Re: [PATCH 3/6] clk: qcom: gdsc: enable optional power domain support

[Bjorn Andersson]

On Fri 02 Jul 02:35 CDT 2021, Rajendra Nayak wrote:

> 
> 
> On 7/2/2021 2:27 AM, Bjorn Andersson wrote:
> > On Thu 01 Jul 15:12 CDT 2021, Dmitry Baryshkov wrote:
> > 
> > > On Thu, 1 Jul 2021 at 07:23, Bjorn Andersson <bjorn.andersson@xxxxxxxxxx> wrote:
> > > > 
> > > > On Wed 30 Jun 15:29 CDT 2021, Dmitry Baryshkov wrote:
> > > > 
> > > > > On Wed, 30 Jun 2021 at 20:11, Bjorn Andersson
> > > > > <bjorn.andersson@xxxxxxxxxx> wrote:
> > > > > > 
> > > > > > On Wed 30 Jun 10:47 CDT 2021, Dmitry Baryshkov wrote:
> > > > > > 
> > > > > > > Hi,
> > > > > > > 
> > > > > > > On Wed, 30 Jun 2021 at 18:00, Bjorn Andersson
> > > > > > > <bjorn.andersson@xxxxxxxxxx> wrote:
> > > > > > > > 
> > > > > > > > On Wed 30 Jun 08:31 CDT 2021, Dmitry Baryshkov wrote:
> > > > > > > > 
> > > > > > > > > On sm8250 dispcc and videocc registers are powered up by the MMCX power
> > > > > > > > > domain. Currently we used a regulator to enable this domain on demand,
> > > > > > > > > however this has some consequences, as genpd code is not reentrant.
> > > > > > > > > 
> > > > > > > > > Teach Qualcomm clock controller code about setting up power domains and
> > > > > > > > > using them for gdsc control.
> > > > > > > > > 
> > > > > > > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
> > > > > > > > 
> > > > > > > > There's a proposal to add a generic binding for statically assigning a
> > > > > > > > performance states here:
> > > > > > > > 
> > > > > > > > https://lore.kernel.org/linux-arm-msm/1622095949-2014-1-git-send-email-rnayak@xxxxxxxxxxxxxx/
> > > > > 
> > > > > I checked this thread. It looks like Rajendra will also switch to the
> > > > > "required-opps" property. So if that series goes in first, we can drop
> > > > > the call to set_performance_state. If this one goes in first, we can
> > > > > drop the set_performance_state call after getting Rajendra's work in.
> > > > > 
> > > > > > > > 
> > > > > > > > 
> > > > > > > > But that said, do you really need this?
> > > > > > > > 
> > > > > > > > The requirement for driving MMCX to LOW_SVS on SM8250 (and NOM on
> > > > > > > > SM8150/SC8180x) seems to only come from the fact that you push MDP_CLK
> > > > > > > > to 460MHz in &mdss.
> > > > > > > > 
> > > > > > > > But then in &mdss_mdp you do the same using an opp-table based on the
> > > > > > > > actual MDP_CLK, which per its power-domains will scale MMCX accordingly.
> > > > > > > 
> > > > > > > MDSS and DSI would bump up MMCX performance state requirements on
> > > > > > > their own, depending on the frequency being selected.
> > > > > > > 
> > > > > > 
> > > > > > Right, but as I copied things from the sm8250.dtsi to come up with
> > > > > > sm8150/sc8180x.dtsi I concluded that as soon as the assigned-clockrate
> > > > > > in &mdss kicks in I need the performance state to be at NOM.
> > > > > > 
> > > > > > So keeping the assigned-clockrate in &mdss means that MMCX will never go
> > > > > > below NOM.
> > > > > 
> > > > > No, because once MDP is fully running, it will lower the clock frequency:
> > > > > 
> > > > > # grep mdp_clk /sys/kernel/debug/clk/clk_summary
> > > > >            disp_cc_mdss_mdp_clk_src       1        1        0
> > > > > 150000000          0     0  50000         ?
> > > > >               disp_cc_mdss_mdp_clk       2        2        0
> > > > > 150000000          0     0  50000         Y
> > > > > 
> > > > 
> > > > But won't that just lower the performance state requested by the
> > > > &mdss_mdp, while the &mdss still votes for NOM - with the outcome being
> > > > that we maintain NOM even if the clock goes down?
> > > 
> > > &mdss doesn't vote on performance state. At least it does not on
> > > msm/msm-next which I have at hand right now.
> > > &mdss toggles mdss_gdsc, but does not assign any performance state.
> > > 
> > 
> > Right, but per the upstream implementation, enabling MDSS_GDSC could in
> > itself fail, because unless something else has driven up the performance
> > state the enable that trickles up won't actually turn on the supply.
> > 
> > > On the other hand &mdss_mdp and &dsi0 clearly vote on mmcx's performance state.
> > > 
> > 
> > Right, but it does so as part of its clock scaling, so this makes
> > perfect sense to me.
> > 
> > > > 
> > > > > > 
> > > > > > > > So wouldn't it be sufficient to ensure that MDSS_GDSC is parented by
> > > > > > > > MMCX and then use opp-tables associated with the devices that scales the
> > > > > > > > clock and thereby actually carries the "required-opps".
> > > > > > > 
> > > > > > > Actually no. I set the performance state in the qcom_cc_map, so that
> > > > > > > further register access is possible. Initially I was doing this in the
> > > > > > > qcom_cc_really_probe() and it was already too late.
> > > > > > > Just to remind: this patchset is not about MDSS_GDSC being parented by
> > > > > > > MMCX, it is about dispcc/videocc registers being gated with MMCX.
> > > > > > > 
> > > > > > 
> > > > > > So you're saying that just enabling MMCX isn't enough to touch the
> > > > > > dispcc/videocc registers? If that's the case it seems like MMCX's
> > > > > > definition of "on" needs to be adjusted - because just specifying MMCX
> > > > > > as the power-domain for dispcc/videocc and enabling pm_runtime should
> > > > > > ensure that MMCX is enabled when the clock registers are accessed (I
> > > > > > don't see anything like that for the GDSC part though).
> > > > > 
> > > > > No, it is not enough. If I comment out the set_performance_state call,
> > > > > the board reboots.
> > > > > 
> > > > > However I can set the opps as low as RET and register access will work.
> > > > > I'll run more experiments and if everything works as expected, I can
> > > > > use retention or min_svs level in the next iteration.
> > > > > Just note that downstream specifies low_svs as minimum voltage level
> > > > > for MMCX regulator.
> > > > > 
> > > > 
> > > > It doesn't make sense to me that a lone power_on on the power-domain
> > > > wouldn't give us enough juice to poke the registers.
> > > > 
> > > > But digging into the rpmhpd implementation answers the question, simply
> > > > invoking rpmhpd_power_on() is a nop, unless
> > > > rpmhpd_set_performance_state() has previously been called, because
> > > > pd->corner is 0. So this explains why enable isn't sufficient.
> > > > 
> > > > Compare this with the rpmpd implementation that will send an
> > > > enable request to the RPM in this case.
> 
> Right, in case of RPMh, there was no separate 'enable' request which
> could be sent, there was just a 'corner' request.
> 
> I don't completely recall, but the reason to not send a 'default corner'
> on enable was perhaps to keep the enable and set_performance orthogonal.
> 
> However, given we then decided to send the lowest possible corner
> in disable, it perhaps makes sense to send a 'lowest non-zero corner' on enable
> as well.
> 

I was slightly worries that the change would dump cx and mx from
whatever level the bootloader put it at down to LOW_SVS during boot.

But both rb3 and rb5 boots fine with this change, so I posted it here:
https://lore.kernel.org/linux-arm-msm/20210703025449.2687201-1-bjorn.andersson@xxxxxxxxxx/

> > > 
> > > Do you think that we should change that to:
> > > 
> > > rpmhpd_aggregate_corner(pd, max(pd->corner, 1)) ?
> > > 
> > > Or
> > > 
> > > rpmhpd_aggregate_corner(pd, max(pd->corner, pd->levels[1])) ?
> > > 
> > 
> > In rpmhpd_power_on() and rpmhpd_set_performance_state() we pass the
> > index of the entry in pd->levels[] that we want, but in
> > rpmhpd_power_off() we pass the value of pd->levels[0].
> > 
> > So I would suggest dropping the if (pd->corner) and doing:
> > 
> >    rpmhpd_aggregate_corner(pd, max(pd->corner, 1));
> 
> So the index value represents the hlvl (0-15) that eventually gets sent to
> rpmh, the pd->levels are the sparse vlvl values that come from the command
> DB mappings.
> 
> What seems sane is to sent the lowest non-zero vlvl. That in most cases
> would be at index 1, but for some which do not support complete off,
> it could be at index 0.
> 

I took this into consideration in above patch, keeping track of the
first non-zero corner and using this when the domain is enabled.

Unfortunately, if the first entry would be say LOW_SVS power_off would
request corner (hlvl) 64. So I fixed that in patch 1/2 in above series.

Regards,
Bjorn

> > 
> > And it seems both rb3 and rb5 still boots with this change (but I need
> > to do some more testing to know for sure).
> > 
> > > > 
> > > > > > I thought our problem you had was that you need to set a
> > > > > > performance_state in order to clock up some of the clocks - e.g.
> > > > > > MDP_CLK.
> > > > > 
> > > > > No, even register access needs proper perf state.
> > > > > 
> > > > 
> > > > Per above finding you're right, enabling a rpmhpd power-domain doesn't
> > > > do anything. And I don't find this intuitive or even in line with the
> > > > expectations of the api...
> > > > 
> > > > 
> > > > 
> > > > A quick test booting rb3 and rb5 seems to indicate that it's possible to
> > > > initialize pd->corner to 1 (to ensure that enable at least gives us the
> > > > lowest level).
> > > > 
> > > > set_performance_state(0) will however then result in voting for "off",
> > > > rather than the lowest enabled level.
> > > 
> > > Well, set_performance_state(0) means that "the device wouldn't
> > > participate anymore to find the target performance state of the
> > > genpd".
> > 
> > I agree.
> > 
> > > Strictly speaking it does not specify whether it is ok to turn
> > > it off or not. (like the regulator with the voltage set to 0V).
> > > But I'd also like to hear a comment from Stephen here.
> > > 
> > 
> > Looking at other power-domains (e.g. gdsc and rpmpd) enabling the
> > power-domain means it is no longer off and if you need some specific
> > performance state you have to vote for that.
> > 
> > So I'm also interested in hearing if there's any reasoning behind how
> > this was written.
> > 
> > Regards,
> > Bjorn
> > 
> 
> -- 
> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
> of Code Aurora Forum, hosted by The Linux Foundation