From: Matthew Herrmann (matthew.herrmann_at_[hidden])
Date: 2007-11-22 17:28:29
I have a question for the author of the logging library on the boost
review queue regarding static initialization and destruction.
I currently use the logging library that was rejected at the last
review, and I noticed some issues with it relating to handling
destruction in a static context. Specifically, I'd get occasional
segfaults if I declared an item in static scope that logged in its
destructor. Once the object was moved to within the main() function, the
issue went away. I'm not sure if it was the logging library, or our use
of it, but it was a bit of an issue since it made the use of logging
less transparent to clients of my library code.
I can see from the documentation that you have given quite a bit of
thought to multithreading concerns. Have you also considered issues
around static initialization and destruction?
In particular, a class that performs logging in its constructor and
destructor should still behave correctly when instantiated in a static
context. I think the problems I had with logging before related to the
logging data structures being torn down in the static destruction phase
before my own objects.
The way I solved that in a (far more limited!) logging library for our
own use was the following:
- Never destroy the logging data structures -- allow them to leak. No
risk of static ordering issues.
- In the static initialization case, initialize the logging structures
on first access, and if not accessed during the static context,
guarantee that they are initialized before main() is called.
To do that, I essentially borrowed the implementation of the boost pool
singleton, except that I removed the destruction:
Anyway, I would appreciate hearing your thoughts on this area as I'd
like to use version 2!
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk