svn commit: r243627 - head/sys/kern

Robert Watson rwatson at FreeBSD.org
Wed Nov 28 08:31:02 UTC 2012


On Wed, 28 Nov 2012, Andre Oppermann wrote:

> Yes, and I didn't really expect you to answer (at least quickly) during your 
> FreeBSD hiatus.  So it was seeking review by chance.
>
> Alas I found and fixed the bug myself within 2.5hrs.  While not optimal, a 
> sign of poor prior testing and too much trust into the submitter of the 
> patch it wasn't an earth shattering event.  Doesn't distract from the fact 
> that it was mea culpa in any case though.

The rapid fix was, of course, extremely appreciated :-).

> For prior review of kern_socket* and netinet/tcp_* related changes it has 
> been on and off by various committers over the past year.  If we do have a 
> policy of prior review required then it should be made official and codified 
> in MAINTAINERS and universally applied to all.

I tend to be of the view that 'maintainers' is a bad idea, and that we should 
just make a regular practice of seeking review for this sort of thing, 
especially as our community grows (and, let us be honest, complexity also 
grows -- your observations about decades of accumulated complexity in the TCP 
stack are not amiss).

I'll try to take a look at this change in detail over the weekend -- 
listen/accept locking is a bit of a sore point; in the original design, I 
didn't have a separate accept lock, but ended up being forced to introduce it 
to solve races along these lines.  In the past we've also relied on the 
pcbinfo lock in the protocol providing significant synchronisation during new 
connection events, and as we reduce the influence of that lock, finding more 
structured solutions is necessary.

Robert


More information about the svn-src-all mailing list