Bruce M. Simpson
bms at FreeBSD.org
Fri May 30 17:00:56 UTC 2008
> Bruce M. Simpson wrote:
>> It could save dirtying an L2 data cache line at the expense of taking
>> a conditional branch,
> Whoa, why don't you take it easy on me :) I'm not that much into
> kernel (or hardware) programming. It's just that reading Ch. 3 of
> TCP/IP Illustrated Vol.2 by Rich Stevens got me digging around FreeBSD
> source code dealing with struct ifnet, where this piece of code caught
> my attention.
It could be red, it could be yellow. It could be 620nm. Who am I to say
what is and what isn't? ;-)
There are bound to be situations where the change is a win, and even
some where there isn't. Context is everything...
>> but to evaluate your suggested change requires a lot more data. Do
>> you plan to do this?
> Perhaps there is already a framework for trying out changes in
> -CURRENT and seeing their relative impact, so perhaps someone more
> experienced than I am can see to this?
All educators are busy right now, please hold and the next available
dogma merchant will be with you as soon as possible. ;-)
(Hint: No, there isn't a framework I know of, unless you wanna make one?
Scientific process applies, reproducible results, etc. You could script
stuff, figure out a way to run the kernel or parts of the network stack
under Valgrind so it can be L2 profiled w/o running it on a real
machine... or hack hwpmc so it can be done live.. anything is possible.)
>> Given how _IF_DEQUEUE() is normally used the impact is likely
> Oh, I see. A nice first attempt of mine anyway ;) Thanks.
Don't take my word for it, down that road lies darkness.
Seriously though -- it's easy to introduce bugs doing things like this,
if anything else it's an exercise in really thinking things through.
More information about the freebsd-net