[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <87wnl74qso.ffs@tglx>
Date: Wed, 17 Nov 2021 00:35:51 +0100
From: Thomas Gleixner <tglx@...utronix.de>
To: Sean Christopherson <seanjc@...gle.com>
Cc: "Liu, Jing2" <jing2.liu@...el.com>,
Paolo Bonzini <pbonzini@...hat.com>,
LKML <linux-kernel@...r.kernel.org>,
"x86@...nel.org" <x86@...nel.org>,
"kvm@...r.kernel.org" <kvm@...r.kernel.org>,
"Nakajima, Jun" <jun.nakajima@...el.com>,
Dave Hansen <dave.hansen@...ux.intel.com>,
Arjan van de Ven <arjan@...ux.intel.com>,
Jing Liu <jing2.liu@...ux.intel.com>,
"Cooper, Andrew" <andrew.cooper3@...rix.com>,
"Bae, Chang Seok" <chang.seok.bae@...el.com>
Subject: Re: Thoughts of AMX KVM support based on latest kernel
On Tue, Nov 16 2021 at 20:12, Sean Christopherson wrote:
> On Tue, Nov 16, 2021, Thomas Gleixner wrote:
>> Now you could argue that the interrupt/softirq XSAVES should also read
>> the XFD MSR and save it in guest_fpstate.xfd. Same in schedule()
>> and kvm_put_guest_fpu(), i.e:
>>
>> XSAVES
>> if (fpstate->is_guest) {
>> rdmsrl(XFD, xfd);
>> fpstate->xfd = xfd;
>> __this_cpu_write(..., xfd);
>> }
>>
>> We can do that, but I'm unhappy about this conditional in schedule(). So
>> I was asking for doing a simple KVM only solution first:
>
> Ah, the schedule() conditional is the part I was missing. Thanks!
As I was missing the preempt notifier... which makes it a different
story, but I still think that the simple version is good enough at least
for a start.
Thanks,
tglx
Powered by blists - more mailing lists