Re: [Qemu-devel] Adding an IPMI BMC device to KVM

[Corey Minyard <tcminyard@gmail.com>]

On 05/18/2012 08:08 AM, Stefan Hajnoczi wrote:
> On Mon, May 7, 2012 at 3:30 PM, Anthony Liguori<anthony@codemonkey.ws>  wrote:
>> On 05/06/2012 09:39 AM, Avi Kivity wrote:
>>> On 05/06/2012 05:35 PM, Anthony Liguori wrote:
>> So what's really the use case here?  Would an IPMI ->  libvirt bridge get you
>> what you need?  I really think that's the best path forward.
Sorry for sending this twice.  I thought my mailer was set to disable HTML.

Many of our customer use the standard watchdog timer and power controls 
on IPMI.  Basically, the watchdog timer needs to be able to power off, 
power cycle, reset, and send an NMI.  The IPMI watchdog can be 
configured to do all those things.  (The NMI is for a "pretimeout" that 
generally triggers a panic.)

Power control is no big surprise.  It's true that you have ACPI to do 
this, but that's not terribly useful on non-x86 (and non ia64, I 
suppose) systems.

> Do you want to expose host sensors - the challenge is that they don't
> reflect the hardware that the virtual machine sees?  Or do you want to
> have synthetic sensors - which virtual sensors are useful to have?

A few sensors, notably the watchdog timer sensor, are not synthetic.  
The others generally are.  Sensors cover things beyond just power and 
temperature.  Important other ones in IPMI deal with the presence of 
FRUs in the system, BIOS/OS state, and intrusion detection.

Two main reasons for synthesized sensors exist.  One, in a legacy 
situation, is to "fool" the management system into thinking everything 
is ok, since it is expecting to see these sensors with specific values.  
You could even possibly reflect the state of real sensors, mapping them 
somehow, if you wanted.

The other reason is for testing.  It's pretty hard to induce some of 
these sensors to go out of range in a real system.  So without 
simulation, you need a hack-ed up management controller in a real system 
to truly test your software.

> Is the IPMI watchdog useful, QEMU already supports the i6300esb PCI
> watchdog (see qemu -watchdog option documentation)?

That's only useful if you can simulate an i6300esb.  Probably not 
possible on non-x86.  Plus legacy systems may be expecting the 
capabilities of the IPMI watchdog.

> Some use cases that illustrate how the guest is going to use IPMI
> would be interesting and could help guide the discussion.

You are probably right.  I've mentioned a couple above.

One other possible one, that someone else mentioned, is the ability to 
control a VM using standard tools like ipmitool or OpenIPMI over network 
interfaces.  People may already have management systems that are 
designed around IPMI, and it would make a move to virtual machines 
easier.  This was the request mentioned that was in the Redhat database.

Such a capability would require re-thinking things a bit.  You have two 
basic options that I see.  Either have the "management controller" run 
outside qemu and define a simple interface to it, or modify qemu to be 
able to run the management controller internally.

My preference in to run it outside qemu, for the following reasons:

  * It decouples things that are IPMI internals from having to go
    through getting into qemu repositories.
  * The management controller then becomes useful for a number of other
    purposes.  I already have one mostly done that I use for testing. 
    It could be used by other VMs.
  * A full management controller is a fairly big piece of software,
    especially if you include the network access.  It probably doesn't
    belong in qemu.
  * No offense, I don't want to muck around inside qemu to accomplish
    this :).

There are, of course, disadvantages.  Some I can think of:

  * There can be confusion about which versions of two things work together.
  * Overall, it is more complex to get working.
  * There are possible security implications.

So I guess those are points we can talk about...

-corey