xref: /DragonOS/kernel/src/sched/pelt.rs (revision 4b0170bd6bb374d0e9699a0076cc23b976ad6db7)
1 use core::intrinsics::unlikely;
2 
3 use alloc::sync::Arc;
4 
5 use crate::process::ProcessControlBlock;
6 
7 use super::{
8     fair::{CfsRunQueue, FairSchedEntity},
9     CpuRunQueue, LoadWeight, SchedPolicy, SCHED_CAPACITY_SCALE, SCHED_CAPACITY_SHIFT,
10 };
11 
12 const RUNNABLE_AVG_Y_N_INV: [u32; 32] = [
13     0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, 0xe0ccdeeb, 0xdbfbb796,
14     0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46,
15     0xb504f333, 0xb123f581, 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
16     0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, 0x85aac367, 0x82cd8698,
17 ];
18 
19 pub const LOAD_AVG_PERIOD: u64 = 32;
20 pub const LOAD_AVG_MAX: usize = 47742;
21 pub const PELT_MIN_DIVIDER: usize = LOAD_AVG_MAX - 1024;
22 
23 #[derive(Debug, Default)]
24 pub struct SchedulerAvg {
25     /// 存储上次更新这些平均值的时间
26     pub last_update_time: u64,
27     /// 存储所有可运行任务的负载之和
28     pub load_sum: u64,
29     /// 存储所有可运行任务的时间之和
30     pub runnable_sum: u64,
31     /// 存储所有运行任务的时间之和
32     pub util_sum: u64,
33     /// 记录周期性任务的贡献值,用于计算平均CPU利用率
34     pub period_contrib: u32,
35 
36     pub load_avg: usize,
37     pub runnable_avg: usize,
38     pub util_avg: usize,
39 }
40 
41 impl SchedulerAvg {
42     #[inline]
43     pub fn get_pelt_divider(&self) -> usize {
44         return PELT_MIN_DIVIDER + self.period_contrib as usize;
45     }
46 
47     pub fn update_load_sum(
48         &mut self,
49         now: u64,
50         load: u32,
51         mut runnable: u32,
52         mut running: u32,
53     ) -> bool {
54         if now < self.last_update_time {
55             self.last_update_time = now;
56             return false;
57         }
58 
59         let mut delta = now - self.last_update_time;
60         delta >>= 10;
61 
62         if delta == 0 {
63             return false;
64         }
65 
66         self.last_update_time += delta << 10;
67 
68         if load == 0 {
69             runnable = 0;
70             running = 0;
71         }
72 
73         self.accumulate_sum(delta, load, runnable, running) != 0
74     }
75 
76     pub fn accumulate_sum(
77         &mut self,
78         mut delta: u64,
79         load: u32,
80         runnable: u32,
81         running: u32,
82     ) -> u64 {
83         let mut contrib = delta as u32;
84 
85         delta += self.period_contrib as u64;
86 
87         let periods = delta / 1024;
88 
89         if periods > 0 {
90             self.load_sum = Self::decay_load(self.load_sum, periods);
91             self.runnable_sum = Self::decay_load(self.runnable_sum, periods);
92             self.util_sum = Self::decay_load(self.util_sum, periods);
93 
94             delta %= 1024;
95             if load > 0 {
96                 contrib = Self::accumulate_pelt_segments(
97                     periods,
98                     1024 - self.period_contrib,
99                     delta as u32,
100                 );
101             }
102         }
103 
104         self.period_contrib = delta as u32;
105 
106         if load > 0 {
107             self.load_sum += (contrib * load) as u64;
108         }
109         if runnable > 0 {
110             self.runnable_sum += (runnable & contrib << SCHED_CAPACITY_SHIFT) as u64;
111         }
112 
113         if running > 0 {
114             self.util_sum += (contrib << SCHED_CAPACITY_SHIFT) as u64;
115         }
116 
117         return periods;
118     }
119 
120     fn decay_load(mut val: u64, n: u64) -> u64 {
121         if unlikely(n > LOAD_AVG_PERIOD) {
122             return 0;
123         }
124 
125         let mut local_n = n;
126 
127         if unlikely(local_n >= LOAD_AVG_PERIOD) {
128             val >>= local_n / LOAD_AVG_PERIOD;
129             local_n %= LOAD_AVG_PERIOD;
130         }
131 
132         ((val as i128 * RUNNABLE_AVG_Y_N_INV[local_n as usize] as i128) >> 32) as u64
133     }
134 
135     fn accumulate_pelt_segments(periods: u64, d1: u32, d3: u32) -> u32 {
136         /* y^0 == 1 */
137         let c3 = d3;
138 
139         /*
140          * c1 = d1 y^p
141          */
142         let c1 = Self::decay_load(d1 as u64, periods) as u32;
143 
144         /*
145          *            p-1
146          * c2 = 1024 \Sum y^n
147          *            n=1
148          *
149          *              inf        inf
150          *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
151          *              n=0        n=p
152          */
153         let c2 = LOAD_AVG_MAX as u32 - Self::decay_load(LOAD_AVG_MAX as u64, periods) as u32 - 1024;
154 
155         return c1 + c2 + c3;
156     }
157 
158     pub fn update_load_avg(&mut self, load: u64) {
159         let divider = self.get_pelt_divider();
160 
161         self.load_avg = (load * self.load_sum) as usize / divider;
162         self.runnable_avg = self.runnable_sum as usize / divider;
163         self.util_avg = self.util_sum as usize / divider;
164     }
165 
166     #[allow(dead_code)]
167     pub fn post_init_entity_util_avg(pcb: &Arc<ProcessControlBlock>) {
168         let se = pcb.sched_info().sched_entity();
169         let cfs_rq = se.cfs_rq();
170         let sa = &mut se.force_mut().avg;
171 
172         // TODO: 这里和架构相关
173         let cpu_scale = SCHED_CAPACITY_SCALE;
174 
175         let cap = (cpu_scale as isize - cfs_rq.avg.util_avg as isize) / 2;
176 
177         if pcb.sched_info().policy() != SchedPolicy::CFS {
178             sa.last_update_time = cfs_rq.cfs_rq_clock_pelt();
179         }
180 
181         if cap > 0 {
182             if cfs_rq.avg.util_avg != 0 {
183                 sa.util_avg = cfs_rq.avg.util_avg * se.load.weight as usize;
184                 sa.util_avg /= cfs_rq.avg.load_avg + 1;
185 
186                 if sa.util_avg as isize > cap {
187                     sa.util_avg = cap as usize;
188                 }
189             } else {
190                 sa.util_avg = cap as usize;
191             }
192         }
193 
194         sa.runnable_avg = sa.util_avg;
195     }
196 }
197 
198 impl CpuRunQueue {
199     pub fn rq_clock_pelt(&self) -> u64 {
200         self.clock_pelt - self.lost_idle_time
201     }
202 }
203 
204 impl CfsRunQueue {
205     pub fn cfs_rq_clock_pelt(&self) -> u64 {
206         if unlikely(self.throttled_count > 0) {
207             return self.throttled_clock_pelt - self.throttled_clock_pelt_time;
208         }
209 
210         let rq = self.rq();
211         let (rq, _guard) = rq.self_lock();
212 
213         return rq.rq_clock_pelt() - self.throttled_clock_pelt_time;
214     }
215 }
216 
217 impl FairSchedEntity {
218     pub fn update_load_avg(&mut self, cfs_rq: &mut CfsRunQueue, now: u64) -> bool {
219         if self.avg.update_load_sum(
220             now,
221             self.on_rq as u32,
222             self.runnable() as u32,
223             cfs_rq.is_curr(&self.self_arc()) as u32,
224         ) {
225             self.avg
226                 .update_load_avg(LoadWeight::scale_load_down(self.load.weight));
227 
228             return true;
229         }
230 
231         return false;
232     }
233 }
234 
235 bitflags! {
236     pub struct UpdateAvgFlags: u8 {
237         /// 更新任务组(task group)信息
238         const UPDATE_TG	= 0x1;
239 
240         /// 跳过年龄和负载的更新
241         const SKIP_AGE_LOAD	= 0x2;
242         /// 执行附加操作
243         const DO_ATTACH	= 0x4;
244         /// 执行分离操作
245         const DO_DETACH	= 0x8;
246     }
247 }
248 
249 pub fn add_positive(x: &mut isize, y: isize) {
250     let res = *x + y;
251     *x = res.max(0);
252 }
253 
254 pub fn sub_positive(x: &mut usize, y: usize) {
255     if *x > y {
256         *x -= y;
257     } else {
258         *x = 0;
259     }
260 }
261