Message-ID: <20190205170415.GG17550@hirez.programming.kicks-ass.net>
Date:   Tue, 5 Feb 2019 18:04:15 +0100
From:   Peter Zijlstra <peterz@...radead.org>
To:     Dave Hansen <dave.hansen@...el.com>
Cc:     Borislav Petkov <bp@...en8.de>,
        Thomas Gleixner <tglx@...utronix.de>,
        Fenghua Yu <fenghua.yu@...el.com>,
        Ingo Molnar <mingo@...hat.com>, H Peter Anvin <hpa@...or.com>,
        Ashok Raj <ashok.raj@...el.com>,
        Michael Chan <michael.chan@...adcom.com>,
        Ravi V Shankar <ravi.v.shankar@...el.com>,
        Ricardo Neri <ricardo.neri@...el.com>,
        linux-kernel <linux-kernel@...r.kernel.org>, x86 <x86@...nel.org>
Subject: Re: [PATCH v3 08/10] x86/setcpuid: Add kernel option setcpuid

On Tue, Feb 05, 2019 at 07:19:16AM -0800, Dave Hansen wrote:
> On 2/5/19 12:48 AM, Peter Zijlstra wrote:
> > On Mon, Feb 04, 2019 at 12:46:30PM -0800, Dave Hansen wrote:
> >> So, the compromise we reached in this case is that Intel will fully
> >> document the future silicon architecture, and then write the kernel
> >> implementation to _that_.  Then, for the weirdo deployments where this
> >> feature is not enumerated, we have the setcpuid= to fake the enumeration
> >> in software.
> > 
> > What user is _EVER_ going to use this? Nobody, I expect the answer to
> > be.
> 
> This is one of the few times that we're pretty confident that folks will
> use this.  The reason we're going to this trouble is that the split lock
> detection is wanted by actual customers, and they want it before it's
> implemented on a processor with real enumeration.

That's big customers that do magic stuff not users.

> This isn't something we want everybody and their grandma to turn on;
> it's a rather specialized feature.  It's really only for folks that care
> about the latency incurred across the entire system on split lock
> operations.

That really should be everyone. That split lock stuff is horrible. There
is no real down-side to having it always enabled. Code that breaks is
bad code you want fixed anyway.

Like I said elsewhere, I wish it would #AC for any unaligned LOCK
prefix, not just cross-line. I see why we'd not want to traditional RISC
#AC for every load/store, but atomics really had better be aligned.

> > Is this some transient state; where a few (early) models will not have
> > the enumeration sorted but all later models will have it all neat and
> > tidy?
> 
> From my understanding, it's not just an early stepping.  It's a
> generational thing.  The current generation lacks the enumeration and
> the next generation will get it.  Both have the silicon to implement the
> feature itself.

I never said stepping, in fact I explicitly said model.

> > If so, we can easily do the FMS solution for this.
> 
> Yeah, we can.  I honestly forget why we didn't do FMS. :)

Right so FMS is fairly horrible; but when it is a stop-gap for a limited
number of models it's waaay better than dodgy cmdline things.

We could of course try to wrmsr_safe() detect the feature; but that
might be a problem is the MSR exists on any other models and has a
different meaning.

> > But a cmdline features thing is not something I can see anybody but
> > a limited set of developers ever using.
> 
> It's not for developers.  This really is for (somewhat niche) end users
> that want split lock detection in production.  This is all really an
> effort to get them running mainline or real distro kernels.

Yes, because cloud service providers and RT are niche products.. *sigh*.

Nobody wants to do live audio on their laptops.. oh wait..

Powered by blists - more mailing lists