1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
8 #include <linux/memcontrol.h>
9 #include <linux/mm_inline.h>
10 #include <linux/writeback.h>
11 #include <linux/shmem_fs.h>
12 #include <linux/pagemap.h>
13 #include <linux/atomic.h>
14 #include <linux/module.h>
15 #include <linux/swap.h>
16 #include <linux/dax.h>
23 * Per node, two clock lists are maintained for file pages: the
24 * inactive and the active list. Freshly faulted pages start out at
25 * the head of the inactive list and page reclaim scans pages from the
26 * tail. Pages that are accessed multiple times on the inactive list
27 * are promoted to the active list, to protect them from reclaim,
28 * whereas active pages are demoted to the inactive list when the
29 * active list grows too big.
31 * fault ------------------------+
33 * +--------------+ | +-------------+
34 * reclaim <- | inactive | <-+-- demotion | active | <--+
35 * +--------------+ +-------------+ |
37 * +-------------- promotion ------------------+
40 * Access frequency and refault distance
42 * A workload is thrashing when its pages are frequently used but they
43 * are evicted from the inactive list every time before another access
44 * would have promoted them to the active list.
46 * In cases where the average access distance between thrashing pages
47 * is bigger than the size of memory there is nothing that can be
48 * done - the thrashing set could never fit into memory under any
51 * However, the average access distance could be bigger than the
52 * inactive list, yet smaller than the size of memory. In this case,
53 * the set could fit into memory if it weren't for the currently
54 * active pages - which may be used more, hopefully less frequently:
56 * +-memory available to cache-+
58 * +-inactive------+-active----+
59 * a b | c d e f g h i | J K L M N |
60 * +---------------+-----------+
62 * It is prohibitively expensive to accurately track access frequency
63 * of pages. But a reasonable approximation can be made to measure
64 * thrashing on the inactive list, after which refaulting pages can be
65 * activated optimistically to compete with the existing active pages.
67 * Approximating inactive page access frequency - Observations:
69 * 1. When a page is accessed for the first time, it is added to the
70 * head of the inactive list, slides every existing inactive page
71 * towards the tail by one slot, and pushes the current tail page
74 * 2. When a page is accessed for the second time, it is promoted to
75 * the active list, shrinking the inactive list by one slot. This
76 * also slides all inactive pages that were faulted into the cache
77 * more recently than the activated page towards the tail of the
82 * 1. The sum of evictions and activations between any two points in
83 * time indicate the minimum number of inactive pages accessed in
86 * 2. Moving one inactive page N page slots towards the tail of the
87 * list requires at least N inactive page accesses.
91 * 1. When a page is finally evicted from memory, the number of
92 * inactive pages accessed while the page was in cache is at least
93 * the number of page slots on the inactive list.
95 * 2. In addition, measuring the sum of evictions and activations (E)
96 * at the time of a page's eviction, and comparing it to another
97 * reading (R) at the time the page faults back into memory tells
98 * the minimum number of accesses while the page was not cached.
99 * This is called the refault distance.
101 * Because the first access of the page was the fault and the second
102 * access the refault, we combine the in-cache distance with the
103 * out-of-cache distance to get the complete minimum access distance
106 * NR_inactive + (R - E)
108 * And knowing the minimum access distance of a page, we can easily
109 * tell if the page would be able to stay in cache assuming all page
110 * slots in the cache were available:
112 * NR_inactive + (R - E) <= NR_inactive + NR_active
114 * which can be further simplified to
116 * (R - E) <= NR_active
118 * Put into words, the refault distance (out-of-cache) can be seen as
119 * a deficit in inactive list space (in-cache). If the inactive list
120 * had (R - E) more page slots, the page would not have been evicted
121 * in between accesses, but activated instead. And on a full system,
122 * the only thing eating into inactive list space is active pages.
125 * Refaulting inactive pages
127 * All that is known about the active list is that the pages have been
128 * accessed more than once in the past. This means that at any given
129 * time there is actually a good chance that pages on the active list
130 * are no longer in active use.
132 * So when a refault distance of (R - E) is observed and there are at
133 * least (R - E) active pages, the refaulting page is activated
134 * optimistically in the hope that (R - E) active pages are actually
135 * used less frequently than the refaulting page - or even not used at
138 * That means if inactive cache is refaulting with a suitable refault
139 * distance, we assume the cache workingset is transitioning and put
140 * pressure on the current active list.
142 * If this is wrong and demotion kicks in, the pages which are truly
143 * used more frequently will be reactivated while the less frequently
144 * used once will be evicted from memory.
146 * But if this is right, the stale pages will be pushed out of memory
147 * and the used pages get to stay in cache.
149 * Refaulting active pages
151 * If on the other hand the refaulting pages have recently been
152 * deactivated, it means that the active list is no longer protecting
153 * actively used cache from reclaim. The cache is NOT transitioning to
154 * a different workingset; the existing workingset is thrashing in the
155 * space allocated to the page cache.
160 * For each node's LRU lists, a counter for inactive evictions and
161 * activations is maintained (node->nonresident_age).
163 * On eviction, a snapshot of this counter (along with some bits to
164 * identify the node) is stored in the now empty page cache
165 * slot of the evicted page. This is called a shadow entry.
167 * On cache misses for which there are shadow entries, an eligible
168 * refault distance will immediately activate the refaulting page.
171 #define WORKINGSET_SHIFT 1
172 #define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
173 WORKINGSET_SHIFT + NODES_SHIFT + \
175 #define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
178 * Eviction timestamps need to be able to cover the full range of
179 * actionable refaults. However, bits are tight in the xarray
180 * entry, and after storing the identifier for the lruvec there might
181 * not be enough left to represent every single actionable refault. In
182 * that case, we have to sacrifice granularity for distance, and group
183 * evictions into coarser buckets by shaving off lower timestamp bits.
185 static unsigned int bucket_order __read_mostly
;
187 static void *pack_shadow(int memcgid
, pg_data_t
*pgdat
, unsigned long eviction
,
190 eviction
>>= bucket_order
;
191 eviction
&= EVICTION_MASK
;
192 eviction
= (eviction
<< MEM_CGROUP_ID_SHIFT
) | memcgid
;
193 eviction
= (eviction
<< NODES_SHIFT
) | pgdat
->node_id
;
194 eviction
= (eviction
<< WORKINGSET_SHIFT
) | workingset
;
196 return xa_mk_value(eviction
);
199 static void unpack_shadow(void *shadow
, int *memcgidp
, pg_data_t
**pgdat
,
200 unsigned long *evictionp
, bool *workingsetp
)
202 unsigned long entry
= xa_to_value(shadow
);
206 workingset
= entry
& ((1UL << WORKINGSET_SHIFT
) - 1);
207 entry
>>= WORKINGSET_SHIFT
;
208 nid
= entry
& ((1UL << NODES_SHIFT
) - 1);
209 entry
>>= NODES_SHIFT
;
210 memcgid
= entry
& ((1UL << MEM_CGROUP_ID_SHIFT
) - 1);
211 entry
>>= MEM_CGROUP_ID_SHIFT
;
214 *pgdat
= NODE_DATA(nid
);
215 *evictionp
= entry
<< bucket_order
;
216 *workingsetp
= workingset
;
220 * workingset_age_nonresident - age non-resident entries as LRU ages
221 * @lruvec: the lruvec that was aged
222 * @nr_pages: the number of pages to count
224 * As in-memory pages are aged, non-resident pages need to be aged as
225 * well, in order for the refault distances later on to be comparable
226 * to the in-memory dimensions. This function allows reclaim and LRU
227 * operations to drive the non-resident aging along in parallel.
229 void workingset_age_nonresident(struct lruvec
*lruvec
, unsigned long nr_pages
)
232 * Reclaiming a cgroup means reclaiming all its children in a
233 * round-robin fashion. That means that each cgroup has an LRU
234 * order that is composed of the LRU orders of its child
235 * cgroups; and every page has an LRU position not just in the
236 * cgroup that owns it, but in all of that group's ancestors.
238 * So when the physical inactive list of a leaf cgroup ages,
239 * the virtual inactive lists of all its parents, including
240 * the root cgroup's, age as well.
243 atomic_long_add(nr_pages
, &lruvec
->nonresident_age
);
244 } while ((lruvec
= parent_lruvec(lruvec
)));
248 * workingset_eviction - note the eviction of a folio from memory
249 * @target_memcg: the cgroup that is causing the reclaim
250 * @folio: the folio being evicted
252 * Return: a shadow entry to be stored in @folio->mapping->i_pages in place
253 * of the evicted @folio so that a later refault can be detected.
255 void *workingset_eviction(struct folio
*folio
, struct mem_cgroup
*target_memcg
)
257 struct pglist_data
*pgdat
= folio_pgdat(folio
);
258 unsigned long eviction
;
259 struct lruvec
*lruvec
;
262 /* Folio is fully exclusive and pins folio's memory cgroup pointer */
263 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
264 VM_BUG_ON_FOLIO(folio_ref_count(folio
), folio
);
265 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
267 lruvec
= mem_cgroup_lruvec(target_memcg
, pgdat
);
268 /* XXX: target_memcg can be NULL, go through lruvec */
269 memcgid
= mem_cgroup_id(lruvec_memcg(lruvec
));
270 eviction
= atomic_long_read(&lruvec
->nonresident_age
);
271 workingset_age_nonresident(lruvec
, folio_nr_pages(folio
));
272 return pack_shadow(memcgid
, pgdat
, eviction
,
273 folio_test_workingset(folio
));
277 * workingset_refault - Evaluate the refault of a previously evicted folio.
278 * @folio: The freshly allocated replacement folio.
279 * @shadow: Shadow entry of the evicted folio.
281 * Calculates and evaluates the refault distance of the previously
282 * evicted folio in the context of the node and the memcg whose memory
283 * pressure caused the eviction.
285 void workingset_refault(struct folio
*folio
, void *shadow
)
287 bool file
= folio_is_file_lru(folio
);
288 struct mem_cgroup
*eviction_memcg
;
289 struct lruvec
*eviction_lruvec
;
290 unsigned long refault_distance
;
291 unsigned long workingset_size
;
292 struct pglist_data
*pgdat
;
293 struct mem_cgroup
*memcg
;
294 unsigned long eviction
;
295 struct lruvec
*lruvec
;
296 unsigned long refault
;
301 unpack_shadow(shadow
, &memcgid
, &pgdat
, &eviction
, &workingset
);
305 * Look up the memcg associated with the stored ID. It might
306 * have been deleted since the folio's eviction.
308 * Note that in rare events the ID could have been recycled
309 * for a new cgroup that refaults a shared folio. This is
310 * impossible to tell from the available data. However, this
311 * should be a rare and limited disturbance, and activations
312 * are always speculative anyway. Ultimately, it's the aging
313 * algorithm's job to shake out the minimum access frequency
314 * for the active cache.
316 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
317 * would be better if the root_mem_cgroup existed in all
318 * configurations instead.
320 eviction_memcg
= mem_cgroup_from_id(memcgid
);
321 if (!mem_cgroup_disabled() && !eviction_memcg
)
323 eviction_lruvec
= mem_cgroup_lruvec(eviction_memcg
, pgdat
);
324 refault
= atomic_long_read(&eviction_lruvec
->nonresident_age
);
327 * Calculate the refault distance
329 * The unsigned subtraction here gives an accurate distance
330 * across nonresident_age overflows in most cases. There is a
331 * special case: usually, shadow entries have a short lifetime
332 * and are either refaulted or reclaimed along with the inode
333 * before they get too old. But it is not impossible for the
334 * nonresident_age to lap a shadow entry in the field, which
335 * can then result in a false small refault distance, leading
336 * to a false activation should this old entry actually
337 * refault again. However, earlier kernels used to deactivate
338 * unconditionally with *every* reclaim invocation for the
339 * longest time, so the occasional inappropriate activation
340 * leading to pressure on the active list is not a problem.
342 refault_distance
= (refault
- eviction
) & EVICTION_MASK
;
345 * The activation decision for this folio is made at the level
346 * where the eviction occurred, as that is where the LRU order
347 * during folio reclaim is being determined.
349 * However, the cgroup that will own the folio is the one that
350 * is actually experiencing the refault event.
352 nr
= folio_nr_pages(folio
);
353 memcg
= folio_memcg(folio
);
354 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
356 mod_lruvec_state(lruvec
, WORKINGSET_REFAULT_BASE
+ file
, nr
);
358 mem_cgroup_flush_stats_delayed();
360 * Compare the distance to the existing workingset size. We
361 * don't activate pages that couldn't stay resident even if
362 * all the memory was available to the workingset. Whether
363 * workingset competition needs to consider anon or not depends
366 workingset_size
= lruvec_page_state(eviction_lruvec
, NR_ACTIVE_FILE
);
368 workingset_size
+= lruvec_page_state(eviction_lruvec
,
371 if (mem_cgroup_get_nr_swap_pages(memcg
) > 0) {
372 workingset_size
+= lruvec_page_state(eviction_lruvec
,
375 workingset_size
+= lruvec_page_state(eviction_lruvec
,
379 if (refault_distance
> workingset_size
)
382 folio_set_active(folio
);
383 workingset_age_nonresident(lruvec
, nr
);
384 mod_lruvec_state(lruvec
, WORKINGSET_ACTIVATE_BASE
+ file
, nr
);
386 /* Folio was active prior to eviction */
388 folio_set_workingset(folio
);
389 /* XXX: Move to lru_cache_add() when it supports new vs putback */
390 lru_note_cost_folio(folio
);
391 mod_lruvec_state(lruvec
, WORKINGSET_RESTORE_BASE
+ file
, nr
);
398 * workingset_activation - note a page activation
399 * @folio: Folio that is being activated.
401 void workingset_activation(struct folio
*folio
)
403 struct mem_cgroup
*memcg
;
407 * Filter non-memcg pages here, e.g. unmap can call
408 * mark_page_accessed() on VDSO pages.
410 * XXX: See workingset_refault() - this should return
411 * root_mem_cgroup even for !CONFIG_MEMCG.
413 memcg
= folio_memcg_rcu(folio
);
414 if (!mem_cgroup_disabled() && !memcg
)
416 workingset_age_nonresident(folio_lruvec(folio
), folio_nr_pages(folio
));
422 * Shadow entries reflect the share of the working set that does not
423 * fit into memory, so their number depends on the access pattern of
424 * the workload. In most cases, they will refault or get reclaimed
425 * along with the inode, but a (malicious) workload that streams
426 * through files with a total size several times that of available
427 * memory, while preventing the inodes from being reclaimed, can
428 * create excessive amounts of shadow nodes. To keep a lid on this,
429 * track shadow nodes and reclaim them when they grow way past the
430 * point where they would still be useful.
433 struct list_lru shadow_nodes
;
435 void workingset_update_node(struct xa_node
*node
)
437 struct address_space
*mapping
;
440 * Track non-empty nodes that contain only shadow entries;
441 * unlink those that contain pages or are being freed.
443 * Avoid acquiring the list_lru lock when the nodes are
444 * already where they should be. The list_empty() test is safe
445 * as node->private_list is protected by the i_pages lock.
447 mapping
= container_of(node
->array
, struct address_space
, i_pages
);
448 lockdep_assert_held(&mapping
->i_pages
.xa_lock
);
450 if (node
->count
&& node
->count
== node
->nr_values
) {
451 if (list_empty(&node
->private_list
)) {
452 list_lru_add(&shadow_nodes
, &node
->private_list
);
453 __inc_lruvec_kmem_state(node
, WORKINGSET_NODES
);
456 if (!list_empty(&node
->private_list
)) {
457 list_lru_del(&shadow_nodes
, &node
->private_list
);
458 __dec_lruvec_kmem_state(node
, WORKINGSET_NODES
);
463 static unsigned long count_shadow_nodes(struct shrinker
*shrinker
,
464 struct shrink_control
*sc
)
466 unsigned long max_nodes
;
470 nodes
= list_lru_shrink_count(&shadow_nodes
, sc
);
475 * Approximate a reasonable limit for the nodes
476 * containing shadow entries. We don't need to keep more
477 * shadow entries than possible pages on the active list,
478 * since refault distances bigger than that are dismissed.
480 * The size of the active list converges toward 100% of
481 * overall page cache as memory grows, with only a tiny
482 * inactive list. Assume the total cache size for that.
484 * Nodes might be sparsely populated, with only one shadow
485 * entry in the extreme case. Obviously, we cannot keep one
486 * node for every eligible shadow entry, so compromise on a
487 * worst-case density of 1/8th. Below that, not all eligible
488 * refaults can be detected anymore.
490 * On 64-bit with 7 xa_nodes per page and 64 slots
491 * each, this will reclaim shadow entries when they consume
492 * ~1.8% of available memory:
494 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
498 struct lruvec
*lruvec
;
501 lruvec
= mem_cgroup_lruvec(sc
->memcg
, NODE_DATA(sc
->nid
));
502 for (pages
= 0, i
= 0; i
< NR_LRU_LISTS
; i
++)
503 pages
+= lruvec_page_state_local(lruvec
,
505 pages
+= lruvec_page_state_local(
506 lruvec
, NR_SLAB_RECLAIMABLE_B
) >> PAGE_SHIFT
;
507 pages
+= lruvec_page_state_local(
508 lruvec
, NR_SLAB_UNRECLAIMABLE_B
) >> PAGE_SHIFT
;
511 pages
= node_present_pages(sc
->nid
);
513 max_nodes
= pages
>> (XA_CHUNK_SHIFT
- 3);
515 if (nodes
<= max_nodes
)
517 return nodes
- max_nodes
;
520 static enum lru_status
shadow_lru_isolate(struct list_head
*item
,
521 struct list_lru_one
*lru
,
522 spinlock_t
*lru_lock
,
523 void *arg
) __must_hold(lru_lock
)
525 struct xa_node
*node
= container_of(item
, struct xa_node
, private_list
);
526 struct address_space
*mapping
;
530 * Page cache insertions and deletions synchronously maintain
531 * the shadow node LRU under the i_pages lock and the
532 * lru_lock. Because the page cache tree is emptied before
533 * the inode can be destroyed, holding the lru_lock pins any
534 * address_space that has nodes on the LRU.
536 * We can then safely transition to the i_pages lock to
537 * pin only the address_space of the particular node we want
538 * to reclaim, take the node off-LRU, and drop the lru_lock.
541 mapping
= container_of(node
->array
, struct address_space
, i_pages
);
543 /* Coming from the list, invert the lock order */
544 if (!xa_trylock(&mapping
->i_pages
)) {
545 spin_unlock_irq(lru_lock
);
550 if (!spin_trylock(&mapping
->host
->i_lock
)) {
551 xa_unlock(&mapping
->i_pages
);
552 spin_unlock_irq(lru_lock
);
557 list_lru_isolate(lru
, item
);
558 __dec_lruvec_kmem_state(node
, WORKINGSET_NODES
);
560 spin_unlock(lru_lock
);
563 * The nodes should only contain one or more shadow entries,
564 * no pages, so we expect to be able to remove them all and
565 * delete and free the empty node afterwards.
567 if (WARN_ON_ONCE(!node
->nr_values
))
569 if (WARN_ON_ONCE(node
->count
!= node
->nr_values
))
571 xa_delete_node(node
, workingset_update_node
);
572 __inc_lruvec_kmem_state(node
, WORKINGSET_NODERECLAIM
);
575 xa_unlock_irq(&mapping
->i_pages
);
576 if (mapping_shrinkable(mapping
))
577 inode_add_lru(mapping
->host
);
578 spin_unlock(&mapping
->host
->i_lock
);
579 ret
= LRU_REMOVED_RETRY
;
582 spin_lock_irq(lru_lock
);
586 static unsigned long scan_shadow_nodes(struct shrinker
*shrinker
,
587 struct shrink_control
*sc
)
589 /* list_lru lock nests inside the IRQ-safe i_pages lock */
590 return list_lru_shrink_walk_irq(&shadow_nodes
, sc
, shadow_lru_isolate
,
594 static struct shrinker workingset_shadow_shrinker
= {
595 .count_objects
= count_shadow_nodes
,
596 .scan_objects
= scan_shadow_nodes
,
597 .seeks
= 0, /* ->count reports only fully expendable nodes */
598 .flags
= SHRINKER_NUMA_AWARE
| SHRINKER_MEMCG_AWARE
,
602 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
605 static struct lock_class_key shadow_nodes_key
;
607 static int __init
workingset_init(void)
609 unsigned int timestamp_bits
;
610 unsigned int max_order
;
613 BUILD_BUG_ON(BITS_PER_LONG
< EVICTION_SHIFT
);
615 * Calculate the eviction bucket size to cover the longest
616 * actionable refault distance, which is currently half of
617 * memory (totalram_pages/2). However, memory hotplug may add
618 * some more pages at runtime, so keep working with up to
619 * double the initial memory by using totalram_pages as-is.
621 timestamp_bits
= BITS_PER_LONG
- EVICTION_SHIFT
;
622 max_order
= fls_long(totalram_pages() - 1);
623 if (max_order
> timestamp_bits
)
624 bucket_order
= max_order
- timestamp_bits
;
625 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
626 timestamp_bits
, max_order
, bucket_order
);
628 ret
= prealloc_shrinker(&workingset_shadow_shrinker
, "mm-shadow");
631 ret
= __list_lru_init(&shadow_nodes
, true, &shadow_nodes_key
,
632 &workingset_shadow_shrinker
);
635 register_shrinker_prepared(&workingset_shadow_shrinker
);
638 free_prealloced_shrinker(&workingset_shadow_shrinker
);
642 module_init(workingset_init
);