The single-scheduler hot path acquired the shared mutex three times per
yield: once to drain timers, once to drain IO completions, and once to pop
the next runnable actor. It also called Instant::now() every loop iteration
to feed timers.pop_due(), even though the pure-yield / pure-compute workloads
never arm a timer.
Confirmed with llmdbg before touching anything: breakpointing
switch_to_scheduler and instruction-counting one scheduler-side yield cycle on
the ctx_switch_probe showed ~6950 instructions / 68 calls of bookkeeping
between consecutive actor resumes, dominated by lock acquisition and the
per-iteration clock read — not by the naked-asm switch itself (~16 insns).
The te!() trace macro compiles to () without the smarm-trace feature, so it
was ruled out as a contributor.
This collapses the two drain `with_shared` calls into one lock hold and reads
the clock only when the timer heap is non-empty. IO completions are still
drained unconditionally while the IO subsystem is live, because the IO/pool
threads push completions onto their own mutex (not `shared`), so the scheduler
cannot know in advance whether any arrived — it must look; that look is a
single empty-VecDeque check on the hot path. No change to timer or IO
semantics; the asm switch is untouched.
Measured on this box (single core, nproc==1), medians over the bench harness,
re-baselined locally first:
yield_many smarm 1-thread 38459 -> 29943 us (-22.1%)
yield_in_hot_loop smarm 1-thread 166220 -> 122629 us (-26.2%)
ping_pong_oneshot smarm 1-thread 1619 -> 1367 us (-15.6%)
chained_spawn smarm 1-thread 481 -> 399 us (-17.0%)
Run-to-run spread on yield_many is ~3-4% (29565-30634), so the 22% move is
well outside the noise. Gap to tokio current_thread closes from ~2.4x to ~1.95x.
All tests green: cargo test --test context (4) plus the full suite (93 total).
The diagram claimed the entry/ret target sits at aligned_top-8 and r15 at
aligned_top-56. The code does (top & ~15) - 8 then a -= 8 before the first
write, so entry actually lands at aligned_top-16 and r15 at aligned_top-64.
Confirmed by single-stepping the shim's ret under llmdbg: the entry slot is
at an address with %16==0, giving the required rsp%16==8 on entry. Code was
correct; only the comment was wrong (and would mislead a maintainer).
A minimal Config::exact(1) program with two actors doing a fixed number of
yield_now() calls. No I/O/timers/channels, so the instruction stream stays
centered on the naked-asm switch. Frame-pointer-friendly. Intended for
single-threaded inspection under a debugger (e.g. llmdbg).
While running benchmarks, a hang surfaced in timer-only workloads — actors
sleeping with no IO in flight. Tracing it down, the race lives in sleep():
between the call to `timers.insert_sleep` and the subsequent `park_current`
yield, the actor is still in State::Runnable. If the timer fires in that
window, the old code's `if matches!(slot.state, State::Parked)` guard silently
drops the wakeup. The actor then parks normally and never gets re-queued —
it sleeps forever.
The fix mirrors what scheduler::unpark() and the IO FdReady path already do:
when the timer fires and the slot is still Runnable, set `pending_unpark`
instead of re-queuing immediately. The upcoming Park yield sees the flag and
re-queues the actor rather than suspending it, closing the race without any
new synchronisation.
Adds a regression test: 100 actors doing pure timer sleeps across ≥2
scheduler threads. The test asserts both correctness (all actors complete)
and timeliness (wall time < 2×sleep duration), which is enough signal to
catch a stuck actor even on a single-core CI runner.
Three changes, each independently measured, landed together.
Fuse pre-resume mutex acquisitions
-----------------------------------
The schedule loop previously took the shared lock three separate times
per actor resume: once to pop the run queue, once to read the stack
pointer, and once to pop the first-resume closure. These are now a
single acquisition that returns everything needed to resume the actor.
As a side effect, pending_closures was changed from a Vec<(Pid, Closure)>
with O(n) linear scan to a Vec<Option<Closure>> indexed by slot index,
making first-closure lookup O(1).
Move stack allocation outside the shared lock
----------------------------------------------
Stack::new() (mmap + mprotect) was previously called inside with_shared,
stalling every other scheduler thread for the duration of two syscalls on
every spawn. It now runs before the lock is acquired.
Stack pool
----------
Rather than munmap-ing a stack when an actor finishes and mmap-ing a fresh
one on the next spawn, stacks are now recycled through a per-Runtime pool
(Mutex<Vec<Stack>>). finalize_actor extracts the stack from the Actor
before clearing the slot and pushes it to the pool outside the shared lock.
spawn_under pops from the pool before falling back to Stack::new().
The pool is unbounded for now (shrink policy TBD) but capped at
stack_pool_cap stacks on return, defaulting to thread_count * 4.
The cap is configurable via Config::stack_pool_cap(n).
Results (24-thread, against stored baseline)
--------------------------------------------
chained_spawn smarm 1-thread: 9763 → 261 µs (-97%)
chained_spawn smarm 24-thread: 23562 → 838 µs (-96%)
ping_pong_oneshot smarm 1-thread: 18409 → 742 µs (-96%)
ping_pong_oneshot smarm 24-thread: 44596 → 1425 µs (-97%)
catch_unwind_panics smarm 24-thread: 267812 → 124094 µs (-54%)
fan_out_compute smarm 24-thread: 2839 → 2226 µs (-22%)
Tokio regressions in the checker output are baseline measurement drift;
no tokio code was changed.
Add `Config::alloc_interval()` and `Config::timeslice_cycles()` so
callers can tune preemption sensitivity at runtime. The values flow
through `RuntimeInner` and are written into per-scheduler-thread locals
via a new `configure_preempt()` call at thread startup, keeping the hot
path free of cross-thread coherency traffic.
Fix unused-variable warnings in channel.rs by inlining `current_pid()`
directly into `te!` macro arguments — since the no-op macro arm never
evaluates its argument, no binding is needed at the call site.
Clean up a handful of dead imports exposed by the refactor.
Config API changes (src/preempt.rs, src/runtime.rs):
- preempt: promote ALLOC_INTERVAL and TIMESLICE_CYCLES from bare consts to
DEFAULT_ALLOC_INTERVAL / DEFAULT_TIMESLICE_CYCLES; store active values in
thread-locals set on each actor resume so multiple runtimes can use
different settings concurrently.
- runtime: add alloc_interval / timeslice_cycles fields to Config; add
Config::alloc_interval(n) and Config::timeslice_cycles(c) builder methods;
thread the values through RuntimeInner to the reset_timeslice() call in
schedule_loop.
Bench changes:
- Add bench_cfg(threads) helper to general/tokio_favored/smarm_favored that
wraps Config::exact and reads SMARM_ALLOC_INTERVAL / SMARM_TIMESLICE_CYCLES
env vars, so the sweep script can vary knobs without recompiling.
Sweep tooling (benches/sweep.py):
- 'run': run the 3-file bench suite once; --save-baseline persists JSON
- 'regress': compare current run against baseline.json, exit 1 on any bench
that regresses >10% vs stored medians
- 'sweep': run the full SWEEP_GRID (10 points), print comparison table,
optional --save-csv; binaries pre-built so no recompile per point
Sweep results (10-point grid, 1-CPU sandbox):
- The preemption knobs have very little effect on this single-CPU machine.
Most benches move <5% across the entire grid.
- Longer timeslices (tc=600k, tc=1200k) reliably hurt spawn_storm_busy
(+11-15%) and catch_unwind_panics (+10-12%) because actors hold the
scheduler mutex longer per timeslice, stalling the storm of joinable tasks.
- Shorter timeslices (tc=150k) give a small improvement on many_timers
(-3-4%) and a wash everywhere else.
- yield_in_hot_loop and uncontended_channel are essentially flat across all
knobs — both are scheduling-dominated and call yield_now explicitly, so
the RDTSC-driven preemption path is irrelevant.
- Conclusion: the knobs matter primarily under contention (multi-core).
Re-run sweep on a multi-core machine before drawing tuning conclusions.
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.
Adds a BinaryHeap of timer entries on SchedulerState. sleep() inserts
an entry and parks; schedule_loop pops due entries each iteration and
unparks them. When the run queue is empty but timers are pending, the
OS thread sleeps until the soonest deadline.
Single-threaded only; thread::sleep is fine because no other thread
can wake us. The IO thread coming next will need a Condvar or pipe
wakeup to break this OS-sleep early.
v0.2 will add stack-frame entries as a second preemption event source.
Both routes share ALLOC_COUNT so the timeslice check rate is the same
whether the actor is alloc-heavy, frame-heavy, or mixed.
Hand-rolled context switching on mmap'd stacks with guard pages,
allocator-driven RDTSC preemption, unbounded MPSC channels, supervision
via per-slot Signal mailboxes, root supervisor as sentinel PID.
Lib + tests + benches clean check/clippy. All 29 tests pass.
Bench: smarm 3.4% over serial baseline, within 160us of tokio
current-thread on prime-counting fan-out.
smarm-dev