On Mon, Feb 29, 2016 at 12:53:18PM -0500, George Spelvin wrote:
> I worry that this is this too ambitious a goal. Who is volunteering
> to actually do this?
>From a quick look, the stuff in the examples was already in the rapl
driver.
> It takes quite a while to find a good OS-level abstraction (remember
> wakelocks?), and MSRs are the CPU architect's equivalent of ioctls.
> So they're a bit of a mess, and there will keep being new ones.
And yet you end up needing only a handful in most cases.
> I agree with you about anything that's going to see widespread use, but
> for specialized (apparently mostly HPC) use where the application really
> is heavily optimized for specific CPU models, perhaps dangerous-but-simple
> is good enough?
If it is that specialized, then it doesn't belong upstream.
> The proposed interface is simple and imposes very little maintenance
> burden on the kernel. My main objection is that it's yet another
> special-case permission system. Are we *sure* we'll never want to have
> to classes of users with different access rights?
The proposed interface is the wrong thing to do. There's no need to talk
about how simple and less of a burden it is.
The burden comes when people start complaining about strange issues and
we go and have to get a full MSR dump at the time the explosion happens
because some userspace tool went nuts and scribbled all over them. No
one wants to be on the receiving end of a bug report like this.
--
Regards/Gruss,
Boris.
ECO tip #101: Trim your mails when you reply.
