On Thu, Jun 08, 2023 at 01:21:02AM +0530, Krishna Kurapati PSSNV wrote: > On 6/7/2023 5:07 PM, Johan Hovold wrote: > > So there at least two issues with this series: > > > > 1. accessing xhci registers from the dwc3 core > > 2. accessing driver data of a child device > > > > 1. The first part about accessing xhci registers goes against the clear > > separation between glue, core and xhci that Felipe tried to maintain. > > > > I'm not entirely against doing this from the core driver before > > registering the xhci platform device as the registers are unmapped > > afterwards. But if this is to be allowed, then the implementation should > > be shared with xhci rather than copied verbatim. > > > > The alternative that avoids this issue entirely could indeed be to > > simply count the number of PHYs described in DT as Rob initially > > suggested. Why would that not work? > > > The reason why I didn't want to read the Phy's from DT is explained in > [1]. I felt it makes the code unreadable and its very tricky to read the > phy's properly, so we decided we would initialize phy's for all ports > and if a phy is missing in DT, the corresponding member in > dwc->usbX_generic_phy[] would be NULL and any phy op on it would be a NOP. That doesn't sound too convincing. Can't you just iterate over the PHYs described in DT and determine the maximum port number used for HS and SS? > Also as per Krzysztof suggestion on [2], we can add a compatible to read > number of phy's / ports present. This avoids accessing xhci members > atleast in driver core. But the layering violations would still be present. Yes, but if the information is already available in DT it's better to use it rather than re-encode it in the driver. > > 2. The driver is already accessing driver data of the child device so > > arguably your series is not really making things much worse than they > > already are. > > > > I've just sent a couple of fixes to address some of the symptoms of > > this layering violation here: > > > > https://lore.kernel.org/lkml/20230607100540.31045-1-johan+linaro@xxxxxxxxxx/ > > > > As you pointed out offline to me that using xhci event notifiers I > mentioned on [3], if it is not acceptable to use them as notifications > to check whether we are in host mode, I believe the only way would be to > use the patches you pushed in. I still think we'll end up using callbacks from the xhci/core into the glue driver, but dedicated ones rather than using usb_register_notify(). The fixes I posted can work as a stop-gap solution until then. > > Getting this fixed properly is going to take a bit of work, and > > depending on how your multiport wake up implementation is going to look > > like, adding support for multiport controllers may not make this much > > harder to address. > > > > So perhaps we can indeed merge support for multiport and then get back > > to cleaning this up. > So, you are referring to use the patches you pushed today as a partial > way to cleanup layering violations and merge multiport on top of it ? If > so, I am fine with it. I can rebase my changes on top of them. Right. A bit depending on how your wakeup implementation looks like, it seems we can merge multiport support and then address the layering issues which are already present in the driver. > [1]: > https://lore.kernel.org/all/4eb26a54-148b-942f-01c6-64e66541de8b@xxxxxxxxxxx/ > [2]: > https://lore.kernel.org/all/ca729f62-672e-d3de-4069-e2205c97e7d8@xxxxxxxxxx/ > [3]: > https://lore.kernel.org/all/37fd026e-ecb1-3584-19f3-f8c1e5a9d20a@xxxxxxxxxxx/ Johan