Received-SPF: pass (google.com: domain of linux-kernel+bounces-143416-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Fri, 12 Apr 2024 15:46:05 -0700
In-Reply-To: <20240412214201.GO3039520@ls.amr.corp.intel.com>
Precedence: bulk
Mime-Version: 1.0
References: <cover.1708933498.git.isaku.yamahata@intel.com>
 <b9fe57ceeaabe650f0aecb21db56ef2b1456dcfe.1708933498.git.isaku.yamahata@intel.com>
 <0c3efffa-8dd5-4231-8e90-e0241f058a20@intel.com> <20240412214201.GO3039520@ls.amr.corp.intel.com>
Message-ID: <Zhm5rYA8eSWIUi36@google.com>
Subject: Re: [PATCH v19 087/130] KVM: TDX: handle vcpu migration over logical processor
From: Sean Christopherson <seanjc@google.com>
To: Isaku Yamahata <isaku.yamahata@intel.com>
Cc: Reinette Chatre <reinette.chatre@intel.com>, kvm@vger.kernel.org, 
	linux-kernel@vger.kernel.org, isaku.yamahata@gmail.com, 
	Paolo Bonzini <pbonzini@redhat.com>, erdemaktas@google.com, Sagi Shahar <sagis@google.com>, 
	Kai Huang <kai.huang@intel.com>, chen.bo@intel.com, hang.yuan@intel.com, 
	tina.zhang@intel.com, isaku.yamahata@linux.intel.com
Content-Type: text/plain; charset="us-ascii"

On Fri, Apr 12, 2024, Isaku Yamahata wrote:
> On Fri, Apr 12, 2024 at 09:15:29AM -0700, Reinette Chatre <reinette.chatre@intel.com> wrote:
> > > +void tdx_mmu_release_hkid(struct kvm *kvm)
> > > +{
> > > +	while (__tdx_mmu_release_hkid(kvm) == -EBUSY)
> > > +		;
> > >  }
> > 
> > As I understand, __tdx_mmu_release_hkid() returns -EBUSY
> > after TDH.VP.FLUSH has been sent for every vCPU followed by
> > TDH.MNG.VPFLUSHDONE, which returns TDX_FLUSHVP_NOT_DONE.
> > 
> > Considering earlier comment that a retry of TDH.VP.FLUSH is not
> > needed, why is this while() loop here that sends the
> > TDH.VP.FLUSH again to all vCPUs instead of just a loop within
> > __tdx_mmu_release_hkid() to _just_ resend TDH.MNG.VPFLUSHDONE?
> > 
> > Could it be possible for a vCPU to appear during this time, thus
> > be missed in one TDH.VP.FLUSH cycle, to require a new cycle of
> > TDH.VP.FLUSH?
> 
> Yes. There is a race between closing KVM vCPU fd and MMU notifier release hook.
> When KVM vCPU fd is closed, vCPU context can be loaded again.

But why is _loading_ a vCPU context problematic?  If I'm reading the TDX module
code correctly, TDX_FLUSHVP_NOT_DONE is returned when a vCPU is "associated" with
a pCPU, and association only happens during TDH.VP_ENTER, TDH.MNG.RD, and TDH.MNG.WR,
none of which I see in tdx_vcpu_load().

Assuming there is something problematic lurking under vcpu_load(), I would love,
love, LOVE an excuse to not do vcpu_{load,put}() in kvm_unload_vcpu_mmu(), i.e.
get rid of that thing entirely.

I have definitely looked into kvm_unload_vcpu_mmu() on more than one occassion,
but I can't remember off the top of my head why I have never yanked out the
vcpu_{load,put}().  Maybe I was just scared of breaking something and didn't have
a good reason to risk breakage?

> The MMU notifier release hook eventually calls tdx_mmu_release_hkid().  Other
> kernel thread (concretely, vhost krenel thread) can get reference count to
> mmu and put it by timer, the MMU notifier release hook can be triggered
> during closing vCPU fd.
> 
> The possible alternative is to make the vCPU closing path complicated not to
> load vCPU context instead f sending IPI on every retry.