On Fri, Jan 16, 2015 at 05:09:28PM +0000, Thierry Reding wrote:
> On Fri, Jan 16, 2015 at 03:07:49PM +0000, Mark Rutland wrote:
> > Hi Paul,
> >
> > /me opens a delicious can of Lumbricus terrestris
> >
> > My comments below aren't so much about this patch alone, but more about
> > the entire strategy with drivers/soc and the kind of code living there.
> >
> > On Fri, Jan 16, 2015 at 11:39:13AM +0000, Paul Walmsley wrote:
> > >
> > > Add "nvidia,tegra132" as a compatible string that denotes a Tegra SoC.
> > >
> > > Signed-off-by: Paul Walmsley <paul@xxxxxxxxx>
> > > Cc: Paul Walmsley <pwalmsley@xxxxxxxxxx>
> > > Cc: Stephen Warren <swarren@xxxxxxxxxxxxx>
> > > Cc: Thierry Reding <thierry.reding@xxxxxxxxx>
> > > Cc: Alexandre Courbot <gnurou@xxxxxxxxx>
> > > ---
> > > drivers/soc/tegra/common.c | 1 +
> > > 1 file changed, 1 insertion(+)
> > >
> > > diff --git a/drivers/soc/tegra/common.c b/drivers/soc/tegra/common.c
> > > index a71cb74..a952986 100644
> > > --- a/drivers/soc/tegra/common.c
> > > +++ b/drivers/soc/tegra/common.c
> > > @@ -15,6 +15,7 @@ static const struct of_device_id tegra_machine_match[] = {
> > > { .compatible = "nvidia,tegra30", },
> > > { .compatible = "nvidia,tegra114", },
> > > { .compatible = "nvidia,tegra124", },
> > > + { .compatible = "nvidia,tegra132", },
> > > { }
> > > };
> >
> > I'm rather worried by this.
> >
> > The only user of this table is soc_is_tegra(), which seems to guard
> > several probe paths in drivers which are DT driven. Given that, I don't
> > see why this is necessary at all, because the presence of the
> > appropriate nodes should be sufficient to handle the early return from
> > an initcall.
>
> We need this to preserve backward-compatibility. There are some
> instances where any device tree nodes that could be used to guard the
> code execution were only added after the fact. So in order to keep
> systems with an old DTB running we need to continue running with hard
> coded physical addresses.
Except that you don't need them for arm64 systems, becuase we don't yet
have that legacy for arm64...
> > More worryingly, if soc_is_tegra() returns true but the node isn't
> > found, the drivers decide to do things. If tegra_pmc_early_init passes
> > soc_is_tegra() (because we happened to match "nvidia,tegra132"), but
> > doesn't find a DTB entry for the pmc, it then pokes an assumed
> > hard-coded physical address.
>
> And that's (currently) safe. It at least makes sure that breakage (if
> any) is restricted to Tegra boards. There are other occurrences where
> drivers don't even bother checking for this at all and simply use an
> initcall to unconditionally create a platform device which then goes
> and tries to access physical addresses on *every* board in a multi-
> platform kernel.
So doing the wrong thing on a subset of boards is fine because it's not
all boards?
Imagine I make a typo in a DT for a 64-bit system, and suddnely hit
youre legacy code. Things happen to work, because I don't botehr reading
the splat in dmesg for whatever reason.
Later we do some cleanup of the existing legacy, for whatever reason,
and we get rid of this particular path for arm64.
Suddenly my board won't boot, and to fix that we have to reintroduce
legacy support code only because said legacy support was enabled where
it should never have been in the first place. Another OS trying to
target/use these DTBs has barely any chance of success.
> > Assuming things because of the _lack_ of a node has been a major pain
> > point on 32-bit as DTBs advanced and code was changed, resulting in boot
> > breaking or DTB compatibility breaking over time as old and new kernels
> > made conflicting assumptions.
>
> So what are you saying? We can't support 64-bit ARM because we need to
> keep backwards-compatibility with old DTBs and weren't smart enough to
> see this through all the way three years ago?
No. I'm saying that the workarounds we have in place for legacy DTBs
should be constrained to those cases where legacy DTBs actually exist.
There is no reason for that legacy to carry across to 64-bit.
> > I don't want to see arm64 kernels making assumptions in that manner (nor
> > would I like to see it in new code for 32-bit ARM); it always leads to
> > boot breakage, and it's completely avoidable. The only thing we can know
> > is what is described in the DTB.
>
> I guess we could always make this conditional on 32-bit ARM. Or we could
> decide that we don't care about DTB backwards-compatibility anymore. In
> fact we're quite likely going to break it soon anyway because of Marc's
> gic_arch_extn rework, so might as well start with a clean plate after
> that...
At the very least I think the soc_is_tegra stuff needs to be constrained
to fallback paths that can only possibly run on 32-bit.
> > So NAK to this addition while soc_is_tegra is in use in that way, and
> > NAK to any DTS patches adding any of the strings that would cause
> > soc_is_tegra to return true on arm64.
> >
> > There's also the worrying pattern of initcalls that look for particular
> > compatible strings -- why can't these use the existing DT probing
> > infrastructure rather than rolling their own? It would be nice to not
> > accumulate initcalls that only make sense on one platform.
>
> You make it sound like the "existing DT probing infrastructure" is in
> any way different from initcalls. Furthermore it's typically associated
> with a specific subsystem or class of devices, so adding custom OF init
> tables for a driver that's used on a single platform doesn't strike me
> as something that people would be willing to merge.
Fair point.
Thanks,
Mark.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
