cvs commit: src/sys/netinet ip_fw2.c

[Robert Watson]

On Tue, 30 Sep 2008, Roman Divacky wrote:

>> Good news.  We'll want to keep an eye on this one as the 7.0 release cycle 
>> progresses, and there may be other unexpected edge case problems from the 
>> rwlock change.  On the whole it seems to have been very successful, but the 
>> view that -CURRENT doesn't receive a whole lot of stress testing is 
>> reinforced...
>
> I think this is a little different case... I guess people are willing to 
> test -CURRENT on their desktops etc. but not on "servers". ie. when you have 
> immediate access to the machine you easily use -CURRENT but not on the 
> remote server.
>
> Also, people don't tend to run firewalls on their desktops (as opposed to 
> servers where they dont). This is why I think this bug slipped. Not that 
> -CURRENT is so badly tested...

Common code paths are reasonably tested by -CURRENT, but coverage of 
not-so-common paths is basically ignored.  Experience with TCP regressions 
last year suggests that little or no interoperability occurs in HEAD, and that 
it's easy for quite serious regressions in stability and performance to go 
unnoticed for extended periods.

This isn't indicative of a "problem", in that it is expected that normal user 
populations don't track the -CURRENT branch, it isn't put into heavy 
production use, etc, but is something we need to recognize: three months in 
HEAD is not enough to declare a feature ready for production.

Robert N M Watson
Computer Laboratory
University of Cambridge