[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <8421f726-0ef0-f253-4d98-ec8a4556cb8e@siemens.com>
Date: Wed, 4 Mar 2020 19:39:58 +0100
From: Jan Kiszka <jan.kiszka@...mens.com>
To: Dave Hansen <dave.hansen@...el.com>, x86 <x86@...nel.org>
Cc: Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: x2apic_wrmsr_fence vs. Intel manual
On 04.03.20 19:27, Dave Hansen wrote:
> On 3/2/20 8:11 AM, Jan Kiszka wrote:
>> The former dates back to ce4e240c279a, but that commit does not mention
>> why lfence is not needed. Did the manual read differently back then? Or
>> why are we safe? To my reading of lfence, it also has a certain
>> instruction serializing effect that mfence does not have.
>
> I asked around Intel about this.
>
> The old "SFENCE, or MFENCE" recommendation was deemed insufficient
> because it has no impact on the ordering of WRMSR since it is not a
> "load or store instruction". LFENCE's instruction-ordering semantic is
> needed because it ensures later ordering of all instructions, not just
> loads and stores.
>
> Jan, do you think you're seeing a bug resulting from WRMSR ordering?
>
Nope, not so far. I'm hunting a race between two guests over Jailhouse
where the kick (sent as IPI) seems to come before the data, but changing
the fences didn't solve it, unfortunately. Along that, I was reading
code and manuals up and down and ran into this inconsistency. That's the
story.
Jan
--
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux
Powered by blists - more mailing lists