Re: [Qemu-devel] [PATCH v5 07/16] spapr_rtas: add ibm, configure-connector RTAS interface

[David Gibson]

On Fri, Feb 27, 2015 at 11:37:56AM -0600, Michael Roth wrote:
> Quoting David Gibson (2015-02-26 23:31:20)
> > On Thu, Feb 26, 2015 at 04:21:55PM -0600, Michael Roth wrote:
> > > Quoting David Gibson (2015-02-24 18:48:23)
> > > > On Tue, Feb 24, 2015 at 02:43:45PM -0600, Michael Roth wrote:
> > > > > Quoting David Gibson (2015-02-24 00:40:32)
> > > > > > On Mon, Feb 16, 2015 at 08:27:43AM -0600, Michael Roth wrote:
> > [snip]
> > > > > > > +    uint64_t wa_offset;
> > > > > > > +    uint32_t drc_index;
> > > > > > > +    sPAPRDRConnector *drc;
> > > > > > > +    sPAPRDRConnectorClass *drck;
> > > > > > > +    sPAPRDRCCResponse resp;
> > > > > > > +    const struct fdt_property *prop = NULL;
> > > > > > > +    char *prop_name = NULL;
> > > > > > > +    int prop_len, rc;
> > > > > > > +
> > > > > > > +    drc_index = rtas_ld(wa_addr, 0);
> > > > > > > +    drc = spapr_dr_connector_by_index(drc_index);
> > > > > > > +    if (!drc) {
> > > > > > > +        DPRINTF("rtas_ibm_configure_connector: invalid 
> > > > > > > sensor/DRC index: %xh\n",
> > > > > > > +                drc_index);
> > > > > > > +        rc = RTAS_OUT_PARAM_ERROR;
> > > > > > > +        goto out;
> > > > > > > +    }
> > > > > > > +    drck = SPAPR_DR_CONNECTOR_GET_CLASS(drc);
> > > > > > > +    resp = drck->configure_connector(drc, &prop_name, &prop, 
> > > > > > > &prop_len);
> > > > > > 
> > > > > > You may have answered this last time round, but if so I forgot the
> > > > > > reason.
> > > > > > 
> > > > > > Why does the awkward iteration need to go down to the drck callback?
> > > > > > Coudln't the drck callback part just supply the fdt fragment blob,
> > > > > > then have generic code which streams it out via iteration?
> > > > > > 
> > > > > > Obviously we have to support the horrid PAPR interface, but it would
> > > > > > be nice to confine the PAPR derived horridness to as small an area 
> > > > > > as
> > > > > > we can.
> > > > > 
> > > > > That horrid interface percolates all the way up the QEMU stack,
> > > > > unfortunately :)
> > > > > 
> > > > > Upon successfully having it's device tree node received, a DRC 
> > > > > transitions
> > > > > to a 'configured' state that's defined in the DR state machine (PAPR+ 
> > > > > 13.4).
> > > > > 
> > > > > We need to track that state, since it's used to differentiate between
> > > > > a case where a device is set to 'isolated' as part of 
> > > > > entity-sense/device
> > > > > configuration, as opposed to 'isolated' as part the unplug path. The
> > > > > overlap between the 2 can occur if we do device_add followed by an 
> > > > > immediate
> > > > > device_del, but since the 'configured' transition must occur before 
> > > > > the latter,
> > > > > it becomes unambiguous.
> > > > > 
> > > > > It's also possible that a guest might be reset in the middle of a 
> > > > > series of
> > > > > calls to configure-connector, in which case that state needs to be 
> > > > > reset. This
> > > > > is currently handled by sPAPRDRConnector's reset hook, so if we moved 
> > > > > that
> > > > > aspect out I think we'd need to wire up a reset hook for the 
> > > > > configure-connector
> > > > > state, which is kinda messy. We'd also need a list of some sort, 
> > > > > keyed by
> > > > > the DRC indexes, to handle the subsequent call-per-iteration's (no 
> > > > > guarantee
> > > > > only one device configuration is 'in-flight' at a time), so we end up
> > > > > duplicating a lot of tracking/functionality.
> > > > 
> > > > Hmm.  You should still be able to handle that with just 2 hooks and 1
> > > > bit of state right?  Say "start_configuration" and "end_configuration"
> > > > or similar.  start_configuration gets the blob from the backend, then
> > > > end_configuration is called once RTAS has finished streaming it out to
> > > > the guest and updates to the cnofigured state.
> > > 
> > > {start,end}_configuration callbacks would work for handling
> > > the normal 'configured' state transitions induced by the call, but we'd 
> > > also
> > > need hooks in the "other direction" for a couple cases:
> > > 
> > > This scenario for instance:
> > > 
> > >       qemu:                         guest:
> > >       add device0 to drc0
> > >                                     drmgr0: rtas-configure-connector drc0
> > >       drc0->start_configuration...
> > >                                     drmgr0: rtas-configure-connector drc0
> > >                                     drmgr0: rtas-configure-connector drc0
> > >                                     ...
> > >                                     drmgr0: rtas-configure-connector drc0
> > >       del device0                   <device0 removal pending>
> > >                                     drmgr0: rtas-configure-connector drc0
> > >       system_reset
> > >       device0 removed by drc0
> > >         reset hook
> > >       add device1 to drc0           drmgr0: rtas-configure-connector drc0
> > >                                     <begins fetching stale FDT>
> > > 
> > > So I think we'd need at least a reset hook wired up to the RTAS state.
> > > 
> > > It's also possible for the guest to force a transition out of the
> > > configured state by ISOLATE'ing the device. This can happen in the
> > > middle of the guests configure-connector calls if there's an error. If
> > > rtas-configure-connector is the one generating the error, it can 
> > > anticipate
> > > this and automatically reset the state, but in some cases the error is
> > > guest internal: the get_node() in src/drmgr/rtas_calls.c can fail for
> > > memory allocation errors, or unexpected workarea structure, after which
> > > point it simply stops calling rtas-configure-connector and ISOLATEs the
> > > device. If we don't track that, a subsequent unplug/plug could also
> > > result in stale FDT fragments being sent to the guest.
> > > 
> > > So we'd the RTAS state tracking code to provide an interface of some
> > > sort for the DRC code to call into, which results in a lot of
> > > duplicated state-tracking.
> > > 
> > > It's a bit unintuitive, but those FDT bits are closely coupled to the DRC
> > > state, so in the end I think that ends up being the most straight-forward
> > > place to manage them.
> > 
> > Hrm.  I'm still not following.  I mean, I see that the back-end needs
> > to be aware of the three-valued state: CC-not-called /
> > CC-calls-in-progress / CC-calls-completed, but I'm not seeing why it
> > needs to be aware of the individual CC calls.
> 
> It's not so much that the backend needs to be aware of the individual
> CC calls, but that the front-end/RTAS needs to be aware of the transitions
> made by the back-end (as with the example scenarios above), which can
> affect the output of those individual CC calls at any point in time.
> 
> But maybe I'm misunderstanding your original suggestion. Is this more to
> do with the awkward drc->connector_connector interface, rather than the
> fact that sPAPRDRConnector manages the FDT/offsets. Because if so...

So, the conceptual operation that needs to take place here is that the
device model needs to convey a device tree fragment to the guest.  If
we were designing an interface to do that today, the obvious method is
to just send a blob of fdt.  One operation, nice and simple.  But PAPR
doesn't do that, because it predates fdt, instead it has this rather
complex and awkward sequence of calls to iterate over the dt fragment.

My basic point here is that we have to live with the PAPR interface -
but we should try to restrict the ugliness to the smallest portion of
the code we can and use a nicer interface internally wherever we can.

So I'd like to restrict knowledge of the CC interface to just the
actual RTAS implementation.  The idea it gets a device tree fragment
from the device model as fdt, then handles the streaming out via
multiple RTAS calls.

> > 
> > Could we have instead a set_state() callback of some sort, that
> > handles both some of the existing architected states like ISOLATED,
> > and also special states like CC-in-progress?
> 
> I suppose we could leave the fdt/fdt_offset/configured values in the
> possession of sPAPRDRConnector, but provide a drc->get_cc_state(),
> which exposes the following to rtas_ibm_configure_connector():
> 
> typedef struct sPAPRDRCCState {
>     /* public */
>     void *fdt;
>     int fdt_offset;
>     int fdt_depth;
>     
>     /* private (or just move these up into sPAPRDRConnector) */
>     int fdt_start_offset;
>     bool configured;
> } sPAPRDRCCState;
> 
> If we do that, we can drop the FDT traversal stuff from sPAPRDRConnector,
> but still allow it to reset fdt/fdt_offset/configured/etc when it needs
> to based on it's internal state transitions.
> 
> And along with that provide a drc->set_configured(void) that
> rtas_ibm_configure_connector() can call when it successfully completes
> the traversal.
> 
> Is that the sort of thing you had in mind?

Um.. I'm not entirely following what you're suggesting.  What I'm
intending is:

RTAS code <-> DTC / device model code

 * Interface is just "give me the fdt" and possibly "set_state"
 * What the RTAS code gets from there is just a void * fdt, no offsets

guest <-> RTAS code

 * This portion fdt_offset and fdt_depth.  It's per-DRC, obviously, so
   it might still live in a DRC linked structure, but it's only ever
   touched by the RTAS code, nothing device backend specific.

 * Clearly, this section of code will need a reset hook so it can
   update its values.

I don't know if that's made anything clearer.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

pgp4rTWqt2e32.pgp
Description: PGP signature