Message-ID: <ada3e064-fdb3-afc5-8c0c-bff54a591189@intel.com>
Date:   Tue, 14 Mar 2023 09:01:18 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.7.1
Subject: Re: [PATCH] x86/amd: Work around Erratum 1386 - XSAVES malfunction on
 context switch
Content-Language: en-US
To:     Andrew Cooper <andrew.cooper3@citrix.com>,
        Borislav Petkov <bp@alien8.de>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        Tavis Ormandy <taviso@gmail.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>,
        Dave Hansen <dave.hansen@linux.intel.com>, x86@kernel.org,
        "H. Peter Anvin" <hpa@zytor.com>,
        Alexander Monakov <amonakov@ispras.ru>
References: <Y/W4x7/KFqmDmmR7@thinkstation.cmpxchg8b.net>
 <20230307174643.1240184-1-andrew.cooper3@citrix.com>
 <20230307175050.GCZAd5eu0/Mk2fdLz5@fat_crate.local>
 <940596cc-a440-181a-a72a-36282a26dd0a@citrix.com>
 <20230307185632.GDZAeI4LwBDUU3/OP9@fat_crate.local>
 <ee866b3e-a53b-4c26-0272-20e798c0650d@citrix.com>
From:   Dave Hansen <dave.hansen@intel.com>
In-Reply-To: <ee866b3e-a53b-4c26-0272-20e798c0650d@citrix.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 3/7/23 12:01, Andrew Cooper wrote:
> On 07/03/2023 6:56 pm, Borislav Petkov wrote:
>> On Tue, Mar 07, 2023 at 06:22:01PM +0000, Andrew Cooper wrote:
>>> Well yes - more and more reports of impacted systems.
>>>
>>> Having the full list of affected models is great and all, but how is it
>>> going to change this patch as a workaround ?
>> We don't have to clear the feature flag on those systems which have
>> a fix.
> Sure, but why is that helpful?
> 
> XSAVES and XSAVEC are functionally identical on Zen1/2 because these
> CPUs don't advertise any supervisor XSAVE states.

On Intel at least, XSAVES uses the modified optimization and XSAVEC does
not.  I'm not sure if you include this in your definition of
"functionally identical", but it is one of those little XSAVE
architecture oddities that's bitten me a time or two.

The AMD docs don't talk about the modified optimization verbatim, but I
think this nugget is alluding to the same behavior:

> 4. XRSTOR loads an internal state value XRSTOR_INFO that can be used to further optimize a
> subsequent XSAVEOPT or XSAVES. This reflects the current privilege level and virtualization
> mode as well as the save area's base address and XCOMP_BV field.