Re: [PATCH 0/3] arm64: cpuinfo: make /proc/cpuinfo more human-readable

[Mark Rutland]

Hi Timur,

On Fri, Oct 13, 2017 at 08:39:09AM -0500, Timur Tabi wrote:
> On Wed, Sep 27, 2017 at 5:34 AM, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > On Tue, Sep 26, 2017 at 04:23:21PM -0600, Al Stone wrote:
> >> As ARMv8 servers get deployed, I keep getting the same set of questions
> >> from end-users of those systems: what do all the hex numbers mean in
> >> /proc/cpuinfo and could you make them so I don't have to carry a cheat
> >> sheet with me all the time?
> >
> > I appreciate that /proc/cpuinfo can be opaque to end users, but I do not
> > believe that this is the right solution to that problem.
> >
> > There are a number of issues stemming from the face that /proc/cpuinfo
> > is ill-defined and overused for a number of cases. Changes to it almost
> > certainly violate brittle de-facto ABI details people are relying on,
> > and there's a very long tail on fallout resulting from this. In
> > addition, many niceties come at an excessive maintenance cost, and are
> > simply unreliable as an ABI.
> >
> > So, as with all other patches modifying /proc/cpuinfo, I must NAK this
> > series.
> 
> Qualcomm Datacenter Technologies is very interested in seeing these
> patches (or some variant of them) accepted. Updates to /proc/cpuinfo
> are long overdue, and I'm asking you to reconsider your objections.

I more than appreciate that there is information that people want access
to, and I'd really like to expose that in a consistent and structured
manner.

So please propose exposing the information elsewhere, which I would be
more than happy with for information that the kernel already has access
to.

I object to /proc/cpuinfo changes because they impose an ABI break, and
because some of the proposed changes impose new and requirements on the
kernel that cannot be maintained in a forwards-compatible manner (which
is liable to result in other ABI breaks).

For better or worse, we're stuck with the existing arch-specific format
of /proc/cpuinfo, as this is parsed by a wide variety of applications,
and any change risks breaking these.

> We're willing to work with distro vendors to get this information
> added to their products while upstream is left behind, but I hope that
> won't be necessary.

Please understand that this is an ABI break, and that is why it is being
NAK'd.

I don't have the power to stop Red Hat or other vendors from
deliberately breaking ABI, but I would very strongly recommend that they
do not.

> I would even go so far as to say that we should be making
> /proc/cpuinfo for ARM match the x86 output as closely as possible,

As has been stated before, /proc/cpuinfo is an existing arch-specific ABI.

For better or worse, we're stuck with it as-is, regardless of what could
be nicer had we done something different from the outset.

It has never been portable across architectures, and applications
relying upon it have never been portable except by chance.

> even using their terminology.  We should be providing information like
> frequencies and product names.

As has been stated before, the problem with exposing these is that we
don't have the relevant information.

We have no consistent mechanism for acquiring product names, and for
those cases where it is truly necessary to identify an implementation,
the MIDR+REVIDR provide the exact same information.

For frequencies, I'd be more than happy to expose this information when
we have it, in new files. In practice today, we do not have this
information most of the time.

> Having a human-readable /proc/cpuinfo with extensive details of the
> CPU spelled out is very useful, and Al's reasoning is valid.  The fact
> that it's "fragile" is not as important as the fact that on x86,
> /proc/cpuinfo is much more useful.

I must disagree.

We care about that fragility so that we do not break userspace, which is
the most important upstream rule.

I certainly agree that exposing the information that we have is useful,
as I have stated several times. I'm not NAKing exposing this information
elsewhere.

If you want a consistent cross-architecture interface for this
information, then you need to propose a new one. That was we can
actually solve the underlying issues, for all architectures, without
breaking ABI.

I would be *very* interested in such an interface, and would be more
than happy to help.

Thanks,
Mark.