On Wed, Aug 22, 2018 at 11:54 PM Sascha Hauer <s.hauer@xxxxxxxxxxxxxx> wrote:
>
> On Wed, Aug 22, 2018 at 06:07:01PM -0700, Andrey Smirnov wrote:
> > On Wed, Aug 22, 2018 at 12:31 AM Sascha Hauer <s.hauer@xxxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Aug 20, 2018 at 11:28:53PM -0700, Andrey Smirnov wrote:
> > > > A number of PLL pairs (e.g. "sys1_pll1" and "sys1_pll2") share the
> > > > same configuration register, so touching PD bit, as is done for
> > > > SCCG_PLL2 in its prepare/unprepare methods will result in shut down of
> > > > both PLLs. This is very undesireable, since attempting to re-parent a
> > > > clock to "sys1_pll2" might result in complete system shutdown due to
> > > > "sys1_pll1" being shut-down as a part of re-parenting process.
> > >
> > > I can imagine that there are problems with the way it is currently
> > > handled, but the scenario you describe shouldn't happen. "sys1_pll1"
> > > will never be shut down because it doesn't have a disable hook:
> > >
> > > static const struct clk_ops clk_sccg_pll1_ops = {
> > > .is_enabled = clk_pll1_is_prepared,
> > > .recalc_rate = clk_pll1_recalc_rate,
> > > .round_rate = clk_pll1_round_rate,
> > > .set_rate = clk_pll1_set_rate,
> > > };
> > >
> > > static const struct clk_ops clk_sccg_pll2_ops = {
> > > .enable = clk_pll1_prepare,
> > > .disable = clk_pll1_unprepare,
> > > .recalc_rate = clk_pll2_recalc_rate,
> > > .round_rate = clk_pll2_round_rate,
> > > .set_rate = clk_pll2_set_rate,
> > > };
> > >
> > > Have I missed something?
> > >
> >
> > Yes, but that's probably due to my explanation being too vague, so let
> > me try again. Both clocks "sys1_pll1" and "sys1_pll2" are
> > controlled/configured via separate bit fields in the _same_ register.
> > That register also has a single "powerdown"/"PD" bit that is shared
> > between both PLLs and affects them both. The way code is currently
> > implemented calling clk_disable() on "sys1_pll2" will clear PD bit and
> > as a result "sys1_pll1" will be shut down as well. This is not that
> > far-fetched of a scenario because a number of derived clocks that go
> > out to individual peripherals are specified with
> > CLK_OPS_PARENT_ENABLE, so trivial clock re-parenting like this one:
> > https://github.com/ndreys/barebox/commit/802d1d5be1f1f3df996ba26b115dd3a55bc10c67
> > will result in hung system.
> >
> > Hopefully this does a better job of explaining what I was trying to fix.
>
> Yeah, I thought you mixed up both PLLs, but reading your initial
> eplanation again you were right from the start.
>
> I agree that we should just remove the enable/disable hooks and see what
> we have to do when somebody actually tries to do something with the
> clocks (other than get the rate from)
>
Hmm, revisiting this in light of a related kernel thread [1], I think
this problem can be avoided if reference counters of critical clocks
are properly initialized. Unless you already applied this patch, I
think we should just wait for the kernel code to land, I'll backport
it and we can work on resolving any issues if they arise then.
Thanks,
Andrey Smirnov
[1] https://lore.kernel.org/lkml/1534945703-4730-4-git-send-email-abel.vesa@xxxxxxx/T/#u
_______________________________________________
barebox mailing list
barebox@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/barebox
