Thread-backed handles: thread blocked indefinitely in an MVar operation

[Simon Marlow]

On 22/06/10 12:43, Dimitry Golubovsky wrote:
> Simon,
>
> On Tue, Jun 22, 2010 at 4:31 AM, Simon Marlow<marlowsd at gmail.com>  wrote:
>
>> So the problem is that fillReadBuffer has to return 0 to indicate EOF, but
>> you are killing the slave thread as soon as it has reached the end of the
>> stream.  So you get one call to fillReadBuffer that returns the data up to
>> the end of the stream, and the next call that should return 0 blocks on the
>> Chan indefinitely because the slave thread has already been killed.
>
> Thanks for your reply.
>
> Updated paste: http://hpaste.org/fastcgi/hpaste.fcgi/view?id=26455
>
> I enclosed readChan in a catchException call which seems to have cured
> the problem: it just simulates EOF on the handle if the thread is
> blocked (channel vanished). Interestingly that in the failing version,
> handle closing was done in a separate thread (thread 3), but in the
> updated version it was the same thread 1.

That will work, but very very slowly, because the exception isn't raised 
until a full GC runs and discovers the thread to be unreachable.  I 
suggest instead having some state to indicate that EOF has been reached.

Cheers,
	Simon