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.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
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 /* Proactive reclaim invoked by userspace through memory.reclaim */
112 unsigned int proactive
:1;
115 * Cgroup memory below memory.low is protected as long as we
116 * don't threaten to OOM. If any cgroup is reclaimed at
117 * reduced force or passed over entirely due to its memory.low
118 * setting (memcg_low_skipped), and nothing is reclaimed as a
119 * result, then go back for one more cycle that reclaims the protected
120 * memory (memcg_low_reclaim) to avert OOM.
122 unsigned int memcg_low_reclaim
:1;
123 unsigned int memcg_low_skipped
:1;
125 unsigned int hibernation_mode
:1;
127 /* One of the zones is ready for compaction */
128 unsigned int compaction_ready
:1;
130 /* There is easily reclaimable cold cache in the current node */
131 unsigned int cache_trim_mode
:1;
133 /* The file folios on the current node are dangerously low */
134 unsigned int file_is_tiny
:1;
136 /* Always discard instead of demoting to lower tier memory */
137 unsigned int no_demotion
:1;
139 /* Allocation order */
142 /* Scan (total_size >> priority) pages at once */
145 /* The highest zone to isolate folios for reclaim from */
148 /* This context's GFP mask */
151 /* Incremented by the number of inactive pages that were scanned */
152 unsigned long nr_scanned
;
154 /* Number of pages freed so far during a call to shrink_zones() */
155 unsigned long nr_reclaimed
;
159 unsigned int unqueued_dirty
;
160 unsigned int congested
;
161 unsigned int writeback
;
162 unsigned int immediate
;
163 unsigned int file_taken
;
167 /* for recording the reclaimed slab by now */
168 struct reclaim_state reclaim_state
;
171 #ifdef ARCH_HAS_PREFETCHW
172 #define prefetchw_prev_lru_folio(_folio, _base, _field) \
174 if ((_folio)->lru.prev != _base) { \
175 struct folio *prev; \
177 prev = lru_to_folio(&(_folio->lru)); \
178 prefetchw(&prev->_field); \
182 #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
186 * From 0 .. 200. Higher means more swappy.
188 int vm_swappiness
= 60;
192 /* Returns true for reclaim through cgroup limits or cgroup interfaces. */
193 static bool cgroup_reclaim(struct scan_control
*sc
)
195 return sc
->target_mem_cgroup
;
199 * Returns true for reclaim on the root cgroup. This is true for direct
200 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
202 static bool root_reclaim(struct scan_control
*sc
)
204 return !sc
->target_mem_cgroup
|| mem_cgroup_is_root(sc
->target_mem_cgroup
);
208 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
209 * @sc: scan_control in question
211 * The normal page dirty throttling mechanism in balance_dirty_pages() is
212 * completely broken with the legacy memcg and direct stalling in
213 * shrink_folio_list() is used for throttling instead, which lacks all the
214 * niceties such as fairness, adaptive pausing, bandwidth proportional
215 * allocation and configurability.
217 * This function tests whether the vmscan currently in progress can assume
218 * that the normal dirty throttling mechanism is operational.
220 static bool writeback_throttling_sane(struct scan_control
*sc
)
222 if (!cgroup_reclaim(sc
))
224 #ifdef CONFIG_CGROUP_WRITEBACK
225 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
231 static bool cgroup_reclaim(struct scan_control
*sc
)
236 static bool root_reclaim(struct scan_control
*sc
)
241 static bool writeback_throttling_sane(struct scan_control
*sc
)
247 static void set_task_reclaim_state(struct task_struct
*task
,
248 struct reclaim_state
*rs
)
250 /* Check for an overwrite */
251 WARN_ON_ONCE(rs
&& task
->reclaim_state
);
253 /* Check for the nulling of an already-nulled member */
254 WARN_ON_ONCE(!rs
&& !task
->reclaim_state
);
256 task
->reclaim_state
= rs
;
260 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
261 * scan_control->nr_reclaimed.
263 static void flush_reclaim_state(struct scan_control
*sc
)
266 * Currently, reclaim_state->reclaimed includes three types of pages
267 * freed outside of vmscan:
269 * (2) Clean file pages from pruned inodes (on highmem systems).
270 * (3) XFS freed buffer pages.
272 * For all of these cases, we cannot universally link the pages to a
273 * single memcg. For example, a memcg-aware shrinker can free one object
274 * charged to the target memcg, causing an entire page to be freed.
275 * If we count the entire page as reclaimed from the memcg, we end up
276 * overestimating the reclaimed amount (potentially under-reclaiming).
278 * Only count such pages for global reclaim to prevent under-reclaiming
279 * from the target memcg; preventing unnecessary retries during memcg
280 * charging and false positives from proactive reclaim.
282 * For uncommon cases where the freed pages were actually mostly
283 * charged to the target memcg, we end up underestimating the reclaimed
284 * amount. This should be fine. The freed pages will be uncharged
285 * anyway, even if they are not counted here properly, and we will be
286 * able to make forward progress in charging (which is usually in a
289 * We can go one step further, and report the uncharged objcg pages in
290 * memcg reclaim, to make reporting more accurate and reduce
291 * underestimation, but it's probably not worth the complexity for now.
293 if (current
->reclaim_state
&& root_reclaim(sc
)) {
294 sc
->nr_reclaimed
+= current
->reclaim_state
->reclaimed
;
295 current
->reclaim_state
->reclaimed
= 0;
299 static bool can_demote(int nid
, struct scan_control
*sc
)
301 if (!numa_demotion_enabled
)
303 if (sc
&& sc
->no_demotion
)
305 if (next_demotion_node(nid
) == NUMA_NO_NODE
)
311 static inline bool can_reclaim_anon_pages(struct mem_cgroup
*memcg
,
313 struct scan_control
*sc
)
317 * For non-memcg reclaim, is there
318 * space in any swap device?
320 if (get_nr_swap_pages() > 0)
323 /* Is the memcg below its swap limit? */
324 if (mem_cgroup_get_nr_swap_pages(memcg
) > 0)
329 * The page can not be swapped.
331 * Can it be reclaimed from this node via demotion?
333 return can_demote(nid
, sc
);
337 * This misses isolated folios which are not accounted for to save counters.
338 * As the data only determines if reclaim or compaction continues, it is
339 * not expected that isolated folios will be a dominating factor.
341 unsigned long zone_reclaimable_pages(struct zone
*zone
)
345 nr
= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_FILE
) +
346 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_FILE
);
347 if (can_reclaim_anon_pages(NULL
, zone_to_nid(zone
), NULL
))
348 nr
+= zone_page_state_snapshot(zone
, NR_ZONE_INACTIVE_ANON
) +
349 zone_page_state_snapshot(zone
, NR_ZONE_ACTIVE_ANON
);
355 * lruvec_lru_size - Returns the number of pages on the given LRU list.
356 * @lruvec: lru vector
358 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
360 static unsigned long lruvec_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
363 unsigned long size
= 0;
366 for (zid
= 0; zid
<= zone_idx
; zid
++) {
367 struct zone
*zone
= &lruvec_pgdat(lruvec
)->node_zones
[zid
];
369 if (!managed_zone(zone
))
372 if (!mem_cgroup_disabled())
373 size
+= mem_cgroup_get_zone_lru_size(lruvec
, lru
, zid
);
375 size
+= zone_page_state(zone
, NR_ZONE_LRU_BASE
+ lru
);
380 static unsigned long drop_slab_node(int nid
)
382 unsigned long freed
= 0;
383 struct mem_cgroup
*memcg
= NULL
;
385 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
387 freed
+= shrink_slab(GFP_KERNEL
, nid
, memcg
, 0);
388 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)) != NULL
);
401 for_each_online_node(nid
) {
402 if (fatal_signal_pending(current
))
405 freed
+= drop_slab_node(nid
);
407 } while ((freed
>> shift
++) > 1);
410 static int reclaimer_offset(void)
412 BUILD_BUG_ON(PGSTEAL_DIRECT
- PGSTEAL_KSWAPD
!=
413 PGDEMOTE_DIRECT
- PGDEMOTE_KSWAPD
);
414 BUILD_BUG_ON(PGSTEAL_DIRECT
- PGSTEAL_KSWAPD
!=
415 PGSCAN_DIRECT
- PGSCAN_KSWAPD
);
416 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED
- PGSTEAL_KSWAPD
!=
417 PGDEMOTE_KHUGEPAGED
- PGDEMOTE_KSWAPD
);
418 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED
- PGSTEAL_KSWAPD
!=
419 PGSCAN_KHUGEPAGED
- PGSCAN_KSWAPD
);
421 if (current_is_kswapd())
423 if (current_is_khugepaged())
424 return PGSTEAL_KHUGEPAGED
- PGSTEAL_KSWAPD
;
425 return PGSTEAL_DIRECT
- PGSTEAL_KSWAPD
;
428 static inline int is_page_cache_freeable(struct folio
*folio
)
431 * A freeable page cache folio is referenced only by the caller
432 * that isolated the folio, the page cache and optional filesystem
433 * private data at folio->private.
435 return folio_ref_count(folio
) - folio_test_private(folio
) ==
436 1 + folio_nr_pages(folio
);
440 * We detected a synchronous write error writing a folio out. Probably
441 * -ENOSPC. We need to propagate that into the address_space for a subsequent
442 * fsync(), msync() or close().
444 * The tricky part is that after writepage we cannot touch the mapping: nothing
445 * prevents it from being freed up. But we have a ref on the folio and once
446 * that folio is locked, the mapping is pinned.
448 * We're allowed to run sleeping folio_lock() here because we know the caller has
451 static void handle_write_error(struct address_space
*mapping
,
452 struct folio
*folio
, int error
)
455 if (folio_mapping(folio
) == mapping
)
456 mapping_set_error(mapping
, error
);
460 static bool skip_throttle_noprogress(pg_data_t
*pgdat
)
462 int reclaimable
= 0, write_pending
= 0;
466 * If kswapd is disabled, reschedule if necessary but do not
467 * throttle as the system is likely near OOM.
469 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
473 * If there are a lot of dirty/writeback folios then do not
474 * throttle as throttling will occur when the folios cycle
475 * towards the end of the LRU if still under writeback.
477 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
478 struct zone
*zone
= pgdat
->node_zones
+ i
;
480 if (!managed_zone(zone
))
483 reclaimable
+= zone_reclaimable_pages(zone
);
484 write_pending
+= zone_page_state_snapshot(zone
,
485 NR_ZONE_WRITE_PENDING
);
487 if (2 * write_pending
<= reclaimable
)
493 void reclaim_throttle(pg_data_t
*pgdat
, enum vmscan_throttle_state reason
)
495 wait_queue_head_t
*wqh
= &pgdat
->reclaim_wait
[reason
];
500 * Do not throttle user workers, kthreads other than kswapd or
501 * workqueues. They may be required for reclaim to make
502 * forward progress (e.g. journalling workqueues or kthreads).
504 if (!current_is_kswapd() &&
505 current
->flags
& (PF_USER_WORKER
|PF_KTHREAD
)) {
511 * These figures are pulled out of thin air.
512 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
513 * parallel reclaimers which is a short-lived event so the timeout is
514 * short. Failing to make progress or waiting on writeback are
515 * potentially long-lived events so use a longer timeout. This is shaky
516 * logic as a failure to make progress could be due to anything from
517 * writeback to a slow device to excessive referenced folios at the tail
518 * of the inactive LRU.
521 case VMSCAN_THROTTLE_WRITEBACK
:
524 if (atomic_inc_return(&pgdat
->nr_writeback_throttled
) == 1) {
525 WRITE_ONCE(pgdat
->nr_reclaim_start
,
526 node_page_state(pgdat
, NR_THROTTLED_WRITTEN
));
530 case VMSCAN_THROTTLE_CONGESTED
:
532 case VMSCAN_THROTTLE_NOPROGRESS
:
533 if (skip_throttle_noprogress(pgdat
)) {
541 case VMSCAN_THROTTLE_ISOLATED
:
550 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
551 ret
= schedule_timeout(timeout
);
552 finish_wait(wqh
, &wait
);
554 if (reason
== VMSCAN_THROTTLE_WRITEBACK
)
555 atomic_dec(&pgdat
->nr_writeback_throttled
);
557 trace_mm_vmscan_throttled(pgdat
->node_id
, jiffies_to_usecs(timeout
),
558 jiffies_to_usecs(timeout
- ret
),
563 * Account for folios written if tasks are throttled waiting on dirty
564 * folios to clean. If enough folios have been cleaned since throttling
565 * started then wakeup the throttled tasks.
567 void __acct_reclaim_writeback(pg_data_t
*pgdat
, struct folio
*folio
,
570 unsigned long nr_written
;
572 node_stat_add_folio(folio
, NR_THROTTLED_WRITTEN
);
575 * This is an inaccurate read as the per-cpu deltas may not
576 * be synchronised. However, given that the system is
577 * writeback throttled, it is not worth taking the penalty
578 * of getting an accurate count. At worst, the throttle
579 * timeout guarantees forward progress.
581 nr_written
= node_page_state(pgdat
, NR_THROTTLED_WRITTEN
) -
582 READ_ONCE(pgdat
->nr_reclaim_start
);
584 if (nr_written
> SWAP_CLUSTER_MAX
* nr_throttled
)
585 wake_up(&pgdat
->reclaim_wait
[VMSCAN_THROTTLE_WRITEBACK
]);
588 /* possible outcome of pageout() */
590 /* failed to write folio out, folio is locked */
592 /* move folio to the active list, folio is locked */
594 /* folio has been sent to the disk successfully, folio is unlocked */
596 /* folio is clean and locked */
601 * pageout is called by shrink_folio_list() for each dirty folio.
602 * Calls ->writepage().
604 static pageout_t
pageout(struct folio
*folio
, struct address_space
*mapping
,
605 struct swap_iocb
**plug
)
608 * If the folio is dirty, only perform writeback if that write
609 * will be non-blocking. To prevent this allocation from being
610 * stalled by pagecache activity. But note that there may be
611 * stalls if we need to run get_block(). We could test
612 * PagePrivate for that.
614 * If this process is currently in __generic_file_write_iter() against
615 * this folio's queue, we can perform writeback even if that
618 * If the folio is swapcache, write it back even if that would
619 * block, for some throttling. This happens by accident, because
620 * swap_backing_dev_info is bust: it doesn't reflect the
621 * congestion state of the swapdevs. Easy to fix, if needed.
623 if (!is_page_cache_freeable(folio
))
627 * Some data journaling orphaned folios can have
628 * folio->mapping == NULL while being dirty with clean buffers.
630 if (folio_test_private(folio
)) {
631 if (try_to_free_buffers(folio
)) {
632 folio_clear_dirty(folio
);
633 pr_info("%s: orphaned folio\n", __func__
);
639 if (mapping
->a_ops
->writepage
== NULL
)
640 return PAGE_ACTIVATE
;
642 if (folio_clear_dirty_for_io(folio
)) {
644 struct writeback_control wbc
= {
645 .sync_mode
= WB_SYNC_NONE
,
646 .nr_to_write
= SWAP_CLUSTER_MAX
,
648 .range_end
= LLONG_MAX
,
653 folio_set_reclaim(folio
);
654 res
= mapping
->a_ops
->writepage(&folio
->page
, &wbc
);
656 handle_write_error(mapping
, folio
, res
);
657 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
658 folio_clear_reclaim(folio
);
659 return PAGE_ACTIVATE
;
662 if (!folio_test_writeback(folio
)) {
663 /* synchronous write or broken a_ops? */
664 folio_clear_reclaim(folio
);
666 trace_mm_vmscan_write_folio(folio
);
667 node_stat_add_folio(folio
, NR_VMSCAN_WRITE
);
675 * Same as remove_mapping, but if the folio is removed from the mapping, it
676 * gets returned with a refcount of 0.
678 static int __remove_mapping(struct address_space
*mapping
, struct folio
*folio
,
679 bool reclaimed
, struct mem_cgroup
*target_memcg
)
684 BUG_ON(!folio_test_locked(folio
));
685 BUG_ON(mapping
!= folio_mapping(folio
));
687 if (!folio_test_swapcache(folio
))
688 spin_lock(&mapping
->host
->i_lock
);
689 xa_lock_irq(&mapping
->i_pages
);
691 * The non racy check for a busy folio.
693 * Must be careful with the order of the tests. When someone has
694 * a ref to the folio, it may be possible that they dirty it then
695 * drop the reference. So if the dirty flag is tested before the
696 * refcount here, then the following race may occur:
698 * get_user_pages(&page);
699 * [user mapping goes away]
701 * !folio_test_dirty(folio) [good]
702 * folio_set_dirty(folio);
704 * !refcount(folio) [good, discard it]
706 * [oops, our write_to data is lost]
708 * Reversing the order of the tests ensures such a situation cannot
709 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
710 * load is not satisfied before that of folio->_refcount.
712 * Note that if the dirty flag is always set via folio_mark_dirty,
713 * and thus under the i_pages lock, then this ordering is not required.
715 refcount
= 1 + folio_nr_pages(folio
);
716 if (!folio_ref_freeze(folio
, refcount
))
718 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
719 if (unlikely(folio_test_dirty(folio
))) {
720 folio_ref_unfreeze(folio
, refcount
);
724 if (folio_test_swapcache(folio
)) {
725 swp_entry_t swap
= folio
->swap
;
727 if (reclaimed
&& !mapping_exiting(mapping
))
728 shadow
= workingset_eviction(folio
, target_memcg
);
729 __delete_from_swap_cache(folio
, swap
, shadow
);
730 mem_cgroup_swapout(folio
, swap
);
731 xa_unlock_irq(&mapping
->i_pages
);
732 put_swap_folio(folio
, swap
);
734 void (*free_folio
)(struct folio
*);
736 free_folio
= mapping
->a_ops
->free_folio
;
738 * Remember a shadow entry for reclaimed file cache in
739 * order to detect refaults, thus thrashing, later on.
741 * But don't store shadows in an address space that is
742 * already exiting. This is not just an optimization,
743 * inode reclaim needs to empty out the radix tree or
744 * the nodes are lost. Don't plant shadows behind its
747 * We also don't store shadows for DAX mappings because the
748 * only page cache folios found in these are zero pages
749 * covering holes, and because we don't want to mix DAX
750 * exceptional entries and shadow exceptional entries in the
751 * same address_space.
753 if (reclaimed
&& folio_is_file_lru(folio
) &&
754 !mapping_exiting(mapping
) && !dax_mapping(mapping
))
755 shadow
= workingset_eviction(folio
, target_memcg
);
756 __filemap_remove_folio(folio
, shadow
);
757 xa_unlock_irq(&mapping
->i_pages
);
758 if (mapping_shrinkable(mapping
))
759 inode_add_lru(mapping
->host
);
760 spin_unlock(&mapping
->host
->i_lock
);
769 xa_unlock_irq(&mapping
->i_pages
);
770 if (!folio_test_swapcache(folio
))
771 spin_unlock(&mapping
->host
->i_lock
);
776 * remove_mapping() - Attempt to remove a folio from its mapping.
777 * @mapping: The address space.
778 * @folio: The folio to remove.
780 * If the folio is dirty, under writeback or if someone else has a ref
781 * on it, removal will fail.
782 * Return: The number of pages removed from the mapping. 0 if the folio
783 * could not be removed.
784 * Context: The caller should have a single refcount on the folio and
787 long remove_mapping(struct address_space
*mapping
, struct folio
*folio
)
789 if (__remove_mapping(mapping
, folio
, false, NULL
)) {
791 * Unfreezing the refcount with 1 effectively
792 * drops the pagecache ref for us without requiring another
795 folio_ref_unfreeze(folio
, 1);
796 return folio_nr_pages(folio
);
802 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
803 * @folio: Folio to be returned to an LRU list.
805 * Add previously isolated @folio to appropriate LRU list.
806 * The folio may still be unevictable for other reasons.
808 * Context: lru_lock must not be held, interrupts must be enabled.
810 void folio_putback_lru(struct folio
*folio
)
812 folio_add_lru(folio
);
813 folio_put(folio
); /* drop ref from isolate */
816 enum folio_references
{
818 FOLIOREF_RECLAIM_CLEAN
,
823 static enum folio_references
folio_check_references(struct folio
*folio
,
824 struct scan_control
*sc
)
826 int referenced_ptes
, referenced_folio
;
827 unsigned long vm_flags
;
829 referenced_ptes
= folio_referenced(folio
, 1, sc
->target_mem_cgroup
,
831 referenced_folio
= folio_test_clear_referenced(folio
);
834 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
835 * Let the folio, now marked Mlocked, be moved to the unevictable list.
837 if (vm_flags
& VM_LOCKED
)
838 return FOLIOREF_ACTIVATE
;
840 /* rmap lock contention: rotate */
841 if (referenced_ptes
== -1)
842 return FOLIOREF_KEEP
;
844 if (referenced_ptes
) {
846 * All mapped folios start out with page table
847 * references from the instantiating fault, so we need
848 * to look twice if a mapped file/anon folio is used more
851 * Mark it and spare it for another trip around the
852 * inactive list. Another page table reference will
853 * lead to its activation.
855 * Note: the mark is set for activated folios as well
856 * so that recently deactivated but used folios are
859 folio_set_referenced(folio
);
861 if (referenced_folio
|| referenced_ptes
> 1)
862 return FOLIOREF_ACTIVATE
;
865 * Activate file-backed executable folios after first usage.
867 if ((vm_flags
& VM_EXEC
) && folio_is_file_lru(folio
))
868 return FOLIOREF_ACTIVATE
;
870 return FOLIOREF_KEEP
;
873 /* Reclaim if clean, defer dirty folios to writeback */
874 if (referenced_folio
&& folio_is_file_lru(folio
))
875 return FOLIOREF_RECLAIM_CLEAN
;
877 return FOLIOREF_RECLAIM
;
880 /* Check if a folio is dirty or under writeback */
881 static void folio_check_dirty_writeback(struct folio
*folio
,
882 bool *dirty
, bool *writeback
)
884 struct address_space
*mapping
;
887 * Anonymous folios are not handled by flushers and must be written
888 * from reclaim context. Do not stall reclaim based on them.
889 * MADV_FREE anonymous folios are put into inactive file list too.
890 * They could be mistakenly treated as file lru. So further anon
893 if (!folio_is_file_lru(folio
) ||
894 (folio_test_anon(folio
) && !folio_test_swapbacked(folio
))) {
900 /* By default assume that the folio flags are accurate */
901 *dirty
= folio_test_dirty(folio
);
902 *writeback
= folio_test_writeback(folio
);
904 /* Verify dirty/writeback state if the filesystem supports it */
905 if (!folio_test_private(folio
))
908 mapping
= folio_mapping(folio
);
909 if (mapping
&& mapping
->a_ops
->is_dirty_writeback
)
910 mapping
->a_ops
->is_dirty_writeback(folio
, dirty
, writeback
);
913 static struct folio
*alloc_demote_folio(struct folio
*src
,
914 unsigned long private)
917 nodemask_t
*allowed_mask
;
918 struct migration_target_control
*mtc
;
920 mtc
= (struct migration_target_control
*)private;
922 allowed_mask
= mtc
->nmask
;
924 * make sure we allocate from the target node first also trying to
925 * demote or reclaim pages from the target node via kswapd if we are
926 * low on free memory on target node. If we don't do this and if
927 * we have free memory on the slower(lower) memtier, we would start
928 * allocating pages from slower(lower) memory tiers without even forcing
929 * a demotion of cold pages from the target memtier. This can result
930 * in the kernel placing hot pages in slower(lower) memory tiers.
933 mtc
->gfp_mask
|= __GFP_THISNODE
;
934 dst
= alloc_migration_target(src
, (unsigned long)mtc
);
938 mtc
->gfp_mask
&= ~__GFP_THISNODE
;
939 mtc
->nmask
= allowed_mask
;
941 return alloc_migration_target(src
, (unsigned long)mtc
);
945 * Take folios on @demote_folios and attempt to demote them to another node.
946 * Folios which are not demoted are left on @demote_folios.
948 static unsigned int demote_folio_list(struct list_head
*demote_folios
,
949 struct pglist_data
*pgdat
)
951 int target_nid
= next_demotion_node(pgdat
->node_id
);
952 unsigned int nr_succeeded
;
953 nodemask_t allowed_mask
;
955 struct migration_target_control mtc
= {
957 * Allocate from 'node', or fail quickly and quietly.
958 * When this happens, 'page' will likely just be discarded
959 * instead of migrated.
961 .gfp_mask
= (GFP_HIGHUSER_MOVABLE
& ~__GFP_RECLAIM
) | __GFP_NOWARN
|
962 __GFP_NOMEMALLOC
| GFP_NOWAIT
,
964 .nmask
= &allowed_mask
967 if (list_empty(demote_folios
))
970 if (target_nid
== NUMA_NO_NODE
)
973 node_get_allowed_targets(pgdat
, &allowed_mask
);
975 /* Demotion ignores all cpuset and mempolicy settings */
976 migrate_pages(demote_folios
, alloc_demote_folio
, NULL
,
977 (unsigned long)&mtc
, MIGRATE_ASYNC
, MR_DEMOTION
,
980 __count_vm_events(PGDEMOTE_KSWAPD
+ reclaimer_offset(), nr_succeeded
);
985 static bool may_enter_fs(struct folio
*folio
, gfp_t gfp_mask
)
987 if (gfp_mask
& __GFP_FS
)
989 if (!folio_test_swapcache(folio
) || !(gfp_mask
& __GFP_IO
))
992 * We can "enter_fs" for swap-cache with only __GFP_IO
993 * providing this isn't SWP_FS_OPS.
994 * ->flags can be updated non-atomicially (scan_swap_map_slots),
995 * but that will never affect SWP_FS_OPS, so the data_race
998 return !data_race(folio_swap_flags(folio
) & SWP_FS_OPS
);
1002 * shrink_folio_list() returns the number of reclaimed pages
1004 static unsigned int shrink_folio_list(struct list_head
*folio_list
,
1005 struct pglist_data
*pgdat
, struct scan_control
*sc
,
1006 struct reclaim_stat
*stat
, bool ignore_references
)
1008 LIST_HEAD(ret_folios
);
1009 LIST_HEAD(free_folios
);
1010 LIST_HEAD(demote_folios
);
1011 unsigned int nr_reclaimed
= 0;
1012 unsigned int pgactivate
= 0;
1013 bool do_demote_pass
;
1014 struct swap_iocb
*plug
= NULL
;
1016 memset(stat
, 0, sizeof(*stat
));
1018 do_demote_pass
= can_demote(pgdat
->node_id
, sc
);
1021 while (!list_empty(folio_list
)) {
1022 struct address_space
*mapping
;
1023 struct folio
*folio
;
1024 enum folio_references references
= FOLIOREF_RECLAIM
;
1025 bool dirty
, writeback
;
1026 unsigned int nr_pages
;
1030 folio
= lru_to_folio(folio_list
);
1031 list_del(&folio
->lru
);
1033 if (!folio_trylock(folio
))
1036 VM_BUG_ON_FOLIO(folio_test_active(folio
), folio
);
1038 nr_pages
= folio_nr_pages(folio
);
1040 /* Account the number of base pages */
1041 sc
->nr_scanned
+= nr_pages
;
1043 if (unlikely(!folio_evictable(folio
)))
1044 goto activate_locked
;
1046 if (!sc
->may_unmap
&& folio_mapped(folio
))
1049 /* folio_update_gen() tried to promote this page? */
1050 if (lru_gen_enabled() && !ignore_references
&&
1051 folio_mapped(folio
) && folio_test_referenced(folio
))
1055 * The number of dirty pages determines if a node is marked
1056 * reclaim_congested. kswapd will stall and start writing
1057 * folios if the tail of the LRU is all dirty unqueued folios.
1059 folio_check_dirty_writeback(folio
, &dirty
, &writeback
);
1060 if (dirty
|| writeback
)
1061 stat
->nr_dirty
+= nr_pages
;
1063 if (dirty
&& !writeback
)
1064 stat
->nr_unqueued_dirty
+= nr_pages
;
1067 * Treat this folio as congested if folios are cycling
1068 * through the LRU so quickly that the folios marked
1069 * for immediate reclaim are making it to the end of
1070 * the LRU a second time.
1072 if (writeback
&& folio_test_reclaim(folio
))
1073 stat
->nr_congested
+= nr_pages
;
1076 * If a folio at the tail of the LRU is under writeback, there
1077 * are three cases to consider.
1079 * 1) If reclaim is encountering an excessive number
1080 * of folios under writeback and this folio has both
1081 * the writeback and reclaim flags set, then it
1082 * indicates that folios are being queued for I/O but
1083 * are being recycled through the LRU before the I/O
1084 * can complete. Waiting on the folio itself risks an
1085 * indefinite stall if it is impossible to writeback
1086 * the folio due to I/O error or disconnected storage
1087 * so instead note that the LRU is being scanned too
1088 * quickly and the caller can stall after the folio
1089 * list has been processed.
1091 * 2) Global or new memcg reclaim encounters a folio that is
1092 * not marked for immediate reclaim, or the caller does not
1093 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1094 * not to fs). In this case mark the folio for immediate
1095 * reclaim and continue scanning.
1097 * Require may_enter_fs() because we would wait on fs, which
1098 * may not have submitted I/O yet. And the loop driver might
1099 * enter reclaim, and deadlock if it waits on a folio for
1100 * which it is needed to do the write (loop masks off
1101 * __GFP_IO|__GFP_FS for this reason); but more thought
1102 * would probably show more reasons.
1104 * 3) Legacy memcg encounters a folio that already has the
1105 * reclaim flag set. memcg does not have any dirty folio
1106 * throttling so we could easily OOM just because too many
1107 * folios are in writeback and there is nothing else to
1108 * reclaim. Wait for the writeback to complete.
1110 * In cases 1) and 2) we activate the folios to get them out of
1111 * the way while we continue scanning for clean folios on the
1112 * inactive list and refilling from the active list. The
1113 * observation here is that waiting for disk writes is more
1114 * expensive than potentially causing reloads down the line.
1115 * Since they're marked for immediate reclaim, they won't put
1116 * memory pressure on the cache working set any longer than it
1117 * takes to write them to disk.
1119 if (folio_test_writeback(folio
)) {
1121 if (current_is_kswapd() &&
1122 folio_test_reclaim(folio
) &&
1123 test_bit(PGDAT_WRITEBACK
, &pgdat
->flags
)) {
1124 stat
->nr_immediate
+= nr_pages
;
1125 goto activate_locked
;
1128 } else if (writeback_throttling_sane(sc
) ||
1129 !folio_test_reclaim(folio
) ||
1130 !may_enter_fs(folio
, sc
->gfp_mask
)) {
1132 * This is slightly racy -
1133 * folio_end_writeback() might have
1134 * just cleared the reclaim flag, then
1135 * setting the reclaim flag here ends up
1136 * interpreted as the readahead flag - but
1137 * that does not matter enough to care.
1138 * What we do want is for this folio to
1139 * have the reclaim flag set next time
1140 * memcg reclaim reaches the tests above,
1141 * so it will then wait for writeback to
1142 * avoid OOM; and it's also appropriate
1143 * in global reclaim.
1145 folio_set_reclaim(folio
);
1146 stat
->nr_writeback
+= nr_pages
;
1147 goto activate_locked
;
1151 folio_unlock(folio
);
1152 folio_wait_writeback(folio
);
1153 /* then go back and try same folio again */
1154 list_add_tail(&folio
->lru
, folio_list
);
1159 if (!ignore_references
)
1160 references
= folio_check_references(folio
, sc
);
1162 switch (references
) {
1163 case FOLIOREF_ACTIVATE
:
1164 goto activate_locked
;
1166 stat
->nr_ref_keep
+= nr_pages
;
1168 case FOLIOREF_RECLAIM
:
1169 case FOLIOREF_RECLAIM_CLEAN
:
1170 ; /* try to reclaim the folio below */
1174 * Before reclaiming the folio, try to relocate
1175 * its contents to another node.
1177 if (do_demote_pass
&&
1178 (thp_migration_supported() || !folio_test_large(folio
))) {
1179 list_add(&folio
->lru
, &demote_folios
);
1180 folio_unlock(folio
);
1185 * Anonymous process memory has backing store?
1186 * Try to allocate it some swap space here.
1187 * Lazyfree folio could be freed directly
1189 if (folio_test_anon(folio
) && folio_test_swapbacked(folio
)) {
1190 if (!folio_test_swapcache(folio
)) {
1191 if (!(sc
->gfp_mask
& __GFP_IO
))
1193 if (folio_maybe_dma_pinned(folio
))
1195 if (folio_test_large(folio
)) {
1196 /* cannot split folio, skip it */
1197 if (!can_split_folio(folio
, NULL
))
1198 goto activate_locked
;
1200 * Split folios without a PMD map right
1201 * away. Chances are some or all of the
1202 * tail pages can be freed without IO.
1204 if (!folio_entire_mapcount(folio
) &&
1205 split_folio_to_list(folio
,
1207 goto activate_locked
;
1209 if (!add_to_swap(folio
)) {
1210 if (!folio_test_large(folio
))
1211 goto activate_locked_split
;
1212 /* Fallback to swap normal pages */
1213 if (split_folio_to_list(folio
,
1215 goto activate_locked
;
1216 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1217 count_memcg_folio_events(folio
, THP_SWPOUT_FALLBACK
, 1);
1218 count_vm_event(THP_SWPOUT_FALLBACK
);
1220 if (!add_to_swap(folio
))
1221 goto activate_locked_split
;
1224 } else if (folio_test_swapbacked(folio
) &&
1225 folio_test_large(folio
)) {
1226 /* Split shmem folio */
1227 if (split_folio_to_list(folio
, folio_list
))
1232 * If the folio was split above, the tail pages will make
1233 * their own pass through this function and be accounted
1236 if ((nr_pages
> 1) && !folio_test_large(folio
)) {
1237 sc
->nr_scanned
-= (nr_pages
- 1);
1242 * The folio is mapped into the page tables of one or more
1243 * processes. Try to unmap it here.
1245 if (folio_mapped(folio
)) {
1246 enum ttu_flags flags
= TTU_BATCH_FLUSH
;
1247 bool was_swapbacked
= folio_test_swapbacked(folio
);
1249 if (folio_test_pmd_mappable(folio
))
1250 flags
|= TTU_SPLIT_HUGE_PMD
;
1252 try_to_unmap(folio
, flags
);
1253 if (folio_mapped(folio
)) {
1254 stat
->nr_unmap_fail
+= nr_pages
;
1255 if (!was_swapbacked
&&
1256 folio_test_swapbacked(folio
))
1257 stat
->nr_lazyfree_fail
+= nr_pages
;
1258 goto activate_locked
;
1263 * Folio is unmapped now so it cannot be newly pinned anymore.
1264 * No point in trying to reclaim folio if it is pinned.
1265 * Furthermore we don't want to reclaim underlying fs metadata
1266 * if the folio is pinned and thus potentially modified by the
1267 * pinning process as that may upset the filesystem.
1269 if (folio_maybe_dma_pinned(folio
))
1270 goto activate_locked
;
1272 mapping
= folio_mapping(folio
);
1273 if (folio_test_dirty(folio
)) {
1275 * Only kswapd can writeback filesystem folios
1276 * to avoid risk of stack overflow. But avoid
1277 * injecting inefficient single-folio I/O into
1278 * flusher writeback as much as possible: only
1279 * write folios when we've encountered many
1280 * dirty folios, and when we've already scanned
1281 * the rest of the LRU for clean folios and see
1282 * the same dirty folios again (with the reclaim
1285 if (folio_is_file_lru(folio
) &&
1286 (!current_is_kswapd() ||
1287 !folio_test_reclaim(folio
) ||
1288 !test_bit(PGDAT_DIRTY
, &pgdat
->flags
))) {
1290 * Immediately reclaim when written back.
1291 * Similar in principle to folio_deactivate()
1292 * except we already have the folio isolated
1293 * and know it's dirty
1295 node_stat_mod_folio(folio
, NR_VMSCAN_IMMEDIATE
,
1297 folio_set_reclaim(folio
);
1299 goto activate_locked
;
1302 if (references
== FOLIOREF_RECLAIM_CLEAN
)
1304 if (!may_enter_fs(folio
, sc
->gfp_mask
))
1306 if (!sc
->may_writepage
)
1310 * Folio is dirty. Flush the TLB if a writable entry
1311 * potentially exists to avoid CPU writes after I/O
1312 * starts and then write it out here.
1314 try_to_unmap_flush_dirty();
1315 switch (pageout(folio
, mapping
, &plug
)) {
1319 goto activate_locked
;
1321 stat
->nr_pageout
+= nr_pages
;
1323 if (folio_test_writeback(folio
))
1325 if (folio_test_dirty(folio
))
1329 * A synchronous write - probably a ramdisk. Go
1330 * ahead and try to reclaim the folio.
1332 if (!folio_trylock(folio
))
1334 if (folio_test_dirty(folio
) ||
1335 folio_test_writeback(folio
))
1337 mapping
= folio_mapping(folio
);
1340 ; /* try to free the folio below */
1345 * If the folio has buffers, try to free the buffer
1346 * mappings associated with this folio. If we succeed
1347 * we try to free the folio as well.
1349 * We do this even if the folio is dirty.
1350 * filemap_release_folio() does not perform I/O, but it
1351 * is possible for a folio to have the dirty flag set,
1352 * but it is actually clean (all its buffers are clean).
1353 * This happens if the buffers were written out directly,
1354 * with submit_bh(). ext3 will do this, as well as
1355 * the blockdev mapping. filemap_release_folio() will
1356 * discover that cleanness and will drop the buffers
1357 * and mark the folio clean - it can be freed.
1359 * Rarely, folios can have buffers and no ->mapping.
1360 * These are the folios which were not successfully
1361 * invalidated in truncate_cleanup_folio(). We try to
1362 * drop those buffers here and if that worked, and the
1363 * folio is no longer mapped into process address space
1364 * (refcount == 1) it can be freed. Otherwise, leave
1365 * the folio on the LRU so it is swappable.
1367 if (folio_needs_release(folio
)) {
1368 if (!filemap_release_folio(folio
, sc
->gfp_mask
))
1369 goto activate_locked
;
1370 if (!mapping
&& folio_ref_count(folio
) == 1) {
1371 folio_unlock(folio
);
1372 if (folio_put_testzero(folio
))
1376 * rare race with speculative reference.
1377 * the speculative reference will free
1378 * this folio shortly, so we may
1379 * increment nr_reclaimed here (and
1380 * leave it off the LRU).
1382 nr_reclaimed
+= nr_pages
;
1388 if (folio_test_anon(folio
) && !folio_test_swapbacked(folio
)) {
1389 /* follow __remove_mapping for reference */
1390 if (!folio_ref_freeze(folio
, 1))
1393 * The folio has only one reference left, which is
1394 * from the isolation. After the caller puts the
1395 * folio back on the lru and drops the reference, the
1396 * folio will be freed anyway. It doesn't matter
1397 * which lru it goes on. So we don't bother checking
1398 * the dirty flag here.
1400 count_vm_events(PGLAZYFREED
, nr_pages
);
1401 count_memcg_folio_events(folio
, PGLAZYFREED
, nr_pages
);
1402 } else if (!mapping
|| !__remove_mapping(mapping
, folio
, true,
1403 sc
->target_mem_cgroup
))
1406 folio_unlock(folio
);
1409 * Folio may get swapped out as a whole, need to account
1412 nr_reclaimed
+= nr_pages
;
1415 * Is there need to periodically free_folio_list? It would
1416 * appear not as the counts should be low
1418 if (unlikely(folio_test_large(folio
)))
1419 destroy_large_folio(folio
);
1421 list_add(&folio
->lru
, &free_folios
);
1424 activate_locked_split
:
1426 * The tail pages that are failed to add into swap cache
1427 * reach here. Fixup nr_scanned and nr_pages.
1430 sc
->nr_scanned
-= (nr_pages
- 1);
1434 /* Not a candidate for swapping, so reclaim swap space. */
1435 if (folio_test_swapcache(folio
) &&
1436 (mem_cgroup_swap_full(folio
) || folio_test_mlocked(folio
)))
1437 folio_free_swap(folio
);
1438 VM_BUG_ON_FOLIO(folio_test_active(folio
), folio
);
1439 if (!folio_test_mlocked(folio
)) {
1440 int type
= folio_is_file_lru(folio
);
1441 folio_set_active(folio
);
1442 stat
->nr_activate
[type
] += nr_pages
;
1443 count_memcg_folio_events(folio
, PGACTIVATE
, nr_pages
);
1446 folio_unlock(folio
);
1448 list_add(&folio
->lru
, &ret_folios
);
1449 VM_BUG_ON_FOLIO(folio_test_lru(folio
) ||
1450 folio_test_unevictable(folio
), folio
);
1452 /* 'folio_list' is always empty here */
1454 /* Migrate folios selected for demotion */
1455 nr_reclaimed
+= demote_folio_list(&demote_folios
, pgdat
);
1456 /* Folios that could not be demoted are still in @demote_folios */
1457 if (!list_empty(&demote_folios
)) {
1458 /* Folios which weren't demoted go back on @folio_list */
1459 list_splice_init(&demote_folios
, folio_list
);
1462 * goto retry to reclaim the undemoted folios in folio_list if
1465 * Reclaiming directly from top tier nodes is not often desired
1466 * due to it breaking the LRU ordering: in general memory
1467 * should be reclaimed from lower tier nodes and demoted from
1470 * However, disabling reclaim from top tier nodes entirely
1471 * would cause ooms in edge scenarios where lower tier memory
1472 * is unreclaimable for whatever reason, eg memory being
1473 * mlocked or too hot to reclaim. We can disable reclaim
1474 * from top tier nodes in proactive reclaim though as that is
1475 * not real memory pressure.
1477 if (!sc
->proactive
) {
1478 do_demote_pass
= false;
1483 pgactivate
= stat
->nr_activate
[0] + stat
->nr_activate
[1];
1485 mem_cgroup_uncharge_list(&free_folios
);
1486 try_to_unmap_flush();
1487 free_unref_page_list(&free_folios
);
1489 list_splice(&ret_folios
, folio_list
);
1490 count_vm_events(PGACTIVATE
, pgactivate
);
1493 swap_write_unplug(plug
);
1494 return nr_reclaimed
;
1497 unsigned int reclaim_clean_pages_from_list(struct zone
*zone
,
1498 struct list_head
*folio_list
)
1500 struct scan_control sc
= {
1501 .gfp_mask
= GFP_KERNEL
,
1504 struct reclaim_stat stat
;
1505 unsigned int nr_reclaimed
;
1506 struct folio
*folio
, *next
;
1507 LIST_HEAD(clean_folios
);
1508 unsigned int noreclaim_flag
;
1510 list_for_each_entry_safe(folio
, next
, folio_list
, lru
) {
1511 if (!folio_test_hugetlb(folio
) && folio_is_file_lru(folio
) &&
1512 !folio_test_dirty(folio
) && !__folio_test_movable(folio
) &&
1513 !folio_test_unevictable(folio
)) {
1514 folio_clear_active(folio
);
1515 list_move(&folio
->lru
, &clean_folios
);
1520 * We should be safe here since we are only dealing with file pages and
1521 * we are not kswapd and therefore cannot write dirty file pages. But
1522 * call memalloc_noreclaim_save() anyway, just in case these conditions
1523 * change in the future.
1525 noreclaim_flag
= memalloc_noreclaim_save();
1526 nr_reclaimed
= shrink_folio_list(&clean_folios
, zone
->zone_pgdat
, &sc
,
1528 memalloc_noreclaim_restore(noreclaim_flag
);
1530 list_splice(&clean_folios
, folio_list
);
1531 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1532 -(long)nr_reclaimed
);
1534 * Since lazyfree pages are isolated from file LRU from the beginning,
1535 * they will rotate back to anonymous LRU in the end if it failed to
1536 * discard so isolated count will be mismatched.
1537 * Compensate the isolated count for both LRU lists.
1539 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
,
1540 stat
.nr_lazyfree_fail
);
1541 mod_node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
,
1542 -(long)stat
.nr_lazyfree_fail
);
1543 return nr_reclaimed
;
1547 * Update LRU sizes after isolating pages. The LRU size updates must
1548 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1550 static __always_inline
void update_lru_sizes(struct lruvec
*lruvec
,
1551 enum lru_list lru
, unsigned long *nr_zone_taken
)
1555 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1556 if (!nr_zone_taken
[zid
])
1559 update_lru_size(lruvec
, lru
, zid
, -nr_zone_taken
[zid
]);
1566 * It is waste of effort to scan and reclaim CMA pages if it is not available
1567 * for current allocation context. Kswapd can not be enrolled as it can not
1568 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1570 static bool skip_cma(struct folio
*folio
, struct scan_control
*sc
)
1572 return !current_is_kswapd() &&
1573 gfp_migratetype(sc
->gfp_mask
) != MIGRATE_MOVABLE
&&
1574 folio_migratetype(folio
) == MIGRATE_CMA
;
1577 static bool skip_cma(struct folio
*folio
, struct scan_control
*sc
)
1584 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1586 * lruvec->lru_lock is heavily contended. Some of the functions that
1587 * shrink the lists perform better by taking out a batch of pages
1588 * and working on them outside the LRU lock.
1590 * For pagecache intensive workloads, this function is the hottest
1591 * spot in the kernel (apart from copy_*_user functions).
1593 * Lru_lock must be held before calling this function.
1595 * @nr_to_scan: The number of eligible pages to look through on the list.
1596 * @lruvec: The LRU vector to pull pages from.
1597 * @dst: The temp list to put pages on to.
1598 * @nr_scanned: The number of pages that were scanned.
1599 * @sc: The scan_control struct for this reclaim session
1600 * @lru: LRU list id for isolating
1602 * returns how many pages were moved onto *@dst.
1604 static unsigned long isolate_lru_folios(unsigned long nr_to_scan
,
1605 struct lruvec
*lruvec
, struct list_head
*dst
,
1606 unsigned long *nr_scanned
, struct scan_control
*sc
,
1609 struct list_head
*src
= &lruvec
->lists
[lru
];
1610 unsigned long nr_taken
= 0;
1611 unsigned long nr_zone_taken
[MAX_NR_ZONES
] = { 0 };
1612 unsigned long nr_skipped
[MAX_NR_ZONES
] = { 0, };
1613 unsigned long skipped
= 0;
1614 unsigned long scan
, total_scan
, nr_pages
;
1615 LIST_HEAD(folios_skipped
);
1619 while (scan
< nr_to_scan
&& !list_empty(src
)) {
1620 struct list_head
*move_to
= src
;
1621 struct folio
*folio
;
1623 folio
= lru_to_folio(src
);
1624 prefetchw_prev_lru_folio(folio
, src
, flags
);
1626 nr_pages
= folio_nr_pages(folio
);
1627 total_scan
+= nr_pages
;
1629 if (folio_zonenum(folio
) > sc
->reclaim_idx
||
1630 skip_cma(folio
, sc
)) {
1631 nr_skipped
[folio_zonenum(folio
)] += nr_pages
;
1632 move_to
= &folios_skipped
;
1637 * Do not count skipped folios because that makes the function
1638 * return with no isolated folios if the LRU mostly contains
1639 * ineligible folios. This causes the VM to not reclaim any
1640 * folios, triggering a premature OOM.
1641 * Account all pages in a folio.
1645 if (!folio_test_lru(folio
))
1647 if (!sc
->may_unmap
&& folio_mapped(folio
))
1651 * Be careful not to clear the lru flag until after we're
1652 * sure the folio is not being freed elsewhere -- the
1653 * folio release code relies on it.
1655 if (unlikely(!folio_try_get(folio
)))
1658 if (!folio_test_clear_lru(folio
)) {
1659 /* Another thread is already isolating this folio */
1664 nr_taken
+= nr_pages
;
1665 nr_zone_taken
[folio_zonenum(folio
)] += nr_pages
;
1668 list_move(&folio
->lru
, move_to
);
1672 * Splice any skipped folios to the start of the LRU list. Note that
1673 * this disrupts the LRU order when reclaiming for lower zones but
1674 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1675 * scanning would soon rescan the same folios to skip and waste lots
1678 if (!list_empty(&folios_skipped
)) {
1681 list_splice(&folios_skipped
, src
);
1682 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1683 if (!nr_skipped
[zid
])
1686 __count_zid_vm_events(PGSCAN_SKIP
, zid
, nr_skipped
[zid
]);
1687 skipped
+= nr_skipped
[zid
];
1690 *nr_scanned
= total_scan
;
1691 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, nr_to_scan
,
1692 total_scan
, skipped
, nr_taken
, lru
);
1693 update_lru_sizes(lruvec
, lru
, nr_zone_taken
);
1698 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1699 * @folio: Folio to isolate from its LRU list.
1701 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1702 * corresponding to whatever LRU list the folio was on.
1704 * The folio will have its LRU flag cleared. If it was found on the
1705 * active list, it will have the Active flag set. If it was found on the
1706 * unevictable list, it will have the Unevictable flag set. These flags
1707 * may need to be cleared by the caller before letting the page go.
1711 * (1) Must be called with an elevated refcount on the folio. This is a
1712 * fundamental difference from isolate_lru_folios() (which is called
1713 * without a stable reference).
1714 * (2) The lru_lock must not be held.
1715 * (3) Interrupts must be enabled.
1717 * Return: true if the folio was removed from an LRU list.
1718 * false if the folio was not on an LRU list.
1720 bool folio_isolate_lru(struct folio
*folio
)
1724 VM_BUG_ON_FOLIO(!folio_ref_count(folio
), folio
);
1726 if (folio_test_clear_lru(folio
)) {
1727 struct lruvec
*lruvec
;
1730 lruvec
= folio_lruvec_lock_irq(folio
);
1731 lruvec_del_folio(lruvec
, folio
);
1732 unlock_page_lruvec_irq(lruvec
);
1740 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1741 * then get rescheduled. When there are massive number of tasks doing page
1742 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1743 * the LRU list will go small and be scanned faster than necessary, leading to
1744 * unnecessary swapping, thrashing and OOM.
1746 static int too_many_isolated(struct pglist_data
*pgdat
, int file
,
1747 struct scan_control
*sc
)
1749 unsigned long inactive
, isolated
;
1752 if (current_is_kswapd())
1755 if (!writeback_throttling_sane(sc
))
1759 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
1760 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
);
1762 inactive
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
1763 isolated
= node_page_state(pgdat
, NR_ISOLATED_ANON
);
1767 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1768 * won't get blocked by normal direct-reclaimers, forming a circular
1771 if (gfp_has_io_fs(sc
->gfp_mask
))
1774 too_many
= isolated
> inactive
;
1776 /* Wake up tasks throttled due to too_many_isolated. */
1778 wake_throttle_isolated(pgdat
);
1784 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1785 * On return, @list is reused as a list of folios to be freed by the caller.
1787 * Returns the number of pages moved to the given lruvec.
1789 static unsigned int move_folios_to_lru(struct lruvec
*lruvec
,
1790 struct list_head
*list
)
1792 int nr_pages
, nr_moved
= 0;
1793 LIST_HEAD(folios_to_free
);
1795 while (!list_empty(list
)) {
1796 struct folio
*folio
= lru_to_folio(list
);
1798 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
1799 list_del(&folio
->lru
);
1800 if (unlikely(!folio_evictable(folio
))) {
1801 spin_unlock_irq(&lruvec
->lru_lock
);
1802 folio_putback_lru(folio
);
1803 spin_lock_irq(&lruvec
->lru_lock
);
1808 * The folio_set_lru needs to be kept here for list integrity.
1810 * #0 move_folios_to_lru #1 release_pages
1811 * if (!folio_put_testzero())
1812 * if (folio_put_testzero())
1813 * !lru //skip lru_lock
1815 * list_add(&folio->lru,)
1816 * list_add(&folio->lru,)
1818 folio_set_lru(folio
);
1820 if (unlikely(folio_put_testzero(folio
))) {
1821 __folio_clear_lru_flags(folio
);
1823 if (unlikely(folio_test_large(folio
))) {
1824 spin_unlock_irq(&lruvec
->lru_lock
);
1825 destroy_large_folio(folio
);
1826 spin_lock_irq(&lruvec
->lru_lock
);
1828 list_add(&folio
->lru
, &folios_to_free
);
1834 * All pages were isolated from the same lruvec (and isolation
1835 * inhibits memcg migration).
1837 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio
, lruvec
), folio
);
1838 lruvec_add_folio(lruvec
, folio
);
1839 nr_pages
= folio_nr_pages(folio
);
1840 nr_moved
+= nr_pages
;
1841 if (folio_test_active(folio
))
1842 workingset_age_nonresident(lruvec
, nr_pages
);
1846 * To save our caller's stack, now use input list for pages to free.
1848 list_splice(&folios_to_free
, list
);
1854 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1855 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1856 * we should not throttle. Otherwise it is safe to do so.
1858 static int current_may_throttle(void)
1860 return !(current
->flags
& PF_LOCAL_THROTTLE
);
1864 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1865 * of reclaimed pages
1867 static unsigned long shrink_inactive_list(unsigned long nr_to_scan
,
1868 struct lruvec
*lruvec
, struct scan_control
*sc
,
1871 LIST_HEAD(folio_list
);
1872 unsigned long nr_scanned
;
1873 unsigned int nr_reclaimed
= 0;
1874 unsigned long nr_taken
;
1875 struct reclaim_stat stat
;
1876 bool file
= is_file_lru(lru
);
1877 enum vm_event_item item
;
1878 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
1879 bool stalled
= false;
1881 while (unlikely(too_many_isolated(pgdat
, file
, sc
))) {
1885 /* wait a bit for the reclaimer. */
1887 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_ISOLATED
);
1889 /* We are about to die and free our memory. Return now. */
1890 if (fatal_signal_pending(current
))
1891 return SWAP_CLUSTER_MAX
;
1896 spin_lock_irq(&lruvec
->lru_lock
);
1898 nr_taken
= isolate_lru_folios(nr_to_scan
, lruvec
, &folio_list
,
1899 &nr_scanned
, sc
, lru
);
1901 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1902 item
= PGSCAN_KSWAPD
+ reclaimer_offset();
1903 if (!cgroup_reclaim(sc
))
1904 __count_vm_events(item
, nr_scanned
);
1905 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_scanned
);
1906 __count_vm_events(PGSCAN_ANON
+ file
, nr_scanned
);
1908 spin_unlock_irq(&lruvec
->lru_lock
);
1913 nr_reclaimed
= shrink_folio_list(&folio_list
, pgdat
, sc
, &stat
, false);
1915 spin_lock_irq(&lruvec
->lru_lock
);
1916 move_folios_to_lru(lruvec
, &folio_list
);
1918 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1919 item
= PGSTEAL_KSWAPD
+ reclaimer_offset();
1920 if (!cgroup_reclaim(sc
))
1921 __count_vm_events(item
, nr_reclaimed
);
1922 __count_memcg_events(lruvec_memcg(lruvec
), item
, nr_reclaimed
);
1923 __count_vm_events(PGSTEAL_ANON
+ file
, nr_reclaimed
);
1924 spin_unlock_irq(&lruvec
->lru_lock
);
1926 lru_note_cost(lruvec
, file
, stat
.nr_pageout
, nr_scanned
- nr_reclaimed
);
1927 mem_cgroup_uncharge_list(&folio_list
);
1928 free_unref_page_list(&folio_list
);
1931 * If dirty folios are scanned that are not queued for IO, it
1932 * implies that flushers are not doing their job. This can
1933 * happen when memory pressure pushes dirty folios to the end of
1934 * the LRU before the dirty limits are breached and the dirty
1935 * data has expired. It can also happen when the proportion of
1936 * dirty folios grows not through writes but through memory
1937 * pressure reclaiming all the clean cache. And in some cases,
1938 * the flushers simply cannot keep up with the allocation
1939 * rate. Nudge the flusher threads in case they are asleep.
1941 if (stat
.nr_unqueued_dirty
== nr_taken
) {
1942 wakeup_flusher_threads(WB_REASON_VMSCAN
);
1944 * For cgroupv1 dirty throttling is achieved by waking up
1945 * the kernel flusher here and later waiting on folios
1946 * which are in writeback to finish (see shrink_folio_list()).
1948 * Flusher may not be able to issue writeback quickly
1949 * enough for cgroupv1 writeback throttling to work
1950 * on a large system.
1952 if (!writeback_throttling_sane(sc
))
1953 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_WRITEBACK
);
1956 sc
->nr
.dirty
+= stat
.nr_dirty
;
1957 sc
->nr
.congested
+= stat
.nr_congested
;
1958 sc
->nr
.unqueued_dirty
+= stat
.nr_unqueued_dirty
;
1959 sc
->nr
.writeback
+= stat
.nr_writeback
;
1960 sc
->nr
.immediate
+= stat
.nr_immediate
;
1961 sc
->nr
.taken
+= nr_taken
;
1963 sc
->nr
.file_taken
+= nr_taken
;
1965 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
1966 nr_scanned
, nr_reclaimed
, &stat
, sc
->priority
, file
);
1967 return nr_reclaimed
;
1971 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1973 * We move them the other way if the folio is referenced by one or more
1976 * If the folios are mostly unmapped, the processing is fast and it is
1977 * appropriate to hold lru_lock across the whole operation. But if
1978 * the folios are mapped, the processing is slow (folio_referenced()), so
1979 * we should drop lru_lock around each folio. It's impossible to balance
1980 * this, so instead we remove the folios from the LRU while processing them.
1981 * It is safe to rely on the active flag against the non-LRU folios in here
1982 * because nobody will play with that bit on a non-LRU folio.
1984 * The downside is that we have to touch folio->_refcount against each folio.
1985 * But we had to alter folio->flags anyway.
1987 static void shrink_active_list(unsigned long nr_to_scan
,
1988 struct lruvec
*lruvec
,
1989 struct scan_control
*sc
,
1992 unsigned long nr_taken
;
1993 unsigned long nr_scanned
;
1994 unsigned long vm_flags
;
1995 LIST_HEAD(l_hold
); /* The folios which were snipped off */
1996 LIST_HEAD(l_active
);
1997 LIST_HEAD(l_inactive
);
1998 unsigned nr_deactivate
, nr_activate
;
1999 unsigned nr_rotated
= 0;
2000 int file
= is_file_lru(lru
);
2001 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2005 spin_lock_irq(&lruvec
->lru_lock
);
2007 nr_taken
= isolate_lru_folios(nr_to_scan
, lruvec
, &l_hold
,
2008 &nr_scanned
, sc
, lru
);
2010 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, nr_taken
);
2012 if (!cgroup_reclaim(sc
))
2013 __count_vm_events(PGREFILL
, nr_scanned
);
2014 __count_memcg_events(lruvec_memcg(lruvec
), PGREFILL
, nr_scanned
);
2016 spin_unlock_irq(&lruvec
->lru_lock
);
2018 while (!list_empty(&l_hold
)) {
2019 struct folio
*folio
;
2022 folio
= lru_to_folio(&l_hold
);
2023 list_del(&folio
->lru
);
2025 if (unlikely(!folio_evictable(folio
))) {
2026 folio_putback_lru(folio
);
2030 if (unlikely(buffer_heads_over_limit
)) {
2031 if (folio_needs_release(folio
) &&
2032 folio_trylock(folio
)) {
2033 filemap_release_folio(folio
, 0);
2034 folio_unlock(folio
);
2038 /* Referenced or rmap lock contention: rotate */
2039 if (folio_referenced(folio
, 0, sc
->target_mem_cgroup
,
2042 * Identify referenced, file-backed active folios and
2043 * give them one more trip around the active list. So
2044 * that executable code get better chances to stay in
2045 * memory under moderate memory pressure. Anon folios
2046 * are not likely to be evicted by use-once streaming
2047 * IO, plus JVM can create lots of anon VM_EXEC folios,
2048 * so we ignore them here.
2050 if ((vm_flags
& VM_EXEC
) && folio_is_file_lru(folio
)) {
2051 nr_rotated
+= folio_nr_pages(folio
);
2052 list_add(&folio
->lru
, &l_active
);
2057 folio_clear_active(folio
); /* we are de-activating */
2058 folio_set_workingset(folio
);
2059 list_add(&folio
->lru
, &l_inactive
);
2063 * Move folios back to the lru list.
2065 spin_lock_irq(&lruvec
->lru_lock
);
2067 nr_activate
= move_folios_to_lru(lruvec
, &l_active
);
2068 nr_deactivate
= move_folios_to_lru(lruvec
, &l_inactive
);
2069 /* Keep all free folios in l_active list */
2070 list_splice(&l_inactive
, &l_active
);
2072 __count_vm_events(PGDEACTIVATE
, nr_deactivate
);
2073 __count_memcg_events(lruvec_memcg(lruvec
), PGDEACTIVATE
, nr_deactivate
);
2075 __mod_node_page_state(pgdat
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
2076 spin_unlock_irq(&lruvec
->lru_lock
);
2079 lru_note_cost(lruvec
, file
, 0, nr_rotated
);
2080 mem_cgroup_uncharge_list(&l_active
);
2081 free_unref_page_list(&l_active
);
2082 trace_mm_vmscan_lru_shrink_active(pgdat
->node_id
, nr_taken
, nr_activate
,
2083 nr_deactivate
, nr_rotated
, sc
->priority
, file
);
2086 static unsigned int reclaim_folio_list(struct list_head
*folio_list
,
2087 struct pglist_data
*pgdat
)
2089 struct reclaim_stat dummy_stat
;
2090 unsigned int nr_reclaimed
;
2091 struct folio
*folio
;
2092 struct scan_control sc
= {
2093 .gfp_mask
= GFP_KERNEL
,
2100 nr_reclaimed
= shrink_folio_list(folio_list
, pgdat
, &sc
, &dummy_stat
, false);
2101 while (!list_empty(folio_list
)) {
2102 folio
= lru_to_folio(folio_list
);
2103 list_del(&folio
->lru
);
2104 folio_putback_lru(folio
);
2107 return nr_reclaimed
;
2110 unsigned long reclaim_pages(struct list_head
*folio_list
)
2113 unsigned int nr_reclaimed
= 0;
2114 LIST_HEAD(node_folio_list
);
2115 unsigned int noreclaim_flag
;
2117 if (list_empty(folio_list
))
2118 return nr_reclaimed
;
2120 noreclaim_flag
= memalloc_noreclaim_save();
2122 nid
= folio_nid(lru_to_folio(folio_list
));
2124 struct folio
*folio
= lru_to_folio(folio_list
);
2126 if (nid
== folio_nid(folio
)) {
2127 folio_clear_active(folio
);
2128 list_move(&folio
->lru
, &node_folio_list
);
2132 nr_reclaimed
+= reclaim_folio_list(&node_folio_list
, NODE_DATA(nid
));
2133 nid
= folio_nid(lru_to_folio(folio_list
));
2134 } while (!list_empty(folio_list
));
2136 nr_reclaimed
+= reclaim_folio_list(&node_folio_list
, NODE_DATA(nid
));
2138 memalloc_noreclaim_restore(noreclaim_flag
);
2140 return nr_reclaimed
;
2143 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
2144 struct lruvec
*lruvec
, struct scan_control
*sc
)
2146 if (is_active_lru(lru
)) {
2147 if (sc
->may_deactivate
& (1 << is_file_lru(lru
)))
2148 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
2150 sc
->skipped_deactivate
= 1;
2154 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
2158 * The inactive anon list should be small enough that the VM never has
2159 * to do too much work.
2161 * The inactive file list should be small enough to leave most memory
2162 * to the established workingset on the scan-resistant active list,
2163 * but large enough to avoid thrashing the aggregate readahead window.
2165 * Both inactive lists should also be large enough that each inactive
2166 * folio has a chance to be referenced again before it is reclaimed.
2168 * If that fails and refaulting is observed, the inactive list grows.
2170 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2171 * on this LRU, maintained by the pageout code. An inactive_ratio
2172 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2175 * memory ratio inactive
2176 * -------------------------------------
2185 static bool inactive_is_low(struct lruvec
*lruvec
, enum lru_list inactive_lru
)
2187 enum lru_list active_lru
= inactive_lru
+ LRU_ACTIVE
;
2188 unsigned long inactive
, active
;
2189 unsigned long inactive_ratio
;
2192 inactive
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ inactive_lru
);
2193 active
= lruvec_page_state(lruvec
, NR_LRU_BASE
+ active_lru
);
2195 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
2197 inactive_ratio
= int_sqrt(10 * gb
);
2201 return inactive
* inactive_ratio
< active
;
2211 static void prepare_scan_control(pg_data_t
*pgdat
, struct scan_control
*sc
)
2214 struct lruvec
*target_lruvec
;
2216 if (lru_gen_enabled())
2219 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
2222 * Flush the memory cgroup stats, so that we read accurate per-memcg
2223 * lruvec stats for heuristics.
2225 mem_cgroup_flush_stats();
2228 * Determine the scan balance between anon and file LRUs.
2230 spin_lock_irq(&target_lruvec
->lru_lock
);
2231 sc
->anon_cost
= target_lruvec
->anon_cost
;
2232 sc
->file_cost
= target_lruvec
->file_cost
;
2233 spin_unlock_irq(&target_lruvec
->lru_lock
);
2236 * Target desirable inactive:active list ratios for the anon
2237 * and file LRU lists.
2239 if (!sc
->force_deactivate
) {
2240 unsigned long refaults
;
2243 * When refaults are being observed, it means a new
2244 * workingset is being established. Deactivate to get
2245 * rid of any stale active pages quickly.
2247 refaults
= lruvec_page_state(target_lruvec
,
2248 WORKINGSET_ACTIVATE_ANON
);
2249 if (refaults
!= target_lruvec
->refaults
[WORKINGSET_ANON
] ||
2250 inactive_is_low(target_lruvec
, LRU_INACTIVE_ANON
))
2251 sc
->may_deactivate
|= DEACTIVATE_ANON
;
2253 sc
->may_deactivate
&= ~DEACTIVATE_ANON
;
2255 refaults
= lruvec_page_state(target_lruvec
,
2256 WORKINGSET_ACTIVATE_FILE
);
2257 if (refaults
!= target_lruvec
->refaults
[WORKINGSET_FILE
] ||
2258 inactive_is_low(target_lruvec
, LRU_INACTIVE_FILE
))
2259 sc
->may_deactivate
|= DEACTIVATE_FILE
;
2261 sc
->may_deactivate
&= ~DEACTIVATE_FILE
;
2263 sc
->may_deactivate
= DEACTIVATE_ANON
| DEACTIVATE_FILE
;
2266 * If we have plenty of inactive file pages that aren't
2267 * thrashing, try to reclaim those first before touching
2270 file
= lruvec_page_state(target_lruvec
, NR_INACTIVE_FILE
);
2271 if (file
>> sc
->priority
&& !(sc
->may_deactivate
& DEACTIVATE_FILE
))
2272 sc
->cache_trim_mode
= 1;
2274 sc
->cache_trim_mode
= 0;
2277 * Prevent the reclaimer from falling into the cache trap: as
2278 * cache pages start out inactive, every cache fault will tip
2279 * the scan balance towards the file LRU. And as the file LRU
2280 * shrinks, so does the window for rotation from references.
2281 * This means we have a runaway feedback loop where a tiny
2282 * thrashing file LRU becomes infinitely more attractive than
2283 * anon pages. Try to detect this based on file LRU size.
2285 if (!cgroup_reclaim(sc
)) {
2286 unsigned long total_high_wmark
= 0;
2287 unsigned long free
, anon
;
2290 free
= sum_zone_node_page_state(pgdat
->node_id
, NR_FREE_PAGES
);
2291 file
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
2292 node_page_state(pgdat
, NR_INACTIVE_FILE
);
2294 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
2295 struct zone
*zone
= &pgdat
->node_zones
[z
];
2297 if (!managed_zone(zone
))
2300 total_high_wmark
+= high_wmark_pages(zone
);
2304 * Consider anon: if that's low too, this isn't a
2305 * runaway file reclaim problem, but rather just
2306 * extreme pressure. Reclaim as per usual then.
2308 anon
= node_page_state(pgdat
, NR_INACTIVE_ANON
);
2311 file
+ free
<= total_high_wmark
&&
2312 !(sc
->may_deactivate
& DEACTIVATE_ANON
) &&
2313 anon
>> sc
->priority
;
2318 * Determine how aggressively the anon and file LRU lists should be
2321 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2322 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2324 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
2327 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2328 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2329 unsigned long anon_cost
, file_cost
, total_cost
;
2330 int swappiness
= mem_cgroup_swappiness(memcg
);
2331 u64 fraction
[ANON_AND_FILE
];
2332 u64 denominator
= 0; /* gcc */
2333 enum scan_balance scan_balance
;
2334 unsigned long ap
, fp
;
2337 /* If we have no swap space, do not bother scanning anon folios. */
2338 if (!sc
->may_swap
|| !can_reclaim_anon_pages(memcg
, pgdat
->node_id
, sc
)) {
2339 scan_balance
= SCAN_FILE
;
2344 * Global reclaim will swap to prevent OOM even with no
2345 * swappiness, but memcg users want to use this knob to
2346 * disable swapping for individual groups completely when
2347 * using the memory controller's swap limit feature would be
2350 if (cgroup_reclaim(sc
) && !swappiness
) {
2351 scan_balance
= SCAN_FILE
;
2356 * Do not apply any pressure balancing cleverness when the
2357 * system is close to OOM, scan both anon and file equally
2358 * (unless the swappiness setting disagrees with swapping).
2360 if (!sc
->priority
&& swappiness
) {
2361 scan_balance
= SCAN_EQUAL
;
2366 * If the system is almost out of file pages, force-scan anon.
2368 if (sc
->file_is_tiny
) {
2369 scan_balance
= SCAN_ANON
;
2374 * If there is enough inactive page cache, we do not reclaim
2375 * anything from the anonymous working right now.
2377 if (sc
->cache_trim_mode
) {
2378 scan_balance
= SCAN_FILE
;
2382 scan_balance
= SCAN_FRACT
;
2384 * Calculate the pressure balance between anon and file pages.
2386 * The amount of pressure we put on each LRU is inversely
2387 * proportional to the cost of reclaiming each list, as
2388 * determined by the share of pages that are refaulting, times
2389 * the relative IO cost of bringing back a swapped out
2390 * anonymous page vs reloading a filesystem page (swappiness).
2392 * Although we limit that influence to ensure no list gets
2393 * left behind completely: at least a third of the pressure is
2394 * applied, before swappiness.
2396 * With swappiness at 100, anon and file have equal IO cost.
2398 total_cost
= sc
->anon_cost
+ sc
->file_cost
;
2399 anon_cost
= total_cost
+ sc
->anon_cost
;
2400 file_cost
= total_cost
+ sc
->file_cost
;
2401 total_cost
= anon_cost
+ file_cost
;
2403 ap
= swappiness
* (total_cost
+ 1);
2404 ap
/= anon_cost
+ 1;
2406 fp
= (200 - swappiness
) * (total_cost
+ 1);
2407 fp
/= file_cost
+ 1;
2411 denominator
= ap
+ fp
;
2413 for_each_evictable_lru(lru
) {
2414 int file
= is_file_lru(lru
);
2415 unsigned long lruvec_size
;
2416 unsigned long low
, min
;
2419 lruvec_size
= lruvec_lru_size(lruvec
, lru
, sc
->reclaim_idx
);
2420 mem_cgroup_protection(sc
->target_mem_cgroup
, memcg
,
2425 * Scale a cgroup's reclaim pressure by proportioning
2426 * its current usage to its memory.low or memory.min
2429 * This is important, as otherwise scanning aggression
2430 * becomes extremely binary -- from nothing as we
2431 * approach the memory protection threshold, to totally
2432 * nominal as we exceed it. This results in requiring
2433 * setting extremely liberal protection thresholds. It
2434 * also means we simply get no protection at all if we
2435 * set it too low, which is not ideal.
2437 * If there is any protection in place, we reduce scan
2438 * pressure by how much of the total memory used is
2439 * within protection thresholds.
2441 * There is one special case: in the first reclaim pass,
2442 * we skip over all groups that are within their low
2443 * protection. If that fails to reclaim enough pages to
2444 * satisfy the reclaim goal, we come back and override
2445 * the best-effort low protection. However, we still
2446 * ideally want to honor how well-behaved groups are in
2447 * that case instead of simply punishing them all
2448 * equally. As such, we reclaim them based on how much
2449 * memory they are using, reducing the scan pressure
2450 * again by how much of the total memory used is under
2453 unsigned long cgroup_size
= mem_cgroup_size(memcg
);
2454 unsigned long protection
;
2456 /* memory.low scaling, make sure we retry before OOM */
2457 if (!sc
->memcg_low_reclaim
&& low
> min
) {
2459 sc
->memcg_low_skipped
= 1;
2464 /* Avoid TOCTOU with earlier protection check */
2465 cgroup_size
= max(cgroup_size
, protection
);
2467 scan
= lruvec_size
- lruvec_size
* protection
/
2471 * Minimally target SWAP_CLUSTER_MAX pages to keep
2472 * reclaim moving forwards, avoiding decrementing
2473 * sc->priority further than desirable.
2475 scan
= max(scan
, SWAP_CLUSTER_MAX
);
2480 scan
>>= sc
->priority
;
2483 * If the cgroup's already been deleted, make sure to
2484 * scrape out the remaining cache.
2486 if (!scan
&& !mem_cgroup_online(memcg
))
2487 scan
= min(lruvec_size
, SWAP_CLUSTER_MAX
);
2489 switch (scan_balance
) {
2491 /* Scan lists relative to size */
2495 * Scan types proportional to swappiness and
2496 * their relative recent reclaim efficiency.
2497 * Make sure we don't miss the last page on
2498 * the offlined memory cgroups because of a
2501 scan
= mem_cgroup_online(memcg
) ?
2502 div64_u64(scan
* fraction
[file
], denominator
) :
2503 DIV64_U64_ROUND_UP(scan
* fraction
[file
],
2508 /* Scan one type exclusively */
2509 if ((scan_balance
== SCAN_FILE
) != file
)
2513 /* Look ma, no brain */
2522 * Anonymous LRU management is a waste if there is
2523 * ultimately no way to reclaim the memory.
2525 static bool can_age_anon_pages(struct pglist_data
*pgdat
,
2526 struct scan_control
*sc
)
2528 /* Aging the anon LRU is valuable if swap is present: */
2529 if (total_swap_pages
> 0)
2532 /* Also valuable if anon pages can be demoted: */
2533 return can_demote(pgdat
->node_id
, sc
);
2536 #ifdef CONFIG_LRU_GEN
2538 #ifdef CONFIG_LRU_GEN_ENABLED
2539 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps
, NR_LRU_GEN_CAPS
);
2540 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2542 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps
, NR_LRU_GEN_CAPS
);
2543 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2546 static bool should_walk_mmu(void)
2548 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK
);
2551 static bool should_clear_pmd_young(void)
2553 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG
);
2556 /******************************************************************************
2558 ******************************************************************************/
2560 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2562 #define DEFINE_MAX_SEQ(lruvec) \
2563 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2565 #define DEFINE_MIN_SEQ(lruvec) \
2566 unsigned long min_seq[ANON_AND_FILE] = { \
2567 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2568 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2571 #define for_each_gen_type_zone(gen, type, zone) \
2572 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2573 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2574 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2576 #define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2577 #define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2579 static struct lruvec
*get_lruvec(struct mem_cgroup
*memcg
, int nid
)
2581 struct pglist_data
*pgdat
= NODE_DATA(nid
);
2585 struct lruvec
*lruvec
= &memcg
->nodeinfo
[nid
]->lruvec
;
2587 /* see the comment in mem_cgroup_lruvec() */
2589 lruvec
->pgdat
= pgdat
;
2594 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2596 return &pgdat
->__lruvec
;
2599 static int get_swappiness(struct lruvec
*lruvec
, struct scan_control
*sc
)
2601 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2602 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
2607 if (!can_demote(pgdat
->node_id
, sc
) &&
2608 mem_cgroup_get_nr_swap_pages(memcg
) < MIN_LRU_BATCH
)
2611 return mem_cgroup_swappiness(memcg
);
2614 static int get_nr_gens(struct lruvec
*lruvec
, int type
)
2616 return lruvec
->lrugen
.max_seq
- lruvec
->lrugen
.min_seq
[type
] + 1;
2619 static bool __maybe_unused
seq_is_valid(struct lruvec
*lruvec
)
2621 /* see the comment on lru_gen_folio */
2622 return get_nr_gens(lruvec
, LRU_GEN_FILE
) >= MIN_NR_GENS
&&
2623 get_nr_gens(lruvec
, LRU_GEN_FILE
) <= get_nr_gens(lruvec
, LRU_GEN_ANON
) &&
2624 get_nr_gens(lruvec
, LRU_GEN_ANON
) <= MAX_NR_GENS
;
2627 /******************************************************************************
2629 ******************************************************************************/
2632 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2633 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2634 * bits in a bitmap, k is the number of hash functions and n is the number of
2637 * Page table walkers use one of the two filters to reduce their search space.
2638 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2639 * aging uses the double-buffering technique to flip to the other filter each
2640 * time it produces a new generation. For non-leaf entries that have enough
2641 * leaf entries, the aging carries them over to the next generation in
2642 * walk_pmd_range(); the eviction also report them when walking the rmap
2643 * in lru_gen_look_around().
2645 * For future optimizations:
2646 * 1. It's not necessary to keep both filters all the time. The spare one can be
2647 * freed after the RCU grace period and reallocated if needed again.
2648 * 2. And when reallocating, it's worth scaling its size according to the number
2649 * of inserted entries in the other filter, to reduce the memory overhead on
2650 * small systems and false positives on large systems.
2651 * 3. Jenkins' hash function is an alternative to Knuth's.
2653 #define BLOOM_FILTER_SHIFT 15
2655 static inline int filter_gen_from_seq(unsigned long seq
)
2657 return seq
% NR_BLOOM_FILTERS
;
2660 static void get_item_key(void *item
, int *key
)
2662 u32 hash
= hash_ptr(item
, BLOOM_FILTER_SHIFT
* 2);
2664 BUILD_BUG_ON(BLOOM_FILTER_SHIFT
* 2 > BITS_PER_TYPE(u32
));
2666 key
[0] = hash
& (BIT(BLOOM_FILTER_SHIFT
) - 1);
2667 key
[1] = hash
>> BLOOM_FILTER_SHIFT
;
2670 static bool test_bloom_filter(struct lruvec
*lruvec
, unsigned long seq
, void *item
)
2673 unsigned long *filter
;
2674 int gen
= filter_gen_from_seq(seq
);
2676 filter
= READ_ONCE(lruvec
->mm_state
.filters
[gen
]);
2680 get_item_key(item
, key
);
2682 return test_bit(key
[0], filter
) && test_bit(key
[1], filter
);
2685 static void update_bloom_filter(struct lruvec
*lruvec
, unsigned long seq
, void *item
)
2688 unsigned long *filter
;
2689 int gen
= filter_gen_from_seq(seq
);
2691 filter
= READ_ONCE(lruvec
->mm_state
.filters
[gen
]);
2695 get_item_key(item
, key
);
2697 if (!test_bit(key
[0], filter
))
2698 set_bit(key
[0], filter
);
2699 if (!test_bit(key
[1], filter
))
2700 set_bit(key
[1], filter
);
2703 static void reset_bloom_filter(struct lruvec
*lruvec
, unsigned long seq
)
2705 unsigned long *filter
;
2706 int gen
= filter_gen_from_seq(seq
);
2708 filter
= lruvec
->mm_state
.filters
[gen
];
2710 bitmap_clear(filter
, 0, BIT(BLOOM_FILTER_SHIFT
));
2714 filter
= bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT
),
2715 __GFP_HIGH
| __GFP_NOMEMALLOC
| __GFP_NOWARN
);
2716 WRITE_ONCE(lruvec
->mm_state
.filters
[gen
], filter
);
2719 /******************************************************************************
2721 ******************************************************************************/
2723 static struct lru_gen_mm_list
*get_mm_list(struct mem_cgroup
*memcg
)
2725 static struct lru_gen_mm_list mm_list
= {
2726 .fifo
= LIST_HEAD_INIT(mm_list
.fifo
),
2727 .lock
= __SPIN_LOCK_UNLOCKED(mm_list
.lock
),
2732 return &memcg
->mm_list
;
2734 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2739 void lru_gen_add_mm(struct mm_struct
*mm
)
2742 struct mem_cgroup
*memcg
= get_mem_cgroup_from_mm(mm
);
2743 struct lru_gen_mm_list
*mm_list
= get_mm_list(memcg
);
2745 VM_WARN_ON_ONCE(!list_empty(&mm
->lru_gen
.list
));
2747 VM_WARN_ON_ONCE(mm
->lru_gen
.memcg
);
2748 mm
->lru_gen
.memcg
= memcg
;
2750 spin_lock(&mm_list
->lock
);
2752 for_each_node_state(nid
, N_MEMORY
) {
2753 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
2755 /* the first addition since the last iteration */
2756 if (lruvec
->mm_state
.tail
== &mm_list
->fifo
)
2757 lruvec
->mm_state
.tail
= &mm
->lru_gen
.list
;
2760 list_add_tail(&mm
->lru_gen
.list
, &mm_list
->fifo
);
2762 spin_unlock(&mm_list
->lock
);
2765 void lru_gen_del_mm(struct mm_struct
*mm
)
2768 struct lru_gen_mm_list
*mm_list
;
2769 struct mem_cgroup
*memcg
= NULL
;
2771 if (list_empty(&mm
->lru_gen
.list
))
2775 memcg
= mm
->lru_gen
.memcg
;
2777 mm_list
= get_mm_list(memcg
);
2779 spin_lock(&mm_list
->lock
);
2781 for_each_node(nid
) {
2782 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
2784 /* where the current iteration continues after */
2785 if (lruvec
->mm_state
.head
== &mm
->lru_gen
.list
)
2786 lruvec
->mm_state
.head
= lruvec
->mm_state
.head
->prev
;
2788 /* where the last iteration ended before */
2789 if (lruvec
->mm_state
.tail
== &mm
->lru_gen
.list
)
2790 lruvec
->mm_state
.tail
= lruvec
->mm_state
.tail
->next
;
2793 list_del_init(&mm
->lru_gen
.list
);
2795 spin_unlock(&mm_list
->lock
);
2798 mem_cgroup_put(mm
->lru_gen
.memcg
);
2799 mm
->lru_gen
.memcg
= NULL
;
2804 void lru_gen_migrate_mm(struct mm_struct
*mm
)
2806 struct mem_cgroup
*memcg
;
2807 struct task_struct
*task
= rcu_dereference_protected(mm
->owner
, true);
2809 VM_WARN_ON_ONCE(task
->mm
!= mm
);
2810 lockdep_assert_held(&task
->alloc_lock
);
2812 /* for mm_update_next_owner() */
2813 if (mem_cgroup_disabled())
2816 /* migration can happen before addition */
2817 if (!mm
->lru_gen
.memcg
)
2821 memcg
= mem_cgroup_from_task(task
);
2823 if (memcg
== mm
->lru_gen
.memcg
)
2826 VM_WARN_ON_ONCE(list_empty(&mm
->lru_gen
.list
));
2833 static void reset_mm_stats(struct lruvec
*lruvec
, struct lru_gen_mm_walk
*walk
, bool last
)
2838 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec
))->lock
);
2841 hist
= lru_hist_from_seq(walk
->max_seq
);
2843 for (i
= 0; i
< NR_MM_STATS
; i
++) {
2844 WRITE_ONCE(lruvec
->mm_state
.stats
[hist
][i
],
2845 lruvec
->mm_state
.stats
[hist
][i
] + walk
->mm_stats
[i
]);
2846 walk
->mm_stats
[i
] = 0;
2850 if (NR_HIST_GENS
> 1 && last
) {
2851 hist
= lru_hist_from_seq(lruvec
->mm_state
.seq
+ 1);
2853 for (i
= 0; i
< NR_MM_STATS
; i
++)
2854 WRITE_ONCE(lruvec
->mm_state
.stats
[hist
][i
], 0);
2858 static bool should_skip_mm(struct mm_struct
*mm
, struct lru_gen_mm_walk
*walk
)
2861 unsigned long size
= 0;
2862 struct pglist_data
*pgdat
= lruvec_pgdat(walk
->lruvec
);
2863 int key
= pgdat
->node_id
% BITS_PER_TYPE(mm
->lru_gen
.bitmap
);
2865 if (!walk
->force_scan
&& !test_bit(key
, &mm
->lru_gen
.bitmap
))
2868 clear_bit(key
, &mm
->lru_gen
.bitmap
);
2870 for (type
= !walk
->can_swap
; type
< ANON_AND_FILE
; type
++) {
2871 size
+= type
? get_mm_counter(mm
, MM_FILEPAGES
) :
2872 get_mm_counter(mm
, MM_ANONPAGES
) +
2873 get_mm_counter(mm
, MM_SHMEMPAGES
);
2876 if (size
< MIN_LRU_BATCH
)
2879 return !mmget_not_zero(mm
);
2882 static bool iterate_mm_list(struct lruvec
*lruvec
, struct lru_gen_mm_walk
*walk
,
2883 struct mm_struct
**iter
)
2887 struct mm_struct
*mm
= NULL
;
2888 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2889 struct lru_gen_mm_list
*mm_list
= get_mm_list(memcg
);
2890 struct lru_gen_mm_state
*mm_state
= &lruvec
->mm_state
;
2893 * mm_state->seq is incremented after each iteration of mm_list. There
2894 * are three interesting cases for this page table walker:
2895 * 1. It tries to start a new iteration with a stale max_seq: there is
2896 * nothing left to do.
2897 * 2. It started the next iteration: it needs to reset the Bloom filter
2898 * so that a fresh set of PTE tables can be recorded.
2899 * 3. It ended the current iteration: it needs to reset the mm stats
2900 * counters and tell its caller to increment max_seq.
2902 spin_lock(&mm_list
->lock
);
2904 VM_WARN_ON_ONCE(mm_state
->seq
+ 1 < walk
->max_seq
);
2906 if (walk
->max_seq
<= mm_state
->seq
)
2909 if (!mm_state
->head
)
2910 mm_state
->head
= &mm_list
->fifo
;
2912 if (mm_state
->head
== &mm_list
->fifo
)
2916 mm_state
->head
= mm_state
->head
->next
;
2917 if (mm_state
->head
== &mm_list
->fifo
) {
2918 WRITE_ONCE(mm_state
->seq
, mm_state
->seq
+ 1);
2923 /* force scan for those added after the last iteration */
2924 if (!mm_state
->tail
|| mm_state
->tail
== mm_state
->head
) {
2925 mm_state
->tail
= mm_state
->head
->next
;
2926 walk
->force_scan
= true;
2929 mm
= list_entry(mm_state
->head
, struct mm_struct
, lru_gen
.list
);
2930 if (should_skip_mm(mm
, walk
))
2935 reset_mm_stats(lruvec
, walk
, last
);
2937 spin_unlock(&mm_list
->lock
);
2940 reset_bloom_filter(lruvec
, walk
->max_seq
+ 1);
2950 static bool iterate_mm_list_nowalk(struct lruvec
*lruvec
, unsigned long max_seq
)
2952 bool success
= false;
2953 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
2954 struct lru_gen_mm_list
*mm_list
= get_mm_list(memcg
);
2955 struct lru_gen_mm_state
*mm_state
= &lruvec
->mm_state
;
2957 spin_lock(&mm_list
->lock
);
2959 VM_WARN_ON_ONCE(mm_state
->seq
+ 1 < max_seq
);
2961 if (max_seq
> mm_state
->seq
) {
2962 mm_state
->head
= NULL
;
2963 mm_state
->tail
= NULL
;
2964 WRITE_ONCE(mm_state
->seq
, mm_state
->seq
+ 1);
2965 reset_mm_stats(lruvec
, NULL
, true);
2969 spin_unlock(&mm_list
->lock
);
2974 /******************************************************************************
2976 ******************************************************************************/
2979 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
2981 * The P term is refaulted/(evicted+protected) from a tier in the generation
2982 * currently being evicted; the I term is the exponential moving average of the
2983 * P term over the generations previously evicted, using the smoothing factor
2984 * 1/2; the D term isn't supported.
2986 * The setpoint (SP) is always the first tier of one type; the process variable
2987 * (PV) is either any tier of the other type or any other tier of the same
2990 * The error is the difference between the SP and the PV; the correction is to
2991 * turn off protection when SP>PV or turn on protection when SP<PV.
2993 * For future optimizations:
2994 * 1. The D term may discount the other two terms over time so that long-lived
2995 * generations can resist stale information.
2998 unsigned long refaulted
;
2999 unsigned long total
;
3003 static void read_ctrl_pos(struct lruvec
*lruvec
, int type
, int tier
, int gain
,
3004 struct ctrl_pos
*pos
)
3006 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3007 int hist
= lru_hist_from_seq(lrugen
->min_seq
[type
]);
3009 pos
->refaulted
= lrugen
->avg_refaulted
[type
][tier
] +
3010 atomic_long_read(&lrugen
->refaulted
[hist
][type
][tier
]);
3011 pos
->total
= lrugen
->avg_total
[type
][tier
] +
3012 atomic_long_read(&lrugen
->evicted
[hist
][type
][tier
]);
3014 pos
->total
+= lrugen
->protected[hist
][type
][tier
- 1];
3018 static void reset_ctrl_pos(struct lruvec
*lruvec
, int type
, bool carryover
)
3021 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3022 bool clear
= carryover
? NR_HIST_GENS
== 1 : NR_HIST_GENS
> 1;
3023 unsigned long seq
= carryover
? lrugen
->min_seq
[type
] : lrugen
->max_seq
+ 1;
3025 lockdep_assert_held(&lruvec
->lru_lock
);
3027 if (!carryover
&& !clear
)
3030 hist
= lru_hist_from_seq(seq
);
3032 for (tier
= 0; tier
< MAX_NR_TIERS
; tier
++) {
3036 sum
= lrugen
->avg_refaulted
[type
][tier
] +
3037 atomic_long_read(&lrugen
->refaulted
[hist
][type
][tier
]);
3038 WRITE_ONCE(lrugen
->avg_refaulted
[type
][tier
], sum
/ 2);
3040 sum
= lrugen
->avg_total
[type
][tier
] +
3041 atomic_long_read(&lrugen
->evicted
[hist
][type
][tier
]);
3043 sum
+= lrugen
->protected[hist
][type
][tier
- 1];
3044 WRITE_ONCE(lrugen
->avg_total
[type
][tier
], sum
/ 2);
3048 atomic_long_set(&lrugen
->refaulted
[hist
][type
][tier
], 0);
3049 atomic_long_set(&lrugen
->evicted
[hist
][type
][tier
], 0);
3051 WRITE_ONCE(lrugen
->protected[hist
][type
][tier
- 1], 0);
3056 static bool positive_ctrl_err(struct ctrl_pos
*sp
, struct ctrl_pos
*pv
)
3059 * Return true if the PV has a limited number of refaults or a lower
3060 * refaulted/total than the SP.
3062 return pv
->refaulted
< MIN_LRU_BATCH
||
3063 pv
->refaulted
* (sp
->total
+ MIN_LRU_BATCH
) * sp
->gain
<=
3064 (sp
->refaulted
+ 1) * pv
->total
* pv
->gain
;
3067 /******************************************************************************
3069 ******************************************************************************/
3071 /* promote pages accessed through page tables */
3072 static int folio_update_gen(struct folio
*folio
, int gen
)
3074 unsigned long new_flags
, old_flags
= READ_ONCE(folio
->flags
);
3076 VM_WARN_ON_ONCE(gen
>= MAX_NR_GENS
);
3077 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3080 /* lru_gen_del_folio() has isolated this page? */
3081 if (!(old_flags
& LRU_GEN_MASK
)) {
3082 /* for shrink_folio_list() */
3083 new_flags
= old_flags
| BIT(PG_referenced
);
3087 new_flags
= old_flags
& ~(LRU_GEN_MASK
| LRU_REFS_MASK
| LRU_REFS_FLAGS
);
3088 new_flags
|= (gen
+ 1UL) << LRU_GEN_PGOFF
;
3089 } while (!try_cmpxchg(&folio
->flags
, &old_flags
, new_flags
));
3091 return ((old_flags
& LRU_GEN_MASK
) >> LRU_GEN_PGOFF
) - 1;
3094 /* protect pages accessed multiple times through file descriptors */
3095 static int folio_inc_gen(struct lruvec
*lruvec
, struct folio
*folio
, bool reclaiming
)
3097 int type
= folio_is_file_lru(folio
);
3098 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3099 int new_gen
, old_gen
= lru_gen_from_seq(lrugen
->min_seq
[type
]);
3100 unsigned long new_flags
, old_flags
= READ_ONCE(folio
->flags
);
3102 VM_WARN_ON_ONCE_FOLIO(!(old_flags
& LRU_GEN_MASK
), folio
);
3105 new_gen
= ((old_flags
& LRU_GEN_MASK
) >> LRU_GEN_PGOFF
) - 1;
3106 /* folio_update_gen() has promoted this page? */
3107 if (new_gen
>= 0 && new_gen
!= old_gen
)
3110 new_gen
= (old_gen
+ 1) % MAX_NR_GENS
;
3112 new_flags
= old_flags
& ~(LRU_GEN_MASK
| LRU_REFS_MASK
| LRU_REFS_FLAGS
);
3113 new_flags
|= (new_gen
+ 1UL) << LRU_GEN_PGOFF
;
3114 /* for folio_end_writeback() */
3116 new_flags
|= BIT(PG_reclaim
);
3117 } while (!try_cmpxchg(&folio
->flags
, &old_flags
, new_flags
));
3119 lru_gen_update_size(lruvec
, folio
, old_gen
, new_gen
);
3124 static void update_batch_size(struct lru_gen_mm_walk
*walk
, struct folio
*folio
,
3125 int old_gen
, int new_gen
)
3127 int type
= folio_is_file_lru(folio
);
3128 int zone
= folio_zonenum(folio
);
3129 int delta
= folio_nr_pages(folio
);
3131 VM_WARN_ON_ONCE(old_gen
>= MAX_NR_GENS
);
3132 VM_WARN_ON_ONCE(new_gen
>= MAX_NR_GENS
);
3136 walk
->nr_pages
[old_gen
][type
][zone
] -= delta
;
3137 walk
->nr_pages
[new_gen
][type
][zone
] += delta
;
3140 static void reset_batch_size(struct lruvec
*lruvec
, struct lru_gen_mm_walk
*walk
)
3142 int gen
, type
, zone
;
3143 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3147 for_each_gen_type_zone(gen
, type
, zone
) {
3148 enum lru_list lru
= type
* LRU_INACTIVE_FILE
;
3149 int delta
= walk
->nr_pages
[gen
][type
][zone
];
3154 walk
->nr_pages
[gen
][type
][zone
] = 0;
3155 WRITE_ONCE(lrugen
->nr_pages
[gen
][type
][zone
],
3156 lrugen
->nr_pages
[gen
][type
][zone
] + delta
);
3158 if (lru_gen_is_active(lruvec
, gen
))
3160 __update_lru_size(lruvec
, lru
, zone
, delta
);
3164 static int should_skip_vma(unsigned long start
, unsigned long end
, struct mm_walk
*args
)
3166 struct address_space
*mapping
;
3167 struct vm_area_struct
*vma
= args
->vma
;
3168 struct lru_gen_mm_walk
*walk
= args
->private;
3170 if (!vma_is_accessible(vma
))
3173 if (is_vm_hugetlb_page(vma
))
3176 if (!vma_has_recency(vma
))
3179 if (vma
->vm_flags
& (VM_LOCKED
| VM_SPECIAL
))
3182 if (vma
== get_gate_vma(vma
->vm_mm
))
3185 if (vma_is_anonymous(vma
))
3186 return !walk
->can_swap
;
3188 if (WARN_ON_ONCE(!vma
->vm_file
|| !vma
->vm_file
->f_mapping
))
3191 mapping
= vma
->vm_file
->f_mapping
;
3192 if (mapping_unevictable(mapping
))
3195 if (shmem_mapping(mapping
))
3196 return !walk
->can_swap
;
3198 /* to exclude special mappings like dax, etc. */
3199 return !mapping
->a_ops
->read_folio
;
3203 * Some userspace memory allocators map many single-page VMAs. Instead of
3204 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3205 * table to reduce zigzags and improve cache performance.
3207 static bool get_next_vma(unsigned long mask
, unsigned long size
, struct mm_walk
*args
,
3208 unsigned long *vm_start
, unsigned long *vm_end
)
3210 unsigned long start
= round_up(*vm_end
, size
);
3211 unsigned long end
= (start
| ~mask
) + 1;
3212 VMA_ITERATOR(vmi
, args
->mm
, start
);
3214 VM_WARN_ON_ONCE(mask
& size
);
3215 VM_WARN_ON_ONCE((start
& mask
) != (*vm_start
& mask
));
3217 for_each_vma(vmi
, args
->vma
) {
3218 if (end
&& end
<= args
->vma
->vm_start
)
3221 if (should_skip_vma(args
->vma
->vm_start
, args
->vma
->vm_end
, args
))
3224 *vm_start
= max(start
, args
->vma
->vm_start
);
3225 *vm_end
= min(end
- 1, args
->vma
->vm_end
- 1) + 1;
3233 static unsigned long get_pte_pfn(pte_t pte
, struct vm_area_struct
*vma
, unsigned long addr
)
3235 unsigned long pfn
= pte_pfn(pte
);
3237 VM_WARN_ON_ONCE(addr
< vma
->vm_start
|| addr
>= vma
->vm_end
);
3239 if (!pte_present(pte
) || is_zero_pfn(pfn
))
3242 if (WARN_ON_ONCE(pte_devmap(pte
) || pte_special(pte
)))
3245 if (WARN_ON_ONCE(!pfn_valid(pfn
)))
3251 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3252 static unsigned long get_pmd_pfn(pmd_t pmd
, struct vm_area_struct
*vma
, unsigned long addr
)
3254 unsigned long pfn
= pmd_pfn(pmd
);
3256 VM_WARN_ON_ONCE(addr
< vma
->vm_start
|| addr
>= vma
->vm_end
);
3258 if (!pmd_present(pmd
) || is_huge_zero_pmd(pmd
))
3261 if (WARN_ON_ONCE(pmd_devmap(pmd
)))
3264 if (WARN_ON_ONCE(!pfn_valid(pfn
)))
3271 static struct folio
*get_pfn_folio(unsigned long pfn
, struct mem_cgroup
*memcg
,
3272 struct pglist_data
*pgdat
, bool can_swap
)
3274 struct folio
*folio
;
3276 /* try to avoid unnecessary memory loads */
3277 if (pfn
< pgdat
->node_start_pfn
|| pfn
>= pgdat_end_pfn(pgdat
))
3280 folio
= pfn_folio(pfn
);
3281 if (folio_nid(folio
) != pgdat
->node_id
)
3284 if (folio_memcg_rcu(folio
) != memcg
)
3287 /* file VMAs can contain anon pages from COW */
3288 if (!folio_is_file_lru(folio
) && !can_swap
)
3294 static bool suitable_to_scan(int total
, int young
)
3296 int n
= clamp_t(int, cache_line_size() / sizeof(pte_t
), 2, 8);
3298 /* suitable if the average number of young PTEs per cacheline is >=1 */
3299 return young
* n
>= total
;
3302 static bool walk_pte_range(pmd_t
*pmd
, unsigned long start
, unsigned long end
,
3303 struct mm_walk
*args
)
3311 struct lru_gen_mm_walk
*walk
= args
->private;
3312 struct mem_cgroup
*memcg
= lruvec_memcg(walk
->lruvec
);
3313 struct pglist_data
*pgdat
= lruvec_pgdat(walk
->lruvec
);
3314 int old_gen
, new_gen
= lru_gen_from_seq(walk
->max_seq
);
3316 pte
= pte_offset_map_nolock(args
->mm
, pmd
, start
& PMD_MASK
, &ptl
);
3319 if (!spin_trylock(ptl
)) {
3324 arch_enter_lazy_mmu_mode();
3326 for (i
= pte_index(start
), addr
= start
; addr
!= end
; i
++, addr
+= PAGE_SIZE
) {
3328 struct folio
*folio
;
3329 pte_t ptent
= ptep_get(pte
+ i
);
3332 walk
->mm_stats
[MM_LEAF_TOTAL
]++;
3334 pfn
= get_pte_pfn(ptent
, args
->vma
, addr
);
3338 if (!pte_young(ptent
)) {
3339 walk
->mm_stats
[MM_LEAF_OLD
]++;
3343 folio
= get_pfn_folio(pfn
, memcg
, pgdat
, walk
->can_swap
);
3347 if (!ptep_test_and_clear_young(args
->vma
, addr
, pte
+ i
))
3348 VM_WARN_ON_ONCE(true);
3351 walk
->mm_stats
[MM_LEAF_YOUNG
]++;
3353 if (pte_dirty(ptent
) && !folio_test_dirty(folio
) &&
3354 !(folio_test_anon(folio
) && folio_test_swapbacked(folio
) &&
3355 !folio_test_swapcache(folio
)))
3356 folio_mark_dirty(folio
);
3358 old_gen
= folio_update_gen(folio
, new_gen
);
3359 if (old_gen
>= 0 && old_gen
!= new_gen
)
3360 update_batch_size(walk
, folio
, old_gen
, new_gen
);
3363 if (i
< PTRS_PER_PTE
&& get_next_vma(PMD_MASK
, PAGE_SIZE
, args
, &start
, &end
))
3366 arch_leave_lazy_mmu_mode();
3367 pte_unmap_unlock(pte
, ptl
);
3369 return suitable_to_scan(total
, young
);
3372 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3373 static void walk_pmd_range_locked(pud_t
*pud
, unsigned long addr
, struct vm_area_struct
*vma
,
3374 struct mm_walk
*args
, unsigned long *bitmap
, unsigned long *first
)
3379 struct lru_gen_mm_walk
*walk
= args
->private;
3380 struct mem_cgroup
*memcg
= lruvec_memcg(walk
->lruvec
);
3381 struct pglist_data
*pgdat
= lruvec_pgdat(walk
->lruvec
);
3382 int old_gen
, new_gen
= lru_gen_from_seq(walk
->max_seq
);
3384 VM_WARN_ON_ONCE(pud_leaf(*pud
));
3386 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3389 bitmap_zero(bitmap
, MIN_LRU_BATCH
);
3393 i
= addr
== -1 ? 0 : pmd_index(addr
) - pmd_index(*first
);
3394 if (i
&& i
<= MIN_LRU_BATCH
) {
3395 __set_bit(i
- 1, bitmap
);
3399 pmd
= pmd_offset(pud
, *first
);
3401 ptl
= pmd_lockptr(args
->mm
, pmd
);
3402 if (!spin_trylock(ptl
))
3405 arch_enter_lazy_mmu_mode();
3409 struct folio
*folio
;
3411 /* don't round down the first address */
3412 addr
= i
? (*first
& PMD_MASK
) + i
* PMD_SIZE
: *first
;
3414 pfn
= get_pmd_pfn(pmd
[i
], vma
, addr
);
3418 if (!pmd_trans_huge(pmd
[i
])) {
3419 if (should_clear_pmd_young())
3420 pmdp_test_and_clear_young(vma
, addr
, pmd
+ i
);
3424 folio
= get_pfn_folio(pfn
, memcg
, pgdat
, walk
->can_swap
);
3428 if (!pmdp_test_and_clear_young(vma
, addr
, pmd
+ i
))
3431 walk
->mm_stats
[MM_LEAF_YOUNG
]++;
3433 if (pmd_dirty(pmd
[i
]) && !folio_test_dirty(folio
) &&
3434 !(folio_test_anon(folio
) && folio_test_swapbacked(folio
) &&
3435 !folio_test_swapcache(folio
)))
3436 folio_mark_dirty(folio
);
3438 old_gen
= folio_update_gen(folio
, new_gen
);
3439 if (old_gen
>= 0 && old_gen
!= new_gen
)
3440 update_batch_size(walk
, folio
, old_gen
, new_gen
);
3442 i
= i
> MIN_LRU_BATCH
? 0 : find_next_bit(bitmap
, MIN_LRU_BATCH
, i
) + 1;
3443 } while (i
<= MIN_LRU_BATCH
);
3445 arch_leave_lazy_mmu_mode();
3451 static void walk_pmd_range_locked(pud_t
*pud
, unsigned long addr
, struct vm_area_struct
*vma
,
3452 struct mm_walk
*args
, unsigned long *bitmap
, unsigned long *first
)
3457 static void walk_pmd_range(pud_t
*pud
, unsigned long start
, unsigned long end
,
3458 struct mm_walk
*args
)
3464 struct vm_area_struct
*vma
;
3465 DECLARE_BITMAP(bitmap
, MIN_LRU_BATCH
);
3466 unsigned long first
= -1;
3467 struct lru_gen_mm_walk
*walk
= args
->private;
3469 VM_WARN_ON_ONCE(pud_leaf(*pud
));
3472 * Finish an entire PMD in two passes: the first only reaches to PTE
3473 * tables to avoid taking the PMD lock; the second, if necessary, takes
3474 * the PMD lock to clear the accessed bit in PMD entries.
3476 pmd
= pmd_offset(pud
, start
& PUD_MASK
);
3478 /* walk_pte_range() may call get_next_vma() */
3480 for (i
= pmd_index(start
), addr
= start
; addr
!= end
; i
++, addr
= next
) {
3481 pmd_t val
= pmdp_get_lockless(pmd
+ i
);
3483 next
= pmd_addr_end(addr
, end
);
3485 if (!pmd_present(val
) || is_huge_zero_pmd(val
)) {
3486 walk
->mm_stats
[MM_LEAF_TOTAL
]++;
3490 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3491 if (pmd_trans_huge(val
)) {
3492 unsigned long pfn
= pmd_pfn(val
);
3493 struct pglist_data
*pgdat
= lruvec_pgdat(walk
->lruvec
);
3495 walk
->mm_stats
[MM_LEAF_TOTAL
]++;
3497 if (!pmd_young(val
)) {
3498 walk
->mm_stats
[MM_LEAF_OLD
]++;
3502 /* try to avoid unnecessary memory loads */
3503 if (pfn
< pgdat
->node_start_pfn
|| pfn
>= pgdat_end_pfn(pgdat
))
3506 walk_pmd_range_locked(pud
, addr
, vma
, args
, bitmap
, &first
);
3510 walk
->mm_stats
[MM_NONLEAF_TOTAL
]++;
3512 if (should_clear_pmd_young()) {
3513 if (!pmd_young(val
))
3516 walk_pmd_range_locked(pud
, addr
, vma
, args
, bitmap
, &first
);
3519 if (!walk
->force_scan
&& !test_bloom_filter(walk
->lruvec
, walk
->max_seq
, pmd
+ i
))
3522 walk
->mm_stats
[MM_NONLEAF_FOUND
]++;
3524 if (!walk_pte_range(&val
, addr
, next
, args
))
3527 walk
->mm_stats
[MM_NONLEAF_ADDED
]++;
3529 /* carry over to the next generation */
3530 update_bloom_filter(walk
->lruvec
, walk
->max_seq
+ 1, pmd
+ i
);
3533 walk_pmd_range_locked(pud
, -1, vma
, args
, bitmap
, &first
);
3535 if (i
< PTRS_PER_PMD
&& get_next_vma(PUD_MASK
, PMD_SIZE
, args
, &start
, &end
))
3539 static int walk_pud_range(p4d_t
*p4d
, unsigned long start
, unsigned long end
,
3540 struct mm_walk
*args
)
3546 struct lru_gen_mm_walk
*walk
= args
->private;
3548 VM_WARN_ON_ONCE(p4d_leaf(*p4d
));
3550 pud
= pud_offset(p4d
, start
& P4D_MASK
);
3552 for (i
= pud_index(start
), addr
= start
; addr
!= end
; i
++, addr
= next
) {
3553 pud_t val
= READ_ONCE(pud
[i
]);
3555 next
= pud_addr_end(addr
, end
);
3557 if (!pud_present(val
) || WARN_ON_ONCE(pud_leaf(val
)))
3560 walk_pmd_range(&val
, addr
, next
, args
);
3562 if (need_resched() || walk
->batched
>= MAX_LRU_BATCH
) {
3563 end
= (addr
| ~PUD_MASK
) + 1;
3568 if (i
< PTRS_PER_PUD
&& get_next_vma(P4D_MASK
, PUD_SIZE
, args
, &start
, &end
))
3571 end
= round_up(end
, P4D_SIZE
);
3573 if (!end
|| !args
->vma
)
3576 walk
->next_addr
= max(end
, args
->vma
->vm_start
);
3581 static void walk_mm(struct lruvec
*lruvec
, struct mm_struct
*mm
, struct lru_gen_mm_walk
*walk
)
3583 static const struct mm_walk_ops mm_walk_ops
= {
3584 .test_walk
= should_skip_vma
,
3585 .p4d_entry
= walk_pud_range
,
3586 .walk_lock
= PGWALK_RDLOCK
,
3590 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
3592 walk
->next_addr
= FIRST_USER_ADDRESS
;
3595 DEFINE_MAX_SEQ(lruvec
);
3599 /* another thread might have called inc_max_seq() */
3600 if (walk
->max_seq
!= max_seq
)
3603 /* folio_update_gen() requires stable folio_memcg() */
3604 if (!mem_cgroup_trylock_pages(memcg
))
3607 /* the caller might be holding the lock for write */
3608 if (mmap_read_trylock(mm
)) {
3609 err
= walk_page_range(mm
, walk
->next_addr
, ULONG_MAX
, &mm_walk_ops
, walk
);
3611 mmap_read_unlock(mm
);
3614 mem_cgroup_unlock_pages();
3616 if (walk
->batched
) {
3617 spin_lock_irq(&lruvec
->lru_lock
);
3618 reset_batch_size(lruvec
, walk
);
3619 spin_unlock_irq(&lruvec
->lru_lock
);
3623 } while (err
== -EAGAIN
);
3626 static struct lru_gen_mm_walk
*set_mm_walk(struct pglist_data
*pgdat
, bool force_alloc
)
3628 struct lru_gen_mm_walk
*walk
= current
->reclaim_state
->mm_walk
;
3630 if (pgdat
&& current_is_kswapd()) {
3631 VM_WARN_ON_ONCE(walk
);
3633 walk
= &pgdat
->mm_walk
;
3634 } else if (!walk
&& force_alloc
) {
3635 VM_WARN_ON_ONCE(current_is_kswapd());
3637 walk
= kzalloc(sizeof(*walk
), __GFP_HIGH
| __GFP_NOMEMALLOC
| __GFP_NOWARN
);
3640 current
->reclaim_state
->mm_walk
= walk
;
3645 static void clear_mm_walk(void)
3647 struct lru_gen_mm_walk
*walk
= current
->reclaim_state
->mm_walk
;
3649 VM_WARN_ON_ONCE(walk
&& memchr_inv(walk
->nr_pages
, 0, sizeof(walk
->nr_pages
)));
3650 VM_WARN_ON_ONCE(walk
&& memchr_inv(walk
->mm_stats
, 0, sizeof(walk
->mm_stats
)));
3652 current
->reclaim_state
->mm_walk
= NULL
;
3654 if (!current_is_kswapd())
3658 static bool inc_min_seq(struct lruvec
*lruvec
, int type
, bool can_swap
)
3661 int remaining
= MAX_LRU_BATCH
;
3662 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3663 int new_gen
, old_gen
= lru_gen_from_seq(lrugen
->min_seq
[type
]);
3665 if (type
== LRU_GEN_ANON
&& !can_swap
)
3668 /* prevent cold/hot inversion if force_scan is true */
3669 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
3670 struct list_head
*head
= &lrugen
->folios
[old_gen
][type
][zone
];
3672 while (!list_empty(head
)) {
3673 struct folio
*folio
= lru_to_folio(head
);
3675 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio
), folio
);
3676 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio
), folio
);
3677 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio
) != type
, folio
);
3678 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio
) != zone
, folio
);
3680 new_gen
= folio_inc_gen(lruvec
, folio
, false);
3681 list_move_tail(&folio
->lru
, &lrugen
->folios
[new_gen
][type
][zone
]);
3688 reset_ctrl_pos(lruvec
, type
, true);
3689 WRITE_ONCE(lrugen
->min_seq
[type
], lrugen
->min_seq
[type
] + 1);
3694 static bool try_to_inc_min_seq(struct lruvec
*lruvec
, bool can_swap
)
3696 int gen
, type
, zone
;
3697 bool success
= false;
3698 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3699 DEFINE_MIN_SEQ(lruvec
);
3701 VM_WARN_ON_ONCE(!seq_is_valid(lruvec
));
3703 /* find the oldest populated generation */
3704 for (type
= !can_swap
; type
< ANON_AND_FILE
; type
++) {
3705 while (min_seq
[type
] + MIN_NR_GENS
<= lrugen
->max_seq
) {
3706 gen
= lru_gen_from_seq(min_seq
[type
]);
3708 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
3709 if (!list_empty(&lrugen
->folios
[gen
][type
][zone
]))
3719 /* see the comment on lru_gen_folio */
3721 min_seq
[LRU_GEN_ANON
] = min(min_seq
[LRU_GEN_ANON
], min_seq
[LRU_GEN_FILE
]);
3722 min_seq
[LRU_GEN_FILE
] = max(min_seq
[LRU_GEN_ANON
], lrugen
->min_seq
[LRU_GEN_FILE
]);
3725 for (type
= !can_swap
; type
< ANON_AND_FILE
; type
++) {
3726 if (min_seq
[type
] == lrugen
->min_seq
[type
])
3729 reset_ctrl_pos(lruvec
, type
, true);
3730 WRITE_ONCE(lrugen
->min_seq
[type
], min_seq
[type
]);
3737 static void inc_max_seq(struct lruvec
*lruvec
, bool can_swap
, bool force_scan
)
3741 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3743 spin_lock_irq(&lruvec
->lru_lock
);
3745 VM_WARN_ON_ONCE(!seq_is_valid(lruvec
));
3747 for (type
= ANON_AND_FILE
- 1; type
>= 0; type
--) {
3748 if (get_nr_gens(lruvec
, type
) != MAX_NR_GENS
)
3751 VM_WARN_ON_ONCE(!force_scan
&& (type
== LRU_GEN_FILE
|| can_swap
));
3753 if (inc_min_seq(lruvec
, type
, can_swap
))
3756 spin_unlock_irq(&lruvec
->lru_lock
);
3762 * Update the active/inactive LRU sizes for compatibility. Both sides of
3763 * the current max_seq need to be covered, since max_seq+1 can overlap
3764 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3765 * overlap, cold/hot inversion happens.
3767 prev
= lru_gen_from_seq(lrugen
->max_seq
- 1);
3768 next
= lru_gen_from_seq(lrugen
->max_seq
+ 1);
3770 for (type
= 0; type
< ANON_AND_FILE
; type
++) {
3771 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
3772 enum lru_list lru
= type
* LRU_INACTIVE_FILE
;
3773 long delta
= lrugen
->nr_pages
[prev
][type
][zone
] -
3774 lrugen
->nr_pages
[next
][type
][zone
];
3779 __update_lru_size(lruvec
, lru
, zone
, delta
);
3780 __update_lru_size(lruvec
, lru
+ LRU_ACTIVE
, zone
, -delta
);
3784 for (type
= 0; type
< ANON_AND_FILE
; type
++)
3785 reset_ctrl_pos(lruvec
, type
, false);
3787 WRITE_ONCE(lrugen
->timestamps
[next
], jiffies
);
3788 /* make sure preceding modifications appear */
3789 smp_store_release(&lrugen
->max_seq
, lrugen
->max_seq
+ 1);
3791 spin_unlock_irq(&lruvec
->lru_lock
);
3794 static bool try_to_inc_max_seq(struct lruvec
*lruvec
, unsigned long max_seq
,
3795 struct scan_control
*sc
, bool can_swap
, bool force_scan
)
3798 struct lru_gen_mm_walk
*walk
;
3799 struct mm_struct
*mm
= NULL
;
3800 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3802 VM_WARN_ON_ONCE(max_seq
> READ_ONCE(lrugen
->max_seq
));
3804 /* see the comment in iterate_mm_list() */
3805 if (max_seq
<= READ_ONCE(lruvec
->mm_state
.seq
)) {
3811 * If the hardware doesn't automatically set the accessed bit, fallback
3812 * to lru_gen_look_around(), which only clears the accessed bit in a
3813 * handful of PTEs. Spreading the work out over a period of time usually
3814 * is less efficient, but it avoids bursty page faults.
3816 if (!should_walk_mmu()) {
3817 success
= iterate_mm_list_nowalk(lruvec
, max_seq
);
3821 walk
= set_mm_walk(NULL
, true);
3823 success
= iterate_mm_list_nowalk(lruvec
, max_seq
);
3827 walk
->lruvec
= lruvec
;
3828 walk
->max_seq
= max_seq
;
3829 walk
->can_swap
= can_swap
;
3830 walk
->force_scan
= force_scan
;
3833 success
= iterate_mm_list(lruvec
, walk
, &mm
);
3835 walk_mm(lruvec
, mm
, walk
);
3839 inc_max_seq(lruvec
, can_swap
, force_scan
);
3844 /******************************************************************************
3845 * working set protection
3846 ******************************************************************************/
3848 static bool lruvec_is_sizable(struct lruvec
*lruvec
, struct scan_control
*sc
)
3850 int gen
, type
, zone
;
3851 unsigned long total
= 0;
3852 bool can_swap
= get_swappiness(lruvec
, sc
);
3853 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
3854 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
3855 DEFINE_MAX_SEQ(lruvec
);
3856 DEFINE_MIN_SEQ(lruvec
);
3858 for (type
= !can_swap
; type
< ANON_AND_FILE
; type
++) {
3861 for (seq
= min_seq
[type
]; seq
<= max_seq
; seq
++) {
3862 gen
= lru_gen_from_seq(seq
);
3864 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++)
3865 total
+= max(READ_ONCE(lrugen
->nr_pages
[gen
][type
][zone
]), 0L);
3869 /* whether the size is big enough to be helpful */
3870 return mem_cgroup_online(memcg
) ? (total
>> sc
->priority
) : total
;
3873 static bool lruvec_is_reclaimable(struct lruvec
*lruvec
, struct scan_control
*sc
,
3874 unsigned long min_ttl
)
3877 unsigned long birth
;
3878 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
3879 DEFINE_MIN_SEQ(lruvec
);
3881 /* see the comment on lru_gen_folio */
3882 gen
= lru_gen_from_seq(min_seq
[LRU_GEN_FILE
]);
3883 birth
= READ_ONCE(lruvec
->lrugen
.timestamps
[gen
]);
3885 if (time_is_after_jiffies(birth
+ min_ttl
))
3888 if (!lruvec_is_sizable(lruvec
, sc
))
3891 mem_cgroup_calculate_protection(NULL
, memcg
);
3893 return !mem_cgroup_below_min(NULL
, memcg
);
3896 /* to protect the working set of the last N jiffies */
3897 static unsigned long lru_gen_min_ttl __read_mostly
;
3899 static void lru_gen_age_node(struct pglist_data
*pgdat
, struct scan_control
*sc
)
3901 struct mem_cgroup
*memcg
;
3902 unsigned long min_ttl
= READ_ONCE(lru_gen_min_ttl
);
3904 VM_WARN_ON_ONCE(!current_is_kswapd());
3906 /* check the order to exclude compaction-induced reclaim */
3907 if (!min_ttl
|| sc
->order
|| sc
->priority
== DEF_PRIORITY
)
3910 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
3912 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3914 if (lruvec_is_reclaimable(lruvec
, sc
, min_ttl
)) {
3915 mem_cgroup_iter_break(NULL
, memcg
);
3920 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)));
3923 * The main goal is to OOM kill if every generation from all memcgs is
3924 * younger than min_ttl. However, another possibility is all memcgs are
3925 * either too small or below min.
3927 if (mutex_trylock(&oom_lock
)) {
3928 struct oom_control oc
= {
3929 .gfp_mask
= sc
->gfp_mask
,
3934 mutex_unlock(&oom_lock
);
3938 /******************************************************************************
3939 * rmap/PT walk feedback
3940 ******************************************************************************/
3943 * This function exploits spatial locality when shrink_folio_list() walks the
3944 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3945 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3946 * the PTE table to the Bloom filter. This forms a feedback loop between the
3947 * eviction and the aging.
3949 void lru_gen_look_around(struct page_vma_mapped_walk
*pvmw
)
3952 unsigned long start
;
3954 struct lru_gen_mm_walk
*walk
;
3956 pte_t
*pte
= pvmw
->pte
;
3957 unsigned long addr
= pvmw
->address
;
3958 struct folio
*folio
= pfn_folio(pvmw
->pfn
);
3959 bool can_swap
= !folio_is_file_lru(folio
);
3960 struct mem_cgroup
*memcg
= folio_memcg(folio
);
3961 struct pglist_data
*pgdat
= folio_pgdat(folio
);
3962 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
3963 DEFINE_MAX_SEQ(lruvec
);
3964 int old_gen
, new_gen
= lru_gen_from_seq(max_seq
);
3966 lockdep_assert_held(pvmw
->ptl
);
3967 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio
), folio
);
3969 if (spin_is_contended(pvmw
->ptl
))
3972 /* avoid taking the LRU lock under the PTL when possible */
3973 walk
= current
->reclaim_state
? current
->reclaim_state
->mm_walk
: NULL
;
3975 start
= max(addr
& PMD_MASK
, pvmw
->vma
->vm_start
);
3976 end
= min(addr
| ~PMD_MASK
, pvmw
->vma
->vm_end
- 1) + 1;
3978 if (end
- start
> MIN_LRU_BATCH
* PAGE_SIZE
) {
3979 if (addr
- start
< MIN_LRU_BATCH
* PAGE_SIZE
/ 2)
3980 end
= start
+ MIN_LRU_BATCH
* PAGE_SIZE
;
3981 else if (end
- addr
< MIN_LRU_BATCH
* PAGE_SIZE
/ 2)
3982 start
= end
- MIN_LRU_BATCH
* PAGE_SIZE
;
3984 start
= addr
- MIN_LRU_BATCH
* PAGE_SIZE
/ 2;
3985 end
= addr
+ MIN_LRU_BATCH
* PAGE_SIZE
/ 2;
3989 /* folio_update_gen() requires stable folio_memcg() */
3990 if (!mem_cgroup_trylock_pages(memcg
))
3993 arch_enter_lazy_mmu_mode();
3995 pte
-= (addr
- start
) / PAGE_SIZE
;
3997 for (i
= 0, addr
= start
; addr
!= end
; i
++, addr
+= PAGE_SIZE
) {
3999 pte_t ptent
= ptep_get(pte
+ i
);
4001 pfn
= get_pte_pfn(ptent
, pvmw
->vma
, addr
);
4005 if (!pte_young(ptent
))
4008 folio
= get_pfn_folio(pfn
, memcg
, pgdat
, can_swap
);
4012 if (!ptep_test_and_clear_young(pvmw
->vma
, addr
, pte
+ i
))
4013 VM_WARN_ON_ONCE(true);
4017 if (pte_dirty(ptent
) && !folio_test_dirty(folio
) &&
4018 !(folio_test_anon(folio
) && folio_test_swapbacked(folio
) &&
4019 !folio_test_swapcache(folio
)))
4020 folio_mark_dirty(folio
);
4023 old_gen
= folio_update_gen(folio
, new_gen
);
4024 if (old_gen
>= 0 && old_gen
!= new_gen
)
4025 update_batch_size(walk
, folio
, old_gen
, new_gen
);
4030 old_gen
= folio_lru_gen(folio
);
4032 folio_set_referenced(folio
);
4033 else if (old_gen
!= new_gen
)
4034 folio_activate(folio
);
4037 arch_leave_lazy_mmu_mode();
4038 mem_cgroup_unlock_pages();
4040 /* feedback from rmap walkers to page table walkers */
4041 if (suitable_to_scan(i
, young
))
4042 update_bloom_filter(lruvec
, max_seq
, pvmw
->pmd
);
4045 /******************************************************************************
4047 ******************************************************************************/
4049 /* see the comment on MEMCG_NR_GENS */
4060 static int lru_gen_memcg_seg(struct lruvec
*lruvec
)
4062 return READ_ONCE(lruvec
->lrugen
.seg
);
4065 static void lru_gen_rotate_memcg(struct lruvec
*lruvec
, int op
)
4069 unsigned long flags
;
4070 int bin
= get_random_u32_below(MEMCG_NR_BINS
);
4071 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
4073 spin_lock_irqsave(&pgdat
->memcg_lru
.lock
, flags
);
4075 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec
->lrugen
.list
));
4078 new = old
= lruvec
->lrugen
.gen
;
4080 /* see the comment on MEMCG_NR_GENS */
4081 if (op
== MEMCG_LRU_HEAD
)
4082 seg
= MEMCG_LRU_HEAD
;
4083 else if (op
== MEMCG_LRU_TAIL
)
4084 seg
= MEMCG_LRU_TAIL
;
4085 else if (op
== MEMCG_LRU_OLD
)
4086 new = get_memcg_gen(pgdat
->memcg_lru
.seq
);
4087 else if (op
== MEMCG_LRU_YOUNG
)
4088 new = get_memcg_gen(pgdat
->memcg_lru
.seq
+ 1);
4090 VM_WARN_ON_ONCE(true);
4092 WRITE_ONCE(lruvec
->lrugen
.seg
, seg
);
4093 WRITE_ONCE(lruvec
->lrugen
.gen
, new);
4095 hlist_nulls_del_rcu(&lruvec
->lrugen
.list
);
4097 if (op
== MEMCG_LRU_HEAD
|| op
== MEMCG_LRU_OLD
)
4098 hlist_nulls_add_head_rcu(&lruvec
->lrugen
.list
, &pgdat
->memcg_lru
.fifo
[new][bin
]);
4100 hlist_nulls_add_tail_rcu(&lruvec
->lrugen
.list
, &pgdat
->memcg_lru
.fifo
[new][bin
]);
4102 pgdat
->memcg_lru
.nr_memcgs
[old
]--;
4103 pgdat
->memcg_lru
.nr_memcgs
[new]++;
4105 if (!pgdat
->memcg_lru
.nr_memcgs
[old
] && old
== get_memcg_gen(pgdat
->memcg_lru
.seq
))
4106 WRITE_ONCE(pgdat
->memcg_lru
.seq
, pgdat
->memcg_lru
.seq
+ 1);
4108 spin_unlock_irqrestore(&pgdat
->memcg_lru
.lock
, flags
);
4111 void lru_gen_online_memcg(struct mem_cgroup
*memcg
)
4115 int bin
= get_random_u32_below(MEMCG_NR_BINS
);
4117 for_each_node(nid
) {
4118 struct pglist_data
*pgdat
= NODE_DATA(nid
);
4119 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
4121 spin_lock_irq(&pgdat
->memcg_lru
.lock
);
4123 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec
->lrugen
.list
));
4125 gen
= get_memcg_gen(pgdat
->memcg_lru
.seq
);
4127 lruvec
->lrugen
.gen
= gen
;
4129 hlist_nulls_add_tail_rcu(&lruvec
->lrugen
.list
, &pgdat
->memcg_lru
.fifo
[gen
][bin
]);
4130 pgdat
->memcg_lru
.nr_memcgs
[gen
]++;
4132 spin_unlock_irq(&pgdat
->memcg_lru
.lock
);
4136 void lru_gen_offline_memcg(struct mem_cgroup
*memcg
)
4140 for_each_node(nid
) {
4141 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
4143 lru_gen_rotate_memcg(lruvec
, MEMCG_LRU_OLD
);
4147 void lru_gen_release_memcg(struct mem_cgroup
*memcg
)
4152 for_each_node(nid
) {
4153 struct pglist_data
*pgdat
= NODE_DATA(nid
);
4154 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
4156 spin_lock_irq(&pgdat
->memcg_lru
.lock
);
4158 if (hlist_nulls_unhashed(&lruvec
->lrugen
.list
))
4161 gen
= lruvec
->lrugen
.gen
;
4163 hlist_nulls_del_init_rcu(&lruvec
->lrugen
.list
);
4164 pgdat
->memcg_lru
.nr_memcgs
[gen
]--;
4166 if (!pgdat
->memcg_lru
.nr_memcgs
[gen
] && gen
== get_memcg_gen(pgdat
->memcg_lru
.seq
))
4167 WRITE_ONCE(pgdat
->memcg_lru
.seq
, pgdat
->memcg_lru
.seq
+ 1);
4169 spin_unlock_irq(&pgdat
->memcg_lru
.lock
);
4173 void lru_gen_soft_reclaim(struct mem_cgroup
*memcg
, int nid
)
4175 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
4177 /* see the comment on MEMCG_NR_GENS */
4178 if (lru_gen_memcg_seg(lruvec
) != MEMCG_LRU_HEAD
)
4179 lru_gen_rotate_memcg(lruvec
, MEMCG_LRU_HEAD
);
4182 #else /* !CONFIG_MEMCG */
4184 static int lru_gen_memcg_seg(struct lruvec
*lruvec
)
4191 /******************************************************************************
4193 ******************************************************************************/
4195 static bool sort_folio(struct lruvec
*lruvec
, struct folio
*folio
, struct scan_control
*sc
,
4199 int gen
= folio_lru_gen(folio
);
4200 int type
= folio_is_file_lru(folio
);
4201 int zone
= folio_zonenum(folio
);
4202 int delta
= folio_nr_pages(folio
);
4203 int refs
= folio_lru_refs(folio
);
4204 int tier
= lru_tier_from_refs(refs
);
4205 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
4207 VM_WARN_ON_ONCE_FOLIO(gen
>= MAX_NR_GENS
, folio
);
4210 if (!folio_evictable(folio
)) {
4211 success
= lru_gen_del_folio(lruvec
, folio
, true);
4212 VM_WARN_ON_ONCE_FOLIO(!success
, folio
);
4213 folio_set_unevictable(folio
);
4214 lruvec_add_folio(lruvec
, folio
);
4215 __count_vm_events(UNEVICTABLE_PGCULLED
, delta
);
4219 /* dirty lazyfree */
4220 if (type
== LRU_GEN_FILE
&& folio_test_anon(folio
) && folio_test_dirty(folio
)) {
4221 success
= lru_gen_del_folio(lruvec
, folio
, true);
4222 VM_WARN_ON_ONCE_FOLIO(!success
, folio
);
4223 folio_set_swapbacked(folio
);
4224 lruvec_add_folio_tail(lruvec
, folio
);
4229 if (gen
!= lru_gen_from_seq(lrugen
->min_seq
[type
])) {
4230 list_move(&folio
->lru
, &lrugen
->folios
[gen
][type
][zone
]);
4235 if (tier
> tier_idx
|| refs
== BIT(LRU_REFS_WIDTH
)) {
4236 int hist
= lru_hist_from_seq(lrugen
->min_seq
[type
]);
4238 gen
= folio_inc_gen(lruvec
, folio
, false);
4239 list_move_tail(&folio
->lru
, &lrugen
->folios
[gen
][type
][zone
]);
4241 WRITE_ONCE(lrugen
->protected[hist
][type
][tier
- 1],
4242 lrugen
->protected[hist
][type
][tier
- 1] + delta
);
4247 if (zone
> sc
->reclaim_idx
|| skip_cma(folio
, sc
)) {
4248 gen
= folio_inc_gen(lruvec
, folio
, false);
4249 list_move_tail(&folio
->lru
, &lrugen
->folios
[gen
][type
][zone
]);
4253 /* waiting for writeback */
4254 if (folio_test_locked(folio
) || folio_test_writeback(folio
) ||
4255 (type
== LRU_GEN_FILE
&& folio_test_dirty(folio
))) {
4256 gen
= folio_inc_gen(lruvec
, folio
, true);
4257 list_move(&folio
->lru
, &lrugen
->folios
[gen
][type
][zone
]);
4264 static bool isolate_folio(struct lruvec
*lruvec
, struct folio
*folio
, struct scan_control
*sc
)
4268 /* swapping inhibited */
4269 if (!(sc
->gfp_mask
& __GFP_IO
) &&
4270 (folio_test_dirty(folio
) ||
4271 (folio_test_anon(folio
) && !folio_test_swapcache(folio
))))
4274 /* raced with release_pages() */
4275 if (!folio_try_get(folio
))
4278 /* raced with another isolation */
4279 if (!folio_test_clear_lru(folio
)) {
4284 /* see the comment on MAX_NR_TIERS */
4285 if (!folio_test_referenced(folio
))
4286 set_mask_bits(&folio
->flags
, LRU_REFS_MASK
| LRU_REFS_FLAGS
, 0);
4288 /* for shrink_folio_list() */
4289 folio_clear_reclaim(folio
);
4290 folio_clear_referenced(folio
);
4292 success
= lru_gen_del_folio(lruvec
, folio
, true);
4293 VM_WARN_ON_ONCE_FOLIO(!success
, folio
);
4298 static int scan_folios(struct lruvec
*lruvec
, struct scan_control
*sc
,
4299 int type
, int tier
, struct list_head
*list
)
4303 enum vm_event_item item
;
4308 int remaining
= MAX_LRU_BATCH
;
4309 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
4310 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
4312 VM_WARN_ON_ONCE(!list_empty(list
));
4314 if (get_nr_gens(lruvec
, type
) == MIN_NR_GENS
)
4317 gen
= lru_gen_from_seq(lrugen
->min_seq
[type
]);
4319 for (i
= MAX_NR_ZONES
; i
> 0; i
--) {
4321 int skipped_zone
= 0;
4322 int zone
= (sc
->reclaim_idx
+ i
) % MAX_NR_ZONES
;
4323 struct list_head
*head
= &lrugen
->folios
[gen
][type
][zone
];
4325 while (!list_empty(head
)) {
4326 struct folio
*folio
= lru_to_folio(head
);
4327 int delta
= folio_nr_pages(folio
);
4329 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio
), folio
);
4330 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio
), folio
);
4331 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio
) != type
, folio
);
4332 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio
) != zone
, folio
);
4336 if (sort_folio(lruvec
, folio
, sc
, tier
))
4338 else if (isolate_folio(lruvec
, folio
, sc
)) {
4339 list_add(&folio
->lru
, list
);
4342 list_move(&folio
->lru
, &moved
);
4343 skipped_zone
+= delta
;
4346 if (!--remaining
|| max(isolated
, skipped_zone
) >= MIN_LRU_BATCH
)
4351 list_splice(&moved
, head
);
4352 __count_zid_vm_events(PGSCAN_SKIP
, zone
, skipped_zone
);
4353 skipped
+= skipped_zone
;
4356 if (!remaining
|| isolated
>= MIN_LRU_BATCH
)
4360 item
= PGSCAN_KSWAPD
+ reclaimer_offset();
4361 if (!cgroup_reclaim(sc
)) {
4362 __count_vm_events(item
, isolated
);
4363 __count_vm_events(PGREFILL
, sorted
);
4365 __count_memcg_events(memcg
, item
, isolated
);
4366 __count_memcg_events(memcg
, PGREFILL
, sorted
);
4367 __count_vm_events(PGSCAN_ANON
+ type
, isolated
);
4368 trace_mm_vmscan_lru_isolate(sc
->reclaim_idx
, sc
->order
, MAX_LRU_BATCH
,
4369 scanned
, skipped
, isolated
,
4370 type
? LRU_INACTIVE_FILE
: LRU_INACTIVE_ANON
);
4373 * There might not be eligible folios due to reclaim_idx. Check the
4374 * remaining to prevent livelock if it's not making progress.
4376 return isolated
|| !remaining
? scanned
: 0;
4379 static int get_tier_idx(struct lruvec
*lruvec
, int type
)
4382 struct ctrl_pos sp
, pv
;
4385 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4386 * This value is chosen because any other tier would have at least twice
4387 * as many refaults as the first tier.
4389 read_ctrl_pos(lruvec
, type
, 0, 1, &sp
);
4390 for (tier
= 1; tier
< MAX_NR_TIERS
; tier
++) {
4391 read_ctrl_pos(lruvec
, type
, tier
, 2, &pv
);
4392 if (!positive_ctrl_err(&sp
, &pv
))
4399 static int get_type_to_scan(struct lruvec
*lruvec
, int swappiness
, int *tier_idx
)
4402 struct ctrl_pos sp
, pv
;
4403 int gain
[ANON_AND_FILE
] = { swappiness
, 200 - swappiness
};
4406 * Compare the first tier of anon with that of file to determine which
4407 * type to scan. Also need to compare other tiers of the selected type
4408 * with the first tier of the other type to determine the last tier (of
4409 * the selected type) to evict.
4411 read_ctrl_pos(lruvec
, LRU_GEN_ANON
, 0, gain
[LRU_GEN_ANON
], &sp
);
4412 read_ctrl_pos(lruvec
, LRU_GEN_FILE
, 0, gain
[LRU_GEN_FILE
], &pv
);
4413 type
= positive_ctrl_err(&sp
, &pv
);
4415 read_ctrl_pos(lruvec
, !type
, 0, gain
[!type
], &sp
);
4416 for (tier
= 1; tier
< MAX_NR_TIERS
; tier
++) {
4417 read_ctrl_pos(lruvec
, type
, tier
, gain
[type
], &pv
);
4418 if (!positive_ctrl_err(&sp
, &pv
))
4422 *tier_idx
= tier
- 1;
4427 static int isolate_folios(struct lruvec
*lruvec
, struct scan_control
*sc
, int swappiness
,
4428 int *type_scanned
, struct list_head
*list
)
4434 DEFINE_MIN_SEQ(lruvec
);
4437 * Try to make the obvious choice first. When anon and file are both
4438 * available from the same generation, interpret swappiness 1 as file
4439 * first and 200 as anon first.
4442 type
= LRU_GEN_FILE
;
4443 else if (min_seq
[LRU_GEN_ANON
] < min_seq
[LRU_GEN_FILE
])
4444 type
= LRU_GEN_ANON
;
4445 else if (swappiness
== 1)
4446 type
= LRU_GEN_FILE
;
4447 else if (swappiness
== 200)
4448 type
= LRU_GEN_ANON
;
4450 type
= get_type_to_scan(lruvec
, swappiness
, &tier
);
4452 for (i
= !swappiness
; i
< ANON_AND_FILE
; i
++) {
4454 tier
= get_tier_idx(lruvec
, type
);
4456 scanned
= scan_folios(lruvec
, sc
, type
, tier
, list
);
4464 *type_scanned
= type
;
4469 static int evict_folios(struct lruvec
*lruvec
, struct scan_control
*sc
, int swappiness
)
4476 struct folio
*folio
;
4478 enum vm_event_item item
;
4479 struct reclaim_stat stat
;
4480 struct lru_gen_mm_walk
*walk
;
4481 bool skip_retry
= false;
4482 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
4483 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
4485 spin_lock_irq(&lruvec
->lru_lock
);
4487 scanned
= isolate_folios(lruvec
, sc
, swappiness
, &type
, &list
);
4489 scanned
+= try_to_inc_min_seq(lruvec
, swappiness
);
4491 if (get_nr_gens(lruvec
, !swappiness
) == MIN_NR_GENS
)
4494 spin_unlock_irq(&lruvec
->lru_lock
);
4496 if (list_empty(&list
))
4499 reclaimed
= shrink_folio_list(&list
, pgdat
, sc
, &stat
, false);
4500 sc
->nr_reclaimed
+= reclaimed
;
4501 trace_mm_vmscan_lru_shrink_inactive(pgdat
->node_id
,
4502 scanned
, reclaimed
, &stat
, sc
->priority
,
4503 type
? LRU_INACTIVE_FILE
: LRU_INACTIVE_ANON
);
4505 list_for_each_entry_safe_reverse(folio
, next
, &list
, lru
) {
4506 if (!folio_evictable(folio
)) {
4507 list_del(&folio
->lru
);
4508 folio_putback_lru(folio
);
4512 if (folio_test_reclaim(folio
) &&
4513 (folio_test_dirty(folio
) || folio_test_writeback(folio
))) {
4514 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4515 if (folio_test_workingset(folio
))
4516 folio_set_referenced(folio
);
4520 if (skip_retry
|| folio_test_active(folio
) || folio_test_referenced(folio
) ||
4521 folio_mapped(folio
) || folio_test_locked(folio
) ||
4522 folio_test_dirty(folio
) || folio_test_writeback(folio
)) {
4523 /* don't add rejected folios to the oldest generation */
4524 set_mask_bits(&folio
->flags
, LRU_REFS_MASK
| LRU_REFS_FLAGS
,
4529 /* retry folios that may have missed folio_rotate_reclaimable() */
4530 list_move(&folio
->lru
, &clean
);
4531 sc
->nr_scanned
-= folio_nr_pages(folio
);
4534 spin_lock_irq(&lruvec
->lru_lock
);
4536 move_folios_to_lru(lruvec
, &list
);
4538 walk
= current
->reclaim_state
->mm_walk
;
4539 if (walk
&& walk
->batched
)
4540 reset_batch_size(lruvec
, walk
);
4542 item
= PGSTEAL_KSWAPD
+ reclaimer_offset();
4543 if (!cgroup_reclaim(sc
))
4544 __count_vm_events(item
, reclaimed
);
4545 __count_memcg_events(memcg
, item
, reclaimed
);
4546 __count_vm_events(PGSTEAL_ANON
+ type
, reclaimed
);
4548 spin_unlock_irq(&lruvec
->lru_lock
);
4550 mem_cgroup_uncharge_list(&list
);
4551 free_unref_page_list(&list
);
4553 INIT_LIST_HEAD(&list
);
4554 list_splice_init(&clean
, &list
);
4556 if (!list_empty(&list
)) {
4564 static bool should_run_aging(struct lruvec
*lruvec
, unsigned long max_seq
,
4565 struct scan_control
*sc
, bool can_swap
, unsigned long *nr_to_scan
)
4567 int gen
, type
, zone
;
4568 unsigned long old
= 0;
4569 unsigned long young
= 0;
4570 unsigned long total
= 0;
4571 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
4572 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
4573 DEFINE_MIN_SEQ(lruvec
);
4575 /* whether this lruvec is completely out of cold folios */
4576 if (min_seq
[!can_swap
] + MIN_NR_GENS
> max_seq
) {
4581 for (type
= !can_swap
; type
< ANON_AND_FILE
; type
++) {
4584 for (seq
= min_seq
[type
]; seq
<= max_seq
; seq
++) {
4585 unsigned long size
= 0;
4587 gen
= lru_gen_from_seq(seq
);
4589 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++)
4590 size
+= max(READ_ONCE(lrugen
->nr_pages
[gen
][type
][zone
]), 0L);
4595 else if (seq
+ MIN_NR_GENS
== max_seq
)
4600 /* try to scrape all its memory if this memcg was deleted */
4601 if (!mem_cgroup_online(memcg
)) {
4602 *nr_to_scan
= total
;
4606 *nr_to_scan
= total
>> sc
->priority
;
4609 * The aging tries to be lazy to reduce the overhead, while the eviction
4610 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4611 * ideal number of generations is MIN_NR_GENS+1.
4613 if (min_seq
[!can_swap
] + MIN_NR_GENS
< max_seq
)
4617 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4618 * of the total number of pages for each generation. A reasonable range
4619 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4620 * aging cares about the upper bound of hot pages, while the eviction
4621 * cares about the lower bound of cold pages.
4623 if (young
* MIN_NR_GENS
> total
)
4625 if (old
* (MIN_NR_GENS
+ 2) < total
)
4632 * For future optimizations:
4633 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4636 static long get_nr_to_scan(struct lruvec
*lruvec
, struct scan_control
*sc
, bool can_swap
)
4638 unsigned long nr_to_scan
;
4639 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
4640 DEFINE_MAX_SEQ(lruvec
);
4642 if (mem_cgroup_below_min(sc
->target_mem_cgroup
, memcg
))
4645 if (!should_run_aging(lruvec
, max_seq
, sc
, can_swap
, &nr_to_scan
))
4648 /* skip the aging path at the default priority */
4649 if (sc
->priority
== DEF_PRIORITY
)
4652 /* skip this lruvec as it's low on cold folios */
4653 return try_to_inc_max_seq(lruvec
, max_seq
, sc
, can_swap
, false) ? -1 : 0;
4656 static bool should_abort_scan(struct lruvec
*lruvec
, struct scan_control
*sc
)
4659 enum zone_watermarks mark
;
4661 /* don't abort memcg reclaim to ensure fairness */
4662 if (!root_reclaim(sc
))
4665 if (sc
->nr_reclaimed
>= max(sc
->nr_to_reclaim
, compact_gap(sc
->order
)))
4668 /* check the order to exclude compaction-induced reclaim */
4669 if (!current_is_kswapd() || sc
->order
)
4672 mark
= sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
?
4673 WMARK_PROMO
: WMARK_HIGH
;
4675 for (i
= 0; i
<= sc
->reclaim_idx
; i
++) {
4676 struct zone
*zone
= lruvec_pgdat(lruvec
)->node_zones
+ i
;
4677 unsigned long size
= wmark_pages(zone
, mark
) + MIN_LRU_BATCH
;
4679 if (managed_zone(zone
) && !zone_watermark_ok(zone
, 0, size
, sc
->reclaim_idx
, 0))
4683 /* kswapd should abort if all eligible zones are safe */
4687 static bool try_to_shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
4690 unsigned long scanned
= 0;
4691 int swappiness
= get_swappiness(lruvec
, sc
);
4693 /* clean file folios are more likely to exist */
4694 if (swappiness
&& !(sc
->gfp_mask
& __GFP_IO
))
4700 nr_to_scan
= get_nr_to_scan(lruvec
, sc
, swappiness
);
4701 if (nr_to_scan
<= 0)
4704 delta
= evict_folios(lruvec
, sc
, swappiness
);
4709 if (scanned
>= nr_to_scan
)
4712 if (should_abort_scan(lruvec
, sc
))
4718 /* whether this lruvec should be rotated */
4719 return nr_to_scan
< 0;
4722 static int shrink_one(struct lruvec
*lruvec
, struct scan_control
*sc
)
4725 unsigned long scanned
= sc
->nr_scanned
;
4726 unsigned long reclaimed
= sc
->nr_reclaimed
;
4727 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
4728 struct pglist_data
*pgdat
= lruvec_pgdat(lruvec
);
4730 mem_cgroup_calculate_protection(NULL
, memcg
);
4732 if (mem_cgroup_below_min(NULL
, memcg
))
4733 return MEMCG_LRU_YOUNG
;
4735 if (mem_cgroup_below_low(NULL
, memcg
)) {
4736 /* see the comment on MEMCG_NR_GENS */
4737 if (lru_gen_memcg_seg(lruvec
) != MEMCG_LRU_TAIL
)
4738 return MEMCG_LRU_TAIL
;
4740 memcg_memory_event(memcg
, MEMCG_LOW
);
4743 success
= try_to_shrink_lruvec(lruvec
, sc
);
4745 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
, sc
->priority
);
4748 vmpressure(sc
->gfp_mask
, memcg
, false, sc
->nr_scanned
- scanned
,
4749 sc
->nr_reclaimed
- reclaimed
);
4751 flush_reclaim_state(sc
);
4753 if (success
&& mem_cgroup_online(memcg
))
4754 return MEMCG_LRU_YOUNG
;
4756 if (!success
&& lruvec_is_sizable(lruvec
, sc
))
4759 /* one retry if offlined or too small */
4760 return lru_gen_memcg_seg(lruvec
) != MEMCG_LRU_TAIL
?
4761 MEMCG_LRU_TAIL
: MEMCG_LRU_YOUNG
;
4766 static void shrink_many(struct pglist_data
*pgdat
, struct scan_control
*sc
)
4772 struct lruvec
*lruvec
;
4773 struct lru_gen_folio
*lrugen
;
4774 struct mem_cgroup
*memcg
;
4775 struct hlist_nulls_node
*pos
;
4777 gen
= get_memcg_gen(READ_ONCE(pgdat
->memcg_lru
.seq
));
4778 bin
= first_bin
= get_random_u32_below(MEMCG_NR_BINS
);
4785 hlist_nulls_for_each_entry_rcu(lrugen
, pos
, &pgdat
->memcg_lru
.fifo
[gen
][bin
], list
) {
4787 lru_gen_rotate_memcg(lruvec
, op
);
4791 mem_cgroup_put(memcg
);
4794 if (gen
!= READ_ONCE(lrugen
->gen
))
4797 lruvec
= container_of(lrugen
, struct lruvec
, lrugen
);
4798 memcg
= lruvec_memcg(lruvec
);
4800 if (!mem_cgroup_tryget(memcg
)) {
4801 lru_gen_release_memcg(memcg
);
4808 op
= shrink_one(lruvec
, sc
);
4812 if (should_abort_scan(lruvec
, sc
))
4819 lru_gen_rotate_memcg(lruvec
, op
);
4821 mem_cgroup_put(memcg
);
4823 if (!is_a_nulls(pos
))
4826 /* restart if raced with lru_gen_rotate_memcg() */
4827 if (gen
!= get_nulls_value(pos
))
4830 /* try the rest of the bins of the current generation */
4831 bin
= get_memcg_bin(bin
+ 1);
4832 if (bin
!= first_bin
)
4836 static void lru_gen_shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
4838 struct blk_plug plug
;
4840 VM_WARN_ON_ONCE(root_reclaim(sc
));
4841 VM_WARN_ON_ONCE(!sc
->may_writepage
|| !sc
->may_unmap
);
4845 blk_start_plug(&plug
);
4847 set_mm_walk(NULL
, sc
->proactive
);
4849 if (try_to_shrink_lruvec(lruvec
, sc
))
4850 lru_gen_rotate_memcg(lruvec
, MEMCG_LRU_YOUNG
);
4854 blk_finish_plug(&plug
);
4857 #else /* !CONFIG_MEMCG */
4859 static void shrink_many(struct pglist_data
*pgdat
, struct scan_control
*sc
)
4864 static void lru_gen_shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
4871 static void set_initial_priority(struct pglist_data
*pgdat
, struct scan_control
*sc
)
4874 unsigned long reclaimable
;
4875 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
4877 if (sc
->priority
!= DEF_PRIORITY
|| sc
->nr_to_reclaim
< MIN_LRU_BATCH
)
4880 * Determine the initial priority based on
4881 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4882 * where reclaimed_to_scanned_ratio = inactive / total.
4884 reclaimable
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
4885 if (get_swappiness(lruvec
, sc
))
4886 reclaimable
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
4888 /* round down reclaimable and round up sc->nr_to_reclaim */
4889 priority
= fls_long(reclaimable
) - 1 - fls_long(sc
->nr_to_reclaim
- 1);
4891 sc
->priority
= clamp(priority
, 0, DEF_PRIORITY
);
4894 static void lru_gen_shrink_node(struct pglist_data
*pgdat
, struct scan_control
*sc
)
4896 struct blk_plug plug
;
4897 unsigned long reclaimed
= sc
->nr_reclaimed
;
4899 VM_WARN_ON_ONCE(!root_reclaim(sc
));
4902 * Unmapped clean folios are already prioritized. Scanning for more of
4903 * them is likely futile and can cause high reclaim latency when there
4904 * is a large number of memcgs.
4906 if (!sc
->may_writepage
|| !sc
->may_unmap
)
4911 blk_start_plug(&plug
);
4913 set_mm_walk(pgdat
, sc
->proactive
);
4915 set_initial_priority(pgdat
, sc
);
4917 if (current_is_kswapd())
4918 sc
->nr_reclaimed
= 0;
4920 if (mem_cgroup_disabled())
4921 shrink_one(&pgdat
->__lruvec
, sc
);
4923 shrink_many(pgdat
, sc
);
4925 if (current_is_kswapd())
4926 sc
->nr_reclaimed
+= reclaimed
;
4930 blk_finish_plug(&plug
);
4932 /* kswapd should never fail */
4933 pgdat
->kswapd_failures
= 0;
4936 /******************************************************************************
4938 ******************************************************************************/
4940 static bool __maybe_unused
state_is_valid(struct lruvec
*lruvec
)
4942 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
4944 if (lrugen
->enabled
) {
4947 for_each_evictable_lru(lru
) {
4948 if (!list_empty(&lruvec
->lists
[lru
]))
4952 int gen
, type
, zone
;
4954 for_each_gen_type_zone(gen
, type
, zone
) {
4955 if (!list_empty(&lrugen
->folios
[gen
][type
][zone
]))
4963 static bool fill_evictable(struct lruvec
*lruvec
)
4966 int remaining
= MAX_LRU_BATCH
;
4968 for_each_evictable_lru(lru
) {
4969 int type
= is_file_lru(lru
);
4970 bool active
= is_active_lru(lru
);
4971 struct list_head
*head
= &lruvec
->lists
[lru
];
4973 while (!list_empty(head
)) {
4975 struct folio
*folio
= lru_to_folio(head
);
4977 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio
), folio
);
4978 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio
) != active
, folio
);
4979 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio
) != type
, folio
);
4980 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio
) != -1, folio
);
4982 lruvec_del_folio(lruvec
, folio
);
4983 success
= lru_gen_add_folio(lruvec
, folio
, false);
4984 VM_WARN_ON_ONCE(!success
);
4994 static bool drain_evictable(struct lruvec
*lruvec
)
4996 int gen
, type
, zone
;
4997 int remaining
= MAX_LRU_BATCH
;
4999 for_each_gen_type_zone(gen
, type
, zone
) {
5000 struct list_head
*head
= &lruvec
->lrugen
.folios
[gen
][type
][zone
];
5002 while (!list_empty(head
)) {
5004 struct folio
*folio
= lru_to_folio(head
);
5006 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio
), folio
);
5007 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio
), folio
);
5008 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio
) != type
, folio
);
5009 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio
) != zone
, folio
);
5011 success
= lru_gen_del_folio(lruvec
, folio
, false);
5012 VM_WARN_ON_ONCE(!success
);
5013 lruvec_add_folio(lruvec
, folio
);
5023 static void lru_gen_change_state(bool enabled
)
5025 static DEFINE_MUTEX(state_mutex
);
5027 struct mem_cgroup
*memcg
;
5032 mutex_lock(&state_mutex
);
5034 if (enabled
== lru_gen_enabled())
5038 static_branch_enable_cpuslocked(&lru_gen_caps
[LRU_GEN_CORE
]);
5040 static_branch_disable_cpuslocked(&lru_gen_caps
[LRU_GEN_CORE
]);
5042 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
5046 for_each_node(nid
) {
5047 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
5049 spin_lock_irq(&lruvec
->lru_lock
);
5051 VM_WARN_ON_ONCE(!seq_is_valid(lruvec
));
5052 VM_WARN_ON_ONCE(!state_is_valid(lruvec
));
5054 lruvec
->lrugen
.enabled
= enabled
;
5056 while (!(enabled
? fill_evictable(lruvec
) : drain_evictable(lruvec
))) {
5057 spin_unlock_irq(&lruvec
->lru_lock
);
5059 spin_lock_irq(&lruvec
->lru_lock
);
5062 spin_unlock_irq(&lruvec
->lru_lock
);
5066 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)));
5068 mutex_unlock(&state_mutex
);
5074 /******************************************************************************
5076 ******************************************************************************/
5078 static ssize_t
min_ttl_ms_show(struct kobject
*kobj
, struct kobj_attribute
*attr
, char *buf
)
5080 return sysfs_emit(buf
, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl
)));
5083 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5084 static ssize_t
min_ttl_ms_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
5085 const char *buf
, size_t len
)
5089 if (kstrtouint(buf
, 0, &msecs
))
5092 WRITE_ONCE(lru_gen_min_ttl
, msecs_to_jiffies(msecs
));
5097 static struct kobj_attribute lru_gen_min_ttl_attr
= __ATTR_RW(min_ttl_ms
);
5099 static ssize_t
enabled_show(struct kobject
*kobj
, struct kobj_attribute
*attr
, char *buf
)
5101 unsigned int caps
= 0;
5103 if (get_cap(LRU_GEN_CORE
))
5104 caps
|= BIT(LRU_GEN_CORE
);
5106 if (should_walk_mmu())
5107 caps
|= BIT(LRU_GEN_MM_WALK
);
5109 if (should_clear_pmd_young())
5110 caps
|= BIT(LRU_GEN_NONLEAF_YOUNG
);
5112 return sysfs_emit(buf
, "0x%04x\n", caps
);
5115 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5116 static ssize_t
enabled_store(struct kobject
*kobj
, struct kobj_attribute
*attr
,
5117 const char *buf
, size_t len
)
5122 if (tolower(*buf
) == 'n')
5124 else if (tolower(*buf
) == 'y')
5126 else if (kstrtouint(buf
, 0, &caps
))
5129 for (i
= 0; i
< NR_LRU_GEN_CAPS
; i
++) {
5130 bool enabled
= caps
& BIT(i
);
5132 if (i
== LRU_GEN_CORE
)
5133 lru_gen_change_state(enabled
);
5135 static_branch_enable(&lru_gen_caps
[i
]);
5137 static_branch_disable(&lru_gen_caps
[i
]);
5143 static struct kobj_attribute lru_gen_enabled_attr
= __ATTR_RW(enabled
);
5145 static struct attribute
*lru_gen_attrs
[] = {
5146 &lru_gen_min_ttl_attr
.attr
,
5147 &lru_gen_enabled_attr
.attr
,
5151 static const struct attribute_group lru_gen_attr_group
= {
5153 .attrs
= lru_gen_attrs
,
5156 /******************************************************************************
5158 ******************************************************************************/
5160 static void *lru_gen_seq_start(struct seq_file
*m
, loff_t
*pos
)
5162 struct mem_cgroup
*memcg
;
5163 loff_t nr_to_skip
= *pos
;
5165 m
->private = kvmalloc(PATH_MAX
, GFP_KERNEL
);
5167 return ERR_PTR(-ENOMEM
);
5169 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
5173 for_each_node_state(nid
, N_MEMORY
) {
5175 return get_lruvec(memcg
, nid
);
5177 } while ((memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
)));
5182 static void lru_gen_seq_stop(struct seq_file
*m
, void *v
)
5184 if (!IS_ERR_OR_NULL(v
))
5185 mem_cgroup_iter_break(NULL
, lruvec_memcg(v
));
5191 static void *lru_gen_seq_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
5193 int nid
= lruvec_pgdat(v
)->node_id
;
5194 struct mem_cgroup
*memcg
= lruvec_memcg(v
);
5198 nid
= next_memory_node(nid
);
5199 if (nid
== MAX_NUMNODES
) {
5200 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
5204 nid
= first_memory_node
;
5207 return get_lruvec(memcg
, nid
);
5210 static void lru_gen_seq_show_full(struct seq_file
*m
, struct lruvec
*lruvec
,
5211 unsigned long max_seq
, unsigned long *min_seq
,
5216 int hist
= lru_hist_from_seq(seq
);
5217 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
5219 for (tier
= 0; tier
< MAX_NR_TIERS
; tier
++) {
5220 seq_printf(m
, " %10d", tier
);
5221 for (type
= 0; type
< ANON_AND_FILE
; type
++) {
5222 const char *s
= " ";
5223 unsigned long n
[3] = {};
5225 if (seq
== max_seq
) {
5227 n
[0] = READ_ONCE(lrugen
->avg_refaulted
[type
][tier
]);
5228 n
[1] = READ_ONCE(lrugen
->avg_total
[type
][tier
]);
5229 } else if (seq
== min_seq
[type
] || NR_HIST_GENS
> 1) {
5231 n
[0] = atomic_long_read(&lrugen
->refaulted
[hist
][type
][tier
]);
5232 n
[1] = atomic_long_read(&lrugen
->evicted
[hist
][type
][tier
]);
5234 n
[2] = READ_ONCE(lrugen
->protected[hist
][type
][tier
- 1]);
5237 for (i
= 0; i
< 3; i
++)
5238 seq_printf(m
, " %10lu%c", n
[i
], s
[i
]);
5244 for (i
= 0; i
< NR_MM_STATS
; i
++) {
5245 const char *s
= " ";
5246 unsigned long n
= 0;
5248 if (seq
== max_seq
&& NR_HIST_GENS
== 1) {
5250 n
= READ_ONCE(lruvec
->mm_state
.stats
[hist
][i
]);
5251 } else if (seq
!= max_seq
&& NR_HIST_GENS
> 1) {
5253 n
= READ_ONCE(lruvec
->mm_state
.stats
[hist
][i
]);
5256 seq_printf(m
, " %10lu%c", n
, s
[i
]);
5261 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5262 static int lru_gen_seq_show(struct seq_file
*m
, void *v
)
5265 bool full
= !debugfs_real_fops(m
->file
)->write
;
5266 struct lruvec
*lruvec
= v
;
5267 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
5268 int nid
= lruvec_pgdat(lruvec
)->node_id
;
5269 struct mem_cgroup
*memcg
= lruvec_memcg(lruvec
);
5270 DEFINE_MAX_SEQ(lruvec
);
5271 DEFINE_MIN_SEQ(lruvec
);
5273 if (nid
== first_memory_node
) {
5274 const char *path
= memcg
? m
->private : "";
5278 cgroup_path(memcg
->css
.cgroup
, m
->private, PATH_MAX
);
5280 seq_printf(m
, "memcg %5hu %s\n", mem_cgroup_id(memcg
), path
);
5283 seq_printf(m
, " node %5d\n", nid
);
5286 seq
= min_seq
[LRU_GEN_ANON
];
5287 else if (max_seq
>= MAX_NR_GENS
)
5288 seq
= max_seq
- MAX_NR_GENS
+ 1;
5292 for (; seq
<= max_seq
; seq
++) {
5294 int gen
= lru_gen_from_seq(seq
);
5295 unsigned long birth
= READ_ONCE(lruvec
->lrugen
.timestamps
[gen
]);
5297 seq_printf(m
, " %10lu %10u", seq
, jiffies_to_msecs(jiffies
- birth
));
5299 for (type
= 0; type
< ANON_AND_FILE
; type
++) {
5300 unsigned long size
= 0;
5301 char mark
= full
&& seq
< min_seq
[type
] ? 'x' : ' ';
5303 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++)
5304 size
+= max(READ_ONCE(lrugen
->nr_pages
[gen
][type
][zone
]), 0L);
5306 seq_printf(m
, " %10lu%c", size
, mark
);
5312 lru_gen_seq_show_full(m
, lruvec
, max_seq
, min_seq
, seq
);
5318 static const struct seq_operations lru_gen_seq_ops
= {
5319 .start
= lru_gen_seq_start
,
5320 .stop
= lru_gen_seq_stop
,
5321 .next
= lru_gen_seq_next
,
5322 .show
= lru_gen_seq_show
,
5325 static int run_aging(struct lruvec
*lruvec
, unsigned long seq
, struct scan_control
*sc
,
5326 bool can_swap
, bool force_scan
)
5328 DEFINE_MAX_SEQ(lruvec
);
5329 DEFINE_MIN_SEQ(lruvec
);
5337 if (!force_scan
&& min_seq
[!can_swap
] + MAX_NR_GENS
- 1 <= max_seq
)
5340 try_to_inc_max_seq(lruvec
, max_seq
, sc
, can_swap
, force_scan
);
5345 static int run_eviction(struct lruvec
*lruvec
, unsigned long seq
, struct scan_control
*sc
,
5346 int swappiness
, unsigned long nr_to_reclaim
)
5348 DEFINE_MAX_SEQ(lruvec
);
5350 if (seq
+ MIN_NR_GENS
> max_seq
)
5353 sc
->nr_reclaimed
= 0;
5355 while (!signal_pending(current
)) {
5356 DEFINE_MIN_SEQ(lruvec
);
5358 if (seq
< min_seq
[!swappiness
])
5361 if (sc
->nr_reclaimed
>= nr_to_reclaim
)
5364 if (!evict_folios(lruvec
, sc
, swappiness
))
5373 static int run_cmd(char cmd
, int memcg_id
, int nid
, unsigned long seq
,
5374 struct scan_control
*sc
, int swappiness
, unsigned long opt
)
5376 struct lruvec
*lruvec
;
5378 struct mem_cgroup
*memcg
= NULL
;
5380 if (nid
< 0 || nid
>= MAX_NUMNODES
|| !node_state(nid
, N_MEMORY
))
5383 if (!mem_cgroup_disabled()) {
5386 memcg
= mem_cgroup_from_id(memcg_id
);
5387 if (!mem_cgroup_tryget(memcg
))
5396 if (memcg_id
!= mem_cgroup_id(memcg
))
5399 lruvec
= get_lruvec(memcg
, nid
);
5402 swappiness
= get_swappiness(lruvec
, sc
);
5403 else if (swappiness
> 200)
5408 err
= run_aging(lruvec
, seq
, sc
, swappiness
, opt
);
5411 err
= run_eviction(lruvec
, seq
, sc
, swappiness
, opt
);
5415 mem_cgroup_put(memcg
);
5420 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5421 static ssize_t
lru_gen_seq_write(struct file
*file
, const char __user
*src
,
5422 size_t len
, loff_t
*pos
)
5427 struct blk_plug plug
;
5429 struct scan_control sc
= {
5430 .may_writepage
= true,
5433 .reclaim_idx
= MAX_NR_ZONES
- 1,
5434 .gfp_mask
= GFP_KERNEL
,
5437 buf
= kvmalloc(len
+ 1, GFP_KERNEL
);
5441 if (copy_from_user(buf
, src
, len
)) {
5446 set_task_reclaim_state(current
, &sc
.reclaim_state
);
5447 flags
= memalloc_noreclaim_save();
5448 blk_start_plug(&plug
);
5449 if (!set_mm_walk(NULL
, true)) {
5457 while ((cur
= strsep(&next
, ",;\n"))) {
5461 unsigned int memcg_id
;
5464 unsigned int swappiness
= -1;
5465 unsigned long opt
= -1;
5467 cur
= skip_spaces(cur
);
5471 n
= sscanf(cur
, "%c %u %u %lu %n %u %n %lu %n", &cmd
, &memcg_id
, &nid
,
5472 &seq
, &end
, &swappiness
, &end
, &opt
, &end
);
5473 if (n
< 4 || cur
[end
]) {
5478 err
= run_cmd(cmd
, memcg_id
, nid
, seq
, &sc
, swappiness
, opt
);
5484 blk_finish_plug(&plug
);
5485 memalloc_noreclaim_restore(flags
);
5486 set_task_reclaim_state(current
, NULL
);
5493 static int lru_gen_seq_open(struct inode
*inode
, struct file
*file
)
5495 return seq_open(file
, &lru_gen_seq_ops
);
5498 static const struct file_operations lru_gen_rw_fops
= {
5499 .open
= lru_gen_seq_open
,
5501 .write
= lru_gen_seq_write
,
5502 .llseek
= seq_lseek
,
5503 .release
= seq_release
,
5506 static const struct file_operations lru_gen_ro_fops
= {
5507 .open
= lru_gen_seq_open
,
5509 .llseek
= seq_lseek
,
5510 .release
= seq_release
,
5513 /******************************************************************************
5515 ******************************************************************************/
5517 void lru_gen_init_lruvec(struct lruvec
*lruvec
)
5520 int gen
, type
, zone
;
5521 struct lru_gen_folio
*lrugen
= &lruvec
->lrugen
;
5523 lrugen
->max_seq
= MIN_NR_GENS
+ 1;
5524 lrugen
->enabled
= lru_gen_enabled();
5526 for (i
= 0; i
<= MIN_NR_GENS
+ 1; i
++)
5527 lrugen
->timestamps
[i
] = jiffies
;
5529 for_each_gen_type_zone(gen
, type
, zone
)
5530 INIT_LIST_HEAD(&lrugen
->folios
[gen
][type
][zone
]);
5532 lruvec
->mm_state
.seq
= MIN_NR_GENS
;
5537 void lru_gen_init_pgdat(struct pglist_data
*pgdat
)
5541 spin_lock_init(&pgdat
->memcg_lru
.lock
);
5543 for (i
= 0; i
< MEMCG_NR_GENS
; i
++) {
5544 for (j
= 0; j
< MEMCG_NR_BINS
; j
++)
5545 INIT_HLIST_NULLS_HEAD(&pgdat
->memcg_lru
.fifo
[i
][j
], i
);
5549 void lru_gen_init_memcg(struct mem_cgroup
*memcg
)
5551 INIT_LIST_HEAD(&memcg
->mm_list
.fifo
);
5552 spin_lock_init(&memcg
->mm_list
.lock
);
5555 void lru_gen_exit_memcg(struct mem_cgroup
*memcg
)
5560 VM_WARN_ON_ONCE(!list_empty(&memcg
->mm_list
.fifo
));
5562 for_each_node(nid
) {
5563 struct lruvec
*lruvec
= get_lruvec(memcg
, nid
);
5565 VM_WARN_ON_ONCE(memchr_inv(lruvec
->lrugen
.nr_pages
, 0,
5566 sizeof(lruvec
->lrugen
.nr_pages
)));
5568 lruvec
->lrugen
.list
.next
= LIST_POISON1
;
5570 for (i
= 0; i
< NR_BLOOM_FILTERS
; i
++) {
5571 bitmap_free(lruvec
->mm_state
.filters
[i
]);
5572 lruvec
->mm_state
.filters
[i
] = NULL
;
5577 #endif /* CONFIG_MEMCG */
5579 static int __init
init_lru_gen(void)
5581 BUILD_BUG_ON(MIN_NR_GENS
+ 1 >= MAX_NR_GENS
);
5582 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH
) <= MAX_NR_GENS
);
5584 if (sysfs_create_group(mm_kobj
, &lru_gen_attr_group
))
5585 pr_err("lru_gen: failed to create sysfs group\n");
5587 debugfs_create_file("lru_gen", 0644, NULL
, NULL
, &lru_gen_rw_fops
);
5588 debugfs_create_file("lru_gen_full", 0444, NULL
, NULL
, &lru_gen_ro_fops
);
5592 late_initcall(init_lru_gen
);
5594 #else /* !CONFIG_LRU_GEN */
5596 static void lru_gen_age_node(struct pglist_data
*pgdat
, struct scan_control
*sc
)
5600 static void lru_gen_shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
5604 static void lru_gen_shrink_node(struct pglist_data
*pgdat
, struct scan_control
*sc
)
5608 #endif /* CONFIG_LRU_GEN */
5610 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
5612 unsigned long nr
[NR_LRU_LISTS
];
5613 unsigned long targets
[NR_LRU_LISTS
];
5614 unsigned long nr_to_scan
;
5616 unsigned long nr_reclaimed
= 0;
5617 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
5618 bool proportional_reclaim
;
5619 struct blk_plug plug
;
5621 if (lru_gen_enabled() && !root_reclaim(sc
)) {
5622 lru_gen_shrink_lruvec(lruvec
, sc
);
5626 get_scan_count(lruvec
, sc
, nr
);
5628 /* Record the original scan target for proportional adjustments later */
5629 memcpy(targets
, nr
, sizeof(nr
));
5632 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5633 * event that can occur when there is little memory pressure e.g.
5634 * multiple streaming readers/writers. Hence, we do not abort scanning
5635 * when the requested number of pages are reclaimed when scanning at
5636 * DEF_PRIORITY on the assumption that the fact we are direct
5637 * reclaiming implies that kswapd is not keeping up and it is best to
5638 * do a batch of work at once. For memcg reclaim one check is made to
5639 * abort proportional reclaim if either the file or anon lru has already
5640 * dropped to zero at the first pass.
5642 proportional_reclaim
= (!cgroup_reclaim(sc
) && !current_is_kswapd() &&
5643 sc
->priority
== DEF_PRIORITY
);
5645 blk_start_plug(&plug
);
5646 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
5647 nr
[LRU_INACTIVE_FILE
]) {
5648 unsigned long nr_anon
, nr_file
, percentage
;
5649 unsigned long nr_scanned
;
5651 for_each_evictable_lru(lru
) {
5653 nr_to_scan
= min(nr
[lru
], SWAP_CLUSTER_MAX
);
5654 nr
[lru
] -= nr_to_scan
;
5656 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
5663 if (nr_reclaimed
< nr_to_reclaim
|| proportional_reclaim
)
5667 * For kswapd and memcg, reclaim at least the number of pages
5668 * requested. Ensure that the anon and file LRUs are scanned
5669 * proportionally what was requested by get_scan_count(). We
5670 * stop reclaiming one LRU and reduce the amount scanning
5671 * proportional to the original scan target.
5673 nr_file
= nr
[LRU_INACTIVE_FILE
] + nr
[LRU_ACTIVE_FILE
];
5674 nr_anon
= nr
[LRU_INACTIVE_ANON
] + nr
[LRU_ACTIVE_ANON
];
5677 * It's just vindictive to attack the larger once the smaller
5678 * has gone to zero. And given the way we stop scanning the
5679 * smaller below, this makes sure that we only make one nudge
5680 * towards proportionality once we've got nr_to_reclaim.
5682 if (!nr_file
|| !nr_anon
)
5685 if (nr_file
> nr_anon
) {
5686 unsigned long scan_target
= targets
[LRU_INACTIVE_ANON
] +
5687 targets
[LRU_ACTIVE_ANON
] + 1;
5689 percentage
= nr_anon
* 100 / scan_target
;
5691 unsigned long scan_target
= targets
[LRU_INACTIVE_FILE
] +
5692 targets
[LRU_ACTIVE_FILE
] + 1;
5694 percentage
= nr_file
* 100 / scan_target
;
5697 /* Stop scanning the smaller of the LRU */
5699 nr
[lru
+ LRU_ACTIVE
] = 0;
5702 * Recalculate the other LRU scan count based on its original
5703 * scan target and the percentage scanning already complete
5705 lru
= (lru
== LRU_FILE
) ? LRU_BASE
: LRU_FILE
;
5706 nr_scanned
= targets
[lru
] - nr
[lru
];
5707 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
5708 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
5711 nr_scanned
= targets
[lru
] - nr
[lru
];
5712 nr
[lru
] = targets
[lru
] * (100 - percentage
) / 100;
5713 nr
[lru
] -= min(nr
[lru
], nr_scanned
);
5715 blk_finish_plug(&plug
);
5716 sc
->nr_reclaimed
+= nr_reclaimed
;
5719 * Even if we did not try to evict anon pages at all, we want to
5720 * rebalance the anon lru active/inactive ratio.
5722 if (can_age_anon_pages(lruvec_pgdat(lruvec
), sc
) &&
5723 inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
5724 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
5725 sc
, LRU_ACTIVE_ANON
);
5728 /* Use reclaim/compaction for costly allocs or under memory pressure */
5729 static bool in_reclaim_compaction(struct scan_control
*sc
)
5731 if (IS_ENABLED(CONFIG_COMPACTION
) && sc
->order
&&
5732 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
5733 sc
->priority
< DEF_PRIORITY
- 2))
5740 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5741 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5742 * true if more pages should be reclaimed such that when the page allocator
5743 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5744 * It will give up earlier than that if there is difficulty reclaiming pages.
5746 static inline bool should_continue_reclaim(struct pglist_data
*pgdat
,
5747 unsigned long nr_reclaimed
,
5748 struct scan_control
*sc
)
5750 unsigned long pages_for_compaction
;
5751 unsigned long inactive_lru_pages
;
5754 /* If not in reclaim/compaction mode, stop */
5755 if (!in_reclaim_compaction(sc
))
5759 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5760 * number of pages that were scanned. This will return to the caller
5761 * with the risk reclaim/compaction and the resulting allocation attempt
5762 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5763 * allocations through requiring that the full LRU list has been scanned
5764 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5765 * scan, but that approximation was wrong, and there were corner cases
5766 * where always a non-zero amount of pages were scanned.
5771 /* If compaction would go ahead or the allocation would succeed, stop */
5772 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
5773 struct zone
*zone
= &pgdat
->node_zones
[z
];
5774 if (!managed_zone(zone
))
5777 /* Allocation can already succeed, nothing to do */
5778 if (zone_watermark_ok(zone
, sc
->order
, min_wmark_pages(zone
),
5779 sc
->reclaim_idx
, 0))
5782 if (compaction_suitable(zone
, sc
->order
, sc
->reclaim_idx
))
5787 * If we have not reclaimed enough pages for compaction and the
5788 * inactive lists are large enough, continue reclaiming
5790 pages_for_compaction
= compact_gap(sc
->order
);
5791 inactive_lru_pages
= node_page_state(pgdat
, NR_INACTIVE_FILE
);
5792 if (can_reclaim_anon_pages(NULL
, pgdat
->node_id
, sc
))
5793 inactive_lru_pages
+= node_page_state(pgdat
, NR_INACTIVE_ANON
);
5795 return inactive_lru_pages
> pages_for_compaction
;
5798 static void shrink_node_memcgs(pg_data_t
*pgdat
, struct scan_control
*sc
)
5800 struct mem_cgroup
*target_memcg
= sc
->target_mem_cgroup
;
5801 struct mem_cgroup
*memcg
;
5803 memcg
= mem_cgroup_iter(target_memcg
, NULL
, NULL
);
5805 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
5806 unsigned long reclaimed
;
5807 unsigned long scanned
;
5810 * This loop can become CPU-bound when target memcgs
5811 * aren't eligible for reclaim - either because they
5812 * don't have any reclaimable pages, or because their
5813 * memory is explicitly protected. Avoid soft lockups.
5817 mem_cgroup_calculate_protection(target_memcg
, memcg
);
5819 if (mem_cgroup_below_min(target_memcg
, memcg
)) {
5822 * If there is no reclaimable memory, OOM.
5825 } else if (mem_cgroup_below_low(target_memcg
, memcg
)) {
5828 * Respect the protection only as long as
5829 * there is an unprotected supply
5830 * of reclaimable memory from other cgroups.
5832 if (!sc
->memcg_low_reclaim
) {
5833 sc
->memcg_low_skipped
= 1;
5836 memcg_memory_event(memcg
, MEMCG_LOW
);
5839 reclaimed
= sc
->nr_reclaimed
;
5840 scanned
= sc
->nr_scanned
;
5842 shrink_lruvec(lruvec
, sc
);
5844 shrink_slab(sc
->gfp_mask
, pgdat
->node_id
, memcg
,
5847 /* Record the group's reclaim efficiency */
5849 vmpressure(sc
->gfp_mask
, memcg
, false,
5850 sc
->nr_scanned
- scanned
,
5851 sc
->nr_reclaimed
- reclaimed
);
5853 } while ((memcg
= mem_cgroup_iter(target_memcg
, memcg
, NULL
)));
5856 static void shrink_node(pg_data_t
*pgdat
, struct scan_control
*sc
)
5858 unsigned long nr_reclaimed
, nr_scanned
, nr_node_reclaimed
;
5859 struct lruvec
*target_lruvec
;
5860 bool reclaimable
= false;
5862 if (lru_gen_enabled() && root_reclaim(sc
)) {
5863 lru_gen_shrink_node(pgdat
, sc
);
5867 target_lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
, pgdat
);
5870 memset(&sc
->nr
, 0, sizeof(sc
->nr
));
5872 nr_reclaimed
= sc
->nr_reclaimed
;
5873 nr_scanned
= sc
->nr_scanned
;
5875 prepare_scan_control(pgdat
, sc
);
5877 shrink_node_memcgs(pgdat
, sc
);
5879 flush_reclaim_state(sc
);
5881 nr_node_reclaimed
= sc
->nr_reclaimed
- nr_reclaimed
;
5883 /* Record the subtree's reclaim efficiency */
5885 vmpressure(sc
->gfp_mask
, sc
->target_mem_cgroup
, true,
5886 sc
->nr_scanned
- nr_scanned
, nr_node_reclaimed
);
5888 if (nr_node_reclaimed
)
5891 if (current_is_kswapd()) {
5893 * If reclaim is isolating dirty pages under writeback,
5894 * it implies that the long-lived page allocation rate
5895 * is exceeding the page laundering rate. Either the
5896 * global limits are not being effective at throttling
5897 * processes due to the page distribution throughout
5898 * zones or there is heavy usage of a slow backing
5899 * device. The only option is to throttle from reclaim
5900 * context which is not ideal as there is no guarantee
5901 * the dirtying process is throttled in the same way
5902 * balance_dirty_pages() manages.
5904 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5905 * count the number of pages under pages flagged for
5906 * immediate reclaim and stall if any are encountered
5907 * in the nr_immediate check below.
5909 if (sc
->nr
.writeback
&& sc
->nr
.writeback
== sc
->nr
.taken
)
5910 set_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
5912 /* Allow kswapd to start writing pages during reclaim.*/
5913 if (sc
->nr
.unqueued_dirty
== sc
->nr
.file_taken
)
5914 set_bit(PGDAT_DIRTY
, &pgdat
->flags
);
5917 * If kswapd scans pages marked for immediate
5918 * reclaim and under writeback (nr_immediate), it
5919 * implies that pages are cycling through the LRU
5920 * faster than they are written so forcibly stall
5921 * until some pages complete writeback.
5923 if (sc
->nr
.immediate
)
5924 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_WRITEBACK
);
5928 * Tag a node/memcg as congested if all the dirty pages were marked
5929 * for writeback and immediate reclaim (counted in nr.congested).
5931 * Legacy memcg will stall in page writeback so avoid forcibly
5932 * stalling in reclaim_throttle().
5934 if (sc
->nr
.dirty
&& sc
->nr
.dirty
== sc
->nr
.congested
) {
5935 if (cgroup_reclaim(sc
) && writeback_throttling_sane(sc
))
5936 set_bit(LRUVEC_CGROUP_CONGESTED
, &target_lruvec
->flags
);
5938 if (current_is_kswapd())
5939 set_bit(LRUVEC_NODE_CONGESTED
, &target_lruvec
->flags
);
5943 * Stall direct reclaim for IO completions if the lruvec is
5944 * node is congested. Allow kswapd to continue until it
5945 * starts encountering unqueued dirty pages or cycling through
5946 * the LRU too quickly.
5948 if (!current_is_kswapd() && current_may_throttle() &&
5949 !sc
->hibernation_mode
&&
5950 (test_bit(LRUVEC_CGROUP_CONGESTED
, &target_lruvec
->flags
) ||
5951 test_bit(LRUVEC_NODE_CONGESTED
, &target_lruvec
->flags
)))
5952 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_CONGESTED
);
5954 if (should_continue_reclaim(pgdat
, nr_node_reclaimed
, sc
))
5958 * Kswapd gives up on balancing particular nodes after too
5959 * many failures to reclaim anything from them and goes to
5960 * sleep. On reclaim progress, reset the failure counter. A
5961 * successful direct reclaim run will revive a dormant kswapd.
5964 pgdat
->kswapd_failures
= 0;
5968 * Returns true if compaction should go ahead for a costly-order request, or
5969 * the allocation would already succeed without compaction. Return false if we
5970 * should reclaim first.
5972 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
5974 unsigned long watermark
;
5976 /* Allocation can already succeed, nothing to do */
5977 if (zone_watermark_ok(zone
, sc
->order
, min_wmark_pages(zone
),
5978 sc
->reclaim_idx
, 0))
5981 /* Compaction cannot yet proceed. Do reclaim. */
5982 if (!compaction_suitable(zone
, sc
->order
, sc
->reclaim_idx
))
5986 * Compaction is already possible, but it takes time to run and there
5987 * are potentially other callers using the pages just freed. So proceed
5988 * with reclaim to make a buffer of free pages available to give
5989 * compaction a reasonable chance of completing and allocating the page.
5990 * Note that we won't actually reclaim the whole buffer in one attempt
5991 * as the target watermark in should_continue_reclaim() is lower. But if
5992 * we are already above the high+gap watermark, don't reclaim at all.
5994 watermark
= high_wmark_pages(zone
) + compact_gap(sc
->order
);
5996 return zone_watermark_ok_safe(zone
, 0, watermark
, sc
->reclaim_idx
);
5999 static void consider_reclaim_throttle(pg_data_t
*pgdat
, struct scan_control
*sc
)
6002 * If reclaim is making progress greater than 12% efficiency then
6003 * wake all the NOPROGRESS throttled tasks.
6005 if (sc
->nr_reclaimed
> (sc
->nr_scanned
>> 3)) {
6006 wait_queue_head_t
*wqh
;
6008 wqh
= &pgdat
->reclaim_wait
[VMSCAN_THROTTLE_NOPROGRESS
];
6009 if (waitqueue_active(wqh
))
6016 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6017 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6018 * under writeback and marked for immediate reclaim at the tail of the
6021 if (current_is_kswapd() || cgroup_reclaim(sc
))
6024 /* Throttle if making no progress at high prioities. */
6025 if (sc
->priority
== 1 && !sc
->nr_reclaimed
)
6026 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_NOPROGRESS
);
6030 * This is the direct reclaim path, for page-allocating processes. We only
6031 * try to reclaim pages from zones which will satisfy the caller's allocation
6034 * If a zone is deemed to be full of pinned pages then just give it a light
6035 * scan then give up on it.
6037 static void shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
6041 unsigned long nr_soft_reclaimed
;
6042 unsigned long nr_soft_scanned
;
6044 pg_data_t
*last_pgdat
= NULL
;
6045 pg_data_t
*first_pgdat
= NULL
;
6048 * If the number of buffer_heads in the machine exceeds the maximum
6049 * allowed level, force direct reclaim to scan the highmem zone as
6050 * highmem pages could be pinning lowmem pages storing buffer_heads
6052 orig_mask
= sc
->gfp_mask
;
6053 if (buffer_heads_over_limit
) {
6054 sc
->gfp_mask
|= __GFP_HIGHMEM
;
6055 sc
->reclaim_idx
= gfp_zone(sc
->gfp_mask
);
6058 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
6059 sc
->reclaim_idx
, sc
->nodemask
) {
6061 * Take care memory controller reclaiming has small influence
6064 if (!cgroup_reclaim(sc
)) {
6065 if (!cpuset_zone_allowed(zone
,
6066 GFP_KERNEL
| __GFP_HARDWALL
))
6070 * If we already have plenty of memory free for
6071 * compaction in this zone, don't free any more.
6072 * Even though compaction is invoked for any
6073 * non-zero order, only frequent costly order
6074 * reclamation is disruptive enough to become a
6075 * noticeable problem, like transparent huge
6078 if (IS_ENABLED(CONFIG_COMPACTION
) &&
6079 sc
->order
> PAGE_ALLOC_COSTLY_ORDER
&&
6080 compaction_ready(zone
, sc
)) {
6081 sc
->compaction_ready
= true;
6086 * Shrink each node in the zonelist once. If the
6087 * zonelist is ordered by zone (not the default) then a
6088 * node may be shrunk multiple times but in that case
6089 * the user prefers lower zones being preserved.
6091 if (zone
->zone_pgdat
== last_pgdat
)
6095 * This steals pages from memory cgroups over softlimit
6096 * and returns the number of reclaimed pages and
6097 * scanned pages. This works for global memory pressure
6098 * and balancing, not for a memcg's limit.
6100 nr_soft_scanned
= 0;
6101 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
->zone_pgdat
,
6102 sc
->order
, sc
->gfp_mask
,
6104 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
6105 sc
->nr_scanned
+= nr_soft_scanned
;
6106 /* need some check for avoid more shrink_zone() */
6110 first_pgdat
= zone
->zone_pgdat
;
6112 /* See comment about same check for global reclaim above */
6113 if (zone
->zone_pgdat
== last_pgdat
)
6115 last_pgdat
= zone
->zone_pgdat
;
6116 shrink_node(zone
->zone_pgdat
, sc
);
6120 consider_reclaim_throttle(first_pgdat
, sc
);
6123 * Restore to original mask to avoid the impact on the caller if we
6124 * promoted it to __GFP_HIGHMEM.
6126 sc
->gfp_mask
= orig_mask
;
6129 static void snapshot_refaults(struct mem_cgroup
*target_memcg
, pg_data_t
*pgdat
)
6131 struct lruvec
*target_lruvec
;
6132 unsigned long refaults
;
6134 if (lru_gen_enabled())
6137 target_lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
6138 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_ANON
);
6139 target_lruvec
->refaults
[WORKINGSET_ANON
] = refaults
;
6140 refaults
= lruvec_page_state(target_lruvec
, WORKINGSET_ACTIVATE_FILE
);
6141 target_lruvec
->refaults
[WORKINGSET_FILE
] = refaults
;
6145 * This is the main entry point to direct page reclaim.
6147 * If a full scan of the inactive list fails to free enough memory then we
6148 * are "out of memory" and something needs to be killed.
6150 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6151 * high - the zone may be full of dirty or under-writeback pages, which this
6152 * caller can't do much about. We kick the writeback threads and take explicit
6153 * naps in the hope that some of these pages can be written. But if the
6154 * allocating task holds filesystem locks which prevent writeout this might not
6155 * work, and the allocation attempt will fail.
6157 * returns: 0, if no pages reclaimed
6158 * else, the number of pages reclaimed
6160 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
6161 struct scan_control
*sc
)
6163 int initial_priority
= sc
->priority
;
6164 pg_data_t
*last_pgdat
;
6168 delayacct_freepages_start();
6170 if (!cgroup_reclaim(sc
))
6171 __count_zid_vm_events(ALLOCSTALL
, sc
->reclaim_idx
, 1);
6175 vmpressure_prio(sc
->gfp_mask
, sc
->target_mem_cgroup
,
6178 shrink_zones(zonelist
, sc
);
6180 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
6183 if (sc
->compaction_ready
)
6187 * If we're getting trouble reclaiming, start doing
6188 * writepage even in laptop mode.
6190 if (sc
->priority
< DEF_PRIORITY
- 2)
6191 sc
->may_writepage
= 1;
6192 } while (--sc
->priority
>= 0);
6195 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, sc
->reclaim_idx
,
6197 if (zone
->zone_pgdat
== last_pgdat
)
6199 last_pgdat
= zone
->zone_pgdat
;
6201 snapshot_refaults(sc
->target_mem_cgroup
, zone
->zone_pgdat
);
6203 if (cgroup_reclaim(sc
)) {
6204 struct lruvec
*lruvec
;
6206 lruvec
= mem_cgroup_lruvec(sc
->target_mem_cgroup
,
6208 clear_bit(LRUVEC_CGROUP_CONGESTED
, &lruvec
->flags
);
6212 delayacct_freepages_end();
6214 if (sc
->nr_reclaimed
)
6215 return sc
->nr_reclaimed
;
6217 /* Aborted reclaim to try compaction? don't OOM, then */
6218 if (sc
->compaction_ready
)
6222 * We make inactive:active ratio decisions based on the node's
6223 * composition of memory, but a restrictive reclaim_idx or a
6224 * memory.low cgroup setting can exempt large amounts of
6225 * memory from reclaim. Neither of which are very common, so
6226 * instead of doing costly eligibility calculations of the
6227 * entire cgroup subtree up front, we assume the estimates are
6228 * good, and retry with forcible deactivation if that fails.
6230 if (sc
->skipped_deactivate
) {
6231 sc
->priority
= initial_priority
;
6232 sc
->force_deactivate
= 1;
6233 sc
->skipped_deactivate
= 0;
6237 /* Untapped cgroup reserves? Don't OOM, retry. */
6238 if (sc
->memcg_low_skipped
) {
6239 sc
->priority
= initial_priority
;
6240 sc
->force_deactivate
= 0;
6241 sc
->memcg_low_reclaim
= 1;
6242 sc
->memcg_low_skipped
= 0;
6249 static bool allow_direct_reclaim(pg_data_t
*pgdat
)
6252 unsigned long pfmemalloc_reserve
= 0;
6253 unsigned long free_pages
= 0;
6257 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
6260 for (i
= 0; i
<= ZONE_NORMAL
; i
++) {
6261 zone
= &pgdat
->node_zones
[i
];
6262 if (!managed_zone(zone
))
6265 if (!zone_reclaimable_pages(zone
))
6268 pfmemalloc_reserve
+= min_wmark_pages(zone
);
6269 free_pages
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
6272 /* If there are no reserves (unexpected config) then do not throttle */
6273 if (!pfmemalloc_reserve
)
6276 wmark_ok
= free_pages
> pfmemalloc_reserve
/ 2;
6278 /* kswapd must be awake if processes are being throttled */
6279 if (!wmark_ok
&& waitqueue_active(&pgdat
->kswapd_wait
)) {
6280 if (READ_ONCE(pgdat
->kswapd_highest_zoneidx
) > ZONE_NORMAL
)
6281 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, ZONE_NORMAL
);
6283 wake_up_interruptible(&pgdat
->kswapd_wait
);
6290 * Throttle direct reclaimers if backing storage is backed by the network
6291 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6292 * depleted. kswapd will continue to make progress and wake the processes
6293 * when the low watermark is reached.
6295 * Returns true if a fatal signal was delivered during throttling. If this
6296 * happens, the page allocator should not consider triggering the OOM killer.
6298 static bool throttle_direct_reclaim(gfp_t gfp_mask
, struct zonelist
*zonelist
,
6299 nodemask_t
*nodemask
)
6303 pg_data_t
*pgdat
= NULL
;
6306 * Kernel threads should not be throttled as they may be indirectly
6307 * responsible for cleaning pages necessary for reclaim to make forward
6308 * progress. kjournald for example may enter direct reclaim while
6309 * committing a transaction where throttling it could forcing other
6310 * processes to block on log_wait_commit().
6312 if (current
->flags
& PF_KTHREAD
)
6316 * If a fatal signal is pending, this process should not throttle.
6317 * It should return quickly so it can exit and free its memory
6319 if (fatal_signal_pending(current
))
6323 * Check if the pfmemalloc reserves are ok by finding the first node
6324 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6325 * GFP_KERNEL will be required for allocating network buffers when
6326 * swapping over the network so ZONE_HIGHMEM is unusable.
6328 * Throttling is based on the first usable node and throttled processes
6329 * wait on a queue until kswapd makes progress and wakes them. There
6330 * is an affinity then between processes waking up and where reclaim
6331 * progress has been made assuming the process wakes on the same node.
6332 * More importantly, processes running on remote nodes will not compete
6333 * for remote pfmemalloc reserves and processes on different nodes
6334 * should make reasonable progress.
6336 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
6337 gfp_zone(gfp_mask
), nodemask
) {
6338 if (zone_idx(zone
) > ZONE_NORMAL
)
6341 /* Throttle based on the first usable node */
6342 pgdat
= zone
->zone_pgdat
;
6343 if (allow_direct_reclaim(pgdat
))
6348 /* If no zone was usable by the allocation flags then do not throttle */
6352 /* Account for the throttling */
6353 count_vm_event(PGSCAN_DIRECT_THROTTLE
);
6356 * If the caller cannot enter the filesystem, it's possible that it
6357 * is due to the caller holding an FS lock or performing a journal
6358 * transaction in the case of a filesystem like ext[3|4]. In this case,
6359 * it is not safe to block on pfmemalloc_wait as kswapd could be
6360 * blocked waiting on the same lock. Instead, throttle for up to a
6361 * second before continuing.
6363 if (!(gfp_mask
& __GFP_FS
))
6364 wait_event_interruptible_timeout(pgdat
->pfmemalloc_wait
,
6365 allow_direct_reclaim(pgdat
), HZ
);
6367 /* Throttle until kswapd wakes the process */
6368 wait_event_killable(zone
->zone_pgdat
->pfmemalloc_wait
,
6369 allow_direct_reclaim(pgdat
));
6371 if (fatal_signal_pending(current
))
6378 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
6379 gfp_t gfp_mask
, nodemask_t
*nodemask
)
6381 unsigned long nr_reclaimed
;
6382 struct scan_control sc
= {
6383 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
6384 .gfp_mask
= current_gfp_context(gfp_mask
),
6385 .reclaim_idx
= gfp_zone(gfp_mask
),
6387 .nodemask
= nodemask
,
6388 .priority
= DEF_PRIORITY
,
6389 .may_writepage
= !laptop_mode
,
6395 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6396 * Confirm they are large enough for max values.
6398 BUILD_BUG_ON(MAX_ORDER
>= S8_MAX
);
6399 BUILD_BUG_ON(DEF_PRIORITY
> S8_MAX
);
6400 BUILD_BUG_ON(MAX_NR_ZONES
> S8_MAX
);
6403 * Do not enter reclaim if fatal signal was delivered while throttled.
6404 * 1 is returned so that the page allocator does not OOM kill at this
6407 if (throttle_direct_reclaim(sc
.gfp_mask
, zonelist
, nodemask
))
6410 set_task_reclaim_state(current
, &sc
.reclaim_state
);
6411 trace_mm_vmscan_direct_reclaim_begin(order
, sc
.gfp_mask
);
6413 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
6415 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
6416 set_task_reclaim_state(current
, NULL
);
6418 return nr_reclaimed
;
6423 /* Only used by soft limit reclaim. Do not reuse for anything else. */
6424 unsigned long mem_cgroup_shrink_node(struct mem_cgroup
*memcg
,
6425 gfp_t gfp_mask
, bool noswap
,
6427 unsigned long *nr_scanned
)
6429 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
6430 struct scan_control sc
= {
6431 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
6432 .target_mem_cgroup
= memcg
,
6433 .may_writepage
= !laptop_mode
,
6435 .reclaim_idx
= MAX_NR_ZONES
- 1,
6436 .may_swap
= !noswap
,
6439 WARN_ON_ONCE(!current
->reclaim_state
);
6441 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
6442 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
6444 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
6448 * NOTE: Although we can get the priority field, using it
6449 * here is not a good idea, since it limits the pages we can scan.
6450 * if we don't reclaim here, the shrink_node from balance_pgdat
6451 * will pick up pages from other mem cgroup's as well. We hack
6452 * the priority and make it zero.
6454 shrink_lruvec(lruvec
, &sc
);
6456 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
6458 *nr_scanned
= sc
.nr_scanned
;
6460 return sc
.nr_reclaimed
;
6463 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
6464 unsigned long nr_pages
,
6466 unsigned int reclaim_options
)
6468 unsigned long nr_reclaimed
;
6469 unsigned int noreclaim_flag
;
6470 struct scan_control sc
= {
6471 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
6472 .gfp_mask
= (current_gfp_context(gfp_mask
) & GFP_RECLAIM_MASK
) |
6473 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
6474 .reclaim_idx
= MAX_NR_ZONES
- 1,
6475 .target_mem_cgroup
= memcg
,
6476 .priority
= DEF_PRIORITY
,
6477 .may_writepage
= !laptop_mode
,
6479 .may_swap
= !!(reclaim_options
& MEMCG_RECLAIM_MAY_SWAP
),
6480 .proactive
= !!(reclaim_options
& MEMCG_RECLAIM_PROACTIVE
),
6483 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6484 * equal pressure on all the nodes. This is based on the assumption that
6485 * the reclaim does not bail out early.
6487 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
6489 set_task_reclaim_state(current
, &sc
.reclaim_state
);
6490 trace_mm_vmscan_memcg_reclaim_begin(0, sc
.gfp_mask
);
6491 noreclaim_flag
= memalloc_noreclaim_save();
6493 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
6495 memalloc_noreclaim_restore(noreclaim_flag
);
6496 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
6497 set_task_reclaim_state(current
, NULL
);
6499 return nr_reclaimed
;
6503 static void kswapd_age_node(struct pglist_data
*pgdat
, struct scan_control
*sc
)
6505 struct mem_cgroup
*memcg
;
6506 struct lruvec
*lruvec
;
6508 if (lru_gen_enabled()) {
6509 lru_gen_age_node(pgdat
, sc
);
6513 if (!can_age_anon_pages(pgdat
, sc
))
6516 lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
6517 if (!inactive_is_low(lruvec
, LRU_INACTIVE_ANON
))
6520 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
6522 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
6523 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
6524 sc
, LRU_ACTIVE_ANON
);
6525 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
6529 static bool pgdat_watermark_boosted(pg_data_t
*pgdat
, int highest_zoneidx
)
6535 * Check for watermark boosts top-down as the higher zones
6536 * are more likely to be boosted. Both watermarks and boosts
6537 * should not be checked at the same time as reclaim would
6538 * start prematurely when there is no boosting and a lower
6541 for (i
= highest_zoneidx
; i
>= 0; i
--) {
6542 zone
= pgdat
->node_zones
+ i
;
6543 if (!managed_zone(zone
))
6546 if (zone
->watermark_boost
)
6554 * Returns true if there is an eligible zone balanced for the request order
6555 * and highest_zoneidx
6557 static bool pgdat_balanced(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
6560 unsigned long mark
= -1;
6564 * Check watermarks bottom-up as lower zones are more likely to
6567 for (i
= 0; i
<= highest_zoneidx
; i
++) {
6568 zone
= pgdat
->node_zones
+ i
;
6570 if (!managed_zone(zone
))
6573 if (sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
)
6574 mark
= wmark_pages(zone
, WMARK_PROMO
);
6576 mark
= high_wmark_pages(zone
);
6577 if (zone_watermark_ok_safe(zone
, order
, mark
, highest_zoneidx
))
6582 * If a node has no managed zone within highest_zoneidx, it does not
6583 * need balancing by definition. This can happen if a zone-restricted
6584 * allocation tries to wake a remote kswapd.
6592 /* Clear pgdat state for congested, dirty or under writeback. */
6593 static void clear_pgdat_congested(pg_data_t
*pgdat
)
6595 struct lruvec
*lruvec
= mem_cgroup_lruvec(NULL
, pgdat
);
6597 clear_bit(LRUVEC_NODE_CONGESTED
, &lruvec
->flags
);
6598 clear_bit(LRUVEC_CGROUP_CONGESTED
, &lruvec
->flags
);
6599 clear_bit(PGDAT_DIRTY
, &pgdat
->flags
);
6600 clear_bit(PGDAT_WRITEBACK
, &pgdat
->flags
);
6604 * Prepare kswapd for sleeping. This verifies that there are no processes
6605 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6607 * Returns true if kswapd is ready to sleep
6609 static bool prepare_kswapd_sleep(pg_data_t
*pgdat
, int order
,
6610 int highest_zoneidx
)
6613 * The throttled processes are normally woken up in balance_pgdat() as
6614 * soon as allow_direct_reclaim() is true. But there is a potential
6615 * race between when kswapd checks the watermarks and a process gets
6616 * throttled. There is also a potential race if processes get
6617 * throttled, kswapd wakes, a large process exits thereby balancing the
6618 * zones, which causes kswapd to exit balance_pgdat() before reaching
6619 * the wake up checks. If kswapd is going to sleep, no process should
6620 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6621 * the wake up is premature, processes will wake kswapd and get
6622 * throttled again. The difference from wake ups in balance_pgdat() is
6623 * that here we are under prepare_to_wait().
6625 if (waitqueue_active(&pgdat
->pfmemalloc_wait
))
6626 wake_up_all(&pgdat
->pfmemalloc_wait
);
6628 /* Hopeless node, leave it to direct reclaim */
6629 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
)
6632 if (pgdat_balanced(pgdat
, order
, highest_zoneidx
)) {
6633 clear_pgdat_congested(pgdat
);
6641 * kswapd shrinks a node of pages that are at or below the highest usable
6642 * zone that is currently unbalanced.
6644 * Returns true if kswapd scanned at least the requested number of pages to
6645 * reclaim or if the lack of progress was due to pages under writeback.
6646 * This is used to determine if the scanning priority needs to be raised.
6648 static bool kswapd_shrink_node(pg_data_t
*pgdat
,
6649 struct scan_control
*sc
)
6654 /* Reclaim a number of pages proportional to the number of zones */
6655 sc
->nr_to_reclaim
= 0;
6656 for (z
= 0; z
<= sc
->reclaim_idx
; z
++) {
6657 zone
= pgdat
->node_zones
+ z
;
6658 if (!managed_zone(zone
))
6661 sc
->nr_to_reclaim
+= max(high_wmark_pages(zone
), SWAP_CLUSTER_MAX
);
6665 * Historically care was taken to put equal pressure on all zones but
6666 * now pressure is applied based on node LRU order.
6668 shrink_node(pgdat
, sc
);
6671 * Fragmentation may mean that the system cannot be rebalanced for
6672 * high-order allocations. If twice the allocation size has been
6673 * reclaimed then recheck watermarks only at order-0 to prevent
6674 * excessive reclaim. Assume that a process requested a high-order
6675 * can direct reclaim/compact.
6677 if (sc
->order
&& sc
->nr_reclaimed
>= compact_gap(sc
->order
))
6680 return sc
->nr_scanned
>= sc
->nr_to_reclaim
;
6683 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6685 update_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
, bool active
)
6690 for (i
= 0; i
<= highest_zoneidx
; i
++) {
6691 zone
= pgdat
->node_zones
+ i
;
6693 if (!managed_zone(zone
))
6697 set_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
6699 clear_bit(ZONE_RECLAIM_ACTIVE
, &zone
->flags
);
6704 set_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
6706 update_reclaim_active(pgdat
, highest_zoneidx
, true);
6710 clear_reclaim_active(pg_data_t
*pgdat
, int highest_zoneidx
)
6712 update_reclaim_active(pgdat
, highest_zoneidx
, false);
6716 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6717 * that are eligible for use by the caller until at least one zone is
6720 * Returns the order kswapd finished reclaiming at.
6722 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6723 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6724 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6725 * or lower is eligible for reclaim until at least one usable zone is
6728 static int balance_pgdat(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
6731 unsigned long nr_soft_reclaimed
;
6732 unsigned long nr_soft_scanned
;
6733 unsigned long pflags
;
6734 unsigned long nr_boost_reclaim
;
6735 unsigned long zone_boosts
[MAX_NR_ZONES
] = { 0, };
6738 struct scan_control sc
= {
6739 .gfp_mask
= GFP_KERNEL
,
6744 set_task_reclaim_state(current
, &sc
.reclaim_state
);
6745 psi_memstall_enter(&pflags
);
6746 __fs_reclaim_acquire(_THIS_IP_
);
6748 count_vm_event(PAGEOUTRUN
);
6751 * Account for the reclaim boost. Note that the zone boost is left in
6752 * place so that parallel allocations that are near the watermark will
6753 * stall or direct reclaim until kswapd is finished.
6755 nr_boost_reclaim
= 0;
6756 for (i
= 0; i
<= highest_zoneidx
; i
++) {
6757 zone
= pgdat
->node_zones
+ i
;
6758 if (!managed_zone(zone
))
6761 nr_boost_reclaim
+= zone
->watermark_boost
;
6762 zone_boosts
[i
] = zone
->watermark_boost
;
6764 boosted
= nr_boost_reclaim
;
6767 set_reclaim_active(pgdat
, highest_zoneidx
);
6768 sc
.priority
= DEF_PRIORITY
;
6770 unsigned long nr_reclaimed
= sc
.nr_reclaimed
;
6771 bool raise_priority
= true;
6775 sc
.reclaim_idx
= highest_zoneidx
;
6778 * If the number of buffer_heads exceeds the maximum allowed
6779 * then consider reclaiming from all zones. This has a dual
6780 * purpose -- on 64-bit systems it is expected that
6781 * buffer_heads are stripped during active rotation. On 32-bit
6782 * systems, highmem pages can pin lowmem memory and shrinking
6783 * buffers can relieve lowmem pressure. Reclaim may still not
6784 * go ahead if all eligible zones for the original allocation
6785 * request are balanced to avoid excessive reclaim from kswapd.
6787 if (buffer_heads_over_limit
) {
6788 for (i
= MAX_NR_ZONES
- 1; i
>= 0; i
--) {
6789 zone
= pgdat
->node_zones
+ i
;
6790 if (!managed_zone(zone
))
6799 * If the pgdat is imbalanced then ignore boosting and preserve
6800 * the watermarks for a later time and restart. Note that the
6801 * zone watermarks will be still reset at the end of balancing
6802 * on the grounds that the normal reclaim should be enough to
6803 * re-evaluate if boosting is required when kswapd next wakes.
6805 balanced
= pgdat_balanced(pgdat
, sc
.order
, highest_zoneidx
);
6806 if (!balanced
&& nr_boost_reclaim
) {
6807 nr_boost_reclaim
= 0;
6812 * If boosting is not active then only reclaim if there are no
6813 * eligible zones. Note that sc.reclaim_idx is not used as
6814 * buffer_heads_over_limit may have adjusted it.
6816 if (!nr_boost_reclaim
&& balanced
)
6819 /* Limit the priority of boosting to avoid reclaim writeback */
6820 if (nr_boost_reclaim
&& sc
.priority
== DEF_PRIORITY
- 2)
6821 raise_priority
= false;
6824 * Do not writeback or swap pages for boosted reclaim. The
6825 * intent is to relieve pressure not issue sub-optimal IO
6826 * from reclaim context. If no pages are reclaimed, the
6827 * reclaim will be aborted.
6829 sc
.may_writepage
= !laptop_mode
&& !nr_boost_reclaim
;
6830 sc
.may_swap
= !nr_boost_reclaim
;
6833 * Do some background aging, to give pages a chance to be
6834 * referenced before reclaiming. All pages are rotated
6835 * regardless of classzone as this is about consistent aging.
6837 kswapd_age_node(pgdat
, &sc
);
6840 * If we're getting trouble reclaiming, start doing writepage
6841 * even in laptop mode.
6843 if (sc
.priority
< DEF_PRIORITY
- 2)
6844 sc
.may_writepage
= 1;
6846 /* Call soft limit reclaim before calling shrink_node. */
6848 nr_soft_scanned
= 0;
6849 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(pgdat
, sc
.order
,
6850 sc
.gfp_mask
, &nr_soft_scanned
);
6851 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
6854 * There should be no need to raise the scanning priority if
6855 * enough pages are already being scanned that that high
6856 * watermark would be met at 100% efficiency.
6858 if (kswapd_shrink_node(pgdat
, &sc
))
6859 raise_priority
= false;
6862 * If the low watermark is met there is no need for processes
6863 * to be throttled on pfmemalloc_wait as they should not be
6864 * able to safely make forward progress. Wake them
6866 if (waitqueue_active(&pgdat
->pfmemalloc_wait
) &&
6867 allow_direct_reclaim(pgdat
))
6868 wake_up_all(&pgdat
->pfmemalloc_wait
);
6870 /* Check if kswapd should be suspending */
6871 __fs_reclaim_release(_THIS_IP_
);
6872 ret
= try_to_freeze();
6873 __fs_reclaim_acquire(_THIS_IP_
);
6874 if (ret
|| kthread_should_stop())
6878 * Raise priority if scanning rate is too low or there was no
6879 * progress in reclaiming pages
6881 nr_reclaimed
= sc
.nr_reclaimed
- nr_reclaimed
;
6882 nr_boost_reclaim
-= min(nr_boost_reclaim
, nr_reclaimed
);
6885 * If reclaim made no progress for a boost, stop reclaim as
6886 * IO cannot be queued and it could be an infinite loop in
6887 * extreme circumstances.
6889 if (nr_boost_reclaim
&& !nr_reclaimed
)
6892 if (raise_priority
|| !nr_reclaimed
)
6894 } while (sc
.priority
>= 1);
6896 if (!sc
.nr_reclaimed
)
6897 pgdat
->kswapd_failures
++;
6900 clear_reclaim_active(pgdat
, highest_zoneidx
);
6902 /* If reclaim was boosted, account for the reclaim done in this pass */
6904 unsigned long flags
;
6906 for (i
= 0; i
<= highest_zoneidx
; i
++) {
6907 if (!zone_boosts
[i
])
6910 /* Increments are under the zone lock */
6911 zone
= pgdat
->node_zones
+ i
;
6912 spin_lock_irqsave(&zone
->lock
, flags
);
6913 zone
->watermark_boost
-= min(zone
->watermark_boost
, zone_boosts
[i
]);
6914 spin_unlock_irqrestore(&zone
->lock
, flags
);
6918 * As there is now likely space, wakeup kcompact to defragment
6921 wakeup_kcompactd(pgdat
, pageblock_order
, highest_zoneidx
);
6924 snapshot_refaults(NULL
, pgdat
);
6925 __fs_reclaim_release(_THIS_IP_
);
6926 psi_memstall_leave(&pflags
);
6927 set_task_reclaim_state(current
, NULL
);
6930 * Return the order kswapd stopped reclaiming at as
6931 * prepare_kswapd_sleep() takes it into account. If another caller
6932 * entered the allocator slow path while kswapd was awake, order will
6933 * remain at the higher level.
6939 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6940 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6941 * not a valid index then either kswapd runs for first time or kswapd couldn't
6942 * sleep after previous reclaim attempt (node is still unbalanced). In that
6943 * case return the zone index of the previous kswapd reclaim cycle.
6945 static enum zone_type
kswapd_highest_zoneidx(pg_data_t
*pgdat
,
6946 enum zone_type prev_highest_zoneidx
)
6948 enum zone_type curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
6950 return curr_idx
== MAX_NR_ZONES
? prev_highest_zoneidx
: curr_idx
;
6953 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int alloc_order
, int reclaim_order
,
6954 unsigned int highest_zoneidx
)
6959 if (freezing(current
) || kthread_should_stop())
6962 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
6965 * Try to sleep for a short interval. Note that kcompactd will only be
6966 * woken if it is possible to sleep for a short interval. This is
6967 * deliberate on the assumption that if reclaim cannot keep an
6968 * eligible zone balanced that it's also unlikely that compaction will
6971 if (prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
6973 * Compaction records what page blocks it recently failed to
6974 * isolate pages from and skips them in the future scanning.
6975 * When kswapd is going to sleep, it is reasonable to assume
6976 * that pages and compaction may succeed so reset the cache.
6978 reset_isolation_suitable(pgdat
);
6981 * We have freed the memory, now we should compact it to make
6982 * allocation of the requested order possible.
6984 wakeup_kcompactd(pgdat
, alloc_order
, highest_zoneidx
);
6986 remaining
= schedule_timeout(HZ
/10);
6989 * If woken prematurely then reset kswapd_highest_zoneidx and
6990 * order. The values will either be from a wakeup request or
6991 * the previous request that slept prematurely.
6994 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
,
6995 kswapd_highest_zoneidx(pgdat
,
6998 if (READ_ONCE(pgdat
->kswapd_order
) < reclaim_order
)
6999 WRITE_ONCE(pgdat
->kswapd_order
, reclaim_order
);
7002 finish_wait(&pgdat
->kswapd_wait
, &wait
);
7003 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
7007 * After a short sleep, check if it was a premature sleep. If not, then
7008 * go fully to sleep until explicitly woken up.
7011 prepare_kswapd_sleep(pgdat
, reclaim_order
, highest_zoneidx
)) {
7012 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
7015 * vmstat counters are not perfectly accurate and the estimated
7016 * value for counters such as NR_FREE_PAGES can deviate from the
7017 * true value by nr_online_cpus * threshold. To avoid the zone
7018 * watermarks being breached while under pressure, we reduce the
7019 * per-cpu vmstat threshold while kswapd is awake and restore
7020 * them before going back to sleep.
7022 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
7024 if (!kthread_should_stop())
7027 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
7030 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
7032 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
7034 finish_wait(&pgdat
->kswapd_wait
, &wait
);
7038 * The background pageout daemon, started as a kernel thread
7039 * from the init process.
7041 * This basically trickles out pages so that we have _some_
7042 * free memory available even if there is no other activity
7043 * that frees anything up. This is needed for things like routing
7044 * etc, where we otherwise might have all activity going on in
7045 * asynchronous contexts that cannot page things out.
7047 * If there are applications that are active memory-allocators
7048 * (most normal use), this basically shouldn't matter.
7050 static int kswapd(void *p
)
7052 unsigned int alloc_order
, reclaim_order
;
7053 unsigned int highest_zoneidx
= MAX_NR_ZONES
- 1;
7054 pg_data_t
*pgdat
= (pg_data_t
*)p
;
7055 struct task_struct
*tsk
= current
;
7056 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
7058 if (!cpumask_empty(cpumask
))
7059 set_cpus_allowed_ptr(tsk
, cpumask
);
7062 * Tell the memory management that we're a "memory allocator",
7063 * and that if we need more memory we should get access to it
7064 * regardless (see "__alloc_pages()"). "kswapd" should
7065 * never get caught in the normal page freeing logic.
7067 * (Kswapd normally doesn't need memory anyway, but sometimes
7068 * you need a small amount of memory in order to be able to
7069 * page out something else, and this flag essentially protects
7070 * us from recursively trying to free more memory as we're
7071 * trying to free the first piece of memory in the first place).
7073 tsk
->flags
|= PF_MEMALLOC
| PF_KSWAPD
;
7076 WRITE_ONCE(pgdat
->kswapd_order
, 0);
7077 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
7078 atomic_set(&pgdat
->nr_writeback_throttled
, 0);
7082 alloc_order
= reclaim_order
= READ_ONCE(pgdat
->kswapd_order
);
7083 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
7087 kswapd_try_to_sleep(pgdat
, alloc_order
, reclaim_order
,
7090 /* Read the new order and highest_zoneidx */
7091 alloc_order
= READ_ONCE(pgdat
->kswapd_order
);
7092 highest_zoneidx
= kswapd_highest_zoneidx(pgdat
,
7094 WRITE_ONCE(pgdat
->kswapd_order
, 0);
7095 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, MAX_NR_ZONES
);
7097 ret
= try_to_freeze();
7098 if (kthread_should_stop())
7102 * We can speed up thawing tasks if we don't call balance_pgdat
7103 * after returning from the refrigerator
7109 * Reclaim begins at the requested order but if a high-order
7110 * reclaim fails then kswapd falls back to reclaiming for
7111 * order-0. If that happens, kswapd will consider sleeping
7112 * for the order it finished reclaiming at (reclaim_order)
7113 * but kcompactd is woken to compact for the original
7114 * request (alloc_order).
7116 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, highest_zoneidx
,
7118 reclaim_order
= balance_pgdat(pgdat
, alloc_order
,
7120 if (reclaim_order
< alloc_order
)
7121 goto kswapd_try_sleep
;
7124 tsk
->flags
&= ~(PF_MEMALLOC
| PF_KSWAPD
);
7130 * A zone is low on free memory or too fragmented for high-order memory. If
7131 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7132 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7133 * has failed or is not needed, still wake up kcompactd if only compaction is
7136 void wakeup_kswapd(struct zone
*zone
, gfp_t gfp_flags
, int order
,
7137 enum zone_type highest_zoneidx
)
7140 enum zone_type curr_idx
;
7142 if (!managed_zone(zone
))
7145 if (!cpuset_zone_allowed(zone
, gfp_flags
))
7148 pgdat
= zone
->zone_pgdat
;
7149 curr_idx
= READ_ONCE(pgdat
->kswapd_highest_zoneidx
);
7151 if (curr_idx
== MAX_NR_ZONES
|| curr_idx
< highest_zoneidx
)
7152 WRITE_ONCE(pgdat
->kswapd_highest_zoneidx
, highest_zoneidx
);
7154 if (READ_ONCE(pgdat
->kswapd_order
) < order
)
7155 WRITE_ONCE(pgdat
->kswapd_order
, order
);
7157 if (!waitqueue_active(&pgdat
->kswapd_wait
))
7160 /* Hopeless node, leave it to direct reclaim if possible */
7161 if (pgdat
->kswapd_failures
>= MAX_RECLAIM_RETRIES
||
7162 (pgdat_balanced(pgdat
, order
, highest_zoneidx
) &&
7163 !pgdat_watermark_boosted(pgdat
, highest_zoneidx
))) {
7165 * There may be plenty of free memory available, but it's too
7166 * fragmented for high-order allocations. Wake up kcompactd
7167 * and rely on compaction_suitable() to determine if it's
7168 * needed. If it fails, it will defer subsequent attempts to
7169 * ratelimit its work.
7171 if (!(gfp_flags
& __GFP_DIRECT_RECLAIM
))
7172 wakeup_kcompactd(pgdat
, order
, highest_zoneidx
);
7176 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, highest_zoneidx
, order
,
7178 wake_up_interruptible(&pgdat
->kswapd_wait
);
7181 #ifdef CONFIG_HIBERNATION
7183 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7186 * Rather than trying to age LRUs the aim is to preserve the overall
7187 * LRU order by reclaiming preferentially
7188 * inactive > active > active referenced > active mapped
7190 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
7192 struct scan_control sc
= {
7193 .nr_to_reclaim
= nr_to_reclaim
,
7194 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
7195 .reclaim_idx
= MAX_NR_ZONES
- 1,
7196 .priority
= DEF_PRIORITY
,
7200 .hibernation_mode
= 1,
7202 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
7203 unsigned long nr_reclaimed
;
7204 unsigned int noreclaim_flag
;
7206 fs_reclaim_acquire(sc
.gfp_mask
);
7207 noreclaim_flag
= memalloc_noreclaim_save();
7208 set_task_reclaim_state(current
, &sc
.reclaim_state
);
7210 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
7212 set_task_reclaim_state(current
, NULL
);
7213 memalloc_noreclaim_restore(noreclaim_flag
);
7214 fs_reclaim_release(sc
.gfp_mask
);
7216 return nr_reclaimed
;
7218 #endif /* CONFIG_HIBERNATION */
7221 * This kswapd start function will be called by init and node-hot-add.
7223 void __meminit
kswapd_run(int nid
)
7225 pg_data_t
*pgdat
= NODE_DATA(nid
);
7227 pgdat_kswapd_lock(pgdat
);
7228 if (!pgdat
->kswapd
) {
7229 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
7230 if (IS_ERR(pgdat
->kswapd
)) {
7231 /* failure at boot is fatal */
7232 pr_err("Failed to start kswapd on node %d,ret=%ld\n",
7233 nid
, PTR_ERR(pgdat
->kswapd
));
7234 BUG_ON(system_state
< SYSTEM_RUNNING
);
7235 pgdat
->kswapd
= NULL
;
7238 pgdat_kswapd_unlock(pgdat
);
7242 * Called by memory hotplug when all memory in a node is offlined. Caller must
7243 * be holding mem_hotplug_begin/done().
7245 void __meminit
kswapd_stop(int nid
)
7247 pg_data_t
*pgdat
= NODE_DATA(nid
);
7248 struct task_struct
*kswapd
;
7250 pgdat_kswapd_lock(pgdat
);
7251 kswapd
= pgdat
->kswapd
;
7253 kthread_stop(kswapd
);
7254 pgdat
->kswapd
= NULL
;
7256 pgdat_kswapd_unlock(pgdat
);
7259 static int __init
kswapd_init(void)
7264 for_each_node_state(nid
, N_MEMORY
)
7269 module_init(kswapd_init
)
7275 * If non-zero call node_reclaim when the number of free pages falls below
7278 int node_reclaim_mode __read_mostly
;
7281 * Priority for NODE_RECLAIM. This determines the fraction of pages
7282 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7285 #define NODE_RECLAIM_PRIORITY 4
7288 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7291 int sysctl_min_unmapped_ratio
= 1;
7294 * If the number of slab pages in a zone grows beyond this percentage then
7295 * slab reclaim needs to occur.
7297 int sysctl_min_slab_ratio
= 5;
7299 static inline unsigned long node_unmapped_file_pages(struct pglist_data
*pgdat
)
7301 unsigned long file_mapped
= node_page_state(pgdat
, NR_FILE_MAPPED
);
7302 unsigned long file_lru
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
7303 node_page_state(pgdat
, NR_ACTIVE_FILE
);
7306 * It's possible for there to be more file mapped pages than
7307 * accounted for by the pages on the file LRU lists because
7308 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7310 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
7313 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
7314 static unsigned long node_pagecache_reclaimable(struct pglist_data
*pgdat
)
7316 unsigned long nr_pagecache_reclaimable
;
7317 unsigned long delta
= 0;
7320 * If RECLAIM_UNMAP is set, then all file pages are considered
7321 * potentially reclaimable. Otherwise, we have to worry about
7322 * pages like swapcache and node_unmapped_file_pages() provides
7325 if (node_reclaim_mode
& RECLAIM_UNMAP
)
7326 nr_pagecache_reclaimable
= node_page_state(pgdat
, NR_FILE_PAGES
);
7328 nr_pagecache_reclaimable
= node_unmapped_file_pages(pgdat
);
7330 /* If we can't clean pages, remove dirty pages from consideration */
7331 if (!(node_reclaim_mode
& RECLAIM_WRITE
))
7332 delta
+= node_page_state(pgdat
, NR_FILE_DIRTY
);
7334 /* Watch for any possible underflows due to delta */
7335 if (unlikely(delta
> nr_pagecache_reclaimable
))
7336 delta
= nr_pagecache_reclaimable
;
7338 return nr_pagecache_reclaimable
- delta
;
7342 * Try to free up some pages from this node through reclaim.
7344 static int __node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
7346 /* Minimum pages needed in order to stay on node */
7347 const unsigned long nr_pages
= 1 << order
;
7348 struct task_struct
*p
= current
;
7349 unsigned int noreclaim_flag
;
7350 struct scan_control sc
= {
7351 .nr_to_reclaim
= max(nr_pages
, SWAP_CLUSTER_MAX
),
7352 .gfp_mask
= current_gfp_context(gfp_mask
),
7354 .priority
= NODE_RECLAIM_PRIORITY
,
7355 .may_writepage
= !!(node_reclaim_mode
& RECLAIM_WRITE
),
7356 .may_unmap
= !!(node_reclaim_mode
& RECLAIM_UNMAP
),
7358 .reclaim_idx
= gfp_zone(gfp_mask
),
7360 unsigned long pflags
;
7362 trace_mm_vmscan_node_reclaim_begin(pgdat
->node_id
, order
,
7366 psi_memstall_enter(&pflags
);
7367 delayacct_freepages_start();
7368 fs_reclaim_acquire(sc
.gfp_mask
);
7370 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7372 noreclaim_flag
= memalloc_noreclaim_save();
7373 set_task_reclaim_state(p
, &sc
.reclaim_state
);
7375 if (node_pagecache_reclaimable(pgdat
) > pgdat
->min_unmapped_pages
||
7376 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) > pgdat
->min_slab_pages
) {
7378 * Free memory by calling shrink node with increasing
7379 * priorities until we have enough memory freed.
7382 shrink_node(pgdat
, &sc
);
7383 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
7386 set_task_reclaim_state(p
, NULL
);
7387 memalloc_noreclaim_restore(noreclaim_flag
);
7388 fs_reclaim_release(sc
.gfp_mask
);
7389 psi_memstall_leave(&pflags
);
7390 delayacct_freepages_end();
7392 trace_mm_vmscan_node_reclaim_end(sc
.nr_reclaimed
);
7394 return sc
.nr_reclaimed
>= nr_pages
;
7397 int node_reclaim(struct pglist_data
*pgdat
, gfp_t gfp_mask
, unsigned int order
)
7402 * Node reclaim reclaims unmapped file backed pages and
7403 * slab pages if we are over the defined limits.
7405 * A small portion of unmapped file backed pages is needed for
7406 * file I/O otherwise pages read by file I/O will be immediately
7407 * thrown out if the node is overallocated. So we do not reclaim
7408 * if less than a specified percentage of the node is used by
7409 * unmapped file backed pages.
7411 if (node_pagecache_reclaimable(pgdat
) <= pgdat
->min_unmapped_pages
&&
7412 node_page_state_pages(pgdat
, NR_SLAB_RECLAIMABLE_B
) <=
7413 pgdat
->min_slab_pages
)
7414 return NODE_RECLAIM_FULL
;
7417 * Do not scan if the allocation should not be delayed.
7419 if (!gfpflags_allow_blocking(gfp_mask
) || (current
->flags
& PF_MEMALLOC
))
7420 return NODE_RECLAIM_NOSCAN
;
7423 * Only run node reclaim on the local node or on nodes that do not
7424 * have associated processors. This will favor the local processor
7425 * over remote processors and spread off node memory allocations
7426 * as wide as possible.
7428 if (node_state(pgdat
->node_id
, N_CPU
) && pgdat
->node_id
!= numa_node_id())
7429 return NODE_RECLAIM_NOSCAN
;
7431 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
))
7432 return NODE_RECLAIM_NOSCAN
;
7434 ret
= __node_reclaim(pgdat
, gfp_mask
, order
);
7435 clear_bit(PGDAT_RECLAIM_LOCKED
, &pgdat
->flags
);
7438 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
7445 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7447 * @fbatch: Batch of lru folios to check.
7449 * Checks folios for evictability, if an evictable folio is in the unevictable
7450 * lru list, moves it to the appropriate evictable lru list. This function
7451 * should be only used for lru folios.
7453 void check_move_unevictable_folios(struct folio_batch
*fbatch
)
7455 struct lruvec
*lruvec
= NULL
;
7460 for (i
= 0; i
< fbatch
->nr
; i
++) {
7461 struct folio
*folio
= fbatch
->folios
[i
];
7462 int nr_pages
= folio_nr_pages(folio
);
7464 pgscanned
+= nr_pages
;
7466 /* block memcg migration while the folio moves between lrus */
7467 if (!folio_test_clear_lru(folio
))
7470 lruvec
= folio_lruvec_relock_irq(folio
, lruvec
);
7471 if (folio_evictable(folio
) && folio_test_unevictable(folio
)) {
7472 lruvec_del_folio(lruvec
, folio
);
7473 folio_clear_unevictable(folio
);
7474 lruvec_add_folio(lruvec
, folio
);
7475 pgrescued
+= nr_pages
;
7477 folio_set_lru(folio
);
7481 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
7482 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
7483 unlock_page_lruvec_irq(lruvec
);
7484 } else if (pgscanned
) {
7485 count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
7488 EXPORT_SYMBOL_GPL(check_move_unevictable_folios
);