Re: [Qemu-devel] [PATCH v6 02/15] spapr_drc: initial implementation of sPAPRDRConnector device

[Michael Roth <mdroth@linux.vnet.ibm.com>]

Quoting Bharata B Rao (2015-02-27 03:52:09)
> On Fri, Feb 27, 2015 at 8:41 AM, Michael Roth <mdroth@linux.vnet.ibm.com> wrote:
> > +static void attach(sPAPRDRConnector *drc, DeviceState *d, void *fdt,
> > +                   int fdt_start_offset, bool coldplug, Error **errp)
> > +{
> > +    DPRINTFN("drc: %x, attach", get_index(drc));
> > +
> > +    if (drc->isolation_state != SPAPR_DR_ISOLATION_STATE_ISOLATED) {
> > +        error_setg(errp, "an attached device is still awaiting release");
> > +        return;
> > +    }
> > +    g_assert(drc->allocation_state == SPAPR_DR_ALLOCATION_STATE_UNUSABLE);
> > +    g_assert(fdt || coldplug);
> > +
> > +    /* NOTE: setting initial isolation state to UNISOLATED means we can't
> > +     * detach unless guest has a userspace/kernel that moves this state
> > +     * back to ISOLATED in response to an unplug event, or this is done
> > +     * manually by the admin prior. if we force things while the guest
> > +     * may be accessing the device, we can easily crash the guest, so we
> > +     * we defer completion of removal in such cases to the reset() hook.
> > +     */
> > +    drc->isolation_state = SPAPR_DR_ISOLATION_STATE_UNISOLATED;
> > +    drc->allocation_state = SPAPR_DR_ALLOCATION_STATE_USABLE;
> 
> BTW shouldn't isolation_state and allocation_state be set to
> UNISOLATED and USABLE in response to guest's rtas-set-indicator calls
> ? At least, that is how it is done/expected for CPU hotplug case. Ref:

Well, for PCI we always start in USABLE (state diagram in PAPR+ 13.4), and
stay there. So set_allocation_state() should actually be a no-op for PCI;
there's no valid transition to UNUSABLE (or EXCHANGE/RECOVER).

For non-PCI, it looks like we should start in UNUSABLE, and let RTAS take it
from there. It also looks like the resource removal shouldn't be finalized
until there's a transition from ISOLATED/USABLE to ISOLATED/UNUSABLE.

I just pushed a patch here that I'm hoping will fix the non-PCI cases. Can
you give it a spin and let me know?

https://github.com/mdroth/qemu/commit/d91b0c6d6f794292e384cf129368aaae90294f5b

> kernel code arch/powerpc/platforms/pseries/dlpar.c:
> dlpar_acquire[release]_drc().
> 
> Same question for detach().
> 
> Regards,
> Bharata.