[PATCH v3 42/46] usb: gadget: move ep_matches() from epautoconf to udc-core

[Alan Stern]

On Sat, 25 Jul 2015, Robert Jarzmik wrote:

> Petr Cvek <petr.cvek at tul.cz> writes:
> 
> > On 23.7.2015 21:46, Alan Stern wrote:
> >>> It seems that it allows using a BULK endpoint for requested INT
> >>> endpoint. For my PXA27x machine the original code returns BULK EP
> >>> even with valid INT endpoint definition (because BULK EPs are defined
> >>> earlier than INT EPs).
> >> 
> >> Yes, it does allow a bulk endpoint to be used when an interrupt 
> >> endpoint was requested.  However, it won't return a bulk endpoint if 
> >> all the bulk endpoints are already in use.
> This cannot work for pxa27x.

Do you mean that on pxa27x, a bulk endpoint cannot be used as an
interrupt endpoint?  Why not?  From the device controller's point of
view, there is no difference between bulk and interrupt (except
possibly for the maxpacket sizes and high-bandwidth usage when running
at high speed).

> The pxa27x IP has a hardware limitation which prevents an endpoint from changing
> its type once the UDC is enabled (see the comment at the beginning of
> pxa27x_udc.c).
> 
> If that patchset implies that for a requested INT endpoint a BULK endpoint can
> be returned, that won't work. Felipe and Robert, is that what this patchset
> implies ?

Sort of.  The matching code has always behaved that way and this
patchset does not change the behavior.

> > A default PXA27x configuration returns BULK for requested INT. Which is
> > unfortunate, because PXA27x supports INT endpoints and has one predefined, but
> > this function find BULK first (one BULK is allocated and INT is never used).
> See above.

See response above.

Besides, let's say the pxa27x has one bulk and one interrupt endpoint.  
Now suppose the gadget driver requests a bulk endpoint first.  The 
matching code will allocate the single bulk endpoint.  Then the gadget 
driver requests an interrupt endpoint.  The matching code cannot 
allocate the bulk endpoint, because that endpoint is already allocated.  
So it will allocate the interrupt endpoint.

Thus, as you can see, under the right conditions everything will work 
as desired.

> > Because if they do, the ep_matches() function works poorly. It returns a BULK
> > for device (gadget) side, but host side (PC) thinks that this endpoint is an INT
> > and handles it in this way. But the PXA27x thinks the endpoint is a BULK and
> > handles it in its way (according to datasheet, settings for a BULK and an INT
> > transfers are not 100% compatible).

How do they differ?

> > I cannot test "INT as BULK" behavior for the gadget functions, because all
> > gadgets which works on PXA27x does not use INT endpoints (some allocate the
> > endpoint but never use it).
> Ah a bit of history here.
> 
> At least gadget zero does, and it's my main testing point for pxa27x_udc.
> Then there should be g_serial (no acm nor obex), but that's something I have not
> tried since 2009 ...
> 
> For history also, there was already an attempt a long time ago for epautoconf
> revamping, done by Rodolfo Giometti IIRC.
> 
> Anyway, I need pxa27x_udc to remain functional, so I'd like to understand if
> something will stop working, Robert B.

Perhaps you could submit a patch that adds a "do not allocate a bulk 
endpoint when an interrupt endpoint is requested" quirk flag to the 
usb_gadget structure, and modify the matching code to take the new flag 
into account.

Alan Stern