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.
On 11/30/2023 3:22 AM, Sean Christopherson wrote:
> 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.
If -EIO for KVM ioctl denotes "the VM is dead" is to be the officially
announced API, I'm fine.
> 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,
I was trying to ask if KVM_EXIT_SHUTDOWN and KVM_EXIT_INTERNAL_ERROR are
treated as fatal error by userspace.
> 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.
On Thu, Nov 30, 2023, Xiaoyao Li wrote:
> On 11/30/2023 3:22 AM, Sean Christopherson wrote:
> > 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.
>
> If -EIO for KVM ioctl denotes "the VM is dead" is to be the officially
> announced API, I'm fine.
Yes, -EIO is effectively ABI at this point. Though there is the caveat that -EIO
doesn't guarantee KVM killed the VM, i.e. KVM could return -EIO for some other
reason (though that's highly unlikely for KVM_RUN at least).
> > 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,
>
> I was trying to ask if KVM_EXIT_SHUTDOWN and KVM_EXIT_INTERNAL_ERROR are
> treated as fatal error by userspace.
Ah. Not really. SHUTDOWN isn't fatal per se, e.g. QEMU emulates a RESET if a
vCPU hits shutdown. INTERNAL_ERROR isn't always fatal on x86, e.g. QEMU ignores
(I think that's what happens) emulation failure when the vCPU is at CPL > 0 so
that guest userspace can't DoS the VM.