Re: [Fwd: Re: stream without asynPort]

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

StreamDevice executes shell commands in the context of one of the three 
standard callback threads, depending on the PRIO field of the record.

Originally I had implemented this feature to be able to change the 
serial parameters of an asyn port if necessary. Of course nothing stops 
anyone to call "dbl>file" or "casr 3" wich may take a long time to 
execute, or even "iocPause.

It is always possible to do bad things during record processing, e.g. in 
a subroutine record as Ralph pointed out. The same is true for SNL code. 
No API is "fool proof". This is just another point in EPICS where "you 
should know what you are doing".

Dirk

Andrew Johnson wrote:
> On Monday 21 February 2011 11:59:42 Ralph Lange wrote:
> > On 21.02.2011 18:45 Eric Norum wrote:
> > > I wrote a simply asynOctet driver to do this.  I've not yet packaged this
> > > up for official release until I hear back from the folks testing this.
> > > I'm not 100% sure that it's a good idea to allow commands to be run as
> > > the result of record processing, but maybe there's a real need out there.
> > Any subroutine record can execute commands in the context of record
> > processing. Where's the big difference?
>
> I think the word "command" implies the distinction.  Many of the older IOC 
> test and debugging commands were coded with the assumption that they would be 
> run from the vxWorks shell, i.e. they're running in a very high priority 
> thread and won't normally be preempted by record processing, so they maybe 
> don't do as much locking as they should (indeed there may even be locks 
> missing from dbBase that would be needed to make them completely reliable and 
> thread-safe when used in that way, but adding such locks could adversely 
> affect the performance of the IOC).
>
> I suspect that the 'dbl' command should be safe to run from record processing 
> since we don't support the ability to add or delete records at run-time, but 
> commands like 'scanppl' or 'dbior' might have subtle interactions that could 
> cause deadlocks, crashes or just wrong results in some circustances.  I 
> haven't examined any source code in writing this response so I could be wrong 
> about my specific examples, but if anyone is going to rely on the ability to 
> run these kinds of commands from record processing they should do their own 
> due diligence about them first.
>
> There's also the issue of blocking which Eric just brought up; if the command 
> takes a long time to run it could prevent other real-time operations from 
> operating normally.
>
> - Andrew