Re: [syzbot] [kvm?] WARNING in __kvm_gpc_refresh

From: Sean Christopherson <seanjc@google.com>
To: David Woodhouse <dwmw2@infradead.org>
Cc: syzbot <syzbot+106a4f72b0474e1d1b33@syzkaller.appspotmail.com>,
	 kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	pbonzini@redhat.com,  syzkaller-bugs@googlegroups.com,
	paul <paul@xen.org>
Subject: Re: [syzbot] [kvm?] WARNING in __kvm_gpc_refresh
Date: Tue, 19 Mar 2024 08:23:30 -0700	[thread overview]
Message-ID: <Zfmt8rxlF1ag1iA_@google.com> (raw)
In-Reply-To: <33bcc5778e39780c6895ffa9f52f4b12cf83ad89.camel@infradead.org>

On Mon, Mar 18, 2024, David Woodhouse wrote:
> On Mon, 2024-03-18 at 14:34 -0700, Sean Christopherson wrote:
> > On Mon, Mar 18, 2024, David Woodhouse wrote:
> > > 
> > >         /* Either gpa or uhva must be valid, but not both */
> > >         if (WARN_ON_ONCE(kvm_is_error_gpa(gpa) == kvm_is_error_hva(uhva)))
> > >                 return -EINVAL;
> > > 
> > > Hm, that comment doesn't match the code. It says "not both", but the
> > > code also catches the "neither" case. I think the gpa is in %rbx and
> > > uhva is in %r12, so this is indeed the 'neither' case.
> > > 
> > > Is it expected that we can end up with a cache marked active, but with
> > > the address not valid? Maybe through a race condition with deactive? or
> > > more likely than that?
> > 
> > It's the darn PV system time MSR, which allows the guest to triggering activation
> > with any GPA value.  That results in the cache being marked active without KVM
> > ever setting the GPA (or any other fields).  The fix I'm testing is to move the
> > offset+len check up into activate() and refresh().
> 
> Not sure I even want a gpc of length 1 to work at INVALID_GPA; I don't
> think it's the offset+length check we want to be looking at?
> 
> If we've activated the gpc with gpa==INVALID_GPA, surely the right

This particular issue isn't due to activating with gpa==INVALID_GPA, it's due to
marking the gpc as active without actually activating it.  The offset+length
check is simply what causes KVM to prematurely bail from activation.

> thing to do is just let it fail (perhaps with an explicit check or just
> letting the memslot lookup fail). After fixing that WARN_ON be
> 
>    if (WARN_ON_ONCE(!kvm_is_error_gpa(gpa) && !kvm_is_error_hva(uhva)))

I really don't want to relax the sanity check, as I feel strongly that KVM needs
an invariant that an active cache is either GPA-based or HVA-based, i.e. that at
least one of GPA or HVA is "valid".  In quotes because the GPA doesn't need to
be fully validated, just something that doesn't trip kvm_is_error_gpa().