Re: [PATCH v2 2/3] usb: chipidea: Hook into mux framework to toggle usb switch

From: Stephen Boyd
Date: Mon Aug 07 2017 - 21:51:43 EST


Quoting Peter Rosin (2017-07-31 03:33:22)
> On 2017-07-14 23:40, Stephen Boyd wrote:
> > @@ -1964,16 +1965,26 @@ void ci_hdrc_gadget_destroy(struct ci_hdrc *ci)
> >
> > static int udc_id_switch_for_device(struct ci_hdrc *ci)
> > {
> > + int ret = 0;
> > +
> > if (ci->is_otg)
> > /* Clear and enable BSV irq */
> > hw_write_otgsc(ci, OTGSC_BSVIS | OTGSC_BSVIE,
> > OTGSC_BSVIS | OTGSC_BSVIE);
> >
> > - return 0;
> > + if (!ci_otg_is_fsm_mode(ci))
> > + ret = mux_control_select(ci->platdata->usb_switch, 0);
> > +
> > + if (ci->is_otg && ret)
> > + hw_write_otgsc(ci, OTGSC_BSVIE | OTGSC_BSVIS, OTGSC_BSVIS);
> > +
> > + return ret;
> > }
> >
> > static void udc_id_switch_for_host(struct ci_hdrc *ci)
> > {
> > + mux_control_deselect(ci->platdata->usb_switch);
> > +
>
> This looks broken. You conditionally lock the mux and you unconditionally
> unlock it. Quoting the mux_control_deselect doc:
>
> * It is required that a single call is made to mux_control_deselect() for
> * each and every successful call made to either of mux_control_select() or
> * mux_control_try_select().
>
> Think of the mux as a semaphore with a max count of one. If you lock it,
> you have to unlock it when you're done. If you didn't lock it, you have
> zero business unlocking it. If you try to lock it but fail, you also have
> no business unlocking it. Just like a semaphore.

Good catch. I've added a if (!ci_otg_is_fsm_mode()) check here.

>
> Another point: I do not know if udc_id_switch_for_host is *only* called
> when there has been a prior call to udc_id_switch_for_device that
> succeeded or how this works exactly. But this all looks fragile. Using
> mux_control_select and mux_control_deselect *must* be done in pairs.
> If you want a mux to be locked for "a while", such as in this case, you
> have to make sure you stay within the rules. There is no room for half
> measures, or you will likely cause a deadlock when something unexpected
> happens.
>

Can you elaborate? Is it bad that we keep it "locked" while we're in
host or device mode? It looked like we paired the start/stop ops with
each other so that the mux is properly managed across these ops. My
testing hasn't shown a problem, but maybe there's some corner case
you're thinking of? I'll double check the code.