On 2018-06-07 14:01, Heikki Krogerus wrote:
On Thu, Jun 07, 2018 at 09:22:56AM +0200, Hans de Goede wrote:
Hi,
On 06-06-18 23:36, Mats Karrman wrote:
Hello Gentlemen,
I'm trying to get my head around USB role switches in connection with Type-C ports
and device-trees. So far I have not found much documentation, e.g. there are no
device-tree bindings documented and really no good examples in existing device
trees, although there has been some attempts, e.g. [1] and [2]. Anyway, so I send
you a couple of questions instead:
1) tcpm uses the port device struct to find a single usb_role_switch but there is
room for three USB busses in the Type-C connector; one high speed and two (?) super-
speed. These would not all come from the same controller (there might even be
separate controllers for host and device mode for each bus).
I believe USB 3.2 spec in practice says that the two superspeed
"lanes" must to go to the same controller. Only one will be used for
link training etc. The second one is pulled in after certain state of
enumeration has been passed.
So we may theoretically have two controllers to deal with, one for
USB2 and another for USB3, but not three.
But in any case, let's not try to fix theoretical problems that may
never exist.
AFAIK the 2nd superspeed USB bus is never used as such. There really is only 1
USB bus on the Type-C connector, the combined USB-2 + the 1st superspeed bus,
physically these are 2 separate busses but that is purely for compatibility
reasons, logically there really only is 1 bus, just like a superspeed Type-A
connector has both busses physically but logically represents a single bus/port.
The case I am working on now only have a single USB2 otg controller so it should
be possible to make that driver register a role switch but for other cases?
I guess theoretically a device could use separate role switches / muxes for
the USB-2 and USB-3 busses, but that would be weird. So lets cross that bridge
when we reach it.
I imagine it would be possible to create a composite driver as a proxy for all role
switches but that would probably be different for every platform/product - not
very elegant. Could the role switch infrastructure be extended to handle arbitrary
sets of coordinated switches?
As said lets cross that bridge when we reach it.
Agreed.
OK, I'm buying what you're saying. After reading some manuals for USB3 controllers
and USB3 equipped SoC's I realize that most USB3 controllers seem to also support
USB2. Makes things easier, although not as flexible as I imagined.
One thing that led me astray was the comment for usb_role_switch_register() and the
separate device struct pointers for usb2/usb3/udc in the usb_role_switch_desc struct.
Still not totally clear to me what they are for?
2) How should the connection between the Type-C port and the switches best be
expressed in a device tree? Using graph I presume, but should it be mixed into the
existing "usb-connector" or should this be a separate block?
I don't know?
I'm not super comfortable proposing things for these bindings because
my knowledge of DT is a bit limited. I mostly work with ACPI based
platforms.
I'm not even completely sure device graph is usable in our case,
though I pretty sure it is.
I think it is unfortunate that the graph use numeric addresses that need to be
fixed by documentation and already I see problems with the current assignment
(0=HS, 1=SS, 2=SBU), e.g. if the host and device mode are handled by different
controllers. Graph do support multiple endpoints for one port but then we have
another level of magic numbers which does not exactly make things easier
(e.g. 0=dual or host controller, 1=device controller, 2=mode switch).
The graph stuff is more Heikki's specialty so I will let Heikki answer this.
I'm not really a graph expert. It is just the only tool in kernel I
can see that allows us to link together all these different components
in HW description, and which has support for both DT and ACPI in Linux
kernel. That is the only reason I've talked about it.
One of the motivations for the device connection API was that it hides
the actual method these components are linked together in HW
description. What ever the method is, device graph or something else
(we can support different ways), the drivers don't need to know about
it.
Ahh, this is complicated and the lack of examples makes it a bit hard to
digest. Is it your expectation that the OF graph parsing code will be calling
device_connection_add() for every graph connection found in the device
tree or maybe for every connection with an "id" property?
(Ugh, just realized that the "reg" numeric value of the endpoint node is
put in the "id" field of the endpoint struct, bad luck...).
// Mats
But in any case, I don't understand why should we depend on the
addresses (index in practice) in this case? We can define an extra
identifier property. Then in code we can walk through the whole graph,
ignoring the address, and checking every remote endpoint in it and use
that property to find what we are looking for, no? I'm pretty sure
somebody already proposed "type" property for that "usb-connector"
and I understood it was meant for identifying the endpoint (maybe I'm
wrong), but I guess it newer went in?
Thanks,
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
