RE: Thoughts of AMX KVM support based on latest kernel

From: "Tian, Kevin" <kevin.tian@intel.com>
To: Paolo Bonzini <pbonzini@redhat.com>,
	"Nakajima, Jun" <jun.nakajima@intel.com>,
	Thomas Gleixner <tglx@linutronix.de>
Cc: "Liu, Jing2" <jing2.liu@intel.com>,
	LKML <linux-kernel@vger.kernel.org>,
	"x86@kernel.org" <x86@kernel.org>,
	"kvm@vger.kernel.org" <kvm@vger.kernel.org>,
	Dave Hansen <dave.hansen@linux.intel.com>,
	"Arjan van de Ven" <arjan@linux.intel.com>,
	Jing Liu <jing2.liu@linux.intel.com>,
	"seanjc@google.com" <seanjc@google.com>,
	"Cooper, Andrew" <andrew.cooper3@citrix.com>,
	"Bae, Chang Seok" <chang.seok.bae@intel.com>
Subject: RE: Thoughts of AMX KVM support based on latest kernel
Date: Wed, 17 Nov 2021 10:15:32 +0000	[thread overview]
Message-ID: <BN9PR11MB5433D6A1368904D9B748846B8C9A9@BN9PR11MB5433.namprd11.prod.outlook.com> (raw)
In-Reply-To: <2e2ab02c-b324-e136-924a-0376040163a8@redhat.com>

> From: Paolo Bonzini <pbonzini@redhat.com>
> Sent: Wednesday, November 17, 2021 3:31 PM
> 
> On 11/17/21 05:52, Nakajima, Jun wrote:
> > If a (creative) guest wants to set XFD[AMX] = 1 for fun while keeping
> > AMX state alive without saving the AXM state, it may lose the state
> > after VM exit/entry.
> 
> I think this should not happen, unless you also document that other
> random events (hypothetically, it could be some other core using AMX?)
> can cause the loss of XTILEDATA if XFD[AMX]=1.  Virtualization should
> not be special, I'd prefer that the guest has the same behavior as bare
> metal in this respect.
> 

The state may be lost with the simple version suggested by Thomas,
because with XFD[AMX]=1 the AMX state won't be stored to 
guest_fpstate when the vcpu thread is preempted and then get lost
when restored. To emulate the hardware behavior this will need
additional trick to force XFD[AMX]=0 if XGETBV(1) is true (ignoring 
guest XFD) in fpu_update_guest_xfd_state(void). It also implies 
that the actual guest XFD needs to be saved in a place before forcing 
XFD[AMX] to 0 and recovered (after interrupt is disabled) before 
entering the guest.

We are not sure whether such trick is worthwhile, since a sane
guest shouldn't set XFD[AMX]=1 before storing the AMX state. This
is why we want to seek SDM change to mark out that the software
should not assume XTILEDATA is still valid when XFD[AMX]=1. 

Then KVM can make a simple assumption that once XFD[AMX]=1
it implies the guest doesn't care about the AMX state. Then the
simple version from Thomas can work perfectly.

Thanks
Kevin