Re: [PATCH v2] usb: gadget: epautoconf: claim smallest endpoints first

[Ruslan Bilovol]

  for On Sat, Jul 4, 2020 at 6:42 AM Peter Chen <peter.chen@xxxxxxx> wrote:
>
> On 20-07-03 13:46:27, Ruslan Bilovol wrote:
> > On Tue, Jun 30, 2020 at 4:58 AM Peter Chen <peter.chen@xxxxxxx> wrote:
> > >
> > > On 20-06-29 23:18:45, Ruslan Bilovol wrote:
> > > > UDC hardware may have endpoints with different maxpacket
> > > > size. Current endpoint matching code takes first matching
> > > > endpoint from the list.
> > > >
> > > > It's always possible that gadget allocates endpoints for
> > > > small transfers (maxpacket size) first, then larger ones.
> > > > That works fine if all matching UDC endpoints have same
> > > > maxpacket size or are big enough to serve that allocation.
> > > >
> > > > However, some UDCs have first endpoints in the list with
> > > > bigger maxpacket size, whereas last endpoints are much
> > > > smaller. In this case endpoint allocation will fail for
> > > > the gadget (which allocates smaller endpoints first) on
> > > > final endpoint allocations.
> > > >
> > > > To make endpoint allocation fair, pick up smallest
> > > > matching endpoints first, leaving bigger ones for
> > > > heavier applications.
> > > >
> > > > Signed-off-by: Ruslan Bilovol <ruslan.bilovol@xxxxxxxxx>
> > > > ---
> > > >
> > > > v2: rebased onto latest balbi/next branch
> > > >
> > > >  drivers/usb/gadget/epautoconf.c | 23 ++++++++++++++++++-----
> > > >  1 file changed, 18 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/drivers/usb/gadget/epautoconf.c b/drivers/usb/gadget/epautoconf.c
> > > > index 1eb4fa2e623f..6c453b5d87bb 100644
> > > > --- a/drivers/usb/gadget/epautoconf.c
> > > > +++ b/drivers/usb/gadget/epautoconf.c
> > > > @@ -66,7 +66,7 @@ struct usb_ep *usb_ep_autoconfig_ss(
> > > >       struct usb_ss_ep_comp_descriptor *ep_comp
> > > >  )
> > > >  {
> > > > -     struct usb_ep   *ep;
> > > > +     struct usb_ep   *ep, *ep_min = NULL;
> > > >
> > > >       if (gadget->ops->match_ep) {
> > > >               ep = gadget->ops->match_ep(gadget, desc, ep_comp);
> > > > @@ -74,14 +74,27 @@ struct usb_ep *usb_ep_autoconfig_ss(
> > > >                       goto found_ep;
> > > >       }
> > > >
> > > > -     /* Second, look at endpoints until an unclaimed one looks usable */
> > > > +     /*
> > > > +      * Second, look at endpoints until an unclaimed one looks usable.
> > > > +      * Try to find one with smallest maxpacket limit, leaving larger
> > > > +      * endpoints for heavier applications
> > > > +      */
> > > >       list_for_each_entry (ep, &gadget->ep_list, ep_list) {
> > > > -             if (usb_gadget_ep_match_desc(gadget, ep, desc, ep_comp))
> > > > -                     goto found_ep;
> > > > +             if (usb_gadget_ep_match_desc(gadget, ep, desc, ep_comp)) {
> > > > +                     if (desc->wMaxPacketSize == 0)
> > > > +                             goto found_ep;
> > >
> > > Why you do special handling for this? You still could give the smallest
> > > maxpacket_limit EP for it, right?
> >
> > Of course it's technically possible. However in case "wMaxPacketSize == 0"
> > gadget driver wants to get maximum possible wMaxPacketSize from endpoint
> > configuration and I was thinking about avoiding regressions if we always provide
> > smaller endpoints.
>
> You may only want to change the match logic, not but the special case.
>
> Currently, it returns the first matched endpoint no matter
> "wMaxPacketSize == 0" or not. And you changed the match logic
> as returning the smallest maxPacketsize endpoint, you also don't need
> to consider whether "wMaxPacketSize == 0" or not, otherwise, it may
> introduce the complexity.
>

The reason I kept the old behavior when "wMaxPacketSize == 0" is because
it's a special case. In this case a gadget driver wants to use a whole
available MaxPacketSize of claimed endpoint. Since it doesn't know
what MaxPacketSize may be in a particular UDC endpoint, it just
relies on epautoconf core and gets what's available.
So the behavior looks like the opposite of claiming a smallest endpoint, but
rather using as big as possible MaxPacketSize provided by the endpoint.

We even may want to implement a 'claim biggest endpoint' policy for the
"wMaxPacketSize == 0" case.

Looking at gadget drivers sources, it seems there is no such usecase
in any driver (at least with simple 'grep' over sources), so I'm not sure
if we need to worry about it at all, that's why I decided to keep an old
behavior here.

Thanks
Ruslan

> >
> > As I can see, providing smallest endpoint that matches requested wMaxPacketSize
> > is OK, but if gadget driver just wants autoconf core to use it with
> > maximum possible
> > value, I'm thinking now if we can even change this part and if wMaxPacketSize
> > is zero, find endpoint with maximum possible wMaxPacketSize
> >
> > Does it make sense?
> >
> > Thanks
> > Ruslan
> >
> > >
> > > Peter
> > >
> > > > +                     else if (!ep_min)
> > > > +                             ep_min = ep;
> > > > +                     else if (ep->maxpacket_limit < ep_min->maxpacket_limit)
> > > > +                             ep_min = ep;
> > > > +             }
> > > >       }
> > > >
> > > >       /* Fail */
> > > > -     return NULL;
> > > > +     if (!ep_min)
> > > > +             return NULL;
> > > > +
> > > > +     ep = ep_min;
> > > >  found_ep:
> > > >
> > > >       /*
> > > > --
> > > > 2.17.1
> > > >
> > >
> > > --
> > >
> > > Thanks,
> > > Peter Chen
>
> --
>
> Thanks,
> Peter Chen