Experimental Physics and
Industrial Control System

Andrew Johnson <[email protected]> · Fri, 01 Jun 2007 18:15:07 -0500

In general I agree with Kay; you have a working design, and any change

now would require spending time deleting all the guard parameters and

changing the tests. I'm not that keen on the idea of using macros to

remove the guards either -- nobody actually builds EPICS with NDEBUG

defined that I know of, and I wouldn't be to surprised if some things

actually stop working when built like that - I've fixed more than one

routine in the last few years that was doing processing inside an

assert() statement.

Eric Norum wrote:

Andrew and I discussed some of the technical issues.

I see no showstoppers, even on multi-CPU, multi-cache, machines --

thought I don't think that my crushing logic has completely convinced

Andrew yet.....

I have don't like going beyond what is strictly guaranteed by the

threading primitive libraries that we're using in portable code, and I

think it's fair to say that we'd be outside what Posix promises.

pthread_mutex_lock() provides a read barrier (think cache-invalidate);

pthread_mutex_unlock() provides a write barrier (cache-flush). The

guaranteed synchronization requirements are that a write barrier must

follow the last write to the protected data, and a read barrier precede

the first read. Attempting to read before the read barrier could give

you anything, and that's what we'd be doing by comparing the

mutex.threadId with ours without.

Another benefit would be that it would be useful to have the task id of

a mutex holder exposed as a API-accessible value when trying to diagnose

deadlocks.

Careful, that task id might not be visible from threads across a CPU

boundary; if we offer the thread ID as a mutex method we'd risk this not

working in some circumstances (or we'd have to have a mutex to protect

the mutex...). Just because you only intend that API for debugging

doesn't mean that it won't get used in regular code...

- Andrew
--
The right to be heard does not automatically include
the right to be taken seriously. -- Hubert H. Humphrey

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System