Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:3 as permitted sender) client-ip=2620:137:e000::3:3;
Date:   Wed, 29 Nov 2023 11:22:10 -0800
In-Reply-To: <22c602c9-4943-4a16-a12e-ffc5db29daa1@intel.com>
Mime-Version: 1.0
References: <20231108010953.560824-1-seanjc@google.com> <20231108010953.560824-3-seanjc@google.com>
 <0ee32216-e285-406f-b20d-dd193b791d2b@intel.com> <ZUuyVfdKZG44T1ba@google.com>
 <22c602c9-4943-4a16-a12e-ffc5db29daa1@intel.com>
Message-ID: <ZWePYnuK65GCOGYU@google.com>
Subject: Re: [PATCH v2 2/2] KVM: selftests: Add logic to detect if ioctl()
 failed because VM was killed
From:   Sean Christopherson <seanjc@google.com>
To:     Xiaoyao Li <xiaoyao.li@intel.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org, Michal Luczaj <mhal@rbox.co>,
        Oliver Upton <oliver.upton@linux.dev>,
        Colton Lewis <coltonlewis@google.com>
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Mon, Nov 13, 2023, Xiaoyao Li wrote:
> On 11/9/2023 12:07 AM, Sean Christopherson wrote:
> > On Wed, Nov 08, 2023, Xiaoyao Li wrote:
> > > On 11/8/2023 9:09 AM, Sean Christopherson wrote:
> > > > Add yet another macro to the VM/vCPU ioctl() framework to detect when an
> > > > ioctl() failed because KVM killed/bugged the VM, i.e. when there was
> > > > nothing wrong with the ioctl() itself.  If KVM kills a VM, e.g. by way of
> > > > a failed KVM_BUG_ON(), all subsequent VM and vCPU ioctl()s will fail with
> > > > -EIO, which can be quite misleading and ultimately waste user/developer
> > > > time.
> > > > 
> > > > Use KVM_CHECK_EXTENSION on KVM_CAP_USER_MEMORY to detect if the VM is
> > > > dead and/or bug, as KVM doesn't provide a dedicated ioctl().  Using a
> > > > heuristic is obviously less than ideal, but practically speaking the logic
> > > > is bulletproof barring a KVM change, and any such change would arguably
> > > > break userspace, e.g. if KVM returns something other than -EIO.
> > > 
> > > We hit similar issue when testing TDX VMs. Most failure of SEMCALL is
> > > handled with a KVM_BUG_ON(), which leads to vm dead. Then the following
> > > IOCTL from userspace (QEMU) and gets -EIO.
> > > 
> > > Can we return a new KVM_EXIT_VM_DEAD on KVM_REQ_VM_DEAD?
> > 
> > Why?  Even if KVM_EXIT_VM_DEAD somehow provided enough information to be useful
> > from an automation perspective, the VM is obviously dead.  I don't see how the
> > VMM can do anything but log the error and tear down the VM.  KVM_BUG_ON() comes
> > with a WARN, which will be far more helpful for a human debugger, e.g. because
> > all vCPUs would exit with KVM_EXIT_VM_DEAD, it wouldn't even identify which vCPU
> > initially triggered the issue.
> 
> It's not about providing more helpful debugging info, but to provide a
> dedicated notification for VMM that "the VM is dead, all the following
> command may not response". With it, VMM can get rid of the tricky detection
> like this patch.

But a VMM doesn't need this tricky detection, because this tricky detections isn't
about detecting that the VM is dead, it's all about helping a human debug why a
test failed.

-EIO already effectively says "the VM is dead", e.g. QEMU isn't going to keep trying
to run vCPUs.  Similarly, selftests assert either way, the goal is purely to print
out a unique error message to minimize the chances of confusing the human running
the test (or looking at results).

> > Definitely a "no" on this one.  As has been established by the guest_memfd series,
> > it's ok to return -1/errno with a valid exit_reason.
> > 
> > > But I'm wondering if any userspace relies on -EIO behavior for VM DEAD case.
> > 
> > I doubt userspace relies on -EIO, but userpsace definitely relies on -1/errno being
> > returned when a fatal error.
> 
> what about KVM_EXIT_SHUTDOWN? Or KVM_EXIT_INTERNAL_ERROR?

I don't follow, those are vcpu_run.exit_reason values, not errno values.  Returning
any flavor of KVM_EXIT_*, which are positive values, would break userspace, e.g.
QEMU explicitly looks for "ret < 0", and glibc only treats small-ish negative
values as errors, i.e. a postive return value will be propagated verbatim up to
QEMU.