RE: [RFC PATCH v1] usb: dwc2: reduce dwc2 driver probe time
From: Kaukab, Yousaf
Date: Thu Feb 12 2015 - 08:22:09 EST
> -----Original Message-----
> From: John Youn [mailto:John.Youn@xxxxxxxxxxxx]
> Sent: Thursday, February 12, 2015 4:33 AM
> To: Roy; John.Youn@xxxxxxxxxxxx; Felipe Balbi
> Cc: Yunzhi Li; jwerner@xxxxxxxxxxxx; Herrero, Gregory; Kaukab, Yousaf;
> r.baldyga@xxxxxxxxxxx; Dinh Nguyen; Eddie Cai; Lin Huang; wulf; æå; Tao
> Huang; walkrain@xxxxxxx; Douglas Anderson; Greg Kroah-Hartman; linux-
> usb@xxxxxxxxxxxxxxx; LKML
> Subject: Re: [RFC PATCH v1] usb: dwc2: reduce dwc2 driver probe time
> On 2/11/2015 3:42 AM, Roy wrote:
> > Hi John Youn:
> > Could you please give some suggestions from your point of view,
> > about this probe time issue ?
> > Thanks a lot.
> > at 2015/2/11 2:23, Julius Werner wrote:
> >>> @@ -2703,7 +2703,7 @@ int dwc2_get_hwparams(struct dwc2_hsotg
> >>> gusbcfg = readl(hsotg->regs + GUSBCFG);
> >>> gusbcfg &= ~GUSBCFG_FORCEHOSTMODE;
> >>> writel(gusbcfg, hsotg->regs + GUSBCFG);
> >>> - usleep_range(100000, 150000);
> >>> + usleep_range(25000, 50000);
> >> The point of usleep_range() is to coalesce multiple timer interrupts
> >> in idle systems for power efficiency. It's pretty pointless/harmful
> >> during probe anyway and there's almost never a reason to make the
> >> span larger than a few milliseconds. You should reduce this to
> >> something reasonable (e.g. usleep_range(25000, 26000) or even
> >> usleep_range(25000, 25000)) to save another chunk of time. Same
> >> applies to other delays above.
> Databook does say 25ms. From what I could gather this has to do with the
> debounce filter time on the IDDIG pin after the ForceHstMode/ForceDevMode
> is programmed. There is no way to poll this. I think the change is acceptable,
> even to lower the range as Julius suggested.
> >>> do you know what's the upper boundary for AHB clock ? How fast can
> >>> it be? It's not wise to change timers because "it works on my RK3288
> >>> board", you need to guarantee that this won't break anybody else.
> >> But this code is already a loop that spins on the AHBIdle bit, right?
> >> It should work correctly regardless of the delay. The only question
> >> is whether the code could be more efficient with a longer sleep...
> >> but since the general recommendation is to delay for ranges less than
> >> 10us, and the AHB clock would need to be lower than 100KHz (the ones
> >> I see are usually in the range of tens or hundreds of MHz) to take
> >> longer than that, this seems reasonable to me.
> Agree with this. It shouldn't take nearly that long and you are polling anyways.
> As for the other change:
> > It seems that usleep_range() at boot time will pick the longest value
> > in the range. In dwc2_core_reset() there is a very long delay takes
> > 200ms, and this function run twice when probe, could any one tell me
> > is this delay time resonable ?
> I'm not sure about this value or the reasoning/history behind it. It is not in our
> internal code. It looks like it is taking into account the delay for the
> ForceHstMode/ForceDevMode programming. However, I think your change is
> conservative and should be ok. Maybe Samsung engineers know about this?
If the delay is due to ForceHstMode/ForceDevMode then it should be reduce to 25ms range. As done in dwc2_get_hwparams for example.