Re: [PATCH v3 3/3] clk: qcom: Add Global Clock controller (GCC) driver for SC7180

[Taniya Das <tdas@codeaurora.org>]

Hi Stephen,

Please find my comments.

On 9/25/2019 4:42 AM, Stephen Boyd wrote:
> Quoting Taniya Das (2019-09-23 01:01:11)
>> Hi Stephen,
>>
>> Thanks for your comments.
>>
>> On 9/19/2019 3:09 AM, Stephen Boyd wrote:
>>> Quoting Taniya Das (2019-09-18 02:50:18)
>>>> diff --git a/drivers/clk/qcom/gcc-sc7180.c b/drivers/clk/qcom/gcc-sc7180.c
>>>> new file mode 100644
>>>> index 000000000000..d47865d5408f
>>>> --- /dev/null
>>>> +++ b/drivers/clk/qcom/gcc-sc7180.c
>>>> @@ -0,0 +1,2515 @@
>>>> +// SPDX-License-Identifier: GPL-2.0
>>>> +/*
>>>> + * Copyright (c) 2019, The Linux Foundation. All rights reserved.
>>>> + */
>>>> +
>>>> +#include <linux/err.h>
>>>> +#include <linux/kernel.h>
>>>> +#include <linux/module.h>
>>>> +#include <linux/of.h>
>>>> +#include <linux/of_device.h>
>>>> +#include <linux/regmap.h>
>>>
>>> include clk-provider.h
>>>
>>
>> will add this header.
>> Currently the <drivers/clk/qcom/clk-regmap.h> already includes it.
> 
> Yes but it should be included in any clk-provider drivers too.
> 
>>>> +
>>>> +/* Leave the clock ON for parent config_noc_clk to be kept enabled */
>>>> +static struct clk_branch gcc_disp_ahb_clk = {
>>>> +       .halt_reg = 0xb00c,
>>>> +       .halt_check = BRANCH_HALT,
>>>> +       .hwcg_reg = 0xb00c,
>>>> +       .hwcg_bit = 1,
>>>> +       .clkr = {
>>>> +               .enable_reg = 0xb00c,
>>>> +               .enable_mask = BIT(0),
>>>> +               .hw.init = &(struct clk_init_data){
>>>> +                       .name = "gcc_disp_ahb_clk",
>>>> +                       .flags = CLK_IS_CRITICAL,
>>>
>>> Does this assume the display is left enabled out of the bootloader? Why
>>> is this critical to system operation? Maybe it is really
>>> CLK_IGNORE_UNUSED?
>>>
>>
>> This clock is not kept enabled by bootloader. But leaving this ON for
>> clients on config noc.
> 
> Please see below comment for the other critical clk with no parent.
> 
>>
>>>> +                       .ops = &clk_branch2_ops,
>>>> +               },
>>>> +       },
>>>> +};
>>>> +
> [...]
>>>> +static struct clk_branch gcc_ufs_phy_phy_aux_clk = {
>>>> +       .halt_reg = 0x77094,
>>>> +       .halt_check = BRANCH_HALT,
>>>> +       .hwcg_reg = 0x77094,
>>>> +       .hwcg_bit = 1,
>>>> +       .clkr = {
>>>> +               .enable_reg = 0x77094,
>>>> +               .enable_mask = BIT(0),
>>>> +               .hw.init = &(struct clk_init_data){
>>>> +                       .name = "gcc_ufs_phy_phy_aux_clk",
>>>> +                       .parent_data = &(const struct clk_parent_data){
>>>> +                               .hw = &gcc_ufs_phy_phy_aux_clk_src.clkr.hw,
>>>> +                       },
>>>> +                       .num_parents = 1,
>>>> +                       .flags = CLK_SET_RATE_PARENT,
>>>> +                       .ops = &clk_branch2_ops,
>>>> +               },
>>>> +       },
>>>> +};
>>>> +
>>>> +static struct clk_branch gcc_ufs_phy_rx_symbol_0_clk = {
>>>> +       .halt_reg = 0x7701c,
>>>> +       .halt_check = BRANCH_HALT_SKIP,
>>>
>>> Again, nobody has fixed the UFS driver to not need to do this halt skip
>>> check for these clks? It's been over a year.
>>>
>>
>> The UFS_PHY_RX/TX clocks could be left enabled due to certain HW boot
>> configuration and thus during the late initcall of clk_disable there
>> could be warnings of "clock stuck ON" in the dmesg. That is the reason
>> also to use the BRANCH_HALT_SKIP flag.
> 
> Oh that's bad. Why do the clks stay on when we try to turn them off?
>

Those could be due to the configuration selected by HW and SW cannot 
override them, so traditionally we have never polled for CLK_OFF for 
these clocks.

>>
>> I would also check internally for the UFS driver fix you are referring here.
> 
> Sure. I keep asking but nothing is done :(
> 
>>
>>>> +       .clkr = {
>>>> +               .enable_reg = 0x7701c,
>>>> +               .enable_mask = BIT(0),
>>>> +               .hw.init = &(struct clk_init_data){
>>>> +                       .name = "gcc_ufs_phy_rx_symbol_0_clk",
>>>> +                       .ops = &clk_branch2_ops,
>>>> +               },
>>>> +       },
>>>> +};
>>>> +
> [...]
>>>> +
>>>> +static struct clk_branch gcc_usb3_prim_phy_pipe_clk = {
>>>> +       .halt_reg = 0xf058,
>>>> +       .halt_check = BRANCH_HALT_SKIP,
>>>
>>> Why does this need halt_skip?
>>
>> This is required as the source is external PHY, so we want to not check
>> for HALT.
> 
> This doesn't really answer my question. If the source is an external phy
> then it should be listed as a clock in the DT binding and the parent
> should be specified here. Unless something doesn't work because of that?
> 

The USB phy is managed by the USB driver and clock driver is not aware 
if USB driver models the phy as a clock. Thus we do want to keep a 
dependency on the parent and not poll for CLK_ENABLE.

>>
>>>
>>>> +       .clkr = {
>>>> +               .enable_reg = 0xf058,
>>>> +               .enable_mask = BIT(0),
>>>> +               .hw.init = &(struct clk_init_data){
>>>> +                       .name = "gcc_usb3_prim_phy_pipe_clk",
>>>> +                       .ops = &clk_branch2_ops,
>>>> +               },
>>>> +       },
>>>> +};
>>>> +
>>>> +static struct clk_branch gcc_usb_phy_cfg_ahb2phy_clk = {
>>>> +       .halt_reg = 0x6a004,
>>>> +       .halt_check = BRANCH_HALT,
>>>> +       .hwcg_reg = 0x6a004,
>>>> +       .hwcg_bit = 1,
>>>> +       .clkr = {
>>>> +               .enable_reg = 0x6a004,
>>>> +               .enable_mask = BIT(0),
>>>> +               .hw.init = &(struct clk_init_data){
>>>> +                       .name = "gcc_usb_phy_cfg_ahb2phy_clk",
>>>> +                       .ops = &clk_branch2_ops,
>>>> +               },
>>>> +       },
>>>> +};
>>>> +
>>>> +/* Leave the clock ON for parent config_noc_clk to be kept enabled */
>>>
>>> There's no parent though... So I guess this means it keeps it enabled
>>> implicitly in hardware?
>>>
>>
>> These are not left enabled, but want to leave them enabled for clients
>> on config NOC.
> 
> Sure. It just doesn't make sense to create clk structures and expose
> them in the kernel when we just want to turn the bits on and leave them
> on forever. Why not just do some register writes in probe for this
> driver? Doesn't that work just as well and use less memory?
> 

Even if I write these registers during probe, the late init check 
'clk_core_is_enabled' would return true and would be turned OFF, that is 
the reason for marking them CRITICAL.

-- 
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation.

--