Add support to debug unresponsive host

[Andrew Geissler]

On Thu, May 16, 2019 at 1:36 AM Deepak Kodihalli
<dkodihal at linux.vnet.ibm.com> wrote:
>
> On 15/05/19 6:09 PM, Jayanth Othayoth wrote:
> > ## Problem Description
> > Issue #457:  Add support to debug unresponsive host.
> >
> > Scope: High level design direction to solve this problem,
> >
> > ## Background and References
> > There are situation at customer places where OPAL/Linux goes
> > unresponsive causing a system hang. And there is no way to figure out
> > what went wrong with Linux kernel or OPAL. Looking for a way to trigger
> > a dump capture on Linux host so that we can capture the OS dump for post
> > analysis.
> >
> > ## Proposed Design for POWER processor based systems:
> > Get all Host CPUs in reset vector and Linux then has a mechanism to
> > patch it into panic-kdump path to trigger dump capture. This will enable
> > us to analyze and fix customer issue where we see Linux hang and
> > unresponsive system.
> >
> > ### Redfish Schema used:
> > * Reference: DSP2046 2018.3,
> > * ComputerSystem 1.6.0 schema provides an action called
> > #ComputerSystem.Reset”, This action is used to reset the system.
> > ResetType parameter is used  for indicating type of reset need to be
> > performed. In this use case we can use “Nmi” type
> >      * Nmi: Generate a Diagnostic Interrupt (usually an NMI on x86
> > systems) to cease normal operations, perform diagnostic actions and
> > typically halt the system.
> > * ### d-bus :
> >
> > Option 1:   Extending  the existing  d-bus interface  state.Host  name
> > space (
> > /openbmc/phosphor-dbus-interfaces/xyz/openbmc_project/State/Host.interface.yaml
> > ) to support new RequestedHostTransition property called  “Nmi”.   d-bus
> > backend can internally invoke processor specific target to do Sreset(
> > equivalent to x86 NMI) and associated  actions.
>
> I don't prefer this option, because this would mean adding host specific
> code in phoshor-state-manager, which I think until now is host agnostic.

Yeah, this was my main concern with tying it into phosphor-state-manager.
The fact Redfish put it in with their other state related commands (which
are implemented by phosphor-state-manager) is the only reason I'm a little
wishy-washy here. We could just create a generic systemd target "host-nmi"
or something and phosphor-state-manager could just call that to abstract
any of the specifics, but it sill doesn't really feel like it fits to me.

I think I prefer option 2, and then we can just map bmcweb to that API when
the Redfish command comes in. Sounds like for ppc64 systems we can just
use pdbg to issue the NMI.

> So for that reason, Option 2 sounds better. There are some good
> questions from Neeraj as well, so I would suggest adding this as a
> design template on Gerrit to gather better feedback.
>
> Thanks,
> Deepak
>
> > Option 2: Introducing new d-bus interface in the control.state namespace
> > (
> > /openbmc/phosphor-dbus-interfaces/xyz/openbmc_project/Control/Host/NMI.interface.yaml)
> > namespace and implement the new d-bus back-end for respective  processor
> > specific targets.
>