1.. SPDX-License-Identifier: GPL-2.0 2 3============= 4Multi-Gen LRU 5============= 6The multi-gen LRU is an alternative LRU implementation that optimizes 7page reclaim and improves performance under memory pressure. Page 8reclaim decides the kernel's caching policy and ability to overcommit 9memory. It directly impacts the kswapd CPU usage and RAM efficiency. 10 11Design overview 12=============== 13Objectives 14---------- 15The design objectives are: 16 17* Good representation of access recency 18* Try to profit from spatial locality 19* Fast paths to make obvious choices 20* Simple self-correcting heuristics 21 22The representation of access recency is at the core of all LRU 23implementations. In the multi-gen LRU, each generation represents a 24group of pages with similar access recency. Generations establish a 25(time-based) common frame of reference and therefore help make better 26choices, e.g., between different memcgs on a computer or different 27computers in a data center (for job scheduling). 28 29Exploiting spatial locality improves efficiency when gathering the 30accessed bit. A rmap walk targets a single page and does not try to 31profit from discovering a young PTE. A page table walk can sweep all 32the young PTEs in an address space, but the address space can be too 33sparse to make a profit. The key is to optimize both methods and use 34them in combination. 35 36Fast paths reduce code complexity and runtime overhead. Unmapped pages 37do not require TLB flushes; clean pages do not require writeback. 38These facts are only helpful when other conditions, e.g., access 39recency, are similar. With generations as a common frame of reference, 40additional factors stand out. But obvious choices might not be good 41choices; thus self-correction is necessary. 42 43The benefits of simple self-correcting heuristics are self-evident. 44Again, with generations as a common frame of reference, this becomes 45attainable. Specifically, pages in the same generation can be 46categorized based on additional factors, and a feedback loop can 47statistically compare the refault percentages across those categories 48and infer which of them are better choices. 49 50Assumptions 51----------- 52The protection of hot pages and the selection of cold pages are based 53on page access channels and patterns. There are two access channels: 54 55* Accesses through page tables 56* Accesses through file descriptors 57 58The protection of the former channel is by design stronger because: 59 601. The uncertainty in determining the access patterns of the former 61 channel is higher due to the approximation of the accessed bit. 622. The cost of evicting the former channel is higher due to the TLB 63 flushes required and the likelihood of encountering the dirty bit. 643. The penalty of underprotecting the former channel is higher because 65 applications usually do not prepare themselves for major page 66 faults like they do for blocked I/O. E.g., GUI applications 67 commonly use dedicated I/O threads to avoid blocking rendering 68 threads. 69 70There are also two access patterns: 71 72* Accesses exhibiting temporal locality 73* Accesses not exhibiting temporal locality 74 75For the reasons listed above, the former channel is assumed to follow 76the former pattern unless ``VM_SEQ_READ`` or ``VM_RAND_READ`` is 77present, and the latter channel is assumed to follow the latter 78pattern unless outlying refaults have been observed. 79 80Workflow overview 81================= 82Evictable pages are divided into multiple generations for each 83``lruvec``. The youngest generation number is stored in 84``lrugen->max_seq`` for both anon and file types as they are aged on 85an equal footing. The oldest generation numbers are stored in 86``lrugen->min_seq[]`` separately for anon and file types as clean file 87pages can be evicted regardless of swap constraints. These three 88variables are monotonically increasing. 89 90Generation numbers are truncated into ``order_base_2(MAX_NR_GENS+1)`` 91bits in order to fit into the gen counter in ``folio->flags``. Each 92truncated generation number is an index to ``lrugen->folios[]``. The 93sliding window technique is used to track at least ``MIN_NR_GENS`` and 94at most ``MAX_NR_GENS`` generations. The gen counter stores a value 95within ``[1, MAX_NR_GENS]`` while a page is on one of 96``lrugen->folios[]``; otherwise it stores zero. 97 98Each generation is divided into multiple tiers. A page accessed ``N`` 99times through file descriptors is in tier ``order_base_2(N)``. Unlike 100generations, tiers do not have dedicated ``lrugen->folios[]``. In 101contrast to moving across generations, which requires the LRU lock, 102moving across tiers only involves atomic operations on 103``folio->flags`` and therefore has a negligible cost. A feedback loop 104modeled after the PID controller monitors refaults over all the tiers 105from anon and file types and decides which tiers from which types to 106evict or protect. The desired effect is to balance refault percentages 107between anon and file types proportional to the swappiness level. 108 109There are two conceptually independent procedures: the aging and the 110eviction. They form a closed-loop system, i.e., the page reclaim. 111 112Aging 113----- 114The aging produces young generations. Given an ``lruvec``, it 115increments ``max_seq`` when ``max_seq-min_seq+1`` approaches 116``MIN_NR_GENS``. The aging promotes hot pages to the youngest 117generation when it finds them accessed through page tables; the 118demotion of cold pages happens consequently when it increments 119``max_seq``. The aging uses page table walks and rmap walks to find 120young PTEs. For the former, it iterates ``lruvec_memcg()->mm_list`` 121and calls ``walk_page_range()`` with each ``mm_struct`` on this list 122to scan PTEs, and after each iteration, it increments ``max_seq``. For 123the latter, when the eviction walks the rmap and finds a young PTE, 124the aging scans the adjacent PTEs. For both, on finding a young PTE, 125the aging clears the accessed bit and updates the gen counter of the 126page mapped by this PTE to ``(max_seq%MAX_NR_GENS)+1``. 127 128Eviction 129-------- 130The eviction consumes old generations. Given an ``lruvec``, it 131increments ``min_seq`` when ``lrugen->folios[]`` indexed by 132``min_seq%MAX_NR_GENS`` becomes empty. To select a type and a tier to 133evict from, it first compares ``min_seq[]`` to select the older type. 134If both types are equally old, it selects the one whose first tier has 135a lower refault percentage. The first tier contains single-use 136unmapped clean pages, which are the best bet. The eviction sorts a 137page according to its gen counter if the aging has found this page 138accessed through page tables and updated its gen counter. It also 139moves a page to the next generation, i.e., ``min_seq+1``, if this page 140was accessed multiple times through file descriptors and the feedback 141loop has detected outlying refaults from the tier this page is in. To 142this end, the feedback loop uses the first tier as the baseline, for 143the reason stated earlier. 144 145Working set protection 146---------------------- 147Each generation is timestamped at birth. If ``lru_gen_min_ttl`` is 148set, an ``lruvec`` is protected from the eviction when its oldest 149generation was born within ``lru_gen_min_ttl`` milliseconds. In other 150words, it prevents the working set of ``lru_gen_min_ttl`` milliseconds 151from getting evicted. The OOM killer is triggered if this working set 152cannot be kept in memory. 153 154This time-based approach has the following advantages: 155 1561. It is easier to configure because it is agnostic to applications 157 and memory sizes. 1582. It is more reliable because it is directly wired to the OOM killer. 159 160``mm_struct`` list 161------------------ 162An ``mm_struct`` list is maintained for each memcg, and an 163``mm_struct`` follows its owner task to the new memcg when this task 164is migrated. 165 166A page table walker iterates ``lruvec_memcg()->mm_list`` and calls 167``walk_page_range()`` with each ``mm_struct`` on this list to scan 168PTEs. When multiple page table walkers iterate the same list, each of 169them gets a unique ``mm_struct``, and therefore they can run in 170parallel. 171 172Page table walkers ignore any misplaced pages, e.g., if an 173``mm_struct`` was migrated, pages left in the previous memcg will be 174ignored when the current memcg is under reclaim. Similarly, page table 175walkers will ignore pages from nodes other than the one under reclaim. 176 177This infrastructure also tracks the usage of ``mm_struct`` between 178context switches so that page table walkers can skip processes that 179have been sleeping since the last iteration. 180 181Rmap/PT walk feedback 182--------------------- 183Searching the rmap for PTEs mapping each page on an LRU list (to test 184and clear the accessed bit) can be expensive because pages from 185different VMAs (PA space) are not cache friendly to the rmap (VA 186space). For workloads mostly using mapped pages, searching the rmap 187can incur the highest CPU cost in the reclaim path. 188 189``lru_gen_look_around()`` exploits spatial locality to reduce the 190trips into the rmap. It scans the adjacent PTEs of a young PTE and 191promotes hot pages. If the scan was done cacheline efficiently, it 192adds the PMD entry pointing to the PTE table to the Bloom filter. This 193forms a feedback loop between the eviction and the aging. 194 195Bloom filters 196------------- 197Bloom filters are a space and memory efficient data structure for set 198membership test, i.e., test if an element is not in the set or may be 199in the set. 200 201In the eviction path, specifically, in ``lru_gen_look_around()``, if a 202PMD has a sufficient number of hot pages, its address is placed in the 203filter. In the aging path, set membership means that the PTE range 204will be scanned for young pages. 205 206Note that Bloom filters are probabilistic on set membership. If a test 207is false positive, the cost is an additional scan of a range of PTEs, 208which may yield hot pages anyway. Parameters of the filter itself can 209control the false positive rate in the limit. 210 211PID controller 212-------------- 213A feedback loop modeled after the Proportional-Integral-Derivative 214(PID) controller monitors refaults over anon and file types and 215decides which type to evict when both types are available from the 216same generation. 217 218The PID controller uses generations rather than the wall clock as the 219time domain because a CPU can scan pages at different rates under 220varying memory pressure. It calculates a moving average for each new 221generation to avoid being permanently locked in a suboptimal state. 222 223Memcg LRU 224--------- 225An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs, 226since each node and memcg combination has an LRU of folios (see 227``mem_cgroup_lruvec()``). Its goal is to improve the scalability of 228global reclaim, which is critical to system-wide memory overcommit in 229data centers. Note that memcg LRU only applies to global reclaim. 230 231The basic structure of an memcg LRU can be understood by an analogy to 232the active/inactive LRU (of folios): 233 2341. It has the young and the old (generations), i.e., the counterparts 235 to the active and the inactive; 2362. The increment of ``max_seq`` triggers promotion, i.e., the 237 counterpart to activation; 2383. Other events trigger similar operations, e.g., offlining an memcg 239 triggers demotion, i.e., the counterpart to deactivation. 240 241In terms of global reclaim, it has two distinct features: 242 2431. Sharding, which allows each thread to start at a random memcg (in 244 the old generation) and improves parallelism; 2452. Eventual fairness, which allows direct reclaim to bail out at will 246 and reduces latency without affecting fairness over some time. 247 248In terms of traversing memcgs during global reclaim, it improves the 249best-case complexity from O(n) to O(1) and does not affect the 250worst-case complexity O(n). Therefore, on average, it has a sublinear 251complexity. 252 253Summary 254------- 255The multi-gen LRU (of folios) can be disassembled into the following 256parts: 257 258* Generations 259* Rmap walks 260* Page table walks via ``mm_struct`` list 261* Bloom filters for rmap/PT walk feedback 262* PID controller for refault feedback 263 264The aging and the eviction form a producer-consumer model; 265specifically, the latter drives the former by the sliding window over 266generations. Within the aging, rmap walks drive page table walks by 267inserting hot densely populated page tables to the Bloom filters. 268Within the eviction, the PID controller uses refaults as the feedback 269to select types to evict and tiers to protect. 270