On Fri, Nov 08, 2019 at 11:40:53AM -0800, Rob Clark wrote:
> On Fri, Nov 8, 2019 at 10:42 AM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> >
> > Quoting Rob Clark (2019-11-08 08:54:23)
> > > On Thu, Nov 7, 2019 at 10:35 PM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > > >
> > > > Quoting Rob Clark (2019-11-07 18:06:19)
> > > > > On Thu, Nov 7, 2019 at 1:06 PM Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > > > > >
> > > > > >
> > > > > > NULL is a valid clk pointer returned by clk_get(). What is the display
> > > > > > driver doing that makes it consider NULL an error?
> > > > > >
> > > > >
> > > > > do we not have an iface clk? I think the driver assumes we should
> > > > > have one, rather than it being an optional thing.. we could ofc change
> > > > > that
> > > >
> > > > I think some sort of AHB clk is always enabled so the plan is to just
> > > > hand back NULL to the caller when they call clk_get() on it and nobody
> > > > should be the wiser when calling clk APIs with a NULL iface clk. The
> > > > common clk APIs typically just return 0 and move along. Of course, we'll
> > > > also turn the clk on in the clk driver so that hardware can function
> > > > properly, but we don't need to expose it as a clk object and all that
> > > > stuff if we're literally just slamming a bit somewhere and never looking
> > > > back.
> > > >
> > > > But it sounds like we can't return NULL for this clk for some reason? I
> > > > haven't tried to track it down yet but I think Matthias has found it
> > > > causes some sort of problem in the display driver.
> > > >
> > >
> > > ok, I guess we can change the dpu code to allow NULL.. but what would
> > > the return be, for example on a different SoC where we do have an
> > > iface clk, but the clk driver isn't enabled? Would that also return
> > > NULL? I guess it would be nice to differentiate between those cases..
> > >
> >
> > So the scenario is DT describes the clk
> >
> > dpu_node {
> > clocks = <&cc AHB_CLK>;
> > clock-names = "iface";
> > }
> >
> > but the &cc node has a driver that doesn't probe?
> >
> > I believe in this scenario we return -EPROBE_DEFER because we assume we
> > should wait for the clk driver to probe and provide the iface clk. See
> > of_clk_get_hw_from_clkspec() and how it looks through a list of clk
> > providers and tries to match the &cc phandle to some provider.
> >
> > Once the driver probes, the match will happen and we'll be able to look
> > up the clk in the provider with __of_clk_get_hw_from_provider(). If
> > the clk provider decides that there isn't a clk object, it will return
> > NULL and then eventually clk_hw_create_clk() will turn the NULL return
> > value into a NULL pointer to return from clk_get().
> >
>
> ok, that was the scenario I was worried about (since unclk'd register
> access tends to be insta-reboot and hard to debug).. so I think it
> should be ok to make dpu just ignore NULL clks.
>
> From a quick look, I think something like the attached (untested).
The driver appears to be happy with it, at least at probe() time.
