Re: [PATCH v2 11/17] xen/arm: PCI host bridge discovery within XEN on ARM

[Julien Grall <julien.grall.oss@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 4059 bytes --]

On Fri, 24 Sep 2021, 23:42 Stefano Stabellini, <sstabellini@kernel.org>
wrote:

> On Fri, 24 Sep 2021, Rahul Singh wrote:
> > Hi Stefano,
> >
> > > On 23 Sep 2021, at 8:12 pm, Stefano Stabellini <sstabellini@kernel.org>
> wrote:
> > >
> > > On Thu, 23 Sep 2021, Rahul Singh wrote:
> > >>>> +            goto err_exit;
> > >>>> +    }
> > >>>
> > >>> This is unnecessary at the moment, right? Can we get rid of
> ops->init ?
> > >>
> > >> No this is required for N1SDP board. Please check below patch.
> > >>
> https://gitlab.com/rahsingh/xen-integration/-/commit/6379ba5764df33d57547087cff4ffc078dc515d5
> > >
> > > OK
> > >
> > >
> > >>>> +int pci_host_common_probe(struct dt_device_node *dev, const void
> *data)
> > >>>> +{
> > >>>> +    struct pci_host_bridge *bridge;
> > >>>> +    struct pci_config_window *cfg;
> > >>>> +    struct pci_ecam_ops *ops;
> > >>>> +    const struct dt_device_match *of_id;
> > >>>> +    int err;
> > >>>> +
> > >>>> +    if ( dt_device_for_passthrough(dev) )
> > >>>> +        return 0;
> > >>>> +
> > >>>> +    of_id = dt_match_node(dev->dev.of_match_table,
> dev->dev.of_node);
> > >>>> +    ops = (struct pci_ecam_ops *) of_id->data;
> > >>>
> > >>> Do we really need dt_match_node and dev->dev.of_match_table to get
> > >>> dt_device_match.data?
> > >>>
> > >>
> > >>> data is passed as a parameter to pci_host_common_probe, isn't it
> enough
> > >>> to do:
> > >>>
> > >>> ops = (struct pci_ecam_ops *) data;
> > >>
> > >> As of now not required but in future we might need it if we implement
> other ecam supported bridge
> > >>
> > >> static const struct dt_device_match gen_pci_dt_match[] = {
>
> > >>    { .compatible = "pci-host-ecam-generic",
>
> > >>      .data =       &pci_generic_ecam_ops },
> > >>
> > >>    { .compatible = "pci-host-cam-generic",
> > >>      .data = &gen_pci_cfg_cam_bus_ops },
>
> > >>
> > >>    { },
>
> > >> };
> > >
> > > Even if we add another ECAM-supported bridge, the following:
> > >
> > > ops = (struct pci_ecam_ops *) data;
> > >
> > > could still work, right? The probe function will directly receive as
> > > parameter the .data pointer. You shouldn't need the indirection via
> > > dt_match_node?
> >
> > As per my understanding probe function will not get .data pointer.Probe
> data argument is NULL in most of the cases in XEN
> > Please have a look once dt_pci_init() -> device_init(..) call flow
> implementation.
>
> You are right. Looking at the code, nobody is currently using
> dt_device_match.data and it is clear why: it is not passed to the
> device_desc.init function at all. As it is today, it is basically
> useless.
>

IIRC it is used by the SMMU driver. But you need to lookup for the desc
manually in each init callback.

If I am not mistaken, this is how Linux is dealing with it as well.
However...


> And there is only one case where device_init has a non-NULL data
> parameter and it is in xen/drivers/char/arm-uart.c. All the others are
> not even using the data parameter of device_init.


> I think we need to change device_init so that dt_device_match.data can
> be useful. Sorry for the scope-creep but I think we should do the
> following:
>
> - do not add of_match_table to struct device
>
> - add one more parameter to device_desc.init:
>   int (*init)(struct dt_device_node *dev, struct device_desc *desc, const
> void *data);
>
> - change device_init to call desc->init with the right parameters:
>   desc->init(dev, desc, data);
>
> This way pci_host_common_probe is just going to get a desc directly as
> parameter. I think it would make a lot more sense from an interface
> perspective. It does require a change in all the DT_DEVICE_START.init
> functions adding a struct device_desc *desc parameter, but it should be
> a mechanical change.
>
> Alternatively we could just change device_init to pass
> device_desc.dt_match.data when the data parameter is NULL but it feels
> like a hack.
>
>
> What do you think?


... I like the idea of passing desc parameter (we could also simply pass
desc.data in an argument named "priv").

Cheers,

>

[-- Attachment #2: Type: text/html, Size: 6519 bytes --]