1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
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.
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
60 #include <asm/tlbflush.h>
61 #include <asm/div64.h>
63 #include <linux/swapops.h>
64 #include <linux/balloon_compaction.h>
65 #include <linux/sched/sysctl.h>
70 #define CREATE_TRACE_POINTS
71 #include <trace/events/vmscan.h>
74 /* How many pages shrink_list() should reclaim */
75 unsigned long nr_to_reclaim;
78 * Nodemask of nodes allowed by the caller. If NULL, all nodes
84 * The memory cgroup that hit its limit and as a result is the
85 * primary target of this reclaim invocation.
87 struct mem_cgroup *target_mem_cgroup;
90 * Scan pressure balancing between anon and file LRUs
92 unsigned long anon_cost;
93 unsigned long file_cost;
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;
102 /* Writepage batching in laptop mode; RECLAIM_WRITE */
103 unsigned int may_writepage:1;
105 /* Can mapped folios be reclaimed? */
106 unsigned int may_unmap:1;
108 /* Can folios be swapped as part of reclaim? */
109 unsigned int may_swap:1;
111 /* Not allow cache_trim_mode to be turned on as part of reclaim? */
112 unsigned int no_cache_trim_mode:1;
114 /* Has cache_trim_mode failed at least once? */
115 unsigned int cache_trim_mode_failed:1;
117 /* Proactive reclaim invoked by userspace through memory.reclaim */
118 unsigned int proactive:1;
121 * Cgroup memory below memory.low is protected as long as we
122 * don't threaten to OOM. If any cgroup is reclaimed at
123 * reduced force or passed over entirely due to its memory.low
124 * setting (memcg_low_skipped), and nothing is reclaimed as a
125 * result, then go back for one more cycle that reclaims the protected
126 * memory (memcg_low_reclaim) to avert OOM.
128 unsigned int memcg_low_reclaim:1;
129 unsigned int memcg_low_skipped:1;
131 unsigned int hibernation_mode:1;
133 /* One of the zones is ready for compaction */
134 unsigned int compaction_ready:1;
136 /* There is easily reclaimable cold cache in the current node */
137 unsigned int cache_trim_mode:1;
139 /* The file folios on the current node are dangerously low */
140 unsigned int file_is_tiny:1;
142 /* Always discard instead of demoting to lower tier memory */
143 unsigned int no_demotion:1;
145 /* Allocation order */
148 /* Scan (total_size >> priority) pages at once */
151 /* The highest zone to isolate folios for reclaim from */
154 /* This context's GFP mask */
157 /* Incremented by the number of inactive pages that were scanned */
158 unsigned long nr_scanned;
160 /* Number of pages freed so far during a call to shrink_zones() */
161 unsigned long nr_reclaimed;
165 unsigned int unqueued_dirty;
166 unsigned int congested;
167 unsigned int writeback;
168 unsigned int immediate;
169 unsigned int file_taken;
173 /* for recording the reclaimed slab by now */
174 struct reclaim_state reclaim_state;
177 #ifdef ARCH_HAS_PREFETCHW
178 #define prefetchw_prev_lru_folio(_folio, _base, _field) \
180 if ((_folio)->lru.prev != _base) { \
181 struct folio *prev; \
183 prev = lru_to_folio(&(_folio->lru)); \
184 prefetchw(&prev->_field); \
188 #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
192 * From 0 .. 200. Higher means more swappy.
194 int vm_swappiness = 60;
198 /* Returns true for reclaim through cgroup limits or cgroup interfaces. */
199 static bool cgroup_reclaim(struct scan_control *sc)
201 return sc->target_mem_cgroup;
205 * Returns true for reclaim on the root cgroup. This is true for direct
206 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
208 static bool root_reclaim(struct scan_control *sc)
210 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
214 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
215 * @sc: scan_control in question
217 * The normal page dirty throttling mechanism in balance_dirty_pages() is
218 * completely broken with the legacy memcg and direct stalling in
219 * shrink_folio_list() is used for throttling instead, which lacks all the
220 * niceties such as fairness, adaptive pausing, bandwidth proportional
221 * allocation and configurability.
223 * This function tests whether the vmscan currently in progress can assume
224 * that the normal dirty throttling mechanism is operational.
226 static bool writeback_throttling_sane(struct scan_control *sc)
228 if (!cgroup_reclaim(sc))
230 #ifdef CONFIG_CGROUP_WRITEBACK
231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
237 static bool cgroup_reclaim(struct scan_control *sc)
242 static bool root_reclaim(struct scan_control *sc)
247 static bool writeback_throttling_sane(struct scan_control *sc)
253 static void set_task_reclaim_state(struct task_struct *task,
254 struct reclaim_state *rs)
256 /* Check for an overwrite */
257 WARN_ON_ONCE(rs && task->reclaim_state);
259 /* Check for the nulling of an already-nulled member */
260 WARN_ON_ONCE(!rs && !task->reclaim_state);
262 task->reclaim_state = rs;
266 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
267 * scan_control->nr_reclaimed.
269 static void flush_reclaim_state(struct scan_control *sc)
272 * Currently, reclaim_state->reclaimed includes three types of pages
273 * freed outside of vmscan:
275 * (2) Clean file pages from pruned inodes (on highmem systems).
276 * (3) XFS freed buffer pages.
278 * For all of these cases, we cannot universally link the pages to a
279 * single memcg. For example, a memcg-aware shrinker can free one object
280 * charged to the target memcg, causing an entire page to be freed.
281 * If we count the entire page as reclaimed from the memcg, we end up
282 * overestimating the reclaimed amount (potentially under-reclaiming).
284 * Only count such pages for global reclaim to prevent under-reclaiming
285 * from the target memcg; preventing unnecessary retries during memcg
286 * charging and false positives from proactive reclaim.
288 * For uncommon cases where the freed pages were actually mostly
289 * charged to the target memcg, we end up underestimating the reclaimed
290 * amount. This should be fine. The freed pages will be uncharged
291 * anyway, even if they are not counted here properly, and we will be
292 * able to make forward progress in charging (which is usually in a
295 * We can go one step further, and report the uncharged objcg pages in
296 * memcg reclaim, to make reporting more accurate and reduce
297 * underestimation, but it's probably not worth the complexity for now.
299 if (current->reclaim_state && root_reclaim(sc)) {
300 sc->nr_reclaimed += current->reclaim_state->reclaimed;
301 current->reclaim_state->reclaimed = 0;
305 static bool can_demote(int nid, struct scan_control *sc)
307 if (!numa_demotion_enabled)
309 if (sc && sc->no_demotion)
311 if (next_demotion_node(nid) == NUMA_NO_NODE)
317 static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
319 struct scan_control *sc)
323 * For non-memcg reclaim, is there
324 * space in any swap device?
326 if (get_nr_swap_pages() > 0)
329 /* Is the memcg below its swap limit? */
330 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
335 * The page can not be swapped.
337 * Can it be reclaimed from this node via demotion?
339 return can_demote(nid, sc);
343 * This misses isolated folios which are not accounted for to save counters.
344 * As the data only determines if reclaim or compaction continues, it is
345 * not expected that isolated folios will be a dominating factor.
347 unsigned long zone_reclaimable_pages(struct zone *zone)
351 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
352 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
353 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
354 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
355 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
361 * lruvec_lru_size - Returns the number of pages on the given LRU list.
362 * @lruvec: lru vector
364 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
366 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
369 unsigned long size = 0;
372 for (zid = 0; zid <= zone_idx; zid++) {
373 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
375 if (!managed_zone(zone))
378 if (!mem_cgroup_disabled())
379 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
381 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
386 static unsigned long drop_slab_node(int nid)
388 unsigned long freed = 0;
389 struct mem_cgroup *memcg = NULL;
391 memcg = mem_cgroup_iter(NULL, NULL, NULL);
393 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
394 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
407 for_each_online_node(nid) {
408 if (fatal_signal_pending(current))
411 freed += drop_slab_node(nid);
413 } while ((freed >> shift++) > 1);
416 static int reclaimer_offset(void)
418 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
419 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
420 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
421 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
422 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
423 PGSCAN_DIRECT - PGSCAN_KSWAPD);
424 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
425 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
427 if (current_is_kswapd())
429 if (current_is_khugepaged())
430 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
431 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
434 static inline int is_page_cache_freeable(struct folio *folio)
437 * A freeable page cache folio is referenced only by the caller
438 * that isolated the folio, the page cache and optional filesystem
439 * private data at folio->private.
441 return folio_ref_count(folio) - folio_test_private(folio) ==
442 1 + folio_nr_pages(folio);
446 * We detected a synchronous write error writing a folio out. Probably
447 * -ENOSPC. We need to propagate that into the address_space for a subsequent
448 * fsync(), msync() or close().
450 * The tricky part is that after writepage we cannot touch the mapping: nothing
451 * prevents it from being freed up. But we have a ref on the folio and once
452 * that folio is locked, the mapping is pinned.
454 * We're allowed to run sleeping folio_lock() here because we know the caller has
457 static void handle_write_error(struct address_space *mapping,
458 struct folio *folio, int error)
461 if (folio_mapping(folio) == mapping)
462 mapping_set_error(mapping, error);
466 static bool skip_throttle_noprogress(pg_data_t *pgdat)
468 int reclaimable = 0, write_pending = 0;
472 * If kswapd is disabled, reschedule if necessary but do not
473 * throttle as the system is likely near OOM.
475 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
479 * If there are a lot of dirty/writeback folios then do not
480 * throttle as throttling will occur when the folios cycle
481 * towards the end of the LRU if still under writeback.
483 for (i = 0; i < MAX_NR_ZONES; i++) {
484 struct zone *zone = pgdat->node_zones + i;
486 if (!managed_zone(zone))
489 reclaimable += zone_reclaimable_pages(zone);
490 write_pending += zone_page_state_snapshot(zone,
491 NR_ZONE_WRITE_PENDING);
493 if (2 * write_pending <= reclaimable)
499 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
501 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
506 * Do not throttle user workers, kthreads other than kswapd or
507 * workqueues. They may be required for reclaim to make
508 * forward progress (e.g. journalling workqueues or kthreads).
510 if (!current_is_kswapd() &&
511 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
517 * These figures are pulled out of thin air.
518 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
519 * parallel reclaimers which is a short-lived event so the timeout is
520 * short. Failing to make progress or waiting on writeback are
521 * potentially long-lived events so use a longer timeout. This is shaky
522 * logic as a failure to make progress could be due to anything from
523 * writeback to a slow device to excessive referenced folios at the tail
524 * of the inactive LRU.
527 case VMSCAN_THROTTLE_WRITEBACK:
530 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
531 WRITE_ONCE(pgdat->nr_reclaim_start,
532 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
536 case VMSCAN_THROTTLE_CONGESTED:
538 case VMSCAN_THROTTLE_NOPROGRESS:
539 if (skip_throttle_noprogress(pgdat)) {
547 case VMSCAN_THROTTLE_ISOLATED:
556 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
557 ret = schedule_timeout(timeout);
558 finish_wait(wqh, &wait);
560 if (reason == VMSCAN_THROTTLE_WRITEBACK)
561 atomic_dec(&pgdat->nr_writeback_throttled);
563 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
564 jiffies_to_usecs(timeout - ret),
569 * Account for folios written if tasks are throttled waiting on dirty
570 * folios to clean. If enough folios have been cleaned since throttling
571 * started then wakeup the throttled tasks.
573 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
576 unsigned long nr_written;
578 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
581 * This is an inaccurate read as the per-cpu deltas may not
582 * be synchronised. However, given that the system is
583 * writeback throttled, it is not worth taking the penalty
584 * of getting an accurate count. At worst, the throttle
585 * timeout guarantees forward progress.
587 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
588 READ_ONCE(pgdat->nr_reclaim_start);
590 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
591 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
594 /* possible outcome of pageout() */
596 /* failed to write folio out, folio is locked */
598 /* move folio to the active list, folio is locked */
600 /* folio has been sent to the disk successfully, folio is unlocked */
602 /* folio is clean and locked */
607 * pageout is called by shrink_folio_list() for each dirty folio.
608 * Calls ->writepage().
610 static pageout_t pageout(struct folio *folio, struct address_space *mapping,
611 struct swap_iocb **plug)
614 * If the folio is dirty, only perform writeback if that write
615 * will be non-blocking. To prevent this allocation from being
616 * stalled by pagecache activity. But note that there may be
617 * stalls if we need to run get_block(). We could test
618 * PagePrivate for that.
620 * If this process is currently in __generic_file_write_iter() against
621 * this folio's queue, we can perform writeback even if that
624 * If the folio is swapcache, write it back even if that would
625 * block, for some throttling. This happens by accident, because
626 * swap_backing_dev_info is bust: it doesn't reflect the
627 * congestion state of the swapdevs. Easy to fix, if needed.
629 if (!is_page_cache_freeable(folio))
633 * Some data journaling orphaned folios can have
634 * folio->mapping == NULL while being dirty with clean buffers.
636 if (folio_test_private(folio)) {
637 if (try_to_free_buffers(folio)) {
638 folio_clear_dirty(folio);
639 pr_info("%s: orphaned folio\n", __func__);
645 if (mapping->a_ops->writepage == NULL)
646 return PAGE_ACTIVATE;
648 if (folio_clear_dirty_for_io(folio)) {
650 struct writeback_control wbc = {
651 .sync_mode = WB_SYNC_NONE,
652 .nr_to_write = SWAP_CLUSTER_MAX,
654 .range_end = LLONG_MAX,
659 folio_set_reclaim(folio);
660 res = mapping->a_ops->writepage(&folio->page, &wbc);
662 handle_write_error(mapping, folio, res);
663 if (res == AOP_WRITEPAGE_ACTIVATE) {
664 folio_clear_reclaim(folio);
665 return PAGE_ACTIVATE;
668 if (!folio_test_writeback(folio)) {
669 /* synchronous write or broken a_ops? */
670 folio_clear_reclaim(folio);
672 trace_mm_vmscan_write_folio(folio);
673 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
681 * Same as remove_mapping, but if the folio is removed from the mapping, it
682 * gets returned with a refcount of 0.
684 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
685 bool reclaimed, struct mem_cgroup *target_memcg)
690 BUG_ON(!folio_test_locked(folio));
691 BUG_ON(mapping != folio_mapping(folio));
693 if (!folio_test_swapcache(folio))
694 spin_lock(&mapping->host->i_lock);
695 xa_lock_irq(&mapping->i_pages);
697 * The non racy check for a busy folio.
699 * Must be careful with the order of the tests. When someone has
700 * a ref to the folio, it may be possible that they dirty it then
701 * drop the reference. So if the dirty flag is tested before the
702 * refcount here, then the following race may occur:
704 * get_user_pages(&page);
705 * [user mapping goes away]
707 * !folio_test_dirty(folio) [good]
708 * folio_set_dirty(folio);
710 * !refcount(folio) [good, discard it]
712 * [oops, our write_to data is lost]
714 * Reversing the order of the tests ensures such a situation cannot
715 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
716 * load is not satisfied before that of folio->_refcount.
718 * Note that if the dirty flag is always set via folio_mark_dirty,
719 * and thus under the i_pages lock, then this ordering is not required.
721 refcount = 1 + folio_nr_pages(folio);
722 if (!folio_ref_freeze(folio, refcount))
724 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
725 if (unlikely(folio_test_dirty(folio))) {
726 folio_ref_unfreeze(folio, refcount);
730 if (folio_test_swapcache(folio)) {
731 swp_entry_t swap = folio->swap;
733 if (reclaimed && !mapping_exiting(mapping))
734 shadow = workingset_eviction(folio, target_memcg);
735 __delete_from_swap_cache(folio, swap, shadow);
736 mem_cgroup_swapout(folio, swap);
737 xa_unlock_irq(&mapping->i_pages);
738 put_swap_folio(folio, swap);
740 void (*free_folio)(struct folio *);
742 free_folio = mapping->a_ops->free_folio;
744 * Remember a shadow entry for reclaimed file cache in
745 * order to detect refaults, thus thrashing, later on.
747 * But don't store shadows in an address space that is
748 * already exiting. This is not just an optimization,
749 * inode reclaim needs to empty out the radix tree or
750 * the nodes are lost. Don't plant shadows behind its
753 * We also don't store shadows for DAX mappings because the
754 * only page cache folios found in these are zero pages
755 * covering holes, and because we don't want to mix DAX
756 * exceptional entries and shadow exceptional entries in the
757 * same address_space.
759 if (reclaimed && folio_is_file_lru(folio) &&
760 !mapping_exiting(mapping) && !dax_mapping(mapping))
761 shadow = workingset_eviction(folio, target_memcg);
762 __filemap_remove_folio(folio, shadow);
763 xa_unlock_irq(&mapping->i_pages);
764 if (mapping_shrinkable(mapping))
765 inode_add_lru(mapping->host);
766 spin_unlock(&mapping->host->i_lock);
775 xa_unlock_irq(&mapping->i_pages);
776 if (!folio_test_swapcache(folio))
777 spin_unlock(&mapping->host->i_lock);
782 * remove_mapping() - Attempt to remove a folio from its mapping.
783 * @mapping: The address space.
784 * @folio: The folio to remove.
786 * If the folio is dirty, under writeback or if someone else has a ref
787 * on it, removal will fail.
788 * Return: The number of pages removed from the mapping. 0 if the folio
789 * could not be removed.
790 * Context: The caller should have a single refcount on the folio and
793 long remove_mapping(struct address_space *mapping, struct folio *folio)
795 if (__remove_mapping(mapping, folio, false, NULL)) {
797 * Unfreezing the refcount with 1 effectively
798 * drops the pagecache ref for us without requiring another
801 folio_ref_unfreeze(folio, 1);
802 return folio_nr_pages(folio);
808 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
809 * @folio: Folio to be returned to an LRU list.
811 * Add previously isolated @folio to appropriate LRU list.
812 * The folio may still be unevictable for other reasons.
814 * Context: lru_lock must not be held, interrupts must be enabled.
816 void folio_putback_lru(struct folio *folio)
818 folio_add_lru(folio);
819 folio_put(folio); /* drop ref from isolate */
822 enum folio_references {
824 FOLIOREF_RECLAIM_CLEAN,
829 static enum folio_references folio_check_references(struct folio *folio,
830 struct scan_control *sc)
832 int referenced_ptes, referenced_folio;
833 unsigned long vm_flags;
835 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
837 referenced_folio = folio_test_clear_referenced(folio);
840 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
841 * Let the folio, now marked Mlocked, be moved to the unevictable list.
843 if (vm_flags & VM_LOCKED)
844 return FOLIOREF_ACTIVATE;
846 /* rmap lock contention: rotate */
847 if (referenced_ptes == -1)
848 return FOLIOREF_KEEP;
850 if (referenced_ptes) {
852 * All mapped folios start out with page table
853 * references from the instantiating fault, so we need
854 * to look twice if a mapped file/anon folio is used more
857 * Mark it and spare it for another trip around the
858 * inactive list. Another page table reference will
859 * lead to its activation.
861 * Note: the mark is set for activated folios as well
862 * so that recently deactivated but used folios are
865 folio_set_referenced(folio);
867 if (referenced_folio || referenced_ptes > 1)
868 return FOLIOREF_ACTIVATE;
871 * Activate file-backed executable folios after first usage.
873 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
874 return FOLIOREF_ACTIVATE;
876 return FOLIOREF_KEEP;
879 /* Reclaim if clean, defer dirty folios to writeback */
880 if (referenced_folio && folio_is_file_lru(folio))
881 return FOLIOREF_RECLAIM_CLEAN;
883 return FOLIOREF_RECLAIM;
886 /* Check if a folio is dirty or under writeback */
887 static void folio_check_dirty_writeback(struct folio *folio,
888 bool *dirty, bool *writeback)
890 struct address_space *mapping;
893 * Anonymous folios are not handled by flushers and must be written
894 * from reclaim context. Do not stall reclaim based on them.
895 * MADV_FREE anonymous folios are put into inactive file list too.
896 * They could be mistakenly treated as file lru. So further anon
899 if (!folio_is_file_lru(folio) ||
900 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
906 /* By default assume that the folio flags are accurate */
907 *dirty = folio_test_dirty(folio);
908 *writeback = folio_test_writeback(folio);
910 /* Verify dirty/writeback state if the filesystem supports it */
911 if (!folio_test_private(folio))
914 mapping = folio_mapping(folio);
915 if (mapping && mapping->a_ops->is_dirty_writeback)
916 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
919 static struct folio *alloc_demote_folio(struct folio *src,
920 unsigned long private)
923 nodemask_t *allowed_mask;
924 struct migration_target_control *mtc;
926 mtc = (struct migration_target_control *)private;
928 allowed_mask = mtc->nmask;
930 * make sure we allocate from the target node first also trying to
931 * demote or reclaim pages from the target node via kswapd if we are
932 * low on free memory on target node. If we don't do this and if
933 * we have free memory on the slower(lower) memtier, we would start
934 * allocating pages from slower(lower) memory tiers without even forcing
935 * a demotion of cold pages from the target memtier. This can result
936 * in the kernel placing hot pages in slower(lower) memory tiers.
939 mtc->gfp_mask |= __GFP_THISNODE;
940 dst = alloc_migration_target(src, (unsigned long)mtc);
944 mtc->gfp_mask &= ~__GFP_THISNODE;
945 mtc->nmask = allowed_mask;
947 return alloc_migration_target(src, (unsigned long)mtc);
951 * Take folios on @demote_folios and attempt to demote them to another node.
952 * Folios which are not demoted are left on @demote_folios.
954 static unsigned int demote_folio_list(struct list_head *demote_folios,
955 struct pglist_data *pgdat)
957 int target_nid = next_demotion_node(pgdat->node_id);
958 unsigned int nr_succeeded;
959 nodemask_t allowed_mask;
961 struct migration_target_control mtc = {
963 * Allocate from 'node', or fail quickly and quietly.
964 * When this happens, 'page' will likely just be discarded
965 * instead of migrated.
967 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
968 __GFP_NOMEMALLOC | GFP_NOWAIT,
970 .nmask = &allowed_mask
973 if (list_empty(demote_folios))
976 if (target_nid == NUMA_NO_NODE)
979 node_get_allowed_targets(pgdat, &allowed_mask);
981 /* Demotion ignores all cpuset and mempolicy settings */
982 migrate_pages(demote_folios, alloc_demote_folio, NULL,
983 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
986 mod_node_page_state(pgdat, PGDEMOTE_KSWAPD + reclaimer_offset(),
992 static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
994 if (gfp_mask & __GFP_FS)
996 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
999 * We can "enter_fs" for swap-cache with only __GFP_IO
1000 * providing this isn't SWP_FS_OPS.
1001 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1002 * but that will never affect SWP_FS_OPS, so the data_race
1005 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1009 * shrink_folio_list() returns the number of reclaimed pages
1011 static unsigned int shrink_folio_list(struct list_head *folio_list,
1012 struct pglist_data *pgdat, struct scan_control *sc,
1013 struct reclaim_stat *stat, bool ignore_references)
1015 struct folio_batch free_folios;
1016 LIST_HEAD(ret_folios);
1017 LIST_HEAD(demote_folios);
1018 unsigned int nr_reclaimed = 0;
1019 unsigned int pgactivate = 0;
1020 bool do_demote_pass;
1021 struct swap_iocb *plug = NULL;
1023 folio_batch_init(&free_folios);
1024 memset(stat, 0, sizeof(*stat));
1026 do_demote_pass = can_demote(pgdat->node_id, sc);
1029 while (!list_empty(folio_list)) {
1030 struct address_space *mapping;
1031 struct folio *folio;
1032 enum folio_references references = FOLIOREF_RECLAIM;
1033 bool dirty, writeback;
1034 unsigned int nr_pages;
1038 folio = lru_to_folio(folio_list);
1039 list_del(&folio->lru);
1041 if (!folio_trylock(folio))
1044 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1046 nr_pages = folio_nr_pages(folio);
1048 /* Account the number of base pages */
1049 sc->nr_scanned += nr_pages;
1051 if (unlikely(!folio_evictable(folio)))
1052 goto activate_locked;
1054 if (!sc->may_unmap && folio_mapped(folio))
1057 /* folio_update_gen() tried to promote this page? */
1058 if (lru_gen_enabled() && !ignore_references &&
1059 folio_mapped(folio) && folio_test_referenced(folio))
1063 * The number of dirty pages determines if a node is marked
1064 * reclaim_congested. kswapd will stall and start writing
1065 * folios if the tail of the LRU is all dirty unqueued folios.
1067 folio_check_dirty_writeback(folio, &dirty, &writeback);
1068 if (dirty || writeback)
1069 stat->nr_dirty += nr_pages;
1071 if (dirty && !writeback)
1072 stat->nr_unqueued_dirty += nr_pages;
1075 * Treat this folio as congested if folios are cycling
1076 * through the LRU so quickly that the folios marked
1077 * for immediate reclaim are making it to the end of
1078 * the LRU a second time.
1080 if (writeback && folio_test_reclaim(folio))
1081 stat->nr_congested += nr_pages;
1084 * If a folio at the tail of the LRU is under writeback, there
1085 * are three cases to consider.
1087 * 1) If reclaim is encountering an excessive number
1088 * of folios under writeback and this folio has both
1089 * the writeback and reclaim flags set, then it
1090 * indicates that folios are being queued for I/O but
1091 * are being recycled through the LRU before the I/O
1092 * can complete. Waiting on the folio itself risks an
1093 * indefinite stall if it is impossible to writeback
1094 * the folio due to I/O error or disconnected storage
1095 * so instead note that the LRU is being scanned too
1096 * quickly and the caller can stall after the folio
1097 * list has been processed.
1099 * 2) Global or new memcg reclaim encounters a folio that is
1100 * not marked for immediate reclaim, or the caller does not
1101 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1102 * not to fs). In this case mark the folio for immediate
1103 * reclaim and continue scanning.
1105 * Require may_enter_fs() because we would wait on fs, which
1106 * may not have submitted I/O yet. And the loop driver might
1107 * enter reclaim, and deadlock if it waits on a folio for
1108 * which it is needed to do the write (loop masks off
1109 * __GFP_IO|__GFP_FS for this reason); but more thought
1110 * would probably show more reasons.
1112 * 3) Legacy memcg encounters a folio that already has the
1113 * reclaim flag set. memcg does not have any dirty folio
1114 * throttling so we could easily OOM just because too many
1115 * folios are in writeback and there is nothing else to
1116 * reclaim. Wait for the writeback to complete.
1118 * In cases 1) and 2) we activate the folios to get them out of
1119 * the way while we continue scanning for clean folios on the
1120 * inactive list and refilling from the active list. The
1121 * observation here is that waiting for disk writes is more
1122 * expensive than potentially causing reloads down the line.
1123 * Since they're marked for immediate reclaim, they won't put
1124 * memory pressure on the cache working set any longer than it
1125 * takes to write them to disk.
1127 if (folio_test_writeback(folio)) {
1129 if (current_is_kswapd() &&
1130 folio_test_reclaim(folio) &&
1131 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1132 stat->nr_immediate += nr_pages;
1133 goto activate_locked;
1136 } else if (writeback_throttling_sane(sc) ||
1137 !folio_test_reclaim(folio) ||
1138 !may_enter_fs(folio, sc->gfp_mask)) {
1140 * This is slightly racy -
1141 * folio_end_writeback() might have
1142 * just cleared the reclaim flag, then
1143 * setting the reclaim flag here ends up
1144 * interpreted as the readahead flag - but
1145 * that does not matter enough to care.
1146 * What we do want is for this folio to
1147 * have the reclaim flag set next time
1148 * memcg reclaim reaches the tests above,
1149 * so it will then wait for writeback to
1150 * avoid OOM; and it's also appropriate
1151 * in global reclaim.
1153 folio_set_reclaim(folio);
1154 stat->nr_writeback += nr_pages;
1155 goto activate_locked;
1159 folio_unlock(folio);
1160 folio_wait_writeback(folio);
1161 /* then go back and try same folio again */
1162 list_add_tail(&folio->lru, folio_list);
1167 if (!ignore_references)
1168 references = folio_check_references(folio, sc);
1170 switch (references) {
1171 case FOLIOREF_ACTIVATE:
1172 goto activate_locked;
1174 stat->nr_ref_keep += nr_pages;
1176 case FOLIOREF_RECLAIM:
1177 case FOLIOREF_RECLAIM_CLEAN:
1178 ; /* try to reclaim the folio below */
1182 * Before reclaiming the folio, try to relocate
1183 * its contents to another node.
1185 if (do_demote_pass &&
1186 (thp_migration_supported() || !folio_test_large(folio))) {
1187 list_add(&folio->lru, &demote_folios);
1188 folio_unlock(folio);
1193 * Anonymous process memory has backing store?
1194 * Try to allocate it some swap space here.
1195 * Lazyfree folio could be freed directly
1197 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1198 if (!folio_test_swapcache(folio)) {
1199 if (!(sc->gfp_mask & __GFP_IO))
1201 if (folio_maybe_dma_pinned(folio))
1203 if (folio_test_large(folio)) {
1204 /* cannot split folio, skip it */
1205 if (!can_split_folio(folio, NULL))
1206 goto activate_locked;
1208 * Split folios without a PMD map right
1209 * away. Chances are some or all of the
1210 * tail pages can be freed without IO.
1212 if (!folio_entire_mapcount(folio) &&
1213 split_folio_to_list(folio,
1215 goto activate_locked;
1217 if (!add_to_swap(folio)) {
1218 if (!folio_test_large(folio))
1219 goto activate_locked_split;
1220 /* Fallback to swap normal pages */
1221 if (split_folio_to_list(folio,
1223 goto activate_locked;
1224 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1225 count_memcg_folio_events(folio, THP_SWPOUT_FALLBACK, 1);
1226 count_vm_event(THP_SWPOUT_FALLBACK);
1228 if (!add_to_swap(folio))
1229 goto activate_locked_split;
1232 } else if (folio_test_swapbacked(folio) &&
1233 folio_test_large(folio)) {
1234 /* Split shmem folio */
1235 if (split_folio_to_list(folio, folio_list))
1240 * If the folio was split above, the tail pages will make
1241 * their own pass through this function and be accounted
1244 if ((nr_pages > 1) && !folio_test_large(folio)) {
1245 sc->nr_scanned -= (nr_pages - 1);
1250 * The folio is mapped into the page tables of one or more
1251 * processes. Try to unmap it here.
1253 if (folio_mapped(folio)) {
1254 enum ttu_flags flags = TTU_BATCH_FLUSH;
1255 bool was_swapbacked = folio_test_swapbacked(folio);
1257 if (folio_test_pmd_mappable(folio))
1258 flags |= TTU_SPLIT_HUGE_PMD;
1260 try_to_unmap(folio, flags);
1261 if (folio_mapped(folio)) {
1262 stat->nr_unmap_fail += nr_pages;
1263 if (!was_swapbacked &&
1264 folio_test_swapbacked(folio))
1265 stat->nr_lazyfree_fail += nr_pages;
1266 goto activate_locked;
1271 * Folio is unmapped now so it cannot be newly pinned anymore.
1272 * No point in trying to reclaim folio if it is pinned.
1273 * Furthermore we don't want to reclaim underlying fs metadata
1274 * if the folio is pinned and thus potentially modified by the
1275 * pinning process as that may upset the filesystem.
1277 if (folio_maybe_dma_pinned(folio))
1278 goto activate_locked;
1280 mapping = folio_mapping(folio);
1281 if (folio_test_dirty(folio)) {
1283 * Only kswapd can writeback filesystem folios
1284 * to avoid risk of stack overflow. But avoid
1285 * injecting inefficient single-folio I/O into
1286 * flusher writeback as much as possible: only
1287 * write folios when we've encountered many
1288 * dirty folios, and when we've already scanned
1289 * the rest of the LRU for clean folios and see
1290 * the same dirty folios again (with the reclaim
1293 if (folio_is_file_lru(folio) &&
1294 (!current_is_kswapd() ||
1295 !folio_test_reclaim(folio) ||
1296 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1298 * Immediately reclaim when written back.
1299 * Similar in principle to folio_deactivate()
1300 * except we already have the folio isolated
1301 * and know it's dirty
1303 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1305 folio_set_reclaim(folio);
1307 goto activate_locked;
1310 if (references == FOLIOREF_RECLAIM_CLEAN)
1312 if (!may_enter_fs(folio, sc->gfp_mask))
1314 if (!sc->may_writepage)
1318 * Folio is dirty. Flush the TLB if a writable entry
1319 * potentially exists to avoid CPU writes after I/O
1320 * starts and then write it out here.
1322 try_to_unmap_flush_dirty();
1323 switch (pageout(folio, mapping, &plug)) {
1327 goto activate_locked;
1329 stat->nr_pageout += nr_pages;
1331 if (folio_test_writeback(folio))
1333 if (folio_test_dirty(folio))
1337 * A synchronous write - probably a ramdisk. Go
1338 * ahead and try to reclaim the folio.
1340 if (!folio_trylock(folio))
1342 if (folio_test_dirty(folio) ||
1343 folio_test_writeback(folio))
1345 mapping = folio_mapping(folio);
1348 ; /* try to free the folio below */
1353 * If the folio has buffers, try to free the buffer
1354 * mappings associated with this folio. If we succeed
1355 * we try to free the folio as well.
1357 * We do this even if the folio is dirty.
1358 * filemap_release_folio() does not perform I/O, but it
1359 * is possible for a folio to have the dirty flag set,
1360 * but it is actually clean (all its buffers are clean).
1361 * This happens if the buffers were written out directly,
1362 * with submit_bh(). ext3 will do this, as well as
1363 * the blockdev mapping. filemap_release_folio() will
1364 * discover that cleanness and will drop the buffers
1365 * and mark the folio clean - it can be freed.
1367 * Rarely, folios can have buffers and no ->mapping.
1368 * These are the folios which were not successfully
1369 * invalidated in truncate_cleanup_folio(). We try to
1370 * drop those buffers here and if that worked, and the
1371 * folio is no longer mapped into process address space
1372 * (refcount == 1) it can be freed. Otherwise, leave
1373 * the folio on the LRU so it is swappable.
1375 if (folio_needs_release(folio)) {
1376 if (!filemap_release_folio(folio, sc->gfp_mask))
1377 goto activate_locked;
1378 if (!mapping && folio_ref_count(folio) == 1) {
1379 folio_unlock(folio);
1380 if (folio_put_testzero(folio))
1384 * rare race with speculative reference.
1385 * the speculative reference will free
1386 * this folio shortly, so we may
1387 * increment nr_reclaimed here (and
1388 * leave it off the LRU).
1390 nr_reclaimed += nr_pages;
1396 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1397 /* follow __remove_mapping for reference */
1398 if (!folio_ref_freeze(folio, 1))
1401 * The folio has only one reference left, which is
1402 * from the isolation. After the caller puts the
1403 * folio back on the lru and drops the reference, the
1404 * folio will be freed anyway. It doesn't matter
1405 * which lru it goes on. So we don't bother checking
1406 * the dirty flag here.
1408 count_vm_events(PGLAZYFREED, nr_pages);
1409 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1410 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1411 sc->target_mem_cgroup))
1414 folio_unlock(folio);
1417 * Folio may get swapped out as a whole, need to account
1420 nr_reclaimed += nr_pages;
1422 if (folio_test_large(folio) &&
1423 folio_test_large_rmappable(folio))
1424 folio_undo_large_rmappable(folio);
1425 if (folio_batch_add(&free_folios, folio) == 0) {
1426 mem_cgroup_uncharge_folios(&free_folios);
1427 try_to_unmap_flush();
1428 free_unref_folios(&free_folios);
1432 activate_locked_split:
1434 * The tail pages that are failed to add into swap cache
1435 * reach here. Fixup nr_scanned and nr_pages.
1438 sc->nr_scanned -= (nr_pages - 1);
1442 /* Not a candidate for swapping, so reclaim swap space. */
1443 if (folio_test_swapcache(folio) &&
1444 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1445 folio_free_swap(folio);
1446 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1447 if (!folio_test_mlocked(folio)) {
1448 int type = folio_is_file_lru(folio);
1449 folio_set_active(folio);
1450 stat->nr_activate[type] += nr_pages;
1451 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1454 folio_unlock(folio);
1456 list_add(&folio->lru, &ret_folios);
1457 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1458 folio_test_unevictable(folio), folio);
1460 /* 'folio_list' is always empty here */
1462 /* Migrate folios selected for demotion */
1463 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
1464 /* Folios that could not be demoted are still in @demote_folios */
1465 if (!list_empty(&demote_folios)) {
1466 /* Folios which weren't demoted go back on @folio_list */
1467 list_splice_init(&demote_folios, folio_list);
1470 * goto retry to reclaim the undemoted folios in folio_list if
1473 * Reclaiming directly from top tier nodes is not often desired
1474 * due to it breaking the LRU ordering: in general memory
1475 * should be reclaimed from lower tier nodes and demoted from
1478 * However, disabling reclaim from top tier nodes entirely
1479 * would cause ooms in edge scenarios where lower tier memory
1480 * is unreclaimable for whatever reason, eg memory being
1481 * mlocked or too hot to reclaim. We can disable reclaim
1482 * from top tier nodes in proactive reclaim though as that is
1483 * not real memory pressure.
1485 if (!sc->proactive) {
1486 do_demote_pass = false;
1491 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1493 mem_cgroup_uncharge_folios(&free_folios);
1494 try_to_unmap_flush();
1495 free_unref_folios(&free_folios);
1497 list_splice(&ret_folios, folio_list);
1498 count_vm_events(PGACTIVATE, pgactivate);
1501 swap_write_unplug(plug);
1502 return nr_reclaimed;
1505 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1506 struct list_head *folio_list)
1508 struct scan_control sc = {
1509 .gfp_mask = GFP_KERNEL,
1512 struct reclaim_stat stat;
1513 unsigned int nr_reclaimed;
1514 struct folio *folio, *next;
1515 LIST_HEAD(clean_folios);
1516 unsigned int noreclaim_flag;
1518 list_for_each_entry_safe(folio, next, folio_list, lru) {
1519 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1520 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1521 !folio_test_unevictable(folio)) {
1522 folio_clear_active(folio);
1523 list_move(&folio->lru, &clean_folios);
1528 * We should be safe here since we are only dealing with file pages and
1529 * we are not kswapd and therefore cannot write dirty file pages. But
1530 * call memalloc_noreclaim_save() anyway, just in case these conditions
1531 * change in the future.
1533 noreclaim_flag = memalloc_noreclaim_save();
1534 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1536 memalloc_noreclaim_restore(noreclaim_flag);
1538 list_splice(&clean_folios, folio_list);
1539 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1540 -(long)nr_reclaimed);
1542 * Since lazyfree pages are isolated from file LRU from the beginning,
1543 * they will rotate back to anonymous LRU in the end if it failed to
1544 * discard so isolated count will be mismatched.
1545 * Compensate the isolated count for both LRU lists.
1547 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1548 stat.nr_lazyfree_fail);
1549 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1550 -(long)stat.nr_lazyfree_fail);
1551 return nr_reclaimed;
1555 * Update LRU sizes after isolating pages. The LRU size updates must
1556 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1558 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1559 enum lru_list lru, unsigned long *nr_zone_taken)
1563 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1564 if (!nr_zone_taken[zid])
1567 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1574 * It is waste of effort to scan and reclaim CMA pages if it is not available
1575 * for current allocation context. Kswapd can not be enrolled as it can not
1576 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1578 static bool skip_cma(struct folio *folio, struct scan_control *sc)
1580 return !current_is_kswapd() &&
1581 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1582 folio_migratetype(folio) == MIGRATE_CMA;
1585 static bool skip_cma(struct folio *folio, struct scan_control *sc)
1592 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1594 * lruvec->lru_lock is heavily contended. Some of the functions that
1595 * shrink the lists perform better by taking out a batch of pages
1596 * and working on them outside the LRU lock.
1598 * For pagecache intensive workloads, this function is the hottest
1599 * spot in the kernel (apart from copy_*_user functions).
1601 * Lru_lock must be held before calling this function.
1603 * @nr_to_scan: The number of eligible pages to look through on the list.
1604 * @lruvec: The LRU vector to pull pages from.
1605 * @dst: The temp list to put pages on to.
1606 * @nr_scanned: The number of pages that were scanned.
1607 * @sc: The scan_control struct for this reclaim session
1608 * @lru: LRU list id for isolating
1610 * returns how many pages were moved onto *@dst.
1612 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1613 struct lruvec *lruvec, struct list_head *dst,
1614 unsigned long *nr_scanned, struct scan_control *sc,
1617 struct list_head *src = &lruvec->lists[lru];
1618 unsigned long nr_taken = 0;
1619 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1620 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1621 unsigned long skipped = 0;
1622 unsigned long scan, total_scan, nr_pages;
1623 LIST_HEAD(folios_skipped);
1627 while (scan < nr_to_scan && !list_empty(src)) {
1628 struct list_head *move_to = src;
1629 struct folio *folio;
1631 folio = lru_to_folio(src);
1632 prefetchw_prev_lru_folio(folio, src, flags);
1634 nr_pages = folio_nr_pages(folio);
1635 total_scan += nr_pages;
1637 if (folio_zonenum(folio) > sc->reclaim_idx ||
1638 skip_cma(folio, sc)) {
1639 nr_skipped[folio_zonenum(folio)] += nr_pages;
1640 move_to = &folios_skipped;
1645 * Do not count skipped folios because that makes the function
1646 * return with no isolated folios if the LRU mostly contains
1647 * ineligible folios. This causes the VM to not reclaim any
1648 * folios, triggering a premature OOM.
1649 * Account all pages in a folio.
1653 if (!folio_test_lru(folio))
1655 if (!sc->may_unmap && folio_mapped(folio))
1659 * Be careful not to clear the lru flag until after we're
1660 * sure the folio is not being freed elsewhere -- the
1661 * folio release code relies on it.
1663 if (unlikely(!folio_try_get(folio)))
1666 if (!folio_test_clear_lru(folio)) {
1667 /* Another thread is already isolating this folio */
1672 nr_taken += nr_pages;
1673 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1676 list_move(&folio->lru, move_to);
1680 * Splice any skipped folios to the start of the LRU list. Note that
1681 * this disrupts the LRU order when reclaiming for lower zones but
1682 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1683 * scanning would soon rescan the same folios to skip and waste lots
1686 if (!list_empty(&folios_skipped)) {
1689 list_splice(&folios_skipped, src);
1690 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1691 if (!nr_skipped[zid])
1694 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1695 skipped += nr_skipped[zid];
1698 *nr_scanned = total_scan;
1699 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1700 total_scan, skipped, nr_taken, lru);
1701 update_lru_sizes(lruvec, lru, nr_zone_taken);
1706 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1707 * @folio: Folio to isolate from its LRU list.
1709 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1710 * corresponding to whatever LRU list the folio was on.
1712 * The folio will have its LRU flag cleared. If it was found on the
1713 * active list, it will have the Active flag set. If it was found on the
1714 * unevictable list, it will have the Unevictable flag set. These flags
1715 * may need to be cleared by the caller before letting the page go.
1719 * (1) Must be called with an elevated refcount on the folio. This is a
1720 * fundamental difference from isolate_lru_folios() (which is called
1721 * without a stable reference).
1722 * (2) The lru_lock must not be held.
1723 * (3) Interrupts must be enabled.
1725 * Return: true if the folio was removed from an LRU list.
1726 * false if the folio was not on an LRU list.
1728 bool folio_isolate_lru(struct folio *folio)
1732 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1734 if (folio_test_clear_lru(folio)) {
1735 struct lruvec *lruvec;
1738 lruvec = folio_lruvec_lock_irq(folio);
1739 lruvec_del_folio(lruvec, folio);
1740 unlock_page_lruvec_irq(lruvec);
1748 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1749 * then get rescheduled. When there are massive number of tasks doing page
1750 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1751 * the LRU list will go small and be scanned faster than necessary, leading to
1752 * unnecessary swapping, thrashing and OOM.
1754 static bool too_many_isolated(struct pglist_data *pgdat, int file,
1755 struct scan_control *sc)
1757 unsigned long inactive, isolated;
1760 if (current_is_kswapd())
1763 if (!writeback_throttling_sane(sc))
1767 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1768 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1770 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1771 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1775 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1776 * won't get blocked by normal direct-reclaimers, forming a circular
1779 if (gfp_has_io_fs(sc->gfp_mask))
1782 too_many = isolated > inactive;
1784 /* Wake up tasks throttled due to too_many_isolated. */
1786 wake_throttle_isolated(pgdat);
1792 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1794 * Returns the number of pages moved to the given lruvec.
1796 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1797 struct list_head *list)
1799 int nr_pages, nr_moved = 0;
1800 struct folio_batch free_folios;
1802 folio_batch_init(&free_folios);
1803 while (!list_empty(list)) {
1804 struct folio *folio = lru_to_folio(list);
1806 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1807 list_del(&folio->lru);
1808 if (unlikely(!folio_evictable(folio))) {
1809 spin_unlock_irq(&lruvec->lru_lock);
1810 folio_putback_lru(folio);
1811 spin_lock_irq(&lruvec->lru_lock);
1816 * The folio_set_lru needs to be kept here for list integrity.
1818 * #0 move_folios_to_lru #1 release_pages
1819 * if (!folio_put_testzero())
1820 * if (folio_put_testzero())
1821 * !lru //skip lru_lock
1823 * list_add(&folio->lru,)
1824 * list_add(&folio->lru,)
1826 folio_set_lru(folio);
1828 if (unlikely(folio_put_testzero(folio))) {
1829 __folio_clear_lru_flags(folio);
1831 if (folio_test_large(folio) &&
1832 folio_test_large_rmappable(folio))
1833 folio_undo_large_rmappable(folio);
1834 if (folio_batch_add(&free_folios, folio) == 0) {
1835 spin_unlock_irq(&lruvec->lru_lock);
1836 mem_cgroup_uncharge_folios(&free_folios);
1837 free_unref_folios(&free_folios);
1838 spin_lock_irq(&lruvec->lru_lock);
1845 * All pages were isolated from the same lruvec (and isolation
1846 * inhibits memcg migration).
1848 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1849 lruvec_add_folio(lruvec, folio);
1850 nr_pages = folio_nr_pages(folio);
1851 nr_moved += nr_pages;
1852 if (folio_test_active(folio))
1853 workingset_age_nonresident(lruvec, nr_pages);
1856 if (free_folios.nr) {
1857 spin_unlock_irq(&lruvec->lru_lock);
1858 mem_cgroup_uncharge_folios(&free_folios);
1859 free_unref_folios(&free_folios);
1860 spin_lock_irq(&lruvec->lru_lock);
1867 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1868 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1869 * we should not throttle. Otherwise it is safe to do so.
1871 static int current_may_throttle(void)
1873 return !(current->flags & PF_LOCAL_THROTTLE);
1877 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1878 * of reclaimed pages
1880 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1881 struct lruvec *lruvec, struct scan_control *sc,
1884 LIST_HEAD(folio_list);
1885 unsigned long nr_scanned;
1886 unsigned int nr_reclaimed = 0;
1887 unsigned long nr_taken;
1888 struct reclaim_stat stat;
1889 bool file = is_file_lru(lru);
1890 enum vm_event_item item;
1891 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1892 bool stalled = false;
1894 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1898 /* wait a bit for the reclaimer. */
1900 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1902 /* We are about to die and free our memory. Return now. */
1903 if (fatal_signal_pending(current))
1904 return SWAP_CLUSTER_MAX;
1909 spin_lock_irq(&lruvec->lru_lock);
1911 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1912 &nr_scanned, sc, lru);
1914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1915 item = PGSCAN_KSWAPD + reclaimer_offset();
1916 if (!cgroup_reclaim(sc))
1917 __count_vm_events(item, nr_scanned);
1918 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1919 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1921 spin_unlock_irq(&lruvec->lru_lock);
1926 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1928 spin_lock_irq(&lruvec->lru_lock);
1929 move_folios_to_lru(lruvec, &folio_list);
1931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1932 item = PGSTEAL_KSWAPD + reclaimer_offset();
1933 if (!cgroup_reclaim(sc))
1934 __count_vm_events(item, nr_reclaimed);
1935 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1936 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1937 spin_unlock_irq(&lruvec->lru_lock);
1939 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1942 * If dirty folios are scanned that are not queued for IO, it
1943 * implies that flushers are not doing their job. This can
1944 * happen when memory pressure pushes dirty folios to the end of
1945 * the LRU before the dirty limits are breached and the dirty
1946 * data has expired. It can also happen when the proportion of
1947 * dirty folios grows not through writes but through memory
1948 * pressure reclaiming all the clean cache. And in some cases,
1949 * the flushers simply cannot keep up with the allocation
1950 * rate. Nudge the flusher threads in case they are asleep.
1952 if (stat.nr_unqueued_dirty == nr_taken) {
1953 wakeup_flusher_threads(WB_REASON_VMSCAN);
1955 * For cgroupv1 dirty throttling is achieved by waking up
1956 * the kernel flusher here and later waiting on folios
1957 * which are in writeback to finish (see shrink_folio_list()).
1959 * Flusher may not be able to issue writeback quickly
1960 * enough for cgroupv1 writeback throttling to work
1961 * on a large system.
1963 if (!writeback_throttling_sane(sc))
1964 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1967 sc->nr.dirty += stat.nr_dirty;
1968 sc->nr.congested += stat.nr_congested;
1969 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1970 sc->nr.writeback += stat.nr_writeback;
1971 sc->nr.immediate += stat.nr_immediate;
1972 sc->nr.taken += nr_taken;
1974 sc->nr.file_taken += nr_taken;
1976 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1977 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1978 return nr_reclaimed;
1982 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1984 * We move them the other way if the folio is referenced by one or more
1987 * If the folios are mostly unmapped, the processing is fast and it is
1988 * appropriate to hold lru_lock across the whole operation. But if
1989 * the folios are mapped, the processing is slow (folio_referenced()), so
1990 * we should drop lru_lock around each folio. It's impossible to balance
1991 * this, so instead we remove the folios from the LRU while processing them.
1992 * It is safe to rely on the active flag against the non-LRU folios in here
1993 * because nobody will play with that bit on a non-LRU folio.
1995 * The downside is that we have to touch folio->_refcount against each folio.
1996 * But we had to alter folio->flags anyway.
1998 static void shrink_active_list(unsigned long nr_to_scan,
1999 struct lruvec *lruvec,
2000 struct scan_control *sc,
2003 unsigned long nr_taken;
2004 unsigned long nr_scanned;
2005 unsigned long vm_flags;
2006 LIST_HEAD(l_hold); /* The folios which were snipped off */
2007 LIST_HEAD(l_active);
2008 LIST_HEAD(l_inactive);
2009 unsigned nr_deactivate, nr_activate;
2010 unsigned nr_rotated = 0;
2011 bool file = is_file_lru(lru);
2012 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2016 spin_lock_irq(&lruvec->lru_lock);
2018 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2019 &nr_scanned, sc, lru);
2021 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2023 if (!cgroup_reclaim(sc))
2024 __count_vm_events(PGREFILL, nr_scanned);
2025 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2027 spin_unlock_irq(&lruvec->lru_lock);
2029 while (!list_empty(&l_hold)) {
2030 struct folio *folio;
2033 folio = lru_to_folio(&l_hold);
2034 list_del(&folio->lru);
2036 if (unlikely(!folio_evictable(folio))) {
2037 folio_putback_lru(folio);
2041 if (unlikely(buffer_heads_over_limit)) {
2042 if (folio_needs_release(folio) &&
2043 folio_trylock(folio)) {
2044 filemap_release_folio(folio, 0);
2045 folio_unlock(folio);
2049 /* Referenced or rmap lock contention: rotate */
2050 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2053 * Identify referenced, file-backed active folios and
2054 * give them one more trip around the active list. So
2055 * that executable code get better chances to stay in
2056 * memory under moderate memory pressure. Anon folios
2057 * are not likely to be evicted by use-once streaming
2058 * IO, plus JVM can create lots of anon VM_EXEC folios,
2059 * so we ignore them here.
2061 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2062 nr_rotated += folio_nr_pages(folio);
2063 list_add(&folio->lru, &l_active);
2068 folio_clear_active(folio); /* we are de-activating */
2069 folio_set_workingset(folio);
2070 list_add(&folio->lru, &l_inactive);
2074 * Move folios back to the lru list.
2076 spin_lock_irq(&lruvec->lru_lock);
2078 nr_activate = move_folios_to_lru(lruvec, &l_active);
2079 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2081 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2082 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2084 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2085 spin_unlock_irq(&lruvec->lru_lock);
2088 lru_note_cost(lruvec, file, 0, nr_rotated);
2089 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2090 nr_deactivate, nr_rotated, sc->priority, file);
2093 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2094 struct pglist_data *pgdat,
2095 bool ignore_references)
2097 struct reclaim_stat dummy_stat;
2098 unsigned int nr_reclaimed;
2099 struct folio *folio;
2100 struct scan_control sc = {
2101 .gfp_mask = GFP_KERNEL,
2108 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, ignore_references);
2109 while (!list_empty(folio_list)) {
2110 folio = lru_to_folio(folio_list);
2111 list_del(&folio->lru);
2112 folio_putback_lru(folio);
2115 return nr_reclaimed;
2118 unsigned long reclaim_pages(struct list_head *folio_list, bool ignore_references)
2121 unsigned int nr_reclaimed = 0;
2122 LIST_HEAD(node_folio_list);
2123 unsigned int noreclaim_flag;
2125 if (list_empty(folio_list))
2126 return nr_reclaimed;
2128 noreclaim_flag = memalloc_noreclaim_save();
2130 nid = folio_nid(lru_to_folio(folio_list));
2132 struct folio *folio = lru_to_folio(folio_list);
2134 if (nid == folio_nid(folio)) {
2135 folio_clear_active(folio);
2136 list_move(&folio->lru, &node_folio_list);
2140 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid),
2142 nid = folio_nid(lru_to_folio(folio_list));
2143 } while (!list_empty(folio_list));
2145 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid), ignore_references);
2147 memalloc_noreclaim_restore(noreclaim_flag);
2149 return nr_reclaimed;
2152 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2153 struct lruvec *lruvec, struct scan_control *sc)
2155 if (is_active_lru(lru)) {
2156 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2157 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2159 sc->skipped_deactivate = 1;
2163 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2167 * The inactive anon list should be small enough that the VM never has
2168 * to do too much work.
2170 * The inactive file list should be small enough to leave most memory
2171 * to the established workingset on the scan-resistant active list,
2172 * but large enough to avoid thrashing the aggregate readahead window.
2174 * Both inactive lists should also be large enough that each inactive
2175 * folio has a chance to be referenced again before it is reclaimed.
2177 * If that fails and refaulting is observed, the inactive list grows.
2179 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2180 * on this LRU, maintained by the pageout code. An inactive_ratio
2181 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2184 * memory ratio inactive
2185 * -------------------------------------
2194 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2196 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2197 unsigned long inactive, active;
2198 unsigned long inactive_ratio;
2201 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2202 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2204 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2206 inactive_ratio = int_sqrt(10 * gb);
2210 return inactive * inactive_ratio < active;
2220 static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2223 struct lruvec *target_lruvec;
2225 if (lru_gen_enabled())
2228 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2231 * Flush the memory cgroup stats, so that we read accurate per-memcg
2232 * lruvec stats for heuristics.
2234 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2237 * Determine the scan balance between anon and file LRUs.
2239 spin_lock_irq(&target_lruvec->lru_lock);
2240 sc->anon_cost = target_lruvec->anon_cost;
2241 sc->file_cost = target_lruvec->file_cost;
2242 spin_unlock_irq(&target_lruvec->lru_lock);
2245 * Target desirable inactive:active list ratios for the anon
2246 * and file LRU lists.
2248 if (!sc->force_deactivate) {
2249 unsigned long refaults;
2252 * When refaults are being observed, it means a new
2253 * workingset is being established. Deactivate to get
2254 * rid of any stale active pages quickly.
2256 refaults = lruvec_page_state(target_lruvec,
2257 WORKINGSET_ACTIVATE_ANON);
2258 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2259 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2260 sc->may_deactivate |= DEACTIVATE_ANON;
2262 sc->may_deactivate &= ~DEACTIVATE_ANON;
2264 refaults = lruvec_page_state(target_lruvec,
2265 WORKINGSET_ACTIVATE_FILE);
2266 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2267 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2268 sc->may_deactivate |= DEACTIVATE_FILE;
2270 sc->may_deactivate &= ~DEACTIVATE_FILE;
2272 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2275 * If we have plenty of inactive file pages that aren't
2276 * thrashing, try to reclaim those first before touching
2279 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2280 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2281 !sc->no_cache_trim_mode)
2282 sc->cache_trim_mode = 1;
2284 sc->cache_trim_mode = 0;
2287 * Prevent the reclaimer from falling into the cache trap: as
2288 * cache pages start out inactive, every cache fault will tip
2289 * the scan balance towards the file LRU. And as the file LRU
2290 * shrinks, so does the window for rotation from references.
2291 * This means we have a runaway feedback loop where a tiny
2292 * thrashing file LRU becomes infinitely more attractive than
2293 * anon pages. Try to detect this based on file LRU size.
2295 if (!cgroup_reclaim(sc)) {
2296 unsigned long total_high_wmark = 0;
2297 unsigned long free, anon;
2300 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2301 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2302 node_page_state(pgdat, NR_INACTIVE_FILE);
2304 for (z = 0; z < MAX_NR_ZONES; z++) {
2305 struct zone *zone = &pgdat->node_zones[z];
2307 if (!managed_zone(zone))
2310 total_high_wmark += high_wmark_pages(zone);
2314 * Consider anon: if that's low too, this isn't a
2315 * runaway file reclaim problem, but rather just
2316 * extreme pressure. Reclaim as per usual then.
2318 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2321 file + free <= total_high_wmark &&
2322 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2323 anon >> sc->priority;
2328 * Determine how aggressively the anon and file LRU lists should be
2331 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2332 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2334 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2337 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2338 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2339 unsigned long anon_cost, file_cost, total_cost;
2340 int swappiness = mem_cgroup_swappiness(memcg);
2341 u64 fraction[ANON_AND_FILE];
2342 u64 denominator = 0; /* gcc */
2343 enum scan_balance scan_balance;
2344 unsigned long ap, fp;
2347 /* If we have no swap space, do not bother scanning anon folios. */
2348 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2349 scan_balance = SCAN_FILE;
2354 * Global reclaim will swap to prevent OOM even with no
2355 * swappiness, but memcg users want to use this knob to
2356 * disable swapping for individual groups completely when
2357 * using the memory controller's swap limit feature would be
2360 if (cgroup_reclaim(sc) && !swappiness) {
2361 scan_balance = SCAN_FILE;
2366 * Do not apply any pressure balancing cleverness when the
2367 * system is close to OOM, scan both anon and file equally
2368 * (unless the swappiness setting disagrees with swapping).
2370 if (!sc->priority && swappiness) {
2371 scan_balance = SCAN_EQUAL;
2376 * If the system is almost out of file pages, force-scan anon.
2378 if (sc->file_is_tiny) {
2379 scan_balance = SCAN_ANON;
2384 * If there is enough inactive page cache, we do not reclaim
2385 * anything from the anonymous working right now.
2387 if (sc->cache_trim_mode) {
2388 scan_balance = SCAN_FILE;
2392 scan_balance = SCAN_FRACT;
2394 * Calculate the pressure balance between anon and file pages.
2396 * The amount of pressure we put on each LRU is inversely
2397 * proportional to the cost of reclaiming each list, as
2398 * determined by the share of pages that are refaulting, times
2399 * the relative IO cost of bringing back a swapped out
2400 * anonymous page vs reloading a filesystem page (swappiness).
2402 * Although we limit that influence to ensure no list gets
2403 * left behind completely: at least a third of the pressure is
2404 * applied, before swappiness.
2406 * With swappiness at 100, anon and file have equal IO cost.
2408 total_cost = sc->anon_cost + sc->file_cost;
2409 anon_cost = total_cost + sc->anon_cost;
2410 file_cost = total_cost + sc->file_cost;
2411 total_cost = anon_cost + file_cost;
2413 ap = swappiness * (total_cost + 1);
2414 ap /= anon_cost + 1;
2416 fp = (200 - swappiness) * (total_cost + 1);
2417 fp /= file_cost + 1;
2421 denominator = ap + fp;
2423 for_each_evictable_lru(lru) {
2424 bool file = is_file_lru(lru);
2425 unsigned long lruvec_size;
2426 unsigned long low, min;
2429 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2430 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2435 * Scale a cgroup's reclaim pressure by proportioning
2436 * its current usage to its memory.low or memory.min
2439 * This is important, as otherwise scanning aggression
2440 * becomes extremely binary -- from nothing as we
2441 * approach the memory protection threshold, to totally
2442 * nominal as we exceed it. This results in requiring
2443 * setting extremely liberal protection thresholds. It
2444 * also means we simply get no protection at all if we
2445 * set it too low, which is not ideal.
2447 * If there is any protection in place, we reduce scan
2448 * pressure by how much of the total memory used is
2449 * within protection thresholds.
2451 * There is one special case: in the first reclaim pass,
2452 * we skip over all groups that are within their low
2453 * protection. If that fails to reclaim enough pages to
2454 * satisfy the reclaim goal, we come back and override
2455 * the best-effort low protection. However, we still
2456 * ideally want to honor how well-behaved groups are in
2457 * that case instead of simply punishing them all
2458 * equally. As such, we reclaim them based on how much
2459 * memory they are using, reducing the scan pressure
2460 * again by how much of the total memory used is under
2463 unsigned long cgroup_size = mem_cgroup_size(memcg);
2464 unsigned long protection;
2466 /* memory.low scaling, make sure we retry before OOM */
2467 if (!sc->memcg_low_reclaim && low > min) {
2469 sc->memcg_low_skipped = 1;
2474 /* Avoid TOCTOU with earlier protection check */
2475 cgroup_size = max(cgroup_size, protection);
2477 scan = lruvec_size - lruvec_size * protection /
2481 * Minimally target SWAP_CLUSTER_MAX pages to keep
2482 * reclaim moving forwards, avoiding decrementing
2483 * sc->priority further than desirable.
2485 scan = max(scan, SWAP_CLUSTER_MAX);
2490 scan >>= sc->priority;
2493 * If the cgroup's already been deleted, make sure to
2494 * scrape out the remaining cache.
2496 if (!scan && !mem_cgroup_online(memcg))
2497 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2499 switch (scan_balance) {
2501 /* Scan lists relative to size */
2505 * Scan types proportional to swappiness and
2506 * their relative recent reclaim efficiency.
2507 * Make sure we don't miss the last page on
2508 * the offlined memory cgroups because of a
2511 scan = mem_cgroup_online(memcg) ?
2512 div64_u64(scan * fraction[file], denominator) :
2513 DIV64_U64_ROUND_UP(scan * fraction[file],
2518 /* Scan one type exclusively */
2519 if ((scan_balance == SCAN_FILE) != file)
2523 /* Look ma, no brain */
2532 * Anonymous LRU management is a waste if there is
2533 * ultimately no way to reclaim the memory.
2535 static bool can_age_anon_pages(struct pglist_data *pgdat,
2536 struct scan_control *sc)
2538 /* Aging the anon LRU is valuable if swap is present: */
2539 if (total_swap_pages > 0)
2542 /* Also valuable if anon pages can be demoted: */
2543 return can_demote(pgdat->node_id, sc);
2546 #ifdef CONFIG_LRU_GEN
2548 #ifdef CONFIG_LRU_GEN_ENABLED
2549 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2550 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2552 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2553 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2556 static bool should_walk_mmu(void)
2558 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2561 static bool should_clear_pmd_young(void)
2563 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2566 /******************************************************************************
2568 ******************************************************************************/
2570 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2572 #define DEFINE_MAX_SEQ(lruvec) \
2573 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2575 #define DEFINE_MIN_SEQ(lruvec) \
2576 unsigned long min_seq[ANON_AND_FILE] = { \
2577 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2578 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2581 #define for_each_gen_type_zone(gen, type, zone) \
2582 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2583 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2584 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2586 #define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2587 #define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2589 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2591 struct pglist_data *pgdat = NODE_DATA(nid);
2595 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2597 /* see the comment in mem_cgroup_lruvec() */
2599 lruvec->pgdat = pgdat;
2604 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2606 return &pgdat->__lruvec;
2609 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2611 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2612 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2617 if (!can_demote(pgdat->node_id, sc) &&
2618 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2621 return mem_cgroup_swappiness(memcg);
2624 static int get_nr_gens(struct lruvec *lruvec, int type)
2626 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2629 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2631 /* see the comment on lru_gen_folio */
2632 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2633 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2634 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2637 /******************************************************************************
2639 ******************************************************************************/
2642 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2643 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2644 * bits in a bitmap, k is the number of hash functions and n is the number of
2647 * Page table walkers use one of the two filters to reduce their search space.
2648 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2649 * aging uses the double-buffering technique to flip to the other filter each
2650 * time it produces a new generation. For non-leaf entries that have enough
2651 * leaf entries, the aging carries them over to the next generation in
2652 * walk_pmd_range(); the eviction also report them when walking the rmap
2653 * in lru_gen_look_around().
2655 * For future optimizations:
2656 * 1. It's not necessary to keep both filters all the time. The spare one can be
2657 * freed after the RCU grace period and reallocated if needed again.
2658 * 2. And when reallocating, it's worth scaling its size according to the number
2659 * of inserted entries in the other filter, to reduce the memory overhead on
2660 * small systems and false positives on large systems.
2661 * 3. Jenkins' hash function is an alternative to Knuth's.
2663 #define BLOOM_FILTER_SHIFT 15
2665 static inline int filter_gen_from_seq(unsigned long seq)
2667 return seq % NR_BLOOM_FILTERS;
2670 static void get_item_key(void *item, int *key)
2672 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2674 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2676 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2677 key[1] = hash >> BLOOM_FILTER_SHIFT;
2680 static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2684 unsigned long *filter;
2685 int gen = filter_gen_from_seq(seq);
2687 filter = READ_ONCE(mm_state->filters[gen]);
2691 get_item_key(item, key);
2693 return test_bit(key[0], filter) && test_bit(key[1], filter);
2696 static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2700 unsigned long *filter;
2701 int gen = filter_gen_from_seq(seq);
2703 filter = READ_ONCE(mm_state->filters[gen]);
2707 get_item_key(item, key);
2709 if (!test_bit(key[0], filter))
2710 set_bit(key[0], filter);
2711 if (!test_bit(key[1], filter))
2712 set_bit(key[1], filter);
2715 static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2717 unsigned long *filter;
2718 int gen = filter_gen_from_seq(seq);
2720 filter = mm_state->filters[gen];
2722 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2726 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2727 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2728 WRITE_ONCE(mm_state->filters[gen], filter);
2731 /******************************************************************************
2733 ******************************************************************************/
2735 #ifdef CONFIG_LRU_GEN_WALKS_MMU
2737 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2739 static struct lru_gen_mm_list mm_list = {
2740 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2741 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2746 return &memcg->mm_list;
2748 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2753 static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2755 return &lruvec->mm_state;
2758 static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2761 struct mm_struct *mm;
2762 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2763 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2765 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2766 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2768 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2771 clear_bit(key, &mm->lru_gen.bitmap);
2773 return mmget_not_zero(mm) ? mm : NULL;
2776 void lru_gen_add_mm(struct mm_struct *mm)
2779 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2780 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2782 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2784 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2785 mm->lru_gen.memcg = memcg;
2787 spin_lock(&mm_list->lock);
2789 for_each_node_state(nid, N_MEMORY) {
2790 struct lruvec *lruvec = get_lruvec(memcg, nid);
2791 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2793 /* the first addition since the last iteration */
2794 if (mm_state->tail == &mm_list->fifo)
2795 mm_state->tail = &mm->lru_gen.list;
2798 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2800 spin_unlock(&mm_list->lock);
2803 void lru_gen_del_mm(struct mm_struct *mm)
2806 struct lru_gen_mm_list *mm_list;
2807 struct mem_cgroup *memcg = NULL;
2809 if (list_empty(&mm->lru_gen.list))
2813 memcg = mm->lru_gen.memcg;
2815 mm_list = get_mm_list(memcg);
2817 spin_lock(&mm_list->lock);
2819 for_each_node(nid) {
2820 struct lruvec *lruvec = get_lruvec(memcg, nid);
2821 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2823 /* where the current iteration continues after */
2824 if (mm_state->head == &mm->lru_gen.list)
2825 mm_state->head = mm_state->head->prev;
2827 /* where the last iteration ended before */
2828 if (mm_state->tail == &mm->lru_gen.list)
2829 mm_state->tail = mm_state->tail->next;
2832 list_del_init(&mm->lru_gen.list);
2834 spin_unlock(&mm_list->lock);
2837 mem_cgroup_put(mm->lru_gen.memcg);
2838 mm->lru_gen.memcg = NULL;
2843 void lru_gen_migrate_mm(struct mm_struct *mm)
2845 struct mem_cgroup *memcg;
2846 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2848 VM_WARN_ON_ONCE(task->mm != mm);
2849 lockdep_assert_held(&task->alloc_lock);
2851 /* for mm_update_next_owner() */
2852 if (mem_cgroup_disabled())
2855 /* migration can happen before addition */
2856 if (!mm->lru_gen.memcg)
2860 memcg = mem_cgroup_from_task(task);
2862 if (memcg == mm->lru_gen.memcg)
2865 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2872 #else /* !CONFIG_LRU_GEN_WALKS_MMU */
2874 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2879 static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2884 static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2891 static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
2895 struct lruvec *lruvec = walk->lruvec;
2896 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2898 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2900 hist = lru_hist_from_seq(walk->seq);
2902 for (i = 0; i < NR_MM_STATS; i++) {
2903 WRITE_ONCE(mm_state->stats[hist][i],
2904 mm_state->stats[hist][i] + walk->mm_stats[i]);
2905 walk->mm_stats[i] = 0;
2908 if (NR_HIST_GENS > 1 && last) {
2909 hist = lru_hist_from_seq(walk->seq + 1);
2911 for (i = 0; i < NR_MM_STATS; i++)
2912 WRITE_ONCE(mm_state->stats[hist][i], 0);
2916 static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
2920 struct mm_struct *mm = NULL;
2921 struct lruvec *lruvec = walk->lruvec;
2922 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2923 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2924 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2927 * mm_state->seq is incremented after each iteration of mm_list. There
2928 * are three interesting cases for this page table walker:
2929 * 1. It tries to start a new iteration with a stale max_seq: there is
2930 * nothing left to do.
2931 * 2. It started the next iteration: it needs to reset the Bloom filter
2932 * so that a fresh set of PTE tables can be recorded.
2933 * 3. It ended the current iteration: it needs to reset the mm stats
2934 * counters and tell its caller to increment max_seq.
2936 spin_lock(&mm_list->lock);
2938 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
2940 if (walk->seq <= mm_state->seq)
2943 if (!mm_state->head)
2944 mm_state->head = &mm_list->fifo;
2946 if (mm_state->head == &mm_list->fifo)
2950 mm_state->head = mm_state->head->next;
2951 if (mm_state->head == &mm_list->fifo) {
2952 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2957 /* force scan for those added after the last iteration */
2958 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2959 mm_state->tail = mm_state->head->next;
2960 walk->force_scan = true;
2962 } while (!(mm = get_next_mm(walk)));
2965 reset_mm_stats(walk, last);
2967 spin_unlock(&mm_list->lock);
2970 reset_bloom_filter(mm_state, walk->seq + 1);
2980 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
2982 bool success = false;
2983 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2984 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2985 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2987 spin_lock(&mm_list->lock);
2989 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
2991 if (seq > mm_state->seq) {
2992 mm_state->head = NULL;
2993 mm_state->tail = NULL;
2994 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2998 spin_unlock(&mm_list->lock);
3003 /******************************************************************************
3005 ******************************************************************************/
3008 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3010 * The P term is refaulted/(evicted+protected) from a tier in the generation
3011 * currently being evicted; the I term is the exponential moving average of the
3012 * P term over the generations previously evicted, using the smoothing factor
3013 * 1/2; the D term isn't supported.
3015 * The setpoint (SP) is always the first tier of one type; the process variable
3016 * (PV) is either any tier of the other type or any other tier of the same
3019 * The error is the difference between the SP and the PV; the correction is to
3020 * turn off protection when SP>PV or turn on protection when SP<PV.
3022 * For future optimizations:
3023 * 1. The D term may discount the other two terms over time so that long-lived
3024 * generations can resist stale information.
3027 unsigned long refaulted;
3028 unsigned long total;
3032 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3033 struct ctrl_pos *pos)
3035 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3036 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3038 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3039 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3040 pos->total = lrugen->avg_total[type][tier] +
3041 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3043 pos->total += lrugen->protected[hist][type][tier - 1];
3047 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3050 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3051 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3052 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3054 lockdep_assert_held(&lruvec->lru_lock);
3056 if (!carryover && !clear)
3059 hist = lru_hist_from_seq(seq);
3061 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3065 sum = lrugen->avg_refaulted[type][tier] +
3066 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3067 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3069 sum = lrugen->avg_total[type][tier] +
3070 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3072 sum += lrugen->protected[hist][type][tier - 1];
3073 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3077 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3078 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3080 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3085 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3088 * Return true if the PV has a limited number of refaults or a lower
3089 * refaulted/total than the SP.
3091 return pv->refaulted < MIN_LRU_BATCH ||
3092 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3093 (sp->refaulted + 1) * pv->total * pv->gain;
3096 /******************************************************************************
3098 ******************************************************************************/
3100 /* promote pages accessed through page tables */
3101 static int folio_update_gen(struct folio *folio, int gen)
3103 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3105 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3106 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3109 /* lru_gen_del_folio() has isolated this page? */
3110 if (!(old_flags & LRU_GEN_MASK)) {
3111 /* for shrink_folio_list() */
3112 new_flags = old_flags | BIT(PG_referenced);
3116 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3117 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3118 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3120 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3123 /* protect pages accessed multiple times through file descriptors */
3124 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3126 int type = folio_is_file_lru(folio);
3127 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3128 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3129 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3131 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3134 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3135 /* folio_update_gen() has promoted this page? */
3136 if (new_gen >= 0 && new_gen != old_gen)
3139 new_gen = (old_gen + 1) % MAX_NR_GENS;
3141 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3142 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3143 /* for folio_end_writeback() */
3145 new_flags |= BIT(PG_reclaim);
3146 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3148 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3153 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3154 int old_gen, int new_gen)
3156 int type = folio_is_file_lru(folio);
3157 int zone = folio_zonenum(folio);
3158 int delta = folio_nr_pages(folio);
3160 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3161 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3165 walk->nr_pages[old_gen][type][zone] -= delta;
3166 walk->nr_pages[new_gen][type][zone] += delta;
3169 static void reset_batch_size(struct lru_gen_mm_walk *walk)
3171 int gen, type, zone;
3172 struct lruvec *lruvec = walk->lruvec;
3173 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3177 for_each_gen_type_zone(gen, type, zone) {
3178 enum lru_list lru = type * LRU_INACTIVE_FILE;
3179 int delta = walk->nr_pages[gen][type][zone];
3184 walk->nr_pages[gen][type][zone] = 0;
3185 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3186 lrugen->nr_pages[gen][type][zone] + delta);
3188 if (lru_gen_is_active(lruvec, gen))
3190 __update_lru_size(lruvec, lru, zone, delta);
3194 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3196 struct address_space *mapping;
3197 struct vm_area_struct *vma = args->vma;
3198 struct lru_gen_mm_walk *walk = args->private;
3200 if (!vma_is_accessible(vma))
3203 if (is_vm_hugetlb_page(vma))
3206 if (!vma_has_recency(vma))
3209 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3212 if (vma == get_gate_vma(vma->vm_mm))
3215 if (vma_is_anonymous(vma))
3216 return !walk->can_swap;
3218 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3221 mapping = vma->vm_file->f_mapping;
3222 if (mapping_unevictable(mapping))
3225 if (shmem_mapping(mapping))
3226 return !walk->can_swap;
3228 /* to exclude special mappings like dax, etc. */
3229 return !mapping->a_ops->read_folio;
3233 * Some userspace memory allocators map many single-page VMAs. Instead of
3234 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3235 * table to reduce zigzags and improve cache performance.
3237 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3238 unsigned long *vm_start, unsigned long *vm_end)
3240 unsigned long start = round_up(*vm_end, size);
3241 unsigned long end = (start | ~mask) + 1;
3242 VMA_ITERATOR(vmi, args->mm, start);
3244 VM_WARN_ON_ONCE(mask & size);
3245 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3247 for_each_vma(vmi, args->vma) {
3248 if (end && end <= args->vma->vm_start)
3251 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3254 *vm_start = max(start, args->vma->vm_start);
3255 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3263 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3265 unsigned long pfn = pte_pfn(pte);
3267 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3269 if (!pte_present(pte) || is_zero_pfn(pfn))
3272 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3275 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3281 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3283 unsigned long pfn = pmd_pfn(pmd);
3285 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3287 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3290 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3293 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3299 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3300 struct pglist_data *pgdat, bool can_swap)
3302 struct folio *folio;
3304 /* try to avoid unnecessary memory loads */
3305 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3308 folio = pfn_folio(pfn);
3309 if (folio_nid(folio) != pgdat->node_id)
3312 if (folio_memcg_rcu(folio) != memcg)
3315 /* file VMAs can contain anon pages from COW */
3316 if (!folio_is_file_lru(folio) && !can_swap)
3322 static bool suitable_to_scan(int total, int young)
3324 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3326 /* suitable if the average number of young PTEs per cacheline is >=1 */
3327 return young * n >= total;
3330 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3331 struct mm_walk *args)
3339 struct lru_gen_mm_walk *walk = args->private;
3340 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3341 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3342 DEFINE_MAX_SEQ(walk->lruvec);
3343 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3345 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3348 if (!spin_trylock(ptl)) {
3353 arch_enter_lazy_mmu_mode();
3355 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3357 struct folio *folio;
3358 pte_t ptent = ptep_get(pte + i);
3361 walk->mm_stats[MM_LEAF_TOTAL]++;
3363 pfn = get_pte_pfn(ptent, args->vma, addr);
3367 if (!pte_young(ptent)) {
3368 walk->mm_stats[MM_LEAF_OLD]++;
3372 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3376 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3377 VM_WARN_ON_ONCE(true);
3380 walk->mm_stats[MM_LEAF_YOUNG]++;
3382 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3383 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3384 !folio_test_swapcache(folio)))
3385 folio_mark_dirty(folio);
3387 old_gen = folio_update_gen(folio, new_gen);
3388 if (old_gen >= 0 && old_gen != new_gen)
3389 update_batch_size(walk, folio, old_gen, new_gen);
3392 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3395 arch_leave_lazy_mmu_mode();
3396 pte_unmap_unlock(pte, ptl);
3398 return suitable_to_scan(total, young);
3401 static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3402 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3407 struct lru_gen_mm_walk *walk = args->private;
3408 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3409 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3410 DEFINE_MAX_SEQ(walk->lruvec);
3411 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3413 VM_WARN_ON_ONCE(pud_leaf(*pud));
3415 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3418 bitmap_zero(bitmap, MIN_LRU_BATCH);
3422 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3423 if (i && i <= MIN_LRU_BATCH) {
3424 __set_bit(i - 1, bitmap);
3428 pmd = pmd_offset(pud, *first);
3430 ptl = pmd_lockptr(args->mm, pmd);
3431 if (!spin_trylock(ptl))
3434 arch_enter_lazy_mmu_mode();
3438 struct folio *folio;
3440 /* don't round down the first address */
3441 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3443 pfn = get_pmd_pfn(pmd[i], vma, addr);
3447 if (!pmd_trans_huge(pmd[i])) {
3448 if (should_clear_pmd_young())
3449 pmdp_test_and_clear_young(vma, addr, pmd + i);
3453 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3457 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3460 walk->mm_stats[MM_LEAF_YOUNG]++;
3462 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3463 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3464 !folio_test_swapcache(folio)))
3465 folio_mark_dirty(folio);
3467 old_gen = folio_update_gen(folio, new_gen);
3468 if (old_gen >= 0 && old_gen != new_gen)
3469 update_batch_size(walk, folio, old_gen, new_gen);
3471 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3472 } while (i <= MIN_LRU_BATCH);
3474 arch_leave_lazy_mmu_mode();
3480 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3481 struct mm_walk *args)
3487 struct vm_area_struct *vma;
3488 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3489 unsigned long first = -1;
3490 struct lru_gen_mm_walk *walk = args->private;
3491 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3493 VM_WARN_ON_ONCE(pud_leaf(*pud));
3496 * Finish an entire PMD in two passes: the first only reaches to PTE
3497 * tables to avoid taking the PMD lock; the second, if necessary, takes
3498 * the PMD lock to clear the accessed bit in PMD entries.
3500 pmd = pmd_offset(pud, start & PUD_MASK);
3502 /* walk_pte_range() may call get_next_vma() */
3504 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3505 pmd_t val = pmdp_get_lockless(pmd + i);
3507 next = pmd_addr_end(addr, end);
3509 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3510 walk->mm_stats[MM_LEAF_TOTAL]++;
3514 if (pmd_trans_huge(val)) {
3515 unsigned long pfn = pmd_pfn(val);
3516 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3518 walk->mm_stats[MM_LEAF_TOTAL]++;
3520 if (!pmd_young(val)) {
3521 walk->mm_stats[MM_LEAF_OLD]++;
3525 /* try to avoid unnecessary memory loads */
3526 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3529 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3533 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3535 if (should_clear_pmd_young()) {
3536 if (!pmd_young(val))
3539 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3542 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3545 walk->mm_stats[MM_NONLEAF_FOUND]++;
3547 if (!walk_pte_range(&val, addr, next, args))
3550 walk->mm_stats[MM_NONLEAF_ADDED]++;
3552 /* carry over to the next generation */
3553 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3556 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3558 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3562 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3563 struct mm_walk *args)
3569 struct lru_gen_mm_walk *walk = args->private;
3571 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3573 pud = pud_offset(p4d, start & P4D_MASK);
3575 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3576 pud_t val = READ_ONCE(pud[i]);
3578 next = pud_addr_end(addr, end);
3580 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3583 walk_pmd_range(&val, addr, next, args);
3585 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3586 end = (addr | ~PUD_MASK) + 1;
3591 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3594 end = round_up(end, P4D_SIZE);
3596 if (!end || !args->vma)
3599 walk->next_addr = max(end, args->vma->vm_start);
3604 static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3606 static const struct mm_walk_ops mm_walk_ops = {
3607 .test_walk = should_skip_vma,
3608 .p4d_entry = walk_pud_range,
3609 .walk_lock = PGWALK_RDLOCK,
3613 struct lruvec *lruvec = walk->lruvec;
3614 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3616 walk->next_addr = FIRST_USER_ADDRESS;
3619 DEFINE_MAX_SEQ(lruvec);
3623 /* another thread might have called inc_max_seq() */
3624 if (walk->seq != max_seq)
3627 /* folio_update_gen() requires stable folio_memcg() */
3628 if (!mem_cgroup_trylock_pages(memcg))
3631 /* the caller might be holding the lock for write */
3632 if (mmap_read_trylock(mm)) {
3633 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3635 mmap_read_unlock(mm);
3638 mem_cgroup_unlock_pages();
3640 if (walk->batched) {
3641 spin_lock_irq(&lruvec->lru_lock);
3642 reset_batch_size(walk);
3643 spin_unlock_irq(&lruvec->lru_lock);
3647 } while (err == -EAGAIN);
3650 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3652 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3654 if (pgdat && current_is_kswapd()) {
3655 VM_WARN_ON_ONCE(walk);
3657 walk = &pgdat->mm_walk;
3658 } else if (!walk && force_alloc) {
3659 VM_WARN_ON_ONCE(current_is_kswapd());
3661 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3664 current->reclaim_state->mm_walk = walk;
3669 static void clear_mm_walk(void)
3671 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3673 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3674 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3676 current->reclaim_state->mm_walk = NULL;
3678 if (!current_is_kswapd())
3682 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3685 int remaining = MAX_LRU_BATCH;
3686 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3687 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3689 if (type == LRU_GEN_ANON && !can_swap)
3692 /* prevent cold/hot inversion if force_scan is true */
3693 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3694 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3696 while (!list_empty(head)) {
3697 struct folio *folio = lru_to_folio(head);
3699 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3700 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3701 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3702 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3704 new_gen = folio_inc_gen(lruvec, folio, false);
3705 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3712 reset_ctrl_pos(lruvec, type, true);
3713 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3718 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3720 int gen, type, zone;
3721 bool success = false;
3722 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3723 DEFINE_MIN_SEQ(lruvec);
3725 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3727 /* find the oldest populated generation */
3728 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3729 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3730 gen = lru_gen_from_seq(min_seq[type]);
3732 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3733 if (!list_empty(&lrugen->folios[gen][type][zone]))
3743 /* see the comment on lru_gen_folio */
3745 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3746 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3749 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3750 if (min_seq[type] == lrugen->min_seq[type])
3753 reset_ctrl_pos(lruvec, type, true);
3754 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3761 static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3762 bool can_swap, bool force_scan)
3767 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3769 if (seq < READ_ONCE(lrugen->max_seq))
3772 spin_lock_irq(&lruvec->lru_lock);
3774 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3776 success = seq == lrugen->max_seq;
3780 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3781 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3784 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3786 if (inc_min_seq(lruvec, type, can_swap))
3789 spin_unlock_irq(&lruvec->lru_lock);
3795 * Update the active/inactive LRU sizes for compatibility. Both sides of
3796 * the current max_seq need to be covered, since max_seq+1 can overlap
3797 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3798 * overlap, cold/hot inversion happens.
3800 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3801 next = lru_gen_from_seq(lrugen->max_seq + 1);
3803 for (type = 0; type < ANON_AND_FILE; type++) {
3804 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3805 enum lru_list lru = type * LRU_INACTIVE_FILE;
3806 long delta = lrugen->nr_pages[prev][type][zone] -
3807 lrugen->nr_pages[next][type][zone];
3812 __update_lru_size(lruvec, lru, zone, delta);
3813 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3817 for (type = 0; type < ANON_AND_FILE; type++)
3818 reset_ctrl_pos(lruvec, type, false);
3820 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3821 /* make sure preceding modifications appear */
3822 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3824 spin_unlock_irq(&lruvec->lru_lock);
3829 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3830 bool can_swap, bool force_scan)
3833 struct lru_gen_mm_walk *walk;
3834 struct mm_struct *mm = NULL;
3835 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3836 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3838 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
3841 return inc_max_seq(lruvec, seq, can_swap, force_scan);
3843 /* see the comment in iterate_mm_list() */
3844 if (seq <= READ_ONCE(mm_state->seq))
3848 * If the hardware doesn't automatically set the accessed bit, fallback
3849 * to lru_gen_look_around(), which only clears the accessed bit in a
3850 * handful of PTEs. Spreading the work out over a period of time usually
3851 * is less efficient, but it avoids bursty page faults.
3853 if (!should_walk_mmu()) {
3854 success = iterate_mm_list_nowalk(lruvec, seq);
3858 walk = set_mm_walk(NULL, true);
3860 success = iterate_mm_list_nowalk(lruvec, seq);
3864 walk->lruvec = lruvec;
3866 walk->can_swap = can_swap;
3867 walk->force_scan = force_scan;
3870 success = iterate_mm_list(walk, &mm);
3876 success = inc_max_seq(lruvec, seq, can_swap, force_scan);
3877 WARN_ON_ONCE(!success);
3883 /******************************************************************************
3884 * working set protection
3885 ******************************************************************************/
3887 static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3889 int gen, type, zone;
3890 unsigned long total = 0;
3891 bool can_swap = get_swappiness(lruvec, sc);
3892 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3893 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3894 DEFINE_MAX_SEQ(lruvec);
3895 DEFINE_MIN_SEQ(lruvec);
3897 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3900 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3901 gen = lru_gen_from_seq(seq);
3903 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3904 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3908 /* whether the size is big enough to be helpful */
3909 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3912 static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3913 unsigned long min_ttl)
3916 unsigned long birth;
3917 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3918 DEFINE_MIN_SEQ(lruvec);
3920 /* see the comment on lru_gen_folio */
3921 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3922 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3924 if (time_is_after_jiffies(birth + min_ttl))
3927 if (!lruvec_is_sizable(lruvec, sc))
3930 mem_cgroup_calculate_protection(NULL, memcg);
3932 return !mem_cgroup_below_min(NULL, memcg);
3935 /* to protect the working set of the last N jiffies */
3936 static unsigned long lru_gen_min_ttl __read_mostly;
3938 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3940 struct mem_cgroup *memcg;
3941 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3943 VM_WARN_ON_ONCE(!current_is_kswapd());
3945 /* check the order to exclude compaction-induced reclaim */
3946 if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
3949 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3951 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3953 if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
3954 mem_cgroup_iter_break(NULL, memcg);
3959 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
3962 * The main goal is to OOM kill if every generation from all memcgs is
3963 * younger than min_ttl. However, another possibility is all memcgs are
3964 * either too small or below min.
3966 if (mutex_trylock(&oom_lock)) {
3967 struct oom_control oc = {
3968 .gfp_mask = sc->gfp_mask,
3973 mutex_unlock(&oom_lock);
3977 /******************************************************************************
3978 * rmap/PT walk feedback
3979 ******************************************************************************/
3982 * This function exploits spatial locality when shrink_folio_list() walks the
3983 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3984 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3985 * the PTE table to the Bloom filter. This forms a feedback loop between the
3986 * eviction and the aging.
3988 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
3991 unsigned long start;
3993 struct lru_gen_mm_walk *walk;
3995 pte_t *pte = pvmw->pte;
3996 unsigned long addr = pvmw->address;
3997 struct vm_area_struct *vma = pvmw->vma;
3998 struct folio *folio = pfn_folio(pvmw->pfn);
3999 bool can_swap = !folio_is_file_lru(folio);
4000 struct mem_cgroup *memcg = folio_memcg(folio);
4001 struct pglist_data *pgdat = folio_pgdat(folio);
4002 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4003 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4004 DEFINE_MAX_SEQ(lruvec);
4005 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4007 lockdep_assert_held(pvmw->ptl);
4008 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4010 if (spin_is_contended(pvmw->ptl))
4013 /* exclude special VMAs containing anon pages from COW */
4014 if (vma->vm_flags & VM_SPECIAL)
4017 /* avoid taking the LRU lock under the PTL when possible */
4018 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4020 start = max(addr & PMD_MASK, vma->vm_start);
4021 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4023 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4024 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4025 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4026 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4027 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4029 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4030 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4034 /* folio_update_gen() requires stable folio_memcg() */
4035 if (!mem_cgroup_trylock_pages(memcg))
4038 arch_enter_lazy_mmu_mode();
4040 pte -= (addr - start) / PAGE_SIZE;
4042 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4044 pte_t ptent = ptep_get(pte + i);
4046 pfn = get_pte_pfn(ptent, vma, addr);
4050 if (!pte_young(ptent))
4053 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4057 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4058 VM_WARN_ON_ONCE(true);
4062 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4063 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4064 !folio_test_swapcache(folio)))
4065 folio_mark_dirty(folio);
4068 old_gen = folio_update_gen(folio, new_gen);
4069 if (old_gen >= 0 && old_gen != new_gen)
4070 update_batch_size(walk, folio, old_gen, new_gen);
4075 old_gen = folio_lru_gen(folio);
4077 folio_set_referenced(folio);
4078 else if (old_gen != new_gen)
4079 folio_activate(folio);
4082 arch_leave_lazy_mmu_mode();
4083 mem_cgroup_unlock_pages();
4085 /* feedback from rmap walkers to page table walkers */
4086 if (mm_state && suitable_to_scan(i, young))
4087 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4090 /******************************************************************************
4092 ******************************************************************************/
4094 /* see the comment on MEMCG_NR_GENS */
4103 static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4107 unsigned long flags;
4108 int bin = get_random_u32_below(MEMCG_NR_BINS);
4109 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4111 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4113 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4116 new = old = lruvec->lrugen.gen;
4118 /* see the comment on MEMCG_NR_GENS */
4119 if (op == MEMCG_LRU_HEAD)
4120 seg = MEMCG_LRU_HEAD;
4121 else if (op == MEMCG_LRU_TAIL)
4122 seg = MEMCG_LRU_TAIL;
4123 else if (op == MEMCG_LRU_OLD)
4124 new = get_memcg_gen(pgdat->memcg_lru.seq);
4125 else if (op == MEMCG_LRU_YOUNG)
4126 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4128 VM_WARN_ON_ONCE(true);
4130 WRITE_ONCE(lruvec->lrugen.seg, seg);
4131 WRITE_ONCE(lruvec->lrugen.gen, new);
4133 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4135 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4136 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4138 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4140 pgdat->memcg_lru.nr_memcgs[old]--;
4141 pgdat->memcg_lru.nr_memcgs[new]++;
4143 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4144 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4146 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4151 void lru_gen_online_memcg(struct mem_cgroup *memcg)
4155 int bin = get_random_u32_below(MEMCG_NR_BINS);
4157 for_each_node(nid) {
4158 struct pglist_data *pgdat = NODE_DATA(nid);
4159 struct lruvec *lruvec = get_lruvec(memcg, nid);
4161 spin_lock_irq(&pgdat->memcg_lru.lock);
4163 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4165 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4167 lruvec->lrugen.gen = gen;
4169 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4170 pgdat->memcg_lru.nr_memcgs[gen]++;
4172 spin_unlock_irq(&pgdat->memcg_lru.lock);
4176 void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4180 for_each_node(nid) {
4181 struct lruvec *lruvec = get_lruvec(memcg, nid);
4183 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4187 void lru_gen_release_memcg(struct mem_cgroup *memcg)
4192 for_each_node(nid) {
4193 struct pglist_data *pgdat = NODE_DATA(nid);
4194 struct lruvec *lruvec = get_lruvec(memcg, nid);
4196 spin_lock_irq(&pgdat->memcg_lru.lock);
4198 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4201 gen = lruvec->lrugen.gen;
4203 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4204 pgdat->memcg_lru.nr_memcgs[gen]--;
4206 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4207 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4209 spin_unlock_irq(&pgdat->memcg_lru.lock);
4213 void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4215 struct lruvec *lruvec = get_lruvec(memcg, nid);
4217 /* see the comment on MEMCG_NR_GENS */
4218 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4219 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4222 #endif /* CONFIG_MEMCG */
4224 /******************************************************************************
4226 ******************************************************************************/
4228 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4232 int gen = folio_lru_gen(folio);
4233 int type = folio_is_file_lru(folio);
4234 int zone = folio_zonenum(folio);
4235 int delta = folio_nr_pages(folio);
4236 int refs = folio_lru_refs(folio);
4237 int tier = lru_tier_from_refs(refs);
4238 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4240 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4243 if (!folio_evictable(folio)) {
4244 success = lru_gen_del_folio(lruvec, folio, true);
4245 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4246 folio_set_unevictable(folio);
4247 lruvec_add_folio(lruvec, folio);
4248 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4252 /* dirty lazyfree */
4253 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4254 success = lru_gen_del_folio(lruvec, folio, true);
4255 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4256 folio_set_swapbacked(folio);
4257 lruvec_add_folio_tail(lruvec, folio);
4262 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4263 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4268 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4269 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4271 gen = folio_inc_gen(lruvec, folio, false);
4272 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4274 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4275 lrugen->protected[hist][type][tier - 1] + delta);
4280 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4281 gen = folio_inc_gen(lruvec, folio, false);
4282 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4286 /* waiting for writeback */
4287 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4288 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4289 gen = folio_inc_gen(lruvec, folio, true);
4290 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4297 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4301 /* swap constrained */
4302 if (!(sc->gfp_mask & __GFP_IO) &&
4303 (folio_test_dirty(folio) ||
4304 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4307 /* raced with release_pages() */
4308 if (!folio_try_get(folio))
4311 /* raced with another isolation */
4312 if (!folio_test_clear_lru(folio)) {
4317 /* see the comment on MAX_NR_TIERS */
4318 if (!folio_test_referenced(folio))
4319 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4321 /* for shrink_folio_list() */
4322 folio_clear_reclaim(folio);
4323 folio_clear_referenced(folio);
4325 success = lru_gen_del_folio(lruvec, folio, true);
4326 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4331 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4332 int type, int tier, struct list_head *list)
4336 enum vm_event_item item;
4341 int remaining = MAX_LRU_BATCH;
4342 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4343 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4345 VM_WARN_ON_ONCE(!list_empty(list));
4347 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4350 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4352 for (i = MAX_NR_ZONES; i > 0; i--) {
4354 int skipped_zone = 0;
4355 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4356 struct list_head *head = &lrugen->folios[gen][type][zone];
4358 while (!list_empty(head)) {
4359 struct folio *folio = lru_to_folio(head);
4360 int delta = folio_nr_pages(folio);
4362 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4363 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4364 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4365 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4369 if (sort_folio(lruvec, folio, sc, tier))
4371 else if (isolate_folio(lruvec, folio, sc)) {
4372 list_add(&folio->lru, list);
4375 list_move(&folio->lru, &moved);
4376 skipped_zone += delta;
4379 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4384 list_splice(&moved, head);
4385 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4386 skipped += skipped_zone;
4389 if (!remaining || isolated >= MIN_LRU_BATCH)
4393 item = PGSCAN_KSWAPD + reclaimer_offset();
4394 if (!cgroup_reclaim(sc)) {
4395 __count_vm_events(item, isolated);
4396 __count_vm_events(PGREFILL, sorted);
4398 __count_memcg_events(memcg, item, isolated);
4399 __count_memcg_events(memcg, PGREFILL, sorted);
4400 __count_vm_events(PGSCAN_ANON + type, isolated);
4401 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4402 scanned, skipped, isolated,
4403 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4406 * There might not be eligible folios due to reclaim_idx. Check the
4407 * remaining to prevent livelock if it's not making progress.
4409 return isolated || !remaining ? scanned : 0;
4412 static int get_tier_idx(struct lruvec *lruvec, int type)
4415 struct ctrl_pos sp, pv;
4418 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4419 * This value is chosen because any other tier would have at least twice
4420 * as many refaults as the first tier.
4422 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4423 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4424 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4425 if (!positive_ctrl_err(&sp, &pv))
4432 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4435 struct ctrl_pos sp, pv;
4436 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4439 * Compare the first tier of anon with that of file to determine which
4440 * type to scan. Also need to compare other tiers of the selected type
4441 * with the first tier of the other type to determine the last tier (of
4442 * the selected type) to evict.
4444 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4445 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4446 type = positive_ctrl_err(&sp, &pv);
4448 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4449 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4450 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4451 if (!positive_ctrl_err(&sp, &pv))
4455 *tier_idx = tier - 1;
4460 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4461 int *type_scanned, struct list_head *list)
4467 DEFINE_MIN_SEQ(lruvec);
4470 * Try to make the obvious choice first, and if anon and file are both
4471 * available from the same generation,
4472 * 1. Interpret swappiness 1 as file first and MAX_SWAPPINESS as anon
4474 * 2. If !__GFP_IO, file first since clean pagecache is more likely to
4475 * exist than clean swapcache.
4478 type = LRU_GEN_FILE;
4479 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4480 type = LRU_GEN_ANON;
4481 else if (swappiness == 1)
4482 type = LRU_GEN_FILE;
4483 else if (swappiness == 200)
4484 type = LRU_GEN_ANON;
4485 else if (!(sc->gfp_mask & __GFP_IO))
4486 type = LRU_GEN_FILE;
4488 type = get_type_to_scan(lruvec, swappiness, &tier);
4490 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4492 tier = get_tier_idx(lruvec, type);
4494 scanned = scan_folios(lruvec, sc, type, tier, list);
4502 *type_scanned = type;
4507 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4514 struct folio *folio;
4516 enum vm_event_item item;
4517 struct reclaim_stat stat;
4518 struct lru_gen_mm_walk *walk;
4519 bool skip_retry = false;
4520 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4521 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4523 spin_lock_irq(&lruvec->lru_lock);
4525 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4527 scanned += try_to_inc_min_seq(lruvec, swappiness);
4529 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4532 spin_unlock_irq(&lruvec->lru_lock);
4534 if (list_empty(&list))
4537 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4538 sc->nr_reclaimed += reclaimed;
4539 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4540 scanned, reclaimed, &stat, sc->priority,
4541 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4543 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4544 if (!folio_evictable(folio)) {
4545 list_del(&folio->lru);
4546 folio_putback_lru(folio);
4550 if (folio_test_reclaim(folio) &&
4551 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4552 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4553 if (folio_test_workingset(folio))
4554 folio_set_referenced(folio);
4558 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4559 folio_mapped(folio) || folio_test_locked(folio) ||
4560 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4561 /* don't add rejected folios to the oldest generation */
4562 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4567 /* retry folios that may have missed folio_rotate_reclaimable() */
4568 list_move(&folio->lru, &clean);
4569 sc->nr_scanned -= folio_nr_pages(folio);
4572 spin_lock_irq(&lruvec->lru_lock);
4574 move_folios_to_lru(lruvec, &list);
4576 walk = current->reclaim_state->mm_walk;
4577 if (walk && walk->batched) {
4578 walk->lruvec = lruvec;
4579 reset_batch_size(walk);
4582 item = PGSTEAL_KSWAPD + reclaimer_offset();
4583 if (!cgroup_reclaim(sc))
4584 __count_vm_events(item, reclaimed);
4585 __count_memcg_events(memcg, item, reclaimed);
4586 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4588 spin_unlock_irq(&lruvec->lru_lock);
4590 list_splice_init(&clean, &list);
4592 if (!list_empty(&list)) {
4600 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4601 bool can_swap, unsigned long *nr_to_scan)
4603 int gen, type, zone;
4604 unsigned long old = 0;
4605 unsigned long young = 0;
4606 unsigned long total = 0;
4607 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4608 DEFINE_MIN_SEQ(lruvec);
4610 /* whether this lruvec is completely out of cold folios */
4611 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4616 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4619 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4620 unsigned long size = 0;
4622 gen = lru_gen_from_seq(seq);
4624 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4625 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4630 else if (seq + MIN_NR_GENS == max_seq)
4635 *nr_to_scan = total;
4638 * The aging tries to be lazy to reduce the overhead, while the eviction
4639 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4640 * ideal number of generations is MIN_NR_GENS+1.
4642 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4646 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4647 * of the total number of pages for each generation. A reasonable range
4648 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4649 * aging cares about the upper bound of hot pages, while the eviction
4650 * cares about the lower bound of cold pages.
4652 if (young * MIN_NR_GENS > total)
4654 if (old * (MIN_NR_GENS + 2) < total)
4661 * For future optimizations:
4662 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4665 static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4668 unsigned long nr_to_scan;
4669 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4670 DEFINE_MAX_SEQ(lruvec);
4672 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4675 success = should_run_aging(lruvec, max_seq, can_swap, &nr_to_scan);
4677 /* try to scrape all its memory if this memcg was deleted */
4678 if (nr_to_scan && !mem_cgroup_online(memcg))
4681 /* try to get away with not aging at the default priority */
4682 if (!success || sc->priority == DEF_PRIORITY)
4683 return nr_to_scan >> sc->priority;
4685 /* stop scanning this lruvec as it's low on cold folios */
4686 return try_to_inc_max_seq(lruvec, max_seq, can_swap, false) ? -1 : 0;
4689 static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4692 enum zone_watermarks mark;
4694 /* don't abort memcg reclaim to ensure fairness */
4695 if (!root_reclaim(sc))
4698 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4701 /* check the order to exclude compaction-induced reclaim */
4702 if (!current_is_kswapd() || sc->order)
4705 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4706 WMARK_PROMO : WMARK_HIGH;
4708 for (i = 0; i <= sc->reclaim_idx; i++) {
4709 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4710 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4712 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4716 /* kswapd should abort if all eligible zones are safe */
4720 static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4723 unsigned long scanned = 0;
4724 int swappiness = get_swappiness(lruvec, sc);
4729 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4730 if (nr_to_scan <= 0)
4733 delta = evict_folios(lruvec, sc, swappiness);
4738 if (scanned >= nr_to_scan)
4741 if (should_abort_scan(lruvec, sc))
4747 /* whether this lruvec should be rotated */
4748 return nr_to_scan < 0;
4751 static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4754 unsigned long scanned = sc->nr_scanned;
4755 unsigned long reclaimed = sc->nr_reclaimed;
4756 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4757 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4759 mem_cgroup_calculate_protection(NULL, memcg);
4761 if (mem_cgroup_below_min(NULL, memcg))
4762 return MEMCG_LRU_YOUNG;
4764 if (mem_cgroup_below_low(NULL, memcg)) {
4765 /* see the comment on MEMCG_NR_GENS */
4766 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4767 return MEMCG_LRU_TAIL;
4769 memcg_memory_event(memcg, MEMCG_LOW);
4772 success = try_to_shrink_lruvec(lruvec, sc);
4774 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4777 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4778 sc->nr_reclaimed - reclaimed);
4780 flush_reclaim_state(sc);
4782 if (success && mem_cgroup_online(memcg))
4783 return MEMCG_LRU_YOUNG;
4785 if (!success && lruvec_is_sizable(lruvec, sc))
4788 /* one retry if offlined or too small */
4789 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4790 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4793 static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4799 struct lruvec *lruvec;
4800 struct lru_gen_folio *lrugen;
4801 struct mem_cgroup *memcg;
4802 struct hlist_nulls_node *pos;
4804 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4805 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4812 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4814 lru_gen_rotate_memcg(lruvec, op);
4818 mem_cgroup_put(memcg);
4821 if (gen != READ_ONCE(lrugen->gen))
4824 lruvec = container_of(lrugen, struct lruvec, lrugen);
4825 memcg = lruvec_memcg(lruvec);
4827 if (!mem_cgroup_tryget(memcg)) {
4828 lru_gen_release_memcg(memcg);
4835 op = shrink_one(lruvec, sc);
4839 if (should_abort_scan(lruvec, sc))
4846 lru_gen_rotate_memcg(lruvec, op);
4848 mem_cgroup_put(memcg);
4850 if (!is_a_nulls(pos))
4853 /* restart if raced with lru_gen_rotate_memcg() */
4854 if (gen != get_nulls_value(pos))
4857 /* try the rest of the bins of the current generation */
4858 bin = get_memcg_bin(bin + 1);
4859 if (bin != first_bin)
4863 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4865 struct blk_plug plug;
4867 VM_WARN_ON_ONCE(root_reclaim(sc));
4868 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4872 blk_start_plug(&plug);
4874 set_mm_walk(NULL, sc->proactive);
4876 if (try_to_shrink_lruvec(lruvec, sc))
4877 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4881 blk_finish_plug(&plug);
4884 static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4887 unsigned long reclaimable;
4889 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4892 * Determine the initial priority based on
4893 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4894 * where reclaimed_to_scanned_ratio = inactive / total.
4896 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4897 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
4898 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4900 /* round down reclaimable and round up sc->nr_to_reclaim */
4901 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4903 sc->priority = clamp(priority, 0, DEF_PRIORITY);
4906 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4908 struct blk_plug plug;
4909 unsigned long reclaimed = sc->nr_reclaimed;
4911 VM_WARN_ON_ONCE(!root_reclaim(sc));
4914 * Unmapped clean folios are already prioritized. Scanning for more of
4915 * them is likely futile and can cause high reclaim latency when there
4916 * is a large number of memcgs.
4918 if (!sc->may_writepage || !sc->may_unmap)
4923 blk_start_plug(&plug);
4925 set_mm_walk(pgdat, sc->proactive);
4927 set_initial_priority(pgdat, sc);
4929 if (current_is_kswapd())
4930 sc->nr_reclaimed = 0;
4932 if (mem_cgroup_disabled())
4933 shrink_one(&pgdat->__lruvec, sc);
4935 shrink_many(pgdat, sc);
4937 if (current_is_kswapd())
4938 sc->nr_reclaimed += reclaimed;
4942 blk_finish_plug(&plug);
4944 /* kswapd should never fail */
4945 pgdat->kswapd_failures = 0;
4948 /******************************************************************************
4950 ******************************************************************************/
4952 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4954 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4956 if (lrugen->enabled) {
4959 for_each_evictable_lru(lru) {
4960 if (!list_empty(&lruvec->lists[lru]))
4964 int gen, type, zone;
4966 for_each_gen_type_zone(gen, type, zone) {
4967 if (!list_empty(&lrugen->folios[gen][type][zone]))
4975 static bool fill_evictable(struct lruvec *lruvec)
4978 int remaining = MAX_LRU_BATCH;
4980 for_each_evictable_lru(lru) {
4981 int type = is_file_lru(lru);
4982 bool active = is_active_lru(lru);
4983 struct list_head *head = &lruvec->lists[lru];
4985 while (!list_empty(head)) {
4987 struct folio *folio = lru_to_folio(head);
4989 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4990 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
4991 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4992 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
4994 lruvec_del_folio(lruvec, folio);
4995 success = lru_gen_add_folio(lruvec, folio, false);
4996 VM_WARN_ON_ONCE(!success);
5006 static bool drain_evictable(struct lruvec *lruvec)
5008 int gen, type, zone;
5009 int remaining = MAX_LRU_BATCH;
5011 for_each_gen_type_zone(gen, type, zone) {
5012 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5014 while (!list_empty(head)) {
5016 struct folio *folio = lru_to_folio(head);
5018 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5019 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5020 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5021 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5023 success = lru_gen_del_folio(lruvec, folio, false);
5024 VM_WARN_ON_ONCE(!success);
5025 lruvec_add_folio(lruvec, folio);
5035 static void lru_gen_change_state(bool enabled)
5037 static DEFINE_MUTEX(state_mutex);
5039 struct mem_cgroup *memcg;
5044 mutex_lock(&state_mutex);
5046 if (enabled == lru_gen_enabled())
5050 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5052 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5054 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5058 for_each_node(nid) {
5059 struct lruvec *lruvec = get_lruvec(memcg, nid);
5061 spin_lock_irq(&lruvec->lru_lock);
5063 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5064 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5066 lruvec->lrugen.enabled = enabled;
5068 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5069 spin_unlock_irq(&lruvec->lru_lock);
5071 spin_lock_irq(&lruvec->lru_lock);
5074 spin_unlock_irq(&lruvec->lru_lock);
5078 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5080 mutex_unlock(&state_mutex);
5086 /******************************************************************************
5088 ******************************************************************************/
5090 static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5092 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5095 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5096 static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5097 const char *buf, size_t len)
5101 if (kstrtouint(buf, 0, &msecs))
5104 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5109 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5111 static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5113 unsigned int caps = 0;
5115 if (get_cap(LRU_GEN_CORE))
5116 caps |= BIT(LRU_GEN_CORE);
5118 if (should_walk_mmu())
5119 caps |= BIT(LRU_GEN_MM_WALK);
5121 if (should_clear_pmd_young())
5122 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5124 return sysfs_emit(buf, "0x%04x\n", caps);
5127 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5128 static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5129 const char *buf, size_t len)
5134 if (tolower(*buf) == 'n')
5136 else if (tolower(*buf) == 'y')
5138 else if (kstrtouint(buf, 0, &caps))
5141 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5142 bool enabled = caps & BIT(i);
5144 if (i == LRU_GEN_CORE)
5145 lru_gen_change_state(enabled);
5147 static_branch_enable(&lru_gen_caps[i]);
5149 static_branch_disable(&lru_gen_caps[i]);
5155 static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5157 static struct attribute *lru_gen_attrs[] = {
5158 &lru_gen_min_ttl_attr.attr,
5159 &lru_gen_enabled_attr.attr,
5163 static const struct attribute_group lru_gen_attr_group = {
5165 .attrs = lru_gen_attrs,
5168 /******************************************************************************
5170 ******************************************************************************/
5172 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5174 struct mem_cgroup *memcg;
5175 loff_t nr_to_skip = *pos;
5177 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5179 return ERR_PTR(-ENOMEM);
5181 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5185 for_each_node_state(nid, N_MEMORY) {
5187 return get_lruvec(memcg, nid);
5189 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5194 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5196 if (!IS_ERR_OR_NULL(v))
5197 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5203 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5205 int nid = lruvec_pgdat(v)->node_id;
5206 struct mem_cgroup *memcg = lruvec_memcg(v);
5210 nid = next_memory_node(nid);
5211 if (nid == MAX_NUMNODES) {
5212 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5216 nid = first_memory_node;
5219 return get_lruvec(memcg, nid);
5222 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5223 unsigned long max_seq, unsigned long *min_seq,
5228 int hist = lru_hist_from_seq(seq);
5229 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5230 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5232 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5233 seq_printf(m, " %10d", tier);
5234 for (type = 0; type < ANON_AND_FILE; type++) {
5235 const char *s = " ";
5236 unsigned long n[3] = {};
5238 if (seq == max_seq) {
5240 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5241 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5242 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5244 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5245 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5247 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5250 for (i = 0; i < 3; i++)
5251 seq_printf(m, " %10lu%c", n[i], s[i]);
5260 for (i = 0; i < NR_MM_STATS; i++) {
5261 const char *s = " ";
5262 unsigned long n = 0;
5264 if (seq == max_seq && NR_HIST_GENS == 1) {
5266 n = READ_ONCE(mm_state->stats[hist][i]);
5267 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5269 n = READ_ONCE(mm_state->stats[hist][i]);
5272 seq_printf(m, " %10lu%c", n, s[i]);
5277 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5278 static int lru_gen_seq_show(struct seq_file *m, void *v)
5281 bool full = !debugfs_real_fops(m->file)->write;
5282 struct lruvec *lruvec = v;
5283 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5284 int nid = lruvec_pgdat(lruvec)->node_id;
5285 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5286 DEFINE_MAX_SEQ(lruvec);
5287 DEFINE_MIN_SEQ(lruvec);
5289 if (nid == first_memory_node) {
5290 const char *path = memcg ? m->private : "";
5294 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5296 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5299 seq_printf(m, " node %5d\n", nid);
5302 seq = min_seq[LRU_GEN_ANON];
5303 else if (max_seq >= MAX_NR_GENS)
5304 seq = max_seq - MAX_NR_GENS + 1;
5308 for (; seq <= max_seq; seq++) {
5310 int gen = lru_gen_from_seq(seq);
5311 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5313 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5315 for (type = 0; type < ANON_AND_FILE; type++) {
5316 unsigned long size = 0;
5317 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5319 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5320 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5322 seq_printf(m, " %10lu%c", size, mark);
5328 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5334 static const struct seq_operations lru_gen_seq_ops = {
5335 .start = lru_gen_seq_start,
5336 .stop = lru_gen_seq_stop,
5337 .next = lru_gen_seq_next,
5338 .show = lru_gen_seq_show,
5341 static int run_aging(struct lruvec *lruvec, unsigned long seq,
5342 bool can_swap, bool force_scan)
5344 DEFINE_MAX_SEQ(lruvec);
5345 DEFINE_MIN_SEQ(lruvec);
5353 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5356 try_to_inc_max_seq(lruvec, max_seq, can_swap, force_scan);
5361 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5362 int swappiness, unsigned long nr_to_reclaim)
5364 DEFINE_MAX_SEQ(lruvec);
5366 if (seq + MIN_NR_GENS > max_seq)
5369 sc->nr_reclaimed = 0;
5371 while (!signal_pending(current)) {
5372 DEFINE_MIN_SEQ(lruvec);
5374 if (seq < min_seq[!swappiness])
5377 if (sc->nr_reclaimed >= nr_to_reclaim)
5380 if (!evict_folios(lruvec, sc, swappiness))
5389 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5390 struct scan_control *sc, int swappiness, unsigned long opt)
5392 struct lruvec *lruvec;
5394 struct mem_cgroup *memcg = NULL;
5396 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5399 if (!mem_cgroup_disabled()) {
5402 memcg = mem_cgroup_from_id(memcg_id);
5403 if (!mem_cgroup_tryget(memcg))
5412 if (memcg_id != mem_cgroup_id(memcg))
5415 lruvec = get_lruvec(memcg, nid);
5418 swappiness = get_swappiness(lruvec, sc);
5419 else if (swappiness > 200)
5424 err = run_aging(lruvec, seq, swappiness, opt);
5427 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5431 mem_cgroup_put(memcg);
5436 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5437 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5438 size_t len, loff_t *pos)
5443 struct blk_plug plug;
5445 struct scan_control sc = {
5446 .may_writepage = true,
5449 .reclaim_idx = MAX_NR_ZONES - 1,
5450 .gfp_mask = GFP_KERNEL,
5453 buf = kvmalloc(len + 1, GFP_KERNEL);
5457 if (copy_from_user(buf, src, len)) {
5462 set_task_reclaim_state(current, &sc.reclaim_state);
5463 flags = memalloc_noreclaim_save();
5464 blk_start_plug(&plug);
5465 if (!set_mm_walk(NULL, true)) {
5473 while ((cur = strsep(&next, ",;\n"))) {
5477 unsigned int memcg_id;
5480 unsigned int swappiness = -1;
5481 unsigned long opt = -1;
5483 cur = skip_spaces(cur);
5487 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5488 &seq, &end, &swappiness, &end, &opt, &end);
5489 if (n < 4 || cur[end]) {
5494 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5500 blk_finish_plug(&plug);
5501 memalloc_noreclaim_restore(flags);
5502 set_task_reclaim_state(current, NULL);
5509 static int lru_gen_seq_open(struct inode *inode, struct file *file)
5511 return seq_open(file, &lru_gen_seq_ops);
5514 static const struct file_operations lru_gen_rw_fops = {
5515 .open = lru_gen_seq_open,
5517 .write = lru_gen_seq_write,
5518 .llseek = seq_lseek,
5519 .release = seq_release,
5522 static const struct file_operations lru_gen_ro_fops = {
5523 .open = lru_gen_seq_open,
5525 .llseek = seq_lseek,
5526 .release = seq_release,
5529 /******************************************************************************
5531 ******************************************************************************/
5533 void lru_gen_init_pgdat(struct pglist_data *pgdat)
5537 spin_lock_init(&pgdat->memcg_lru.lock);
5539 for (i = 0; i < MEMCG_NR_GENS; i++) {
5540 for (j = 0; j < MEMCG_NR_BINS; j++)
5541 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5545 void lru_gen_init_lruvec(struct lruvec *lruvec)
5548 int gen, type, zone;
5549 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5550 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5552 lrugen->max_seq = MIN_NR_GENS + 1;
5553 lrugen->enabled = lru_gen_enabled();
5555 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5556 lrugen->timestamps[i] = jiffies;
5558 for_each_gen_type_zone(gen, type, zone)
5559 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5562 mm_state->seq = MIN_NR_GENS;
5567 void lru_gen_init_memcg(struct mem_cgroup *memcg)
5569 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5574 INIT_LIST_HEAD(&mm_list->fifo);
5575 spin_lock_init(&mm_list->lock);
5578 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5582 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5584 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5586 for_each_node(nid) {
5587 struct lruvec *lruvec = get_lruvec(memcg, nid);
5588 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5590 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5591 sizeof(lruvec->lrugen.nr_pages)));
5593 lruvec->lrugen.list.next = LIST_POISON1;
5598 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5599 bitmap_free(mm_state->filters[i]);
5600 mm_state->filters[i] = NULL;
5605 #endif /* CONFIG_MEMCG */
5607 static int __init init_lru_gen(void)
5609 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5610 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5612 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5613 pr_err("lru_gen: failed to create sysfs group\n");
5615 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5616 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5620 late_initcall(init_lru_gen);
5622 #else /* !CONFIG_LRU_GEN */
5624 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5629 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5634 static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5639 #endif /* CONFIG_LRU_GEN */
5641 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5643 unsigned long nr[NR_LRU_LISTS];
5644 unsigned long targets[NR_LRU_LISTS];
5645 unsigned long nr_to_scan;
5647 unsigned long nr_reclaimed = 0;
5648 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5649 bool proportional_reclaim;
5650 struct blk_plug plug;
5652 if (lru_gen_enabled() && !root_reclaim(sc)) {
5653 lru_gen_shrink_lruvec(lruvec, sc);
5657 get_scan_count(lruvec, sc, nr);
5659 /* Record the original scan target for proportional adjustments later */
5660 memcpy(targets, nr, sizeof(nr));
5663 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5664 * event that can occur when there is little memory pressure e.g.
5665 * multiple streaming readers/writers. Hence, we do not abort scanning
5666 * when the requested number of pages are reclaimed when scanning at
5667 * DEF_PRIORITY on the assumption that the fact we are direct
5668 * reclaiming implies that kswapd is not keeping up and it is best to
5669 * do a batch of work at once. For memcg reclaim one check is made to
5670 * abort proportional reclaim if either the file or anon lru has already
5671 * dropped to zero at the first pass.
5673 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5674 sc->priority == DEF_PRIORITY);
5676 blk_start_plug(&plug);
5677 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5678 nr[LRU_INACTIVE_FILE]) {
5679 unsigned long nr_anon, nr_file, percentage;
5680 unsigned long nr_scanned;
5682 for_each_evictable_lru(lru) {
5684 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5685 nr[lru] -= nr_to_scan;
5687 nr_reclaimed += shrink_list(lru, nr_to_scan,
5694 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5698 * For kswapd and memcg, reclaim at least the number of pages
5699 * requested. Ensure that the anon and file LRUs are scanned
5700 * proportionally what was requested by get_scan_count(). We
5701 * stop reclaiming one LRU and reduce the amount scanning
5702 * proportional to the original scan target.
5704 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5705 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5708 * It's just vindictive to attack the larger once the smaller
5709 * has gone to zero. And given the way we stop scanning the
5710 * smaller below, this makes sure that we only make one nudge
5711 * towards proportionality once we've got nr_to_reclaim.
5713 if (!nr_file || !nr_anon)
5716 if (nr_file > nr_anon) {
5717 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5718 targets[LRU_ACTIVE_ANON] + 1;
5720 percentage = nr_anon * 100 / scan_target;
5722 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5723 targets[LRU_ACTIVE_FILE] + 1;
5725 percentage = nr_file * 100 / scan_target;
5728 /* Stop scanning the smaller of the LRU */
5730 nr[lru + LRU_ACTIVE] = 0;
5733 * Recalculate the other LRU scan count based on its original
5734 * scan target and the percentage scanning already complete
5736 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5737 nr_scanned = targets[lru] - nr[lru];
5738 nr[lru] = targets[lru] * (100 - percentage) / 100;
5739 nr[lru] -= min(nr[lru], nr_scanned);
5742 nr_scanned = targets[lru] - nr[lru];
5743 nr[lru] = targets[lru] * (100 - percentage) / 100;
5744 nr[lru] -= min(nr[lru], nr_scanned);
5746 blk_finish_plug(&plug);
5747 sc->nr_reclaimed += nr_reclaimed;
5750 * Even if we did not try to evict anon pages at all, we want to
5751 * rebalance the anon lru active/inactive ratio.
5753 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5754 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5755 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5756 sc, LRU_ACTIVE_ANON);
5759 /* Use reclaim/compaction for costly allocs or under memory pressure */
5760 static bool in_reclaim_compaction(struct scan_control *sc)
5762 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5763 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5764 sc->priority < DEF_PRIORITY - 2))
5771 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5772 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5773 * true if more pages should be reclaimed such that when the page allocator
5774 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5775 * It will give up earlier than that if there is difficulty reclaiming pages.
5777 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5778 unsigned long nr_reclaimed,
5779 struct scan_control *sc)
5781 unsigned long pages_for_compaction;
5782 unsigned long inactive_lru_pages;
5785 /* If not in reclaim/compaction mode, stop */
5786 if (!in_reclaim_compaction(sc))
5790 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5791 * number of pages that were scanned. This will return to the caller
5792 * with the risk reclaim/compaction and the resulting allocation attempt
5793 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5794 * allocations through requiring that the full LRU list has been scanned
5795 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5796 * scan, but that approximation was wrong, and there were corner cases
5797 * where always a non-zero amount of pages were scanned.
5802 /* If compaction would go ahead or the allocation would succeed, stop */
5803 for (z = 0; z <= sc->reclaim_idx; z++) {
5804 struct zone *zone = &pgdat->node_zones[z];
5805 if (!managed_zone(zone))
5808 /* Allocation can already succeed, nothing to do */
5809 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5810 sc->reclaim_idx, 0))
5813 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5818 * If we have not reclaimed enough pages for compaction and the
5819 * inactive lists are large enough, continue reclaiming
5821 pages_for_compaction = compact_gap(sc->order);
5822 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5823 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5824 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5826 return inactive_lru_pages > pages_for_compaction;
5829 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5831 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5832 struct mem_cgroup *memcg;
5834 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
5836 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5837 unsigned long reclaimed;
5838 unsigned long scanned;
5841 * This loop can become CPU-bound when target memcgs
5842 * aren't eligible for reclaim - either because they
5843 * don't have any reclaimable pages, or because their
5844 * memory is explicitly protected. Avoid soft lockups.
5848 mem_cgroup_calculate_protection(target_memcg, memcg);
5850 if (mem_cgroup_below_min(target_memcg, memcg)) {
5853 * If there is no reclaimable memory, OOM.
5856 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5859 * Respect the protection only as long as
5860 * there is an unprotected supply
5861 * of reclaimable memory from other cgroups.
5863 if (!sc->memcg_low_reclaim) {
5864 sc->memcg_low_skipped = 1;
5867 memcg_memory_event(memcg, MEMCG_LOW);
5870 reclaimed = sc->nr_reclaimed;
5871 scanned = sc->nr_scanned;
5873 shrink_lruvec(lruvec, sc);
5875 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5878 /* Record the group's reclaim efficiency */
5880 vmpressure(sc->gfp_mask, memcg, false,
5881 sc->nr_scanned - scanned,
5882 sc->nr_reclaimed - reclaimed);
5884 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5887 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5889 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5890 struct lruvec *target_lruvec;
5891 bool reclaimable = false;
5893 if (lru_gen_enabled() && root_reclaim(sc)) {
5894 lru_gen_shrink_node(pgdat, sc);
5898 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5901 memset(&sc->nr, 0, sizeof(sc->nr));
5903 nr_reclaimed = sc->nr_reclaimed;
5904 nr_scanned = sc->nr_scanned;
5906 prepare_scan_control(pgdat, sc);
5908 shrink_node_memcgs(pgdat, sc);
5910 flush_reclaim_state(sc);
5912 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5914 /* Record the subtree's reclaim efficiency */
5916 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5917 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5919 if (nr_node_reclaimed)
5922 if (current_is_kswapd()) {
5924 * If reclaim is isolating dirty pages under writeback,
5925 * it implies that the long-lived page allocation rate
5926 * is exceeding the page laundering rate. Either the
5927 * global limits are not being effective at throttling
5928 * processes due to the page distribution throughout
5929 * zones or there is heavy usage of a slow backing
5930 * device. The only option is to throttle from reclaim
5931 * context which is not ideal as there is no guarantee
5932 * the dirtying process is throttled in the same way
5933 * balance_dirty_pages() manages.
5935 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5936 * count the number of pages under pages flagged for
5937 * immediate reclaim and stall if any are encountered
5938 * in the nr_immediate check below.
5940 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5941 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5943 /* Allow kswapd to start writing pages during reclaim.*/
5944 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5945 set_bit(PGDAT_DIRTY, &pgdat->flags);
5948 * If kswapd scans pages marked for immediate
5949 * reclaim and under writeback (nr_immediate), it
5950 * implies that pages are cycling through the LRU
5951 * faster than they are written so forcibly stall
5952 * until some pages complete writeback.
5954 if (sc->nr.immediate)
5955 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5959 * Tag a node/memcg as congested if all the dirty pages were marked
5960 * for writeback and immediate reclaim (counted in nr.congested).
5962 * Legacy memcg will stall in page writeback so avoid forcibly
5963 * stalling in reclaim_throttle().
5965 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
5966 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
5967 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
5969 if (current_is_kswapd())
5970 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
5974 * Stall direct reclaim for IO completions if the lruvec is
5975 * node is congested. Allow kswapd to continue until it
5976 * starts encountering unqueued dirty pages or cycling through
5977 * the LRU too quickly.
5979 if (!current_is_kswapd() && current_may_throttle() &&
5980 !sc->hibernation_mode &&
5981 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
5982 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
5983 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
5985 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
5989 * Kswapd gives up on balancing particular nodes after too
5990 * many failures to reclaim anything from them and goes to
5991 * sleep. On reclaim progress, reset the failure counter. A
5992 * successful direct reclaim run will revive a dormant kswapd.
5995 pgdat->kswapd_failures = 0;
5996 else if (sc->cache_trim_mode)
5997 sc->cache_trim_mode_failed = 1;
6001 * Returns true if compaction should go ahead for a costly-order request, or
6002 * the allocation would already succeed without compaction. Return false if we
6003 * should reclaim first.
6005 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6007 unsigned long watermark;
6009 if (!gfp_compaction_allowed(sc->gfp_mask))
6012 /* Allocation can already succeed, nothing to do */
6013 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6014 sc->reclaim_idx, 0))
6017 /* Compaction cannot yet proceed. Do reclaim. */
6018 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6022 * Compaction is already possible, but it takes time to run and there
6023 * are potentially other callers using the pages just freed. So proceed
6024 * with reclaim to make a buffer of free pages available to give
6025 * compaction a reasonable chance of completing and allocating the page.
6026 * Note that we won't actually reclaim the whole buffer in one attempt
6027 * as the target watermark in should_continue_reclaim() is lower. But if
6028 * we are already above the high+gap watermark, don't reclaim at all.
6030 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6032 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6035 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6038 * If reclaim is making progress greater than 12% efficiency then
6039 * wake all the NOPROGRESS throttled tasks.
6041 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6042 wait_queue_head_t *wqh;
6044 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6045 if (waitqueue_active(wqh))
6052 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6053 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6054 * under writeback and marked for immediate reclaim at the tail of the
6057 if (current_is_kswapd() || cgroup_reclaim(sc))
6060 /* Throttle if making no progress at high prioities. */
6061 if (sc->priority == 1 && !sc->nr_reclaimed)
6062 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6066 * This is the direct reclaim path, for page-allocating processes. We only
6067 * try to reclaim pages from zones which will satisfy the caller's allocation
6070 * If a zone is deemed to be full of pinned pages then just give it a light
6071 * scan then give up on it.
6073 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6077 unsigned long nr_soft_reclaimed;
6078 unsigned long nr_soft_scanned;
6080 pg_data_t *last_pgdat = NULL;
6081 pg_data_t *first_pgdat = NULL;
6084 * If the number of buffer_heads in the machine exceeds the maximum
6085 * allowed level, force direct reclaim to scan the highmem zone as
6086 * highmem pages could be pinning lowmem pages storing buffer_heads
6088 orig_mask = sc->gfp_mask;
6089 if (buffer_heads_over_limit) {
6090 sc->gfp_mask |= __GFP_HIGHMEM;
6091 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6094 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6095 sc->reclaim_idx, sc->nodemask) {
6097 * Take care memory controller reclaiming has small influence
6100 if (!cgroup_reclaim(sc)) {
6101 if (!cpuset_zone_allowed(zone,
6102 GFP_KERNEL | __GFP_HARDWALL))
6106 * If we already have plenty of memory free for
6107 * compaction in this zone, don't free any more.
6108 * Even though compaction is invoked for any
6109 * non-zero order, only frequent costly order
6110 * reclamation is disruptive enough to become a
6111 * noticeable problem, like transparent huge
6114 if (IS_ENABLED(CONFIG_COMPACTION) &&
6115 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6116 compaction_ready(zone, sc)) {
6117 sc->compaction_ready = true;
6122 * Shrink each node in the zonelist once. If the
6123 * zonelist is ordered by zone (not the default) then a
6124 * node may be shrunk multiple times but in that case
6125 * the user prefers lower zones being preserved.
6127 if (zone->zone_pgdat == last_pgdat)
6131 * This steals pages from memory cgroups over softlimit
6132 * and returns the number of reclaimed pages and
6133 * scanned pages. This works for global memory pressure
6134 * and balancing, not for a memcg's limit.
6136 nr_soft_scanned = 0;
6137 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6138 sc->order, sc->gfp_mask,
6140 sc->nr_reclaimed += nr_soft_reclaimed;
6141 sc->nr_scanned += nr_soft_scanned;
6142 /* need some check for avoid more shrink_zone() */
6146 first_pgdat = zone->zone_pgdat;
6148 /* See comment about same check for global reclaim above */
6149 if (zone->zone_pgdat == last_pgdat)
6151 last_pgdat = zone->zone_pgdat;
6152 shrink_node(zone->zone_pgdat, sc);
6156 consider_reclaim_throttle(first_pgdat, sc);
6159 * Restore to original mask to avoid the impact on the caller if we
6160 * promoted it to __GFP_HIGHMEM.
6162 sc->gfp_mask = orig_mask;
6165 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6167 struct lruvec *target_lruvec;
6168 unsigned long refaults;
6170 if (lru_gen_enabled())
6173 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6174 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6175 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6176 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6177 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6181 * This is the main entry point to direct page reclaim.
6183 * If a full scan of the inactive list fails to free enough memory then we
6184 * are "out of memory" and something needs to be killed.
6186 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6187 * high - the zone may be full of dirty or under-writeback pages, which this
6188 * caller can't do much about. We kick the writeback threads and take explicit
6189 * naps in the hope that some of these pages can be written. But if the
6190 * allocating task holds filesystem locks which prevent writeout this might not
6191 * work, and the allocation attempt will fail.
6193 * returns: 0, if no pages reclaimed
6194 * else, the number of pages reclaimed
6196 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6197 struct scan_control *sc)
6199 int initial_priority = sc->priority;
6200 pg_data_t *last_pgdat;
6204 delayacct_freepages_start();
6206 if (!cgroup_reclaim(sc))
6207 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6211 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6214 shrink_zones(zonelist, sc);
6216 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6219 if (sc->compaction_ready)
6223 * If we're getting trouble reclaiming, start doing
6224 * writepage even in laptop mode.
6226 if (sc->priority < DEF_PRIORITY - 2)
6227 sc->may_writepage = 1;
6228 } while (--sc->priority >= 0);
6231 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6233 if (zone->zone_pgdat == last_pgdat)
6235 last_pgdat = zone->zone_pgdat;
6237 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6239 if (cgroup_reclaim(sc)) {
6240 struct lruvec *lruvec;
6242 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6244 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6248 delayacct_freepages_end();
6250 if (sc->nr_reclaimed)
6251 return sc->nr_reclaimed;
6253 /* Aborted reclaim to try compaction? don't OOM, then */
6254 if (sc->compaction_ready)
6258 * We make inactive:active ratio decisions based on the node's
6259 * composition of memory, but a restrictive reclaim_idx or a
6260 * memory.low cgroup setting can exempt large amounts of
6261 * memory from reclaim. Neither of which are very common, so
6262 * instead of doing costly eligibility calculations of the
6263 * entire cgroup subtree up front, we assume the estimates are
6264 * good, and retry with forcible deactivation if that fails.
6266 if (sc->skipped_deactivate) {
6267 sc->priority = initial_priority;
6268 sc->force_deactivate = 1;
6269 sc->skipped_deactivate = 0;
6273 /* Untapped cgroup reserves? Don't OOM, retry. */
6274 if (sc->memcg_low_skipped) {
6275 sc->priority = initial_priority;
6276 sc->force_deactivate = 0;
6277 sc->memcg_low_reclaim = 1;
6278 sc->memcg_low_skipped = 0;
6285 static bool allow_direct_reclaim(pg_data_t *pgdat)
6288 unsigned long pfmemalloc_reserve = 0;
6289 unsigned long free_pages = 0;
6293 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6296 for (i = 0; i <= ZONE_NORMAL; i++) {
6297 zone = &pgdat->node_zones[i];
6298 if (!managed_zone(zone))
6301 if (!zone_reclaimable_pages(zone))
6304 pfmemalloc_reserve += min_wmark_pages(zone);
6305 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6308 /* If there are no reserves (unexpected config) then do not throttle */
6309 if (!pfmemalloc_reserve)
6312 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6314 /* kswapd must be awake if processes are being throttled */
6315 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6316 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6317 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6319 wake_up_interruptible(&pgdat->kswapd_wait);
6326 * Throttle direct reclaimers if backing storage is backed by the network
6327 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6328 * depleted. kswapd will continue to make progress and wake the processes
6329 * when the low watermark is reached.
6331 * Returns true if a fatal signal was delivered during throttling. If this
6332 * happens, the page allocator should not consider triggering the OOM killer.
6334 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6335 nodemask_t *nodemask)
6339 pg_data_t *pgdat = NULL;
6342 * Kernel threads should not be throttled as they may be indirectly
6343 * responsible for cleaning pages necessary for reclaim to make forward
6344 * progress. kjournald for example may enter direct reclaim while
6345 * committing a transaction where throttling it could forcing other
6346 * processes to block on log_wait_commit().
6348 if (current->flags & PF_KTHREAD)
6352 * If a fatal signal is pending, this process should not throttle.
6353 * It should return quickly so it can exit and free its memory
6355 if (fatal_signal_pending(current))
6359 * Check if the pfmemalloc reserves are ok by finding the first node
6360 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6361 * GFP_KERNEL will be required for allocating network buffers when
6362 * swapping over the network so ZONE_HIGHMEM is unusable.
6364 * Throttling is based on the first usable node and throttled processes
6365 * wait on a queue until kswapd makes progress and wakes them. There
6366 * is an affinity then between processes waking up and where reclaim
6367 * progress has been made assuming the process wakes on the same node.
6368 * More importantly, processes running on remote nodes will not compete
6369 * for remote pfmemalloc reserves and processes on different nodes
6370 * should make reasonable progress.
6372 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6373 gfp_zone(gfp_mask), nodemask) {
6374 if (zone_idx(zone) > ZONE_NORMAL)
6377 /* Throttle based on the first usable node */
6378 pgdat = zone->zone_pgdat;
6379 if (allow_direct_reclaim(pgdat))
6384 /* If no zone was usable by the allocation flags then do not throttle */
6388 /* Account for the throttling */
6389 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6392 * If the caller cannot enter the filesystem, it's possible that it
6393 * is due to the caller holding an FS lock or performing a journal
6394 * transaction in the case of a filesystem like ext[3|4]. In this case,
6395 * it is not safe to block on pfmemalloc_wait as kswapd could be
6396 * blocked waiting on the same lock. Instead, throttle for up to a
6397 * second before continuing.
6399 if (!(gfp_mask & __GFP_FS))
6400 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6401 allow_direct_reclaim(pgdat), HZ);
6403 /* Throttle until kswapd wakes the process */
6404 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6405 allow_direct_reclaim(pgdat));
6407 if (fatal_signal_pending(current))
6414 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6415 gfp_t gfp_mask, nodemask_t *nodemask)
6417 unsigned long nr_reclaimed;
6418 struct scan_control sc = {
6419 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6420 .gfp_mask = current_gfp_context(gfp_mask),
6421 .reclaim_idx = gfp_zone(gfp_mask),
6423 .nodemask = nodemask,
6424 .priority = DEF_PRIORITY,
6425 .may_writepage = !laptop_mode,
6431 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6432 * Confirm they are large enough for max values.
6434 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6435 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6436 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6439 * Do not enter reclaim if fatal signal was delivered while throttled.
6440 * 1 is returned so that the page allocator does not OOM kill at this
6443 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6446 set_task_reclaim_state(current, &sc.reclaim_state);
6447 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6449 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6451 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6452 set_task_reclaim_state(current, NULL);
6454 return nr_reclaimed;
6459 /* Only used by soft limit reclaim. Do not reuse for anything else. */
6460 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6461 gfp_t gfp_mask, bool noswap,
6463 unsigned long *nr_scanned)
6465 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6466 struct scan_control sc = {
6467 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6468 .target_mem_cgroup = memcg,
6469 .may_writepage = !laptop_mode,
6471 .reclaim_idx = MAX_NR_ZONES - 1,
6472 .may_swap = !noswap,
6475 WARN_ON_ONCE(!current->reclaim_state);
6477 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6478 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6480 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6484 * NOTE: Although we can get the priority field, using it
6485 * here is not a good idea, since it limits the pages we can scan.
6486 * if we don't reclaim here, the shrink_node from balance_pgdat
6487 * will pick up pages from other mem cgroup's as well. We hack
6488 * the priority and make it zero.
6490 shrink_lruvec(lruvec, &sc);
6492 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6494 *nr_scanned = sc.nr_scanned;
6496 return sc.nr_reclaimed;
6499 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6500 unsigned long nr_pages,
6502 unsigned int reclaim_options)
6504 unsigned long nr_reclaimed;
6505 unsigned int noreclaim_flag;
6506 struct scan_control sc = {
6507 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6508 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6509 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6510 .reclaim_idx = MAX_NR_ZONES - 1,
6511 .target_mem_cgroup = memcg,
6512 .priority = DEF_PRIORITY,
6513 .may_writepage = !laptop_mode,
6515 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6516 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6519 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6520 * equal pressure on all the nodes. This is based on the assumption that
6521 * the reclaim does not bail out early.
6523 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6525 set_task_reclaim_state(current, &sc.reclaim_state);
6526 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6527 noreclaim_flag = memalloc_noreclaim_save();
6529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6531 memalloc_noreclaim_restore(noreclaim_flag);
6532 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6533 set_task_reclaim_state(current, NULL);
6535 return nr_reclaimed;
6539 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6541 struct mem_cgroup *memcg;
6542 struct lruvec *lruvec;
6544 if (lru_gen_enabled()) {
6545 lru_gen_age_node(pgdat, sc);
6549 if (!can_age_anon_pages(pgdat, sc))
6552 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6553 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6556 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6558 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6559 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6560 sc, LRU_ACTIVE_ANON);
6561 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6565 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6571 * Check for watermark boosts top-down as the higher zones
6572 * are more likely to be boosted. Both watermarks and boosts
6573 * should not be checked at the same time as reclaim would
6574 * start prematurely when there is no boosting and a lower
6577 for (i = highest_zoneidx; i >= 0; i--) {
6578 zone = pgdat->node_zones + i;
6579 if (!managed_zone(zone))
6582 if (zone->watermark_boost)
6590 * Returns true if there is an eligible zone balanced for the request order
6591 * and highest_zoneidx
6593 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6596 unsigned long mark = -1;
6600 * Check watermarks bottom-up as lower zones are more likely to
6603 for (i = 0; i <= highest_zoneidx; i++) {
6604 zone = pgdat->node_zones + i;
6606 if (!managed_zone(zone))
6609 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6610 mark = wmark_pages(zone, WMARK_PROMO);
6612 mark = high_wmark_pages(zone);
6613 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6618 * If a node has no managed zone within highest_zoneidx, it does not
6619 * need balancing by definition. This can happen if a zone-restricted
6620 * allocation tries to wake a remote kswapd.
6628 /* Clear pgdat state for congested, dirty or under writeback. */
6629 static void clear_pgdat_congested(pg_data_t *pgdat)
6631 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6633 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6634 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6635 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6636 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6640 * Prepare kswapd for sleeping. This verifies that there are no processes
6641 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6643 * Returns true if kswapd is ready to sleep
6645 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6646 int highest_zoneidx)
6649 * The throttled processes are normally woken up in balance_pgdat() as
6650 * soon as allow_direct_reclaim() is true. But there is a potential
6651 * race between when kswapd checks the watermarks and a process gets
6652 * throttled. There is also a potential race if processes get
6653 * throttled, kswapd wakes, a large process exits thereby balancing the
6654 * zones, which causes kswapd to exit balance_pgdat() before reaching
6655 * the wake up checks. If kswapd is going to sleep, no process should
6656 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6657 * the wake up is premature, processes will wake kswapd and get
6658 * throttled again. The difference from wake ups in balance_pgdat() is
6659 * that here we are under prepare_to_wait().
6661 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6662 wake_up_all(&pgdat->pfmemalloc_wait);
6664 /* Hopeless node, leave it to direct reclaim */
6665 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6668 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6669 clear_pgdat_congested(pgdat);
6677 * kswapd shrinks a node of pages that are at or below the highest usable
6678 * zone that is currently unbalanced.
6680 * Returns true if kswapd scanned at least the requested number of pages to
6681 * reclaim or if the lack of progress was due to pages under writeback.
6682 * This is used to determine if the scanning priority needs to be raised.
6684 static bool kswapd_shrink_node(pg_data_t *pgdat,
6685 struct scan_control *sc)
6690 /* Reclaim a number of pages proportional to the number of zones */
6691 sc->nr_to_reclaim = 0;
6692 for (z = 0; z <= sc->reclaim_idx; z++) {
6693 zone = pgdat->node_zones + z;
6694 if (!managed_zone(zone))
6697 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6701 * Historically care was taken to put equal pressure on all zones but
6702 * now pressure is applied based on node LRU order.
6704 shrink_node(pgdat, sc);
6707 * Fragmentation may mean that the system cannot be rebalanced for
6708 * high-order allocations. If twice the allocation size has been
6709 * reclaimed then recheck watermarks only at order-0 to prevent
6710 * excessive reclaim. Assume that a process requested a high-order
6711 * can direct reclaim/compact.
6713 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6716 return sc->nr_scanned >= sc->nr_to_reclaim;
6719 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6721 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6726 for (i = 0; i <= highest_zoneidx; i++) {
6727 zone = pgdat->node_zones + i;
6729 if (!managed_zone(zone))
6733 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6735 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6740 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6742 update_reclaim_active(pgdat, highest_zoneidx, true);
6746 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6748 update_reclaim_active(pgdat, highest_zoneidx, false);
6752 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6753 * that are eligible for use by the caller until at least one zone is
6756 * Returns the order kswapd finished reclaiming at.
6758 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6759 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6760 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6761 * or lower is eligible for reclaim until at least one usable zone is
6764 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6767 unsigned long nr_soft_reclaimed;
6768 unsigned long nr_soft_scanned;
6769 unsigned long pflags;
6770 unsigned long nr_boost_reclaim;
6771 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6774 struct scan_control sc = {
6775 .gfp_mask = GFP_KERNEL,
6780 set_task_reclaim_state(current, &sc.reclaim_state);
6781 psi_memstall_enter(&pflags);
6782 __fs_reclaim_acquire(_THIS_IP_);
6784 count_vm_event(PAGEOUTRUN);
6787 * Account for the reclaim boost. Note that the zone boost is left in
6788 * place so that parallel allocations that are near the watermark will
6789 * stall or direct reclaim until kswapd is finished.
6791 nr_boost_reclaim = 0;
6792 for (i = 0; i <= highest_zoneidx; i++) {
6793 zone = pgdat->node_zones + i;
6794 if (!managed_zone(zone))
6797 nr_boost_reclaim += zone->watermark_boost;
6798 zone_boosts[i] = zone->watermark_boost;
6800 boosted = nr_boost_reclaim;
6803 set_reclaim_active(pgdat, highest_zoneidx);
6804 sc.priority = DEF_PRIORITY;
6806 unsigned long nr_reclaimed = sc.nr_reclaimed;
6807 bool raise_priority = true;
6812 sc.reclaim_idx = highest_zoneidx;
6815 * If the number of buffer_heads exceeds the maximum allowed
6816 * then consider reclaiming from all zones. This has a dual
6817 * purpose -- on 64-bit systems it is expected that
6818 * buffer_heads are stripped during active rotation. On 32-bit
6819 * systems, highmem pages can pin lowmem memory and shrinking
6820 * buffers can relieve lowmem pressure. Reclaim may still not
6821 * go ahead if all eligible zones for the original allocation
6822 * request are balanced to avoid excessive reclaim from kswapd.
6824 if (buffer_heads_over_limit) {
6825 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6826 zone = pgdat->node_zones + i;
6827 if (!managed_zone(zone))
6836 * If the pgdat is imbalanced then ignore boosting and preserve
6837 * the watermarks for a later time and restart. Note that the
6838 * zone watermarks will be still reset at the end of balancing
6839 * on the grounds that the normal reclaim should be enough to
6840 * re-evaluate if boosting is required when kswapd next wakes.
6842 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6843 if (!balanced && nr_boost_reclaim) {
6844 nr_boost_reclaim = 0;
6849 * If boosting is not active then only reclaim if there are no
6850 * eligible zones. Note that sc.reclaim_idx is not used as
6851 * buffer_heads_over_limit may have adjusted it.
6853 if (!nr_boost_reclaim && balanced)
6856 /* Limit the priority of boosting to avoid reclaim writeback */
6857 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6858 raise_priority = false;
6861 * Do not writeback or swap pages for boosted reclaim. The
6862 * intent is to relieve pressure not issue sub-optimal IO
6863 * from reclaim context. If no pages are reclaimed, the
6864 * reclaim will be aborted.
6866 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6867 sc.may_swap = !nr_boost_reclaim;
6870 * Do some background aging, to give pages a chance to be
6871 * referenced before reclaiming. All pages are rotated
6872 * regardless of classzone as this is about consistent aging.
6874 kswapd_age_node(pgdat, &sc);
6877 * If we're getting trouble reclaiming, start doing writepage
6878 * even in laptop mode.
6880 if (sc.priority < DEF_PRIORITY - 2)
6881 sc.may_writepage = 1;
6883 /* Call soft limit reclaim before calling shrink_node. */
6885 nr_soft_scanned = 0;
6886 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
6887 sc.gfp_mask, &nr_soft_scanned);
6888 sc.nr_reclaimed += nr_soft_reclaimed;
6891 * There should be no need to raise the scanning priority if
6892 * enough pages are already being scanned that that high
6893 * watermark would be met at 100% efficiency.
6895 if (kswapd_shrink_node(pgdat, &sc))
6896 raise_priority = false;
6899 * If the low watermark is met there is no need for processes
6900 * to be throttled on pfmemalloc_wait as they should not be
6901 * able to safely make forward progress. Wake them
6903 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6904 allow_direct_reclaim(pgdat))
6905 wake_up_all(&pgdat->pfmemalloc_wait);
6907 /* Check if kswapd should be suspending */
6908 __fs_reclaim_release(_THIS_IP_);
6909 ret = kthread_freezable_should_stop(&was_frozen);
6910 __fs_reclaim_acquire(_THIS_IP_);
6911 if (was_frozen || ret)
6915 * Raise priority if scanning rate is too low or there was no
6916 * progress in reclaiming pages
6918 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6919 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6922 * If reclaim made no progress for a boost, stop reclaim as
6923 * IO cannot be queued and it could be an infinite loop in
6924 * extreme circumstances.
6926 if (nr_boost_reclaim && !nr_reclaimed)
6929 if (raise_priority || !nr_reclaimed)
6931 } while (sc.priority >= 1);
6934 * Restart only if it went through the priority loop all the way,
6935 * but cache_trim_mode didn't work.
6937 if (!sc.nr_reclaimed && sc.priority < 1 &&
6938 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
6939 sc.no_cache_trim_mode = 1;
6943 if (!sc.nr_reclaimed)
6944 pgdat->kswapd_failures++;
6947 clear_reclaim_active(pgdat, highest_zoneidx);
6949 /* If reclaim was boosted, account for the reclaim done in this pass */
6951 unsigned long flags;
6953 for (i = 0; i <= highest_zoneidx; i++) {
6954 if (!zone_boosts[i])
6957 /* Increments are under the zone lock */
6958 zone = pgdat->node_zones + i;
6959 spin_lock_irqsave(&zone->lock, flags);
6960 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6961 spin_unlock_irqrestore(&zone->lock, flags);
6965 * As there is now likely space, wakeup kcompact to defragment
6968 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
6971 snapshot_refaults(NULL, pgdat);
6972 __fs_reclaim_release(_THIS_IP_);
6973 psi_memstall_leave(&pflags);
6974 set_task_reclaim_state(current, NULL);
6977 * Return the order kswapd stopped reclaiming at as
6978 * prepare_kswapd_sleep() takes it into account. If another caller
6979 * entered the allocator slow path while kswapd was awake, order will
6980 * remain at the higher level.
6986 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6987 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6988 * not a valid index then either kswapd runs for first time or kswapd couldn't
6989 * sleep after previous reclaim attempt (node is still unbalanced). In that
6990 * case return the zone index of the previous kswapd reclaim cycle.
6992 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6993 enum zone_type prev_highest_zoneidx)
6995 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
6997 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7000 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7001 unsigned int highest_zoneidx)
7006 if (freezing(current) || kthread_should_stop())
7009 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7012 * Try to sleep for a short interval. Note that kcompactd will only be
7013 * woken if it is possible to sleep for a short interval. This is
7014 * deliberate on the assumption that if reclaim cannot keep an
7015 * eligible zone balanced that it's also unlikely that compaction will
7018 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7020 * Compaction records what page blocks it recently failed to
7021 * isolate pages from and skips them in the future scanning.
7022 * When kswapd is going to sleep, it is reasonable to assume
7023 * that pages and compaction may succeed so reset the cache.
7025 reset_isolation_suitable(pgdat);
7028 * We have freed the memory, now we should compact it to make
7029 * allocation of the requested order possible.
7031 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7033 remaining = schedule_timeout(HZ/10);
7036 * If woken prematurely then reset kswapd_highest_zoneidx and
7037 * order. The values will either be from a wakeup request or
7038 * the previous request that slept prematurely.
7041 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7042 kswapd_highest_zoneidx(pgdat,
7045 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7046 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7049 finish_wait(&pgdat->kswapd_wait, &wait);
7050 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7054 * After a short sleep, check if it was a premature sleep. If not, then
7055 * go fully to sleep until explicitly woken up.
7058 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7059 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7062 * vmstat counters are not perfectly accurate and the estimated
7063 * value for counters such as NR_FREE_PAGES can deviate from the
7064 * true value by nr_online_cpus * threshold. To avoid the zone
7065 * watermarks being breached while under pressure, we reduce the
7066 * per-cpu vmstat threshold while kswapd is awake and restore
7067 * them before going back to sleep.
7069 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7071 if (!kthread_should_stop())
7074 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7077 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7079 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7081 finish_wait(&pgdat->kswapd_wait, &wait);
7085 * The background pageout daemon, started as a kernel thread
7086 * from the init process.
7088 * This basically trickles out pages so that we have _some_
7089 * free memory available even if there is no other activity
7090 * that frees anything up. This is needed for things like routing
7091 * etc, where we otherwise might have all activity going on in
7092 * asynchronous contexts that cannot page things out.
7094 * If there are applications that are active memory-allocators
7095 * (most normal use), this basically shouldn't matter.
7097 static int kswapd(void *p)
7099 unsigned int alloc_order, reclaim_order;
7100 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7101 pg_data_t *pgdat = (pg_data_t *)p;
7102 struct task_struct *tsk = current;
7103 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7105 if (!cpumask_empty(cpumask))
7106 set_cpus_allowed_ptr(tsk, cpumask);
7109 * Tell the memory management that we're a "memory allocator",
7110 * and that if we need more memory we should get access to it
7111 * regardless (see "__alloc_pages()"). "kswapd" should
7112 * never get caught in the normal page freeing logic.
7114 * (Kswapd normally doesn't need memory anyway, but sometimes
7115 * you need a small amount of memory in order to be able to
7116 * page out something else, and this flag essentially protects
7117 * us from recursively trying to free more memory as we're
7118 * trying to free the first piece of memory in the first place).
7120 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7123 WRITE_ONCE(pgdat->kswapd_order, 0);
7124 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7125 atomic_set(&pgdat->nr_writeback_throttled, 0);
7129 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7130 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7134 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7137 /* Read the new order and highest_zoneidx */
7138 alloc_order = READ_ONCE(pgdat->kswapd_order);
7139 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7141 WRITE_ONCE(pgdat->kswapd_order, 0);
7142 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7144 if (kthread_freezable_should_stop(&was_frozen))
7148 * We can speed up thawing tasks if we don't call balance_pgdat
7149 * after returning from the refrigerator
7155 * Reclaim begins at the requested order but if a high-order
7156 * reclaim fails then kswapd falls back to reclaiming for
7157 * order-0. If that happens, kswapd will consider sleeping
7158 * for the order it finished reclaiming at (reclaim_order)
7159 * but kcompactd is woken to compact for the original
7160 * request (alloc_order).
7162 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7164 reclaim_order = balance_pgdat(pgdat, alloc_order,
7166 if (reclaim_order < alloc_order)
7167 goto kswapd_try_sleep;
7170 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7176 * A zone is low on free memory or too fragmented for high-order memory. If
7177 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7178 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7179 * has failed or is not needed, still wake up kcompactd if only compaction is
7182 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7183 enum zone_type highest_zoneidx)
7186 enum zone_type curr_idx;
7188 if (!managed_zone(zone))
7191 if (!cpuset_zone_allowed(zone, gfp_flags))
7194 pgdat = zone->zone_pgdat;
7195 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7197 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7198 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7200 if (READ_ONCE(pgdat->kswapd_order) < order)
7201 WRITE_ONCE(pgdat->kswapd_order, order);
7203 if (!waitqueue_active(&pgdat->kswapd_wait))
7206 /* Hopeless node, leave it to direct reclaim if possible */
7207 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7208 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7209 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7211 * There may be plenty of free memory available, but it's too
7212 * fragmented for high-order allocations. Wake up kcompactd
7213 * and rely on compaction_suitable() to determine if it's
7214 * needed. If it fails, it will defer subsequent attempts to
7215 * ratelimit its work.
7217 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7218 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7222 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7224 wake_up_interruptible(&pgdat->kswapd_wait);
7227 #ifdef CONFIG_HIBERNATION
7229 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7232 * Rather than trying to age LRUs the aim is to preserve the overall
7233 * LRU order by reclaiming preferentially
7234 * inactive > active > active referenced > active mapped
7236 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7238 struct scan_control sc = {
7239 .nr_to_reclaim = nr_to_reclaim,
7240 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7241 .reclaim_idx = MAX_NR_ZONES - 1,
7242 .priority = DEF_PRIORITY,
7246 .hibernation_mode = 1,
7248 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7249 unsigned long nr_reclaimed;
7250 unsigned int noreclaim_flag;
7252 fs_reclaim_acquire(sc.gfp_mask);
7253 noreclaim_flag = memalloc_noreclaim_save();
7254 set_task_reclaim_state(current, &sc.reclaim_state);
7256 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7258 set_task_reclaim_state(current, NULL);
7259 memalloc_noreclaim_restore(noreclaim_flag);
7260 fs_reclaim_release(sc.gfp_mask);
7262 return nr_reclaimed;
7264 #endif /* CONFIG_HIBERNATION */
7267 * This kswapd start function will be called by init and node-hot-add.
7269 void __meminit kswapd_run(int nid)
7271 pg_data_t *pgdat = NODE_DATA(nid);
7273 pgdat_kswapd_lock(pgdat);
7274 if (!pgdat->kswapd) {
7275 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7276 if (IS_ERR(pgdat->kswapd)) {
7277 /* failure at boot is fatal */
7278 pr_err("Failed to start kswapd on node %d,ret=%ld\n",
7279 nid, PTR_ERR(pgdat->kswapd));
7280 BUG_ON(system_state < SYSTEM_RUNNING);
7281 pgdat->kswapd = NULL;
7284 pgdat_kswapd_unlock(pgdat);
7288 * Called by memory hotplug when all memory in a node is offlined. Caller must
7289 * be holding mem_hotplug_begin/done().
7291 void __meminit kswapd_stop(int nid)
7293 pg_data_t *pgdat = NODE_DATA(nid);
7294 struct task_struct *kswapd;
7296 pgdat_kswapd_lock(pgdat);
7297 kswapd = pgdat->kswapd;
7299 kthread_stop(kswapd);
7300 pgdat->kswapd = NULL;
7302 pgdat_kswapd_unlock(pgdat);
7305 static int __init kswapd_init(void)
7310 for_each_node_state(nid, N_MEMORY)
7315 module_init(kswapd_init)
7321 * If non-zero call node_reclaim when the number of free pages falls below
7324 int node_reclaim_mode __read_mostly;
7327 * Priority for NODE_RECLAIM. This determines the fraction of pages
7328 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7331 #define NODE_RECLAIM_PRIORITY 4
7334 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7337 int sysctl_min_unmapped_ratio = 1;
7340 * If the number of slab pages in a zone grows beyond this percentage then
7341 * slab reclaim needs to occur.
7343 int sysctl_min_slab_ratio = 5;
7345 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7347 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7348 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7349 node_page_state(pgdat, NR_ACTIVE_FILE);
7352 * It's possible for there to be more file mapped pages than
7353 * accounted for by the pages on the file LRU lists because
7354 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7356 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7359 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
7360 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7362 unsigned long nr_pagecache_reclaimable;
7363 unsigned long delta = 0;
7366 * If RECLAIM_UNMAP is set, then all file pages are considered
7367 * potentially reclaimable. Otherwise, we have to worry about
7368 * pages like swapcache and node_unmapped_file_pages() provides
7371 if (node_reclaim_mode & RECLAIM_UNMAP)
7372 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7374 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7376 /* If we can't clean pages, remove dirty pages from consideration */
7377 if (!(node_reclaim_mode & RECLAIM_WRITE))
7378 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7380 /* Watch for any possible underflows due to delta */
7381 if (unlikely(delta > nr_pagecache_reclaimable))
7382 delta = nr_pagecache_reclaimable;
7384 return nr_pagecache_reclaimable - delta;
7388 * Try to free up some pages from this node through reclaim.
7390 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7392 /* Minimum pages needed in order to stay on node */
7393 const unsigned long nr_pages = 1 << order;
7394 struct task_struct *p = current;
7395 unsigned int noreclaim_flag;
7396 struct scan_control sc = {
7397 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7398 .gfp_mask = current_gfp_context(gfp_mask),
7400 .priority = NODE_RECLAIM_PRIORITY,
7401 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7402 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7404 .reclaim_idx = gfp_zone(gfp_mask),
7406 unsigned long pflags;
7408 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7412 psi_memstall_enter(&pflags);
7413 delayacct_freepages_start();
7414 fs_reclaim_acquire(sc.gfp_mask);
7416 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7418 noreclaim_flag = memalloc_noreclaim_save();
7419 set_task_reclaim_state(p, &sc.reclaim_state);
7421 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7422 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7424 * Free memory by calling shrink node with increasing
7425 * priorities until we have enough memory freed.
7428 shrink_node(pgdat, &sc);
7429 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7432 set_task_reclaim_state(p, NULL);
7433 memalloc_noreclaim_restore(noreclaim_flag);
7434 fs_reclaim_release(sc.gfp_mask);
7435 psi_memstall_leave(&pflags);
7436 delayacct_freepages_end();
7438 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7440 return sc.nr_reclaimed >= nr_pages;
7443 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7448 * Node reclaim reclaims unmapped file backed pages and
7449 * slab pages if we are over the defined limits.
7451 * A small portion of unmapped file backed pages is needed for
7452 * file I/O otherwise pages read by file I/O will be immediately
7453 * thrown out if the node is overallocated. So we do not reclaim
7454 * if less than a specified percentage of the node is used by
7455 * unmapped file backed pages.
7457 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7458 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7459 pgdat->min_slab_pages)
7460 return NODE_RECLAIM_FULL;
7463 * Do not scan if the allocation should not be delayed.
7465 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7466 return NODE_RECLAIM_NOSCAN;
7469 * Only run node reclaim on the local node or on nodes that do not
7470 * have associated processors. This will favor the local processor
7471 * over remote processors and spread off node memory allocations
7472 * as wide as possible.
7474 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7475 return NODE_RECLAIM_NOSCAN;
7477 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7478 return NODE_RECLAIM_NOSCAN;
7480 ret = __node_reclaim(pgdat, gfp_mask, order);
7481 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7484 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7491 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7493 * @fbatch: Batch of lru folios to check.
7495 * Checks folios for evictability, if an evictable folio is in the unevictable
7496 * lru list, moves it to the appropriate evictable lru list. This function
7497 * should be only used for lru folios.
7499 void check_move_unevictable_folios(struct folio_batch *fbatch)
7501 struct lruvec *lruvec = NULL;
7506 for (i = 0; i < fbatch->nr; i++) {
7507 struct folio *folio = fbatch->folios[i];
7508 int nr_pages = folio_nr_pages(folio);
7510 pgscanned += nr_pages;
7512 /* block memcg migration while the folio moves between lrus */
7513 if (!folio_test_clear_lru(folio))
7516 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7517 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7518 lruvec_del_folio(lruvec, folio);
7519 folio_clear_unevictable(folio);
7520 lruvec_add_folio(lruvec, folio);
7521 pgrescued += nr_pages;
7523 folio_set_lru(folio);
7527 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7528 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7529 unlock_page_lruvec_irq(lruvec);
7530 } else if (pgscanned) {
7531 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7534 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);