Messages in this thread | | | Date | Tue, 30 Jun 2015 14:57:36 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH] x86, msr: Allow read access to /dev/cpu/X/msr |
| |
* Peter Zijlstra <peterz@infradead.org> wrote:
> On Tue, Jun 30, 2015 at 08:20:55AM -0400, Prarit Bhargava wrote: > > it seems like visiting changes on each of these packages (and the other > > packages that I'm sure I've missed) will be moderately difficult. > > > > Thoughts? > > Start by changing the ones users want to run most and leave the rest > requiring root privs until someone has the time to convert them.
Yes, start by just converting the ones used by a single tool, say turbostat, and then convert turbostat (while keeping /dev/msr fall-back code as well) just to see what the tooling fallout is.
> Its not like we can remove the msr driver any time soon anyway.
Yes.
> So I would suggest starting with the perf MSR driver thingy for all > those MSRs that count things and see if you can convert say > turbostat/cpufrequtils/powertop over to that.
One of those tools would be enough, to keep the complexity of the initial submission low - and to allow a change of plans if necessary.
> I suspect there's MSR that are useful to expose but are not counting, I'm not > sure perf is the right interface for those. > > Making an inventory on which MSRs are required by these tools and what kind of > data they provide might give a good idea on how to continue. > > If most of these tools only use counting MSRs that can be serviced with the > perf-msr driver then that would be great.
Fully agreed!
Thanks,
Ingo
| |