Re: console-client signal handling

[Marco Gerards]

Marcus Brinkmann <marcus.brinkmann@ruhr-uni-bochum.de> writes:

> At Wed, 21 Jul 2004 13:15:07 +0200,
> marco_g wrote:
>> 
>> Marcus Brinkmann <marcus.brinkmann@ruhr-uni-bochum.de> writes:
>> > However, I have a big concern:
>> >
>> > I have doubts if calling console_exit, and thus driver_fini, is safe
>> > at all from a signal handler.  You might have to set some global flag,
>> > possibly protected by a spin lock (not mutex, I guess), which is then
>> > checked periodically by some other thread.
>> 
>> Can't I use conditions instead?
>
> Unfortunately not:
>
> 32578            It is not safe to use the pthread_cond_signal( ) function in 
> a signal handler that is invoked
> 32579            asynchronously. Even if it were safe, there would still be a 
> race between the test of the Boolean
> 32580            pthread_cond_wait( ) that could not be efficiently 
> eliminated.
> 32581            Mutexes and condition variables are thus not suitable for 
> releasing a waiting thread by signaling
> 32582            from code running in a signal handler.

Ok.  I just had cthreads in mind.  AFAIK there is no problem with
using the libcthread condition_signal().  Or don't you want to use that
because we will switch to pthreads eventually?

> It's also questionable if you really want another thread to wait on a
> condition all the time.  OTOH, there probably is no suitable other
> thread.  So, if we need another thread, then that's ok.  But block
> signals for all other threads, and have that thread also handle the
> signals you want to catch.  That's best because then memory is already
> synchronized.  Having that thread sleep most of the time and check for the
>
> volatile sig_atomic_t should_we_die;

Unfortunately there is not a single thread that is active in the
console client (libcons gets control of the main thread).  So I will
make a new thread for this.

I will implement this as you proposed now.

> global flag every second should be ok.  A max delay of 1 second at kill is
> acceptable, IMHO.

Absolutely.  AFAIK signals are not reliable anyway.

Thanks,
Marco