diff mbox series

clk: qcom: gcc-msm8953: fix stuck gcc_usb30_master_clk

Message ID 20231002170021.192740-1-trabarni@gmail.com (mailing list archive)
State Changes Requested
Headers show
Series clk: qcom: gcc-msm8953: fix stuck gcc_usb30_master_clk | expand

Commit Message

Barnabás Czémán Oct. 2, 2023, 5 p.m. UTC
According to downstream dwc3-msm source this clock has FSM dependency on
gcc_pcnoc_usb30_clk so enabling it would fail if latter isn't enabled.
This patch add works around this issue by changing parent of
gcc_usb30_master_clk to gcc_pcnoc_usb30_clk. This is acceptable because
both clocks have same parent and are branches/gates.

Signed-off-by: Barnabás Czémán <trabarni@gmail.com>
---
 drivers/clk/qcom/gcc-msm8953.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Konrad Dybcio Oct. 6, 2023, 11:50 p.m. UTC | #1
On 2.10.2023 19:00, Barnabás Czémán wrote:
> According to downstream dwc3-msm source this clock has FSM dependency on
> gcc_pcnoc_usb30_clk so enabling it would fail if latter isn't enabled.
> This patch add works around this issue by changing parent of
> gcc_usb30_master_clk to gcc_pcnoc_usb30_clk. This is acceptable because
> both clocks have same parent and are branches/gates.
> 
> Signed-off-by: Barnabás Czémán <trabarni@gmail.com>
> ---
"meh"

There are multiple cases, especially with qcom, where there are some
magic "dependencies" without parent-child relationship. The common
clock framework doesn't currently have any good way to handle this,
other than some mind gymnastics like you had to do here with matching
them against a common parent/ancestor..

Stephen, what do you say?

Konrad
Stephen Boyd Oct. 24, 2023, 2:59 a.m. UTC | #2
Quoting Konrad Dybcio (2023-10-06 16:50:18)
> On 2.10.2023 19:00, Barnabás Czémán wrote:
> > According to downstream dwc3-msm source this clock has FSM dependency on
> > gcc_pcnoc_usb30_clk so enabling it would fail if latter isn't enabled.
> > This patch add works around this issue by changing parent of
> > gcc_usb30_master_clk to gcc_pcnoc_usb30_clk. This is acceptable because
> > both clocks have same parent and are branches/gates.
> > 
> > Signed-off-by: Barnabás Czémán <trabarni@gmail.com>
> > ---
> "meh"
> 
> There are multiple cases, especially with qcom, where there are some
> magic "dependencies" without parent-child relationship. The common
> clock framework doesn't currently have any good way to handle this,
> other than some mind gymnastics like you had to do here with matching
> them against a common parent/ancestor..
> 
> Stephen, what do you say?
> 

You can't change the parent to be not the actual parent. The consumer of
the branch probably wants to call clk_set_rate() on the branch and have
it propagate up to the parent to set the actual rate. Can the axi clk
simply be left enabled all the time? That seems simpler. Otherwise we
probably need to leave the axi clk control to the interconnect driver
and make sure drivers enable interconnects before enabling this branch.

When things start to get this tangled I tend to think that we need to
remove control of the clk from the general drivers and put the logic to
control interconnects and clks into some SoC glue driver and expose a
single interface, like genpd power_on/power_off so that general drivers
can't get the sequence of steps wrong. Instead all they can do is "power
on" their device, and the SoC glue driver can do the proper sequence of
framework calls to power up the device.
Konrad Dybcio Nov. 18, 2023, 12:48 a.m. UTC | #3
On 24.10.2023 04:59, Stephen Boyd wrote:
> Quoting Konrad Dybcio (2023-10-06 16:50:18)
>> On 2.10.2023 19:00, Barnabás Czémán wrote:
>>> According to downstream dwc3-msm source this clock has FSM dependency on
>>> gcc_pcnoc_usb30_clk so enabling it would fail if latter isn't enabled.
>>> This patch add works around this issue by changing parent of
>>> gcc_usb30_master_clk to gcc_pcnoc_usb30_clk. This is acceptable because
>>> both clocks have same parent and are branches/gates.
>>>
>>> Signed-off-by: Barnabás Czémán <trabarni@gmail.com>
>>> ---
>> "meh"
>>
>> There are multiple cases, especially with qcom, where there are some
>> magic "dependencies" without parent-child relationship. The common
>> clock framework doesn't currently have any good way to handle this,
>> other than some mind gymnastics like you had to do here with matching
>> them against a common parent/ancestor..
>>
>> Stephen, what do you say?
>>
> 
> You can't change the parent to be not the actual parent. The consumer of
> the branch probably wants to call clk_set_rate() on the branch and have
> it propagate up to the parent to set the actual rate. Can the axi clk
> simply be left enabled all the time? That seems simpler. Otherwise we
> probably need to leave the axi clk control to the interconnect driver
> and make sure drivers enable interconnects before enabling this branch.
Yeah I'm almost inclined to think adding even more ifs to the icc driver
will consume more power than just leaving the AXI hanging..

> 
> When things start to get this tangled I tend to think that we need to
> remove control of the clk from the general drivers and put the logic to
> control interconnects and clks into some SoC glue driver and expose a
> single interface, like genpd power_on/power_off so that general drivers
> can't get the sequence of steps wrong. Instead all they can do is "power
> on" their device, and the SoC glue driver can do the proper sequence of
> framework calls to power up the device.
That too, given the structure of qcom SoCs, it should almost look like:

xyznoc-bus {
	compatible = "simple-pm-bus";
	clocks = <&gcc xyznoc_ahb>,
		 <&gcc xyznoc_axi>;
	...

	xyznoc-node@abcd {};
}

etc.

I've actually discussed this with Bjorn, but we came to a conclusion
that it's not trivial to determine which peripheral lives on which NoC
+ many of them seem to sorta overlap more than one..

Konrad
Krishna Kurapati Nov. 22, 2023, 6:09 a.m. UTC | #4
On 11/18/2023 6:18 AM, Konrad Dybcio wrote:
> On 24.10.2023 04:59, Stephen Boyd wrote:
>> Quoting Konrad Dybcio (2023-10-06 16:50:18)
>>> On 2.10.2023 19:00, Barnabás Czémán wrote:
>>>> According to downstream dwc3-msm source this clock has FSM dependency on
>>>> gcc_pcnoc_usb30_clk so enabling it would fail if latter isn't enabled.
>>>> This patch add works around this issue by changing parent of
>>>> gcc_usb30_master_clk to gcc_pcnoc_usb30_clk. This is acceptable because
>>>> both clocks have same parent and are branches/gates.
>>>>
>>>> Signed-off-by: Barnabás Czémán <trabarni@gmail.com>
>>>> ---
>>> "meh"
>>>
>>> There are multiple cases, especially with qcom, where there are some
>>> magic "dependencies" without parent-child relationship. The common
>>> clock framework doesn't currently have any good way to handle this,
>>> other than some mind gymnastics like you had to do here with matching
>>> them against a common parent/ancestor..
>>>
>>> Stephen, what do you say?
>>>
>>
>> You can't change the parent to be not the actual parent. The consumer of
>> the branch probably wants to call clk_set_rate() on the branch and have
>> it propagate up to the parent to set the actual rate. Can the axi clk
>> simply be left enabled all the time? That seems simpler. Otherwise we
>> probably need to leave the axi clk control to the interconnect driver
>> and make sure drivers enable interconnects before enabling this branch.
> Yeah I'm almost inclined to think adding even more ifs to the icc driver
> will consume more power than just leaving the AXI hanging..
> 
>>
>> When things start to get this tangled I tend to think that we need to
>> remove control of the clk from the general drivers and put the logic to
>> control interconnects and clks into some SoC glue driver and expose a
>> single interface, like genpd power_on/power_off so that general drivers
>> can't get the sequence of steps wrong. Instead all they can do is "power
>> on" their device, and the SoC glue driver can do the proper sequence of
>> framework calls to power up the device.
> That too, given the structure of qcom SoCs, it should almost look like:
> 
> xyznoc-bus {
> 	compatible = "simple-pm-bus";
> 	clocks = <&gcc xyznoc_ahb>,
> 		 <&gcc xyznoc_axi>;
> 	...
> 
> 	xyznoc-node@abcd {};
> }
> 
> etc.
> 
> I've actually discussed this with Bjorn, but we came to a conclusion
> that it's not trivial to determine which peripheral lives on which NoC
> + many of them seem to sorta overlap more than one..

Are we seeing the clk getting stuck during suspend/resume or during 
clk_prepare_enable in probe ?

Regards,
Krishna,
diff mbox series

Patch

diff --git a/drivers/clk/qcom/gcc-msm8953.c b/drivers/clk/qcom/gcc-msm8953.c
index 3e5a8cb14d4d..20639340e8a6 100644
--- a/drivers/clk/qcom/gcc-msm8953.c
+++ b/drivers/clk/qcom/gcc-msm8953.c
@@ -3645,7 +3645,7 @@  static struct clk_branch gcc_usb30_master_clk = {
 		.hw.init = &(struct clk_init_data) {
 			.name = "gcc_usb30_master_clk",
 			.parent_hws = (const struct clk_hw*[]){
-				&usb30_master_clk_src.clkr.hw,
+				&gcc_pcnoc_usb3_axi_clk.clkr.hw,
 			},
 			.num_parents = 1,
 			.ops = &clk_branch2_ops,