Re: CA and threads

[Dirk Zimoch <dirk.zimoch@psi.ch>]

Hello Andrew,

Thank, you for your explanation.

Unfortunately this behavior bites me now. After upgrading to a new OS 
and thus Tcl/Tk version, it turned out that monitors in our EPICS-Tcl 
interface did not work any more. This is because Tcl now can be compiled 
"thread aware" (which has been done by the distributor) and refuses to 
accept commands that come from a different thread.
Somehow I now have to get the Tcl callbacks back into the context of the 
main thread which runs an event loop that polls the ca library.

It would be nice to have the old 3.13 style behavior back where 
everything that interacts with the user process is the same (main) 
thread, even if CA uses multiple threads internally.

Dirk

Andrew Johnson wrote:
> On Monday 28 March 2011 09:57:27 Dirk Zimoch wrote:
> > If I create a CA client context with
> >    ca_context_create (ca_disable_preemptive_callback);
> > or, what is the same, with
> >    ca_task_initialize();
> > in which thread context will then the callback handlers be called? Is it
> > in the context of the thread that calls ca_pend_io() or still in the
> > context of a separate thread? How was it implemented in 3.13?
>
> In 3.14 the CA client callbacks are *always* made in the context of an 
> internal CA thread, and never of the thread that called ca_pend_io() / 
> ca_pend_event().  If you use the above calls to initialize the library the 
> callbacks are only ever called while the other thread is still executing the 
> ca_pend_*() library routine, so you do not have to include mutexes in your 
> code.
>
> It occurs to me that there *could* be some circumstances where that's not 
> strictly true, if the ca_pend_* thread responds to a signal, but that would be 
> fairly unusual and would probably break the rules for signal handlers anyway.
>
> In 3.13 the callbacks were made from the thread that called ca_pend_*(), CA 
> was not multi-threaded internally and did all its I/O using the select() OS 
> call.  The CAS still uses this approach.
>
> - Andrew