Am Donnerstag, den 04.10.2018, 15:25 +0200 schrieb Philipp Zabel:
> On Thu, 2018-10-04 at 13:20 +0000, Leonard Crestez wrote:
> > On Thu, 2018-10-04 at 10:59 +0200, Lucas Stach wrote:
> > > Am Montag, den 01.10.2018, 22:53 +0300 schrieb Leonard Crestez:
> > > > When the root complex suspends it must send a PME_Turn_Off TLP.
> > > > Implement this by asserting the "turnoff" reset.
> > > >
> > > > +static void imx6_pcie_pm_turnoff(struct imx6_pcie *imx6_pcie)
> > > > +{
> > > > > > > > + reset_control_assert(imx6_pcie->turnoff_reset);
> > > > + reset_control_deassert(imx6_pcie->turnoff_reset);
> > >
> > > I'm a bit surprised to see no timing requirements here. I would have
> > > expected that there is a minimum time from asserting the reset, so the
> > > turnoff message gets transmitted to the EP before the clocks are
> > > stopped.
> >
> > According to the PCI standard after PME_Turn_Off is sent all components
> > must respond with PME_TO_Ack. There doesn't seem to be any bit exposed
> > to check if acks were received but the standard recommends a 1-10ms
> > timeout after which you should proceed anyway.
> >
> > It seems the NXP vendor tree has an udelay(1000) which I missed, it
> > seems like an acceptable solution. Or maybe it should be msleep(10)?
>
> Maybe usleep_range(1000, 10000) ?
Yep, we aren't in atomic context here and the timeout is relatively
short so usleep_range() looks like the correct thing to use.
Regards,
Lucas
