Skip to main content

rusty_rtos_kernel_core/
system.rs

1//! The static face: declare the system, and the geometry is the declaration.
2//!
3//! [`crate::typed`] made one *object* safe to hold — a `Queue<T, N>` that
4//! moves `T`s and cannot be minted from an integer. This makes the whole
5//! *system* safe to declare. You write down the tasks and queues an
6//! application has; the macro writes the kernel's geometry from them and
7//! hands back a struct whose fields are those tasks and queues, by name and
8//! by type.
9//!
10//! ```ignore
11//! rusty_rtos_kernel_core::system! {
12//!     mod blinky use PosixDemoConfig;
13//!     tasks { consumer: 2, producer: 2 }
14//!     queues { data: u16; 10 }
15//! }
16//! ```
17//!
18//! # What that buys, and why each part of it is a compile-time thing
19//!
20//! **Exact `.bss` by construction.** The geometry const generics —
21//! `TASKS`, `ITEMS`, `LISTS`, `QUEUES`, `SLOTS` — are *computed from the
22//! declaration* rather than picked by hand and rounded up. `SLOTS` is the
23//! sum of the declared queue lengths, not a guess; `TASKS` is what you
24//! declared plus the two the kernel makes for itself. A kernel sized this
25//! way holds exactly the system and nothing else, and
26//! `blinky::Kernel::<..>::SIZE` is a number you can put in a footprint
27//! table rather than an allowance you hope covers it.
28//!
29//! **No create-failure path an application can reach.** Every way
30//! `queue_create` and `create_task` can answer `Err` is eliminated *before
31//! the program runs*:
32//!
33//! | how a create fails | why it cannot here |
34//! |---|---|
35//! | the task arena is full | `TASKS` **is** the declared count plus the kernel's two |
36//! | the queue arena is full | `QUEUES` **is** the declared count |
37//! | the item slots are exhausted | `SLOTS` **is** the sum of the declared lengths |
38//! | a list is missing | `LISTS` is `lists_for` over the same numbers |
39//! | the priority is out of range | a `const` assertion per task, so it is a **compile error** |
40//!
41//! [`System::build`] still answers `Result`, because the kernel calls it
42//! makes do and this crate may not panic — but the `Err` arm is now
43//! unreachable by construction rather than merely unlikely, and the table
44//! above says which construction makes each one unreachable. That is the
45//! difference between a runtime check and a proof, and it is the whole
46//! point of the exercise.
47//!
48//! **The handles never enter the application.** `build` is the only place
49//! that touches a create call, and it happens once, at startup. Afterwards
50//! the application holds `system.data` — a `Queue<u16, 10>` — and there is
51//! no constructor anywhere that invents one from an integer. That is what
52//! makes the four Kani harnesses that pass a *symbolic handle* into a
53//! kernel call unwritable against this face: the state they explore has no
54//! way to come into being (`docs/LEDGER.md`, "The topology work,
55//! measured").
56
57/// Declare a system: its tasks, its queues, and the geometry they imply.
58///
59/// See the [module docs](self) for what the expansion contains and why.
60///
61/// The config is named in the header because the number of lists depends
62/// on `Config::MAX_PRIORITIES`, and a `const` may read an associated const
63/// of a *concrete* type — which is why this is a declaration and not a
64/// generic parameter.
65///
66/// # What a declaration refuses, that a hand-sized kernel cannot
67///
68/// Each of these is paired with the working line it is one character
69/// from, so neither can pass for the wrong reason.
70///
71/// **A priority the config has no ready list for.** In C this is a
72/// `configASSERT` on the bench, if `configASSERT` is compiled in, on the
73/// day that task first runs. Here it is a build failure:
74///
75/// ```compile_fail
76/// # use rusty_rtos_core::config::Config;
77/// # use rusty_rtos_core::tick::Bits32;
78/// # #[derive(Debug, Clone, Copy, Default)]
79/// # pub struct Cfg;
80/// # impl Config for Cfg {
81/// #     type Tick = Bits32;
82/// #     const TICK_RATE_HZ: u32 = 100;
83/// #     const MAX_PRIORITIES: u8 = 4;
84/// #     const MINIMAL_STACK_SIZE: usize = 1;
85/// #     const MAX_TASK_NAME_LEN: usize = 8;
86/// #     const TIMER_TASK_PRIORITY: u8 = 3;
87/// #     const TIMER_TASK_STACK_DEPTH: usize = 1;
88/// #     const TIMER_QUEUE_LENGTH: usize = 1;
89/// #     const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
90/// # }
91/// rusty_rtos_kernel_core::system! {
92///     mod bad use Cfg;
93///     tasks { spinner: 4 }
94///     queues { }
95/// }
96/// # fn main() {}
97/// ```
98///
99/// The control arm — the same declaration, one priority lower, which is
100/// the highest `MAX_PRIORITIES = 4` actually has:
101///
102/// ```
103/// # use rusty_rtos_core::config::Config;
104/// # use rusty_rtos_core::tick::Bits32;
105/// # #[derive(Debug, Clone, Copy, Default)]
106/// # pub struct Cfg;
107/// # impl Config for Cfg {
108/// #     type Tick = Bits32;
109/// #     const TICK_RATE_HZ: u32 = 100;
110/// #     const MAX_PRIORITIES: u8 = 4;
111/// #     const MINIMAL_STACK_SIZE: usize = 1;
112/// #     const MAX_TASK_NAME_LEN: usize = 8;
113/// #     const TIMER_TASK_PRIORITY: u8 = 3;
114/// #     const TIMER_TASK_STACK_DEPTH: usize = 1;
115/// #     const TIMER_QUEUE_LENGTH: usize = 1;
116/// #     const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
117/// # }
118/// rusty_rtos_kernel_core::system! {
119///     mod good use Cfg;
120///     tasks { spinner: 3 }
121///     queues { }
122/// }
123/// # fn main() { assert_eq!(good::TASKS, 3); }
124/// ```
125///
126/// **A queue told to carry what it was not declared for.** The
127/// declaration named `data` as a queue of `u16`, so the field is a
128/// `Queue<u16, 4>` and a `u32` is not a `u16`:
129///
130/// ```compile_fail
131/// # use rusty_rtos_core::config::Config;
132/// # use rusty_rtos_core::tick::Bits32;
133/// # #[derive(Debug, Clone, Copy, Default)]
134/// # pub struct Cfg;
135/// # impl Config for Cfg {
136/// #     type Tick = Bits32;
137/// #     const TICK_RATE_HZ: u32 = 100;
138/// #     const MAX_PRIORITIES: u8 = 4;
139/// #     const MINIMAL_STACK_SIZE: usize = 1;
140/// #     const MAX_TASK_NAME_LEN: usize = 8;
141/// #     const TIMER_TASK_PRIORITY: u8 = 3;
142/// #     const TIMER_TASK_STACK_DEPTH: usize = 1;
143/// #     const TIMER_QUEUE_LENGTH: usize = 1;
144/// #     const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
145/// # }
146/// rusty_rtos_kernel_core::system! {
147///     mod app use Cfg;
148///     tasks { worker: 1 }
149///     queues { data: u16; 4 }
150/// }
151/// # fn main() {}
152/// fn wrong<P, T, H>(k: &mut app::Kernel<P, T, H>, s: &mut app::System)
153/// where
154///     P: rusty_rtos_core::port::Port,
155///     T: rusty_rtos_core::trace::Trace,
156///     H: rusty_rtos_kernel_core::TickHook<app::Kernel<P, T, H>>,
157/// {
158///     let _ = s.data.send(k, 70_000u32, 0);
159/// }
160/// ```
161///
162/// The control arm, one type name different:
163///
164/// ```
165/// # use rusty_rtos_core::config::Config;
166/// # use rusty_rtos_core::tick::Bits32;
167/// # #[derive(Debug, Clone, Copy, Default)]
168/// # pub struct Cfg;
169/// # impl Config for Cfg {
170/// #     type Tick = Bits32;
171/// #     const TICK_RATE_HZ: u32 = 100;
172/// #     const MAX_PRIORITIES: u8 = 4;
173/// #     const MINIMAL_STACK_SIZE: usize = 1;
174/// #     const MAX_TASK_NAME_LEN: usize = 8;
175/// #     const TIMER_TASK_PRIORITY: u8 = 3;
176/// #     const TIMER_TASK_STACK_DEPTH: usize = 1;
177/// #     const TIMER_QUEUE_LENGTH: usize = 1;
178/// #     const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
179/// # }
180/// rusty_rtos_kernel_core::system! {
181///     mod app use Cfg;
182///     tasks { worker: 1 }
183///     queues { data: u16; 4 }
184/// }
185/// # fn main() {}
186/// fn right<P, T, H>(k: &mut app::Kernel<P, T, H>, s: &mut app::System)
187/// where
188///     P: rusty_rtos_core::port::Port,
189///     T: rusty_rtos_core::trace::Trace,
190///     H: rusty_rtos_kernel_core::TickHook<app::Kernel<P, T, H>>,
191/// {
192///     let _ = s.data.send(k, 7_000u16, 0);
193/// }
194/// ```
195///
196/// **And there is no third way in.** A `System` has no public
197/// constructor but [`System::build`], and a `Queue<T, N>` has none but
198/// `create`, so no application can hold a handle it invented — which is
199/// what makes the symbolic-handle proofs unwritable against this face
200/// rather than merely slow.
201#[macro_export]
202macro_rules! system {
203    (
204        $(#[$meta:meta])*
205        $vis:vis mod $name:ident use $cfg:ty;
206        tasks { $($task:ident : $prio:expr),* $(,)? }
207        queues { $($queue:ident : $qty:ty ; $qlen:expr),* $(,)? }
208    ) => {
209        $(#[$meta])*
210        $vis mod $name {
211            #![allow(unused_imports, clippy::allow_attributes)]
212            use super::*;
213
214            /// The tasks this system declares. The kernel adds its own two.
215            pub const APP_TASKS: usize = 0usize $(+ { let _ = $prio; 1usize })*;
216            /// Room for every declared task plus the idle task and the
217            /// timer daemon, and for nothing else.
218            pub const TASKS: usize = APP_TASKS + $crate::OVERHEAD_TASKS;
219            /// The declared queues, and no spare.
220            pub const QUEUES: usize = 0usize $(+ { let _ = $qlen; 1usize })*;
221            /// The sum of the declared queue lengths: every slot is spoken
222            /// for and none is spare.
223            pub const SLOTS: usize = 0usize $(+ $qlen)*;
224            /// Software timers. Not declared yet; a system that wants them
225            /// sizes its own kernel.
226            pub const TIMERS: usize = 0;
227            /// Event groups, as `TIMERS`.
228            pub const GROUPS: usize = 0;
229            /// Stream buffers, as `TIMERS`.
230            pub const BUFFERS: usize = 0;
231            /// Stream-buffer bytes, as `TIMERS`.
232            pub const BYTES: usize = 0;
233            /// A ready list per priority, the fixed six, two per queue and
234            /// one per event group.
235            ///
236            /// Declared BEFORE `ITEMS`, which now depends on it: the list
237            /// arena holds one end-marker node per list alongside the items.
238            pub const LISTS: usize = $crate::lists_for(
239                <$cfg as ::rusty_rtos_core::config::Config>::MAX_PRIORITIES,
240                QUEUES,
241                GROUPS,
242            );
243            /// Slots for two list items per task, one per timer, and one
244            /// end marker per list — rounded up to a power of two.
245            pub const ITEMS: usize = $crate::list_slots_for(TASKS, TIMERS, LISTS);
246            /// The timer daemon's mailbox, taken from the config rather than
247            /// from a hardcoded ceiling.
248            ///
249            /// It was `MAX_TIMER_COMMANDS`, a fixed 32, which scaled with
250            /// nothing and so cost the same 768 bytes in a two-task blinker as
251            /// in a twenty-four-task corpus — 39 % of the former's whole
252            /// static footprint. A declared system gets what it declared.
253            pub const TIMER_CMDS: usize =
254                <$cfg as ::rusty_rtos_core::config::Config>::TIMER_QUEUE_LENGTH;
255
256            // A declared priority the config has no ready list for is a
257            // compile error, not a create that answers `Err` at startup on
258            // the bench. This is the one create-failure path that is an
259            // application mistake rather than a geometry mistake, so it is
260            // the one worth refusing here.
261            $(
262                const _: () = assert!(
263                    ($prio as u8) < <$cfg as ::rusty_rtos_core::config::Config>::MAX_PRIORITIES,
264                    concat!(
265                        "task `", stringify!($task),
266                        "` declares a priority its Config has no ready list for"
267                    ),
268                );
269            )*
270
271            /// This system's kernel: the geometry above, and nothing to
272            /// choose. `P`, `T` and `H` are still the port, the trace sink
273            /// and the tick hook, which are the platform's business.
274            pub type Kernel<P, T, H> = $crate::Kernel<
275                $cfg, P, T, H,
276                TASKS, ITEMS, LISTS, QUEUES, SLOTS, BUFFERS, BYTES, TIMERS, GROUPS,
277                TIMER_CMDS,
278            >;
279
280            /// The declared system: every task and every queue, by name.
281            ///
282            /// Built once, by [`System::build`]. After that an application
283            /// holds this and never a handle it could have invented.
284            #[derive(Debug)]
285            pub struct System {
286                $(
287                    /// A declared task.
288                    pub $task: ::rusty_rtos_core::handle::TaskHandle,
289                )*
290                $(
291                    /// A declared queue, typed and length-checked.
292                    pub $queue: $crate::typed::Queue<$qty, { $qlen }>,
293                )*
294            }
295
296            impl System {
297                /// Create everything the declaration names, in the order it
298                /// names it.
299                ///
300                /// # Errors
301                /// None an application can reach: see the module docs for
302                /// why each create-failure path is closed by construction.
303                /// The `Result` is here because this crate may not panic,
304                /// not because there is a case to handle.
305                pub fn build<P, T, H>(kernel: &mut Kernel<P, T, H>) -> ::rusty_rtos_core::error::Result<Self>
306                where
307                    P: ::rusty_rtos_core::port::Port,
308                    T: ::rusty_rtos_core::trace::Trace,
309                    H: $crate::TickHook<Kernel<P, T, H>>,
310                {
311                    $(
312                        let $queue = $crate::typed::Queue::<$qty, { $qlen }>::create(kernel)?;
313                    )*
314                    $(
315                        let $task = kernel.create_task(stringify!($task), $prio)?;
316                    )*
317                    Ok(Self { $($task,)* $($queue,)* })
318                }
319            }
320        }
321    };
322}
323
324#[cfg(test)]
325// `dead_code`: a `system!` declaration generates a `System` and its
326// `build`, and `deep`'s exists to SIZE a kernel rather than to furnish one
327// -- its tasks are created directly so the declared queue slot stays free
328// for the timer daemon.
329#[allow(clippy::unwrap_used, clippy::expect_used, clippy::panic, dead_code)]
330// `pub(crate)` so `timer.rs`, `stream.rs` and `events.rs` can test against
331// the same config, port and sink rather than declaring three more of each
332// -- three copies of a test harness is three places for it to drift (H3).
333pub(crate) mod tests {
334    use rusty_rtos_core::config::Config;
335    use rusty_rtos_core::hooks::NoTickHook;
336    use rusty_rtos_core::isr::Woken;
337    use rusty_rtos_core::port::Port;
338    use rusty_rtos_core::tick::Bits32;
339    use rusty_rtos_core::trace::{Event, Scheduling, Trace};
340
341    use crate::typed::Sent;
342
343    /// Four priorities, so a declaration can get one wrong.
344    #[derive(Debug, Clone, Copy, Default)]
345    pub struct TestConfig;
346
347    impl Config for TestConfig {
348        type Tick = Bits32;
349        const TICK_RATE_HZ: u32 = 100;
350        const MAX_PRIORITIES: u8 = 4;
351        const MINIMAL_STACK_SIZE: usize = 1;
352        const MAX_TASK_NAME_LEN: usize = 8;
353        const TIMER_TASK_PRIORITY: u8 = 3;
354        const TIMER_TASK_STACK_DEPTH: usize = 1;
355        const TIMER_QUEUE_LENGTH: usize = 1;
356        const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
357        const USE_TIME_SLICING: bool = false;
358    }
359
360    /// [`TestConfig`] with `configUSE_TICKLESS_IDLE` on, and nothing else
361    /// changed — so a difference between the two is the tickless path and
362    /// cannot be anything else.
363    pub struct TicklessConfig;
364
365    impl Config for TicklessConfig {
366        type Tick = Bits32;
367        const TICK_RATE_HZ: u32 = 100;
368        const MAX_PRIORITIES: u8 = 4;
369        const MINIMAL_STACK_SIZE: usize = 1;
370        const MAX_TASK_NAME_LEN: usize = 8;
371        const TIMER_TASK_PRIORITY: u8 = 3;
372        const TIMER_TASK_STACK_DEPTH: usize = 1;
373        const TIMER_QUEUE_LENGTH: usize = 1;
374        const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
375        const USE_TIME_SLICING: bool = false;
376        const USE_TICKLESS_IDLE: bool = true;
377    }
378
379    /// A port that takes every whole window it is offered — the most
380    /// aggressive sleep there is, and so the hardest case for the kernel to
381    /// stay correct across.
382    #[derive(Debug, Default)]
383    pub struct SleepyPort {
384        nesting: core::cell::Cell<u32>,
385        /// Ticks this port has been asked to sleep, in total.
386        pub slept: core::cell::Cell<u64>,
387        /// Times it was asked at all.
388        pub sleeps: core::cell::Cell<u64>,
389    }
390
391    impl Port for SleepyPort {
392        fn yield_now(&self) {}
393        fn yield_from_isr(&self, _woken: Woken) {}
394        fn enter_critical(&self) {
395            self.nesting.set(self.nesting.get().saturating_add(1));
396        }
397        fn exit_critical(&self) {
398            self.nesting.set(self.nesting.get().saturating_sub(1));
399        }
400        fn set_interrupt_mask_from_isr(&self) -> u32 {
401            0
402        }
403        fn clear_interrupt_mask_from_isr(&self, _saved: u32) {}
404        fn in_isr(&self) -> bool {
405            false
406        }
407        fn set_in_tick_entry(&self, _yes: bool) {}
408
409        fn suppress_ticks_and_sleep(&self, expected_idle_ticks: u64) -> u64 {
410            self.sleeps.set(self.sleeps.get().saturating_add(1));
411            self.slept
412                .set(self.slept.get().saturating_add(expected_idle_ticks));
413            expected_idle_ticks
414        }
415    }
416
417    /// A port that OVERSLEEPS, to prove the kernel clamps rather than
418    /// trusting it.
419    #[derive(Debug, Default)]
420    pub struct OversleepingPort {
421        nesting: core::cell::Cell<u32>,
422    }
423
424    impl Port for OversleepingPort {
425        fn yield_now(&self) {}
426        fn yield_from_isr(&self, _woken: Woken) {}
427        fn enter_critical(&self) {
428            self.nesting.set(self.nesting.get().saturating_add(1));
429        }
430        fn exit_critical(&self) {
431            self.nesting.set(self.nesting.get().saturating_sub(1));
432        }
433        fn set_interrupt_mask_from_isr(&self) -> u32 {
434            0
435        }
436        fn clear_interrupt_mask_from_isr(&self, _saved: u32) {}
437        fn in_isr(&self) -> bool {
438            false
439        }
440        fn set_in_tick_entry(&self, _yes: bool) {}
441
442        fn suppress_ticks_and_sleep(&self, expected_idle_ticks: u64) -> u64 {
443            expected_idle_ticks.saturating_mul(4).saturating_add(9)
444        }
445    }
446
447    #[derive(Debug, Default)]
448    pub struct TestPort {
449        nesting: core::cell::Cell<u32>,
450    }
451
452    impl Port for TestPort {
453        fn yield_now(&self) {}
454        fn yield_from_isr(&self, _woken: Woken) {}
455        fn enter_critical(&self) {
456            self.nesting.set(self.nesting.get().saturating_add(1));
457        }
458        fn exit_critical(&self) {
459            self.nesting.set(self.nesting.get().saturating_sub(1));
460        }
461        fn set_interrupt_mask_from_isr(&self) -> u32 {
462            0
463        }
464        fn clear_interrupt_mask_from_isr(&self, _saved: u32) {}
465        fn in_isr(&self) -> bool {
466            false
467        }
468        fn set_in_tick_entry(&self, _yes: bool) {}
469    }
470
471    /// A [`TestPort`] that COUNTS the switches it was asked for.
472    ///
473    /// `TestPort::yield_now` is empty, which is right for tests about what
474    /// the kernel computes and useless for a test about whether it asked
475    /// for a switch at all. Those are different questions and the second
476    /// one has no other instrument.
477    #[derive(Debug, Default)]
478    pub struct CountingPort {
479        nesting: core::cell::Cell<u32>,
480        yields: core::cell::Cell<u32>,
481    }
482
483    impl Port for CountingPort {
484        /// Count the kernel ASKING for a switch, not one style of taking it.
485        ///
486        /// The first version of this counted `yield_now`, and measured the
487        /// wrong thing: `Port::COMMITS_SWITCH` defaults to `false`, so
488        /// `port_yield` calls `switch_context` itself and never reaches
489        /// `yield_now` at all. The test read zero and looked like a kernel
490        /// defect. `count_yield` is on the unconditional path, so it
491        /// answers "did the kernel decide a switch was needed" whichever
492        /// way the port takes it — which is the question.
493        fn count_yield(&self) {
494            self.yields.set(self.yields.get().saturating_add(1));
495        }
496        fn yield_now(&self) {}
497        fn yield_from_isr(&self, _woken: Woken) {}
498        fn enter_critical(&self) {
499            self.nesting.set(self.nesting.get().saturating_add(1));
500        }
501        fn exit_critical(&self) {
502            self.nesting.set(self.nesting.get().saturating_sub(1));
503        }
504        fn set_interrupt_mask_from_isr(&self) -> u32 {
505            0
506        }
507        fn clear_interrupt_mask_from_isr(&self, _saved: u32) {}
508        fn in_isr(&self) -> bool {
509            false
510        }
511        fn set_in_tick_entry(&self, _yes: bool) {}
512    }
513
514    #[derive(Debug, Default)]
515    pub struct NoTrace;
516
517    impl Trace for NoTrace {
518        // Nothing here reads a task name, so the kernel is told not to build one.
519        // Without this the trait default is `true` and every traced event costs a
520        // name lookup plus a UTF-8 validation for a sink that drops it: measured
521        // at 3.86x on one row (2026-09-21).
522        const WANTS_NAMES: bool = false;
523
524        fn event(&mut self, _tick: u64, _event: Event<'_>) {}
525    }
526
527    // The declaration under test: two tasks and two queues of different
528    // types and different lengths, so nothing about the geometry can come
529    // out right by symmetry.
530    crate::system! {
531        mod demo use TestConfig;
532        tasks { consumer: 2, producer: 1 }
533        queues { data: u16; 10, marks: u8; 3 }
534    }
535
536    type K = demo::Kernel<TestPort, NoTrace, NoTickHook>;
537
538    // ---- tickless idle ---------------------------------------------------
539
540    crate::system! {
541        mod sleepy use TicklessConfig;
542        tasks { worker: 1 }
543        queues { mark: u8; 1 }
544    }
545
546    crate::system! {
547        mod awake use TestConfig;
548        tasks { worker: 1 }
549        queues { mark: u8; 1 }
550    }
551
552    /// A sink that keeps the shape of a trace: every event, with the tick it
553    /// carried. Comparing two of these compares two schedules.
554    #[derive(Debug, Default, PartialEq, Eq)]
555    pub struct Recorder {
556        pub lines: std::vec::Vec<(u64, &'static str)>,
557    }
558
559    impl Trace for Recorder {
560        fn event(&mut self, tick: u64, event: Event<'_>) {
561            self.lines.push((tick, event.name()));
562        }
563    }
564
565    /// `worker` delays, leaving only the idle task ready — which is the
566    /// precondition tickless idle exists for.
567    const DELAY: u64 = 10;
568
569    /// Both declarations build what they declare.
570    ///
571    /// The tests that matter below start a scheduler, and `System::build` and
572    /// a started scheduler cannot both have the one queue slot a declaration
573    /// reserves -- `system!` sizes `QUEUES` with "no spare", and the timer
574    /// daemon needs it. So they create the one task they need by hand, and
575    /// this is where the generated builders are exercised.
576    #[test]
577    fn both_tickless_declarations_build_what_they_declare() {
578        let mut asleep = sleepy_kernel();
579        let slept = sleepy::System::build(&mut asleep).expect("the tickless declaration builds");
580
581        let mut ticking =
582            awake::Kernel::<TestPort, NoTrace, NoTickHook>::new(TestPort::default(), NoTrace)
583                .expect("the declared geometry adds up");
584        let ticked = awake::System::build(&mut ticking).expect("the ticking declaration builds");
585
586        // The two declarations are the same text over two configurations, so
587        // they must hand back the same handle for the same task. If they did
588        // not, the arms of the invariance test below would be comparing two
589        // different systems and the comparison would mean nothing.
590        assert_eq!(
591            slept.worker, ticked.worker,
592            "identical declarations gave different task handles"
593        );
594        let _ = (&slept.mark, &ticked.mark);
595    }
596
597    /// A kernel with tickless on and a port that takes every window.
598    fn sleepy_kernel() -> sleepy::Kernel<SleepyPort, Scheduling<Recorder>, NoTickHook> {
599        sleepy::Kernel::new(SleepyPort::default(), Scheduling::new(Recorder::default()))
600            .expect("the declared geometry adds up")
601    }
602
603    /// The port asks for the whole window, and the kernel winds the clock
604    /// over it — all but the last tick, which is left pended so the delayed
605    /// task is woken by the same code that would have woken it.
606    #[test]
607    fn a_sleeping_port_is_asked_for_the_whole_idle_window() {
608        let mut k = sleepy_kernel();
609        k.create_task("worker", 1).expect("worker");
610        let started = k.start_scheduler().expect("start");
611        // The timer daemon is created READY at `TIMER_TASK_PRIORITY`, and
612        // in a running system it blocks on its command queue immediately.
613        // Nothing runs it here, so park it: a task ready above the idle
614        // priority is precisely what `expected_idle_time` refuses to sleep
615        // through, and leaving it ready would test that refusal instead of
616        // the sleep.
617        k.suspend(Some(started.timer))
618            .expect("park the timer daemon");
619        k.delay(DELAY).expect("the worker delays");
620
621        let before = k.tick_count();
622        k.idle_suppress_ticks();
623
624        assert_eq!(k.port().sleeps.get(), 1, "the port was asked once");
625        assert_eq!(
626            k.port().slept.get(),
627            DELAY,
628            "it was offered the whole window"
629        );
630        // Nine ticks were STEPPED and the tenth was left pended, and then
631        // `resume_all` unwound it through `increment_tick` -- which is what
632        // wakes the delayed task, by the same code that would have woken it
633        // had the kernel ticked all ten. So the clock lands on the wake
634        // either way, and that is the whole reason the schedule survives.
635        assert_eq!(
636            k.tick_count(),
637            before.saturating_add(DELAY),
638            "the clock did not land on the worker's wake"
639        );
640    }
641
642    /// The `configUSE_TICKLESS_IDLE` gate: the same port, the same script,
643    /// a configuration with it off, and the port is never asked.
644    #[test]
645    fn tickless_off_never_asks_the_port_to_sleep() {
646        let mut k =
647            awake::Kernel::<SleepyPort, NoTrace, NoTickHook>::new(SleepyPort::default(), NoTrace)
648                .expect("the declared geometry adds up");
649        k.create_task("worker", 1).expect("worker");
650        let started = k.start_scheduler().expect("start");
651        // The timer daemon is created READY at `TIMER_TASK_PRIORITY`, and
652        // in a running system it blocks on its command queue immediately.
653        // Nothing runs it here, so park it: a task ready above the idle
654        // priority is precisely what `expected_idle_time` refuses to sleep
655        // through, and leaving it ready would test that refusal instead of
656        // the sleep.
657        k.suspend(Some(started.timer))
658            .expect("park the timer daemon");
659        k.delay(DELAY).expect("the worker delays");
660
661        let before = k.tick_count();
662        k.idle_suppress_ticks();
663
664        assert_eq!(k.port().sleeps.get(), 0, "the port was asked");
665        assert_eq!(k.tick_count(), before, "the clock moved");
666    }
667
668    /// A port that oversleeps is a port bug, and the kernel refuses to wind
669    /// the clock past the wake it would lose. The C `configASSERT`s this;
670    /// a kernel that forbids panicking has to clamp instead.
671    #[test]
672    fn the_kernel_clamps_a_port_that_oversleeps() {
673        let mut k = sleepy::Kernel::<OversleepingPort, NoTrace, NoTickHook>::new(
674            OversleepingPort::default(),
675            NoTrace,
676        )
677        .expect("the declared geometry adds up");
678        k.create_task("worker", 1).expect("worker");
679        let started = k.start_scheduler().expect("start");
680        // The timer daemon is created READY at `TIMER_TASK_PRIORITY`, and
681        // in a running system it blocks on its command queue immediately.
682        // Nothing runs it here, so park it: a task ready above the idle
683        // priority is precisely what `expected_idle_time` refuses to sleep
684        // through, and leaving it ready would test that refusal instead of
685        // the sleep.
686        k.suspend(Some(started.timer))
687            .expect("park the timer daemon");
688        k.delay(DELAY).expect("the worker delays");
689
690        let before = k.tick_count();
691        k.idle_suppress_ticks();
692
693        // The port claimed `4 * DELAY + 9` = 49 ticks. Unclamped the clock
694        // would be 49 ticks on and the worker's wake would have been stepped
695        // straight past; clamped it lands on the wake exactly, like a port
696        // that behaved.
697        assert_eq!(
698            k.tick_count(),
699            before.saturating_add(DELAY),
700            "a port claiming four times its window moved the clock past the wake"
701        );
702    }
703
704    /// ★ The claim the whole mission rests on.
705    ///
706    /// Two kernels, identical but for `configUSE_TICKLESS_IDLE`, run the
707    /// same script. One sleeps through the idle window; the other ticks
708    /// through it. Their traces differ — that is the point, and it is why a
709    /// byte-diff cannot gate a tickless port. Their SCHEDULES do not.
710    ///
711    /// `Scheduling` is the projection that says so, and this is it applied
712    /// to a kernel that really did suppress its ticks rather than to a trace
713    /// file standing in for one.
714    #[test]
715    fn the_schedule_is_the_same_with_and_without_tickless() {
716        let target = DELAY.saturating_add(2);
717
718        let mut asleep = sleepy_kernel();
719        asleep.create_task("worker", 1).expect("worker");
720        let started = asleep.start_scheduler().expect("start");
721        asleep
722            .suspend(Some(started.timer))
723            .expect("park the timer daemon");
724        asleep.delay(DELAY).expect("the worker delays");
725        asleep.idle_suppress_ticks();
726        while asleep.tick_count() < target {
727            if asleep.increment_tick() {
728                asleep.switch_context();
729            }
730        }
731
732        let mut ticking = awake::Kernel::<TestPort, Scheduling<Recorder>, NoTickHook>::new(
733            TestPort::default(),
734            Scheduling::new(Recorder::default()),
735        )
736        .expect("the declared geometry adds up");
737        ticking.create_task("worker", 1).expect("worker");
738        let started = ticking.start_scheduler().expect("start");
739        ticking
740            .suspend(Some(started.timer))
741            .expect("park the timer daemon");
742        ticking.delay(DELAY).expect("the worker delays");
743        ticking.idle_suppress_ticks();
744        while ticking.tick_count() < target {
745            if ticking.increment_tick() {
746                ticking.switch_context();
747            }
748        }
749
750        assert!(
751            asleep.port().sleeps.get() > 0,
752            "the sleeping arm never slept, so this proves nothing"
753        );
754
755        let slept_schedule = asleep.into_trace().into_inner();
756        let ticked_schedule = ticking.into_trace().into_inner();
757
758        assert_eq!(
759            slept_schedule, ticked_schedule,
760            "sleeping through the idle window moved the schedule"
761        );
762    }
763
764    // ---- a delayed list deeper than the conformance corpus ever builds ---
765
766    // `System` here is never built: the tasks are created directly so the
767    // declared queue slot stays free for `xTimerCreateTimerTask`. The
768    // declaration is still what sizes the kernel.
769    crate::system! {
770        mod deep use TestConfig;
771        tasks {
772            d0: 1, d1: 1, d2: 1, d3: 1, d4: 1,
773            d5: 1, d6: 1, d7: 1, d8: 1, d9: 1
774        }
775        queues { unused: u8; 1 }
776    }
777
778    /// Ten tasks blocked at once, waking in wake-time order.
779    ///
780    /// **This exists because the conformance gate cannot reach here.** That
781    /// gate is the strongest evidence in the project -- 19 scenarios, every
782    /// scheduling decision compared line-for-line against the instrumented C
783    /// kernel -- and its delayed list never gets deeper than FOUR. Measured,
784    /// by replaying each scenario's own trace and tracking membership:
785    ///
786    /// ```text
787    ///   BlockQ 4  death 4  dynamic 3  PollQ 3  semtest 3  recmutex 3
788    ///   blocktim 3  IntSemTest 3  TimerDemo 3  QPeek 2  GenQTest 2
789    ///   MessageBufferAMP 2  countsem 1  QueueOverwrite 1
790    ///   QueueSetPolling 1  StreamBufferInterrupt 1  EventGroupsDemo 1
791    /// ```
792    ///
793    /// `vListInsert` is the only operation in the kernel whose cost is its
794    /// depth, and the corpus never tests it past n=4. Anything that goes
795    /// wrong only when the list is deep -- an ordering bug, an off-by-one in
796    /// the walk, a fast path that appends where it should splice -- passes
797    /// that gate green.
798    ///
799    /// The delays are deliberately NOT monotonic: a shortcut that appends
800    /// unconditionally orders them wrongly, and ascending delays would let
801    /// it through. They are distinct, because two tasks sharing a wake time
802    /// wake on the same tick and this probe cannot see which left first.
803    #[test]
804    #[allow(clippy::indexing_slicing)]
805    fn ten_blocked_tasks_wake_in_wake_time_order() {
806        let mut k =
807            deep::Kernel::<TestPort, NoTrace, NoTickHook>::new(TestPort::default(), NoTrace)
808                .expect("the declared geometry adds up");
809        // Created directly rather than through `System::build`. Build would
810        // also create the declared queue, and that slot is the one
811        // `xTimerCreateTimerTask` needs at startup -- the declaration is
812        // here to SIZE the kernel, not to furnish it.
813        let names = ["d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d8", "d9"];
814        let mut all = [::rusty_rtos_core::handle::TaskHandle::from_raw(0); 10];
815        for (slot, name) in all.iter_mut().zip(names) {
816            *slot = k.create_task(name, 1).expect("task");
817        }
818        let started = k.start_scheduler().expect("start");
819        // Park the timer daemon. It sits at TIMER_TASK_PRIORITY, above these
820        // tasks, so leaving it ready means it is the one that runs and the
821        // ten never get a turn. The tickless tests park it for the same
822        // reason.
823        k.suspend(Some(started.timer)).expect("park the daemon");
824
825        // Out of order on purpose: a shortcut that appends unconditionally
826        // gets these wrong, and ascending delays would let it through.
827        //
828        // DISTINCT, though. Two tasks with the same wake time wake on the
829        // same TICK, and `task_state_get` cannot say which of them left the
830        // delayed list first -- the probe below would record them in index
831        // order whatever the list did. The tie rule is real and it is tested
832        // where it is observable, directly on the list, in
833        // `list::tests::a_value_equal_to_the_tail_goes_after_it`.
834        let delays = [9u64, 3, 14, 7, 1, 11, 5, 8, 2, 13];
835
836        // Block each in turn. `delay` acts on whichever task is RUNNING, so
837        // the schedule picks the order and this records what it picked --
838        // asserting the kernel's own choice rather than assuming one.
839        let mut blocked = [(0u64, 0usize); 10];
840        let mut n = 0usize;
841        for _ in 0..200u32 {
842            if n == all.len() {
843                break;
844            }
845            let running = k.current();
846            if let Some(which) = all.iter().position(|t| *t == running) {
847                if !blocked[..n].iter().any(|(_, w)| *w == which) {
848                    k.delay(delays[which]).expect("blocks");
849                    blocked[n] = (delays[which], which);
850                    n = n.saturating_add(1);
851                }
852            }
853            k.switch_context();
854        }
855        assert_eq!(n, all.len(), "every task should have blocked");
856
857        for (i, task) in all.iter().enumerate() {
858            assert_eq!(
859                k.task_state_get(*task).expect("state"),
860                crate::kernel::TaskState::Blocked,
861                "task {i} should be blocked; all ten are, at once, which is                  more than twice as deep as any conformance scenario reaches"
862            );
863        }
864
865        // Tick forward and record the order they come back.
866        let mut woke = [0usize; 10];
867        let mut w = 0usize;
868        let mut asleep = [true; 10];
869        for _ in 0..40u32 {
870            k.tick_from_isr();
871            for (i, task) in all.iter().enumerate() {
872                if asleep[i]
873                    && k.task_state_get(*task).expect("state") != crate::kernel::TaskState::Blocked
874                {
875                    asleep[i] = false;
876                    woke[w] = i;
877                    w = w.saturating_add(1);
878                }
879            }
880        }
881        assert_eq!(w, all.len(), "every task should have woken");
882
883        // Ascending wake time; ties in the order they blocked. That is what
884        // `vListInsert` promises and the only order the list guarantees.
885        let mut want = blocked;
886        want.sort_by_key(|(ticks, _)| *ticks);
887        let want = want.map(|(_, which)| which);
888        assert_eq!(
889            woke, want,
890            "the delayed list did not wake in wake-time order"
891        );
892    }
893
894    /// Creating a task at a HIGHER priority than the running one must ask
895    /// for a switch; creating one LOWER must not.
896    ///
897    /// **This closes a gap the ledger measured and left open.** `cargo
898    /// mutants` over `kernel.rs`, judged by the conformance corpus, left six
899    /// survivors, and five were this one line in `prvAddNewTaskToReadyList`:
900    ///
901    /// ```text
902    /// if self.running && C::USE_PREEMPTION && self.current_priority() < priority
903    /// ```
904    ///
905    /// `death` was added so the corpus would create tasks after the
906    /// scheduler had started, and it narrowed six survivors to four. It
907    /// could not do better, for a reason visible in its own source:
908    /// `vCreateTasks` creates both suicidal tasks at `uxTaskPriorityGet(
909    /// NULL )` -- its OWN priority -- so the new task never outranks the
910    /// running one. The guard is reached and evaluated on every create and
911    /// is never TRUE. Mutations that make it fire spuriously die at once;
912    /// mutations that suppress a yield which never happens are invisible.
913    /// Reachable is not killed.
914    ///
915    /// No corpus scenario creates a higher-priority task while running, and
916    /// none can be added without a C original to diff against. A unit test
917    /// can do it, and this is it. The four survivors:
918    ///
919    /// | mutation | caught by |
920    /// |---|---|
921    /// | condition → `false` | the HIGHER half |
922    /// | body removed (no yield) | the HIGHER half |
923    /// | `<` → `>` | the HIGHER half |
924    /// | `<` → `!=` | the LOWER half |
925    ///
926    /// Which is why both halves are here. The lower half looks like a
927    /// formality and is the only thing that kills `!=`.
928    #[test]
929    fn creating_a_higher_priority_task_asks_for_a_switch_and_a_lower_one_does_not() {
930        let mut k = CountingKernel::new(CountingPort::default(), NoTrace)
931            .expect("the declared geometry adds up");
932        let mid = k.create_task("mid", 2).expect("task mid");
933        k.start_scheduler().expect("start");
934        assert_eq!(k.current(), mid, "the only app task runs");
935
936        // ---- LOWER: no switch is owed -----------------------------------
937        let before = k.port().yields.get();
938        let _low = k.create_task("low", 1).expect("task low");
939        assert_eq!(
940            k.port().yields.get(),
941            before,
942            "a task created BELOW the running priority cannot preempt it, so              the kernel must not ask for a switch"
943        );
944
945        // ---- HIGHER: a switch is owed -----------------------------------
946        let before = k.port().yields.get();
947        let _high = k.create_task("high", 3).expect("task high");
948        assert!(
949            k.port().yields.get() > before,
950            "a task created ABOVE the running priority preempts it, and the              kernel must ask for the switch"
951        );
952    }
953
954    /// A kernel with room for a stream buffer, which `system!` never gives
955    /// one: the macro sizes `BUFFERS` and `BYTES` at zero.
956    type BufferKernel = crate::Kernel<
957        TestConfig,
958        TestPort,
959        NoTrace,
960        NoTickHook,
961        4,
962        { crate::list_slots_for(4, 0, crate::lists_for(TestConfig::MAX_PRIORITIES, 1, 0)) },
963        { crate::lists_for(TestConfig::MAX_PRIORITIES, 1, 0) },
964        1,
965        1,
966        2,
967        64,
968        0,
969        0,
970        { <TestConfig as ::rusty_rtos_core::config::Config>::TIMER_QUEUE_LENGTH },
971    >;
972
973    /// `xStreamBufferSetTriggerLevel`'s contract, pinned.
974    ///
975    /// The C coerces a zero trigger to one BEFORE bounds-checking:
976    ///
977    /// ```c
978    /// if( xTriggerLevel == ( size_t ) 0 ) { xTriggerLevel = 1; }
979    /// if( xTriggerLevel < pxStreamBuffer->xLength ) { ...pdPASS }
980    /// else { xReturn = pdFALSE; }
981    /// ```
982    ///
983    /// This kernel had the two the other way round. The orders disagree on
984    /// exactly one input -- a zero trigger against a length of ONE -- and
985    /// that input cannot be built: `length` is `size + 1` (the spare byte a
986    /// ring buffer needs to tell full from empty) and `size == 0` is refused
987    /// at creation. So it was an inconsistency with the C, not a live bug,
988    /// and the order has been matched anyway rather than left resting on an
989    /// arithmetic detail in another function.
990    ///
991    /// What this test pins is the contract at the lengths that exist. There
992    /// is no conformance scenario for this API and no C demo in the vendored
993    /// tree that calls it, so a differential is not available --
994    /// `docs/HOLES.md` H2 -- and this is the only evidence there is.
995    #[test]
996    fn the_trigger_level_contract() {
997        let mut k =
998            BufferKernel::new(TestPort::default(), NoTrace).expect("the declared geometry adds up");
999
1000        // `create(1, 1)` is a length of TWO: one byte of payload and the
1001        // spare. The smallest buffer a caller can actually make.
1002        let small = k.stream_buffer_create(1, 1).expect("a one-byte buffer");
1003        assert_eq!(
1004            k.stream_buffer_set_trigger_level(small, 0),
1005            Ok(true),
1006            "zero becomes one, and one IS below a length of two"
1007        );
1008        assert_eq!(
1009            k.stream_buffer_set_trigger_level(small, 2),
1010            Ok(false),
1011            "two is not below a length of two"
1012        );
1013
1014        let wide = k.stream_buffer_create(8, 1).expect("an eight-byte buffer");
1015        assert_eq!(k.stream_buffer_set_trigger_level(wide, 0), Ok(true));
1016        assert_eq!(
1017            k.stream_buffer_set_trigger_level(wide, 8),
1018            Ok(true),
1019            "8 < 9"
1020        );
1021        assert_eq!(
1022            k.stream_buffer_set_trigger_level(wide, 9),
1023            Ok(false),
1024            "nine is not below a length of nine"
1025        );
1026    }
1027
1028    fn kernel() -> K {
1029        K::new(TestPort::default(), NoTrace).expect("the declared geometry adds up")
1030    }
1031
1032    /// `configUSE_TIME_SLICING`: two READY tasks of the same priority
1033    /// must take turns.
1034    ///
1035    /// Fixed priority says what happens between DIFFERENT priorities and
1036    /// nothing at all about two tasks at the same one. The C fills that in
1037    /// twice over: `xTaskIncrementTick` asks for a switch whenever more
1038    /// than one task is ready at the running task's priority, and
1039    /// `taskSELECT_HIGHEST_PRIORITY_TASK` then takes the NEXT entry rather
1040    /// than the head.
1041    ///
1042    /// Without it a task that never blocks starves its equal-priority
1043    /// peers completely — not slowly, completely — and the failure looks
1044    /// like somebody else's demo being broken. `rusty_rtos-capi`'s host
1045    /// cell found it that way: two demo files whose tasks sit at the same
1046    /// priority, one of them getting 13.5 million turns and the other
1047    /// ZERO.
1048    #[derive(Debug, Clone, Copy, Default)]
1049    pub struct SliceConfig;
1050
1051    impl Config for SliceConfig {
1052        type Tick = Bits32;
1053        const TICK_RATE_HZ: u32 = 100;
1054        const MAX_PRIORITIES: u8 = 4;
1055        const MINIMAL_STACK_SIZE: usize = 1;
1056        const MAX_TASK_NAME_LEN: usize = 8;
1057        // The daemon sits BELOW the two tasks under test. Nothing drives
1058        // it here, so at its usual priority it would simply be the current
1059        // task for ever and the test would measure the daemon.
1060        const TIMER_TASK_PRIORITY: u8 = 0;
1061        const TIMER_TASK_STACK_DEPTH: usize = 1;
1062        const TIMER_QUEUE_LENGTH: usize = 1;
1063        const NOTIFICATION_ARRAY_ENTRIES: usize = 1;
1064        // The point of this config: the default, stated out loud.
1065        const USE_TIME_SLICING: bool = true;
1066        const USE_PREEMPTION: bool = true;
1067    }
1068
1069    /// Built from the const generics directly rather than through
1070    /// `system!`, because that macro declares its tasks for you and this
1071    /// test needs to create them itself.
1072    type SliceKernel = crate::Kernel<
1073        SliceConfig,
1074        TestPort,
1075        NoTrace,
1076        NoTickHook,
1077        5,
1078        { crate::list_slots_for(5, 1, crate::lists_for(SliceConfig::MAX_PRIORITIES, 1, 1)) },
1079        { crate::lists_for(SliceConfig::MAX_PRIORITIES, 1, 1) },
1080        1,
1081        1,
1082        1,
1083        8,
1084        1,
1085        1,
1086        { <SliceConfig as ::rusty_rtos_core::config::Config>::TIMER_QUEUE_LENGTH },
1087    >;
1088
1089    /// The same declaration driven by a port that COUNTS switch requests,
1090    /// with room for a queue of our own beside the timer daemon's.
1091    ///
1092    /// `QUEUES` is 2 and `SLOTS` 8 for exactly that reason: the timer queue
1093    /// is created by `Kernel::new` and takes the first of each, so a test
1094    /// that makes its own queue needs the second.
1095    type CountingKernel = crate::Kernel<
1096        SliceConfig,
1097        CountingPort,
1098        NoTrace,
1099        NoTickHook,
1100        5,
1101        { crate::list_slots_for(5, 1, crate::lists_for(SliceConfig::MAX_PRIORITIES, 2, 1)) },
1102        { crate::lists_for(SliceConfig::MAX_PRIORITIES, 2, 1) },
1103        2,
1104        8,
1105        1,
1106        8,
1107        1,
1108        1,
1109        { <SliceConfig as ::rusty_rtos_core::config::Config>::TIMER_QUEUE_LENGTH },
1110    >;
1111
1112    /// Tick a kernel whose two equal-priority tasks never block, and count
1113    /// how many ticks each of them was the current task for.
1114    fn turns_over(ticks: u32) -> (u32, u32) {
1115        let mut k =
1116            SliceKernel::new(TestPort::default(), NoTrace).expect("the declared geometry adds up");
1117        let a = k.create_task("a", 2).expect("task a");
1118        let b = k.create_task("b", 2).expect("task b");
1119        k.start_scheduler().expect("start");
1120
1121        let (mut for_a, mut for_b) = (0u32, 0u32);
1122        for _ in 0..ticks {
1123            if k.increment_tick() {
1124                k.switch_context();
1125            }
1126            let now = k.current();
1127            if now == a {
1128                for_a = for_a.saturating_add(1);
1129            } else if now == b {
1130                for_b = for_b.saturating_add(1);
1131            }
1132        }
1133        (for_a, for_b)
1134    }
1135
1136    // A third test belongs here and is not written yet: equal-priority
1137    // rotation with a HIGHER-priority task cutting in between selections,
1138    // which is the shape every real system has. Two attempts at it were
1139    // withdrawn because they measured their own call pattern rather than
1140    // the scheduler -- `delay` yields as part of its contract, so a loop
1141    // that calls `delay` and then `switch_context` makes two selections
1142    // and can only sample after the second.
1143    //
1144    // What is known: driven directly, `switch_context` alternates
1145    // correctly between two equal-priority tasks with a third cutting in
1146    // (observed at the list, cursor 0 -> 1 -> 0). What is not known is why
1147    // `rusty_rtos_port/firmware/host-kernel` sees one of a pair get ZERO
1148    // turns in exactly that shape. That cell fails today and is the live
1149    // reproduction; this is the note that says so rather than a test that
1150    // asserts something unproven.
1151
1152    #[test]
1153    fn two_ready_tasks_of_equal_priority_take_turns() {
1154        let (a, b) = turns_over(200);
1155        assert!(a > 0, "task a never ran: {a} vs {b}");
1156        assert!(b > 0, "task b never ran: {a} vs {b}");
1157    }
1158
1159    /// `queueYIELD_IF_USING_PREEMPTION`: a send that readies a
1160    /// HIGHER-priority receiver must ask for a switch.
1161    ///
1162    /// This is the whole of `TimerDemo.c:469`, with no C anywhere near it.
1163    /// That demo stops a timer and asserts on the next line that the timer
1164    /// is inactive, and its own comment says why it may: "this will appear
1165    /// to happen immediately to this task because this task is running at
1166    /// a priority below the timer service task". `xTimerStop` is a send to
1167    /// the daemon's queue, so the assertion is really this test.
1168    ///
1169    /// It has two halves and the second is the one that bites. A send must
1170    /// ask for a switch when it wakes a higher-priority receiver — AND the
1171    /// receiver has to be WAITING ON THE QUEUE for there to be anything to
1172    /// wake. A daemon that polls its queue with a zero block time is
1173    /// parked nowhere, so the send finds no waiter, asks for nothing, and
1174    /// being higher priority buys it precisely nothing. Both halves are
1175    /// asserted below, because the first one passing is what makes the
1176    /// second one's absence invisible.
1177    #[test]
1178    fn a_send_that_wakes_a_higher_priority_receiver_asks_for_a_switch() {
1179        let mut k = CountingKernel::new(CountingPort::default(), NoTrace)
1180            .expect("the declared geometry adds up");
1181        let hi = k.create_task("hi", 2).expect("task hi");
1182        let lo = k.create_task("lo", 1).expect("task lo");
1183        let q = k.queue_create(4).expect("a queue");
1184        k.start_scheduler().expect("start");
1185
1186        // The high-priority task waits ON the queue, which is what makes
1187        // it a waiter rather than merely a ready task of higher priority.
1188        assert_eq!(k.current(), hi, "the higher priority task runs first");
1189        assert!(
1190            matches!(k.queue_receive(q, 100), Ok(crate::queue::Wait::Blocked)),
1191            "an empty queue with a block time must park the receiver"
1192        );
1193        k.switch_context();
1194        assert_eq!(k.current(), lo, "with hi parked, lo runs");
1195
1196        // Now the send, from the LOWER-priority task.
1197        let before = k.port().yields.get();
1198        assert!(matches!(
1199            k.queue_send(q, 7, 0),
1200            Ok(crate::queue::Wait::Ready(()))
1201        ));
1202        let asked = k.port().yields.get().saturating_sub(before);
1203        assert_eq!(
1204            asked, 1,
1205            "a send that readies a higher-priority receiver must ask for a switch"
1206        );
1207    }
1208
1209    /// The other half: a receiver that did NOT wait cannot be woken, so
1210    /// the send asks for nothing — which is correct, and is exactly the
1211    /// trap a polling daemon falls into.
1212    ///
1213    /// Recorded as an assertion rather than a comment because it is the
1214    /// behaviour a reader will otherwise call a bug in the send path,
1215    /// having found the send not yielding. The send is right; the caller
1216    /// that polls instead of waiting is wrong.
1217    #[test]
1218    fn a_send_asks_for_nothing_when_the_higher_priority_task_never_waited() {
1219        let mut k = CountingKernel::new(CountingPort::default(), NoTrace)
1220            .expect("the declared geometry adds up");
1221        let _hi = k.create_task("hi", 2).expect("task hi");
1222        let _lo = k.create_task("lo", 1).expect("task lo");
1223        let q = k.queue_create(4).expect("a queue");
1224        k.start_scheduler().expect("start");
1225
1226        // `hi` polls. A zero block time on an empty queue is `Empty`, NOT
1227        // `Blocked` — and the difference is the whole point: `Blocked`
1228        // means "I have been parked and will be woken", `Empty` means
1229        // "there was nothing there and I was not parked". A poller gets
1230        // the second, so there is nothing for a later send to wake.
1231        assert!(
1232            matches!(
1233                k.queue_receive(q, 0),
1234                Err(rusty_rtos_core::error::Error::Empty)
1235            ),
1236            "a zero-wait receive on an empty queue reports Empty, not Blocked"
1237        );
1238        k.switch_context();
1239
1240        let before = k.port().yields.get();
1241        assert!(matches!(
1242            k.queue_send(q, 7, 0),
1243            Ok(crate::queue::Wait::Ready(()))
1244        ));
1245        assert_eq!(
1246            k.port().yields.get(),
1247            before,
1248            "there was no waiter to wake, so nothing should have been asked for"
1249        );
1250    }
1251
1252    #[test]
1253    fn neither_equal_priority_task_starves_the_other() {
1254        let (a, b) = turns_over(200);
1255        // A round robin gives them one tick each in turn. Allowing a 4x
1256        // imbalance is generous; a scheduler that never rotates gives one
1257        // of them zero, which is what this is here to catch.
1258        let (lo, hi) = (a.min(b), a.max(b));
1259        assert!(
1260            lo.saturating_mul(4) >= hi,
1261            "one starved the other: {a} against {b}"
1262        );
1263    }
1264
1265    /// **The geometry is the declaration**, arithmetic and all.
1266    ///
1267    /// This is the `.bss` claim as a number rather than a hope: every
1268    /// const is exactly what the two `tasks` and two `queues` lines say,
1269    /// with no spare anywhere.
1270    #[test]
1271    fn the_geometry_is_derived_from_the_declaration() {
1272        assert_eq!(demo::APP_TASKS, 2, "the two declared tasks");
1273        assert_eq!(demo::TASKS, 4, "plus the idle task and the timer daemon");
1274        assert_eq!(demo::QUEUES, 2, "the two declared queues");
1275        assert_eq!(demo::SLOTS, 13, "10 + 3, the declared lengths and no spare");
1276        // `ITEMS` is the SLOT count now, not the item count: two list items
1277        // per task (8), plus one end-marker node per list (14), rounded up
1278        // to a power of two. 22 -> 32.
1279        assert_eq!(
1280            demo::ITEMS,
1281            32,
1282            "8 items + 14 markers, rounded to a power of two"
1283        );
1284        assert_eq!(
1285            demo::ITEMS,
1286            crate::list_slots_for(demo::TASKS, demo::TIMERS, demo::LISTS),
1287            "and it is derived, not written by hand"
1288        );
1289        // The slack is the price of the mask, and it is worth being able to
1290        // see: this configuration carries 10 spare nodes.
1291        assert_eq!(
1292            demo::ITEMS - demo::LISTS,
1293            18,
1294            "item capacity: 8 needed, 18 available"
1295        );
1296        assert_eq!(
1297            demo::LISTS,
1298            crate::lists_for(TestConfig::MAX_PRIORITIES, 2, 0),
1299            "a ready list per priority, the fixed six, and two per queue"
1300        );
1301    }
1302
1303    /// A declared system builds, and `Kernel::new` accepts the geometry —
1304    /// which is the check that `items_for`/`lists_for` were applied to the
1305    /// same numbers the struct was built from.
1306    #[test]
1307    fn a_declared_system_builds_and_the_kernel_accepts_its_geometry() {
1308        let mut k = kernel();
1309        let system = demo::System::build(&mut k).expect("nothing here can fail");
1310        // The handles are the kernel's, in declaration order.
1311        assert!(!system.consumer.is_null());
1312        assert!(!system.producer.is_null());
1313        assert_ne!(system.consumer, system.producer);
1314        assert_eq!(k.task_count(), 2, "no task the declaration did not name");
1315    }
1316
1317    /// The queues come back typed, and carry what they were declared to.
1318    #[test]
1319    fn the_declared_queues_move_their_own_type() {
1320        let mut k = kernel();
1321        let mut system = demo::System::build(&mut k).expect("nothing here can fail");
1322        assert!(matches!(system.data.send(&mut k, 4_242u16, 0), Sent::Ok));
1323        assert!(matches!(system.marks.send(&mut k, 7u8, 0), Sent::Ok));
1324        assert_eq!(system.data.len(&mut k), Ok(1));
1325        assert_eq!(system.marks.len(&mut k), Ok(1));
1326    }
1327
1328    /// **The `.bss` claim, as a number.**
1329    ///
1330    /// The alternative to declaring is what every scenario in this repo
1331    /// does today: pick a geometry by hand that covers the biggest thing
1332    /// you might do. The demo's kernel is sized `TASKS = 24`,
1333    /// `QUEUES = 12`, `SLOTS = 128`, `TIMERS = 32` because it has to hold
1334    /// eighteen scenarios at once. A system that declares two tasks and
1335    /// thirteen slots gets a kernel sized for two tasks and thirteen
1336    /// slots, and the difference is not small.
1337    ///
1338    /// This is a counter, not a benchmark: `size_of` is exact, decided at
1339    /// compile time, and identical on every machine.
1340    #[test]
1341    fn a_declared_kernel_is_sized_for_what_was_declared() {
1342        /// The demo's hand-picked geometry, which is what you use when
1343        /// you cannot declare.
1344        type Blanket = crate::Kernel<
1345            TestConfig,
1346            TestPort,
1347            NoTrace,
1348            NoTickHook,
1349            24,
1350            { crate::list_slots_for(24, 32, crate::lists_for(TestConfig::MAX_PRIORITIES, 12, 4)) },
1351            { crate::lists_for(TestConfig::MAX_PRIORITIES, 12, 4) },
1352            12,
1353            128,
1354            8,
1355            2048,
1356            32,
1357            4,
1358            { <TestConfig as ::rusty_rtos_core::config::Config>::TIMER_QUEUE_LENGTH },
1359        >;
1360
1361        let declared = core::mem::size_of::<K>();
1362        let blanket = core::mem::size_of::<Blanket>();
1363        // Printed so the footprint row can be read off a test run.
1364        let ratio = blanket as f64 / declared as f64;
1365        std::eprintln!("declared {declared} bytes, hand-sized {blanket} bytes, {ratio:.1}x");
1366        assert!(
1367            declared < blanket,
1368            "a declared kernel ({declared}) should be smaller than one              sized for everything ({blanket})"
1369        );
1370        // The declaration is the only input: change the declaration and
1371        // this number changes with it, which is the whole claim.
1372        assert_eq!(
1373            declared,
1374            core::mem::size_of::<demo::Kernel<TestPort, NoTrace, NoTickHook>>()
1375        );
1376    }
1377
1378    /// **The slot count is exact, and the exactness is load-bearing.**
1379    ///
1380    /// `SLOTS` is the sum of the declared lengths, so filling every
1381    /// declared queue to its declared length uses the arena to the last
1382    /// slot and no further. A geometry that was rounded up would pass a
1383    /// weaker version of this; this one would fail if `SLOTS` were 12.
1384    #[test]
1385    fn every_declared_slot_is_usable_and_there_is_not_one_spare() {
1386        let mut k = kernel();
1387        let mut system = demo::System::build(&mut k).expect("nothing here can fail");
1388        for i in 0..10u16 {
1389            assert!(
1390                matches!(system.data.send(&mut k, i, 0), Sent::Ok),
1391                "slot {i} of the declared ten"
1392            );
1393        }
1394        for i in 0..3u8 {
1395            assert!(
1396                matches!(system.marks.send(&mut k, i, 0), Sent::Ok),
1397                "slot {i} of the declared three"
1398            );
1399        }
1400        // Both are now exactly full, which is `SLOTS` exactly consumed.
1401        assert!(matches!(system.data.send(&mut k, 99, 0), Sent::Full(99)));
1402        assert!(matches!(system.marks.send(&mut k, 9, 0), Sent::Full(9)));
1403        assert_eq!(demo::SLOTS, 13);
1404    }
1405}