Experimental Physics and
Industrial Control System

On Sep 10, 2021, at 10:05 AM, Matt Newville <newville at cars.uchicago.edu> wrote:

On Thu, Sep 9, 2021 at 7:08 PM Johnson, Andrew N. wrote:

However I do think that the access fault you’re seeing is because a thread that is calling libca doesn’t have a CA context.

…

Does pyepics give access to the libca routines ca_current_context() and ca_attach_context() ? If so I would recommend having that watchdog thread explicitly attach to the context that was created for the main thread. I think that will probably solve the problem if those APIs are available.

Yes, those are available.  There is also a simple subclass of threading.Thread (epics.ca.CAThread) that explicitly attaches the thread to the initial CA context.  My recollection was that Mark tried this and it didn't help.  

I would definitely recommend using your subclass then, even if it doesn’t fix this problem by itself.

The situation that I am concerned about is there being 2 copies of libca and libCom linked into the same process — that could lead to all kinds of problems when symbols are looked up dynamically if the second copy overrides calls to the first once it gets loaded. This problem is being reported on Windows, and the behavior of dynamic loading there will probably be different than on Linux.

Is there any way to get a stack-trace from the C/C++ world when this dies? Mark’s traceback only showed the Python side, and it would be helpful to see where this is coming from in the underlying C/C++ libraries too:

  File "C:\Users\rivers\Anaconda3\envs\tomoscan\lib\site-packages\epics\ca.py", line 923, in poll

    pend_event(evt)

  File "C:\Users\rivers\Anaconda3\envs\tomoscan\lib\site-packages\epics\ca.py", line 910, in pend_event

    ret = libca.ca_pend_event(timeout)

OSError: exception: access violation reading 0x0000000000000048

The access violation there looks very much like an offset from a NULL pointer, but knowing what C++ routine is running at the time would be a lot more helpful. I might be wrong about this being the CA context, but we’ll probably need more information to find out what’s actually happening.

- Andrew