On Sat, Aug 23, 2025, Sven Peter wrote: > On 22.08.25 00:28, Thinh Nguyen wrote: > > On Thu, Aug 21, 2025, Sven Peter wrote: > > > From: Janne Grunau <j@xxxxxxxxxx> > > [...] > > > > /* > > > * Some platforms need to power off all Root hub ports immediately after DWC3 set to host > > > * mode to avoid VBUS glitch happen when xhci get reset later. > > > + * On Apple platforms we must not touch any MMIO yet because dwc3 > > > + * will not work correctly before its PHY has been initialized. > > > > This doesn't make sense, the phy should've been initialized by this > > point. We already access MMIO before this. Even your [PATCH RFC 7/22] > > attempts to access MMIO before this, yet accessing here causes problem. > > Do we know when the phy get initialized then? > > You're absolutely right, this doesn't make any sense! > > dwc3_power_off_all_roothub_ports uses ioremap which will map the XHCI region > as Device-nGnRE. That just doesn't work on these machines and actually > explains the SErrors. > Resources on these machines generally have the IORESOURCE_MEM_NONPOSTED flag > set (via the nonposted-mmio dt property) and then use ioremap_np to map them > using Device-nGnRnE. > Ok. So the failure does not seem to be because the phy hasn't been initialized yet, but it's because of io mapping. Did you try to check against IORESOURCE_MEM_NONPOSTED to use ioremap_np instead? That keeps the logic consistent between different platforms. BR, Thinh
