Re: [RFC PATCH v2] tools/x86: add kcpuid tool to show raw CPU features

[Feng Tang]

Hi Arvind,

On Tue, Sep 22, 2020 at 06:15:23PM -0400, Arvind Sankar wrote:
> On Tue, Sep 22, 2020 at 10:10:24PM +0200, Borislav Petkov wrote:
> > + AMD folks.
> > 
> > On Tue, Sep 22, 2020 at 01:27:50PM +0800, Feng Tang wrote:
> > > End users frequently want to know what features their processor
> > > supports, independent of what the kernel supports.
> > > 
> > > /proc/cpuinfo is great. It is omnipresent and since it is provided by
> > > the kernel it is always as up to date as the kernel. But, it could be
> > > ambiguous about processor features which can be disabled by the kernel
> > > at boot-time or compile-time.
> > > 
> > > There are some user space tools showing more raw features, but they are
> > > not bound with kernel, and go with distros. Many end users are still
> > > using old distros with new kernels (upgraded by themselves), and may
> > > not upgrade the distros only to get a newer tool.
> > > 
> > > So here arise the need for a new tool, which
> > >   * Shows raw cpu features got from running cpuid
> > >   * Be easier to obtain updates for compared to existing userspace
> > >     tooling (perhaps distributed like perf)
> > >   * Inherits "modern" kernel development process, in contrast to some
> > >     of the existing userspace cpuid tools which are still being developed
> > >     without git and distributed in tarballs from non-https sites.
> > >   * Can produce output consistent with /proc/cpuinfo to make comparison
> > >     easier.
> 
> Rather than a tool, would additional file(s) in, say,
> /sys/devices/system/cpu/cpu<n> be nicer? They could show the raw CPUID
> features, one file per leaf or sub-leaf, maybe even along with whether
> they were disabled at boot-time.

My thought is we already have in-kernel powerful /proc/cpuinfo, while 
a user space tool could be more flexible for text parsing/layout, and
show different info on user's demand/options.

> > >   * Be in-kernel, could leverage kernel enabling, and even
> > >     theoretically consume arch/x86/boot/cpustr.h so it could pick up
> > >     new features directly from one-line X86_FEATURE_* definitions.
> 
> That's arch/x86/include/asm/cpufeatures.h right -- cpustr.h is generated
> from that. The table there already has comments which could be extracted
> as the one-line description.

Thanks for the hint! I found the comments in cpufeatures.h is much better
than what I extraced from SDM :), which I should use instead.

One other thing as Boris has mentioned, cpu feature is mixture of raw
silicon features and kernel software ones. Also, cpufeatures.h only
contains shows ont-bit boolean flag, while cpuid has multiple-bits field
containing numbers.

Thanks,
Feng