Add CPUID subleaf capability to cpuctl/cpucontrol

[Konstantin Belousov]

On Thu, Mar 27, 2014 at 12:45:59PM +0800, Jia-Shiun Li wrote:
> On Wed, Mar 26, 2014 at 11:54 PM, Konstantin Belousov
> <kostikbel at gmail.com> wrote:
> > On Wed, Mar 26, 2014 at 11:18:02PM +0800, Jia-Shiun Li wrote:
> >> Hi all,
> >>
> >> I am recently writing a small tool playing msr with cpuctl(4).
> >> Meanwhile I found that it is currently not passing value but 0 in ECX
> >> register to CPUID instruction as input. So I have the attached patch
> >> to do it.
> >>
> >> ECX is used to specify sub-leaf for some EAX leaf value. For example
> >>   EAX=0x04: Deterministic Cache Parameters Leaf
> >>   EAX=0x0b: Extended Topology Enumeration Leaf
> >> with the attached patch user will be able to get subleaf info by
> >> writing applications using cpuctl(4) ioctl or by using cpucontrol(8)
> >> with -s flag.
> >>
> >> Please comment.
> >>
> >> Regards,
> >> Jia-Shiun.
> >
> >> Index: sys/dev/cpuctl/cpuctl.c
> >> ===================================================================
> >> --- sys/dev/cpuctl/cpuctl.c   (revision 263420)
> >> +++ sys/dev/cpuctl/cpuctl.c   (working copy)
> >> @@ -204,7 +204,7 @@
> >>       oldcpu = td->td_oncpu;
> >>       is_bound = cpu_sched_is_bound(td);
> >>       set_cpu(cpu, td);
> >> -     cpuid_count(data->level, 0, data->data);
> >> +     cpuid_count(data->level, data->sublevel, data->data);
> >>       restore_cpu(oldcpu, is_bound, td);
> >>       return (0);
> >>  }
> >> Index: sys/sys/cpuctl.h
> >> ===================================================================
> >> --- sys/sys/cpuctl.h  (revision 263420)
> >> +++ sys/sys/cpuctl.h  (working copy)
> >> @@ -36,6 +36,7 @@
> >>
> >>  typedef struct {
> >>       int             level;  /* CPUID level */
> >> +     int             sublevel; /* sublevel */
> >>       uint32_t        data[4];
> >>  } cpuctl_cpuid_args_t;
> >
> > This breaks the ABI.  If you want to extend the CPUCTL_CPUID,
> > define new ioctl number for the extended structure, and keep the
> > old arg structure and old number handled by the compat shims. Due
> > to the construction of the CPUCTL_CPUID using _IOWR(), you in fact
> > get the new number due to the structure size change.
> >
> > But IMO adding this functionality to a driver does not make much sense
> > at all.  I remember that at time when initial version of cpuctl(4)
> > was written, we did not have useful implementation of cpuset(2).
> > Now, when it is possible to bind thread to a core from usermode,
> > you do not need a driver providing interface to CPUID, since CPUID
> > is usermode instruction.
> 
> 
> thank you. I did not realize the ABI issue. I will try using affinity
> to get CPUID
> info in user space. It looks better with both regards.

From what I remember, in the multi-socket configuration Intel allows
CPUs to only differ by stepping. Due to this, I see only one possible
uses of executing CPUID on specific core, to get the stepping of the
package, to either test for the presence of the specific CPU bug, or to
select the proper microcode update.

That said, I am curious why do you want this feature.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 834 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20140327/267d5c9f/attachment.sig>