threaded rts deadlock

[Sigbjorn Finne]

Hi,

thanks for a fine report; well spotted. I believe the
Scheduler changes I've just checked in avoids
running into this race condition.

--sigbjorn

----- Original Message ----- 
From: "Wolfgang Thaller" <wolfgang.thaller at gmx.net>
To: <glasgow-haskell-bugs at haskell.org>
Sent: Tuesday, June 18, 2002 11:47
Subject: threaded rts deadlock


> If a callback into Haskell code finishes very quickly, a deadlock 
> occurs. Adding a small delay helps.
> 
> This is what I think happens:
> 
> rts_evalIO creates a haskell thread
> rts_evalIO calls scheduleExtThread to spawn a new OS thread. When the 
> callback is finished, schedule() calls broadcastCondition on m->wakeup. 
> But because the callback took very little time, this happens _before_ 
> rts_evalIO calls waitThread.
> waitThread is responsible for initializing m->wakeup and waiting for a 
> broadcast, but broadcastCondition might already have been called 
> _before_ that. It won't be called again, so nothing happens anymore...
> Some locking to prevent this race condition is probably necessary...
> 
> Cheers,
> 
> Wolfgang
>