162306a36Sopenharmony_ci// SPDX-License-Identifier: GPL-2.0 262306a36Sopenharmony_ci/* 362306a36Sopenharmony_ci * Workingset detection 462306a36Sopenharmony_ci * 562306a36Sopenharmony_ci * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner 662306a36Sopenharmony_ci */ 762306a36Sopenharmony_ci 862306a36Sopenharmony_ci#include <linux/memcontrol.h> 962306a36Sopenharmony_ci#include <linux/mm_inline.h> 1062306a36Sopenharmony_ci#include <linux/writeback.h> 1162306a36Sopenharmony_ci#include <linux/shmem_fs.h> 1262306a36Sopenharmony_ci#include <linux/pagemap.h> 1362306a36Sopenharmony_ci#include <linux/atomic.h> 1462306a36Sopenharmony_ci#include <linux/module.h> 1562306a36Sopenharmony_ci#include <linux/swap.h> 1662306a36Sopenharmony_ci#include <linux/dax.h> 1762306a36Sopenharmony_ci#include <linux/fs.h> 1862306a36Sopenharmony_ci#include <linux/mm.h> 1962306a36Sopenharmony_ci 2062306a36Sopenharmony_ci/* 2162306a36Sopenharmony_ci * Double CLOCK lists 2262306a36Sopenharmony_ci * 2362306a36Sopenharmony_ci * Per node, two clock lists are maintained for file pages: the 2462306a36Sopenharmony_ci * inactive and the active list. Freshly faulted pages start out at 2562306a36Sopenharmony_ci * the head of the inactive list and page reclaim scans pages from the 2662306a36Sopenharmony_ci * tail. Pages that are accessed multiple times on the inactive list 2762306a36Sopenharmony_ci * are promoted to the active list, to protect them from reclaim, 2862306a36Sopenharmony_ci * whereas active pages are demoted to the inactive list when the 2962306a36Sopenharmony_ci * active list grows too big. 3062306a36Sopenharmony_ci * 3162306a36Sopenharmony_ci * fault ------------------------+ 3262306a36Sopenharmony_ci * | 3362306a36Sopenharmony_ci * +--------------+ | +-------------+ 3462306a36Sopenharmony_ci * reclaim <- | inactive | <-+-- demotion | active | <--+ 3562306a36Sopenharmony_ci * +--------------+ +-------------+ | 3662306a36Sopenharmony_ci * | | 3762306a36Sopenharmony_ci * +-------------- promotion ------------------+ 3862306a36Sopenharmony_ci * 3962306a36Sopenharmony_ci * 4062306a36Sopenharmony_ci * Access frequency and refault distance 4162306a36Sopenharmony_ci * 4262306a36Sopenharmony_ci * A workload is thrashing when its pages are frequently used but they 4362306a36Sopenharmony_ci * are evicted from the inactive list every time before another access 4462306a36Sopenharmony_ci * would have promoted them to the active list. 4562306a36Sopenharmony_ci * 4662306a36Sopenharmony_ci * In cases where the average access distance between thrashing pages 4762306a36Sopenharmony_ci * is bigger than the size of memory there is nothing that can be 4862306a36Sopenharmony_ci * done - the thrashing set could never fit into memory under any 4962306a36Sopenharmony_ci * circumstance. 5062306a36Sopenharmony_ci * 5162306a36Sopenharmony_ci * However, the average access distance could be bigger than the 5262306a36Sopenharmony_ci * inactive list, yet smaller than the size of memory. In this case, 5362306a36Sopenharmony_ci * the set could fit into memory if it weren't for the currently 5462306a36Sopenharmony_ci * active pages - which may be used more, hopefully less frequently: 5562306a36Sopenharmony_ci * 5662306a36Sopenharmony_ci * +-memory available to cache-+ 5762306a36Sopenharmony_ci * | | 5862306a36Sopenharmony_ci * +-inactive------+-active----+ 5962306a36Sopenharmony_ci * a b | c d e f g h i | J K L M N | 6062306a36Sopenharmony_ci * +---------------+-----------+ 6162306a36Sopenharmony_ci * 6262306a36Sopenharmony_ci * It is prohibitively expensive to accurately track access frequency 6362306a36Sopenharmony_ci * of pages. But a reasonable approximation can be made to measure 6462306a36Sopenharmony_ci * thrashing on the inactive list, after which refaulting pages can be 6562306a36Sopenharmony_ci * activated optimistically to compete with the existing active pages. 6662306a36Sopenharmony_ci * 6762306a36Sopenharmony_ci * Approximating inactive page access frequency - Observations: 6862306a36Sopenharmony_ci * 6962306a36Sopenharmony_ci * 1. When a page is accessed for the first time, it is added to the 7062306a36Sopenharmony_ci * head of the inactive list, slides every existing inactive page 7162306a36Sopenharmony_ci * towards the tail by one slot, and pushes the current tail page 7262306a36Sopenharmony_ci * out of memory. 7362306a36Sopenharmony_ci * 7462306a36Sopenharmony_ci * 2. When a page is accessed for the second time, it is promoted to 7562306a36Sopenharmony_ci * the active list, shrinking the inactive list by one slot. This 7662306a36Sopenharmony_ci * also slides all inactive pages that were faulted into the cache 7762306a36Sopenharmony_ci * more recently than the activated page towards the tail of the 7862306a36Sopenharmony_ci * inactive list. 7962306a36Sopenharmony_ci * 8062306a36Sopenharmony_ci * Thus: 8162306a36Sopenharmony_ci * 8262306a36Sopenharmony_ci * 1. The sum of evictions and activations between any two points in 8362306a36Sopenharmony_ci * time indicate the minimum number of inactive pages accessed in 8462306a36Sopenharmony_ci * between. 8562306a36Sopenharmony_ci * 8662306a36Sopenharmony_ci * 2. Moving one inactive page N page slots towards the tail of the 8762306a36Sopenharmony_ci * list requires at least N inactive page accesses. 8862306a36Sopenharmony_ci * 8962306a36Sopenharmony_ci * Combining these: 9062306a36Sopenharmony_ci * 9162306a36Sopenharmony_ci * 1. When a page is finally evicted from memory, the number of 9262306a36Sopenharmony_ci * inactive pages accessed while the page was in cache is at least 9362306a36Sopenharmony_ci * the number of page slots on the inactive list. 9462306a36Sopenharmony_ci * 9562306a36Sopenharmony_ci * 2. In addition, measuring the sum of evictions and activations (E) 9662306a36Sopenharmony_ci * at the time of a page's eviction, and comparing it to another 9762306a36Sopenharmony_ci * reading (R) at the time the page faults back into memory tells 9862306a36Sopenharmony_ci * the minimum number of accesses while the page was not cached. 9962306a36Sopenharmony_ci * This is called the refault distance. 10062306a36Sopenharmony_ci * 10162306a36Sopenharmony_ci * Because the first access of the page was the fault and the second 10262306a36Sopenharmony_ci * access the refault, we combine the in-cache distance with the 10362306a36Sopenharmony_ci * out-of-cache distance to get the complete minimum access distance 10462306a36Sopenharmony_ci * of this page: 10562306a36Sopenharmony_ci * 10662306a36Sopenharmony_ci * NR_inactive + (R - E) 10762306a36Sopenharmony_ci * 10862306a36Sopenharmony_ci * And knowing the minimum access distance of a page, we can easily 10962306a36Sopenharmony_ci * tell if the page would be able to stay in cache assuming all page 11062306a36Sopenharmony_ci * slots in the cache were available: 11162306a36Sopenharmony_ci * 11262306a36Sopenharmony_ci * NR_inactive + (R - E) <= NR_inactive + NR_active 11362306a36Sopenharmony_ci * 11462306a36Sopenharmony_ci * If we have swap we should consider about NR_inactive_anon and 11562306a36Sopenharmony_ci * NR_active_anon, so for page cache and anonymous respectively: 11662306a36Sopenharmony_ci * 11762306a36Sopenharmony_ci * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file 11862306a36Sopenharmony_ci * + NR_inactive_anon + NR_active_anon 11962306a36Sopenharmony_ci * 12062306a36Sopenharmony_ci * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon 12162306a36Sopenharmony_ci * + NR_inactive_file + NR_active_file 12262306a36Sopenharmony_ci * 12362306a36Sopenharmony_ci * Which can be further simplified to: 12462306a36Sopenharmony_ci * 12562306a36Sopenharmony_ci * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon 12662306a36Sopenharmony_ci * 12762306a36Sopenharmony_ci * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file 12862306a36Sopenharmony_ci * 12962306a36Sopenharmony_ci * Put into words, the refault distance (out-of-cache) can be seen as 13062306a36Sopenharmony_ci * a deficit in inactive list space (in-cache). If the inactive list 13162306a36Sopenharmony_ci * had (R - E) more page slots, the page would not have been evicted 13262306a36Sopenharmony_ci * in between accesses, but activated instead. And on a full system, 13362306a36Sopenharmony_ci * the only thing eating into inactive list space is active pages. 13462306a36Sopenharmony_ci * 13562306a36Sopenharmony_ci * 13662306a36Sopenharmony_ci * Refaulting inactive pages 13762306a36Sopenharmony_ci * 13862306a36Sopenharmony_ci * All that is known about the active list is that the pages have been 13962306a36Sopenharmony_ci * accessed more than once in the past. This means that at any given 14062306a36Sopenharmony_ci * time there is actually a good chance that pages on the active list 14162306a36Sopenharmony_ci * are no longer in active use. 14262306a36Sopenharmony_ci * 14362306a36Sopenharmony_ci * So when a refault distance of (R - E) is observed and there are at 14462306a36Sopenharmony_ci * least (R - E) pages in the userspace workingset, the refaulting page 14562306a36Sopenharmony_ci * is activated optimistically in the hope that (R - E) pages are actually 14662306a36Sopenharmony_ci * used less frequently than the refaulting page - or even not used at 14762306a36Sopenharmony_ci * all anymore. 14862306a36Sopenharmony_ci * 14962306a36Sopenharmony_ci * That means if inactive cache is refaulting with a suitable refault 15062306a36Sopenharmony_ci * distance, we assume the cache workingset is transitioning and put 15162306a36Sopenharmony_ci * pressure on the current workingset. 15262306a36Sopenharmony_ci * 15362306a36Sopenharmony_ci * If this is wrong and demotion kicks in, the pages which are truly 15462306a36Sopenharmony_ci * used more frequently will be reactivated while the less frequently 15562306a36Sopenharmony_ci * used once will be evicted from memory. 15662306a36Sopenharmony_ci * 15762306a36Sopenharmony_ci * But if this is right, the stale pages will be pushed out of memory 15862306a36Sopenharmony_ci * and the used pages get to stay in cache. 15962306a36Sopenharmony_ci * 16062306a36Sopenharmony_ci * Refaulting active pages 16162306a36Sopenharmony_ci * 16262306a36Sopenharmony_ci * If on the other hand the refaulting pages have recently been 16362306a36Sopenharmony_ci * deactivated, it means that the active list is no longer protecting 16462306a36Sopenharmony_ci * actively used cache from reclaim. The cache is NOT transitioning to 16562306a36Sopenharmony_ci * a different workingset; the existing workingset is thrashing in the 16662306a36Sopenharmony_ci * space allocated to the page cache. 16762306a36Sopenharmony_ci * 16862306a36Sopenharmony_ci * 16962306a36Sopenharmony_ci * Implementation 17062306a36Sopenharmony_ci * 17162306a36Sopenharmony_ci * For each node's LRU lists, a counter for inactive evictions and 17262306a36Sopenharmony_ci * activations is maintained (node->nonresident_age). 17362306a36Sopenharmony_ci * 17462306a36Sopenharmony_ci * On eviction, a snapshot of this counter (along with some bits to 17562306a36Sopenharmony_ci * identify the node) is stored in the now empty page cache 17662306a36Sopenharmony_ci * slot of the evicted page. This is called a shadow entry. 17762306a36Sopenharmony_ci * 17862306a36Sopenharmony_ci * On cache misses for which there are shadow entries, an eligible 17962306a36Sopenharmony_ci * refault distance will immediately activate the refaulting page. 18062306a36Sopenharmony_ci */ 18162306a36Sopenharmony_ci 18262306a36Sopenharmony_ci#define WORKINGSET_SHIFT 1 18362306a36Sopenharmony_ci#define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \ 18462306a36Sopenharmony_ci WORKINGSET_SHIFT + NODES_SHIFT + \ 18562306a36Sopenharmony_ci MEM_CGROUP_ID_SHIFT) 18662306a36Sopenharmony_ci#define EVICTION_MASK (~0UL >> EVICTION_SHIFT) 18762306a36Sopenharmony_ci 18862306a36Sopenharmony_ci/* 18962306a36Sopenharmony_ci * Eviction timestamps need to be able to cover the full range of 19062306a36Sopenharmony_ci * actionable refaults. However, bits are tight in the xarray 19162306a36Sopenharmony_ci * entry, and after storing the identifier for the lruvec there might 19262306a36Sopenharmony_ci * not be enough left to represent every single actionable refault. In 19362306a36Sopenharmony_ci * that case, we have to sacrifice granularity for distance, and group 19462306a36Sopenharmony_ci * evictions into coarser buckets by shaving off lower timestamp bits. 19562306a36Sopenharmony_ci */ 19662306a36Sopenharmony_cistatic unsigned int bucket_order __read_mostly; 19762306a36Sopenharmony_ci 19862306a36Sopenharmony_cistatic void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction, 19962306a36Sopenharmony_ci bool workingset) 20062306a36Sopenharmony_ci{ 20162306a36Sopenharmony_ci eviction &= EVICTION_MASK; 20262306a36Sopenharmony_ci eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid; 20362306a36Sopenharmony_ci eviction = (eviction << NODES_SHIFT) | pgdat->node_id; 20462306a36Sopenharmony_ci eviction = (eviction << WORKINGSET_SHIFT) | workingset; 20562306a36Sopenharmony_ci 20662306a36Sopenharmony_ci return xa_mk_value(eviction); 20762306a36Sopenharmony_ci} 20862306a36Sopenharmony_ci 20962306a36Sopenharmony_cistatic void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat, 21062306a36Sopenharmony_ci unsigned long *evictionp, bool *workingsetp) 21162306a36Sopenharmony_ci{ 21262306a36Sopenharmony_ci unsigned long entry = xa_to_value(shadow); 21362306a36Sopenharmony_ci int memcgid, nid; 21462306a36Sopenharmony_ci bool workingset; 21562306a36Sopenharmony_ci 21662306a36Sopenharmony_ci workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1); 21762306a36Sopenharmony_ci entry >>= WORKINGSET_SHIFT; 21862306a36Sopenharmony_ci nid = entry & ((1UL << NODES_SHIFT) - 1); 21962306a36Sopenharmony_ci entry >>= NODES_SHIFT; 22062306a36Sopenharmony_ci memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1); 22162306a36Sopenharmony_ci entry >>= MEM_CGROUP_ID_SHIFT; 22262306a36Sopenharmony_ci 22362306a36Sopenharmony_ci *memcgidp = memcgid; 22462306a36Sopenharmony_ci *pgdat = NODE_DATA(nid); 22562306a36Sopenharmony_ci *evictionp = entry; 22662306a36Sopenharmony_ci *workingsetp = workingset; 22762306a36Sopenharmony_ci} 22862306a36Sopenharmony_ci 22962306a36Sopenharmony_ci#ifdef CONFIG_LRU_GEN 23062306a36Sopenharmony_ci 23162306a36Sopenharmony_cistatic void *lru_gen_eviction(struct folio *folio) 23262306a36Sopenharmony_ci{ 23362306a36Sopenharmony_ci int hist; 23462306a36Sopenharmony_ci unsigned long token; 23562306a36Sopenharmony_ci unsigned long min_seq; 23662306a36Sopenharmony_ci struct lruvec *lruvec; 23762306a36Sopenharmony_ci struct lru_gen_folio *lrugen; 23862306a36Sopenharmony_ci int type = folio_is_file_lru(folio); 23962306a36Sopenharmony_ci int delta = folio_nr_pages(folio); 24062306a36Sopenharmony_ci int refs = folio_lru_refs(folio); 24162306a36Sopenharmony_ci int tier = lru_tier_from_refs(refs); 24262306a36Sopenharmony_ci struct mem_cgroup *memcg = folio_memcg(folio); 24362306a36Sopenharmony_ci struct pglist_data *pgdat = folio_pgdat(folio); 24462306a36Sopenharmony_ci 24562306a36Sopenharmony_ci BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT); 24662306a36Sopenharmony_ci 24762306a36Sopenharmony_ci lruvec = mem_cgroup_lruvec(memcg, pgdat); 24862306a36Sopenharmony_ci lrugen = &lruvec->lrugen; 24962306a36Sopenharmony_ci min_seq = READ_ONCE(lrugen->min_seq[type]); 25062306a36Sopenharmony_ci token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0); 25162306a36Sopenharmony_ci 25262306a36Sopenharmony_ci hist = lru_hist_from_seq(min_seq); 25362306a36Sopenharmony_ci atomic_long_add(delta, &lrugen->evicted[hist][type][tier]); 25462306a36Sopenharmony_ci 25562306a36Sopenharmony_ci return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs); 25662306a36Sopenharmony_ci} 25762306a36Sopenharmony_ci 25862306a36Sopenharmony_ci/* 25962306a36Sopenharmony_ci * Tests if the shadow entry is for a folio that was recently evicted. 26062306a36Sopenharmony_ci * Fills in @lruvec, @token, @workingset with the values unpacked from shadow. 26162306a36Sopenharmony_ci */ 26262306a36Sopenharmony_cistatic bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, 26362306a36Sopenharmony_ci unsigned long *token, bool *workingset) 26462306a36Sopenharmony_ci{ 26562306a36Sopenharmony_ci int memcg_id; 26662306a36Sopenharmony_ci unsigned long min_seq; 26762306a36Sopenharmony_ci struct mem_cgroup *memcg; 26862306a36Sopenharmony_ci struct pglist_data *pgdat; 26962306a36Sopenharmony_ci 27062306a36Sopenharmony_ci unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset); 27162306a36Sopenharmony_ci 27262306a36Sopenharmony_ci memcg = mem_cgroup_from_id(memcg_id); 27362306a36Sopenharmony_ci *lruvec = mem_cgroup_lruvec(memcg, pgdat); 27462306a36Sopenharmony_ci 27562306a36Sopenharmony_ci min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]); 27662306a36Sopenharmony_ci return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)); 27762306a36Sopenharmony_ci} 27862306a36Sopenharmony_ci 27962306a36Sopenharmony_cistatic void lru_gen_refault(struct folio *folio, void *shadow) 28062306a36Sopenharmony_ci{ 28162306a36Sopenharmony_ci bool recent; 28262306a36Sopenharmony_ci int hist, tier, refs; 28362306a36Sopenharmony_ci bool workingset; 28462306a36Sopenharmony_ci unsigned long token; 28562306a36Sopenharmony_ci struct lruvec *lruvec; 28662306a36Sopenharmony_ci struct lru_gen_folio *lrugen; 28762306a36Sopenharmony_ci int type = folio_is_file_lru(folio); 28862306a36Sopenharmony_ci int delta = folio_nr_pages(folio); 28962306a36Sopenharmony_ci 29062306a36Sopenharmony_ci rcu_read_lock(); 29162306a36Sopenharmony_ci 29262306a36Sopenharmony_ci recent = lru_gen_test_recent(shadow, type, &lruvec, &token, &workingset); 29362306a36Sopenharmony_ci if (lruvec != folio_lruvec(folio)) 29462306a36Sopenharmony_ci goto unlock; 29562306a36Sopenharmony_ci 29662306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta); 29762306a36Sopenharmony_ci 29862306a36Sopenharmony_ci if (!recent) 29962306a36Sopenharmony_ci goto unlock; 30062306a36Sopenharmony_ci 30162306a36Sopenharmony_ci lrugen = &lruvec->lrugen; 30262306a36Sopenharmony_ci 30362306a36Sopenharmony_ci hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type])); 30462306a36Sopenharmony_ci /* see the comment in folio_lru_refs() */ 30562306a36Sopenharmony_ci refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset; 30662306a36Sopenharmony_ci tier = lru_tier_from_refs(refs); 30762306a36Sopenharmony_ci 30862306a36Sopenharmony_ci atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]); 30962306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta); 31062306a36Sopenharmony_ci 31162306a36Sopenharmony_ci /* 31262306a36Sopenharmony_ci * Count the following two cases as stalls: 31362306a36Sopenharmony_ci * 1. For pages accessed through page tables, hotter pages pushed out 31462306a36Sopenharmony_ci * hot pages which refaulted immediately. 31562306a36Sopenharmony_ci * 2. For pages accessed multiple times through file descriptors, 31662306a36Sopenharmony_ci * they would have been protected by sort_folio(). 31762306a36Sopenharmony_ci */ 31862306a36Sopenharmony_ci if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) { 31962306a36Sopenharmony_ci set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset)); 32062306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta); 32162306a36Sopenharmony_ci } 32262306a36Sopenharmony_ciunlock: 32362306a36Sopenharmony_ci rcu_read_unlock(); 32462306a36Sopenharmony_ci} 32562306a36Sopenharmony_ci 32662306a36Sopenharmony_ci#else /* !CONFIG_LRU_GEN */ 32762306a36Sopenharmony_ci 32862306a36Sopenharmony_cistatic void *lru_gen_eviction(struct folio *folio) 32962306a36Sopenharmony_ci{ 33062306a36Sopenharmony_ci return NULL; 33162306a36Sopenharmony_ci} 33262306a36Sopenharmony_ci 33362306a36Sopenharmony_cistatic bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec, 33462306a36Sopenharmony_ci unsigned long *token, bool *workingset) 33562306a36Sopenharmony_ci{ 33662306a36Sopenharmony_ci return false; 33762306a36Sopenharmony_ci} 33862306a36Sopenharmony_ci 33962306a36Sopenharmony_cistatic void lru_gen_refault(struct folio *folio, void *shadow) 34062306a36Sopenharmony_ci{ 34162306a36Sopenharmony_ci} 34262306a36Sopenharmony_ci 34362306a36Sopenharmony_ci#endif /* CONFIG_LRU_GEN */ 34462306a36Sopenharmony_ci 34562306a36Sopenharmony_ci/** 34662306a36Sopenharmony_ci * workingset_age_nonresident - age non-resident entries as LRU ages 34762306a36Sopenharmony_ci * @lruvec: the lruvec that was aged 34862306a36Sopenharmony_ci * @nr_pages: the number of pages to count 34962306a36Sopenharmony_ci * 35062306a36Sopenharmony_ci * As in-memory pages are aged, non-resident pages need to be aged as 35162306a36Sopenharmony_ci * well, in order for the refault distances later on to be comparable 35262306a36Sopenharmony_ci * to the in-memory dimensions. This function allows reclaim and LRU 35362306a36Sopenharmony_ci * operations to drive the non-resident aging along in parallel. 35462306a36Sopenharmony_ci */ 35562306a36Sopenharmony_civoid workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages) 35662306a36Sopenharmony_ci{ 35762306a36Sopenharmony_ci /* 35862306a36Sopenharmony_ci * Reclaiming a cgroup means reclaiming all its children in a 35962306a36Sopenharmony_ci * round-robin fashion. That means that each cgroup has an LRU 36062306a36Sopenharmony_ci * order that is composed of the LRU orders of its child 36162306a36Sopenharmony_ci * cgroups; and every page has an LRU position not just in the 36262306a36Sopenharmony_ci * cgroup that owns it, but in all of that group's ancestors. 36362306a36Sopenharmony_ci * 36462306a36Sopenharmony_ci * So when the physical inactive list of a leaf cgroup ages, 36562306a36Sopenharmony_ci * the virtual inactive lists of all its parents, including 36662306a36Sopenharmony_ci * the root cgroup's, age as well. 36762306a36Sopenharmony_ci */ 36862306a36Sopenharmony_ci do { 36962306a36Sopenharmony_ci atomic_long_add(nr_pages, &lruvec->nonresident_age); 37062306a36Sopenharmony_ci } while ((lruvec = parent_lruvec(lruvec))); 37162306a36Sopenharmony_ci} 37262306a36Sopenharmony_ci 37362306a36Sopenharmony_ci/** 37462306a36Sopenharmony_ci * workingset_eviction - note the eviction of a folio from memory 37562306a36Sopenharmony_ci * @target_memcg: the cgroup that is causing the reclaim 37662306a36Sopenharmony_ci * @folio: the folio being evicted 37762306a36Sopenharmony_ci * 37862306a36Sopenharmony_ci * Return: a shadow entry to be stored in @folio->mapping->i_pages in place 37962306a36Sopenharmony_ci * of the evicted @folio so that a later refault can be detected. 38062306a36Sopenharmony_ci */ 38162306a36Sopenharmony_civoid *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg) 38262306a36Sopenharmony_ci{ 38362306a36Sopenharmony_ci struct pglist_data *pgdat = folio_pgdat(folio); 38462306a36Sopenharmony_ci unsigned long eviction; 38562306a36Sopenharmony_ci struct lruvec *lruvec; 38662306a36Sopenharmony_ci int memcgid; 38762306a36Sopenharmony_ci 38862306a36Sopenharmony_ci /* Folio is fully exclusive and pins folio's memory cgroup pointer */ 38962306a36Sopenharmony_ci VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); 39062306a36Sopenharmony_ci VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 39162306a36Sopenharmony_ci VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 39262306a36Sopenharmony_ci 39362306a36Sopenharmony_ci if (lru_gen_enabled()) 39462306a36Sopenharmony_ci return lru_gen_eviction(folio); 39562306a36Sopenharmony_ci 39662306a36Sopenharmony_ci lruvec = mem_cgroup_lruvec(target_memcg, pgdat); 39762306a36Sopenharmony_ci /* XXX: target_memcg can be NULL, go through lruvec */ 39862306a36Sopenharmony_ci memcgid = mem_cgroup_id(lruvec_memcg(lruvec)); 39962306a36Sopenharmony_ci eviction = atomic_long_read(&lruvec->nonresident_age); 40062306a36Sopenharmony_ci eviction >>= bucket_order; 40162306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 40262306a36Sopenharmony_ci if (!is_prot_page(folio_page(folio, 0)) && page_is_file_lru(folio_page(folio, 0))) { 40362306a36Sopenharmony_ci lruvec = folio_lruvec(folio); 40462306a36Sopenharmony_ci workingset_age_nonresident(lruvec, folio_nr_pages(folio)); 40562306a36Sopenharmony_ci } else { 40662306a36Sopenharmony_ci workingset_age_nonresident(lruvec, folio_nr_pages(folio)); 40762306a36Sopenharmony_ci } 40862306a36Sopenharmony_ci#else 40962306a36Sopenharmony_ci workingset_age_nonresident(lruvec, folio_nr_pages(folio)); 41062306a36Sopenharmony_ci#endif 41162306a36Sopenharmony_ci return pack_shadow(memcgid, pgdat, eviction, 41262306a36Sopenharmony_ci folio_test_workingset(folio)); 41362306a36Sopenharmony_ci} 41462306a36Sopenharmony_ci 41562306a36Sopenharmony_ci/** 41662306a36Sopenharmony_ci * workingset_test_recent - tests if the shadow entry is for a folio that was 41762306a36Sopenharmony_ci * recently evicted. Also fills in @workingset with the value unpacked from 41862306a36Sopenharmony_ci * shadow. 41962306a36Sopenharmony_ci * @shadow: the shadow entry to be tested. 42062306a36Sopenharmony_ci * @file: whether the corresponding folio is from the file lru. 42162306a36Sopenharmony_ci * @workingset: where the workingset value unpacked from shadow should 42262306a36Sopenharmony_ci * be stored. 42362306a36Sopenharmony_ci * 42462306a36Sopenharmony_ci * Return: true if the shadow is for a recently evicted folio; false otherwise. 42562306a36Sopenharmony_ci */ 42662306a36Sopenharmony_cibool workingset_test_recent(void *shadow, bool file, bool *workingset) 42762306a36Sopenharmony_ci{ 42862306a36Sopenharmony_ci struct mem_cgroup *eviction_memcg; 42962306a36Sopenharmony_ci struct lruvec *eviction_lruvec; 43062306a36Sopenharmony_ci unsigned long refault_distance; 43162306a36Sopenharmony_ci unsigned long workingset_size; 43262306a36Sopenharmony_ci unsigned long refault; 43362306a36Sopenharmony_ci int memcgid; 43462306a36Sopenharmony_ci struct pglist_data *pgdat; 43562306a36Sopenharmony_ci unsigned long eviction; 43662306a36Sopenharmony_ci 43762306a36Sopenharmony_ci if (lru_gen_enabled()) 43862306a36Sopenharmony_ci return lru_gen_test_recent(shadow, file, &eviction_lruvec, &eviction, workingset); 43962306a36Sopenharmony_ci 44062306a36Sopenharmony_ci unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset); 44162306a36Sopenharmony_ci eviction <<= bucket_order; 44262306a36Sopenharmony_ci 44362306a36Sopenharmony_ci /* 44462306a36Sopenharmony_ci * Look up the memcg associated with the stored ID. It might 44562306a36Sopenharmony_ci * have been deleted since the folio's eviction. 44662306a36Sopenharmony_ci * 44762306a36Sopenharmony_ci * Note that in rare events the ID could have been recycled 44862306a36Sopenharmony_ci * for a new cgroup that refaults a shared folio. This is 44962306a36Sopenharmony_ci * impossible to tell from the available data. However, this 45062306a36Sopenharmony_ci * should be a rare and limited disturbance, and activations 45162306a36Sopenharmony_ci * are always speculative anyway. Ultimately, it's the aging 45262306a36Sopenharmony_ci * algorithm's job to shake out the minimum access frequency 45362306a36Sopenharmony_ci * for the active cache. 45462306a36Sopenharmony_ci * 45562306a36Sopenharmony_ci * XXX: On !CONFIG_MEMCG, this will always return NULL; it 45662306a36Sopenharmony_ci * would be better if the root_mem_cgroup existed in all 45762306a36Sopenharmony_ci * configurations instead. 45862306a36Sopenharmony_ci */ 45962306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 46062306a36Sopenharmony_ci if (memcgid != -1) { 46162306a36Sopenharmony_ci eviction_memcg = mem_cgroup_from_id(memcgid); 46262306a36Sopenharmony_ci if (!mem_cgroup_disabled() && !eviction_memcg) 46362306a36Sopenharmony_ci return false; 46462306a36Sopenharmony_ci } 46562306a36Sopenharmony_ci#else 46662306a36Sopenharmony_ci eviction_memcg = mem_cgroup_from_id(memcgid); 46762306a36Sopenharmony_ci if (!mem_cgroup_disabled() && !eviction_memcg) 46862306a36Sopenharmony_ci return false; 46962306a36Sopenharmony_ci#endif 47062306a36Sopenharmony_ci 47162306a36Sopenharmony_ci eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat); 47262306a36Sopenharmony_ci refault = atomic_long_read(&eviction_lruvec->nonresident_age); 47362306a36Sopenharmony_ci 47462306a36Sopenharmony_ci /* 47562306a36Sopenharmony_ci * Calculate the refault distance 47662306a36Sopenharmony_ci * 47762306a36Sopenharmony_ci * The unsigned subtraction here gives an accurate distance 47862306a36Sopenharmony_ci * across nonresident_age overflows in most cases. There is a 47962306a36Sopenharmony_ci * special case: usually, shadow entries have a short lifetime 48062306a36Sopenharmony_ci * and are either refaulted or reclaimed along with the inode 48162306a36Sopenharmony_ci * before they get too old. But it is not impossible for the 48262306a36Sopenharmony_ci * nonresident_age to lap a shadow entry in the field, which 48362306a36Sopenharmony_ci * can then result in a false small refault distance, leading 48462306a36Sopenharmony_ci * to a false activation should this old entry actually 48562306a36Sopenharmony_ci * refault again. However, earlier kernels used to deactivate 48662306a36Sopenharmony_ci * unconditionally with *every* reclaim invocation for the 48762306a36Sopenharmony_ci * longest time, so the occasional inappropriate activation 48862306a36Sopenharmony_ci * leading to pressure on the active list is not a problem. 48962306a36Sopenharmony_ci */ 49062306a36Sopenharmony_ci refault_distance = (refault - eviction) & EVICTION_MASK; 49162306a36Sopenharmony_ci 49262306a36Sopenharmony_ci /* 49362306a36Sopenharmony_ci * Compare the distance to the existing workingset size. We 49462306a36Sopenharmony_ci * don't activate pages that couldn't stay resident even if 49562306a36Sopenharmony_ci * all the memory was available to the workingset. Whether 49662306a36Sopenharmony_ci * workingset competition needs to consider anon or not depends 49762306a36Sopenharmony_ci * on having free swap space. 49862306a36Sopenharmony_ci */ 49962306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 50062306a36Sopenharmony_ci workingset_size = lruvec_page_state(node_lruvec(pgdat), NR_ACTIVE_FILE); 50162306a36Sopenharmony_ci#else 50262306a36Sopenharmony_ci workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE); 50362306a36Sopenharmony_ci#endif 50462306a36Sopenharmony_ci 50562306a36Sopenharmony_ci if (!file) { 50662306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 50762306a36Sopenharmony_ci workingset_size += lruvec_page_state(node_lruvec(pgdat), 50862306a36Sopenharmony_ci NR_INACTIVE_FILE); 50962306a36Sopenharmony_ci#else 51062306a36Sopenharmony_ci 51162306a36Sopenharmony_ci workingset_size += lruvec_page_state(eviction_lruvec, 51262306a36Sopenharmony_ci NR_INACTIVE_FILE); 51362306a36Sopenharmony_ci#endif 51462306a36Sopenharmony_ci } 51562306a36Sopenharmony_ci if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) { 51662306a36Sopenharmony_ci workingset_size += lruvec_page_state(eviction_lruvec, 51762306a36Sopenharmony_ci NR_ACTIVE_ANON); 51862306a36Sopenharmony_ci if (file) { 51962306a36Sopenharmony_ci workingset_size += lruvec_page_state(eviction_lruvec, 52062306a36Sopenharmony_ci NR_INACTIVE_ANON); 52162306a36Sopenharmony_ci } 52262306a36Sopenharmony_ci } 52362306a36Sopenharmony_ci 52462306a36Sopenharmony_ci return refault_distance <= workingset_size; 52562306a36Sopenharmony_ci} 52662306a36Sopenharmony_ci 52762306a36Sopenharmony_ci/** 52862306a36Sopenharmony_ci * workingset_refault - Evaluate the refault of a previously evicted folio. 52962306a36Sopenharmony_ci * @folio: The freshly allocated replacement folio. 53062306a36Sopenharmony_ci * @shadow: Shadow entry of the evicted folio. 53162306a36Sopenharmony_ci * 53262306a36Sopenharmony_ci * Calculates and evaluates the refault distance of the previously 53362306a36Sopenharmony_ci * evicted folio in the context of the node and the memcg whose memory 53462306a36Sopenharmony_ci * pressure caused the eviction. 53562306a36Sopenharmony_ci */ 53662306a36Sopenharmony_civoid workingset_refault(struct folio *folio, void *shadow) 53762306a36Sopenharmony_ci{ 53862306a36Sopenharmony_ci bool file = folio_is_file_lru(folio); 53962306a36Sopenharmony_ci struct pglist_data *pgdat; 54062306a36Sopenharmony_ci struct mem_cgroup *memcg; 54162306a36Sopenharmony_ci struct lruvec *lruvec; 54262306a36Sopenharmony_ci bool workingset; 54362306a36Sopenharmony_ci long nr; 54462306a36Sopenharmony_ci 54562306a36Sopenharmony_ci if (lru_gen_enabled()) { 54662306a36Sopenharmony_ci lru_gen_refault(folio, shadow); 54762306a36Sopenharmony_ci return; 54862306a36Sopenharmony_ci } 54962306a36Sopenharmony_ci 55062306a36Sopenharmony_ci /* Flush stats (and potentially sleep) before holding RCU read lock */ 55162306a36Sopenharmony_ci mem_cgroup_flush_stats_ratelimited(); 55262306a36Sopenharmony_ci 55362306a36Sopenharmony_ci rcu_read_lock(); 55462306a36Sopenharmony_ci 55562306a36Sopenharmony_ci /* 55662306a36Sopenharmony_ci * The activation decision for this folio is made at the level 55762306a36Sopenharmony_ci * where the eviction occurred, as that is where the LRU order 55862306a36Sopenharmony_ci * during folio reclaim is being determined. 55962306a36Sopenharmony_ci * 56062306a36Sopenharmony_ci * However, the cgroup that will own the folio is the one that 56162306a36Sopenharmony_ci * is actually experiencing the refault event. 56262306a36Sopenharmony_ci */ 56362306a36Sopenharmony_ci nr = folio_nr_pages(folio); 56462306a36Sopenharmony_ci memcg = folio_memcg(folio); 56562306a36Sopenharmony_ci pgdat = folio_pgdat(folio); 56662306a36Sopenharmony_ci lruvec = mem_cgroup_lruvec(memcg, pgdat); 56762306a36Sopenharmony_ci 56862306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 56962306a36Sopenharmony_ci if (!is_prot_page(folio_page(folio, 0)) && file) 57062306a36Sopenharmony_ci mod_lruvec_state(node_lruvec(pgdat), 57162306a36Sopenharmony_ci WORKINGSET_REFAULT_BASE + file, folio_nr_pages(folio)); 57262306a36Sopenharmony_ci else 57362306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr); 57462306a36Sopenharmony_ci#else 57562306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr); 57662306a36Sopenharmony_ci#endif 57762306a36Sopenharmony_ci 57862306a36Sopenharmony_ci if (!workingset_test_recent(shadow, file, &workingset)) 57962306a36Sopenharmony_ci goto out; 58062306a36Sopenharmony_ci 58162306a36Sopenharmony_ci folio_set_active(folio); 58262306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 58362306a36Sopenharmony_ci if (!is_prot_page(folio_page(folio, 0)) && file) { 58462306a36Sopenharmony_ci workingset_age_nonresident(node_lruvec(pgdat), 58562306a36Sopenharmony_ci folio_nr_pages(folio)); 58662306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, folio_nr_pages(folio)); 58762306a36Sopenharmony_ci } else { 58862306a36Sopenharmony_ci workingset_age_nonresident(lruvec, nr); 58962306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr); 59062306a36Sopenharmony_ci } 59162306a36Sopenharmony_ci#else 59262306a36Sopenharmony_ci workingset_age_nonresident(lruvec, nr); 59362306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr); 59462306a36Sopenharmony_ci#endif 59562306a36Sopenharmony_ci 59662306a36Sopenharmony_ci /* Folio was active prior to eviction */ 59762306a36Sopenharmony_ci if (workingset) { 59862306a36Sopenharmony_ci folio_set_workingset(folio); 59962306a36Sopenharmony_ci /* 60062306a36Sopenharmony_ci * XXX: Move to folio_add_lru() when it supports new vs 60162306a36Sopenharmony_ci * putback 60262306a36Sopenharmony_ci */ 60362306a36Sopenharmony_ci lru_note_cost_refault(folio); 60462306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 60562306a36Sopenharmony_ci if (!is_prot_page(folio_page(folio, 0)) && file) 60662306a36Sopenharmony_ci mod_lruvec_state(node_lruvec(pgdat), WORKINGSET_RESTORE_BASE + file, folio_nr_pages(folio)); 60762306a36Sopenharmony_ci else 60862306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr); 60962306a36Sopenharmony_ci#else 61062306a36Sopenharmony_ci mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr); 61162306a36Sopenharmony_ci#endif 61262306a36Sopenharmony_ci } 61362306a36Sopenharmony_ciout: 61462306a36Sopenharmony_ci rcu_read_unlock(); 61562306a36Sopenharmony_ci} 61662306a36Sopenharmony_ci 61762306a36Sopenharmony_ci/** 61862306a36Sopenharmony_ci * workingset_activation - note a page activation 61962306a36Sopenharmony_ci * @folio: Folio that is being activated. 62062306a36Sopenharmony_ci */ 62162306a36Sopenharmony_civoid workingset_activation(struct folio *folio) 62262306a36Sopenharmony_ci{ 62362306a36Sopenharmony_ci struct mem_cgroup *memcg; 62462306a36Sopenharmony_ci struct lruvec *lruvec; 62562306a36Sopenharmony_ci 62662306a36Sopenharmony_ci rcu_read_lock(); 62762306a36Sopenharmony_ci /* 62862306a36Sopenharmony_ci * Filter non-memcg pages here, e.g. unmap can call 62962306a36Sopenharmony_ci * mark_page_accessed() on VDSO pages. 63062306a36Sopenharmony_ci * 63162306a36Sopenharmony_ci * XXX: See workingset_refault() - this should return 63262306a36Sopenharmony_ci * root_mem_cgroup even for !CONFIG_MEMCG. 63362306a36Sopenharmony_ci */ 63462306a36Sopenharmony_ci memcg = folio_memcg_rcu(folio); 63562306a36Sopenharmony_ci if (!mem_cgroup_disabled() && !memcg) 63662306a36Sopenharmony_ci goto out; 63762306a36Sopenharmony_ci#ifdef CONFIG_HYPERHOLD_FILE_LRU 63862306a36Sopenharmony_ci if (!is_prot_page(folio_page(folio, 0)) && page_is_file_lru(folio_page(folio, 0))) { 63962306a36Sopenharmony_ci lruvec = folio_lruvec(folio); 64062306a36Sopenharmony_ci workingset_age_nonresident(lruvec, folio_nr_pages(folio)); 64162306a36Sopenharmony_ci } else { 64262306a36Sopenharmony_ci workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio)); 64362306a36Sopenharmony_ci } 64462306a36Sopenharmony_ci#else 64562306a36Sopenharmony_ci workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio)); 64662306a36Sopenharmony_ci#endif 64762306a36Sopenharmony_ciout: 64862306a36Sopenharmony_ci rcu_read_unlock(); 64962306a36Sopenharmony_ci} 65062306a36Sopenharmony_ci 65162306a36Sopenharmony_ci/* 65262306a36Sopenharmony_ci * Shadow entries reflect the share of the working set that does not 65362306a36Sopenharmony_ci * fit into memory, so their number depends on the access pattern of 65462306a36Sopenharmony_ci * the workload. In most cases, they will refault or get reclaimed 65562306a36Sopenharmony_ci * along with the inode, but a (malicious) workload that streams 65662306a36Sopenharmony_ci * through files with a total size several times that of available 65762306a36Sopenharmony_ci * memory, while preventing the inodes from being reclaimed, can 65862306a36Sopenharmony_ci * create excessive amounts of shadow nodes. To keep a lid on this, 65962306a36Sopenharmony_ci * track shadow nodes and reclaim them when they grow way past the 66062306a36Sopenharmony_ci * point where they would still be useful. 66162306a36Sopenharmony_ci */ 66262306a36Sopenharmony_ci 66362306a36Sopenharmony_cistruct list_lru shadow_nodes; 66462306a36Sopenharmony_ci 66562306a36Sopenharmony_civoid workingset_update_node(struct xa_node *node) 66662306a36Sopenharmony_ci{ 66762306a36Sopenharmony_ci struct address_space *mapping; 66862306a36Sopenharmony_ci 66962306a36Sopenharmony_ci /* 67062306a36Sopenharmony_ci * Track non-empty nodes that contain only shadow entries; 67162306a36Sopenharmony_ci * unlink those that contain pages or are being freed. 67262306a36Sopenharmony_ci * 67362306a36Sopenharmony_ci * Avoid acquiring the list_lru lock when the nodes are 67462306a36Sopenharmony_ci * already where they should be. The list_empty() test is safe 67562306a36Sopenharmony_ci * as node->private_list is protected by the i_pages lock. 67662306a36Sopenharmony_ci */ 67762306a36Sopenharmony_ci mapping = container_of(node->array, struct address_space, i_pages); 67862306a36Sopenharmony_ci lockdep_assert_held(&mapping->i_pages.xa_lock); 67962306a36Sopenharmony_ci 68062306a36Sopenharmony_ci if (node->count && node->count == node->nr_values) { 68162306a36Sopenharmony_ci if (list_empty(&node->private_list)) { 68262306a36Sopenharmony_ci list_lru_add(&shadow_nodes, &node->private_list); 68362306a36Sopenharmony_ci __inc_lruvec_kmem_state(node, WORKINGSET_NODES); 68462306a36Sopenharmony_ci } 68562306a36Sopenharmony_ci } else { 68662306a36Sopenharmony_ci if (!list_empty(&node->private_list)) { 68762306a36Sopenharmony_ci list_lru_del(&shadow_nodes, &node->private_list); 68862306a36Sopenharmony_ci __dec_lruvec_kmem_state(node, WORKINGSET_NODES); 68962306a36Sopenharmony_ci } 69062306a36Sopenharmony_ci } 69162306a36Sopenharmony_ci} 69262306a36Sopenharmony_ci 69362306a36Sopenharmony_cistatic unsigned long count_shadow_nodes(struct shrinker *shrinker, 69462306a36Sopenharmony_ci struct shrink_control *sc) 69562306a36Sopenharmony_ci{ 69662306a36Sopenharmony_ci unsigned long max_nodes; 69762306a36Sopenharmony_ci unsigned long nodes; 69862306a36Sopenharmony_ci unsigned long pages; 69962306a36Sopenharmony_ci 70062306a36Sopenharmony_ci nodes = list_lru_shrink_count(&shadow_nodes, sc); 70162306a36Sopenharmony_ci if (!nodes) 70262306a36Sopenharmony_ci return SHRINK_EMPTY; 70362306a36Sopenharmony_ci 70462306a36Sopenharmony_ci /* 70562306a36Sopenharmony_ci * Approximate a reasonable limit for the nodes 70662306a36Sopenharmony_ci * containing shadow entries. We don't need to keep more 70762306a36Sopenharmony_ci * shadow entries than possible pages on the active list, 70862306a36Sopenharmony_ci * since refault distances bigger than that are dismissed. 70962306a36Sopenharmony_ci * 71062306a36Sopenharmony_ci * The size of the active list converges toward 100% of 71162306a36Sopenharmony_ci * overall page cache as memory grows, with only a tiny 71262306a36Sopenharmony_ci * inactive list. Assume the total cache size for that. 71362306a36Sopenharmony_ci * 71462306a36Sopenharmony_ci * Nodes might be sparsely populated, with only one shadow 71562306a36Sopenharmony_ci * entry in the extreme case. Obviously, we cannot keep one 71662306a36Sopenharmony_ci * node for every eligible shadow entry, so compromise on a 71762306a36Sopenharmony_ci * worst-case density of 1/8th. Below that, not all eligible 71862306a36Sopenharmony_ci * refaults can be detected anymore. 71962306a36Sopenharmony_ci * 72062306a36Sopenharmony_ci * On 64-bit with 7 xa_nodes per page and 64 slots 72162306a36Sopenharmony_ci * each, this will reclaim shadow entries when they consume 72262306a36Sopenharmony_ci * ~1.8% of available memory: 72362306a36Sopenharmony_ci * 72462306a36Sopenharmony_ci * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE 72562306a36Sopenharmony_ci */ 72662306a36Sopenharmony_ci#ifdef CONFIG_MEMCG 72762306a36Sopenharmony_ci#ifndef CONFIG_HYPERHOLD_FILE_LRU 72862306a36Sopenharmony_ci if (sc->memcg) { 72962306a36Sopenharmony_ci struct lruvec *lruvec; 73062306a36Sopenharmony_ci int i; 73162306a36Sopenharmony_ci 73262306a36Sopenharmony_ci mem_cgroup_flush_stats(); 73362306a36Sopenharmony_ci lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid)); 73462306a36Sopenharmony_ci for (pages = 0, i = 0; i < NR_LRU_LISTS; i++) 73562306a36Sopenharmony_ci pages += lruvec_page_state_local(lruvec, 73662306a36Sopenharmony_ci NR_LRU_BASE + i); 73762306a36Sopenharmony_ci pages += lruvec_page_state_local( 73862306a36Sopenharmony_ci lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT; 73962306a36Sopenharmony_ci pages += lruvec_page_state_local( 74062306a36Sopenharmony_ci lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT; 74162306a36Sopenharmony_ci } else 74262306a36Sopenharmony_ci#endif 74362306a36Sopenharmony_ci#endif 74462306a36Sopenharmony_ci pages = node_present_pages(sc->nid); 74562306a36Sopenharmony_ci 74662306a36Sopenharmony_ci max_nodes = pages >> (XA_CHUNK_SHIFT - 3); 74762306a36Sopenharmony_ci 74862306a36Sopenharmony_ci if (nodes <= max_nodes) 74962306a36Sopenharmony_ci return 0; 75062306a36Sopenharmony_ci return nodes - max_nodes; 75162306a36Sopenharmony_ci} 75262306a36Sopenharmony_ci 75362306a36Sopenharmony_cistatic enum lru_status shadow_lru_isolate(struct list_head *item, 75462306a36Sopenharmony_ci struct list_lru_one *lru, 75562306a36Sopenharmony_ci spinlock_t *lru_lock, 75662306a36Sopenharmony_ci void *arg) __must_hold(lru_lock) 75762306a36Sopenharmony_ci{ 75862306a36Sopenharmony_ci struct xa_node *node = container_of(item, struct xa_node, private_list); 75962306a36Sopenharmony_ci struct address_space *mapping; 76062306a36Sopenharmony_ci int ret; 76162306a36Sopenharmony_ci 76262306a36Sopenharmony_ci /* 76362306a36Sopenharmony_ci * Page cache insertions and deletions synchronously maintain 76462306a36Sopenharmony_ci * the shadow node LRU under the i_pages lock and the 76562306a36Sopenharmony_ci * lru_lock. Because the page cache tree is emptied before 76662306a36Sopenharmony_ci * the inode can be destroyed, holding the lru_lock pins any 76762306a36Sopenharmony_ci * address_space that has nodes on the LRU. 76862306a36Sopenharmony_ci * 76962306a36Sopenharmony_ci * We can then safely transition to the i_pages lock to 77062306a36Sopenharmony_ci * pin only the address_space of the particular node we want 77162306a36Sopenharmony_ci * to reclaim, take the node off-LRU, and drop the lru_lock. 77262306a36Sopenharmony_ci */ 77362306a36Sopenharmony_ci 77462306a36Sopenharmony_ci mapping = container_of(node->array, struct address_space, i_pages); 77562306a36Sopenharmony_ci 77662306a36Sopenharmony_ci /* Coming from the list, invert the lock order */ 77762306a36Sopenharmony_ci if (!xa_trylock(&mapping->i_pages)) { 77862306a36Sopenharmony_ci spin_unlock_irq(lru_lock); 77962306a36Sopenharmony_ci ret = LRU_RETRY; 78062306a36Sopenharmony_ci goto out; 78162306a36Sopenharmony_ci } 78262306a36Sopenharmony_ci 78362306a36Sopenharmony_ci /* For page cache we need to hold i_lock */ 78462306a36Sopenharmony_ci if (mapping->host != NULL) { 78562306a36Sopenharmony_ci if (!spin_trylock(&mapping->host->i_lock)) { 78662306a36Sopenharmony_ci xa_unlock(&mapping->i_pages); 78762306a36Sopenharmony_ci spin_unlock_irq(lru_lock); 78862306a36Sopenharmony_ci ret = LRU_RETRY; 78962306a36Sopenharmony_ci goto out; 79062306a36Sopenharmony_ci } 79162306a36Sopenharmony_ci } 79262306a36Sopenharmony_ci 79362306a36Sopenharmony_ci list_lru_isolate(lru, item); 79462306a36Sopenharmony_ci __dec_lruvec_kmem_state(node, WORKINGSET_NODES); 79562306a36Sopenharmony_ci 79662306a36Sopenharmony_ci spin_unlock(lru_lock); 79762306a36Sopenharmony_ci 79862306a36Sopenharmony_ci /* 79962306a36Sopenharmony_ci * The nodes should only contain one or more shadow entries, 80062306a36Sopenharmony_ci * no pages, so we expect to be able to remove them all and 80162306a36Sopenharmony_ci * delete and free the empty node afterwards. 80262306a36Sopenharmony_ci */ 80362306a36Sopenharmony_ci if (WARN_ON_ONCE(!node->nr_values)) 80462306a36Sopenharmony_ci goto out_invalid; 80562306a36Sopenharmony_ci if (WARN_ON_ONCE(node->count != node->nr_values)) 80662306a36Sopenharmony_ci goto out_invalid; 80762306a36Sopenharmony_ci xa_delete_node(node, workingset_update_node); 80862306a36Sopenharmony_ci __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM); 80962306a36Sopenharmony_ci 81062306a36Sopenharmony_ciout_invalid: 81162306a36Sopenharmony_ci xa_unlock_irq(&mapping->i_pages); 81262306a36Sopenharmony_ci if (mapping->host != NULL) { 81362306a36Sopenharmony_ci if (mapping_shrinkable(mapping)) 81462306a36Sopenharmony_ci inode_add_lru(mapping->host); 81562306a36Sopenharmony_ci spin_unlock(&mapping->host->i_lock); 81662306a36Sopenharmony_ci } 81762306a36Sopenharmony_ci ret = LRU_REMOVED_RETRY; 81862306a36Sopenharmony_ciout: 81962306a36Sopenharmony_ci cond_resched(); 82062306a36Sopenharmony_ci spin_lock_irq(lru_lock); 82162306a36Sopenharmony_ci return ret; 82262306a36Sopenharmony_ci} 82362306a36Sopenharmony_ci 82462306a36Sopenharmony_cistatic unsigned long scan_shadow_nodes(struct shrinker *shrinker, 82562306a36Sopenharmony_ci struct shrink_control *sc) 82662306a36Sopenharmony_ci{ 82762306a36Sopenharmony_ci /* list_lru lock nests inside the IRQ-safe i_pages lock */ 82862306a36Sopenharmony_ci return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate, 82962306a36Sopenharmony_ci NULL); 83062306a36Sopenharmony_ci} 83162306a36Sopenharmony_ci 83262306a36Sopenharmony_cistatic struct shrinker workingset_shadow_shrinker = { 83362306a36Sopenharmony_ci .count_objects = count_shadow_nodes, 83462306a36Sopenharmony_ci .scan_objects = scan_shadow_nodes, 83562306a36Sopenharmony_ci .seeks = 0, /* ->count reports only fully expendable nodes */ 83662306a36Sopenharmony_ci .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, 83762306a36Sopenharmony_ci}; 83862306a36Sopenharmony_ci 83962306a36Sopenharmony_ci/* 84062306a36Sopenharmony_ci * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe 84162306a36Sopenharmony_ci * i_pages lock. 84262306a36Sopenharmony_ci */ 84362306a36Sopenharmony_cistatic struct lock_class_key shadow_nodes_key; 84462306a36Sopenharmony_ci 84562306a36Sopenharmony_cistatic int __init workingset_init(void) 84662306a36Sopenharmony_ci{ 84762306a36Sopenharmony_ci unsigned int timestamp_bits; 84862306a36Sopenharmony_ci unsigned int max_order; 84962306a36Sopenharmony_ci int ret; 85062306a36Sopenharmony_ci 85162306a36Sopenharmony_ci BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); 85262306a36Sopenharmony_ci /* 85362306a36Sopenharmony_ci * Calculate the eviction bucket size to cover the longest 85462306a36Sopenharmony_ci * actionable refault distance, which is currently half of 85562306a36Sopenharmony_ci * memory (totalram_pages/2). However, memory hotplug may add 85662306a36Sopenharmony_ci * some more pages at runtime, so keep working with up to 85762306a36Sopenharmony_ci * double the initial memory by using totalram_pages as-is. 85862306a36Sopenharmony_ci */ 85962306a36Sopenharmony_ci timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; 86062306a36Sopenharmony_ci max_order = fls_long(totalram_pages() - 1); 86162306a36Sopenharmony_ci if (max_order > timestamp_bits) 86262306a36Sopenharmony_ci bucket_order = max_order - timestamp_bits; 86362306a36Sopenharmony_ci pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n", 86462306a36Sopenharmony_ci timestamp_bits, max_order, bucket_order); 86562306a36Sopenharmony_ci 86662306a36Sopenharmony_ci ret = prealloc_shrinker(&workingset_shadow_shrinker, "mm-shadow"); 86762306a36Sopenharmony_ci if (ret) 86862306a36Sopenharmony_ci goto err; 86962306a36Sopenharmony_ci ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key, 87062306a36Sopenharmony_ci &workingset_shadow_shrinker); 87162306a36Sopenharmony_ci if (ret) 87262306a36Sopenharmony_ci goto err_list_lru; 87362306a36Sopenharmony_ci register_shrinker_prepared(&workingset_shadow_shrinker); 87462306a36Sopenharmony_ci return 0; 87562306a36Sopenharmony_cierr_list_lru: 87662306a36Sopenharmony_ci free_prealloced_shrinker(&workingset_shadow_shrinker); 87762306a36Sopenharmony_cierr: 87862306a36Sopenharmony_ci return ret; 87962306a36Sopenharmony_ci} 88062306a36Sopenharmony_cimodule_init(workingset_init); 881