_mask = GFP_KERNEL,
1503 .may_unmap = 1,
1504 };
1505 struct reclaim_stat stat;
1506 unsigned int nr_reclaimed;
1507 struct folio *folio, *next;
1508 LIST_HEAD(clean_folios);
1509 unsigned int noreclaim_flag;
1510
1511 list_for_each_entry_safe(folio, next, folio_list, lru) {
1512 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1513 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1514 !folio_test_unevictable(folio)) {
1515 folio_clear_active(folio);
1516 list_move(&folio->lru, &clean_folios);
1517 }
1518 }
1519
1520 /*
1521 * We should be safe here since we are only dealing with file pages and
1522 * we are not kswapd and therefore cannot write dirty file pages. But
1523 * call memalloc_noreclaim_save() anyway, just in case these conditions
1524 * change in the future.
1525 */
1526 noreclaim_flag = memalloc_noreclaim_save();
1527 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1528 &stat, true);
1529 memalloc_noreclaim_restore(noreclaim_flag);
1530
1531 list_splice(&clean_folios, folio_list);
1532 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1533 -(long)nr_reclaimed);
1534 /*
1535 * Since lazyfree pages are isolated from file LRU from the beginning,
1536 * they will rotate back to anonymous LRU in the end if it failed to
1537 * discard so isolated count will be mismatched.
1538 * Compensate the isolated count for both LRU lists.
1539 */
1540 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1541 stat.nr_lazyfree_fail);
1542 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1543 -(long)stat.nr_lazyfree_fail);
1544 return nr_reclaimed;
1545}
1546
1547/*
1548 * Update LRU sizes after isolating pages. The LRU size updates must
1549 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1550 */
1551static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1552 enum lru_list lru, unsigned long *nr_zone_taken)
1553{
1554 int zid;
1555
1556 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1557 if (!nr_zone_taken[zid])
1558 continue;
1559
1560 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1561 }
1562
1563}
1564
1565#ifdef CONFIG_CMA
1566/*
1567 * It is waste of effort to scan and reclaim CMA pages if it is not available
1568 * for current allocation context. Kswapd can not be enrolled as it can not
1569 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1570 */
1571static bool skip_cma(struct folio *folio, struct scan_control *sc)
1572{
1573 return !current_is_kswapd() &&
1574 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1575 folio_migratetype(folio) == MIGRATE_CMA;
1576}
1577#else
1578static bool skip_cma(struct folio *folio, struct scan_control *sc)
1579{
1580 return false;
1581}
1582#endif
1583
1584/*
1585 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1586 *
1587 * lruvec->lru_lock is heavily contended. Some of the functions that
1588 * shrink the lists perform better by taking out a batch of pages
1589 * and working on them outside the LRU lock.
1590 *
1591 * For pagecache intensive workloads, this function is the hottest
1592 * spot in the kernel (apart from copy_*_user functions).
1593 *
1594 * Lru_lock must be held before calling this function.
1595 *
1596 * @nr_to_scan: The number of eligible pages to look through on the list.
1597 * @lruvec: The LRU vector to pull pages from.
1598 * @dst: The temp list to put pages on to.
1599 * @nr_scanned: The number of pages that were scanned.
1600 * @sc: The scan_control struct for this reclaim session
1601 * @lru: LRU list id for isolating
1602 *
1603 * returns how many pages were moved onto *@dst.
1604 */
1605static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1606 struct lruvec *lruvec, struct list_head *dst,
1607 unsigned long *nr_scanned, struct scan_control *sc,
1608 enum lru_list lru)
1609{
1610 struct list_head *src = &lruvec->lists[lru];
1611 unsigned long nr_taken = 0;
1612 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1613 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1614 unsigned long skipped = 0;
1615 unsigned long scan, total_scan, nr_pages;
1616 LIST_HEAD(folios_skipped);
1617
1618 total_scan = 0;
1619 scan = 0;
1620 while (scan < nr_to_scan && !list_empty(src)) {
1621 struct list_head *move_to = src;
1622 struct folio *folio;
1623
1624 folio = lru_to_folio(src);
1625 prefetchw_prev_lru_folio(folio, src, flags);
1626
1627 nr_pages = folio_nr_pages(folio);
1628 total_scan += nr_pages;
1629
1630 if (folio_zonenum(folio) > sc->reclaim_idx ||
1631 skip_cma(folio, sc)) {
1632 nr_skipped[folio_zonenum(folio)] += nr_pages;
1633 move_to = &folios_skipped;
1634 goto move;
1635 }
1636
1637 /*
1638 * Do not count skipped folios because that makes the function
1639 * return with no isolated folios if the LRU mostly contains
1640 * ineligible folios. This causes the VM to not reclaim any
1641 * folios, triggering a premature OOM.
1642 * Account all pages in a folio.
1643 */
1644 scan += nr_pages;
1645
1646 if (!folio_test_lru(folio))
1647 goto move;
1648 if (!sc->may_unmap && folio_mapped(folio))
1649 goto move;
1650
1651 /*
1652 * Be careful not to clear the lru flag until after we're
1653 * sure the folio is not being freed elsewhere -- the
1654 * folio release code relies on it.
1655 */
1656 if (unlikely(!folio_try_get(folio)))
1657 goto move;
1658
1659 if (!folio_test_clear_lru(folio)) {
1660 /* Another thread is already isolating this folio */
1661 folio_put(folio);
1662 goto move;
1663 }
1664
1665 nr_taken += nr_pages;
1666 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1667 move_to = dst;
1668move:
1669 list_move(&folio->lru, move_to);
1670 }
1671
1672 /*
1673 * Splice any skipped folios to the start of the LRU list. Note that
1674 * this disrupts the LRU order when reclaiming for lower zones but
1675 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1676 * scanning would soon rescan the same folios to skip and waste lots
1677 * of cpu cycles.
1678 */
1679 if (!list_empty(&folios_skipped)) {
1680 int zid;
1681
1682 list_splice(&folios_skipped, src);
1683 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1684 if (!nr_skipped[zid])
1685 continue;
1686
1687 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1688 skipped += nr_skipped[zid];
1689 }
1690 }
1691 *nr_scanned = total_scan;
1692 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1693 total_scan, skipped, nr_taken, lru);
1694 update_lru_sizes(lruvec, lru, nr_zone_taken);
1695 return nr_taken;
1696}
1697
1698/**
1699 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1700 * @folio: Folio to isolate from its LRU list.
1701 *
1702 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1703 * corresponding to whatever LRU list the folio was on.
1704 *
1705 * The folio will have its LRU flag cleared. If it was found on the
1706 * active list, it will have the Active flag set. If it was found on the
1707 * unevictable list, it will have the Unevictable flag set. These flags
1708 * may need to be cleared by the caller before letting the page go.
1709 *
1710 * Context:
1711 *
1712 * (1) Must be called with an elevated refcount on the folio. This is a
1713 * fundamental difference from isolate_lru_folios() (which is called
1714 * without a stable reference).
1715 * (2) The lru_lock must not be held.
1716 * (3) Interrupts must be enabled.
1717 *
1718 * Return: true if the folio was removed from an LRU list.
1719 * false if the folio was not on an LRU list.
1720 */
1721bool folio_isolate_lru(struct folio *folio)
1722{
1723 bool ret = false;
1724
1725 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1726
1727 if (folio_test_clear_lru(folio)) {
1728 struct lruvec *lruvec;
1729
1730 folio_get(folio);
1731 lruvec = folio_lruvec_lock_irq(folio);
1732 lruvec_del_folio(lruvec, folio);
1733 unlock_page_lruvec_irq(lruvec);
1734 ret = true;
1735 }
1736
1737 return ret;
1738}
1739
1740/*
1741 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1742 * then get rescheduled. When there are massive number of tasks doing page
1743 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1744 * the LRU list will go small and be scanned faster than necessary, leading to
1745 * unnecessary swapping, thrashing and OOM.
1746 */
1747static int too_many_isolated(struct pglist_data *pgdat, int file,
1748 struct scan_control *sc)
1749{
1750 unsigned long inactive, isolated;
1751 bool too_many;
1752
1753 if (current_is_kswapd())
1754 return 0;
1755
1756 if (!writeback_throttling_sane(sc))
1757 return 0;
1758
1759 if (file) {
1760 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1761 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1762 } else {
1763 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1764 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1765 }
1766
1767 /*
1768 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1769 * won't get blocked by normal direct-reclaimers, forming a circular
1770 * deadlock.
1771 */
1772 if (gfp_has_io_fs(sc->gfp_mask))
1773 inactive >>= 3;
1774
1775 too_many = isolated > inactive;
1776
1777 /* Wake up tasks throttled due to too_many_isolated. */
1778 if (!too_many)
1779 wake_throttle_isolated(pgdat);
1780
1781 return too_many;
1782}
1783
1784/*
1785 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1786 * On return, @list is reused as a list of folios to be freed by the caller.
1787 *
1788 * Returns the number of pages moved to the given lruvec.
1789 */
1790static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1791 struct list_head *list)
1792{
1793 int nr_pages, nr_moved = 0;
1794 LIST_HEAD(folios_to_free);
1795
1796 while (!list_empty(list)) {
1797 struct folio *folio = lru_to_folio(list);
1798
1799 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1800 list_del(&folio->lru);
1801 if (unlikely(!folio_evictable(folio))) {
1802 spin_unlock_irq(&lruvec->lru_lock);
1803 folio_putback_lru(folio);
1804 spin_lock_irq(&lruvec->lru_lock);
1805 continue;
1806 }
1807
1808 /*
1809 * The folio_set_lru needs to be kept here for list integrity.
1810 * Otherwise:
1811 * #0 move_folios_to_lru #1 release_pages
1812 * if (!folio_put_testzero())
1813 * if (folio_put_testzero())
1814 * !lru //skip lru_lock
1815 * folio_set_lru()
1816 * list_add(&folio->lru,)
1817 * list_add(&folio->lru,)
1818 */
1819 folio_set_lru(folio);
1820
1821 if (unlikely(folio_put_testzero(folio))) {
1822 __folio_clear_lru_flags(folio);
1823
1824 if (unlikely(folio_test_large(folio))) {
1825 spin_unlock_irq(&lruvec->lru_lock);
1826 destroy_large_folio(folio);
1827 spin_lock_irq(&lruvec->lru_lock);
1828 } else
1829 list_add(&folio->lru, &folios_to_free);
1830
1831 continue;
1832 }
1833
1834 /*
1835 * All pages were isolated from the same lruvec (and isolation
1836 * inhibits memcg migration).
1837 */
1838 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1839 lruvec_add_folio(lruvec, folio);
1840 nr_pages = folio_nr_pages(folio);
1841 nr_moved += nr_pages;
1842 if (folio_test_active(folio))
1843 workingset_age_nonresident(lruvec, nr_pages);
1844 }
1845
1846 /*
1847 * To save our caller's stack, now use input list for pages to free.
1848 */
1849 list_splice(&folios_to_free, list);
1850
1851 return nr_moved;
1852}
1853
1854/*
1855 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1856 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1857 * we should not throttle. Otherwise it is safe to do so.
1858 */
1859static int current_may_throttle(void)
1860{
1861 return !(current->flags & PF_LOCAL_THROTTLE);
1862}
1863
1864/*
1865 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1866 * of reclaimed pages
1867 */
1868static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1869 struct lruvec *lruvec, struct scan_control *sc,
1870 enum lru_list lru)
1871{
1872 LIST_HEAD(folio_list);
1873 unsigned long nr_scanned;
1874 unsigned int nr_reclaimed = 0;
1875 unsigned long nr_taken;
1876 struct reclaim_stat stat;
1877 bool file = is_file_lru(lru);
1878 enum vm_event_item item;
1879 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1880 bool stalled = false;
1881
1882 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1883 if (stalled)
1884 return 0;
1885
1886 /* wait a bit for the reclaimer. */
1887 stalled = true;
1888 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1889
1890 /* We are about to die and free our memory. Return now. */
1891 if (fatal_signal_pending(current))
1892 return SWAP_CLUSTER_MAX;
1893 }
1894
1895 lru_add_drain();
1896
1897 spin_lock_irq(&lruvec->lru_lock);
1898
1899 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1900 &nr_scanned, sc, lru);
1901
1902 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1903 item = PGSCAN_KSWAPD + reclaimer_offset();
1904 if (!cgroup_reclaim(sc))
1905 __count_vm_events(item, nr_scanned);
1906 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1907 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1908
1909 spin_unlock_irq(&lruvec->lru_lock);
1910
1911 if (nr_taken == 0)
1912 return 0;
1913
1914 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1915
1916 spin_lock_irq(&lruvec->lru_lock);
1917 move_folios_to_lru(lruvec, &folio_list);
1918
1919 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1920 item = PGSTEAL_KSWAPD + reclaimer_offset();
1921 if (!cgroup_reclaim(sc))
1922 __count_vm_events(item, nr_reclaimed);
1923 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1924 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1925 spin_unlock_irq(&lruvec->lru_lock);
1926
1927 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1928 mem_cgroup_uncharge_list(&folio_list);
1929 free_unref_page_list(&folio_list);
1930
1931 /*
1932 * If dirty folios are scanned that are not queued for IO, it
1933 * implies that flushers are not doing their job. This can
1934 * happen when memory pressure pushes dirty folios to the end of
1935 * the LRU before the dirty limits are breached and the dirty
1936 * data has expired. It can also happen when the proportion of
1937 * dirty folios grows not through writes but through memory
1938 * pressure reclaiming all the clean cache. And in some cases,
1939 * the flushers simply cannot keep up with the allocation
1940 * rate. Nudge the flusher threads in case they are asleep.
1941 */
1942 if (stat.nr_unqueued_dirty == nr_taken) {
1943 wakeup_flusher_threads(WB_REASON_VMSCAN);
1944 /*
1945 * For cgroupv1 dirty throttling is achieved by waking up
1946 * the kernel flusher here and later waiting on folios
1947 * which are in writeback to finish (see shrink_folio_list()).
1948 *
1949 * Flusher may not be able to issue writeback quickly
1950 * enough for cgroupv1 writeback throttling to work
1951 * on a large system.
1952 */
1953 if (!writeback_throttling_sane(sc))
1954 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1955 }
1956
1957 sc->nr.dirty += stat.nr_dirty;
1958 sc->nr.congested += stat.nr_congested;
1959 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1960 sc->nr.writeback += stat.nr_writeback;
1961 sc->nr.immediate += stat.nr_immediate;
1962 sc->nr.taken += nr_taken;
1963 if (file)
1964 sc->nr.file_taken += nr_taken;
1965
1966 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1967 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1968 return nr_reclaimed;
1969}
1970
1971/*
1972 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1973 *
1974 * We move them the other way if the folio is referenced by one or more
1975 * processes.
1976 *
1977 * If the folios are mostly unmapped, the processing is fast and it is
1978 * appropriate to hold lru_lock across the whole operation. But if
1979 * the folios are mapped, the processing is slow (folio_referenced()), so
1980 * we should drop lru_lock around each folio. It's impossible to balance
1981 * this, so instead we remove the folios from the LRU while processing them.
1982 * It is safe to rely on the active flag against the non-LRU folios in here
1983 * because nobody will play with that bit on a non-LRU folio.
1984 *
1985 * The downside is that we have to touch folio->_refcount against each folio.
1986 * But we had to alter folio->flags anyway.
1987 */
1988static void shrink_active_list(unsigned long nr_to_scan,
1989 struct lruvec *lruvec,
1990 struct scan_control *sc,
1991 enum lru_list lru)
1992{
1993 unsigned long nr_taken;
1994 unsigned long nr_scanned;
1995 unsigned long vm_flags;
1996 LIST_HEAD(l_hold); /* The folios which were snipped off */
1997 LIST_HEAD(l_active);
1998 LIST_HEAD(l_inactive);
1999 unsigned nr_deactivate, nr_activate;
2000 unsigned nr_rotated = 0;
2001 int file = is_file_lru(lru);
2002 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2003
2004 lru_add_drain();
2005
2006 spin_lock_irq(&lruvec->lru_lock);
2007
2008 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2009 &nr_scanned, sc, lru);
2010
2011 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2012
2013 if (!cgroup_reclaim(sc))
2014 __count_vm_events(PGREFILL, nr_scanned);
2015 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2016
2017 spin_unlock_irq(&lruvec->lru_lock);
2018
2019 while (!list_empty(&l_hold)) {
2020 struct folio *folio;
2021
2022 cond_resched();
2023 folio = lru_to_folio(&l_hold);
2024 list_del(&folio->lru);
2025
2026 if (unlikely(!folio_evictable(folio))) {
2027 folio_putback_lru(folio);
2028 continue;
2029 }
2030
2031 if (unlikely(buffer_heads_over_limit)) {
2032 if (folio_needs_release(folio) &&
2033 folio_trylock(folio)) {
2034 filemap_release_folio(folio, 0);
2035 folio_unlock(folio);
2036 }
2037 }
2038
2039 /* Referenced or rmap lock contention: rotate */
2040 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2041 &vm_flags) != 0) {
2042 /*
2043 * Identify referenced, file-backed active folios and
2044 * give them one more trip around the active list. So
2045 * that executable code get better chances to stay in
2046 * memory under moderate memory pressure. Anon folios
2047 * are not likely to be evicted by use-once streaming
2048 * IO, plus JVM can create lots of anon VM_EXEC folios,
2049 * so we ignore them here.
2050 */
2051 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2052 nr_rotated += folio_nr_pages(folio);
2053 list_add(&folio->lru, &l_active);
2054 continue;
2055 }
2056 }
2057
2058 folio_clear_active(folio); /* we are de-activating */
2059 folio_set_workingset(folio);
2060 list_add(&folio->lru, &l_inactive);
2061 }
2062
2063 /*
2064 * Move folios back to the lru list.
2065 */
2066 spin_lock_irq(&lruvec->lru_lock);
2067
2068 nr_activate = move_folios_to_lru(lruvec, &l_active);
2069 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2070 /* Keep all free folios in l_active list */
2071 list_splice(&l_inactive, &l_active);
2072
2073 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2074 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2075
2076 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2077 spin_unlock_irq(&lruvec->lru_lock);
2078
2079 if (nr_rotated)
2080 lru_note_cost(lruvec, file, 0, nr_rotated);
2081 mem_cgroup_uncharge_list(&l_active);
2082 free_unref_page_list(&l_active);
2083 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2084 nr_deactivate, nr_rotated, sc->priority, file);
2085}
2086
2087static unsigned int reclaim_folio_list(struct list_head *folio_list,
2088 struct pglist_data *pgdat)
2089{
2090 struct reclaim_stat dummy_stat;
2091 unsigned int nr_reclaimed;
2092 struct folio *folio;
2093 struct scan_control sc = {
2094 .gfp_mask = GFP_KERNEL,
2095 .may_writepage = 1,
2096 .may_unmap = 1,
2097 .may_swap = 1,
2098 .no_demotion = 1,
2099 };
2100
2101 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2102 while (!list_empty(folio_list)) {
2103 folio = lru_to_folio(folio_list);
2104 list_del(&folio->lru);
2105 folio_putback_lru(folio);
2106 }
2107
2108 return nr_reclaimed;
2109}
2110
2111unsigned long reclaim_pages(struct list_head *folio_list)
2112{
2113 int nid;
2114 unsigned int nr_reclaimed = 0;
2115 LIST_HEAD(node_folio_list);
2116 unsigned int noreclaim_flag;
2117
2118 if (list_empty(folio_list))
2119 return nr_reclaimed;
2120
2121 noreclaim_flag = memalloc_noreclaim_save();
2122
2123 nid = folio_nid(lru_to_folio(folio_list));
2124 do {
2125 struct folio *folio = lru_to_folio(folio_list);
2126
2127 if (nid == folio_nid(folio)) {
2128 folio_clear_active(folio);
2129 list_move(&folio->lru, &node_folio_list);
2130 continue;
2131 }
2132
2133 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2134 nid = folio_nid(lru_to_folio(folio_list));
2135 } while (!list_empty(folio_list));
2136
2137 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2138
2139 memalloc_noreclaim_restore(noreclaim_flag);
2140
2141 return nr_reclaimed;
2142}
2143
2144static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2145 struct lruvec *lruvec, struct scan_control *sc)
2146{
2147 if (is_active_lru(lru)) {
2148 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2149 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2150 else
2151 sc->skipped_deactivate = 1;
2152 return 0;
2153 }
2154
2155 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2156}
2157
2158/*
2159 * The inactive anon list should be small enough that the VM never has
2160 * to do too much work.
2161 *
2162 * The inactive file list should be small enough to leave most memory
2163 * to the established workingset on the scan-resistant active list,
2164 * but large enough to avoid thrashing the aggregate readahead window.
2165 *
2166 * Both inactive lists should also be large enough that each inactive
2167 * folio has a chance to be referenced again before it is reclaimed.
2168 *
2169 * If that fails and refaulting is observed, the inactive list grows.
2170 *
2171 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2172 * on this LRU, maintained by the pageout code. An inactive_ratio
2173 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2174 *
2175 * total target max
2176 * memory ratio inactive
2177 * -------------------------------------
2178 * 10MB 1 5MB
2179 * 100MB 1 50MB
2180 * 1GB 3 250MB
2181 * 10GB 10 0.9GB
2182 * 100GB 31 3GB
2183 * 1TB 101 10GB
2184 * 10TB 320 32GB
2185 */
2186static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2187{
2188 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2189 unsigned long inactive, active;
2190 unsigned long inactive_ratio;
2191 unsigned long gb;
2192
2193 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2194 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2195
2196 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2197 if (gb)
2198 inactive_ratio = int_sqrt(10 * gb);
2199 else
2200 inactive_ratio = 1;
2201
2202 return inactive * inactive_ratio < active;
2203}
2204
2205enum scan_balance {
2206 SCAN_EQUAL,
2207 SCAN_FRACT,
2208 SCAN_ANON,
2209 SCAN_FILE,
2210};
2211
2212static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2213{
2214 unsigned long file;
2215 struct lruvec *target_lruvec;
2216
2217 if (lru_gen_enabled())
2218 return;
2219
2220 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2221
2222 /*
2223 * Flush the memory cgroup stats, so that we read accurate per-memcg
2224 * lruvec stats for heuristics.
2225 */
2226 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2227
2228 /*
2229 * Determine the scan balance between anon and file LRUs.
2230 */
2231 spin_lock_irq(&target_lruvec->lru_lock);
2232 sc->anon_cost = target_lruvec->anon_cost;
2233 sc->file_cost = target_lruvec->file_cost;
2234 spin_unlock_irq(&target_lruvec->lru_lock);
2235
2236 /*
2237 * Target desirable inactive:active list ratios for the anon
2238 * and file LRU lists.
2239 */
2240 if (!sc->force_deactivate) {
2241 unsigned long refaults;
2242
2243 /*
2244 * When refaults are being observed, it means a new
2245 * workingset is being established. Deactivate to get
2246 * rid of any stale active pages quickly.
2247 */
2248 refaults = lruvec_page_state(target_lruvec,
2249 WORKINGSET_ACTIVATE_ANON);
2250 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2251 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2252 sc->may_deactivate |= DEACTIVATE_ANON;
2253 else
2254 sc->may_deactivate &= ~DEACTIVATE_ANON;
2255
2256 refaults = lruvec_page_state(target_lruvec,
2257 WORKINGSET_ACTIVATE_FILE);
2258 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2259 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2260 sc->may_deactivate |= DEACTIVATE_FILE;
2261 else
2262 sc->may_deactivate &= ~DEACTIVATE_FILE;
2263 } else
2264 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2265
2266 /*
2267 * If we have plenty of inactive file pages that aren't
2268 * thrashing, try to reclaim those first before touching
2269 * anonymous pages.
2270 */
2271 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2272 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2273 sc->cache_trim_mode = 1;
2274 else
2275 sc->cache_trim_mode = 0;
2276
2277 /*
2278 * Prevent the reclaimer from falling into the cache trap: as
2279 * cache pages start out inactive, every cache fault will tip
2280 * the scan balance towards the file LRU. And as the file LRU
2281 * shrinks, so does the window for rotation from references.
2282 * This means we have a runaway feedback loop where a tiny
2283 * thrashing file LRU becomes infinitely more attractive than
2284 * anon pages. Try to detect this based on file LRU size.
2285 */
2286 if (!cgroup_reclaim(sc)) {
2287 unsigned long total_high_wmark = 0;
2288 unsigned long free, anon;
2289 int z;
2290
2291 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2292 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2293 node_page_state(pgdat, NR_INACTIVE_FILE);
2294
2295 for (z = 0; z < MAX_NR_ZONES; z++) {
2296 struct zone *zone = &pgdat->node_zones[z];
2297
2298 if (!managed_zone(zone))
2299 continue;
2300
2301 total_high_wmark += high_wmark_pages(zone);
2302 }
2303
2304 /*
2305 * Consider anon: if that's low too, this isn't a
2306 * runaway file reclaim problem, but rather just
2307 * extreme pressure. Reclaim as per usual then.
2308 */
2309 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2310
2311 sc->file_is_tiny =
2312 file + free <= total_high_wmark &&
2313 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2314 anon >> sc->priority;
2315 }
2316}
2317
2318/*
2319 * Determine how aggressively the anon and file LRU lists should be
2320 * scanned.
2321 *
2322 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2323 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2324 */
2325static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2326 unsigned long *nr)
2327{
2328 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2329 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2330 unsigned long anon_cost, file_cost, total_cost;
2331 int swappiness = mem_cgroup_swappiness(memcg);
2332 u64 fraction[ANON_AND_FILE];
2333 u64 denominator = 0; /* gcc */
2334 enum scan_balance scan_balance;
2335 unsigned long ap, fp;
2336 enum lru_list lru;
2337
2338 /* If we have no swap space, do not bother scanning anon folios. */
2339 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2340 scan_balance = SCAN_FILE;
2341 goto out;
2342 }
2343
2344 /*
2345 * Global reclaim will swap to prevent OOM even with no
2346 * swappiness, but memcg users want to use this knob to
2347 * disable swapping for individual groups completely when
2348 * using the memory controller's swap limit feature would be
2349 * too expensive.
2350 */
2351 if (cgroup_reclaim(sc) && !swappiness) {
2352 scan_balance = SCAN_FILE;
2353 goto out;
2354 }
2355
2356 /*
2357 * Do not apply any pressure balancing cleverness when the
2358 * system is close to OOM, scan both anon and file equally
2359 * (unless the swappiness setting disagrees with swapping).
2360 */
2361 if (!sc->priority && swappiness) {
2362 scan_balance = SCAN_EQUAL;
2363 goto out;
2364 }
2365
2366 /*
2367 * If the system is almost out of file pages, force-scan anon.
2368 */
2369 if (sc->file_is_tiny) {
2370 scan_balance = SCAN_ANON;
2371 goto out;
2372 }
2373
2374 /*
2375 * If there is enough inactive page cache, we do not reclaim
2376 * anything from the anonymous working right now.
2377 */
2378 if (sc->cache_trim_mode) {
2379 scan_balance = SCAN_FILE;
2380 goto out;
2381 }
2382
2383 scan_balance = SCAN_FRACT;
2384 /*
2385 * Calculate the pressure balance between anon and file pages.
2386 *
2387 * The amount of pressure we put on each LRU is inversely
2388 * proportional to the cost of reclaiming each list, as
2389 * determined by the share of pages that are refaulting, times
2390 * the relative IO cost of bringing back a swapped out
2391 * anonymous page vs reloading a filesystem page (swappiness).
2392 *
2393 * Although we limit that influence to ensure no list gets
2394 * left behind completely: at least a third of the pressure is
2395 * applied, before swappiness.
2396 *
2397 * With swappiness at 100, anon and file have equal IO cost.
2398 */
2399 total_cost = sc->anon_cost + sc->file_cost;
2400 anon_cost = total_cost + sc->anon_cost;
2401 file_cost = total_cost + sc->file_cost;
2402 total_cost = anon_cost + file_cost;
2403
2404 ap = swappiness * (total_cost + 1);
2405 ap /= anon_cost + 1;
2406
2407 fp = (200 - swappiness) * (total_cost + 1);
2408 fp /= file_cost + 1;
2409
2410 fraction[0] = ap;
2411 fraction[1] = fp;
2412 denominator = ap + fp;
2413out:
2414 for_each_evictable_lru(lru) {
2415 int file = is_file_lru(lru);
2416 unsigned long lruvec_size;
2417 unsigned long low, min;
2418 unsigned long scan;
2419
2420 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2421 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2422 &min, &low);
2423
2424 if (min || low) {
2425 /*
2426 * Scale a cgroup's reclaim pressure by proportioning
2427 * its current usage to its memory.low or memory.min
2428 * setting.
2429 *
2430 * This is important, as otherwise scanning aggression
2431 * becomes extremely binary -- from nothing as we
2432 * approach the memory protection threshold, to totally
2433 * nominal as we exceed it. This results in requiring
2434 * setting extremely liberal protection thresholds. It
2435 * also means we simply get no protection at all if we
2436 * set it too low, which is not ideal.
2437 *
2438 * If there is any protection in place, we reduce scan
2439 * pressure by how much of the total memory used is
2440 * within protection thresholds.
2441 *
2442 * There is one special case: in the first reclaim pass,
2443 * we skip over all groups that are within their low
2444 * protection. If that fails to reclaim enough pages to
2445 * satisfy the reclaim goal, we come back and override
2446 * the best-effort low protection. However, we still
2447 * ideally want to honor how well-behaved groups are in
2448 * that case instead of simply punishing them all
2449 * equally. As such, we reclaim them based on how much
2450 * memory they are using, reducing the scan pressure
2451 * again by how much of the total memory used is under
2452 * hard protection.
2453 */
2454 unsigned long cgroup_size = mem_cgroup_size(memcg);
2455 unsigned long protection;
2456
2457 /* memory.low scaling, make sure we retry before OOM */
2458 if (!sc->memcg_low_reclaim && low > min) {
2459 protection = low;
2460 sc->memcg_low_skipped = 1;
2461 } else {
2462 protection = min;
2463 }
2464
2465 /* Avoid TOCTOU with earlier protection check */
2466 cgroup_size = max(cgroup_size, protection);
2467
2468 scan = lruvec_size - lruvec_size * protection /
2469 (cgroup_size + 1);
2470
2471 /*
2472 * Minimally target SWAP_CLUSTER_MAX pages to keep
2473 * reclaim moving forwards, avoiding decrementing
2474 * sc->priority further than desirable.
2475 */
2476 scan = max(scan, SWAP_CLUSTER_MAX);
2477 } else {
2478 scan = lruvec_size;
2479 }
2480
2481 scan >>= sc->priority;
2482
2483 /*
2484 * If the cgroup's already been deleted, make sure to
2485 * scrape out the remaining cache.
2486 */
2487 if (!scan && !mem_cgroup_online(memcg))
2488 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2489
2490 switch (scan_balance) {
2491 case SCAN_EQUAL:
2492 /* Scan lists relative to size */
2493 break;
2494 case SCAN_FRACT:
2495 /*
2496 * Scan types proportional to swappiness and
2497 * their relative recent reclaim efficiency.
2498 * Make sure we don't miss the last page on
2499 * the offlined memory cgroups because of a
2500 * round-off error.
2501 */
2502 scan = mem_cgroup_online(memcg) ?
2503 div64_u64(scan * fraction[file], denominator) :
2504 DIV64_U64_ROUND_UP(scan * fraction[file],
2505 denominator);
2506 break;
2507 case SCAN_FILE:
2508 case SCAN_ANON:
2509 /* Scan one type exclusively */
2510 if ((scan_balance == SCAN_FILE) != file)
2511 scan = 0;
2512 break;
2513 default:
2514 /* Look ma, no brain */
2515 BUG();
2516 }
2517
2518 nr[lru] = scan;
2519 }
2520}
2521
2522/*
2523 * Anonymous LRU management is a waste if there is
2524 * ultimately no way to reclaim the memory.
2525 */
2526static bool can_age_anon_pages(struct pglist_data *pgdat,
2527 struct scan_control *sc)
2528{
2529 /* Aging the anon LRU is valuable if swap is present: */
2530 if (total_swap_pages > 0)
2531 return true;
2532
2533 /* Also valuable if anon pages can be demoted: */
2534 return can_demote(pgdat->node_id, sc);
2535}
2536
2537#ifdef CONFIG_LRU_GEN
2538
2539#ifdef CONFIG_LRU_GEN_ENABLED
2540DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2541#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2542#else
2543DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2544#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2545#endif
2546
2547static bool should_walk_mmu(void)
2548{
2549 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2550}
2551
2552static bool should_clear_pmd_young(void)
2553{
2554 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2555}
2556
2557/******************************************************************************
2558 * shorthand helpers
2559 ******************************************************************************/
2560
2561#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2562
2563#define DEFINE_MAX_SEQ(lruvec) \
2564 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2565
2566#define DEFINE_MIN_SEQ(lruvec) \
2567 unsigned long min_seq[ANON_AND_FILE] = { \
2568 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2569 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2570 }
2571
2572#define for_each_gen_type_zone(gen, type, zone) \
2573 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2574 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2575 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2576
2577#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2578#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2579
2580static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2581{
2582 struct pglist_data *pgdat = NODE_DATA(nid);
2583
2584#ifdef CONFIG_MEMCG
2585 if (memcg) {
2586 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2587
2588 /* see the comment in mem_cgroup_lruvec() */
2589 if (!lruvec->pgdat)
2590 lruvec->pgdat = pgdat;
2591
2592 return lruvec;
2593 }
2594#endif
2595 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2596
2597 return &pgdat->__lruvec;
2598}
2599
2600static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2601{
2602 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2603 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2604
2605 if (!sc->may_swap)
2606 return 0;
2607
2608 if (!can_demote(pgdat->node_id, sc) &&
2609 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2610 return 0;
2611
2612 return mem_cgroup_swappiness(memcg);
2613}
2614
2615static int get_nr_gens(struct lruvec *lruvec, int type)
2616{
2617 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2618}
2619
2620static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2621{
2622 /* see the comment on lru_gen_folio */
2623 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2624 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2625 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2626}
2627
2628/******************************************************************************
2629 * Bloom filters
2630 ******************************************************************************/
2631
2632/*
2633 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2634 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2635 * bits in a bitmap, k is the number of hash functions and n is the number of
2636 * inserted items.
2637 *
2638 * Page table walkers use one of the two filters to reduce their search space.
2639 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2640 * aging uses the double-buffering technique to flip to the other filter each
2641 * time it produces a new generation. For non-leaf entries that have enough
2642 * leaf entries, the aging carries them over to the next generation in
2643 * walk_pmd_range(); the eviction also report them when walking the rmap
2644 * in lru_gen_look_around().
2645 *
2646 * For future optimizations:
2647 * 1. It's not necessary to keep both filters all the time. The spare one can be
2648 * freed after the RCU grace period and reallocated if needed again.
2649 * 2. And when reallocating, it's worth scaling its size according to the number
2650 * of inserted entries in the other filter, to reduce the memory overhead on
2651 * small systems and false positives on large systems.
2652 * 3. Jenkins' hash function is an alternative to Knuth's.
2653 */
2654#define BLOOM_FILTER_SHIFT 15
2655
2656static inline int filter_gen_from_seq(unsigned long seq)
2657{
2658 return seq % NR_BLOOM_FILTERS;
2659}
2660
2661static void get_item_key(void *item, int *key)
2662{
2663 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2664
2665 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2666
2667 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2668 key[1] = hash >> BLOOM_FILTER_SHIFT;
2669}
2670
2671static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2672 void *item)
2673{
2674 int key[2];
2675 unsigned long *filter;
2676 int gen = filter_gen_from_seq(seq);
2677
2678 filter = READ_ONCE(mm_state->filters[gen]);
2679 if (!filter)
2680 return true;
2681
2682 get_item_key(item, key);
2683
2684 return test_bit(key[0], filter) && test_bit(key[1], filter);
2685}
2686
2687static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2688 void *item)
2689{
2690 int key[2];
2691 unsigned long *filter;
2692 int gen = filter_gen_from_seq(seq);
2693
2694 filter = READ_ONCE(mm_state->filters[gen]);
2695 if (!filter)
2696 return;
2697
2698 get_item_key(item, key);
2699
2700 if (!test_bit(key[0], filter))
2701 set_bit(key[0], filter);
2702 if (!test_bit(key[1], filter))
2703 set_bit(key[1], filter);
2704}
2705
2706static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2707{
2708 unsigned long *filter;
2709 int gen = filter_gen_from_seq(seq);
2710
2711 filter = mm_state->filters[gen];
2712 if (filter) {
2713 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2714 return;
2715 }
2716
2717 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2718 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2719 WRITE_ONCE(mm_state->filters[gen], filter);
2720}
2721
2722/******************************************************************************
2723 * mm_struct list
2724 ******************************************************************************/
2725
2726#ifdef CONFIG_LRU_GEN_WALKS_MMU
2727
2728static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2729{
2730 static struct lru_gen_mm_list mm_list = {
2731 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2732 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2733 };
2734
2735#ifdef CONFIG_MEMCG
2736 if (memcg)
2737 return &memcg->mm_list;
2738#endif
2739 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2740
2741 return &mm_list;
2742}
2743
2744static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2745{
2746 return &lruvec->mm_state;
2747}
2748
2749static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2750{
2751 int key;
2752 struct mm_struct *mm;
2753 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2754 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2755
2756 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2757 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2758
2759 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2760 return NULL;
2761
2762 clear_bit(key, &mm->lru_gen.bitmap);
2763
2764 return mmget_not_zero(mm) ? mm : NULL;
2765}
2766
2767void lru_gen_add_mm(struct mm_struct *mm)
2768{
2769 int nid;
2770 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2771 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2772
2773 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2774#ifdef CONFIG_MEMCG
2775 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2776 mm->lru_gen.memcg = memcg;
2777#endif
2778 spin_lock(&mm_list->lock);
2779
2780 for_each_node_state(nid, N_MEMORY) {
2781 struct lruvec *lruvec = get_lruvec(memcg, nid);
2782 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2783
2784 /* the first addition since the last iteration */
2785 if (mm_state->tail == &mm_list->fifo)
2786 mm_state->tail = &mm->lru_gen.list;
2787 }
2788
2789 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2790
2791 spin_unlock(&mm_list->lock);
2792}
2793
2794void lru_gen_del_mm(struct mm_struct *mm)
2795{
2796 int nid;
2797 struct lru_gen_mm_list *mm_list;
2798 struct mem_cgroup *memcg = NULL;
2799
2800 if (list_empty(&mm->lru_gen.list))
2801 return;
2802
2803#ifdef CONFIG_MEMCG
2804 memcg = mm->lru_gen.memcg;
2805#endif
2806 mm_list = get_mm_list(memcg);
2807
2808 spin_lock(&mm_list->lock);
2809
2810 for_each_node(nid) {
2811 struct lruvec *lruvec = get_lruvec(memcg, nid);
2812 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2813
2814 /* where the current iteration continues after */
2815 if (mm_state->head == &mm->lru_gen.list)
2816 mm_state->head = mm_state->head->prev;
2817
2818 /* where the last iteration ended before */
2819 if (mm_state->tail == &mm->lru_gen.list)
2820 mm_state->tail = mm_state->tail->next;
2821 }
2822
2823 list_del_init(&mm->lru_gen.list);
2824
2825 spin_unlock(&mm_list->lock);
2826
2827#ifdef CONFIG_MEMCG
2828 mem_cgroup_put(mm->lru_gen.memcg);
2829 mm->lru_gen.memcg = NULL;
2830#endif
2831}
2832
2833#ifdef CONFIG_MEMCG
2834void lru_gen_migrate_mm(struct mm_struct *mm)
2835{
2836 struct mem_cgroup *memcg;
2837 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2838
2839 VM_WARN_ON_ONCE(task->mm != mm);
2840 lockdep_assert_held(&task->alloc_lock);
2841
2842 /* for mm_update_next_owner() */
2843 if (mem_cgroup_disabled())
2844 return;
2845
2846 /* migration can happen before addition */
2847 if (!mm->lru_gen.memcg)
2848 return;
2849
2850 rcu_read_lock();
2851 memcg = mem_cgroup_from_task(task);
2852 rcu_read_unlock();
2853 if (memcg == mm->lru_gen.memcg)
2854 return;
2855
2856 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2857
2858 lru_gen_del_mm(mm);
2859 lru_gen_add_mm(mm);
2860}
2861#endif
2862
2863#else /* !CONFIG_LRU_GEN_WALKS_MMU */
2864
2865static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2866{
2867 return NULL;
2868}
2869
2870static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2871{
2872 return NULL;
2873}
2874
2875static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2876{
2877 return NULL;
2878}
2879
2880#endif
2881
2882static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
2883{
2884 int i;
2885 int hist;
2886 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2887
2888 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2889
2890 if (walk) {
2891 hist = lru_hist_from_seq(walk->max_seq);
2892
2893 for (i = 0; i < NR_MM_STATS; i++) {
2894 WRITE_ONCE(mm_state->stats[hist][i],
2895 mm_state->stats[hist][i] + walk->mm_stats[i]);
2896 walk->mm_stats[i] = 0;
2897 }
2898 }
2899
2900 if (NR_HIST_GENS > 1 && last) {
2901 hist = lru_hist_from_seq(mm_state->seq + 1);
2902
2903 for (i = 0; i < NR_MM_STATS; i++)
2904 WRITE_ONCE(mm_state->stats[hist][i], 0);
2905 }
2906}
2907
2908static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
2909 struct mm_struct **iter)
2910{
2911 bool first = false;
2912 bool last = false;
2913 struct mm_struct *mm = NULL;
2914 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2915 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2916 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2917
2918 /*
2919 * mm_state->seq is incremented after each iteration of mm_list. There
2920 * are three interesting cases for this page table walker:
2921 * 1. It tries to start a new iteration with a stale max_seq: there is
2922 * nothing left to do.
2923 * 2. It started the next iteration: it needs to reset the Bloom filter
2924 * so that a fresh set of PTE tables can be recorded.
2925 * 3. It ended the current iteration: it needs to reset the mm stats
2926 * counters and tell its caller to increment max_seq.
2927 */
2928 spin_lock(&mm_list->lock);
2929
2930 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
2931
2932 if (walk->max_seq <= mm_state->seq)
2933 goto done;
2934
2935 if (!mm_state->head)
2936 mm_state->head = &mm_list->fifo;
2937
2938 if (mm_state->head == &mm_list->fifo)
2939 first = true;
2940
2941 do {
2942 mm_state->head = mm_state->head->next;
2943 if (mm_state->head == &mm_list->fifo) {
2944 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2945 last = true;
2946 break;
2947 }
2948
2949 /* force scan for those added after the last iteration */
2950 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2951 mm_state->tail = mm_state->head->next;
2952 walk->force_scan = true;
2953 }
2954 } while (!(mm = get_next_mm(walk)));
2955done:
2956 if (*iter || last)
2957 reset_mm_stats(lruvec, walk, last);
2958
2959 spin_unlock(&mm_list->lock);
2960
2961 if (mm && first)
2962 reset_bloom_filter(mm_state, walk->max_seq + 1);
2963
2964 if (*iter)
2965 mmput_async(*iter);
2966
2967 *iter = mm;
2968
2969 return last;
2970}
2971
2972static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
2973{
2974 bool success = false;
2975 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2976 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2977 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2978
2979 spin_lock(&mm_list->lock);
2980
2981 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
2982
2983 if (max_seq > mm_state->seq) {
2984 mm_state->head = NULL;
2985 mm_state->tail = NULL;
2986 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2987 reset_mm_stats(lruvec, NULL, true);
2988 success = true;
2989 }
2990
2991 spin_unlock(&mm_list->lock);
2992
2993 return success;
2994}
2995
2996/******************************************************************************
2997 * PID controller
2998 ******************************************************************************/
2999
3000/*
3001 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3002 *
3003 * The P term is refaulted/(evicted+protected) from a tier in the generation
3004 * currently being evicted; the I term is the exponential moving average of the
3005 * P term over the generations previously evicted, using the smoothing factor
3006 * 1/2; the D term isn't supported.
3007 *
3008 * The setpoint (SP) is always the first tier of one type; the process variable
3009 * (PV) is either any tier of the other type or any other tier of the same
3010 * type.
3011 *
3012 * The error is the difference between the SP and the PV; the correction is to
3013 * turn off protection when SP>PV or turn on protection when SP<PV.
3014 *
3015 * For future optimizations:
3016 * 1. The D term may discount the other two terms over time so that long-lived
3017 * generations can resist stale information.
3018 */
3019struct ctrl_pos {
3020 unsigned long refaulted;
3021 unsigned long total;
3022 int gain;
3023};
3024
3025static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3026 struct ctrl_pos *pos)
3027{
3028 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3029 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3030
3031 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3032 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3033 pos->total = lrugen->avg_total[type][tier] +
3034 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3035 if (tier)
3036 pos->total += lrugen->protected[hist][type][tier - 1];
3037 pos->gain = gain;
3038}
3039
3040static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3041{
3042 int hist, tier;
3043 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3044 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3045 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3046
3047 lockdep_assert_held(&lruvec->lru_lock);
3048
3049 if (!carryover && !clear)
3050 return;
3051
3052 hist = lru_hist_from_seq(seq);
3053
3054 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3055 if (carryover) {
3056 unsigned long sum;
3057
3058 sum = lrugen->avg_refaulted[type][tier] +
3059 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3060 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3061
3062 sum = lrugen->avg_total[type][tier] +
3063 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3064 if (tier)
3065 sum += lrugen->protected[hist][type][tier - 1];
3066 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3067 }
3068
3069 if (clear) {
3070 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3071 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3072 if (tier)
3073 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3074 }
3075 }
3076}
3077
3078static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3079{
3080 /*
3081 * Return true if the PV has a limited number of refaults or a lower
3082 * refaulted/total than the SP.
3083 */
3084 return pv->refaulted < MIN_LRU_BATCH ||
3085 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3086 (sp->refaulted + 1) * pv->total * pv->gain;
3087}
3088
3089/******************************************************************************
3090 * the aging
3091 ******************************************************************************/
3092
3093/* promote pages accessed through page tables */
3094static int folio_update_gen(struct folio *folio, int gen)
3095{
3096 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3097
3098 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3099 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3100
3101 do {
3102 /* lru_gen_del_folio() has isolated this page? */
3103 if (!(old_flags & LRU_GEN_MASK)) {
3104 /* for shrink_folio_list() */
3105 new_flags = old_flags | BIT(PG_referenced);
3106 continue;
3107 }
3108
3109 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3110 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3111 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3112
3113 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3114}
3115
3116/* protect pages accessed multiple times through file descriptors */
3117static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3118{
3119 int type = folio_is_file_lru(folio);
3120 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3121 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3122 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3123
3124 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3125
3126 do {
3127 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3128 /* folio_update_gen() has promoted this page? */
3129 if (new_gen >= 0 && new_gen != old_gen)
3130 return new_gen;
3131
3132 new_gen = (old_gen + 1) % MAX_NR_GENS;
3133
3134 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3135 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3136 /* for folio_end_writeback() */
3137 if (reclaiming)
3138 new_flags |= BIT(PG_reclaim);
3139 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3140
3141 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3142
3143 return new_gen;
3144}
3145
3146static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3147 int old_gen, int new_gen)
3148{
3149 int type = folio_is_file_lru(folio);
3150 int zone = folio_zonenum(folio);
3151 int delta = folio_nr_pages(folio);
3152
3153 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3154 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3155
3156 walk->batched++;
3157
3158 walk->nr_pages[old_gen][type][zone] -= delta;
3159 walk->nr_pages[new_gen][type][zone] += delta;
3160}
3161
3162static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3163{
3164 int gen, type, zone;
3165 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3166
3167 walk->batched = 0;
3168
3169 for_each_gen_type_zone(gen, type, zone) {
3170 enum lru_list lru = type * LRU_INACTIVE_FILE;
3171 int delta = walk->nr_pages[gen][type][zone];
3172
3173 if (!delta)
3174 continue;
3175
3176 walk->nr_pages[gen][type][zone] = 0;
3177 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3178 lrugen->nr_pages[gen][type][zone] + delta);
3179
3180 if (lru_gen_is_active(lruvec, gen))
3181 lru += LRU_ACTIVE;
3182 __update_lru_size(lruvec, lru, zone, delta);
3183 }
3184}
3185
3186static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3187{
3188 struct address_space *mapping;
3189 struct vm_area_struct *vma = args->vma;
3190 struct lru_gen_mm_walk *walk = args->private;
3191
3192 if (!vma_is_accessible(vma))
3193 return true;
3194
3195 if (is_vm_hugetlb_page(vma))
3196 return true;
3197
3198 if (!vma_has_recency(vma))
3199 return true;
3200
3201 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3202 return true;
3203
3204 if (vma == get_gate_vma(vma->vm_mm))
3205 return true;
3206
3207 if (vma_is_anonymous(vma))
3208 return !walk->can_swap;
3209
3210 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3211 return true;
3212
3213 mapping = vma->vm_file->f_mapping;
3214 if (mapping_unevictable(mapping))
3215 return true;
3216
3217 if (shmem_mapping(mapping))
3218 return !walk->can_swap;
3219
3220 /* to exclude special mappings like dax, etc. */
3221 return !mapping->a_ops->read_folio;
3222}
3223
3224/*
3225 * Some userspace memory allocators map many single-page VMAs. Instead of
3226 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3227 * table to reduce zigzags and improve cache performance.
3228 */
3229static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3230 unsigned long *vm_start, unsigned long *vm_end)
3231{
3232 unsigned long start = round_up(*vm_end, size);
3233 unsigned long end = (start | ~mask) + 1;
3234 VMA_ITERATOR(vmi, args->mm, start);
3235
3236 VM_WARN_ON_ONCE(mask & size);
3237 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3238
3239 for_each_vma(vmi, args->vma) {
3240 if (end && end <= args->vma->vm_start)
3241 return false;
3242
3243 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3244 continue;
3245
3246 *vm_start = max(start, args->vma->vm_start);
3247 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3248
3249 return true;
3250 }
3251
3252 return false;
3253}
3254
3255static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3256{
3257 unsigned long pfn = pte_pfn(pte);
3258
3259 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3260
3261 if (!pte_present(pte) || is_zero_pfn(pfn))
3262 return -1;
3263
3264 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3265 return -1;
3266
3267 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3268 return -1;
3269
3270 return pfn;
3271}
3272
3273static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3274{
3275 unsigned long pfn = pmd_pfn(pmd);
3276
3277 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3278
3279 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3280 return -1;
3281
3282 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3283 return -1;
3284
3285 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3286 return -1;
3287
3288 return pfn;
3289}
3290
3291static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3292 struct pglist_data *pgdat, bool can_swap)
3293{
3294 struct folio *folio;
3295
3296 /* try to avoid unnecessary memory loads */
3297 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3298 return NULL;
3299
3300 folio = pfn_folio(pfn);
3301 if (folio_nid(folio) != pgdat->node_id)
3302 return NULL;
3303
3304 if (folio_memcg_rcu(folio) != memcg)
3305 return NULL;
3306
3307 /* file VMAs can contain anon pages from COW */
3308 if (!folio_is_file_lru(folio) && !can_swap)
3309 return NULL;
3310
3311 return folio;
3312}
3313
3314static bool suitable_to_scan(int total, int young)
3315{
3316 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3317
3318 /* suitable if the average number of young PTEs per cacheline is >=1 */
3319 return young * n >= total;
3320}
3321
3322static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3323 struct mm_walk *args)
3324{
3325 int i;
3326 pte_t *pte;
3327 spinlock_t *ptl;
3328 unsigned long addr;
3329 int total = 0;
3330 int young = 0;
3331 struct lru_gen_mm_walk *walk = args->private;
3332 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3333 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3334 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3335
3336 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3337 if (!pte)
3338 return false;
3339 if (!spin_trylock(ptl)) {
3340 pte_unmap(pte);
3341 return false;
3342 }
3343
3344 arch_enter_lazy_mmu_mode();
3345restart:
3346 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3347 unsigned long pfn;
3348 struct folio *folio;
3349 pte_t ptent = ptep_get(pte + i);
3350
3351 total++;
3352 walk->mm_stats[MM_LEAF_TOTAL]++;
3353
3354 pfn = get_pte_pfn(ptent, args->vma, addr);
3355 if (pfn == -1)
3356 continue;
3357
3358 if (!pte_young(ptent)) {
3359 walk->mm_stats[MM_LEAF_OLD]++;
3360 continue;
3361 }
3362
3363 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3364 if (!folio)
3365 continue;
3366
3367 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3368 VM_WARN_ON_ONCE(true);
3369
3370 young++;
3371 walk->mm_stats[MM_LEAF_YOUNG]++;
3372
3373 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3374 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3375 !folio_test_swapcache(folio)))
3376 folio_mark_dirty(folio);
3377
3378 old_gen = folio_update_gen(folio, new_gen);
3379 if (old_gen >= 0 && old_gen != new_gen)
3380 update_batch_size(walk, folio, old_gen, new_gen);
3381 }
3382
3383 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3384 goto restart;
3385
3386 arch_leave_lazy_mmu_mode();
3387 pte_unmap_unlock(pte, ptl);
3388
3389 return suitable_to_scan(total, young);
3390}
3391
3392static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3393 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3394{
3395 int i;
3396 pmd_t *pmd;
3397 spinlock_t *ptl;
3398 struct lru_gen_mm_walk *walk = args->private;
3399 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3400 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3401 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3402
3403 VM_WARN_ON_ONCE(pud_leaf(*pud));
3404
3405 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3406 if (*first == -1) {
3407 *first = addr;
3408 bitmap_zero(bitmap, MIN_LRU_BATCH);
3409 return;
3410 }
3411
3412 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3413 if (i && i <= MIN_LRU_BATCH) {
3414 __set_bit(i - 1, bitmap);
3415 return;
3416 }
3417
3418 pmd = pmd_offset(pud, *first);
3419
3420 ptl = pmd_lockptr(args->mm, pmd);
3421 if (!spin_trylock(ptl))
3422 goto done;
3423
3424 arch_enter_lazy_mmu_mode();
3425
3426 do {
3427 unsigned long pfn;
3428 struct folio *folio;
3429
3430 /* don't round down the first address */
3431 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3432
3433 pfn = get_pmd_pfn(pmd[i], vma, addr);
3434 if (pfn == -1)
3435 goto next;
3436
3437 if (!pmd_trans_huge(pmd[i])) {
3438 if (should_clear_pmd_young())
3439 pmdp_test_and_clear_young(vma, addr, pmd + i);
3440 goto next;
3441 }
3442
3443 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3444 if (!folio)
3445 goto next;
3446
3447 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3448 goto next;
3449
3450 walk->mm_stats[MM_LEAF_YOUNG]++;
3451
3452 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3453 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3454 !folio_test_swapcache(folio)))
3455 folio_mark_dirty(folio);
3456
3457 old_gen = folio_update_gen(folio, new_gen);
3458 if (old_gen >= 0 && old_gen != new_gen)
3459 update_batch_size(walk, folio, old_gen, new_gen);
3460next:
3461 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3462 } while (i <= MIN_LRU_BATCH);
3463
3464 arch_leave_lazy_mmu_mode();
3465 spin_unlock(ptl);
3466done:
3467 *first = -1;
3468}
3469
3470static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3471 struct mm_walk *args)
3472{
3473 int i;
3474 pmd_t *pmd;
3475 unsigned long next;
3476 unsigned long addr;
3477 struct vm_area_struct *vma;
3478 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3479 unsigned long first = -1;
3480 struct lru_gen_mm_walk *walk = args->private;
3481 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3482
3483 VM_WARN_ON_ONCE(pud_leaf(*pud));
3484
3485 /*
3486 * Finish an entire PMD in two passes: the first only reaches to PTE
3487 * tables to avoid taking the PMD lock; the second, if necessary, takes
3488 * the PMD lock to clear the accessed bit in PMD entries.
3489 */
3490 pmd = pmd_offset(pud, start & PUD_MASK);
3491restart:
3492 /* walk_pte_range() may call get_next_vma() */
3493 vma = args->vma;
3494 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3495 pmd_t val = pmdp_get_lockless(pmd + i);
3496
3497 next = pmd_addr_end(addr, end);
3498
3499 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3500 walk->mm_stats[MM_LEAF_TOTAL]++;
3501 continue;
3502 }
3503
3504 if (pmd_trans_huge(val)) {
3505 unsigned long pfn = pmd_pfn(val);
3506 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3507
3508 walk->mm_stats[MM_LEAF_TOTAL]++;
3509
3510 if (!pmd_young(val)) {
3511 walk->mm_stats[MM_LEAF_OLD]++;
3512 continue;
3513 }
3514
3515 /* try to avoid unnecessary memory loads */
3516 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3517 continue;
3518
3519 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3520 continue;
3521 }
3522
3523 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3524
3525 if (should_clear_pmd_young()) {
3526 if (!pmd_young(val))
3527 continue;
3528
3529 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3530 }
3531
3532 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->max_seq, pmd + i))
3533 continue;
3534
3535 walk->mm_stats[MM_NONLEAF_FOUND]++;
3536
3537 if (!walk_pte_range(&val, addr, next, args))
3538 continue;
3539
3540 walk->mm_stats[MM_NONLEAF_ADDED]++;
3541
3542 /* carry over to the next generation */
3543 update_bloom_filter(mm_state, walk->max_seq + 1, pmd + i);
3544 }
3545
3546 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3547
3548 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3549 goto restart;
3550}
3551
3552static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3553 struct mm_walk *args)
3554{
3555 int i;
3556 pud_t *pud;
3557 unsigned long addr;
3558 unsigned long next;
3559 struct lru_gen_mm_walk *walk = args->private;
3560
3561 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3562
3563 pud = pud_offset(p4d, start & P4D_MASK);
3564restart:
3565 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3566 pud_t val = READ_ONCE(pud[i]);
3567
3568 next = pud_addr_end(addr, end);
3569
3570 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3571 continue;
3572
3573 walk_pmd_range(&val, addr, next, args);
3574
3575 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3576 end = (addr | ~PUD_MASK) + 1;
3577 goto done;
3578 }
3579 }
3580
3581 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3582 goto restart;
3583
3584 end = round_up(end, P4D_SIZE);
3585done:
3586 if (!end || !args->vma)
3587 return 1;
3588
3589 walk->next_addr = max(end, args->vma->vm_start);
3590
3591 return -EAGAIN;
3592}
3593
3594static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3595{
3596 static const struct mm_walk_ops mm_walk_ops = {
3597 .test_walk = should_skip_vma,
3598 .p4d_entry = walk_pud_range,
3599 .walk_lock = PGWALK_RDLOCK,
3600 };
3601
3602 int err;
3603 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3604
3605 walk->next_addr = FIRST_USER_ADDRESS;
3606
3607 do {
3608 DEFINE_MAX_SEQ(lruvec);
3609
3610 err = -EBUSY;
3611
3612 /* another thread might have called inc_max_seq() */
3613 if (walk->max_seq != max_seq)
3614 break;
3615
3616 /* folio_update_gen() requires stable folio_memcg() */
3617 if (!mem_cgroup_trylock_pages(memcg))
3618 break;
3619
3620 /* the caller might be holding the lock for write */
3621 if (mmap_read_trylock(mm)) {
3622 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3623
3624 mmap_read_unlock(mm);
3625 }
3626
3627 mem_cgroup_unlock_pages();
3628
3629 if (walk->batched) {
3630 spin_lock_irq(&lruvec->lru_lock);
3631 reset_batch_size(lruvec, walk);
3632 spin_unlock_irq(&lruvec->lru_lock);
3633 }
3634
3635 cond_resched();
3636 } while (err == -EAGAIN);
3637}
3638
3639static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3640{
3641 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3642
3643 if (pgdat && current_is_kswapd()) {
3644 VM_WARN_ON_ONCE(walk);
3645
3646 walk = &pgdat->mm_walk;
3647 } else if (!walk && force_alloc) {
3648 VM_WARN_ON_ONCE(current_is_kswapd());
3649
3650 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3651 }
3652
3653 current->reclaim_state->mm_walk = walk;
3654
3655 return walk;
3656}
3657
3658static void clear_mm_walk(void)
3659{
3660 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3661
3662 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3663 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3664
3665 current->reclaim_state->mm_walk = NULL;
3666
3667 if (!current_is_kswapd())
3668 kfree(walk);
3669}
3670
3671static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3672{
3673 int zone;
3674 int remaining = MAX_LRU_BATCH;
3675 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3676 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3677
3678 if (type == LRU_GEN_ANON && !can_swap)
3679 goto done;
3680
3681 /* prevent cold/hot inversion if force_scan is true */
3682 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3683 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3684
3685 while (!list_empty(head)) {
3686 struct folio *folio = lru_to_folio(head);
3687
3688 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3689 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3690 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3691 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3692
3693 new_gen = folio_inc_gen(lruvec, folio, false);
3694 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3695
3696 if (!--remaining)
3697 return false;
3698 }
3699 }
3700done:
3701 reset_ctrl_pos(lruvec, type, true);
3702 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3703
3704 return true;
3705}
3706
3707static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3708{
3709 int gen, type, zone;
3710 bool success = false;
3711 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3712 DEFINE_MIN_SEQ(lruvec);
3713
3714 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3715
3716 /* find the oldest populated generation */
3717 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3718 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3719 gen = lru_gen_from_seq(min_seq[type]);
3720
3721 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3722 if (!list_empty(&lrugen->folios[gen][type][zone]))
3723 goto next;
3724 }
3725
3726 min_seq[type]++;
3727 }
3728next:
3729 ;
3730 }
3731
3732 /* see the comment on lru_gen_folio */
3733 if (can_swap) {
3734 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3735 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3736 }
3737
3738 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3739 if (min_seq[type] == lrugen->min_seq[type])
3740 continue;
3741
3742 reset_ctrl_pos(lruvec, type, true);
3743 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3744 success = true;
3745 }
3746
3747 return success;
3748}
3749
3750static bool inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
3751 bool can_swap, bool force_scan)
3752{
3753 bool success;
3754 int prev, next;
3755 int type, zone;
3756 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3757restart:
3758 if (max_seq < READ_ONCE(lrugen->max_seq))
3759 return false;
3760
3761 spin_lock_irq(&lruvec->lru_lock);
3762
3763 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3764
3765 success = max_seq == lrugen->max_seq;
3766 if (!success)
3767 goto unlock;
3768
3769 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3770 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3771 continue;
3772
3773 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3774
3775 if (inc_min_seq(lruvec, type, can_swap))
3776 continue;
3777
3778 spin_unlock_irq(&lruvec->lru_lock);
3779 cond_resched();
3780 goto restart;
3781 }
3782
3783 /*
3784 * Update the active/inactive LRU sizes for compatibility. Both sides of
3785 * the current max_seq need to be covered, since max_seq+1 can overlap
3786 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3787 * overlap, cold/hot inversion happens.
3788 */
3789 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3790 next = lru_gen_from_seq(lrugen->max_seq + 1);
3791
3792 for (type = 0; type < ANON_AND_FILE; type++) {
3793 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3794 enum lru_list lru = type * LRU_INACTIVE_FILE;
3795 long delta = lrugen->nr_pages[prev][type][zone] -
3796 lrugen->nr_pages[next][type][zone];
3797
3798 if (!delta)
3799 continue;
3800
3801 __update_lru_size(lruvec, lru, zone, delta);
3802 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3803 }
3804 }
3805
3806 for (type = 0; type < ANON_AND_FILE; type++)
3807 reset_ctrl_pos(lruvec, type, false);
3808
3809 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3810 /* make sure preceding modifications appear */
3811 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3812unlock:
3813 spin_unlock_irq(&lruvec->lru_lock);
3814
3815 return success;
3816}
3817
3818static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
3819 struct scan_control *sc, bool can_swap, bool force_scan)
3820{
3821 bool success;
3822 struct lru_gen_mm_walk *walk;
3823 struct mm_struct *mm = NULL;
3824 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3825 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3826
3827 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
3828
3829 if (!mm_state)
3830 return inc_max_seq(lruvec, max_seq, can_swap, force_scan);
3831
3832 /* see the comment in iterate_mm_list() */
3833 if (max_seq <= READ_ONCE(mm_state->seq))
3834 return false;
3835
3836 /*
3837 * If the hardware doesn't automatically set the accessed bit, fallback
3838 * to lru_gen_look_around(), which only clears the accessed bit in a
3839 * handful of PTEs. Spreading the work out over a period of time usually
3840 * is less efficient, but it avoids bursty page faults.
3841 */
3842 if (!should_walk_mmu()) {
3843 success = iterate_mm_list_nowalk(lruvec, max_seq);
3844 goto done;
3845 }
3846
3847 walk = set_mm_walk(NULL, true);
3848 if (!walk) {
3849 success = iterate_mm_list_nowalk(lruvec, max_seq);
3850 goto done;
3851 }
3852
3853 walk->lruvec = lruvec;
3854 walk->max_seq = max_seq;
3855 walk->can_swap = can_swap;
3856 walk->force_scan = force_scan;
3857
3858 do {
3859 success = iterate_mm_list(lruvec, walk, &mm);
3860 if (mm)
3861 walk_mm(lruvec, mm, walk);
3862 } while (mm);
3863done:
3864 if (success) {
3865 success = inc_max_seq(lruvec, max_seq, can_swap, force_scan);
3866 WARN_ON_ONCE(!success);
3867 }
3868
3869 return success;
3870}
3871
3872/******************************************************************************
3873 * working set protection
3874 ******************************************************************************/
3875
3876static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3877{
3878 int gen, type, zone;
3879 unsigned long total = 0;
3880 bool can_swap = get_swappiness(lruvec, sc);
3881 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3882 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3883 DEFINE_MAX_SEQ(lruvec);
3884 DEFINE_MIN_SEQ(lruvec);
3885
3886 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3887 unsigned long seq;
3888
3889 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3890 gen = lru_gen_from_seq(seq);
3891
3892 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3893 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3894 }
3895 }
3896
3897 /* whether the size is big enough to be helpful */
3898 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3899}
3900
3901static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3902 unsigned long min_ttl)
3903{
3904 int gen;
3905 unsigned long birth;
3906 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3907 DEFINE_MIN_SEQ(lruvec);
3908
3909 /* see the comment on lru_gen_folio */
3910 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3911 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3912
3913 if (time_is_after_jiffies(birth + min_ttl))
3914 return false;
3915
3916 if (!lruvec_is_sizable(lruvec, sc))
3917 return false;
3918
3919 mem_cgroup_calculate_protection(NULL, memcg);
3920
3921 return !mem_cgroup_below_min(NULL, memcg);
3922}
3923
3924/* to protect the working set of the last N jiffies */
3925static unsigned long lru_gen_min_ttl __read_mostly;
3926
3927static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3928{
3929 struct mem_cgroup *memcg;
3930 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3931
3932 VM_WARN_ON_ONCE(!current_is_kswapd());
3933
3934 /* check the order to exclude compaction-induced reclaim */
3935 if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
3936 return;
3937
3938 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3939 do {
3940 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3941
3942 if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
3943 mem_cgroup_iter_break(NULL, memcg);
3944 return;
3945 }
3946
3947 cond_resched();
3948 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
3949
3950 /*
3951 * The main goal is to OOM kill if every generation from all memcgs is
3952 * younger than min_ttl. However, another possibility is all memcgs are
3953 * either too small or below min.
3954 */
3955 if (mutex_trylock(&oom_lock)) {
3956 struct oom_control oc = {
3957 .gfp_mask = sc->gfp_mask,
3958 };
3959
3960 out_of_memory(&oc);
3961
3962 mutex_unlock(&oom_lock);
3963 }
3964}
3965
3966/******************************************************************************
3967 * rmap/PT walk feedback
3968 ******************************************************************************/
3969
3970/*
3971 * This function exploits spatial locality when shrink_folio_list() walks the
3972 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3973 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3974 * the PTE table to the Bloom filter. This forms a feedback loop between the
3975 * eviction and the aging.
3976 */
3977void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
3978{
3979 int i;
3980 unsigned long start;
3981 unsigned long end;
3982 struct lru_gen_mm_walk *walk;
3983 int young = 0;
3984 pte_t *pte = pvmw->pte;
3985 unsigned long addr = pvmw->address;
3986 struct vm_area_struct *vma = pvmw->vma;
3987 struct folio *folio = pfn_folio(pvmw->pfn);
3988 bool can_swap = !folio_is_file_lru(folio);
3989 struct mem_cgroup *memcg = folio_memcg(folio);
3990 struct pglist_data *pgdat = folio_pgdat(folio);
3991 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3992 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3993 DEFINE_MAX_SEQ(lruvec);
3994 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3995
3996 lockdep_assert_held(pvmw->ptl);
3997 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
3998
3999 if (spin_is_contended(pvmw->ptl))
4000 return;
4001
4002 /* exclude special VMAs containing anon pages from COW */
4003 if (vma->vm_flags & VM_SPECIAL)
4004 return;
4005
4006 /* avoid taking the LRU lock under the PTL when possible */
4007 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4008
4009 start = max(addr & PMD_MASK, vma->vm_start);
4010 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4011
4012 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4013 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4014 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4015 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4016 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4017 else {
4018 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4019 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4020 }
4021 }
4022
4023 /* folio_update_gen() requires stable folio_memcg() */
4024 if (!mem_cgroup_trylock_pages(memcg))
4025 return;
4026
4027 arch_enter_lazy_mmu_mode();
4028
4029 pte -= (addr - start) / PAGE_SIZE;
4030
4031 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4032 unsigned long pfn;
4033 pte_t ptent = ptep_get(pte + i);
4034
4035 pfn = get_pte_pfn(ptent, vma, addr);
4036 if (pfn == -1)
4037 continue;
4038
4039 if (!pte_young(ptent))
4040 continue;
4041
4042 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4043 if (!folio)
4044 continue;
4045
4046 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4047 VM_WARN_ON_ONCE(true);
4048
4049 young++;
4050
4051 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4052 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4053 !folio_test_swapcache(folio)))
4054 folio_mark_dirty(folio);
4055
4056 if (walk) {
4057 old_gen = folio_update_gen(folio, new_gen);
4058 if (old_gen >= 0 && old_gen != new_gen)
4059 update_batch_size(walk, folio, old_gen, new_gen);
4060
4061 continue;
4062 }
4063
4064 old_gen = folio_lru_gen(folio);
4065 if (old_gen < 0)
4066 folio_set_referenced(folio);
4067 else if (old_gen != new_gen)
4068 folio_activate(folio);
4069 }
4070
4071 arch_leave_lazy_mmu_mode();
4072 mem_cgroup_unlock_pages();
4073
4074 /* feedback from rmap walkers to page table walkers */
4075 if (mm_state && suitable_to_scan(i, young))
4076 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4077}
4078
4079/******************************************************************************
4080 * memcg LRU
4081 ******************************************************************************/
4082
4083/* see the comment on MEMCG_NR_GENS */
4084enum {
4085 MEMCG_LRU_NOP,
4086 MEMCG_LRU_HEAD,
4087 MEMCG_LRU_TAIL,
4088 MEMCG_LRU_OLD,
4089 MEMCG_LRU_YOUNG,
4090};
4091
4092static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4093{
4094 int seg;
4095 int old, new;
4096 unsigned long flags;
4097 int bin = get_random_u32_below(MEMCG_NR_BINS);
4098 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4099
4100 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4101
4102 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4103
4104 seg = 0;
4105 new = old = lruvec->lrugen.gen;
4106
4107 /* see the comment on MEMCG_NR_GENS */
4108 if (op == MEMCG_LRU_HEAD)
4109 seg = MEMCG_LRU_HEAD;
4110 else if (op == MEMCG_LRU_TAIL)
4111 seg = MEMCG_LRU_TAIL;
4112 else if (op == MEMCG_LRU_OLD)
4113 new = get_memcg_gen(pgdat->memcg_lru.seq);
4114 else if (op == MEMCG_LRU_YOUNG)
4115 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4116 else
4117 VM_WARN_ON_ONCE(true);
4118
4119 WRITE_ONCE(lruvec->lrugen.seg, seg);
4120 WRITE_ONCE(lruvec->lrugen.gen, new);
4121
4122 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4123
4124 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4125 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4126 else
4127 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4128
4129 pgdat->memcg_lru.nr_memcgs[old]--;
4130 pgdat->memcg_lru.nr_memcgs[new]++;
4131
4132 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4133 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4134
4135 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4136}
4137
4138#ifdef CONFIG_MEMCG
4139
4140void lru_gen_online_memcg(struct mem_cgroup *memcg)
4141{
4142 int gen;
4143 int nid;
4144 int bin = get_random_u32_below(MEMCG_NR_BINS);
4145
4146 for_each_node(nid) {
4147 struct pglist_data *pgdat = NODE_DATA(nid);
4148 struct lruvec *lruvec = get_lruvec(memcg, nid);
4149
4150 spin_lock_irq(&pgdat->memcg_lru.lock);
4151
4152 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4153
4154 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4155
4156 lruvec->lrugen.gen = gen;
4157
4158 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4159 pgdat->memcg_lru.nr_memcgs[gen]++;
4160
4161 spin_unlock_irq(&pgdat->memcg_lru.lock);
4162 }
4163}
4164
4165void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4166{
4167 int nid;
4168
4169 for_each_node(nid) {
4170 struct lruvec *lruvec = get_lruvec(memcg, nid);
4171
4172 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4173 }
4174}
4175
4176void lru_gen_release_memcg(struct mem_cgroup *memcg)
4177{
4178 int gen;
4179 int nid;
4180
4181 for_each_node(nid) {
4182 struct pglist_data *pgdat = NODE_DATA(nid);
4183 struct lruvec *lruvec = get_lruvec(memcg, nid);
4184
4185 spin_lock_irq(&pgdat->memcg_lru.lock);
4186
4187 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4188 goto unlock;
4189
4190 gen = lruvec->lrugen.gen;
4191
4192 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4193 pgdat->memcg_lru.nr_memcgs[gen]--;
4194
4195 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4196 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4197unlock:
4198 spin_unlock_irq(&pgdat->memcg_lru.lock);
4199 }
4200}
4201
4202void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4203{
4204 struct lruvec *lruvec = get_lruvec(memcg, nid);
4205
4206 /* see the comment on MEMCG_NR_GENS */
4207 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4208 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4209}
4210
4211#endif /* CONFIG_MEMCG */
4212
4213/******************************************************************************
4214 * the eviction
4215 ******************************************************************************/
4216
4217static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4218 int tier_idx)
4219{
4220 bool success;
4221 int gen = folio_lru_gen(folio);
4222 int type = folio_is_file_lru(folio);
4223 int zone = folio_zonenum(folio);
4224 int delta = folio_nr_pages(folio);
4225 int refs = folio_lru_refs(folio);
4226 int tier = lru_tier_from_refs(refs);
4227 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4228
4229 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4230
4231 /* unevictable */
4232 if (!folio_evictable(folio)) {
4233 success = lru_gen_del_folio(lruvec, folio, true);
4234 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4235 folio_set_unevictable(folio);
4236 lruvec_add_folio(lruvec, folio);
4237 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4238 return true;
4239 }
4240
4241 /* dirty lazyfree */
4242 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4243 success = lru_gen_del_folio(lruvec, folio, true);
4244 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4245 folio_set_swapbacked(folio);
4246 lruvec_add_folio_tail(lruvec, folio);
4247 return true;
4248 }
4249
4250 /* promoted */
4251 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4252 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4253 return true;
4254 }
4255
4256 /* protected */
4257 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4258 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4259
4260 gen = folio_inc_gen(lruvec, folio, false);
4261 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4262
4263 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4264 lrugen->protected[hist][type][tier - 1] + delta);
4265 return true;
4266 }
4267
4268 /* ineligible */
4269 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4270 gen = folio_inc_gen(lruvec, folio, false);
4271 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4272 return true;
4273 }
4274
4275 /* waiting for writeback */
4276 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4277 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4278 gen = folio_inc_gen(lruvec, folio, true);
4279 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4280 return true;
4281 }
4282
4283 return false;
4284}
4285
4286static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4287{
4288 bool success;
4289
4290 /* swapping inhibited */
4291 if (!(sc->gfp_mask & __GFP_IO) &&
4292 (folio_test_dirty(folio) ||
4293 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4294 return false;
4295
4296 /* raced with release_pages() */
4297 if (!folio_try_get(folio))
4298 return false;
4299
4300 /* raced with another isolation */
4301 if (!folio_test_clear_lru(folio)) {
4302 folio_put(folio);
4303 return false;
4304 }
4305
4306 /* see the comment on MAX_NR_TIERS */
4307 if (!folio_test_referenced(folio))
4308 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4309
4310 /* for shrink_folio_list() */
4311 folio_clear_reclaim(folio);
4312 folio_clear_referenced(folio);
4313
4314 success = lru_gen_del_folio(lruvec, folio, true);
4315 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4316
4317 return true;
4318}
4319
4320static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4321 int type, int tier, struct list_head *list)
4322{
4323 int i;
4324 int gen;
4325 enum vm_event_item item;
4326 int sorted = 0;
4327 int scanned = 0;
4328 int isolated = 0;
4329 int skipped = 0;
4330 int remaining = MAX_LRU_BATCH;
4331 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4332 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4333
4334 VM_WARN_ON_ONCE(!list_empty(list));
4335
4336 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4337 return 0;
4338
4339 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4340
4341 for (i = MAX_NR_ZONES; i > 0; i--) {
4342 LIST_HEAD(moved);
4343 int skipped_zone = 0;
4344 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4345 struct list_head *head = &lrugen->folios[gen][type][zone];
4346
4347 while (!list_empty(head)) {
4348 struct folio *folio = lru_to_folio(head);
4349 int delta = folio_nr_pages(folio);
4350
4351 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4352 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4353 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4354 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4355
4356 scanned += delta;
4357
4358 if (sort_folio(lruvec, folio, sc, tier))
4359 sorted += delta;
4360 else if (isolate_folio(lruvec, folio, sc)) {
4361 list_add(&folio->lru, list);
4362 isolated += delta;
4363 } else {
4364 list_move(&folio->lru, &moved);
4365 skipped_zone += delta;
4366 }
4367
4368 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4369 break;
4370 }
4371
4372 if (skipped_zone) {
4373 list_splice(&moved, head);
4374 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4375 skipped += skipped_zone;
4376 }
4377
4378 if (!remaining || isolated >= MIN_LRU_BATCH)
4379 break;
4380 }
4381
4382 item = PGSCAN_KSWAPD + reclaimer_offset();
4383 if (!cgroup_reclaim(sc)) {
4384 __count_vm_events(item, isolated);
4385 __count_vm_events(PGREFILL, sorted);
4386 }
4387 __count_memcg_events(memcg, item, isolated);
4388 __count_memcg_events(memcg, PGREFILL, sorted);
4389 __count_vm_events(PGSCAN_ANON + type, isolated);
4390 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4391 scanned, skipped, isolated,
4392 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4393
4394 /*
4395 * There might not be eligible folios due to reclaim_idx. Check the
4396 * remaining to prevent livelock if it's not making progress.
4397 */
4398 return isolated || !remaining ? scanned : 0;
4399}
4400
4401static int get_tier_idx(struct lruvec *lruvec, int type)
4402{
4403 int tier;
4404 struct ctrl_pos sp, pv;
4405
4406 /*
4407 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4408 * This value is chosen because any other tier would have at least twice
4409 * as many refaults as the first tier.
4410 */
4411 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4412 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4413 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4414 if (!positive_ctrl_err(&sp, &pv))
4415 break;
4416 }
4417
4418 return tier - 1;
4419}
4420
4421static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4422{
4423 int type, tier;
4424 struct ctrl_pos sp, pv;
4425 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4426
4427 /*
4428 * Compare the first tier of anon with that of file to determine which
4429 * type to scan. Also need to compare other tiers of the selected type
4430 * with the first tier of the other type to determine the last tier (of
4431 * the selected type) to evict.
4432 */
4433 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4434 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4435 type = positive_ctrl_err(&sp, &pv);
4436
4437 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4438 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4439 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4440 if (!positive_ctrl_err(&sp, &pv))
4441 break;
4442 }
4443
4444 *tier_idx = tier - 1;
4445
4446 return type;
4447}
4448
4449static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4450 int *type_scanned, struct list_head *list)
4451{
4452 int i;
4453 int type;
4454 int scanned;
4455 int tier = -1;
4456 DEFINE_MIN_SEQ(lruvec);
4457
4458 /*
4459 * Try to make the obvious choice first. When anon and file are both
4460 * available from the same generation, interpret swappiness 1 as file
4461 * first and 200 as anon first.
4462 */
4463 if (!swappiness)
4464 type = LRU_GEN_FILE;
4465 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4466 type = LRU_GEN_ANON;
4467 else if (swappiness == 1)
4468 type = LRU_GEN_FILE;
4469 else if (swappiness == 200)
4470 type = LRU_GEN_ANON;
4471 else
4472 type = get_type_to_scan(lruvec, swappiness, &tier);
4473
4474 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4475 if (tier < 0)
4476 tier = get_tier_idx(lruvec, type);
4477
4478 scanned = scan_folios(lruvec, sc, type, tier, list);
4479 if (scanned)
4480 break;
4481
4482 type = !type;
4483 tier = -1;
4484 }
4485
4486 *type_scanned = type;
4487
4488 return scanned;
4489}
4490
4491static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4492{
4493 int type;
4494 int scanned;
4495 int reclaimed;
4496 LIST_HEAD(list);
4497 LIST_HEAD(clean);
4498 struct folio *folio;
4499 struct folio *next;
4500 enum vm_event_item item;
4501 struct reclaim_stat stat;
4502 struct lru_gen_mm_walk *walk;
4503 bool skip_retry = false;
4504 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4505 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4506
4507 spin_lock_irq(&lruvec->lru_lock);
4508
4509 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4510
4511 scanned += try_to_inc_min_seq(lruvec, swappiness);
4512
4513 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4514 scanned = 0;
4515
4516 spin_unlock_irq(&lruvec->lru_lock);
4517
4518 if (list_empty(&list))
4519 return scanned;
4520retry:
4521 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4522 sc->nr_reclaimed += reclaimed;
4523 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4524 scanned, reclaimed, &stat, sc->priority,
4525 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4526
4527 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4528 if (!folio_evictable(folio)) {
4529 list_del(&folio->lru);
4530 folio_putback_lru(folio);
4531 continue;
4532 }
4533
4534 if (folio_test_reclaim(folio) &&
4535 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4536 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4537 if (folio_test_workingset(folio))
4538 folio_set_referenced(folio);
4539 continue;
4540 }
4541
4542 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4543 folio_mapped(folio) || folio_test_locked(folio) ||
4544 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4545 /* don't add rejected folios to the oldest generation */
4546 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4547 BIT(PG_active));
4548 continue;
4549 }
4550
4551 /* retry folios that may have missed folio_rotate_reclaimable() */
4552 list_move(&folio->lru, &clean);
4553 sc->nr_scanned -= folio_nr_pages(folio);
4554 }
4555
4556 spin_lock_irq(&lruvec->lru_lock);
4557
4558 move_folios_to_lru(lruvec, &list);
4559
4560 walk = current->reclaim_state->mm_walk;
4561 if (walk && walk->batched)
4562 reset_batch_size(lruvec, walk);
4563
4564 item = PGSTEAL_KSWAPD + reclaimer_offset();
4565 if (!cgroup_reclaim(sc))
4566 __count_vm_events(item, reclaimed);
4567 __count_memcg_events(memcg, item, reclaimed);
4568 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4569
4570 spin_unlock_irq(&lruvec->lru_lock);
4571
4572 mem_cgroup_uncharge_list(&list);
4573 free_unref_page_list(&list);
4574
4575 INIT_LIST_HEAD(&list);
4576 list_splice_init(&clean, &list);
4577
4578 if (!list_empty(&list)) {
4579 skip_retry = true;
4580 goto retry;
4581 }
4582
4583 return scanned;
4584}
4585
4586static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4587 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4588{
4589 int gen, type, zone;
4590 unsigned long old = 0;
4591 unsigned long young = 0;
4592 unsigned long total = 0;
4593 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4594 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4595 DEFINE_MIN_SEQ(lruvec);
4596
4597 /* whether this lruvec is completely out of cold folios */
4598 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4599 *nr_to_scan = 0;
4600 return true;
4601 }
4602
4603 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4604 unsigned long seq;
4605
4606 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4607 unsigned long size = 0;
4608
4609 gen = lru_gen_from_seq(seq);
4610
4611 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4612 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4613
4614 total += size;
4615 if (seq == max_seq)
4616 young += size;
4617 else if (seq + MIN_NR_GENS == max_seq)
4618 old += size;
4619 }
4620 }
4621
4622 /* try to scrape all its memory if this memcg was deleted */
4623 if (!mem_cgroup_online(memcg)) {
4624 *nr_to_scan = total;
4625 return false;
4626 }
4627
4628 *nr_to_scan = total >> sc->priority;
4629
4630 /*
4631 * The aging tries to be lazy to reduce the overhead, while the eviction
4632 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4633 * ideal number of generations is MIN_NR_GENS+1.
4634 */
4635 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4636 return false;
4637
4638 /*
4639 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4640 * of the total number of pages for each generation. A reasonable range
4641 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4642 * aging cares about the upper bound of hot pages, while the eviction
4643 * cares about the lower bound of cold pages.
4644 */
4645 if (young * MIN_NR_GENS > total)
4646 return true;
4647 if (old * (MIN_NR_GENS + 2) < total)
4648 return true;
4649
4650 return false;
4651}
4652
4653/*
4654 * For future optimizations:
4655 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4656 * reclaim.
4657 */
4658static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4659{
4660 unsigned long nr_to_scan;
4661 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4662 DEFINE_MAX_SEQ(lruvec);
4663
4664 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4665 return -1;
4666
4667 if (!should_run_aging(lruvec, max_seq, sc, can_swap, &nr_to_scan))
4668 return nr_to_scan;
4669
4670 /* skip the aging path at the default priority */
4671 if (sc->priority == DEF_PRIORITY)
4672 return nr_to_scan;
4673
4674 /* skip this lruvec as it's low on cold folios */
4675 return try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false) ? -1 : 0;
4676}
4677
4678static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4679{
4680 int i;
4681 enum zone_watermarks mark;
4682
4683 /* don't abort memcg reclaim to ensure fairness */
4684 if (!root_reclaim(sc))
4685 return false;
4686
4687 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4688 return true;
4689
4690 /* check the order to exclude compaction-induced reclaim */
4691 if (!current_is_kswapd() || sc->order)
4692 return false;
4693
4694 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4695 WMARK_PROMO : WMARK_HIGH;
4696
4697 for (i = 0; i <= sc->reclaim_idx; i++) {
4698 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4699 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4700
4701 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4702 return false;
4703 }
4704
4705 /* kswapd should abort if all eligible zones are safe */
4706 return true;
4707}
4708
4709static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4710{
4711 long nr_to_scan;
4712 unsigned long scanned = 0;
4713 int swappiness = get_swappiness(lruvec, sc);
4714
4715 /* clean file folios are more likely to exist */
4716 if (swappiness && !(sc->gfp_mask & __GFP_IO))
4717 swappiness = 1;
4718
4719 while (true) {
4720 int delta;
4721
4722 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4723 if (nr_to_scan <= 0)
4724 break;
4725
4726 delta = evict_folios(lruvec, sc, swappiness);
4727 if (!delta)
4728 break;
4729
4730 scanned += delta;
4731 if (scanned >= nr_to_scan)
4732 break;
4733
4734 if (should_abort_scan(lruvec, sc))
4735 break;
4736
4737 cond_resched();
4738 }
4739
4740 /* whether this lruvec should be rotated */
4741 return nr_to_scan < 0;
4742}
4743
4744static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4745{
4746 bool success;
4747 unsigned long scanned = sc->nr_scanned;
4748 unsigned long reclaimed = sc->nr_reclaimed;
4749 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4750 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4751
4752 mem_cgroup_calculate_protection(NULL, memcg);
4753
4754 if (mem_cgroup_below_min(NULL, memcg))
4755 return MEMCG_LRU_YOUNG;
4756
4757 if (mem_cgroup_below_low(NULL, memcg)) {
4758 /* see the comment on MEMCG_NR_GENS */
4759 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4760 return MEMCG_LRU_TAIL;
4761
4762 memcg_memory_event(memcg, MEMCG_LOW);
4763 }
4764
4765 success = try_to_shrink_lruvec(lruvec, sc);
4766
4767 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4768
4769 if (!sc->proactive)
4770 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4771 sc->nr_reclaimed - reclaimed);
4772
4773 flush_reclaim_state(sc);
4774
4775 if (success && mem_cgroup_online(memcg))
4776 return MEMCG_LRU_YOUNG;
4777
4778 if (!success && lruvec_is_sizable(lruvec, sc))
4779 return 0;
4780
4781 /* one retry if offlined or too small */
4782 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4783 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4784}
4785
4786static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4787{
4788 int op;
4789 int gen;
4790 int bin;
4791 int first_bin;
4792 struct lruvec *lruvec;
4793 struct lru_gen_folio *lrugen;
4794 struct mem_cgroup *memcg;
4795 struct hlist_nulls_node *pos;
4796
4797 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4798 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4799restart:
4800 op = 0;
4801 memcg = NULL;
4802
4803 rcu_read_lock();
4804
4805 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4806 if (op) {
4807 lru_gen_rotate_memcg(lruvec, op);
4808 op = 0;
4809 }
4810
4811 mem_cgroup_put(memcg);
4812 memcg = NULL;
4813
4814 if (gen != READ_ONCE(lrugen->gen))
4815 continue;
4816
4817 lruvec = container_of(lrugen, struct lruvec, lrugen);
4818 memcg = lruvec_memcg(lruvec);
4819
4820 if (!mem_cgroup_tryget(memcg)) {
4821 lru_gen_release_memcg(memcg);
4822 memcg = NULL;
4823 continue;
4824 }
4825
4826 rcu_read_unlock();
4827
4828 op = shrink_one(lruvec, sc);
4829
4830 rcu_read_lock();
4831
4832 if (should_abort_scan(lruvec, sc))
4833 break;
4834 }
4835
4836 rcu_read_unlock();
4837
4838 if (op)
4839 lru_gen_rotate_memcg(lruvec, op);
4840
4841 mem_cgroup_put(memcg);
4842
4843 if (!is_a_nulls(pos))
4844 return;
4845
4846 /* restart if raced with lru_gen_rotate_memcg() */
4847 if (gen != get_nulls_value(pos))
4848 goto restart;
4849
4850 /* try the rest of the bins of the current generation */
4851 bin = get_memcg_bin(bin + 1);
4852 if (bin != first_bin)
4853 goto restart;
4854}
4855
4856static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4857{
4858 struct blk_plug plug;
4859
4860 VM_WARN_ON_ONCE(root_reclaim(sc));
4861 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4862
4863 lru_add_drain();
4864
4865 blk_start_plug(&plug);
4866
4867 set_mm_walk(NULL, sc->proactive);
4868
4869 if (try_to_shrink_lruvec(lruvec, sc))
4870 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4871
4872 clear_mm_walk();
4873
4874 blk_finish_plug(&plug);
4875}
4876
4877static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4878{
4879 int priority;
4880 unsigned long reclaimable;
4881 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
4882
4883 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4884 return;
4885 /*
4886 * Determine the initial priority based on
4887 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4888 * where reclaimed_to_scanned_ratio = inactive / total.
4889 */
4890 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4891 if (get_swappiness(lruvec, sc))
4892 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4893
4894 /* round down reclaimable and round up sc->nr_to_reclaim */
4895 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4896
4897 sc->priority = clamp(priority, 0, DEF_PRIORITY);
4898}
4899
4900static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4901{
4902 struct blk_plug plug;
4903 unsigned long reclaimed = sc->nr_reclaimed;
4904
4905 VM_WARN_ON_ONCE(!root_reclaim(sc));
4906
4907 /*
4908 * Unmapped clean folios are already prioritized. Scanning for more of
4909 * them is likely futile and can cause high reclaim latency when there
4910 * is a large number of memcgs.
4911 */
4912 if (!sc->may_writepage || !sc->may_unmap)
4913 goto done;
4914
4915 lru_add_drain();
4916
4917 blk_start_plug(&plug);
4918
4919 set_mm_walk(pgdat, sc->proactive);
4920
4921 set_initial_priority(pgdat, sc);
4922
4923 if (current_is_kswapd())
4924 sc->nr_reclaimed = 0;
4925
4926 if (mem_cgroup_disabled())
4927 shrink_one(&pgdat->__lruvec, sc);
4928 else
4929 shrink_many(pgdat, sc);
4930
4931 if (current_is_kswapd())
4932 sc->nr_reclaimed += reclaimed;
4933
4934 clear_mm_walk();
4935
4936 blk_finish_plug(&plug);
4937done:
4938 /* kswapd should never fail */
4939 pgdat->kswapd_failures = 0;
4940}
4941
4942/******************************************************************************
4943 * state change
4944 ******************************************************************************/
4945
4946static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4947{
4948 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4949
4950 if (lrugen->enabled) {
4951 enum lru_list lru;
4952
4953 for_each_evictable_lru(lru) {
4954 if (!list_empty(&lruvec->lists[lru]))
4955 return false;
4956 }
4957 } else {
4958 int gen, type, zone;
4959
4960 for_each_gen_type_zone(gen, type, zone) {
4961 if (!list_empty(&lrugen->folios[gen][type][zone]))
4962 return false;
4963 }
4964 }
4965
4966 return true;
4967}
4968
4969static bool fill_evictable(struct lruvec *lruvec)
4970{
4971 enum lru_list lru;
4972 int remaining = MAX_LRU_BATCH;
4973
4974 for_each_evictable_lru(lru) {
4975 int type = is_file_lru(lru);
4976 bool active = is_active_lru(lru);
4977 struct list_head *head = &lruvec->lists[lru];
4978
4979 while (!list_empty(head)) {
4980 bool success;
4981 struct folio *folio = lru_to_folio(head);
4982
4983 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4984 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
4985 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4986 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
4987
4988 lruvec_del_folio(lruvec, folio);
4989 success = lru_gen_add_folio(lruvec, folio, false);
4990 VM_WARN_ON_ONCE(!success);
4991
4992 if (!--remaining)
4993 return false;
4994 }
4995 }
4996
4997 return true;
4998}
4999
5000static bool drain_evictable(struct lruvec *lruvec)
5001{
5002 int gen, type, zone;
5003 int remaining = MAX_LRU_BATCH;
5004
5005 for_each_gen_type_zone(gen, type, zone) {
5006 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5007
5008 while (!list_empty(head)) {
5009 bool success;
5010 struct folio *folio = lru_to_folio(head);
5011
5012 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5013 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5014 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5015 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5016
5017 success = lru_gen_del_folio(lruvec, folio, false);
5018 VM_WARN_ON_ONCE(!success);
5019 lruvec_add_folio(lruvec, folio);
5020
5021 if (!--remaining)
5022 return false;
5023 }
5024 }
5025
5026 return true;
5027}
5028
5029static void lru_gen_change_state(bool enabled)
5030{
5031 static DEFINE_MUTEX(state_mutex);
5032
5033 struct mem_cgroup *memcg;
5034
5035 cgroup_lock();
5036 cpus_read_lock();
5037 get_online_mems();
5038 mutex_lock(&state_mutex);
5039
5040 if (enabled == lru_gen_enabled())
5041 goto unlock;
5042
5043 if (enabled)
5044 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5045 else
5046 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5047
5048 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5049 do {
5050 int nid;
5051
5052 for_each_node(nid) {
5053 struct lruvec *lruvec = get_lruvec(memcg, nid);
5054
5055 spin_lock_irq(&lruvec->lru_lock);
5056
5057 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5058 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5059
5060 lruvec->lrugen.enabled = enabled;
5061
5062 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5063 spin_unlock_irq(&lruvec->lru_lock);
5064 cond_resched();
5065 spin_lock_irq(&lruvec->lru_lock);
5066 }
5067
5068 spin_unlock_irq(&lruvec->lru_lock);
5069 }
5070
5071 cond_resched();
5072 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5073unlock:
5074 mutex_unlock(&state_mutex);
5075 put_online_mems();
5076 cpus_read_unlock();
5077 cgroup_unlock();
5078}
5079
5080/******************************************************************************
5081 * sysfs interface
5082 ******************************************************************************/
5083
5084static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5085{
5086 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5087}
5088
5089/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5090static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5091 const char *buf, size_t len)
5092{
5093 unsigned int msecs;
5094
5095 if (kstrtouint(buf, 0, &msecs))
5096 return -EINVAL;
5097
5098 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5099
5100 return len;
5101}
5102
5103static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5104
5105static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5106{
5107 unsigned int caps = 0;
5108
5109 if (get_cap(LRU_GEN_CORE))
5110 caps |= BIT(LRU_GEN_CORE);
5111
5112 if (should_walk_mmu())
5113 caps |= BIT(LRU_GEN_MM_WALK);
5114
5115 if (should_clear_pmd_young())
5116 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5117
5118 return sysfs_emit(buf, "0x%04x\n", caps);
5119}
5120
5121/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5122static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5123 const char *buf, size_t len)
5124{
5125 int i;
5126 unsigned int caps;
5127
5128 if (tolower(*buf) == 'n')
5129 caps = 0;
5130 else if (tolower(*buf) == 'y')
5131 caps = -1;
5132 else if (kstrtouint(buf, 0, &caps))
5133 return -EINVAL;
5134
5135 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5136 bool enabled = caps & BIT(i);
5137
5138 if (i == LRU_GEN_CORE)
5139 lru_gen_change_state(enabled);
5140 else if (enabled)
5141 static_branch_enable(&lru_gen_caps[i]);
5142 else
5143 static_branch_disable(&lru_gen_caps[i]);
5144 }
5145
5146 return len;
5147}
5148
5149static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5150
5151static struct attribute *lru_gen_attrs[] = {
5152 &lru_gen_min_ttl_attr.attr,
5153 &lru_gen_enabled_attr.attr,
5154 NULL
5155};
5156
5157static const struct attribute_group lru_gen_attr_group = {
5158 .name = "lru_gen",
5159 .attrs = lru_gen_attrs,
5160};
5161
5162/******************************************************************************
5163 * debugfs interface
5164 ******************************************************************************/
5165
5166static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5167{
5168 struct mem_cgroup *memcg;
5169 loff_t nr_to_skip = *pos;
5170
5171 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5172 if (!m->private)
5173 return ERR_PTR(-ENOMEM);
5174
5175 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5176 do {
5177 int nid;
5178
5179 for_each_node_state(nid, N_MEMORY) {
5180 if (!nr_to_skip--)
5181 return get_lruvec(memcg, nid);
5182 }
5183 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5184
5185 return NULL;
5186}
5187
5188static void lru_gen_seq_stop(struct seq_file *m, void *v)
5189{
5190 if (!IS_ERR_OR_NULL(v))
5191 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5192
5193 kvfree(m->private);
5194 m->private = NULL;
5195}
5196
5197static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5198{
5199 int nid = lruvec_pgdat(v)->node_id;
5200 struct mem_cgroup *memcg = lruvec_memcg(v);
5201
5202 ++*pos;
5203
5204 nid = next_memory_node(nid);
5205 if (nid == MAX_NUMNODES) {
5206 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5207 if (!memcg)
5208 return NULL;
5209
5210 nid = first_memory_node;
5211 }
5212
5213 return get_lruvec(memcg, nid);
5214}
5215
5216static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5217 unsigned long max_seq, unsigned long *min_seq,
5218 unsigned long seq)
5219{
5220 int i;
5221 int type, tier;
5222 int hist = lru_hist_from_seq(seq);
5223 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5224 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5225
5226 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5227 seq_printf(m, " %10d", tier);
5228 for (type = 0; type < ANON_AND_FILE; type++) {
5229 const char *s = " ";
5230 unsigned long n[3] = {};
5231
5232 if (seq == max_seq) {
5233 s = "RT ";
5234 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5235 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5236 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5237 s = "rep";
5238 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5239 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5240 if (tier)
5241 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5242 }
5243
5244 for (i = 0; i < 3; i++)
5245 seq_printf(m, " %10lu%c", n[i], s[i]);
5246 }
5247 seq_putc(m, '\n');
5248 }
5249
5250 if (!mm_state)
5251 return;
5252
5253 seq_puts(m, " ");
5254 for (i = 0; i < NR_MM_STATS; i++) {
5255 const char *s = " ";
5256 unsigned long n = 0;
5257
5258 if (seq == max_seq && NR_HIST_GENS == 1) {
5259 s = "LOYNFA";
5260 n = READ_ONCE(mm_state->stats[hist][i]);
5261 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5262 s = "loynfa";
5263 n = READ_ONCE(mm_state->stats[hist][i]);
5264 }
5265
5266 seq_printf(m, " %10lu%c", n, s[i]);
5267 }
5268 seq_putc(m, '\n');
5269}
5270
5271/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5272static int lru_gen_seq_show(struct seq_file *m, void *v)
5273{
5274 unsigned long seq;
5275 bool full = !debugfs_real_fops(m->file)->write;
5276 struct lruvec *lruvec = v;
5277 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5278 int nid = lruvec_pgdat(lruvec)->node_id;
5279 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5280 DEFINE_MAX_SEQ(lruvec);
5281 DEFINE_MIN_SEQ(lruvec);
5282
5283 if (nid == first_memory_node) {
5284 const char *path = memcg ? m->private : "";
5285
5286#ifdef CONFIG_MEMCG
5287 if (memcg)
5288 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5289#endif
5290 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5291 }
5292
5293 seq_printf(m, " node %5d\n", nid);
5294
5295 if (!full)
5296 seq = min_seq[LRU_GEN_ANON];
5297 else if (max_seq >= MAX_NR_GENS)
5298 seq = max_seq - MAX_NR_GENS + 1;
5299 else
5300 seq = 0;
5301
5302 for (; seq <= max_seq; seq++) {
5303 int type, zone;
5304 int gen = lru_gen_from_seq(seq);
5305 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5306
5307 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5308
5309 for (type = 0; type < ANON_AND_FILE; type++) {
5310 unsigned long size = 0;
5311 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5312
5313 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5314 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5315
5316 seq_printf(m, " %10lu%c", size, mark);
5317 }
5318
5319 seq_putc(m, '\n');
5320
5321 if (full)
5322 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5323 }
5324
5325 return 0;
5326}
5327
5328static const struct seq_operations lru_gen_seq_ops = {
5329 .start = lru_gen_seq_start,
5330 .stop = lru_gen_seq_stop,
5331 .next = lru_gen_seq_next,
5332 .show = lru_gen_seq_show,
5333};
5334
5335static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5336 bool can_swap, bool force_scan)
5337{
5338 DEFINE_MAX_SEQ(lruvec);
5339 DEFINE_MIN_SEQ(lruvec);
5340
5341 if (seq < max_seq)
5342 return 0;
5343
5344 if (seq > max_seq)
5345 return -EINVAL;
5346
5347 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5348 return -ERANGE;
5349
5350 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5351
5352 return 0;
5353}
5354
5355static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5356 int swappiness, unsigned long nr_to_reclaim)
5357{
5358 DEFINE_MAX_SEQ(lruvec);
5359
5360 if (seq + MIN_NR_GENS > max_seq)
5361 return -EINVAL;
5362
5363 sc->nr_reclaimed = 0;
5364
5365 while (!signal_pending(current)) {
5366 DEFINE_MIN_SEQ(lruvec);
5367
5368 if (seq < min_seq[!swappiness])
5369 return 0;
5370
5371 if (sc->nr_reclaimed >= nr_to_reclaim)
5372 return 0;
5373
5374 if (!evict_folios(lruvec, sc, swappiness))
5375 return 0;
5376
5377 cond_resched();
5378 }
5379
5380 return -EINTR;
5381}
5382
5383static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5384 struct scan_control *sc, int swappiness, unsigned long opt)
5385{
5386 struct lruvec *lruvec;
5387 int err = -EINVAL;
5388 struct mem_cgroup *memcg = NULL;
5389
5390 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5391 return -EINVAL;
5392
5393 if (!mem_cgroup_disabled()) {
5394 rcu_read_lock();
5395
5396 memcg = mem_cgroup_from_id(memcg_id);
5397 if (!mem_cgroup_tryget(memcg))
5398 memcg = NULL;
5399
5400 rcu_read_unlock();
5401
5402 if (!memcg)
5403 return -EINVAL;
5404 }
5405
5406 if (memcg_id != mem_cgroup_id(memcg))
5407 goto done;
5408
5409 lruvec = get_lruvec(memcg, nid);
5410
5411 if (swappiness < 0)
5412 swappiness = get_swappiness(lruvec, sc);
5413 else if (swappiness > 200)
5414 goto done;
5415
5416 switch (cmd) {
5417 case '+':
5418 err = run_aging(lruvec, seq, sc, swappiness, opt);
5419 break;
5420 case '-':
5421 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5422 break;
5423 }
5424done:
5425 mem_cgroup_put(memcg);
5426
5427 return err;
5428}
5429
5430/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5431static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5432 size_t len, loff_t *pos)
5433{
5434 void *buf;
5435 char *cur, *next;
5436 unsigned int flags;
5437 struct blk_plug plug;
5438 int err = -EINVAL;
5439 struct scan_control sc = {
5440 .may_writepage = true,
5441 .may_unmap = true,
5442 .may_swap = true,
5443 .reclaim_idx = MAX_NR_ZONES - 1,
5444 .gfp_mask = GFP_KERNEL,
5445 };
5446
5447 buf = kvmalloc(len + 1, GFP_KERNEL);
5448 if (!buf)
5449 return -ENOMEM;
5450
5451 if (copy_from_user(buf, src, len)) {
5452 kvfree(buf);
5453 return -EFAULT;
5454 }
5455
5456 set_task_reclaim_state(current, &sc.reclaim_state);
5457 flags = memalloc_noreclaim_save();
5458 blk_start_plug(&plug);
5459 if (!set_mm_walk(NULL, true)) {
5460 err = -ENOMEM;
5461 goto done;
5462 }
5463
5464 next = buf;
5465 next[len] = '\0';
5466
5467 while ((cur = strsep(&next, ",;\n"))) {
5468 int n;
5469 int end;
5470 char cmd;
5471 unsigned int memcg_id;
5472 unsigned int nid;
5473 unsigned long seq;
5474 unsigned int swappiness = -1;
5475 unsigned long opt = -1;
5476
5477 cur = skip_spaces(cur);
5478 if (!*cur)
5479 continue;
5480
5481 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5482 &seq, &end, &swappiness, &end, &opt, &end);
5483 if (n < 4 || cur[end]) {
5484 err = -EINVAL;
5485 break;
5486 }
5487
5488 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5489 if (err)
5490 break;
5491 }
5492done:
5493 clear_mm_walk();
5494 blk_finish_plug(&plug);
5495 memalloc_noreclaim_restore(flags);
5496 set_task_reclaim_state(current, NULL);
5497
5498 kvfree(buf);
5499
5500 return err ? : len;
5501}
5502
5503static int lru_gen_seq_open(struct inode *inode, struct file *file)
5504{
5505 return seq_open(file, &lru_gen_seq_ops);
5506}
5507
5508static const struct file_operations lru_gen_rw_fops = {
5509 .open = lru_gen_seq_open,
5510 .read = seq_read,
5511 .write = lru_gen_seq_write,
5512 .llseek = seq_lseek,
5513 .release = seq_release,
5514};
5515
5516static const struct file_operations lru_gen_ro_fops = {
5517 .open = lru_gen_seq_open,
5518 .read = seq_read,
5519 .llseek = seq_lseek,
5520 .release = seq_release,
5521};
5522
5523/******************************************************************************
5524 * initialization
5525 ******************************************************************************/
5526
5527void lru_gen_init_pgdat(struct pglist_data *pgdat)
5528{
5529 int i, j;
5530
5531 spin_lock_init(&pgdat->memcg_lru.lock);
5532
5533 for (i = 0; i < MEMCG_NR_GENS; i++) {
5534 for (j = 0; j < MEMCG_NR_BINS; j++)
5535 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5536 }
5537}
5538
5539void lru_gen_init_lruvec(struct lruvec *lruvec)
5540{
5541 int i;
5542 int gen, type, zone;
5543 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5544 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5545
5546 lrugen->max_seq = MIN_NR_GENS + 1;
5547 lrugen->enabled = lru_gen_enabled();
5548
5549 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5550 lrugen->timestamps[i] = jiffies;
5551
5552 for_each_gen_type_zone(gen, type, zone)
5553 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5554
5555 if (mm_state)
5556 mm_state->seq = MIN_NR_GENS;
5557}
5558
5559#ifdef CONFIG_MEMCG
5560
5561void lru_gen_init_memcg(struct mem_cgroup *memcg)
5562{
5563 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5564
5565 if (!mm_list)
5566 return;
5567
5568 INIT_LIST_HEAD(&mm_list->fifo);
5569 spin_lock_init(&mm_list->lock);
5570}
5571
5572void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5573{
5574 int i;
5575 int nid;
5576 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5577
5578 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5579
5580 for_each_node(nid) {
5581 struct lruvec *lruvec = get_lruvec(memcg, nid);
5582 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5583
5584 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5585 sizeof(lruvec->lrugen.nr_pages)));
5586
5587 lruvec->lrugen.list.next = LIST_POISON1;
5588
5589 if (!mm_state)
5590 continue;
5591
5592 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5593 bitmap_free(mm_state->filters[i]);
5594 mm_state->filters[i] = NULL;
5595 }
5596 }
5597}
5598
5599#endif /* CONFIG_MEMCG */
5600
5601static int __init init_lru_gen(void)
5602{
5603 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5604 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5605
5606 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5607 pr_err("lru_gen: failed to create sysfs group\n");
5608
5609 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5610 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5611
5612 return 0;
5613};
5614late_initcall(init_lru_gen);
5615
5616#else /* !CONFIG_LRU_GEN */
5617
5618static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5619{
5620 BUILD_BUG();
5621}
5622
5623static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5624{
5625 BUILD_BUG();
5626}
5627
5628static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5629{
5630 BUILD_BUG();
5631}
5632
5633#endif /* CONFIG_LRU_GEN */
5634
5635static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5636{
5637 unsigned long nr[NR_LRU_LISTS];
5638 unsigned long targets[NR_LRU_LISTS];
5639 unsigned long nr_to_scan;
5640 enum lru_list lru;
5641 unsigned long nr_reclaimed = 0;
5642 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5643 bool proportional_reclaim;
5644 struct blk_plug plug;
5645
5646 if (lru_gen_enabled() && !root_reclaim(sc)) {
5647 lru_gen_shrink_lruvec(lruvec, sc);
5648 return;
5649 }
5650
5651 get_scan_count(lruvec, sc, nr);
5652
5653 /* Record the original scan target for proportional adjustments later */
5654 memcpy(targets, nr, sizeof(nr));
5655
5656 /*
5657 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5658 * event that can occur when there is little memory pressure e.g.
5659 * multiple streaming readers/writers. Hence, we do not abort scanning
5660 * when the requested number of pages are reclaimed when scanning at
5661 * DEF_PRIORITY on the assumption that the fact we are direct
5662 * reclaiming implies that kswapd is not keeping up and it is best to
5663 * do a batch of work at once. For memcg reclaim one check is made to
5664 * abort proportional reclaim if either the file or anon lru has already
5665 * dropped to zero at the first pass.
5666 */
5667 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5668 sc->priority == DEF_PRIORITY);
5669
5670 blk_start_plug(&plug);
5671 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5672 nr[LRU_INACTIVE_FILE]) {
5673 unsigned long nr_anon, nr_file, percentage;
5674 unsigned long nr_scanned;
5675
5676 for_each_evictable_lru(lru) {
5677 if (nr[lru]) {
5678 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5679 nr[lru] -= nr_to_scan;
5680
5681 nr_reclaimed += shrink_list(lru, nr_to_scan,
5682 lruvec, sc);
5683 }
5684 }
5685
5686 cond_resched();
5687
5688 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5689 continue;
5690
5691 /*
5692 * For kswapd and memcg, reclaim at least the number of pages
5693 * requested. Ensure that the anon and file LRUs are scanned
5694 * proportionally what was requested by get_scan_count(). We
5695 * stop reclaiming one LRU and reduce the amount scanning
5696 * proportional to the original scan target.
5697 */
5698 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5699 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5700
5701 /*
5702 * It's just vindictive to attack the larger once the smaller
5703 * has gone to zero. And given the way we stop scanning the
5704 * smaller below, this makes sure that we only make one nudge
5705 * towards proportionality once we've got nr_to_reclaim.
5706 */
5707 if (!nr_file || !nr_anon)
5708 break;
5709
5710 if (nr_file > nr_anon) {
5711 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5712 targets[LRU_ACTIVE_ANON] + 1;
5713 lru = LRU_BASE;
5714 percentage = nr_anon * 100 / scan_target;
5715 } else {
5716 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5717 targets[LRU_ACTIVE_FILE] + 1;
5718 lru = LRU_FILE;
5719 percentage = nr_file * 100 / scan_target;
5720 }
5721
5722 /* Stop scanning the smaller of the LRU */
5723 nr[lru] = 0;
5724 nr[lru + LRU_ACTIVE] = 0;
5725
5726 /*
5727 * Recalculate the other LRU scan count based on its original
5728 * scan target and the percentage scanning already complete
5729 */
5730 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5731 nr_scanned = targets[lru] - nr[lru];
5732 nr[lru] = targets[lru] * (100 - percentage) / 100;
5733 nr[lru] -= min(nr[lru], nr_scanned);
5734
5735 lru += LRU_ACTIVE;
5736 nr_scanned = targets[lru] - nr[lru];
5737 nr[lru] = targets[lru] * (100 - percentage) / 100;
5738 nr[lru] -= min(nr[lru], nr_scanned);
5739 }
5740 blk_finish_plug(&plug);
5741 sc->nr_reclaimed += nr_reclaimed;
5742
5743 /*
5744 * Even if we did not try to evict anon pages at all, we want to
5745 * rebalance the anon lru active/inactive ratio.
5746 */
5747 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5748 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5749 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5750 sc, LRU_ACTIVE_ANON);
5751}
5752
5753/* Use reclaim/compaction for costly allocs or under memory pressure */
5754static bool in_reclaim_compaction(struct scan_control *sc)
5755{
5756 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5757 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5758 sc->priority < DEF_PRIORITY - 2))
5759 return true;
5760
5761 return false;
5762}
5763
5764/*
5765 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5766 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5767 * true if more pages should be reclaimed such that when the page allocator
5768 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5769 * It will give up earlier than that if there is difficulty reclaiming pages.
5770 */
5771static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5772 unsigned long nr_reclaimed,
5773 struct scan_control *sc)
5774{
5775 unsigned long pages_for_compaction;
5776 unsigned long inactive_lru_pages;
5777 int z;
5778
5779 /* If not in reclaim/compaction mode, stop */
5780 if (!in_reclaim_compaction(sc))
5781 return false;
5782
5783 /*
5784 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5785 * number of pages that were scanned. This will return to the caller
5786 * with the risk reclaim/compaction and the resulting allocation attempt
5787 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5788 * allocations through requiring that the full LRU list has been scanned
5789 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5790 * scan, but that approximation was wrong, and there were corner cases
5791 * where always a non-zero amount of pages were scanned.
5792 */
5793 if (!nr_reclaimed)
5794 return false;
5795
5796 /* If compaction would go ahead or the allocation would succeed, stop */
5797 for (z = 0; z <= sc->reclaim_idx; z++) {
5798 struct zone *zone = &pgdat->node_zones[z];
5799 if (!managed_zone(zone))
5800 continue;
5801
5802 /* Allocation can already succeed, nothing to do */
5803 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5804 sc->reclaim_idx, 0))
5805 return false;
5806
5807 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5808 return false;
5809 }
5810
5811 /*
5812 * If we have not reclaimed enough pages for compaction and the
5813 * inactive lists are large enough, continue reclaiming
5814 */
5815 pages_for_compaction = compact_gap(sc->order);
5816 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5817 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5818 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5819
5820 return inactive_lru_pages > pages_for_compaction;
5821}
5822
5823static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5824{
5825 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5826 struct mem_cgroup *memcg;
5827
5828 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
5829 do {
5830 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5831 unsigned long reclaimed;
5832 unsigned long scanned;
5833
5834 /*
5835 * This loop can become CPU-bound when target memcgs
5836 * aren't eligible for reclaim - either because they
5837 * don't have any reclaimable pages, or because their
5838 * memory is explicitly protected. Avoid soft lockups.
5839 */
5840 cond_resched();
5841
5842 mem_cgroup_calculate_protection(target_memcg, memcg);
5843
5844 if (mem_cgroup_below_min(target_memcg, memcg)) {
5845 /*
5846 * Hard protection.
5847 * If there is no reclaimable memory, OOM.
5848 */
5849 continue;
5850 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5851 /*
5852 * Soft protection.
5853 * Respect the protection only as long as
5854 * there is an unprotected supply
5855 * of reclaimable memory from other cgroups.
5856 */
5857 if (!sc->memcg_low_reclaim) {
5858 sc->memcg_low_skipped = 1;
5859 continue;
5860 }
5861 memcg_memory_event(memcg, MEMCG_LOW);
5862 }
5863
5864 reclaimed = sc->nr_reclaimed;
5865 scanned = sc->nr_scanned;
5866
5867 shrink_lruvec(lruvec, sc);
5868
5869 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5870 sc->priority);
5871
5872 /* Record the group's reclaim efficiency */
5873 if (!sc->proactive)
5874 vmpressure(sc->gfp_mask, memcg, false,
5875 sc->nr_scanned - scanned,
5876 sc->nr_reclaimed - reclaimed);
5877
5878 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5879}
5880
5881static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5882{
5883 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5884 struct lruvec *target_lruvec;
5885 bool reclaimable = false;
5886
5887 if (lru_gen_enabled() && root_reclaim(sc)) {
5888 lru_gen_shrink_node(pgdat, sc);
5889 return;
5890 }
5891
5892 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5893
5894again:
5895 memset(&sc->nr, 0, sizeof(sc->nr));
5896
5897 nr_reclaimed = sc->nr_reclaimed;
5898 nr_scanned = sc->nr_scanned;
5899
5900 prepare_scan_control(pgdat, sc);
5901
5902 shrink_node_memcgs(pgdat, sc);
5903
5904 flush_reclaim_state(sc);
5905
5906 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5907
5908 /* Record the subtree's reclaim efficiency */
5909 if (!sc->proactive)
5910 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5911 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5912
5913 if (nr_node_reclaimed)
5914 reclaimable = true;
5915
5916 if (current_is_kswapd()) {
5917 /*
5918 * If reclaim is isolating dirty pages under writeback,
5919 * it implies that the long-lived page allocation rate
5920 * is exceeding the page laundering rate. Either the
5921 * global limits are not being effective at throttling
5922 * processes due to the page distribution throughout
5923 * zones or there is heavy usage of a slow backing
5924 * device. The only option is to throttle from reclaim
5925 * context which is not ideal as there is no guarantee
5926 * the dirtying process is throttled in the same way
5927 * balance_dirty_pages() manages.
5928 *
5929 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5930 * count the number of pages under pages flagged for
5931 * immediate reclaim and stall if any are encountered
5932 * in the nr_immediate check below.
5933 */
5934 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5935 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5936
5937 /* Allow kswapd to start writing pages during reclaim.*/
5938 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5939 set_bit(PGDAT_DIRTY, &pgdat->flags);
5940
5941 /*
5942 * If kswapd scans pages marked for immediate
5943 * reclaim and under writeback (nr_immediate), it
5944 * implies that pages are cycling through the LRU
5945 * faster than they are written so forcibly stall
5946 * until some pages complete writeback.
5947 */
5948 if (sc->nr.immediate)
5949 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5950 }
5951
5952 /*
5953 * Tag a node/memcg as congested if all the dirty pages were marked
5954 * for writeback and immediate reclaim (counted in nr.congested).
5955 *
5956 * Legacy memcg will stall in page writeback so avoid forcibly
5957 * stalling in reclaim_throttle().
5958 */
5959 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
5960 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
5961 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
5962
5963 if (current_is_kswapd())
5964 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
5965 }
5966
5967 /*
5968 * Stall direct reclaim for IO completions if the lruvec is
5969 * node is congested. Allow kswapd to continue until it
5970 * starts encountering unqueued dirty pages or cycling through
5971 * the LRU too quickly.
5972 */
5973 if (!current_is_kswapd() && current_may_throttle() &&
5974 !sc->hibernation_mode &&
5975 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
5976 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
5977 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
5978
5979 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
5980 goto again;
5981
5982 /*
5983 * Kswapd gives up on balancing particular nodes after too
5984 * many failures to reclaim anything from them and goes to
5985 * sleep. On reclaim progress, reset the failure counter. A
5986 * successful direct reclaim run will revive a dormant kswapd.
5987 */
5988 if (reclaimable)
5989 pgdat->kswapd_failures = 0;
5990}
5991
5992/*
5993 * Returns true if compaction should go ahead for a costly-order request, or
5994 * the allocation would already succeed without compaction. Return false if we
5995 * should reclaim first.
5996 */
5997static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
5998{
5999 unsigned long watermark;
6000
6001 if (!gfp_compaction_allowed(sc->gfp_mask))
6002 return false;
6003
6004 /* Allocation can already succeed, nothing to do */
6005 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6006 sc->reclaim_idx, 0))
6007 return true;
6008
6009 /* Compaction cannot yet proceed. Do reclaim. */
6010 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6011 return false;
6012
6013 /*
6014 * Compaction is already possible, but it takes time to run and there
6015 * are potentially other callers using the pages just freed. So proceed
6016 * with reclaim to make a buffer of free pages available to give
6017 * compaction a reasonable chance of completing and allocating the page.
6018 * Note that we won't actually reclaim the whole buffer in one attempt
6019 * as the target watermark in should_continue_reclaim() is lower. But if
6020 * we are already above the high+gap watermark, don't reclaim at all.
6021 */
6022 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6023
6024 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6025}
6026
6027static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6028{
6029 /*
6030 * If reclaim is making progress greater than 12% efficiency then
6031 * wake all the NOPROGRESS throttled tasks.
6032 */
6033 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6034 wait_queue_head_t *wqh;
6035
6036 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6037 if (waitqueue_active(wqh))
6038 wake_up(wqh);
6039
6040 return;
6041 }
6042
6043 /*
6044 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6045 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6046 * under writeback and marked for immediate reclaim at the tail of the
6047 * LRU.
6048 */
6049 if (current_is_kswapd() || cgroup_reclaim(sc))
6050 return;
6051
6052 /* Throttle if making no progress at high prioities. */
6053 if (sc->priority == 1 && !sc->nr_reclaimed)
6054 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6055}
6056
6057/*
6058 * This is the direct reclaim path, for page-allocating processes. We only
6059 * try to reclaim pages from zones which will satisfy the caller's allocation
6060 * request.
6061 *
6062 * If a zone is deemed to be full of pinned pages then just give it a light
6063 * scan then give up on it.
6064 */
6065static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6066{
6067 struct zoneref *z;
6068 struct zone *zone;
6069 unsigned long nr_soft_reclaimed;
6070 unsigned long nr_soft_scanned;
6071 gfp_t orig_mask;
6072 pg_data_t *last_pgdat = NULL;
6073 pg_data_t *first_pgdat = NULL;
6074
6075 /*
6076 * If the number of buffer_heads in the machine exceeds the maximum
6077 * allowed level, force direct reclaim to scan the highmem zone as
6078 * highmem pages could be pinning lowmem pages storing buffer_heads
6079 */
6080 orig_mask = sc->gfp_mask;
6081 if (buffer_heads_over_limit) {
6082 sc->gfp_mask |= __GFP_HIGHMEM;
6083 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6084 }
6085
6086 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6087 sc->reclaim_idx, sc->nodemask) {
6088 /*
6089 * Take care memory controller reclaiming has small influence
6090 * to global LRU.
6091 */
6092 if (!cgroup_reclaim(sc)) {
6093 if (!cpuset_zone_allowed(zone,
6094 GFP_KERNEL | __GFP_HARDWALL))
6095 continue;
6096
6097 /*
6098 * If we already have plenty of memory free for
6099 * compaction in this zone, don't free any more.
6100 * Even though compaction is invoked for any
6101 * non-zero order, only frequent costly order
6102 * reclamation is disruptive enough to become a
6103 * noticeable problem, like transparent huge
6104 * page allocations.
6105 */
6106 if (IS_ENABLED(CONFIG_COMPACTION) &&
6107 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6108 compaction_ready(zone, sc)) {
6109 sc->compaction_ready = true;
6110 continue;
6111 }
6112
6113 /*
6114 * Shrink each node in the zonelist once. If the
6115 * zonelist is ordered by zone (not the default) then a
6116 * node may be shrunk multiple times but in that case
6117 * the user prefers lower zones being preserved.
6118 */
6119 if (zone->zone_pgdat == last_pgdat)
6120 continue;
6121
6122 /*
6123 * This steals pages from memory cgroups over softlimit
6124 * and returns the number of reclaimed pages and
6125 * scanned pages. This works for global memory pressure
6126 * and balancing, not for a memcg's limit.
6127 */
6128 nr_soft_scanned = 0;
6129 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6130 sc->order, sc->gfp_mask,
6131 &nr_soft_scanned);
6132 sc->nr_reclaimed += nr_soft_reclaimed;
6133 sc->nr_scanned += nr_soft_scanned;
6134 /* need some check for avoid more shrink_zone() */
6135 }
6136
6137 if (!first_pgdat)
6138 first_pgdat = zone->zone_pgdat;
6139
6140 /* See comment about same check for global reclaim above */
6141 if (zone->zone_pgdat == last_pgdat)
6142 continue;
6143 last_pgdat = zone->zone_pgdat;
6144 shrink_node(zone->zone_pgdat, sc);
6145 }
6146
6147 if (first_pgdat)
6148 consider_reclaim_throttle(first_pgdat, sc);
6149
6150 /*
6151 * Restore to original mask to avoid the impact on the caller if we
6152 * promoted it to __GFP_HIGHMEM.
6153 */
6154 sc->gfp_mask = orig_mask;
6155}
6156
6157static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6158{
6159 struct lruvec *target_lruvec;
6160 unsigned long refaults;
6161
6162 if (lru_gen_enabled())
6163 return;
6164
6165 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6166 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6167 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6168 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6169 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6170}
6171
6172/*
6173 * This is the main entry point to direct page reclaim.
6174 *
6175 * If a full scan of the inactive list fails to free enough memory then we
6176 * are "out of memory" and something needs to be killed.
6177 *
6178 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6179 * high - the zone may be full of dirty or under-writeback pages, which this
6180 * caller can't do much about. We kick the writeback threads and take explicit
6181 * naps in the hope that some of these pages can be written. But if the
6182 * allocating task holds filesystem locks which prevent writeout this might not
6183 * work, and the allocation attempt will fail.
6184 *
6185 * returns: 0, if no pages reclaimed
6186 * else, the number of pages reclaimed
6187 */
6188static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6189 struct scan_control *sc)
6190{
6191 int initial_priority = sc->priority;
6192 pg_data_t *last_pgdat;
6193 struct zoneref *z;
6194 struct zone *zone;
6195retry:
6196 delayacct_freepages_start();
6197
6198 if (!cgroup_reclaim(sc))
6199 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6200
6201 do {
6202 if (!sc->proactive)
6203 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6204 sc->priority);
6205 sc->nr_scanned = 0;
6206 shrink_zones(zonelist, sc);
6207
6208 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6209 break;
6210
6211 if (sc->compaction_ready)
6212 break;
6213
6214 /*
6215 * If we're getting trouble reclaiming, start doing
6216 * writepage even in laptop mode.
6217 */
6218 if (sc->priority < DEF_PRIORITY - 2)
6219 sc->may_writepage = 1;
6220 } while (--sc->priority >= 0);
6221
6222 last_pgdat = NULL;
6223 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6224 sc->nodemask) {
6225 if (zone->zone_pgdat == last_pgdat)
6226 continue;
6227 last_pgdat = zone->zone_pgdat;
6228
6229 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6230
6231 if (cgroup_reclaim(sc)) {
6232 struct lruvec *lruvec;
6233
6234 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6235 zone->zone_pgdat);
6236 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6237 }
6238 }
6239
6240 delayacct_freepages_end();
6241
6242 if (sc->nr_reclaimed)
6243 return sc->nr_reclaimed;
6244
6245 /* Aborted reclaim to try compaction? don't OOM, then */
6246 if (sc->compaction_ready)
6247 return 1;
6248
6249 /*
6250 * We make inactive:active ratio decisions based on the node's
6251 * composition of memory, but a restrictive reclaim_idx or a
6252 * memory.low cgroup setting can exempt large amounts of
6253 * memory from reclaim. Neither of which are very common, so
6254 * instead of doing costly eligibility calculations of the
6255 * entire cgroup subtree up front, we assume the estimates are
6256 * good, and retry with forcible deactivation if that fails.
6257 */
6258 if (sc->skipped_deactivate) {
6259 sc->priority = initial_priority;
6260 sc->force_deactivate = 1;
6261 sc->skipped_deactivate = 0;
6262 goto retry;
6263 }
6264
6265 /* Untapped cgroup reserves? Don't OOM, retry. */
6266 if (sc->memcg_low_skipped) {
6267 sc->priority = initial_priority;
6268 sc->force_deactivate = 0;
6269 sc->memcg_low_reclaim = 1;
6270 sc->memcg_low_skipped = 0;
6271 goto retry;
6272 }
6273
6274 return 0;
6275}
6276
6277static bool allow_direct_reclaim(pg_data_t *pgdat)
6278{
6279 struct zone *zone;
6280 unsigned long pfmemalloc_reserve = 0;
6281 unsigned long free_pages = 0;
6282 int i;
6283 bool wmark_ok;
6284
6285 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6286 return true;
6287
6288 for (i = 0; i <= ZONE_NORMAL; i++) {
6289 zone = &pgdat->node_zones[i];
6290 if (!managed_zone(zone))
6291 continue;
6292
6293 if (!zone_reclaimable_pages(zone))
6294 continue;
6295
6296 pfmemalloc_reserve += min_wmark_pages(zone);
6297 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6298 }
6299
6300 /* If there are no reserves (unexpected config) then do not throttle */
6301 if (!pfmemalloc_reserve)
6302 return true;
6303
6304 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6305
6306 /* kswapd must be awake if processes are being throttled */
6307 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6308 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6309 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6310
6311 wake_up_interruptible(&pgdat->kswapd_wait);
6312 }
6313
6314 return wmark_ok;
6315}
6316
6317/*
6318 * Throttle direct reclaimers if backing storage is backed by the network
6319 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6320 * depleted. kswapd will continue to make progress and wake the processes
6321 * when the low watermark is reached.
6322 *
6323 * Returns true if a fatal signal was delivered during throttling. If this
6324 * happens, the page allocator should not consider triggering the OOM killer.
6325 */
6326static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6327 nodemask_t *nodemask)
6328{
6329 struct zoneref *z;
6330 struct zone *zone;
6331 pg_data_t *pgdat = NULL;
6332
6333 /*
6334 * Kernel threads should not be throttled as they may be indirectly
6335 * responsible for cleaning pages necessary for reclaim to make forward
6336 * progress. kjournald for example may enter direct reclaim while
6337 * committing a transaction where throttling it could forcing other
6338 * processes to block on log_wait_commit().
6339 */
6340 if (current->flags & PF_KTHREAD)
6341 goto out;
6342
6343 /*
6344 * If a fatal signal is pending, this process should not throttle.
6345 * It should return quickly so it can exit and free its memory
6346 */
6347 if (fatal_signal_pending(current))
6348 goto out;
6349
6350 /*
6351 * Check if the pfmemalloc reserves are ok by finding the first node
6352 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6353 * GFP_KERNEL will be required for allocating network buffers when
6354 * swapping over the network so ZONE_HIGHMEM is unusable.
6355 *
6356 * Throttling is based on the first usable node and throttled processes
6357 * wait on a queue until kswapd makes progress and wakes them. There
6358 * is an affinity then between processes waking up and where reclaim
6359 * progress has been made assuming the process wakes on the same node.
6360 * More importantly, processes running on remote nodes will not compete
6361 * for remote pfmemalloc reserves and processes on different nodes
6362 * should make reasonable progress.
6363 */
6364 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6365 gfp_zone(gfp_mask), nodemask) {
6366 if (zone_idx(zone) > ZONE_NORMAL)
6367 continue;
6368
6369 /* Throttle based on the first usable node */
6370 pgdat = zone->zone_pgdat;
6371 if (allow_direct_reclaim(pgdat))
6372 goto out;
6373 break;
6374 }
6375
6376 /* If no zone was usable by the allocation flags then do not throttle */
6377 if (!pgdat)
6378 goto out;
6379
6380 /* Account for the throttling */
6381 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6382
6383 /*
6384 * If the caller cannot enter the filesystem, it's possible that it
6385 * is due to the caller holding an FS lock or performing a journal
6386 * transaction in the case of a filesystem like ext[3|4]. In this case,
6387 * it is not safe to block on pfmemalloc_wait as kswapd could be
6388 * blocked waiting on the same lock. Instead, throttle for up to a
6389 * second before continuing.
6390 */
6391 if (!(gfp_mask & __GFP_FS))
6392 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6393 allow_direct_reclaim(pgdat), HZ);
6394 else
6395 /* Throttle until kswapd wakes the process */
6396 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6397 allow_direct_reclaim(pgdat));
6398
6399 if (fatal_signal_pending(current))
6400 return true;
6401
6402out:
6403 return false;
6404}
6405
6406unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6407 gfp_t gfp_mask, nodemask_t *nodemask)
6408{
6409 unsigned long nr_reclaimed;
6410 struct scan_control sc = {
6411 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6412 .gfp_mask = current_gfp_context(gfp_mask),
6413 .reclaim_idx = gfp_zone(gfp_mask),
6414 .order = order,
6415 .nodemask = nodemask,
6416 .priority = DEF_PRIORITY,
6417 .may_writepage = !laptop_mode,
6418 .may_unmap = 1,
6419 .may_swap = 1,
6420 };
6421
6422 /*
6423 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6424 * Confirm they are large enough for max values.
6425 */
6426 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6427 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6428 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6429
6430 /*
6431 * Do not enter reclaim if fatal signal was delivered while throttled.
6432 * 1 is returned so that the page allocator does not OOM kill at this
6433 * point.
6434 */
6435 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6436 return 1;
6437
6438 set_task_reclaim_state(current, &sc.reclaim_state);
6439 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6440
6441 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6442
6443 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6444 set_task_reclaim_state(current, NULL);
6445
6446 return nr_reclaimed;
6447}
6448
6449#ifdef CONFIG_MEMCG
6450
6451/* Only used by soft limit reclaim. Do not reuse for anything else. */
6452unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6453 gfp_t gfp_mask, bool noswap,
6454 pg_data_t *pgdat,
6455 unsigned long *nr_scanned)
6456{
6457 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6458 struct scan_control sc = {
6459 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6460 .target_mem_cgroup = memcg,
6461 .may_writepage = !laptop_mode,
6462 .may_unmap = 1,
6463 .reclaim_idx = MAX_NR_ZONES - 1,
6464 .may_swap = !noswap,
6465 };
6466
6467 WARN_ON_ONCE(!current->reclaim_state);
6468
6469 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6470 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6471
6472 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6473 sc.gfp_mask);
6474
6475 /*
6476 * NOTE: Although we can get the priority field, using it
6477 * here is not a good idea, since it limits the pages we can scan.
6478 * if we don't reclaim here, the shrink_node from balance_pgdat
6479 * will pick up pages from other mem cgroup's as well. We hack
6480 * the priority and make it zero.
6481 */
6482 shrink_lruvec(lruvec, &sc);
6483
6484 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6485
6486 *nr_scanned = sc.nr_scanned;
6487
6488 return sc.nr_reclaimed;
6489}
6490
6491unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6492 unsigned long nr_pages,
6493 gfp_t gfp_mask,
6494 unsigned int reclaim_options)
6495{
6496 unsigned long nr_reclaimed;
6497 unsigned int noreclaim_flag;
6498 struct scan_control sc = {
6499 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6500 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6501 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6502 .reclaim_idx = MAX_NR_ZONES - 1,
6503 .target_mem_cgroup = memcg,
6504 .priority = DEF_PRIORITY,
6505 .may_writepage = !laptop_mode,
6506 .may_unmap = 1,
6507 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6508 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6509 };
6510 /*
6511 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6512 * equal pressure on all the nodes. This is based on the assumption that
6513 * the reclaim does not bail out early.
6514 */
6515 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6516
6517 set_task_reclaim_state(current, &sc.reclaim_state);
6518 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6519 noreclaim_flag = memalloc_noreclaim_save();
6520
6521 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6522
6523 memalloc_noreclaim_restore(noreclaim_flag);
6524 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6525 set_task_reclaim_state(current, NULL);
6526
6527 return nr_reclaimed;
6528}
6529#endif
6530
6531static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6532{
6533 struct mem_cgroup *memcg;
6534 struct lruvec *lruvec;
6535
6536 if (lru_gen_enabled()) {
6537 lru_gen_age_node(pgdat, sc);
6538 return;
6539 }
6540
6541 if (!can_age_anon_pages(pgdat, sc))
6542 return;
6543
6544 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6545 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6546 return;
6547
6548 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6549 do {
6550 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6551 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6552 sc, LRU_ACTIVE_ANON);
6553 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6554 } while (memcg);
6555}
6556
6557static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6558{
6559 int i;
6560 struct zone *zone;
6561
6562 /*
6563 * Check for watermark boosts top-down as the higher zones
6564 * are more likely to be boosted. Both watermarks and boosts
6565 * should not be checked at the same time as reclaim would
6566 * start prematurely when there is no boosting and a lower
6567 * zone is balanced.
6568 */
6569 for (i = highest_zoneidx; i >= 0; i--) {
6570 zone = pgdat->node_zones + i;
6571 if (!managed_zone(zone))
6572 continue;
6573
6574 if (zone->watermark_boost)
6575 return true;
6576 }
6577
6578 return false;
6579}
6580
6581/*
6582 * Returns true if there is an eligible zone balanced for the request order
6583 * and highest_zoneidx
6584 */
6585static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6586{
6587 int i;
6588 unsigned long mark = -1;
6589 struct zone *zone;
6590
6591 /*
6592 * Check watermarks bottom-up as lower zones are more likely to
6593 * meet watermarks.
6594 */
6595 for (i = 0; i <= highest_zoneidx; i++) {
6596 zone = pgdat->node_zones + i;
6597
6598 if (!managed_zone(zone))
6599 continue;
6600
6601 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6602 mark = wmark_pages(zone, WMARK_PROMO);
6603 else
6604 mark = high_wmark_pages(zone);
6605 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6606 return true;
6607 }
6608
6609 /*
6610 * If a node has no managed zone within highest_zoneidx, it does not
6611 * need balancing by definition. This can happen if a zone-restricted
6612 * allocation tries to wake a remote kswapd.
6613 */
6614 if (mark == -1)
6615 return true;
6616
6617 return false;
6618}
6619
6620/* Clear pgdat state for congested, dirty or under writeback. */
6621static void clear_pgdat_congested(pg_data_t *pgdat)
6622{
6623 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6624
6625 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6626 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6627 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6628 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6629}
6630
6631/*
6632 * Prepare kswapd for sleeping. This verifies that there are no processes
6633 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6634 *
6635 * Returns true if kswapd is ready to sleep
6636 */
6637static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6638 int highest_zoneidx)
6639{
6640 /*
6641 * The throttled processes are normally woken up in balance_pgdat() as
6642 * soon as allow_direct_reclaim() is true. But there is a potential
6643 * race between when kswapd checks the watermarks and a process gets
6644 * throttled. There is also a potential race if processes get
6645 * throttled, kswapd wakes, a large process exits thereby balancing the
6646 * zones, which causes kswapd to exit balance_pgdat() before reaching
6647 * the wake up checks. If kswapd is going to sleep, no process should
6648 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6649 * the wake up is premature, processes will wake kswapd and get
6650 * throttled again. The difference from wake ups in balance_pgdat() is
6651 * that here we are under prepare_to_wait().
6652 */
6653 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6654 wake_up_all(&pgdat->pfmemalloc_wait);
6655
6656 /* Hopeless node, leave it to direct reclaim */
6657 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6658 return true;
6659
6660 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6661 clear_pgdat_congested(pgdat);
6662 return true;
6663 }
6664
6665 return false;
6666}
6667
6668/*
6669 * kswapd shrinks a node of pages that are at or below the highest usable
6670 * zone that is currently unbalanced.
6671 *
6672 * Returns true if kswapd scanned at least the requested number of pages to
6673 * reclaim or if the lack of progress was due to pages under writeback.
6674 * This is used to determine if the scanning priority needs to be raised.
6675 */
6676static bool kswapd_shrink_node(pg_data_t *pgdat,
6677 struct scan_control *sc)
6678{
6679 struct zone *zone;
6680 int z;
6681
6682 /* Reclaim a number of pages proportional to the number of zones */
6683 sc->nr_to_reclaim = 0;
6684 for (z = 0; z <= sc->reclaim_idx; z++) {
6685 zone = pgdat->node_zones + z;
6686 if (!managed_zone(zone))
6687 continue;
6688
6689 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6690 }
6691
6692 /*
6693 * Historically care was taken to put equal pressure on all zones but
6694 * now pressure is applied based on node LRU order.
6695 */
6696 shrink_node(pgdat, sc);
6697
6698 /*
6699 * Fragmentation may mean that the system cannot be rebalanced for
6700 * high-order allocations. If twice the allocation size has been
6701 * reclaimed then recheck watermarks only at order-0 to prevent
6702 * excessive reclaim. Assume that a process requested a high-order
6703 * can direct reclaim/compact.
6704 */
6705 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6706 sc->order = 0;
6707
6708 return sc->nr_scanned >= sc->nr_to_reclaim;
6709}
6710
6711/* Page allocator PCP high watermark is lowered if reclaim is active. */
6712static inline void
6713update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6714{
6715 int i;
6716 struct zone *zone;
6717
6718 for (i = 0; i <= highest_zoneidx; i++) {
6719 zone = pgdat->node_zones + i;
6720
6721 if (!managed_zone(zone))
6722 continue;
6723
6724 if (active)
6725 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6726 else
6727 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6728 }
6729}
6730
6731static inline void
6732set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6733{
6734 update_reclaim_active(pgdat, highest_zoneidx, true);
6735}
6736
6737static inline void
6738clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6739{
6740 update_reclaim_active(pgdat, highest_zoneidx, false);
6741}
6742
6743/*
6744 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6745 * that are eligible for use by the caller until at least one zone is
6746 * balanced.
6747 *
6748 * Returns the order kswapd finished reclaiming at.
6749 *
6750 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6751 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6752 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6753 * or lower is eligible for reclaim until at least one usable zone is
6754 * balanced.
6755 */
6756static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6757{
6758 int i;
6759 unsigned long nr_soft_reclaimed;
6760 unsigned long nr_soft_scanned;
6761 unsigned long pflags;
6762 unsigned long nr_boost_reclaim;
6763 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6764 bool boosted;
6765 struct zone *zone;
6766 struct scan_control sc = {
6767 .gfp_mask = GFP_KERNEL,
6768 .order = order,
6769 .may_unmap = 1,
6770 };
6771
6772 set_task_reclaim_state(current, &sc.reclaim_state);
6773 psi_memstall_enter(&pflags);
6774 __fs_reclaim_acquire(_THIS_IP_);
6775
6776 count_vm_event(PAGEOUTRUN);
6777
6778 /*
6779 * Account for the reclaim boost. Note that the zone boost is left in
6780 * place so that parallel allocations that are near the watermark will
6781 * stall or direct reclaim until kswapd is finished.
6782 */
6783 nr_boost_reclaim = 0;
6784 for (i = 0; i <= highest_zoneidx; i++) {
6785 zone = pgdat->node_zones + i;
6786 if (!managed_zone(zone))
6787 continue;
6788
6789 nr_boost_reclaim += zone->watermark_boost;
6790 zone_boosts[i] = zone->watermark_boost;
6791 }
6792 boosted = nr_boost_reclaim;
6793
6794restart:
6795 set_reclaim_active(pgdat, highest_zoneidx);
6796 sc.priority = DEF_PRIORITY;
6797 do {
6798 unsigned long nr_reclaimed = sc.nr_reclaimed;
6799 bool raise_priority = true;
6800 bool balanced;
6801 bool ret;
6802
6803 sc.reclaim_idx = highest_zoneidx;
6804
6805 /*
6806 * If the number of buffer_heads exceeds the maximum allowed
6807 * then consider reclaiming from all zones. This has a dual
6808 * purpose -- on 64-bit systems it is expected that
6809 * buffer_heads are stripped during active rotation. On 32-bit
6810 * systems, highmem pages can pin lowmem memory and shrinking
6811 * buffers can relieve lowmem pressure. Reclaim may still not
6812 * go ahead if all eligible zones for the original allocation
6813 * request are balanced to avoid excessive reclaim from kswapd.
6814 */
6815 if (buffer_heads_over_limit) {
6816 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6817 zone = pgdat->node_zones + i;
6818 if (!managed_zone(zone))
6819 continue;
6820
6821 sc.reclaim_idx = i;
6822 break;
6823 }
6824 }
6825
6826 /*
6827 * If the pgdat is imbalanced then ignore boosting and preserve
6828 * the watermarks for a later time and restart. Note that the
6829 * zone watermarks will be still reset at the end of balancing
6830 * on the grounds that the normal reclaim should be enough to
6831 * re-evaluate if boosting is required when kswapd next wakes.
6832 */
6833 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6834 if (!balanced && nr_boost_reclaim) {
6835 nr_boost_reclaim = 0;
6836 goto restart;
6837 }
6838
6839 /*
6840 * If boosting is not active then only reclaim if there are no
6841 * eligible zones. Note that sc.reclaim_idx is not used as
6842 * buffer_heads_over_limit may have adjusted it.
6843 */
6844 if (!nr_boost_reclaim && balanced)
6845 goto out;
6846
6847 /* Limit the priority of boosting to avoid reclaim writeback */
6848 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6849 raise_priority = false;
6850
6851 /*
6852 * Do not writeback or swap pages for boosted reclaim. The
6853 * intent is to relieve pressure not issue sub-optimal IO
6854 * from reclaim context. If no pages are reclaimed, the
6855 * reclaim will be aborted.
6856 */
6857 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6858 sc.may_swap = !nr_boost_reclaim;
6859
6860 /*
6861 * Do some background aging, to give pages a chance to be
6862 * referenced before reclaiming. All pages are rotated
6863 * regardless of classzone as this is about consistent aging.
6864 */
6865 kswapd_age_node(pgdat, &sc);
6866
6867 /*
6868 * If we're getting trouble reclaiming, start doing writepage
6869 * even in laptop mode.
6870 */
6871 if (sc.priority < DEF_PRIORITY - 2)
6872 sc.may_writepage = 1;
6873
6874 /* Call soft limit reclaim before calling shrink_node. */
6875 sc.nr_scanned = 0;
6876 nr_soft_scanned = 0;
6877 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
6878 sc.gfp_mask, &nr_soft_scanned);
6879 sc.nr_reclaimed += nr_soft_reclaimed;
6880
6881 /*
6882 * There should be no need to raise the scanning priority if
6883 * enough pages are already being scanned that that high
6884 * watermark would be met at 100% efficiency.
6885 */
6886 if (kswapd_shrink_node(pgdat, &sc))
6887 raise_priority = false;
6888
6889 /*
6890 * If the low watermark is met there is no need for processes
6891 * to be throttled on pfmemalloc_wait as they should not be
6892 * able to safely make forward progress. Wake them
6893 */
6894 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6895 allow_direct_reclaim(pgdat))
6896 wake_up_all(&pgdat->pfmemalloc_wait);
6897
6898 /* Check if kswapd should be suspending */
6899 __fs_reclaim_release(_THIS_IP_);
6900 ret = try_to_freeze();
6901 __fs_reclaim_acquire(_THIS_IP_);
6902 if (ret || kthread_should_stop())
6903 break;
6904
6905 /*
6906 * Raise priority if scanning rate is too low or there was no
6907 * progress in reclaiming pages
6908 */
6909 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6910 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6911
6912 /*
6913 * If reclaim made no progress for a boost, stop reclaim as
6914 * IO cannot be queued and it could be an infinite loop in
6915 * extreme circumstances.
6916 */
6917 if (nr_boost_reclaim && !nr_reclaimed)
6918 break;
6919
6920 if (raise_priority || !nr_reclaimed)
6921 sc.priority--;
6922 } while (sc.priority >= 1);
6923
6924 if (!sc.nr_reclaimed)
6925 pgdat->kswapd_failures++;
6926
6927out:
6928 clear_reclaim_active(pgdat, highest_zoneidx);
6929
6930 /* If reclaim was boosted, account for the reclaim done in this pass */
6931 if (boosted) {
6932 unsigned long flags;
6933
6934 for (i = 0; i <= highest_zoneidx; i++) {
6935 if (!zone_boosts[i])
6936 continue;
6937
6938 /* Increments are under the zone lock */
6939 zone = pgdat->node_zones + i;
6940 spin_lock_irqsave(&zone->lock, flags);
6941 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6942 spin_unlock_irqrestore(&zone->lock, flags);
6943 }
6944
6945 /*
6946 * As there is now likely space, wakeup kcompact to defragment
6947 * pageblocks.
6948 */
6949 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
6950 }
6951
6952 snapshot_refaults(NULL, pgdat);
6953 __fs_reclaim_release(_THIS_IP_);
6954 psi_memstall_leave(&pflags);
6955 set_task_reclaim_state(current, NULL);
6956
6957 /*
6958 * Return the order kswapd stopped reclaiming at as
6959 * prepare_kswapd_sleep() takes it into account. If another caller
6960 * entered the allocator slow path while kswapd was awake, order will
6961 * remain at the higher level.
6962 */
6963 return sc.order;
6964}
6965
6966/*
6967 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6968 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6969 * not a valid index then either kswapd runs for first time or kswapd couldn't
6970 * sleep after previous reclaim attempt (node is still unbalanced). In that
6971 * case return the zone index of the previous kswapd reclaim cycle.
6972 */
6973static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6974 enum zone_type prev_highest_zoneidx)
6975{
6976 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
6977
6978 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
6979}
6980
6981static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
6982 unsigned int highest_zoneidx)
6983{
6984 long remaining = 0;
6985 DEFINE_WAIT(wait);
6986
6987 if (freezing(current) || kthread_should_stop())
6988 return;
6989
6990 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
6991
6992 /*
6993 * Try to sleep for a short interval. Note that kcompactd will only be
6994 * woken if it is possible to sleep for a short interval. This is
6995 * deliberate on the assumption that if reclaim cannot keep an
6996 * eligible zone balanced that it's also unlikely that compaction will
6997 * succeed.
6998 */
6999 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7000 /*
7001 * Compaction records what page blocks it recently failed to
7002 * isolate pages from and skips them in the future scanning.
7003 * When kswapd is going to sleep, it is reasonable to assume
7004 * that pages and compaction may succeed so reset the cache.
7005 */
7006 reset_isolation_suitable(pgdat);
7007
7008 /*
7009 * We have freed the memory, now we should compact it to make
7010 * allocation of the requested order possible.
7011 */
7012 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7013
7014 remaining = schedule_timeout(HZ/10);
7015
7016 /*
7017 * If woken prematurely then reset kswapd_highest_zoneidx and
7018 * order. The values will either be from a wakeup request or
7019 * the previous request that slept prematurely.
7020 */
7021 if (remaining) {
7022 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7023 kswapd_highest_zoneidx(pgdat,
7024 highest_zoneidx));
7025
7026 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7027 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7028 }
7029
7030 finish_wait(&pgdat->kswapd_wait, &wait);
7031 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7032 }
7033
7034 /*
7035 * After a short sleep, check if it was a premature sleep. If not, then
7036 * go fully to sleep until explicitly woken up.
7037 */
7038 if (!remaining &&
7039 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7040 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7041
7042 /*
7043 * vmstat counters are not perfectly accurate and the estimated
7044 * value for counters such as NR_FREE_PAGES can deviate from the
7045 * true value by nr_online_cpus * threshold. To avoid the zone
7046 * watermarks being breached while under pressure, we reduce the
7047 * per-cpu vmstat threshold while kswapd is awake and restore
7048 * them before going back to sleep.
7049 */
7050 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7051
7052 if (!kthread_should_stop())
7053 schedule();
7054
7055 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7056 } else {
7057 if (remaining)
7058 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7059 else
7060 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7061 }
7062 finish_wait(&pgdat->kswapd_wait, &wait);
7063}
7064
7065/*
7066 * The background pageout daemon, started as a kernel thread
7067 * from the init process.
7068 *
7069 * This basically trickles out pages so that we have _some_
7070 * free memory available even if there is no other activity
7071 * that frees anything up. This is needed for things like routing
7072 * etc, where we otherwise might have all activity going on in
7073 * asynchronous contexts that cannot page things out.
7074 *
7075 * If there are applications that are active memory-allocators
7076 * (most normal use), this basically shouldn't matter.
7077 */
7078static int kswapd(void *p)
7079{
7080 unsigned int alloc_order, reclaim_order;
7081 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7082 pg_data_t *pgdat = (pg_data_t *)p;
7083 struct task_struct *tsk = current;
7084 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7085
7086 if (!cpumask_empty(cpumask))
7087 set_cpus_allowed_ptr(tsk, cpumask);
7088
7089 /*
7090 * Tell the memory management that we're a "memory allocator",
7091 * and that if we need more memory we should get access to it
7092 * regardless (see "__alloc_pages()"). "kswapd" should
7093 * never get caught in the normal page freeing logic.
7094 *
7095 * (Kswapd normally doesn't need memory anyway, but sometimes
7096 * you need a small amount of memory in order to be able to
7097 * page out something else, and this flag essentially protects
7098 * us from recursively trying to free more memory as we're
7099 * trying to free the first piece of memory in the first place).
7100 */
7101 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7102 set_freezable();
7103
7104 WRITE_ONCE(pgdat->kswapd_order, 0);
7105 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7106 atomic_set(&pgdat->nr_writeback_throttled, 0);
7107 for ( ; ; ) {
7108 bool ret;
7109
7110 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7111 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7112 highest_zoneidx);
7113
7114kswapd_try_sleep:
7115 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7116 highest_zoneidx);
7117
7118 /* Read the new order and highest_zoneidx */
7119 alloc_order = READ_ONCE(pgdat->kswapd_order);
7120 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7121 highest_zoneidx);
7122 WRITE_ONCE(pgdat->kswapd_order, 0);
7123 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7124
7125 ret = try_to_freeze();
7126 if (kthread_should_stop())
7127 break;
7128
7129 /*
7130 * We can speed up thawing tasks if we don't call balance_pgdat
7131 * after returning from the refrigerator
7132 */
7133 if (ret)
7134 continue;
7135
7136 /*
7137 * Reclaim begins at the requested order but if a high-order
7138 * reclaim fails then kswapd falls back to reclaiming for
7139 * order-0. If that happens, kswapd will consider sleeping
7140 * for the order it finished reclaiming at (reclaim_order)
7141 * but kcompactd is woken to compact for the original
7142 * request (alloc_order).
7143 */
7144 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7145 alloc_order);
7146 reclaim_order = balance_pgdat(pgdat, alloc_order,
7147 highest_zoneidx);
7148 if (reclaim_order < alloc_order)
7149 goto kswapd_try_sleep;
7150 }
7151
7152 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7153
7154 return 0;
7155}
7156
7157/*
7158 * A zone is low on free memory or too fragmented for high-order memory. If
7159 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7160 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7161 * has failed or is not needed, still wake up kcompactd if only compaction is
7162 * needed.
7163 */
7164void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7165 enum zone_type highest_zoneidx)
7166{
7167 pg_data_t *pgdat;
7168 enum zone_type curr_idx;
7169
7170 if (!managed_zone(zone))
7171 return;
7172
7173 if (!cpuset_zone_allowed(zone, gfp_flags))
7174 return;
7175
7176 pgdat = zone->zone_pgdat;
7177 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7178
7179 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7180 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7181
7182 if (READ_ONCE(pgdat->kswapd_order) < order)
7183 WRITE_ONCE(pgdat->kswapd_order, order);
7184
7185 if (!waitqueue_active(&pgdat->kswapd_wait))
7186 return;
7187
7188 /* Hopeless node, leave it to direct reclaim if possible */
7189 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7190 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7191 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7192 /*
7193 * There may be plenty of free memory available, but it's too
7194 * fragmented for high-order allocations. Wake up kcompactd
7195 * and rely on compaction_suitable() to determine if it's
7196 * needed. If it fails, it will defer subsequent attempts to
7197 * ratelimit its work.
7198 */
7199 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7200 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7201 return;
7202 }
7203
7204 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7205 gfp_flags);
7206 wake_up_interruptible(&pgdat->kswapd_wait);
7207}
7208
7209#ifdef CONFIG_HIBERNATION
7210/*
7211 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7212 * freed pages.
7213 *
7214 * Rather than trying to age LRUs the aim is to preserve the overall
7215 * LRU order by reclaiming preferentially
7216 * inactive > active > active referenced > active mapped
7217 */
7218unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7219{
7220 struct scan_control sc = {
7221 .nr_to_reclaim = nr_to_reclaim,
7222 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7223 .reclaim_idx = MAX_NR_ZONES - 1,
7224 .priority = DEF_PRIORITY,
7225 .may_writepage = 1,
7226 .may_unmap = 1,
7227 .may_swap = 1,
7228 .hibernation_mode = 1,
7229 };
7230 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7231 unsigned long nr_reclaimed;
7232 unsigned int noreclaim_flag;
7233
7234 fs_reclaim_acquire(sc.gfp_mask);
7235 noreclaim_flag = memalloc_noreclaim_save();
7236 set_task_reclaim_state(current, &sc.reclaim_state);
7237
7238 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7239
7240 set_task_reclaim_state(current, NULL);
7241 memalloc_noreclaim_restore(noreclaim_flag);
7242 fs_reclaim_release(sc.gfp_mask);
7243
7244 return nr_reclaimed;
7245}
7246#endif /* CONFIG_HIBERNATION */
7247
7248/*
7249 * This kswapd start function will be called by init and node-hot-add.
7250 */
7251void __meminit kswapd_run(int nid)
7252{
7253 pg_data_t *pgdat = NODE_DATA(nid);
7254
7255 pgdat_kswapd_lock(pgdat);
7256 if (!pgdat->kswapd) {
7257 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7258 if (IS_ERR(pgdat->kswapd)) {
7259 /* failure at boot is fatal */
7260 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7261 nid, PTR_ERR(pgdat->kswapd));
7262 BUG_ON(system_state < SYSTEM_RUNNING);
7263 pgdat->kswapd = NULL;
7264 }
7265 }
7266 pgdat_kswapd_unlock(pgdat);
7267}
7268
7269/*
7270 * Called by memory hotplug when all memory in a node is offlined. Caller must
7271 * be holding mem_hotplug_begin/done().
7272 */
7273void __meminit kswapd_stop(int nid)
7274{
7275 pg_data_t *pgdat = NODE_DATA(nid);
7276 struct task_struct *kswapd;
7277
7278 pgdat_kswapd_lock(pgdat);
7279 kswapd = pgdat->kswapd;
7280 if (kswapd) {
7281 kthread_stop(kswapd);
7282 pgdat->kswapd = NULL;
7283 }
7284 pgdat_kswapd_unlock(pgdat);
7285}
7286
7287static int __init kswapd_init(void)
7288{
7289 int nid;
7290
7291 swap_setup();
7292 for_each_node_state(nid, N_MEMORY)
7293 kswapd_run(nid);
7294 return 0;
7295}
7296
7297module_init(kswapd_init)
7298
7299#ifdef CONFIG_NUMA
7300/*
7301 * Node reclaim mode
7302 *
7303 * If non-zero call node_reclaim when the number of free pages falls below
7304 * the watermarks.
7305 */
7306int node_reclaim_mode __read_mostly;
7307
7308/*
7309 * Priority for NODE_RECLAIM. This determines the fraction of pages
7310 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7311 * a zone.
7312 */
7313#define NODE_RECLAIM_PRIORITY 4
7314
7315/*
7316 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7317 * occur.
7318 */
7319int sysctl_min_unmapped_ratio = 1;
7320
7321/*
7322 * If the number of slab pages in a zone grows beyond this percentage then
7323 * slab reclaim needs to occur.
7324 */
7325int sysctl_min_slab_ratio = 5;
7326
7327static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7328{
7329 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7330 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7331 node_page_state(pgdat, NR_ACTIVE_FILE);
7332
7333 /*
7334 * It's possible for there to be more file mapped pages than
7335 * accounted for by the pages on the file LRU lists because
7336 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7337 */
7338 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7339}
7340
7341/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7342static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7343{
7344 unsigned long nr_pagecache_reclaimable;
7345 unsigned long delta = 0;
7346
7347 /*
7348 * If RECLAIM_UNMAP is set, then all file pages are considered
7349 * potentially reclaimable. Otherwise, we have to worry about
7350 * pages like swapcache and node_unmapped_file_pages() provides
7351 * a better estimate
7352 */
7353 if (node_reclaim_mode & RECLAIM_UNMAP)
7354 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7355 else
7356 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7357
7358 /* If we can't clean pages, remove dirty pages from consideration */
7359 if (!(node_reclaim_mode & RECLAIM_WRITE))
7360 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7361
7362 /* Watch for any possible underflows due to delta */
7363 if (unlikely(delta > nr_pagecache_reclaimable))
7364 delta = nr_pagecache_reclaimable;
7365
7366 return nr_pagecache_reclaimable - delta;
7367}
7368
7369/*
7370 * Try to free up some pages from this node through reclaim.
7371 */
7372static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7373{
7374 /* Minimum pages needed in order to stay on node */
7375 const unsigned long nr_pages = 1 << order;
7376 struct task_struct *p = current;
7377 unsigned int noreclaim_flag;
7378 struct scan_control sc = {
7379 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7380 .gfp_mask = current_gfp_context(gfp_mask),
7381 .order = order,
7382 .priority = NODE_RECLAIM_PRIORITY,
7383 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7384 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7385 .may_swap = 1,
7386 .reclaim_idx = gfp_zone(gfp_mask),
7387 };
7388 unsigned long pflags;
7389
7390 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7391 sc.gfp_mask);
7392
7393 cond_resched();
7394 psi_memstall_enter(&pflags);
7395 delayacct_freepages_start();
7396 fs_reclaim_acquire(sc.gfp_mask);
7397 /*
7398 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7399 */
7400 noreclaim_flag = memalloc_noreclaim_save();
7401 set_task_reclaim_state(p, &sc.reclaim_state);
7402
7403 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7404 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7405 /*
7406 * Free memory by calling shrink node with increasing
7407 * priorities until we have enough memory freed.
7408 */
7409 do {
7410 shrink_node(pgdat, &sc);
7411 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7412 }
7413
7414 set_task_reclaim_state(p, NULL);
7415 memalloc_noreclaim_restore(noreclaim_flag);
7416 fs_reclaim_release(sc.gfp_mask);
7417 psi_memstall_leave(&pflags);
7418 delayacct_freepages_end();
7419
7420 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7421
7422 return sc.nr_reclaimed >= nr_pages;
7423}
7424
7425int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7426{
7427 int ret;
7428
7429 /*
7430 * Node reclaim reclaims unmapped file backed pages and
7431 * slab pages if we are over the defined limits.
7432 *
7433 * A small portion of unmapped file backed pages is needed for
7434 * file I/O otherwise pages read by file I/O will be immediately
7435 * thrown out if the node is overallocated. So we do not reclaim
7436 * if less than a specified percentage of the node is used by
7437 * unmapped file backed pages.
7438 */
7439 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7440 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7441 pgdat->min_slab_pages)
7442 return NODE_RECLAIM_FULL;
7443
7444 /*
7445 * Do not scan if the allocation should not be delayed.
7446 */
7447 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7448 return NODE_RECLAIM_NOSCAN;
7449
7450 /*
7451 * Only run node reclaim on the local node or on nodes that do not
7452 * have associated processors. This will favor the local processor
7453 * over remote processors and spread off node memory allocations
7454 * as wide as possible.
7455 */
7456 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7457 return NODE_RECLAIM_NOSCAN;
7458
7459 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7460 return NODE_RECLAIM_NOSCAN;
7461
7462 ret = __node_reclaim(pgdat, gfp_mask, order);
7463 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7464
7465 if (!ret)
7466 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7467
7468 return ret;
7469}
7470#endif
7471
7472/**
7473 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7474 * lru list
7475 * @fbatch: Batch of lru folios to check.
7476 *
7477 * Checks folios for evictability, if an evictable folio is in the unevictable
7478 * lru list, moves it to the appropriate evictable lru list. This function
7479 * should be only used for lru folios.
7480 */
7481void check_move_unevictable_folios(struct folio_batch *fbatch)
7482{
7483 struct lruvec *lruvec = NULL;
7484 int pgscanned = 0;
7485 int pgrescued = 0;
7486 int i;
7487
7488 for (i = 0; i < fbatch->nr; i++) {
7489 struct folio *folio = fbatch->folios[i];
7490 int nr_pages = folio_nr_pages(folio);
7491
7492 pgscanned += nr_pages;
7493
7494 /* block memcg migration while the folio moves between lrus */
7495 if (!folio_test_clear_lru(folio))
7496 continue;
7497
7498 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7499 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7500 lruvec_del_folio(lruvec, folio);
7501 folio_clear_unevictable(folio);
7502 lruvec_add_folio(lruvec, folio);
7503 pgrescued += nr_pages;
7504 }
7505 folio_set_lru(folio);
7506 }
7507
7508 if (lruvec) {
7509 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7510 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7511 unlock_page_lruvec_irq(lruvec);
7512 } else if (pgscanned) {
7513 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7514 }
7515}
7516EXPORT_SYMBOL_GPL(check_move_unevictable_folios);