ax-task 0.8.1

OS-independent IRQ-safe SMP task scheduling core
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229

//! Root-domain RT period callback and online-rq replenishment.

use super::*;

impl RootDomain {
    pub(in crate::sched::system::task_system) fn rt_bandwidth_enabled(&self) -> bool {
        self.rt_bandwidth.enabled()
    }

    pub(in crate::sched::system::task_system) fn activate_rt_period(
        &self,
        cpu: CpuId,
        sample_now: impl FnOnce() -> MonotonicInstant,
    ) -> bool {
        self.rt_bandwidth.activate(cpu, sample_now)
    }

    /// Runs the Linux `do_sched_rt_period_timer()` transaction.
    pub(super) fn service_rt_period(
        &self,
        system: &TaskSystem,
        cpu: CpuId,
        now: MonotonicInstant,
    ) -> bool {
        let Some(firing) = self.rt_bandwidth.begin_period(cpu, now) else {
            return false;
        };
        let overruns = firing.overruns();
        let mut keep_active = false;
        let mut rescheduled = false;

        for remote in &self.runqueues {
            if !remote.is_online() {
                continue;
            }
            let snapshot = *remote.lock_rt_bandwidth();
            let run_queue = remote.lock_run_queue(RunQueueGuardSource::RtAccounting);
            let runnable = run_queue.has_runnable_rt();
            let throttled = run_queue.rt_is_throttled();
            drop(run_queue);
            // The ledger snapshot is optimistic and precedes the rq read. A
            // concurrent owner may publish throttling between them, so the rq
            // fact must veto the empty-ledger fast path.
            if snapshot.time_ns() == 0 && !runnable && !throttled {
                continue;
            }
            let mut transaction = OwnerRqTxn::begin(system, remote);
            let charged_runtime_active = remote.lock_rt_bandwidth().time_ns() != 0;
            if !charged_runtime_active {
                let runtime_active = transaction.rt_is_throttled();
                let runnable = transaction.has_runnable_rt();
                keep_active |= runtime_active || runnable;
                transaction.commit();
                continue;
            }
            if transaction.rt_is_throttled() {
                // Linux balances a charged, throttled rq before subtracting
                // elapsed periods. A zero rt_time never enters balance_runtime
                // and therefore cannot borrow quota or clear throttling here.
                self.balance_rt_runtime(remote.owner());
            }
            let mut bandwidth = remote.lock_rt_bandwidth();
            let below_runtime = bandwidth.replenish(overruns);
            let charged_runtime_active = bandwidth.time_ns() != 0;
            drop(bandwidth);
            let unthrottled = transaction.rt_is_throttled() && below_runtime;
            if unthrottled {
                transaction.set_rt_throttled(false);
            }
            let runtime_active = charged_runtime_active || transaction.rt_is_throttled();
            let runnable = transaction.has_runnable_rt();
            keep_active |= runtime_active || runnable;
            transaction.commit();
            if !unthrottled || !runnable {
                continue;
            }
            rescheduled = true;
            if remote.owner() == cpu {
                remote.request_reschedule(RescheduleKind::Immediate);
            } else {
                remote.request_remote_reschedule(RescheduleKind::Immediate);
            }
        }

        self.rt_bandwidth.finish_period(firing, keep_active);
        rescheduled
    }

    /// Linux `sched_rt_runtime_exceeded()` plus `do_balance_runtime()`.
    /// The rq caller owns local execution accounting; the root lock is entered
    /// only on the quota edge and serializes transfers among independent
    /// per-rq runtime locks.
    pub(super) fn charge_rt_runtime(
        &self,
        cpu: CpuId,
        runtime_ns: u64,
        already_throttled: bool,
    ) -> bool {
        if !self.rt_bandwidth.enabled() {
            return false;
        }
        let remote = &self.runqueues[cpu.as_usize()];
        if !remote.lock_rt_bandwidth().account(runtime_ns) {
            return false;
        }
        // Linux `sched_rt_runtime_exceeded()` keeps accounting PI-boosted
        // execution on a throttled rq, but returns before `balance_runtime()`.
        // A sticky throttle therefore cannot repeatedly borrow runtime or scan
        // the root-domain span on every later charge.
        if already_throttled {
            return true;
        }
        self.balance_rt_runtime(cpu);
        remote.lock_rt_bandwidth().should_throttle()
    }

    fn balance_rt_runtime(&self, receiver: CpuId) {
        let period_ns = self.rt_bandwidth.period_ns();
        let _root = self.rt_bandwidth.lock_runtime();
        let span_weight = self
            .runqueues
            .iter()
            .filter(|remote| remote.is_online())
            .count();
        if span_weight == 0 {
            return;
        }
        for donor in &self.runqueues {
            if donor.owner() == receiver || !donor.is_online() {
                continue;
            }
            let receiver_runtime = self.runqueues[receiver.as_usize()]
                .lock_rt_bandwidth()
                .runtime_ns();
            let room = period_ns.saturating_sub(receiver_runtime);
            if room == 0 {
                break;
            }
            let amount = {
                let mut runtime = donor.lock_rt_bandwidth();
                if !runtime.enabled() {
                    continue;
                }
                let amount = (runtime.spare_runtime_ns() / span_weight as u64).min(room);
                if amount != 0 {
                    runtime.lend_runtime(amount);
                }
                amount
            };
            if amount != 0 {
                self.runqueues[receiver.as_usize()]
                    .lock_rt_bandwidth()
                    .borrow_runtime(amount, period_ns);
            }
            if self.runqueues[receiver.as_usize()]
                .lock_rt_bandwidth()
                .runtime_ns()
                == period_ns
            {
                break;
            }
        }
    }

    pub(in crate::sched::system::task_system) fn enable_rt_runtime(&self, cpu: CpuId) {
        let remote = &self.runqueues[cpu.as_usize()];
        let mut run_queue = remote.lock_run_queue(RunQueueGuardSource::RtAccounting);
        remote.lock_rt_bandwidth().enable(
            self.rt_bandwidth.period_ns(),
            self.rt_bandwidth.runtime_ns(),
        );
        run_queue.set_rt_throttled(false);
    }

    pub(in crate::sched::system::task_system) fn disable_rt_runtime(&self, cpu: CpuId) {
        let root_guard = self.rt_bandwidth.lock_runtime();
        let base = self.rt_bandwidth.runtime_ns();
        let current = self.runqueues[cpu.as_usize()]
            .lock_rt_bandwidth()
            .runtime_ns();
        let mut want = i128::from(base) - i128::from(current);
        for remote in &self.runqueues {
            if remote.owner() == cpu || !remote.is_online() || want == 0 {
                continue;
            }
            let mut runtime = remote.lock_rt_bandwidth();
            if !runtime.enabled() {
                continue;
            }
            if want > 0 {
                let reclaim = u64::try_from(want)
                    .expect("positive RT reclaim must fit u64")
                    .min(runtime.runtime_ns());
                runtime.adjust_runtime(-i128::from(reclaim));
                want -= i128::from(reclaim);
            } else {
                let returned = u64::try_from(-want).expect("negative RT reclaim must fit u64");
                runtime.adjust_runtime(i128::from(returned));
                want = 0;
            }
        }
        assert_eq!(want, 0, "root-domain RT runtime loan leaked across hotplug");
        self.runqueues[cpu.as_usize()].lock_rt_bandwidth().disable();
        drop(root_guard);
        self.runqueues[cpu.as_usize()]
            .lock_run_queue(RunQueueGuardSource::RtAccounting)
            .set_rt_throttled(false);
    }
}

impl TaskSystem {
    pub(crate) fn rt_bandwidth_enabled(&self) -> bool {
        self.root_domain.rt_bandwidth_enabled()
    }

    pub(crate) fn service_rt_period(&self, cpu: &CpuLocal, now: MonotonicInstant) -> bool {
        self.root_domain.service_rt_period(self, cpu.owner(), now)
    }

    pub(crate) fn charge_rt_runtime(
        &self,
        cpu: CpuId,
        runtime_ns: u64,
        already_throttled: bool,
    ) -> bool {
        self.root_domain
            .charge_rt_runtime(cpu, runtime_ns, already_throttled)
    }
}