| Message ID | 20190625164733.11091-9-jorge.ramirez-ortiz@linaro.org (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
| Series | Support CPU frequency scaling on QCS404 | expand |
On Tue 25 Jun 09:47 PDT 2019, Jorge Ramirez-Ortiz wrote: > When COMMON_CLK_DISABLED_UNUSED is set, in an effort to save power and > to keep the software model of the clock in line with reality, the > framework transverses the clock tree and disables those clocks that > were enabled by the firmware but have not been enabled by any device > driver. > > If CPUFREQ is enabled, early during the system boot, it might attempt > to change the CPU frequency ("set_rate"). If the HFPLL is selected as > a provider, it will then change the rate for this clock. > > As boot continues, clk_disable_unused_subtree will run. Since it wont > find a valid counter (enable_count) for a clock that is actually > enabled it will attempt to disable it which will cause the CPU to > stop. But if CPUfreq has acquired the CPU clock and the hfpll is the currently selected input, why does the clock framework not know about this clock being used? > Notice that in this driver, calls to check whether the clock is > enabled are routed via the is_enabled callback which queries the > hardware. > > The following commit, rather than marking the clock critical and > forcing the clock to be always enabled, addresses the above scenario > making sure the clock is not disabled but it continues to rely on the > firmware to enable the clock. > > Co-developed-by: Niklas Cassel <niklas.cassel@linaro.org> > Signed-off-by: Niklas Cassel <niklas.cassel@linaro.org> > Signed-off-by: Jorge Ramirez-Ortiz <jorge.ramirez-ortiz@linaro.org> I can see that we have a real issue in the case where CPUfreq is not enabled and hence there are no clients, according to Linux. And that I don't know another way to guard against. Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org> Regards, Bjorn > --- > drivers/clk/qcom/hfpll.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/clk/qcom/hfpll.c b/drivers/clk/qcom/hfpll.c > index 0ffed0d41c50..d5fd27938e7b 100644 > --- a/drivers/clk/qcom/hfpll.c > +++ b/drivers/clk/qcom/hfpll.c > @@ -58,6 +58,13 @@ static int qcom_hfpll_probe(struct platform_device *pdev) > .parent_names = (const char *[]){ "xo" }, > .num_parents = 1, > .ops = &clk_ops_hfpll, > + /* > + * rather than marking the clock critical and forcing the clock > + * to be always enabled, we make sure that the clock is not > + * disabled: the firmware remains responsible of enabling this > + * clock (for more info check the commit log) > + */ > + .flags = CLK_IGNORE_UNUSED, > }; > > h = devm_kzalloc(dev, sizeof(*h), GFP_KERNEL); > -- > 2.21.0 >
On 7/11/19 17:16, Bjorn Andersson wrote: > On Tue 25 Jun 09:47 PDT 2019, Jorge Ramirez-Ortiz wrote: > >> When COMMON_CLK_DISABLED_UNUSED is set, in an effort to save power and >> to keep the software model of the clock in line with reality, the >> framework transverses the clock tree and disables those clocks that >> were enabled by the firmware but have not been enabled by any device >> driver. >> >> If CPUFREQ is enabled, early during the system boot, it might attempt >> to change the CPU frequency ("set_rate"). If the HFPLL is selected as >> a provider, it will then change the rate for this clock. >> >> As boot continues, clk_disable_unused_subtree will run. Since it wont >> find a valid counter (enable_count) for a clock that is actually >> enabled it will attempt to disable it which will cause the CPU to >> stop. > > But if CPUfreq has acquired the CPU clock and the hfpll is the currently > selected input, why does the clock framework not know about this clock > being used? right, see the comment right below - maybe I should have been more explicit at the time. sorry about it. > >> Notice that in this driver, calls to check whether the clock is >> enabled are routed via the is_enabled callback which queries the >> hardware. calls to check whether the clock is enabled dont use the usage counter but a hardware read. IIRC the clock framework will check some counter to know if the clock is being used. >> >> The following commit, rather than marking the clock critical and >> forcing the clock to be always enabled, addresses the above scenario >> making sure the clock is not disabled but it continues to rely on the >> firmware to enable the clock. >> >> Co-developed-by: Niklas Cassel <niklas.cassel@linaro.org> >> Signed-off-by: Niklas Cassel <niklas.cassel@linaro.org> >> Signed-off-by: Jorge Ramirez-Ortiz <jorge.ramirez-ortiz@linaro.org> > > > I can see that we have a real issue in the case where CPUfreq is not > enabled and hence there are no clients, according to Linux. And that I > don't know another way to guard against. the issue is there when CPUfreq is enabled that is for sure (if we just remove this commit the system will not boot due to the situation I tried to describe above). > > Reviewed-by: Bjorn Andersson <bjorn.andersson@linaro.org> > > Regards, > Bjorn > >> --- >> drivers/clk/qcom/hfpll.c | 7 +++++++ >> 1 file changed, 7 insertions(+) >> >> diff --git a/drivers/clk/qcom/hfpll.c b/drivers/clk/qcom/hfpll.c >> index 0ffed0d41c50..d5fd27938e7b 100644 >> --- a/drivers/clk/qcom/hfpll.c >> +++ b/drivers/clk/qcom/hfpll.c >> @@ -58,6 +58,13 @@ static int qcom_hfpll_probe(struct platform_device *pdev) >> .parent_names = (const char *[]){ "xo" }, >> .num_parents = 1, >> .ops = &clk_ops_hfpll, >> + /* >> + * rather than marking the clock critical and forcing the clock >> + * to be always enabled, we make sure that the clock is not >> + * disabled: the firmware remains responsible of enabling this >> + * clock (for more info check the commit log) >> + */ >> + .flags = CLK_IGNORE_UNUSED, >> }; >> >> h = devm_kzalloc(dev, sizeof(*h), GFP_KERNEL); >> -- >> 2.21.0 >> >
diff --git a/drivers/clk/qcom/hfpll.c b/drivers/clk/qcom/hfpll.c index 0ffed0d41c50..d5fd27938e7b 100644 --- a/drivers/clk/qcom/hfpll.c +++ b/drivers/clk/qcom/hfpll.c @@ -58,6 +58,13 @@ static int qcom_hfpll_probe(struct platform_device *pdev) .parent_names = (const char *[]){ "xo" }, .num_parents = 1, .ops = &clk_ops_hfpll, + /* + * rather than marking the clock critical and forcing the clock + * to be always enabled, we make sure that the clock is not + * disabled: the firmware remains responsible of enabling this + * clock (for more info check the commit log) + */ + .flags = CLK_IGNORE_UNUSED, }; h = devm_kzalloc(dev, sizeof(*h), GFP_KERNEL);
When COMMON_CLK_DISABLED_UNUSED is set, in an effort to save power and to keep the software model of the clock in line with reality, the framework transverses the clock tree and disables those clocks that were enabled by the firmware but have not been enabled by any device driver. If CPUFREQ is enabled, early during the system boot, it might attempt to change the CPU frequency ("set_rate"). If the HFPLL is selected as a provider, it will then change the rate for this clock. As boot continues, clk_disable_unused_subtree will run. Since it wont find a valid counter (enable_count) for a clock that is actually enabled it will attempt to disable it which will cause the CPU to stop. Notice that in this driver, calls to check whether the clock is enabled are routed via the is_enabled callback which queries the hardware. The following commit, rather than marking the clock critical and forcing the clock to be always enabled, addresses the above scenario making sure the clock is not disabled but it continues to rely on the firmware to enable the clock. Co-developed-by: Niklas Cassel <niklas.cassel@linaro.org> Signed-off-by: Niklas Cassel <niklas.cassel@linaro.org> Signed-off-by: Jorge Ramirez-Ortiz <jorge.ramirez-ortiz@linaro.org> --- drivers/clk/qcom/hfpll.c | 7 +++++++ 1 file changed, 7 insertions(+)