You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Errors inside spawned processes are swallowed and misreported as "deadlock: no process is runnable" #174
A runtime error inside a spawned process is swallowed. The process dies, and the program reports the consequence — no runnable process — instead of the cause, with no location.
The same expression, in the main process and inside a spawn:
# die1.kex — main process
main do
IO.printLine([1].append(2))
end
Internal error: die1.kex:2:26: runtime error: Undefined method: append for [Int]
# die2.kex — inside spawn
type Msg = | Ping(Process)
main do
let p = spawn do
receive do
Ping(who) => who.send([1].append(2))
end
end
p.send(Ping(Process.self))
receive do
r => IO.printLine("got ${r}")
end
end
Internal error: :0:0: runtime error: deadlock: no process is runnable
Same defect, same line of code. In one case the message names the method, the receiver type, and the source location; in the other it names none of them and misattributes the failure to the scheduler. :0:0 is not a real location.
"Deadlock" is also the wrong diagnosis. Nothing deadlocked — a process crashed, and the parent then blocked forever in receive on a reply that could never come. The scheduler is reporting the second-order effect.
Why this is worth fixing beyond the message
This is how every bug inside a spawned process presents, so any process-based code is debugged with no cause and no line number. It cost me a while to find in a hand-written spec fake; the same shape appears in any receive loop.
It also gets worse under the capability/mocking design in #143: state-carrying test fakes have to be backed by processes (no global mutable state), so ordinary mistakes in test doubles will land in spawned processes by construction, and will surface as deadlock: no process is runnable.
Suggested direction
The crashing process should report its own error — method, receiver, and source location, as the main process already does — and the parent's subsequent block should either name the dead process or, at minimum, not be reported as a deadlock.
Backend note
Confirmed on the walker (interpreter). On BEAM (-R) the same program hangs rather than terminating; I did not confirm whether the underlying cause is the same, so that half needs checking before the fix is scoped.
Activity
Keep IO mock working, closing #181, #174