On Wed, Jan 08, 2020 at 02:19:29PM +0100, Wolfram Sang wrote:
>
> > > > As I said to Laurent, too, I think the risk that a bus is not fully
> > > > described is higher than a device which does not respond to a read_byte.
> > > > In both cases, we would wrongly use an address in use.
> >
> > I don't fully agree with this, I think we shouldn't impose a penalty on
> > every user because some device trees don't fully describe the hardware.
>
> I haven't decided yet. However, my general preference is that for a
> generic OS like Linux, saftey comes first, then performance. If you have
> a fully described DT, then the overhead will be 1 read_byte transaction
> per requested alias at probe time. We could talk about using quick_read
> to half the overhead. You could even patch it away, if it is too much
> for $customer.
>
> > I think we should, at the very least, skip the probe and rely on DT if
> > DT explicitly states that all used addresses are listed. We discussed a
> > property to report addresses used by devices not described in DT, if
> > that property is listed I would prefer trusting DT.
>
> Yeah, we discussed this property and I have no intentions of dropping
> it. I haven't though of including it into this series, but it probably
> makes sense. We don't have to define much anyhow, just state what
> already exists, I guess.
>
> From Documentation/devicetree/bindings/i2c/i2c-ocores.txt:
>
> dummy@60 {
> compatible = "dummy";
> reg = <0x60>;
> };
>
> I think "dummy" is generic enough to be described in i2c.txt.
We may want a compatible value that guarantees noone will ever match it
:-) I was imagining a single property at the bus level with multiple
ranges instead, but dummy nodes could be OK too.
--
Regards,
Laurent Pinchart
