On Tue, Apr 28, 2020 at 3:27 PM Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote: > > Better late than never... > > On Tue, 28 Apr 2020, Andrey Konovalov wrote: > > > On Tue, Apr 28, 2020 at 2:50 AM Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote: > > > > > > On Mon, Apr 27, 2020 at 10:47 PM Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote: > > > > > > > > On Mon, Apr 27, 2020 at 9:51 PM Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote: > > > > > > Another question, which actually seems to be a major problem. > > > > > > > > > > It looks like automatic endpoint selection based on required features > > > > > doesn't work in raw-gadget. The way it tries to do that is iterating > > > > > over the list of available endpoints and finding one that has the > > > > > right direction and transfer type. But it seems that the right way to > > > > > do that (like it's done in usb_ep_autoconfig()) is to also patch the > > > > > bEndpointAddress field of the endpoint descriptor. > > > > > > > > > > Currently with raw-gadget the endpoints are supposed to be enabled > > > > > after a set_configuration/set_interface request from the host, so it's > > > > > too late to patch the endpoint descriptor (the one that was sent to > > > > > the host during enumeration). I'm guessing that sending one endpoint > > > > > descriptor during enumeration and then using a different one (with > > > > > patched bEndpointAddress) to set ep->desc before doing usb_ep_enable() > > > > > won't work (as there's going to be mismatch in endpoint addresses > > > > > between the host and the UDC), right? > > It certainly won't work. > > > > > > I wonder what we can do about that. It seems that we actually need to > > > > > parse the descriptors and figure out the right endpoints before the > > > > > enumeration starts. > > That's right. In the case of raw-gadget, this would be done around the > time the user tells the driver what the config descriptors are. > > > > > Or maybe somehow expose all endpoints and their features (like > > > > GadgetFS) and let userspace put proper addresses into endpoint > > > > descriptors before enumeration starts. > > And if the user messes up the assignment, his driver won't work. > > > > I think this is the way to go, giving the userspace maximum control > > > over what's going on. Seems not too hard to implement too. The only > > > part that is not clear to me is how to figure out the endpoint address > > > that a particular UDC endpoint expects to have. > > > > > > Most of the time endpoints are named as something like "ep1in" or > > > "ep2out", and AFAIU the number in the endpoint name is the address. > > > However net2280 seems to have weird endpoint names "ep-a", "ep-b", > > > etc. Looking at usb_ep_autoconfig_ss(), those endpoints supposedly > > > have sequential addresses starting from 1 (for each in/out type). But > > > then in the GadgetFS example [1] e.g. "ep-a" corresponds to address 7, > > > which I don't understand. My guess is that this is done to handle both > > > dummy_udc and net2280 in a single if case, but I'm not sure. > > > > > > [1] http://www.linux-usb.org/gadget/usb.c > > > > Oh, there's actually a comment that explains this [1] in dummy_hcd.c > > (exactly the place where one would look for it :). So "ep-a" and > > others are fully configurable and accept any endpoint address. OK, I > > think I've figured out what to do here, will send a patch. > > > > [1] https://elixir.bootlin.com/linux/v5.7-rc3/source/drivers/usb/gadget/udc/dummy_hcd.c#L113 Hi Alan, I've been looking at this a little more. Do I understand correctly that even though Dummy UDC names endpoints as "ep1in", etc. it actually allows to assign endpoints addresses different from what is specified in the endpoint names (it uses find_endpoint() to find the right endpoint based on ep->desc)? E.g. you can technically assign endpoint with address 2 (| USB_DIR_IN) to "ep1in". If this is correct, this kind of limits Dummy UDC usage with Raw Gadget the way it is currently implemented, as Raw Gadget assumes that the endpoint address must be fixed when the endpoint is named as ep1in. Would it be acceptable to add another mode to Dummy UDC that names the endpoints as "ep-a"? Perhaps enabled with a module parameter. I'm not sure if this kind of naming would be technically correct, as "ep-a" name assumes that we can assign arbitrary transfer type to the endpoint as well, which isn't possible with Dummy UDC, as it doesn't support ISO transfers. Or do you think there can be another way to expose the fact that Dummy UDC allows arbitrary addresses? I could hardcode this into Raw Gadget, but it doesn't feel like the right approach. Thanks!
