Skip to contentSkip to conversation

clk: qcom: gcc-msm8953: fix stuck gcc_usb30_master_clk

PATCH In review Barnabás Czémán 4 participants 5 messages Oct 2, 2023 – Nov 22, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 clk: qcom: gcc-msm8953: fix stuck gcc_usb30_master_clk

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.

---
 drivers/clk/qcom/gcc-msm8953.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/clk/qcom/gcc-msm8953.c +1 -1
index 3e5a8cb14d4d..20639340e8a6 100644
@@ -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,
Signature · 1 line
2.42.0
#1 lore raw
Quoted 8 lines · Barnabás Czémán

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.

---

"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

maintainer
#2 lore raw

Quoting Konrad Dybcio (2023-10-06 16:50:18)

Quoted 11 lines · Barnabás Czémán
Quoted 8 lines

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.

---

"meh"

Quoted 8 lines

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.

#3 lore raw
Quoted 21 lines · Stephen Boyd

Quoting Konrad Dybcio (2023-10-06 16:50:18)

Quoted 11 lines · Barnabás Czémán
Quoted 8 lines

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.

---

"meh"

Quoted 8 lines

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?

Quoted 6 lines

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..

Quoted 8 lines

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

#4 lore raw
Quoted 50 lines · Konrad Dybcio
Quoted 21 lines · Stephen Boyd

Quoting Konrad Dybcio (2023-10-06 16:50:18)

Quoted 11 lines · Barnabás Czémán
Quoted 8 lines

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.

---

"meh"

Quoted 8 lines

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?

Quoted 6 lines

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..

Quoted 8 lines

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 {};
}
Quoted 5 lines

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,