On Mon, 31 May 2021 at 22:07, Dmitry Osipenko <digetx@xxxxxxxxx> wrote:
>
> 31.05.2021 16:17, Ulf Hansson пишет:
> > On Mon, 24 May 2021 at 22:23, Dmitry Osipenko <digetx@xxxxxxxxx> wrote:
> >>
> >> 24.05.2021 20:04, Ulf Hansson пишет:
> >>> On Mon, 24 May 2021 at 01:13, Dmitry Osipenko <digetx@xxxxxxxxx> wrote:
> >>>>
> >>>> NVIDIA Tegra SoCs have multiple power domains, each domain corresponds
> >>>> to an external SoC power rail. Core power domain covers vast majority of
> >>>> hardware blocks within a Tegra SoC. The voltage of a power domain should
> >>>> be set to a level which satisfies all devices within the power domain.
> >>>> Add support for the core power domain which controls voltage state of the
> >>>> domain. This allows us to support system-wide DVFS on Tegra20-210 SoCs.
> >>>> The PMC powergate domains now are sub-domains of the core domain, this
> >>>> requires device-tree updating, older DTBs are unaffected.
> >>>>
> >>>> Tested-by: Peter Geis <pgwipeout@xxxxxxxxx> # Ouya T30
> >>>> Tested-by: Paul Fertser <fercerpav@xxxxxxxxx> # PAZ00 T20
> >>>> Tested-by: Nicolas Chauvet <kwizart@xxxxxxxxx> # PAZ00 T20 and TK1 T124
> >>>> Tested-by: Matt Merhar <mattmerhar@xxxxxxxxxxxxxx> # Ouya T30
> >>>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
> >>>
> >>> [...]
> >>>
> >>>> +
> >>>> +static int tegra_pmc_core_pd_add(struct tegra_pmc *pmc, struct device_node *np)
> >>>> +{
> >>>> + static struct lock_class_key tegra_core_domain_lock_class;
> >>>> + struct generic_pm_domain *genpd;
> >>>> + const char *rname = "core";
> >>>> + int err;
> >>>> +
> >>>> + genpd = devm_kzalloc(pmc->dev, sizeof(*genpd), GFP_KERNEL);
> >>>> + if (!genpd)
> >>>> + return -ENOMEM;
> >>>> +
> >>>> + genpd->name = np->name;
> >>>> + genpd->set_performance_state = tegra_pmc_core_pd_set_performance_state;
> >>>> + genpd->opp_to_performance_state = tegra_pmc_core_pd_opp_to_performance_state;
> >>>> +
> >>>> + err = devm_pm_opp_set_regulators(pmc->dev, &rname, 1);
> >>>> + if (err)
> >>>> + return dev_err_probe(pmc->dev, err,
> >>>> + "failed to set core OPP regulator\n");
> >>>> +
> >>>> + err = pm_genpd_init(genpd, NULL, false);
> >>>> + if (err) {
> >>>> + dev_err(pmc->dev, "failed to init core genpd: %d\n", err);
> >>>> + return err;
> >>>> + }
> >>>> +
> >>>> + /*
> >>>> + * We have a "PMC pwrgate -> Core" hierarchy of the power domains
> >>>> + * where PMC needs to resume and change performance (voltage) of the
> >>>> + * Core domain from the PMC GENPD on/off callbacks, hence we need
> >>>> + * to annotate the lock in order to remove confusion from the
> >>>> + * lockdep checker when a nested access happens.
> >>>> + */
> >>>
> >>> Can you elaborate a bit more on this?
> >>>
> >>> Are you saying that when the child domain (PMC pwrgate) gets powered
> >>> off, you want to drop its aggregated votes it may hold for the
> >>> performance state, as otherwise it may affect the parent domain (core
> >>> domain)?
> >>
> >> Yes, in particular we want to remove/add the performance vote when clk is disabled/enabled, see tegra_clock_runtime_resume/suspend() of the clk-runtimePM driver [1]. I'll send that clk patch separately once this series and some other tegra-clk patches will be merged, otherwise there are too many dependencies and it's too difficult to review.
> >
> > You are likely correct from the merging point of view, but for
> > completeness I would prefer to look at the whole series. Would you
> > mind folding in some of these changes too?
>
> Sure, I will send those changes once this fundamental patchset will be merged.
>
> >> [1] https://patchwork.ozlabs.org/project/linux-tegra/patch/20201217180638.22748-33-digetx@xxxxxxxxx/
> >
> > Hmm. In general, the new changes to genpd and the opp library with
> > "performance states" for DVFS, should help to avoid using clock rate
> > notifications.
> >
> > Instead of updating the performance votes from the clock provider
> > driver, the more proper thing would be to let the clock consumer
> > driver (various drivers) to call dev_pm_opp_set_rate() when it needs
> > to move to a new OPP. This also means calling dev_pm_opp_set_rate(dev,
> > 0) when the votes for an OPP can be dropped.
> >
> > In this way, the opp library will call genpd to update the performance
> > state vote for the corresponding device.
>
> This is not sufficient for Tegra because we have individual OPP tables for the root PLLs, system clocks and device clocks. The device clocks could be muxed to a different PLLs, depending on clk requirements of a particular board.
Are you saying that the clock providers for the "root PLLs" and
"system clocks" have OPP tables themselves? If so, would you mind
posting a patch for an updated DT binding for these changes, so it can
be discussed separately?
>
> >> Please see the example lockdep trace in the end of the email that is fixed by the mutex annotation. What we have there is the tegra-host1x device driver that resumes PMC powergate domain on Tegra30, the PMC driver enables clock from the genpd.power_on callback of the powergate domain and this propagates to the clock's RPM callback which sets the performance vote of the core domain. Hence core domain vote is set from within of the powergate domain.
> >
> > Right, this sounds like a fragile looking path. Are you sure it can't
> > lead into deadlock situations?
>
> The locking was a concern, but turned out that it works nicely for the Tegra clk/power tree. This approach may not be suitable for other SoCs.
>
> > In any case, we designed dev_pm_opp_set_rate() (and its friends in
> > genpd) with these locking issues in mind.
>
> The device drivers don't manage the parent clocks directly and OPP core doesn't support this use-case where OPP needs to be applied to a generic/parent PLL clock. Moving the OPP management to the clk driver is the easy solution which works good in practice for Tegra, it also removes a need to switch each driver to dev_pm_opp_set_rate() usage.
I admit, if clock consumer drivers could avoid calling
dev_pm_opp_set_rate|opp(), that would be nice. But, as I stated, it's
a fragile path from locking point of view, to call
dev_pm_opp_set_rate|opp() from a clock provider driver. Personally, I
think it's better to avoid it.
More importantly, you also need to convince the clock subsystem
maintainers, that setting an OPP internally from the clock provider
driver is a good idea. As far as I can tell, they have said *no* to
this, since the common clock framework was invented, I believe for
good reasons.
>
> >>> I guess this would be a valid scenario to optimize for, especially if
> >>> you have more than one child domain of the core power domain, right?
> >>>
> >>> If you have only one child domain, would it be sufficient to assign
> >>> ->power_on|off() callbacks for the core domain and deal with the
> >>> performance stare votes from there instead?
> >>
> >> The core domain is the parent domain of the PMC domains + some devices directly belong to the core domain. The GENPD core aggregates the performance votes from the children domains and from devices of the parent core, this all works great already.
> >>
> >> It sounds to me that you're suggesting to reinvent the aggregation logic within the PMC driver and create a fake hierarchy that doesn't match hardware. It won't help the lockdep warning anyways.
> >>
> >> I actually don't quite understand what problem you're trying to solve, could you please explain? The lockdep warning is harmless, we just need to annotate the mutex lock class.
> >>
> >> If you don't feel comfortable with the usage of lockdep_set_class() in the driver, then maybe it should be possible to make it a part of the pm_genpd_init(). For example like we did that for tegra-host1x driver recently [2].
> >
> > I was not trying to solve a problem, but was just curious and wanted
> > to ask about the reasons for the lockdep_set_class(), as it simply
> > caught my attention while reviewing.
> >
> > Looks like there may be something fishy going on, but let's see, I may be wrong.
> >
> >>
> >> [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a24f98176d1efae2c37d3438c57a624d530d9c33
> >>
> >>
> >> LOCKDEP
> >> ============================================
> ...
> >
> > Thanks for sharing the log.
> >
> > Could you perhaps point me to the corresponding DTS files. I would
> > like to understand more about the PM domain hierarchy and where the
> > clock controller may be located.
>
> https://github.com/grate-driver/linux/tree/master has all the latest versions of patches applied.
>
> For example please see clock@60006000 and pmc@7000e400 nodes of [1].
>
> [1] https://github.com/grate-driver/linux/blob/master/arch/arm/boot/dts/tegra30.dtsi
Thanks, that certainly helped me understand better!
I see that you want to add OPP tables to clock provider nodes. As I
said above, an updated DT binding is probably a good idea to discuss
separately.
>
> This is how output of pm_genpd_summary looks like on Tegra30 Nexus 7, where performance=voltage:
>
> domain status children performance
> /device runtime status
> ----------------------------------------------------------------------------------------------
> heg on 1000000
> /devices/soc0/50000000.host1x active 1000000
> /devices/soc0/50000000.host1x/540c0000.epp suspended 0
> /devices/soc0/50000000.host1x/54140000.gr2d suspended 0
> mpe off-0 0
> /devices/soc0/50000000.host1x/54040000.mpe suspended 0
> vdec off-0 0
> /devices/soc0/6001a000.vde suspended 0
> venc off-0 0
> /devices/soc0/50000000.host1x/54080000.vi suspended 0
> /devices/soc0/50000000.host1x/54100000.isp suspended 0
> 3d1 off-0 0
> /devices/genpd:1:54180000.gr3d suspended 0
> 3d0 off-0 0
> /devices/genpd:0:54180000.gr3d suspended 0
> core-domain on 1000000
> 3d0, 3d1, venc, vdec, mpe, heg
> /devices/soc0/7000f400.memory-controller unsupported 1000000
> /devices/platform/tegra_clk_pll_c active 1000000
> /devices/platform/tegra_clk_pll_e suspended 0
> /devices/platform/tegra_clk_pll_m active 1000000
> /devices/platform/tegra_clk_sclk active 1000000
> /devices/platform/tegra_clk_dsia suspended 0
> /devices/platform/tegra_clk_dsib suspended 0
> /devices/platform/tegra_clk_spdif_out suspended 0
> /devices/platform/tegra_clk_se suspended 0
> /devices/platform/tegra_clk_hdmi suspended 0
> /devices/platform/tegra_clk_pwm active 1000000
> /devices/platform/tegra_clk_pcie suspended 0
> /devices/platform/tegra_clk_afi suspended 0
> /devices/platform/tegra_clk_vde suspended 0
> /devices/platform/tegra_clk_vi suspended 0
> /devices/platform/tegra_clk_epp suspended 0
> /devices/platform/tegra_clk_mpe suspended 0
> /devices/platform/tegra_clk_nor suspended 0
> /devices/platform/tegra_clk_sdmmc1 suspended 0
> /devices/platform/tegra_clk_sdmmc3 active 950000
> /devices/platform/tegra_clk_mipi suspended 0
> /devices/platform/tegra_clk_sbc1 suspended 0
> /devices/platform/tegra_clk_sbc2 suspended 0
> /devices/platform/tegra_clk_sbc3 suspended 0
> /devices/platform/tegra_clk_sbc4 suspended 0
> /devices/platform/tegra_clk_sbc5 suspended 0
> /devices/platform/tegra_clk_sbc6 suspended 0
> /devices/platform/tegra_clk_cve suspended 0
> /devices/platform/tegra_clk_tvo suspended 0
> /devices/platform/tegra_clk_tvdac suspended 0
> /devices/platform/tegra_clk_ndflash suspended 0
> /devices/platform/tegra_clk_sata_oob suspended 0
> /devices/platform/tegra_clk_sata suspended 0
> /devices/platform/tegra_clk_fuse_burn suspended 0
> /devices/platform/tegra_clk_usbd active 1000000
> /devices/platform/tegra_clk_usb2 suspended 0
> /devices/platform/tegra_clk_usb3 suspended 0
> /devices/soc0/50000000.host1x/54200000.dc active 1000000
> /devices/soc0/50000000.host1x/54240000.dc suspended 0
Okay, to not stall things from moving forward, may I suggest that you
simply drop the call to lockdep_set_class() (and the corresponding
comment) for now.
Then you can continue to post the next parts - and if it turns out
that lockdep_set_class() becomes needed, you can always add it back
then.
Kind regards
Uffe
