preempt: explicit check!() macro for no-alloc loops
Stable Rust emits stack probes inline (subq/movq/jne loop) rather than calling __rust_probestack, so there's no transparent hook for stack- frame preemption. Override of __rust_probestack links cleanly but never runs. Falling back to an explicit check!() that users drop into hot compute loops. check!() decrements the same ALLOC_COUNT counter as the heap path, so both event sources fire timeslice checks at the same rate. Documents the prep-to-park invariant on maybe_preempt — library code that registers a wakeup and then parks must keep that window alloc-free and check-free, or a preemption-driven yield in the middle would lose the wakeup.
This commit is contained in:
22
src/lib.rs
22
src/lib.rs
@@ -42,3 +42,25 @@ pub use scheduler::{
|
||||
block_on_io, run, self_pid, sleep, spawn, spawn_under, yield_now, JoinError, JoinHandle,
|
||||
};
|
||||
pub use supervisor::Signal;
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// check!() — explicit preemption point for tight no-alloc loops.
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/// Voluntarily check whether this actor's timeslice has expired, yielding
|
||||
/// if so. Drop this into hot compute loops that don't allocate (heap or
|
||||
/// large stack frames) — without it, such loops monopolise the scheduler
|
||||
/// until they return.
|
||||
///
|
||||
/// Decrements the same per-actor event counter as the heap allocator's
|
||||
/// preemption hook, so the check rate is identical regardless of whether
|
||||
/// the actor is alloc-heavy, check-heavy, or mixed.
|
||||
///
|
||||
/// No-op outside an actor (the runtime's `PREEMPTION_ENABLED` flag is
|
||||
/// false there).
|
||||
#[macro_export]
|
||||
macro_rules! check {
|
||||
() => {
|
||||
$crate::preempt::maybe_preempt()
|
||||
};
|
||||
}
|
||||
|
||||
@@ -6,10 +6,16 @@
|
||||
//! `switch_to_scheduler` to yield. Resetting the counter to `ALLOC_INTERVAL`
|
||||
//! amortises the RDTSC across many cheap events.
|
||||
//!
|
||||
//! Events today are heap allocations (via `PreemptingAllocator`). v0.2 will
|
||||
//! add stack-frame entries as a second event source — frames are stack
|
||||
//! allocations, the counter naming still fits — sharing this same counter
|
||||
//! so both routes behave consistently.
|
||||
//! Two event sources today:
|
||||
//! - `PreemptingAllocator` — heap allocations.
|
||||
//! - `smarm::check!()` — explicit preemption point for tight no-alloc
|
||||
//! loops, since stable Rust gives us no transparent way to preempt
|
||||
//! such loops (`__rust_probestack` is emitted inline by LLVM and not
|
||||
//! called at runtime).
|
||||
//!
|
||||
//! Both sources share `ALLOC_COUNT`, so the timeslice check fires at the
|
||||
//! same rate regardless of whether the actor is alloc-heavy, check-heavy,
|
||||
//! or mixed.
|
||||
//!
|
||||
//! All state is thread-local. The scheduler enables preemption on resume
|
||||
//! and disables it on the return path, so the scheduler can never preempt
|
||||
@@ -80,9 +86,17 @@ unsafe impl GlobalAlloc for PreemptingAllocator {
|
||||
}
|
||||
|
||||
/// Shared preemption check. Called by every preemption event source — the
|
||||
/// heap allocator today, the stack-frame entry hook in v0.2. Decrements
|
||||
/// `ALLOC_COUNT`; every `ALLOC_INTERVAL` calls reads the timeslice clock
|
||||
/// and yields if expired.
|
||||
/// heap allocator today, `smarm::check!()` for tight no-alloc loops.
|
||||
/// Decrements `ALLOC_COUNT`; every `ALLOC_INTERVAL` calls reads the
|
||||
/// timeslice clock and yields if expired.
|
||||
///
|
||||
/// **Invariant**: must not be called inside a "prep-to-park" region —
|
||||
/// e.g. between registering as a channel's parked receiver and calling
|
||||
/// `park_current()`. A preemption-driven yield in that window would
|
||||
/// reach the scheduler with state=Runnable, the unparker would no-op,
|
||||
/// the actor would then park, and the wakeup would be lost. Library
|
||||
/// code that touches the parking primitives must keep its prep-to-park
|
||||
/// regions allocation-free and check!()-free.
|
||||
#[inline(always)]
|
||||
pub fn maybe_preempt() {
|
||||
ALLOC_COUNT.with(|c| {
|
||||
|
||||
66
tests/preempt.rs
Normal file
66
tests/preempt.rs
Normal file
@@ -0,0 +1,66 @@
|
||||
//! Tests for explicit preemption via `smarm::check!()`.
|
||||
|
||||
use smarm::{run, spawn};
|
||||
use std::sync::atomic::{AtomicU64, Ordering};
|
||||
use std::sync::Arc;
|
||||
|
||||
#[test]
|
||||
fn check_yields_when_timeslice_expired() {
|
||||
// A single actor that drives the timeslice clock to zero manually,
|
||||
// then calls check!() and expects to yield. The scheduler has nothing
|
||||
// else to run, so it just re-queues us. To prove we actually yielded,
|
||||
// observe the run counter on the slot... we don't have one. So
|
||||
// instead: spawn a second actor that increments a counter and joins
|
||||
// it; verify both actors made progress in interleaved order under
|
||||
// forced timeslice expiry.
|
||||
let order: Arc<std::sync::Mutex<Vec<u8>>> = Arc::new(std::sync::Mutex::new(Vec::new()));
|
||||
let o1 = order.clone();
|
||||
let o2 = order.clone();
|
||||
|
||||
run(move || {
|
||||
let a = spawn(move || {
|
||||
o1.lock().unwrap().push(b'A');
|
||||
// Force the timeslice to be considered expired.
|
||||
smarm::preempt::expire_timeslice_for_test();
|
||||
smarm::check!();
|
||||
o1.lock().unwrap().push(b'a');
|
||||
});
|
||||
let b = spawn(move || {
|
||||
o2.lock().unwrap().push(b'B');
|
||||
smarm::preempt::expire_timeslice_for_test();
|
||||
smarm::check!();
|
||||
o2.lock().unwrap().push(b'b');
|
||||
});
|
||||
a.join().unwrap();
|
||||
b.join().unwrap();
|
||||
});
|
||||
|
||||
// FIFO scheduling + forced preemption: A starts, expires, yields to B;
|
||||
// B starts, expires, yields to A; A finishes, B finishes.
|
||||
// Required: both uppercase letters appear before either lowercase.
|
||||
let v = order.lock().unwrap();
|
||||
let pos_big_a = v.iter().position(|&c| c == b'A').unwrap();
|
||||
let pos_big_b = v.iter().position(|&c| c == b'B').unwrap();
|
||||
let pos_lit_a = v.iter().position(|&c| c == b'a').unwrap();
|
||||
let pos_lit_b = v.iter().position(|&c| c == b'b').unwrap();
|
||||
assert!(pos_big_a < pos_lit_a, "A's tail ran before B's head: {:?}", *v);
|
||||
assert!(pos_big_b < pos_lit_b, "B's tail ran before A's head: {:?}", *v);
|
||||
assert!(pos_big_a.max(pos_big_b) < pos_lit_a.min(pos_lit_b),
|
||||
"preemption didn't interleave: {:?}", *v);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn check_is_a_noop_when_timeslice_not_expired() {
|
||||
// After a fresh resume, check!() should be cheap and not yield. Run
|
||||
// a single actor that calls check!() many times; it should complete
|
||||
// promptly.
|
||||
let count = Arc::new(AtomicU64::new(0));
|
||||
let c = count.clone();
|
||||
run(move || {
|
||||
for _ in 0..1_000 {
|
||||
smarm::check!();
|
||||
c.fetch_add(1, Ordering::Relaxed);
|
||||
}
|
||||
});
|
||||
assert_eq!(count.load(Ordering::Relaxed), 1_000);
|
||||
}
|
||||
Reference in New Issue
Block a user