diff mbox series

[V3,4/6] tty: serial: meson: The UART baud rate calculation is described using the common clock code. Also added S4 chip uart Compatible.

Message ID 20211230102110.3861-5-yu.tu@amlogic.com (mailing list archive)
State New, archived
Headers show
Series the UART driver compatible with the Amlogic Meson | expand

Commit Message

Yu Tu Dec. 30, 2021, 10:21 a.m. UTC
Using the common Clock code to describe the UART baud rate clock makes
it easier for the UART driver to be compatible with the baud rate
requirements of the UART IP on different meson chips

Signed-off-by: Yu Tu <yu.tu@amlogic.com>
---
 drivers/tty/serial/Kconfig      |   1 +
 drivers/tty/serial/meson_uart.c | 311 ++++++++++++++++++++++++++------
 2 files changed, 257 insertions(+), 55 deletions(-)

Comments

Greg KH Dec. 30, 2021, 12:27 p.m. UTC | #1
On Thu, Dec 30, 2021 at 06:21:08PM +0800, Yu Tu wrote:
> Using the common Clock code to describe the UART baud rate clock makes
> it easier for the UART driver to be compatible with the baud rate
> requirements of the UART IP on different meson chips
> 
> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
> ---

Your subject line is very odd, please fix up.

thanks,

greg k-h
Martin Blumenstingl Dec. 30, 2021, 11:13 p.m. UTC | #2
On Thu, Dec 30, 2021 at 11:21 AM Yu Tu <yu.tu@amlogic.com> wrote:
[...]
> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>  static int meson_uart_startup(struct uart_port *port)
>  {
>         u32 val;
> -       int ret = 0;
> +       int ret;
>
> -       val = readl(port->membase + AML_UART_CONTROL);
> -       val |= AML_UART_CLEAR_ERR;
> -       writel(val, port->membase + AML_UART_CONTROL);
> -       val &= ~AML_UART_CLEAR_ERR;
> -       writel(val, port->membase + AML_UART_CONTROL);
> +       meson_uart_reset(port);
I suggest splitting this into a separate patch. In general I think
it's a good idea to re-use meson_uart_reset here if possible.
However, if during testing it turns out that this doesn't work then we
can revert this single patch which updates meson_uart_startup() only -
instead of reverting the whole transition to the common clock
framework.

[...]
>  static int meson_uart_request_port(struct uart_port *port)
>  {
> +       struct meson_uart_data *private_data = port->private_data;
> +       int ret;
> +
> +       ret = clk_prepare_enable(private_data->pclk);
> +       if (ret)
> +               return ret;
> +
> +       ret = clk_prepare_enable(private_data->baud_clk);
> +       if (ret) {
> +               clk_disable_unprepare(private_data->pclk);
> +               return ret;
> +       }
This code is from my original suggestion - and I had a doubt there
which I forgot to add as a comment originally:
Can you confirm that accessing the UART controller registers works
even when "pclk" is turned off?
I am asking this because the common clock framework can access the
clocks at any time.
And I have seen SoCs which would hang when trying to access a module's
registers while the module's pclk is turned off.

[...]
>         port->fifosize = 64;
commit 27d44e05d7b85d ("tty: serial: meson: retrieve port FIFO size
from DT") [0] from May 2021 has changed this line to:
  port->fifosize = fifosize;
So your patch currently does not apply to linux-next (or even Linus'
mainline tree).

[...]
>  static const struct of_device_id meson_uart_dt_match[] = {
> -       { .compatible = "amlogic,meson6-uart" },
> -       { .compatible = "amlogic,meson8-uart" },
> -       { .compatible = "amlogic,meson8b-uart" },
> -       { .compatible = "amlogic,meson-gx-uart" },
> +       {
> +               .compatible = "amlogic,meson6-uart",
> +               .data = (void *)false,
> +       },
> +       {
> +               .compatible = "amlogic,meson8-uart",
> +               .data = (void *)false,
> +       },
> +       {
> +               .compatible = "amlogic,meson8b-uart",
> +               .data = (void *)false,
> +       },
> +       {
> +               .compatible = "amlogic,meson-gxbb-uart",
> +               .data = (void *)false,
> +       },
> +       {
> +               .compatible = "amlogic,meson-gxl-uart",
> +               .data = (void *)true,
> +       },
> +       {
> +               .compatible = "amlogic,meson-g12a-uart",
> +               .data = (void *)true,
> +       },
> +       {
> +               .compatible = "amlogic,meson-s4-uart",
> +               .data = (void *)true,
> +       },
> +       /*
> +        * deprecated, don't use anymore because it doesn't differentiate
> +        * between GXBB and GXL which have different revisions of the UART IP.
> +        */
> +       {
> +               .compatible = "amlogic,meson-gx-uart",
> +               .data = (void *)false,
> +       },
For this change I think it's also best to split it into separate
changes, similar to the dt-bindings:
1) deprecate and replace "amlogic,meson-gx-uart"
2) another one to add the S4 compatible string
3) and a third one with the big common clock framework change (adding
the .data attributes)

This is about the "Separate each logical change into a separate patch"
rule from the Linux kernel patch submission guide [1]
Also I hope that it will make it (at least a bit) easier for others to
also review this patch.


Best regards,
Martin


[0] https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/drivers/tty/serial/meson_uart.c?id=27d44e05d7b85d9d4cfe0a3c0663ea49752ece93
[1] https://www.kernel.org/doc/html/v4.17/process/submitting-patches.html#separate-your-changes
Yu Tu Dec. 31, 2021, 10:23 a.m. UTC | #3
On 2021/12/30 20:27, Greg Kroah-Hartman wrote:
> [ EXTERNAL EMAIL ]
> 
> On Thu, Dec 30, 2021 at 06:21:08PM +0800, Yu Tu wrote:
>> Using the common Clock code to describe the UART baud rate clock makes
>> it easier for the UART driver to be compatible with the baud rate
>> requirements of the UART IP on different meson chips
>>
>> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
>> ---
> 
> Your subject line is very odd, please fix up.
> 
I will correct.
> thanks,
> 
> greg k-h
>
Yu Tu Dec. 31, 2021, 11:24 a.m. UTC | #4
On 2021/12/31 7:13, Martin Blumenstingl wrote:
> [ EXTERNAL EMAIL ]
> 
> On Thu, Dec 30, 2021 at 11:21 AM Yu Tu <yu.tu@amlogic.com> wrote:
> [...]
>> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>>   static int meson_uart_startup(struct uart_port *port)
>>   {
>>          u32 val;
>> -       int ret = 0;
>> +       int ret;
>>
>> -       val = readl(port->membase + AML_UART_CONTROL);
>> -       val |= AML_UART_CLEAR_ERR;
>> -       writel(val, port->membase + AML_UART_CONTROL);
>> -       val &= ~AML_UART_CLEAR_ERR;
>> -       writel(val, port->membase + AML_UART_CONTROL);
>> +       meson_uart_reset(port);
> I suggest splitting this into a separate patch. In general I think
> it's a good idea to re-use meson_uart_reset here if possible.
> However, if during testing it turns out that this doesn't work then we
> can revert this single patch which updates meson_uart_startup() only -
> instead of reverting the whole transition to the common clock
> framework.
> 
I approve of your suggestion. I will make the next patch.

> [...]
>>   static int meson_uart_request_port(struct uart_port *port)
>>   {
>> +       struct meson_uart_data *private_data = port->private_data;
>> +       int ret;
>> +
>> +       ret = clk_prepare_enable(private_data->pclk);
>> +       if (ret)
>> +               return ret;
>> +
>> +       ret = clk_prepare_enable(private_data->baud_clk);
>> +       if (ret) {
>> +               clk_disable_unprepare(private_data->pclk);
>> +               return ret;
>> +       }
> This code is from my original suggestion - and I had a doubt there
> which I forgot to add as a comment originally:
> Can you confirm that accessing the UART controller registers works
> even when "pclk" is turned off?
> I am asking this because the common clock framework can access the
> clocks at any time.
> And I have seen SoCs which would hang when trying to access a module's
> registers while the module's pclk is turned off.
On all meson platforms, the default pclk for all UART is turned on 
during the u-boot phase. When registering uart pclk in the kernel phase, 
the CLK_IGNORE_UNUSED flag is added. So the real shutdown is when the 
standby goes down, the parent clk shuts down.
.
> 
> [...]
>>          port->fifosize = 64;
> commit 27d44e05d7b85d ("tty: serial: meson: retrieve port FIFO size
> from DT") [0] from May 2021 has changed this line to:
>    port->fifosize = fifosize;
> So your patch currently does not apply to linux-next (or even Linus'
> mainline tree).
> 
So do I need to wait for [0] patch merged before I can continue to make 
changes ?
What can I do before?
> [...]
>>   static const struct of_device_id meson_uart_dt_match[] = {
>> -       { .compatible = "amlogic,meson6-uart" },
>> -       { .compatible = "amlogic,meson8-uart" },
>> -       { .compatible = "amlogic,meson8b-uart" },
>> -       { .compatible = "amlogic,meson-gx-uart" },
>> +       {
>> +               .compatible = "amlogic,meson6-uart",
>> +               .data = (void *)false,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson8-uart",
>> +               .data = (void *)false,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson8b-uart",
>> +               .data = (void *)false,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson-gxbb-uart",
>> +               .data = (void *)false,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson-gxl-uart",
>> +               .data = (void *)true,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson-g12a-uart",
>> +               .data = (void *)true,
>> +       },
>> +       {
>> +               .compatible = "amlogic,meson-s4-uart",
>> +               .data = (void *)true,
>> +       },
>> +       /*
>> +        * deprecated, don't use anymore because it doesn't differentiate
>> +        * between GXBB and GXL which have different revisions of the UART IP.
>> +        */
>> +       {
>> +               .compatible = "amlogic,meson-gx-uart",
>> +               .data = (void *)false,
>> +       },
> For this change I think it's also best to split it into separate
> changes, similar to the dt-bindings:
> 1) deprecate and replace "amlogic,meson-gx-uart"
> 2) another one to add the S4 compatible string
> 3) and a third one with the big common clock framework change (adding
> the .data attributes)
> 
> This is about the "Separate each logical change into a separate patch"
> rule from the Linux kernel patch submission guide [1]
> Also I hope that it will make it (at least a bit) easier for others to
> also review this patch.
> 
I quite agree with your suggestion. I'll do as you suggest.
> 
> Best regards,
> Martin
> 
> 
> [0] https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/drivers/tty/serial/meson_uart.c?id=27d44e05d7b85d9d4cfe0a3c0663ea49752ece93
> [1] https://www.kernel.org/doc/html/v4.17/process/submitting-patches.html#separate-your-changes
>
Martin Blumenstingl Dec. 31, 2021, 3:32 p.m. UTC | #5
On Fri, Dec 31, 2021 at 12:24 PM Yu Tu <yu.tu@amlogic.com> wrote:
[...]
> >>   static int meson_uart_request_port(struct uart_port *port)
> >>   {
> >> +       struct meson_uart_data *private_data = port->private_data;
> >> +       int ret;
> >> +
> >> +       ret = clk_prepare_enable(private_data->pclk);
> >> +       if (ret)
> >> +               return ret;
> >> +
> >> +       ret = clk_prepare_enable(private_data->baud_clk);
> >> +       if (ret) {
> >> +               clk_disable_unprepare(private_data->pclk);
> >> +               return ret;
> >> +       }
> > This code is from my original suggestion - and I had a doubt there
> > which I forgot to add as a comment originally:
> > Can you confirm that accessing the UART controller registers works
> > even when "pclk" is turned off?
> > I am asking this because the common clock framework can access the
> > clocks at any time.
> > And I have seen SoCs which would hang when trying to access a module's
> > registers while the module's pclk is turned off.
> On all meson platforms, the default pclk for all UART is turned on
> during the u-boot phase. When registering uart pclk in the kernel phase,
> the CLK_IGNORE_UNUSED flag is added. So the real shutdown is when the
> standby goes down, the parent clk shuts down.
Interesting, thanks for sharing that u-boot turns these clocks on.
Let's say someone wanted to make u-boot save power and turn off all
UART clocks except the one for uart_AO (where we typically connect the
serial console).
In that case the pclk of uart_C (just to choose an example here) is
turned off. Would there be a problem then accessing the registers of
uart_C before clk_prepare_enable is called?

[...]
> >>          port->fifosize = 64;
> > commit 27d44e05d7b85d ("tty: serial: meson: retrieve port FIFO size
> > from DT") [0] from May 2021 has changed this line to:
> >    port->fifosize = fifosize;
> > So your patch currently does not apply to linux-next (or even Linus'
> > mainline tree).
> >
> So do I need to wait for [0] patch merged before I can continue to make
> changes ?
These changes are already merged.

> What can I do before?
You should base your changes on top of the tty.git/tty-next branch [1]
where Greg (the maintainer of this tree) will pick up the patches once
they are good (got enough Acked-by/Reviewed-by, etc.).
I suspect that you based your changes on an older or stable kernel
version (let's say 5.10). New functionality should always get into the
-next tree where various auto-build robots will compile-test the
changes and we even have Kernel CI where changes are tested on real
hardware (BayLibre even maintains Amlogic boards in their Kernel CI
labs). Let's say Amlogic updates to Linux 5.17 next year then the
patches are already included in that kernel version - instead of being
only available in Linux 5.10.


Best regards,
Martin


[1] https://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty.git/log/?h=tty-next
Yu Tu Jan. 1, 2022, 1:30 p.m. UTC | #6
Hi Martin,
     Thank you very much for your reply.

On 2021/12/31 23:32, Martin Blumenstingl wrote:
> [ EXTERNAL EMAIL ]
> 
> On Fri, Dec 31, 2021 at 12:24 PM Yu Tu <yu.tu@amlogic.com> wrote:
> [...]
>>>>    static int meson_uart_request_port(struct uart_port *port)
>>>>    {
>>>> +       struct meson_uart_data *private_data = port->private_data;
>>>> +       int ret;
>>>> +
>>>> +       ret = clk_prepare_enable(private_data->pclk);
>>>> +       if (ret)
>>>> +               return ret;
>>>> +
>>>> +       ret = clk_prepare_enable(private_data->baud_clk);
>>>> +       if (ret) {
>>>> +               clk_disable_unprepare(private_data->pclk);
>>>> +               return ret;
>>>> +       }
>>> This code is from my original suggestion - and I had a doubt there
>>> which I forgot to add as a comment originally:
>>> Can you confirm that accessing the UART controller registers works
>>> even when "pclk" is turned off?
>>> I am asking this because the common clock framework can access the
>>> clocks at any time.
>>> And I have seen SoCs which would hang when trying to access a module's
>>> registers while the module's pclk is turned off.
>> On all meson platforms, the default pclk for all UART is turned on
>> during the u-boot phase. When registering uart pclk in the kernel phase,
>> the CLK_IGNORE_UNUSED flag is added. So the real shutdown is when the
>> standby goes down, the parent clk shuts down.
> Interesting, thanks for sharing that u-boot turns these clocks on.
> Let's say someone wanted to make u-boot save power and turn off all
> UART clocks except the one for uart_AO (where we typically connect the
> serial console).
> In that case the pclk of uart_C (just to choose an example here) is
> turned off. Would there be a problem then accessing the registers of
> uart_C before clk_prepare_enable is called?
The way you describe it, it does hang. This would not be recommended on 
actual projects.

At present, AmLogic chips are older than S4 Soc, and we have no way to 
deal with this problem. We have to tell customers not to use it in this 
way。Customers rarely use it in real projects.On the S4 SOC we will use 
a clock like the UART pclk to control the shutdown using two registers, 
one safe (need to operate in EL3) and one normal (EL1). It will only be 
closed if both registers are closed. This mainly prevents misoperation.

With your experience, I'd like to know how you deal with this kind of 
problem.
> 
> [...]
>>>>           port->fifosize = 64;
>>> commit 27d44e05d7b85d ("tty: serial: meson: retrieve port FIFO size
>>> from DT") [0] from May 2021 has changed this line to:
>>>     port->fifosize = fifosize;
>>> So your patch currently does not apply to linux-next (or even Linus'
>>> mainline tree).
>>>
>> So do I need to wait for [0] patch merged before I can continue to make
>> changes ?
> These changes are already merged.
> 
>> What can I do before?
> You should base your changes on top of the tty.git/tty-next branch [1]
> where Greg (the maintainer of this tree) will pick up the patches once
> they are good (got enough Acked-by/Reviewed-by, etc.).
> I suspect that you based your changes on an older or stable kernel
> version (let's say 5.10). New functionality should always get into the
> -next tree where various auto-build robots will compile-test the
> changes and we even have Kernel CI where changes are tested on real
> hardware (BayLibre even maintains Amlogic boards in their Kernel CI
> labs). Let's say Amlogic updates to Linux 5.17 next year then the
> patches are already included in that kernel version - instead of being
> only available in Linux 5.10.
> 
I'm sorry, I did branch confirm there was a mistake, I have corrected.
> 
> Best regards,
> Martin
> 
> 
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty.git/log/?h=tty-next
>
Martin Blumenstingl Jan. 2, 2022, 7:36 p.m. UTC | #7
Hi,

On Sat, Jan 1, 2022 at 2:30 PM Yu Tu <yu.tu@amlogic.com> wrote:
[...]
> > Interesting, thanks for sharing that u-boot turns these clocks on.
> > Let's say someone wanted to make u-boot save power and turn off all
> > UART clocks except the one for uart_AO (where we typically connect the
> > serial console).
> > In that case the pclk of uart_C (just to choose an example here) is
> > turned off. Would there be a problem then accessing the registers of
> > uart_C before clk_prepare_enable is called?
> The way you describe it, it does hang. This would not be recommended on
> actual projects.
>
> At present, AmLogic chips are older than S4 Soc, and we have no way to
> deal with this problem. We have to tell customers not to use it in this
> way。Customers rarely use it in real projects.On the S4 SOC we will use
> a clock like the UART pclk to control the shutdown using two registers,
> one safe (need to operate in EL3) and one normal (EL1). It will only be
> closed if both registers are closed. This mainly prevents misoperation.
oh, interesting that there's some updates specifically with the S4 SoCs :-)

> With your experience, I'd like to know how you deal with this kind of
> problem.
Before this patch the driver simply turns on the clock from within
meson_uart_probe() (specifically it does so in
meson_uart_probe_clock()).
I think there's advanced power-saving techniques. Maybe for now we
keep it simple and just enable the clock(s) at probe time and disable
them at driver remove time. What do you think?


Best regards,
Martin
Jerome Brunet Jan. 3, 2022, 12:40 p.m. UTC | #8
On Thu 30 Dec 2021 at 18:21, Yu Tu <yu.tu@amlogic.com> wrote:

> Using the common Clock code to describe the UART baud rate clock makes
> it easier for the UART driver to be compatible with the baud rate
> requirements of the UART IP on different meson chips
>
> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
> ---
>  drivers/tty/serial/Kconfig      |   1 +
>  drivers/tty/serial/meson_uart.c | 311 ++++++++++++++++++++++++++------
>  2 files changed, 257 insertions(+), 55 deletions(-)
>
> diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
> index 780908d43557..32e238173036 100644
> --- a/drivers/tty/serial/Kconfig
> +++ b/drivers/tty/serial/Kconfig
> @@ -198,6 +198,7 @@ config SERIAL_KGDB_NMI
>  config SERIAL_MESON
>  	tristate "Meson serial port support"
>  	depends on ARCH_MESON
> +	depends on COMMON_CLK
>  	select SERIAL_CORE
>  	help
>  	  This enables the driver for the on-chip UARTs of the Amlogic
> diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
> index 99efe62a1507..07eb1f40aaaa 100644
> --- a/drivers/tty/serial/meson_uart.c
> +++ b/drivers/tty/serial/meson_uart.c
> @@ -6,6 +6,7 @@
>   */
>  
>  #include <linux/clk.h>
> +#include <linux/clk-provider.h>
>  #include <linux/console.h>
>  #include <linux/delay.h>
>  #include <linux/init.h>
> @@ -65,9 +66,7 @@
>  #define AML_UART_RECV_IRQ(c)		((c) & 0xff)
>  
>  /* AML_UART_REG5 bits */
> -#define AML_UART_BAUD_MASK		0x7fffff
>  #define AML_UART_BAUD_USE		BIT(23)
> -#define AML_UART_BAUD_XTAL		BIT(24)
>  
>  #define AML_UART_PORT_NUM		12
>  #define AML_UART_PORT_OFFSET		6
> @@ -76,6 +75,21 @@
>  #define AML_UART_POLL_USEC		5
>  #define AML_UART_TIMEOUT_USEC		10000
>  
> +struct meson_uart_data {
> +	struct uart_port	port;
> +	struct clk		*pclk;
> +	struct clk		*baud_clk;
> +	struct clk_divider	baud_div;
> +	struct clk_mux		use_xtal_mux;
> +	struct clk_mux		xtal_clk_sel_mux;
> +	struct clk_mux		xtal2_clk_sel_mux;
> +	struct clk_fixed_factor	xtal_div2;
> +	struct clk_fixed_factor	xtal_div3;
> +	struct clk_fixed_factor	clk81_div4;

Keeping all these internal elements around is not useful since they are
registered using devm_

> +	bool			no_clk81_input;

What is this ?

> +	bool			has_xtal_clk_sel;
> +};
> +
>  static struct uart_driver meson_uart_driver;
>  
>  static struct uart_port *meson_ports[AML_UART_PORT_NUM];
> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>  static int meson_uart_startup(struct uart_port *port)
>  {
>  	u32 val;
> -	int ret = 0;
> +	int ret;
>  
> -	val = readl(port->membase + AML_UART_CONTROL);
> -	val |= AML_UART_CLEAR_ERR;
> -	writel(val, port->membase + AML_UART_CONTROL);
> -	val &= ~AML_UART_CLEAR_ERR;
> -	writel(val, port->membase + AML_UART_CONTROL);
> +	meson_uart_reset(port);
>  
> +	val = readl(port->membase + AML_UART_CONTROL);
>  	val |= (AML_UART_RX_EN | AML_UART_TX_EN);
>  	writel(val, port->membase + AML_UART_CONTROL);
>  
> @@ -295,19 +306,17 @@ static int meson_uart_startup(struct uart_port *port)
>  
>  static void meson_uart_change_speed(struct uart_port *port, unsigned long baud)
>  {
> +	struct meson_uart_data *private_data = port->private_data;
>  	u32 val;
>  
>  	while (!meson_uart_tx_empty(port))
>  		cpu_relax();
>  
> -	if (port->uartclk == 24000000) {

This check shows that previous code assumed bit 24 was left untouched
Below you can see that GXBB and newer used the XTAL path while older
used the other

Your change makes this dynamic which is another "unexpected" change.
Please make the bit 24 mux RO to start with so the behavior remains unchanged
for older SoCs.

> -		val = ((port->uartclk / 3) / baud) - 1;
> -		val |= AML_UART_BAUD_XTAL;
> -	} else {
> -		val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
> -	}
> +	val = readl(port->membase + AML_UART_REG5);
>  	val |= AML_UART_BAUD_USE;
>  	writel(val, port->membase + AML_UART_REG5);
> +
> +	clk_set_rate(private_data->baud_clk, baud);
>  }
>  
>  static void meson_uart_set_termios(struct uart_port *port,
> @@ -397,11 +406,27 @@ static int meson_uart_verify_port(struct uart_port *port,
>  
>  static void meson_uart_release_port(struct uart_port *port)
>  {
> -	/* nothing to do */
> +	struct meson_uart_data *private_data = port->private_data;
> +
> +	clk_disable_unprepare(private_data->baud_clk);
> +	clk_disable_unprepare(private_data->pclk);
>  }
>  
>  static int meson_uart_request_port(struct uart_port *port)
>  {
> +	struct meson_uart_data *private_data = port->private_data;
> +	int ret;
> +
> +	ret = clk_prepare_enable(private_data->pclk);
> +	if (ret)
> +		return ret;
> +
> +	ret = clk_prepare_enable(private_data->baud_clk);
> +	if (ret) {
> +		clk_disable_unprepare(private_data->pclk);
> +		return ret;
> +	}
> +
>  	return 0;
>  }
>  
> @@ -629,56 +654,175 @@ static struct uart_driver meson_uart_driver = {
>  	.cons		= MESON_SERIAL_CONSOLE,
>  };
>  
> -static inline struct clk *meson_uart_probe_clock(struct device *dev,
> -						 const char *id)
> +static int meson_uart_register_clk(struct uart_port *port,
> +				   const char *name_suffix,
> +				   const struct clk_parent_data *parent_data,
> +				   unsigned int num_parents,
> +				   const struct clk_ops *ops,
> +				   struct clk_hw *hw)
>  {
> -	struct clk *clk = NULL;
> +	struct clk_init_data init = { };
> +	char clk_name[32];
>  	int ret;
>  
> -	clk = devm_clk_get(dev, id);
> -	if (IS_ERR(clk))
> -		return clk;
> +	snprintf(clk_name, sizeof(clk_name), "%s#%s", dev_name(port->dev),
> +		 name_suffix);
>  
> -	ret = clk_prepare_enable(clk);
> -	if (ret) {
> -		dev_err(dev, "couldn't enable clk\n");
> -		return ERR_PTR(ret);
> -	}
> +	init.name = clk_name;
> +	init.ops = ops;
> +	init.flags = CLK_SET_RATE_PARENT;
> +	init.parent_data = parent_data;
> +	init.num_parents = num_parents;
> +
> +	hw->init = &init;
>  
> -	devm_add_action_or_reset(dev,
> -			(void(*)(void *))clk_disable_unprepare,
> -			clk);
> +	ret = devm_clk_hw_register(port->dev, hw);
> +	if (ret)
> +		return dev_err_probe(port->dev, ret,
> +				     "Failed to register the '%s' clock\n",
> +				     clk_name);
>  
> -	return clk;
> +	return ret;
>  }
>  
> -static int meson_uart_probe_clocks(struct platform_device *pdev,
> -				   struct uart_port *port)
> -{
> -	struct clk *clk_xtal = NULL;
> -	struct clk *clk_pclk = NULL;
> -	struct clk *clk_baud = NULL;
> +static int meson_uart_probe_clocks(struct uart_port *port,
> +				   bool register_clk81_div4)
> +{
> +	struct meson_uart_data *private_data = port->private_data;
> +	struct clk_parent_data use_xtal_mux_parents[2] = {
> +		{ .index = -1, },
> +		{ .index = -1, },
> +	};
> +	struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
> +	struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
> +	struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
> +	struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
> +	struct clk_parent_data baud_div_parent = { };
> +	struct clk *clk_baud, *clk_xtal;
> +	int ret;
>  
> -	clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
> -	if (IS_ERR(clk_pclk))
> -		return PTR_ERR(clk_pclk);
> +	private_data->pclk = devm_clk_get(port->dev, "pclk");
> +	if (IS_ERR(private_data->pclk))
> +		return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
> +				     "Failed to get the 'pclk' clock\n");
> +
> +	clk_baud = devm_clk_get(port->dev, "baud");
> +	if (IS_ERR(clk_baud))
> +		return dev_err_probe(port->dev, PTR_ERR(clk_baud),
> +				     "Failed to get the 'baud' clock\n");
>  
> -	clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
> +	clk_xtal = devm_clk_get(port->dev, "xtal");
>  	if (IS_ERR(clk_xtal))
> -		return PTR_ERR(clk_xtal);
> +		return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
> +				     "Failed to get the 'xtal' clock\n");
> +
> +	private_data->xtal_div3.mult = 1;
> +	private_data->xtal_div3.div = 3;
> +	ret = meson_uart_register_clk(port, "xtal_div3", &xtal_div_parent,
> +				      1, &clk_fixed_factor_ops,
> +				      &private_data->xtal_div3.hw);
> +	if (ret)
> +		return ret;
>  
> -	clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
> -	if (IS_ERR(clk_baud))
> -		return PTR_ERR(clk_baud);
> +	if (register_clk81_div4) {

Clock dividers represent HW elements. The presence of an HW element cannot
dependent on the current of the input clocks. There is no way this is right

> +		private_data->clk81_div4.mult = 1;
> +		private_data->clk81_div4.div = 4;
> +		ret = meson_uart_register_clk(port, "clk81_div4",
> +					      &clk81_div_parent, 1,
> +					      &clk_fixed_factor_ops,
> +					      &private_data->clk81_div4.hw);
> +		if (ret)
> +			return ret;
> +
> +		use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;

This is only going to be used on meson8 and older. Worst case it should depend
on this compatible but I don't think this would be right here.

IMO, this shows that the UART "baud" input was actually fed by a
derivation of "fclk_div4" instead of "clk_81" on these older SoCs.

So instead of registering what I suspect to be a fake element, the
meson8 clock controller driver and DT should be fixed a this should go away.

> +	}
>  
> -	port->uartclk = clk_get_rate(clk_baud);
> +	if (private_data->has_xtal_clk_sel) {
> +		private_data->xtal_div2.mult = 1;
> +		private_data->xtal_div2.div = 2;
> +		ret = meson_uart_register_clk(port, "xtal_div2",
> +					      &xtal_div_parent, 1,
> +					      &clk_fixed_factor_ops,
> +					      &private_data->xtal_div2.hw);
> +		if (ret)
> +			return ret;
> +
> +		xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
> +		xtal_clk_sel_mux_parents[1].fw_name = "xtal";
> +
> +		private_data->xtal_clk_sel_mux.reg = port->membase + AML_UART_REG5;
> +		private_data->xtal_clk_sel_mux.mask = 0x1;
> +		private_data->xtal_clk_sel_mux.shift = 26;
> +		private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
> +		ret = meson_uart_register_clk(port, "xtal_clk_sel",
> +					      xtal_clk_sel_mux_parents,
> +					      ARRAY_SIZE(xtal_clk_sel_mux_parents),
> +					      &clk_mux_ops,
> +					      &private_data->xtal_clk_sel_mux.hw);
> +		if (ret)
> +			return ret;
> +
> +		xtal2_clk_sel_mux_parents[0].hw = &private_data->xtal_clk_sel_mux.hw;
> +		xtal2_clk_sel_mux_parents[1].hw = &private_data->xtal_div2.hw;
> +
> +		private_data->xtal2_clk_sel_mux.reg = port->membase + AML_UART_REG5;
> +		private_data->xtal2_clk_sel_mux.mask = 0x1;
> +		private_data->xtal2_clk_sel_mux.shift = 27;
> +		private_data->xtal2_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
> +		ret = meson_uart_register_clk(port, "xtal2_clk_sel",
> +					      xtal2_clk_sel_mux_parents,
> +					      ARRAY_SIZE(xtal2_clk_sel_mux_parents),
> +					      &clk_mux_ops,
> +					      &private_data->xtal2_clk_sel_mux.hw);
> +		if (ret)
> +			return ret;
> +
> +		use_xtal_mux_parents[1].hw = &private_data->xtal2_clk_sel_mux.hw;
> +	} else {
> +		use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;

Well the above is a bit over-complicated. If I summarize: 

GXBB and older used a fixed divider of 3. Bits 26 and 27 read
0 according to the documentation.

Chips after GXBB have 2 configurable divider of 2 and 3. bits 26 and 27
selects which of these dividers is used.

So the above could be replaced with a single divider covering bits 26 and 27
with the following divider table { 2, 2, 1, 3 }. The divider should use
RO ops and not have CLK_SET_RATE_PARENT. It can be used for all chips
variant like this, including the older ones.

The only information you need to carry if whether or not you want to
make this divider modifiable. This means using the dt data to store
clk_div_ops or clk_div_ro_ops pointer.

To avoid changing the behavior of the older platforms in this patch, I would
suggest to make everything use clk_div_ro_ops first, and make another
patch to use clk_div_ops if necessary.


> +	}
> +
> +	private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
> +	private_data->use_xtal_mux.mask = 0x1;
> +	private_data->use_xtal_mux.shift = 24;
> +	private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
> +	ret = meson_uart_register_clk(port, "use_xtal", use_xtal_mux_parents,
> +				      ARRAY_SIZE(use_xtal_mux_parents),
> +				      &clk_mux_ops,

Use RO ops here to start with.
You can make this writable with another patch explicitly describing the change.

> +				      &private_data->use_xtal_mux.hw);
> +	if (ret)
> +		return ret;
> +
> +	baud_div_parent.hw = &private_data->use_xtal_mux.hw;
> +
> +	private_data->baud_div.reg = port->membase + AML_UART_REG5;
> +	private_data->baud_div.shift = 0;
> +	private_data->baud_div.width = 23;
> +	private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
> +	ret = meson_uart_register_clk(port, "baud_div",
> +				      &baud_div_parent, 1,
> +				      &clk_divider_ops,
> +				      &private_data->baud_div.hw);
> +	if (ret)
> +		return ret;
> +
> +	private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
> +						     &private_data->baud_div.hw,
> +						     "baud_rate");

There is a problem with this function in CCF. It will pin the driver to
itself, making it unremovable. It is an ongoing topic. For now, just use
"hw->clk" to get the struct *clk.

> +	if (IS_ERR(private_data->baud_clk))
> +		return dev_err_probe(port->dev,

I don't think anything can defer here, so dev_err_probe() is not
necessary I think

> +				     PTR_ERR(private_data->baud_clk),
> +				     "Failed to request the 'baud_rate' clock\n");
>  
>  	return 0;
>  }
>  
>  static int meson_uart_probe(struct platform_device *pdev)
>  {
> +	struct meson_uart_data *private_data;
>  	struct resource *res_mem, *res_irq;
> +	struct clk *clk_baud, *clk_xtal;
> +	bool register_clk81_div4;
>  	struct uart_port *port;
>  	int ret = 0;
>  	int id = -1;
> @@ -711,18 +855,37 @@ static int meson_uart_probe(struct platform_device *pdev)
>  		return -EBUSY;
>  	}
>  
> -	port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port), GFP_KERNEL);
> -	if (!port)
> +	private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
> +				    GFP_KERNEL);
> +	if (!private_data)
>  		return -ENOMEM;
>  
> +	if (device_get_match_data(&pdev->dev))
> +		private_data->has_xtal_clk_sel = true;
> +
> +	private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
> +	if (IS_ERR(private_data->pclk))
> +		return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
> +				     "Failed to get the 'pclk' clock\n");
> +
> +	clk_baud = devm_clk_get(&pdev->dev, "baud");
> +	if (IS_ERR(clk_baud))
> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
> +				     "Failed to get the 'baud' clock\n");
> +
> +	clk_xtal = devm_clk_get(&pdev->dev, "xtal");
> +	if (IS_ERR(clk_xtal))
> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
> +				     "Failed to get the 'xtal' clock\n");
> +

This is second time you call devm_clk_get() on these clocks. One
instance has to go away

> +	register_clk81_div4 = clk_get_rate(clk_xtal) != clk_get_rate(clk_baud);
> +

The above is a ugly way to distinguish the meson8 (32bit) SoC family
from the rest. This definitely not the way to achieve it.

The right is the compatible data but here I think it is not
necessary. The clock input should be fixed.

> +	port = &private_data->port;
> +
>  	port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
>  	if (IS_ERR(port->membase))
>  		return PTR_ERR(port->membase);
>  
> -	ret = meson_uart_probe_clocks(pdev, port);
> -	if (ret)
> -		return ret;
> -
>  	port->iotype = UPIO_MEM;
>  	port->mapbase = res_mem->start;
>  	port->mapsize = resource_size(res_mem);
> @@ -735,6 +898,12 @@ static int meson_uart_probe(struct platform_device *pdev)
>  	port->x_char = 0;
>  	port->ops = &meson_uart_ops;
>  	port->fifosize = 64;
> +	port->uartclk = clk_get_rate(clk_baud);
> +	port->private_data = private_data;
> +
> +	ret = meson_uart_probe_clocks(port, register_clk81_div4);
> +	if (ret)
> +		return ret;
>  
>  	meson_ports[pdev->id] = port;
>  	platform_set_drvdata(pdev, port);
> @@ -761,10 +930,42 @@ static int meson_uart_remove(struct platform_device *pdev)
>  }
>  
>  static const struct of_device_id meson_uart_dt_match[] = {
> -	{ .compatible = "amlogic,meson6-uart" },
> -	{ .compatible = "amlogic,meson8-uart" },
> -	{ .compatible = "amlogic,meson8b-uart" },
> -	{ .compatible = "amlogic,meson-gx-uart" },
> +	{
> +		.compatible = "amlogic,meson6-uart",
> +		.data = (void *)false,
> +	},
> +	{
> +		.compatible = "amlogic,meson8-uart",
> +		.data = (void *)false,
> +	},
> +	{
> +		.compatible = "amlogic,meson8b-uart",
> +		.data = (void *)false,
> +	},
> +	{
> +		.compatible = "amlogic,meson-gxbb-uart",
> +		.data = (void *)false,
> +	},
> +	{
> +		.compatible = "amlogic,meson-gxl-uart",
> +		.data = (void *)true,
> +	},
> +	{
> +		.compatible = "amlogic,meson-g12a-uart",
> +		.data = (void *)true,
> +	},
> +	{
> +		.compatible = "amlogic,meson-s4-uart",
> +		.data = (void *)true,
> +	},
> +	/*
> +	 * deprecated, don't use anymore because it doesn't differentiate
> +	 * between GXBB and GXL which have different revisions of the UART IP.
> +	 */
> +	{
> +		.compatible = "amlogic,meson-gx-uart",
> +		.data = (void *)false,
> +	},
>  	{ /* sentinel */ },
>  };
>  MODULE_DEVICE_TABLE(of, meson_uart_dt_match);
Jerome Brunet Jan. 3, 2022, 1:50 p.m. UTC | #9
On Sat 01 Jan 2022 at 21:30, Yu Tu <yu.tu@amlogic.com> wrote:

> Hi Martin,
>     Thank you very much for your reply.
>
> On 2021/12/31 23:32, Martin Blumenstingl wrote:
>> [ EXTERNAL EMAIL ]
>> On Fri, Dec 31, 2021 at 12:24 PM Yu Tu <yu.tu@amlogic.com> wrote:
>> [...]
>>>>>    static int meson_uart_request_port(struct uart_port *port)
>>>>>    {
>>>>> +       struct meson_uart_data *private_data = port->private_data;
>>>>> +       int ret;
>>>>> +
>>>>> +       ret = clk_prepare_enable(private_data->pclk);
>>>>> +       if (ret)
>>>>> +               return ret;
>>>>> +
>>>>> +       ret = clk_prepare_enable(private_data->baud_clk);
>>>>> +       if (ret) {
>>>>> +               clk_disable_unprepare(private_data->pclk);
>>>>> +               return ret;
>>>>> +       }
>>>> This code is from my original suggestion - and I had a doubt there
>>>> which I forgot to add as a comment originally:
>>>> Can you confirm that accessing the UART controller registers works
>>>> even when "pclk" is turned off?
>>>> I am asking this because the common clock framework can access the
>>>> clocks at any time.
>>>> And I have seen SoCs which would hang when trying to access a module's
>>>> registers while the module's pclk is turned off.
>>> On all meson platforms, the default pclk for all UART is turned on
>>> during the u-boot phase. When registering uart pclk in the kernel phase,
>>> the CLK_IGNORE_UNUSED flag is added. So the real shutdown is when the
>>> standby goes down, the parent clk shuts down.
>> Interesting, thanks for sharing that u-boot turns these clocks on.
>> Let's say someone wanted to make u-boot save power and turn off all
>> UART clocks except the one for uart_AO (where we typically connect the
>> serial console).
>> In that case the pclk of uart_C (just to choose an example here) is
>> turned off. Would there be a problem then accessing the registers of
>> uart_C before clk_prepare_enable is called?
> The way you describe it, it does hang. This would not be recommended on
> actual projects.
>
> At present, AmLogic chips are older than S4 Soc, and we have no way to deal
> with this problem. We have to tell customers not to use it in this way。
> Customers rarely use it in real projects.On the S4 SOC we will use a clock
> like the UART pclk to control the shutdown using two registers, one safe
> (need to operate in EL3) and one normal (EL1). It will only be closed if
> both registers are closed. This mainly prevents misoperation.
>
> With your experience, I'd like to know how you deal with this kind of
> problem.

Relying on the CLK_IGNORE_UNUSED and the boot loader is unsafe.
As Martin is suggesting, the driver must ensure pclk is on before
touching any register.

This was done in the probe before in probe before your change which
seems safe but maybe not optimal.

Again, If you wish to optimize that, please do so in another dedicated
change so we can discuss it and make sure it is still safe

>> [...]
>>>>>           port->fifosize = 64;
>>>> commit 27d44e05d7b85d ("tty: serial: meson: retrieve port FIFO size
>>>> from DT") [0] from May 2021 has changed this line to:
>>>>     port->fifosize = fifosize;
>>>> So your patch currently does not apply to linux-next (or even Linus'
>>>> mainline tree).
>>>>
>>> So do I need to wait for [0] patch merged before I can continue to make
>>> changes ?
>> These changes are already merged.
>> 
>>> What can I do before?
>> You should base your changes on top of the tty.git/tty-next branch [1]
>> where Greg (the maintainer of this tree) will pick up the patches once
>> they are good (got enough Acked-by/Reviewed-by, etc.).
>> I suspect that you based your changes on an older or stable kernel
>> version (let's say 5.10). New functionality should always get into the
>> -next tree where various auto-build robots will compile-test the
>> changes and we even have Kernel CI where changes are tested on real
>> hardware (BayLibre even maintains Amlogic boards in their Kernel CI
>> labs). Let's say Amlogic updates to Linux 5.17 next year then the
>> patches are already included in that kernel version - instead of being
>> only available in Linux 5.10.
>> 
> I'm sorry, I did branch confirm there was a mistake, I have corrected.
>> Best regards,
>> Martin
>> 
>> [1] https://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty.git/log/?h=tty-next
>>
Yu Tu Jan. 4, 2022, 8:20 a.m. UTC | #10
Hi Martin,
	Thank you very much for your reply.
On 2022/1/3 3:36, Martin Blumenstingl wrote:
> [ EXTERNAL EMAIL ]
> 
> Hi,
> 
> On Sat, Jan 1, 2022 at 2:30 PM Yu Tu <yu.tu@amlogic.com> wrote:
> [...]
>>> Interesting, thanks for sharing that u-boot turns these clocks on.
>>> Let's say someone wanted to make u-boot save power and turn off all
>>> UART clocks except the one for uart_AO (where we typically connect the
>>> serial console).
>>> In that case the pclk of uart_C (just to choose an example here) is
>>> turned off. Would there be a problem then accessing the registers of
>>> uart_C before clk_prepare_enable is called?
>> The way you describe it, it does hang. This would not be recommended on
>> actual projects.
>>
>> At present, AmLogic chips are older than S4 Soc, and we have no way to
>> deal with this problem. We have to tell customers not to use it in this
>> way。Customers rarely use it in real projects.On the S4 SOC we will use
>> a clock like the UART pclk to control the shutdown using two registers,
>> one safe (need to operate in EL3) and one normal (EL1). It will only be
>> closed if both registers are closed. This mainly prevents misoperation.
> oh, interesting that there's some updates specifically with the S4 SoCs :-)
> 
>> With your experience, I'd like to know how you deal with this kind of
>> problem.
> Before this patch the driver simply turns on the clock from within
> meson_uart_probe() (specifically it does so in
> meson_uart_probe_clock()).
> I think there's advanced power-saving techniques. Maybe for now we
> keep it simple and just enable the clock(s) at probe time and disable
> them at driver remove time. What do you think?
> 
I agree with you.
> 
> Best regards,
> Martin
>
Yu Tu Jan. 4, 2022, 9:57 a.m. UTC | #11
Hi Jerome,
	Thank you very much for your reply.

On 2022/1/3 20:40, Jerome Brunet wrote:
> [ EXTERNAL EMAIL ]
> 
> 
> On Thu 30 Dec 2021 at 18:21, Yu Tu <yu.tu@amlogic.com> wrote:
> 
>> Using the common Clock code to describe the UART baud rate clock makes
>> it easier for the UART driver to be compatible with the baud rate
>> requirements of the UART IP on different meson chips
>>
>> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
>> ---
>>   drivers/tty/serial/Kconfig      |   1 +
>>   drivers/tty/serial/meson_uart.c | 311 ++++++++++++++++++++++++++------
>>   2 files changed, 257 insertions(+), 55 deletions(-)
>>
>> diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
>> index 780908d43557..32e238173036 100644
>> --- a/drivers/tty/serial/Kconfig
>> +++ b/drivers/tty/serial/Kconfig
>> @@ -198,6 +198,7 @@ config SERIAL_KGDB_NMI
>>   config SERIAL_MESON
>>   	tristate "Meson serial port support"
>>   	depends on ARCH_MESON
>> +	depends on COMMON_CLK
>>   	select SERIAL_CORE
>>   	help
>>   	  This enables the driver for the on-chip UARTs of the Amlogic
>> diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
>> index 99efe62a1507..07eb1f40aaaa 100644
>> --- a/drivers/tty/serial/meson_uart.c
>> +++ b/drivers/tty/serial/meson_uart.c
>> @@ -6,6 +6,7 @@
>>    */
>>   
>>   #include <linux/clk.h>
>> +#include <linux/clk-provider.h>
>>   #include <linux/console.h>
>>   #include <linux/delay.h>
>>   #include <linux/init.h>
>> @@ -65,9 +66,7 @@
>>   #define AML_UART_RECV_IRQ(c)		((c) & 0xff)
>>   
>>   /* AML_UART_REG5 bits */
>> -#define AML_UART_BAUD_MASK		0x7fffff
>>   #define AML_UART_BAUD_USE		BIT(23)
>> -#define AML_UART_BAUD_XTAL		BIT(24)
>>   
>>   #define AML_UART_PORT_NUM		12
>>   #define AML_UART_PORT_OFFSET		6
>> @@ -76,6 +75,21 @@
>>   #define AML_UART_POLL_USEC		5
>>   #define AML_UART_TIMEOUT_USEC		10000
>>   
>> +struct meson_uart_data {
>> +	struct uart_port	port;
>> +	struct clk		*pclk;
>> +	struct clk		*baud_clk;
>> +	struct clk_divider	baud_div;
>> +	struct clk_mux		use_xtal_mux;
>> +	struct clk_mux		xtal_clk_sel_mux;
>> +	struct clk_mux		xtal2_clk_sel_mux;
>> +	struct clk_fixed_factor	xtal_div2;
>> +	struct clk_fixed_factor	xtal_div3;
>> +	struct clk_fixed_factor	clk81_div4;
> 
> Keeping all these internal elements around is not useful since they are
> registered using devm_
> 
I'm sorry. I don't know what you mean. That's exactly what you said, but 
what's wrong, I don't understand. Do you have any better suggestions, 
please specify specific points, preferably give examples.
>> +	bool			no_clk81_input;
> 
> What is this ?
> 
To distinguish between clK81 and XTAL.
>> +	bool			has_xtal_clk_sel;
>> +};
>> +
>>   static struct uart_driver meson_uart_driver;
>>   
>>   static struct uart_port *meson_ports[AML_UART_PORT_NUM];
>> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>>   static int meson_uart_startup(struct uart_port *port)
>>   {
>>   	u32 val;
>> -	int ret = 0;
>> +	int ret;
>>   
>> -	val = readl(port->membase + AML_UART_CONTROL);
>> -	val |= AML_UART_CLEAR_ERR;
>> -	writel(val, port->membase + AML_UART_CONTROL);
>> -	val &= ~AML_UART_CLEAR_ERR;
>> -	writel(val, port->membase + AML_UART_CONTROL);
>> +	meson_uart_reset(port);
>>   
>> +	val = readl(port->membase + AML_UART_CONTROL);
>>   	val |= (AML_UART_RX_EN | AML_UART_TX_EN);
>>   	writel(val, port->membase + AML_UART_CONTROL);
>>   
>> @@ -295,19 +306,17 @@ static int meson_uart_startup(struct uart_port *port)
>>   
>>   static void meson_uart_change_speed(struct uart_port *port, unsigned long baud)
>>   {
>> +	struct meson_uart_data *private_data = port->private_data;
>>   	u32 val;
>>   
>>   	while (!meson_uart_tx_empty(port))
>>   		cpu_relax();
>>   
>> -	if (port->uartclk == 24000000) {
> 
> This check shows that previous code assumed bit 24 was left untouched
> Below you can see that GXBB and newer used the XTAL path while older
> used the other
> 
> Your change makes this dynamic which is another "unexpected" change.
> Please make the bit 24 mux RO to start with so the behavior remains unchanged
> for older SoCs.
I agree with what you say, and I will correct it.
> 
>> -		val = ((port->uartclk / 3) / baud) - 1;
>> -		val |= AML_UART_BAUD_XTAL;
>> -	} else {
>> -		val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
>> -	}
>> +	val = readl(port->membase + AML_UART_REG5);
>>   	val |= AML_UART_BAUD_USE;
>>   	writel(val, port->membase + AML_UART_REG5);
>> +
>> +	clk_set_rate(private_data->baud_clk, baud);
>>   }
>>   
>>   static void meson_uart_set_termios(struct uart_port *port,
>> @@ -397,11 +406,27 @@ static int meson_uart_verify_port(struct uart_port *port,
>>   
>>   static void meson_uart_release_port(struct uart_port *port)
>>   {
>> -	/* nothing to do */
>> +	struct meson_uart_data *private_data = port->private_data;
>> +
>> +	clk_disable_unprepare(private_data->baud_clk);
>> +	clk_disable_unprepare(private_data->pclk);
>>   }
>>   
>>   static int meson_uart_request_port(struct uart_port *port)
>>   {
>> +	struct meson_uart_data *private_data = port->private_data;
>> +	int ret;
>> +
>> +	ret = clk_prepare_enable(private_data->pclk);
>> +	if (ret)
>> +		return ret;
>> +
>> +	ret = clk_prepare_enable(private_data->baud_clk);
>> +	if (ret) {
>> +		clk_disable_unprepare(private_data->pclk);
>> +		return ret;
>> +	}
>> +
>>   	return 0;
>>   }
>>   
>> @@ -629,56 +654,175 @@ static struct uart_driver meson_uart_driver = {
>>   	.cons		= MESON_SERIAL_CONSOLE,
>>   };
>>   
>> -static inline struct clk *meson_uart_probe_clock(struct device *dev,
>> -						 const char *id)
>> +static int meson_uart_register_clk(struct uart_port *port,
>> +				   const char *name_suffix,
>> +				   const struct clk_parent_data *parent_data,
>> +				   unsigned int num_parents,
>> +				   const struct clk_ops *ops,
>> +				   struct clk_hw *hw)
>>   {
>> -	struct clk *clk = NULL;
>> +	struct clk_init_data init = { };
>> +	char clk_name[32];
>>   	int ret;
>>   
>> -	clk = devm_clk_get(dev, id);
>> -	if (IS_ERR(clk))
>> -		return clk;
>> +	snprintf(clk_name, sizeof(clk_name), "%s#%s", dev_name(port->dev),
>> +		 name_suffix);
>>   
>> -	ret = clk_prepare_enable(clk);
>> -	if (ret) {
>> -		dev_err(dev, "couldn't enable clk\n");
>> -		return ERR_PTR(ret);
>> -	}
>> +	init.name = clk_name;
>> +	init.ops = ops;
>> +	init.flags = CLK_SET_RATE_PARENT;
>> +	init.parent_data = parent_data;
>> +	init.num_parents = num_parents;
>> +
>> +	hw->init = &init;
>>   
>> -	devm_add_action_or_reset(dev,
>> -			(void(*)(void *))clk_disable_unprepare,
>> -			clk);
>> +	ret = devm_clk_hw_register(port->dev, hw);
>> +	if (ret)
>> +		return dev_err_probe(port->dev, ret,
>> +				     "Failed to register the '%s' clock\n",
>> +				     clk_name);
>>   
>> -	return clk;
>> +	return ret;
>>   }
>>   
>> -static int meson_uart_probe_clocks(struct platform_device *pdev,
>> -				   struct uart_port *port)
>> -{
>> -	struct clk *clk_xtal = NULL;
>> -	struct clk *clk_pclk = NULL;
>> -	struct clk *clk_baud = NULL;
>> +static int meson_uart_probe_clocks(struct uart_port *port,
>> +				   bool register_clk81_div4)
>> +{
>> +	struct meson_uart_data *private_data = port->private_data;
>> +	struct clk_parent_data use_xtal_mux_parents[2] = {
>> +		{ .index = -1, },
>> +		{ .index = -1, },
>> +	};
>> +	struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
>> +	struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
>> +	struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
>> +	struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
>> +	struct clk_parent_data baud_div_parent = { };
>> +	struct clk *clk_baud, *clk_xtal;
>> +	int ret;
>>   
>> -	clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
>> -	if (IS_ERR(clk_pclk))
>> -		return PTR_ERR(clk_pclk);
>> +	private_data->pclk = devm_clk_get(port->dev, "pclk");
>> +	if (IS_ERR(private_data->pclk))
>> +		return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
>> +				     "Failed to get the 'pclk' clock\n");
>> +
>> +	clk_baud = devm_clk_get(port->dev, "baud");
>> +	if (IS_ERR(clk_baud))
>> +		return dev_err_probe(port->dev, PTR_ERR(clk_baud),
>> +				     "Failed to get the 'baud' clock\n");
>>   
>> -	clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
>> +	clk_xtal = devm_clk_get(port->dev, "xtal");
>>   	if (IS_ERR(clk_xtal))
>> -		return PTR_ERR(clk_xtal);
>> +		return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
>> +				     "Failed to get the 'xtal' clock\n");
>> +
>> +	private_data->xtal_div3.mult = 1;
>> +	private_data->xtal_div3.div = 3;
>> +	ret = meson_uart_register_clk(port, "xtal_div3", &xtal_div_parent,
>> +				      1, &clk_fixed_factor_ops,
>> +				      &private_data->xtal_div3.hw);
>> +	if (ret)
>> +		return ret;
>>   
>> -	clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
>> -	if (IS_ERR(clk_baud))
>> -		return PTR_ERR(clk_baud);
>> +	if (register_clk81_div4) {
> 
> Clock dividers represent HW elements. The presence of an HW element cannot
> dependent on the current of the input clocks. There is no way this is right
> 
>> +		private_data->clk81_div4.mult = 1;
>> +		private_data->clk81_div4.div = 4;
>> +		ret = meson_uart_register_clk(port, "clk81_div4",
>> +					      &clk81_div_parent, 1,
>> +					      &clk_fixed_factor_ops,
>> +					      &private_data->clk81_div4.hw);
>> +		if (ret)
>> +			return ret;
>> +
>> +		use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;
> 
> This is only going to be used on meson8 and older. Worst case it should depend
> on this compatible but I don't think this would be right here.
> 
> IMO, this shows that the UART "baud" input was actually fed by a
> derivation of "fclk_div4" instead of "clk_81" on these older SoCs.
> 
> So instead of registering what I suspect to be a fake element, the
> meson8 clock controller driver and DT should be fixed a this should go away.
> 
Virtually all UART controllers are supported, but they are not used 
after Meson8. So the description is more reasonable in UART driver.
>> +	}
>>   
>> -	port->uartclk = clk_get_rate(clk_baud);
>> +	if (private_data->has_xtal_clk_sel) {
>> +		private_data->xtal_div2.mult = 1;
>> +		private_data->xtal_div2.div = 2;
>> +		ret = meson_uart_register_clk(port, "xtal_div2",
>> +					      &xtal_div_parent, 1,
>> +					      &clk_fixed_factor_ops,
>> +					      &private_data->xtal_div2.hw);
>> +		if (ret)
>> +			return ret;
>> +
>> +		xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
>> +		xtal_clk_sel_mux_parents[1].fw_name = "xtal";
>> +
>> +		private_data->xtal_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>> +		private_data->xtal_clk_sel_mux.mask = 0x1;
>> +		private_data->xtal_clk_sel_mux.shift = 26;
>> +		private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>> +		ret = meson_uart_register_clk(port, "xtal_clk_sel",
>> +					      xtal_clk_sel_mux_parents,
>> +					      ARRAY_SIZE(xtal_clk_sel_mux_parents),
>> +					      &clk_mux_ops,
>> +					      &private_data->xtal_clk_sel_mux.hw);
>> +		if (ret)
>> +			return ret;
>> +
>> +		xtal2_clk_sel_mux_parents[0].hw = &private_data->xtal_clk_sel_mux.hw;
>> +		xtal2_clk_sel_mux_parents[1].hw = &private_data->xtal_div2.hw;
>> +
>> +		private_data->xtal2_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>> +		private_data->xtal2_clk_sel_mux.mask = 0x1;
>> +		private_data->xtal2_clk_sel_mux.shift = 27;
>> +		private_data->xtal2_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>> +		ret = meson_uart_register_clk(port, "xtal2_clk_sel",
>> +					      xtal2_clk_sel_mux_parents,
>> +					      ARRAY_SIZE(xtal2_clk_sel_mux_parents),
>> +					      &clk_mux_ops,
>> +					      &private_data->xtal2_clk_sel_mux.hw);
>> +		if (ret)
>> +			return ret;
>> +
>> +		use_xtal_mux_parents[1].hw = &private_data->xtal2_clk_sel_mux.hw;
>> +	} else {
>> +		use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;
> 
> Well the above is a bit over-complicated. If I summarize:
> 
> GXBB and older used a fixed divider of 3. Bits 26 and 27 read
> 0 according to the documentation.
> 
> Chips after GXBB have 2 configurable divider of 2 and 3. bits 26 and 27
> selects which of these dividers is used.
> 
> So the above could be replaced with a single divider covering bits 26 and 27
> with the following divider table { 2, 2, 1, 3 }. The divider should use
> RO ops and not have CLK_SET_RATE_PARENT. It can be used for all chips
> variant like this, including the older ones.
> 
As you say, the way you say it is complicated. I think you're 
complicating things.I don't understand what you're trying to achieve.
Can you be more specific?

> The only information you need to carry if whether or not you want to
> make this divider modifiable. This means using the dt data to store
> clk_div_ops or clk_div_ro_ops pointer.
> 
> To avoid changing the behavior of the older platforms in this patch, I would
> suggest to make everything use clk_div_ro_ops first, and make another
> patch to use clk_div_ops if necessary.
> 
> 
>> +	}
>> +
>> +	private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
>> +	private_data->use_xtal_mux.mask = 0x1;
>> +	private_data->use_xtal_mux.shift = 24;
>> +	private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
>> +	ret = meson_uart_register_clk(port, "use_xtal", use_xtal_mux_parents,
>> +				      ARRAY_SIZE(use_xtal_mux_parents),
>> +				      &clk_mux_ops,
> 
> Use RO ops here to start with.
> You can make this writable with another patch explicitly describing the change.
I agree with what you say, and I will correct it.
> 
>> +				      &private_data->use_xtal_mux.hw);
>> +	if (ret)
>> +		return ret;
>> +
>> +	baud_div_parent.hw = &private_data->use_xtal_mux.hw;
>> +
>> +	private_data->baud_div.reg = port->membase + AML_UART_REG5;
>> +	private_data->baud_div.shift = 0;
>> +	private_data->baud_div.width = 23;
>> +	private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
>> +	ret = meson_uart_register_clk(port, "baud_div",
>> +				      &baud_div_parent, 1,
>> +				      &clk_divider_ops,
>> +				      &private_data->baud_div.hw);
>> +	if (ret)
>> +		return ret;
>> +
>> +	private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
>> +						     &private_data->baud_div.hw,
>> +						     "baud_rate");
> 
> There is a problem with this function in CCF. It will pin the driver to
> itself, making it unremovable. It is an ongoing topic. For now, just use
> "hw->clk" to get the struct *clk.
> 
>> +	if (IS_ERR(private_data->baud_clk))
>> +		return dev_err_probe(port->dev,
> 
> I don't think anything can defer here, so dev_err_probe() is not
> necessary I think
> 
I agree with what you say, and I will correct it.
>> +				     PTR_ERR(private_data->baud_clk),
>> +				     "Failed to request the 'baud_rate' clock\n");
>>   
>>   	return 0;
>>   }
>>   
>>   static int meson_uart_probe(struct platform_device *pdev)
>>   {
>> +	struct meson_uart_data *private_data;
>>   	struct resource *res_mem, *res_irq;
>> +	struct clk *clk_baud, *clk_xtal;
>> +	bool register_clk81_div4;
>>   	struct uart_port *port;
>>   	int ret = 0;
>>   	int id = -1;
>> @@ -711,18 +855,37 @@ static int meson_uart_probe(struct platform_device *pdev)
>>   		return -EBUSY;
>>   	}
>>   
>> -	port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port), GFP_KERNEL);
>> -	if (!port)
>> +	private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
>> +				    GFP_KERNEL);
>> +	if (!private_data)
>>   		return -ENOMEM;
>>   
>> +	if (device_get_match_data(&pdev->dev))
>> +		private_data->has_xtal_clk_sel = true;
>> +
>> +	private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
>> +	if (IS_ERR(private_data->pclk))
>> +		return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
>> +				     "Failed to get the 'pclk' clock\n");
>> +
>> +	clk_baud = devm_clk_get(&pdev->dev, "baud");
>> +	if (IS_ERR(clk_baud))
>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
>> +				     "Failed to get the 'baud' clock\n");
>> +
>> +	clk_xtal = devm_clk_get(&pdev->dev, "xtal");
>> +	if (IS_ERR(clk_xtal))
>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
>> +				     "Failed to get the 'xtal' clock\n");
>> +
> 
> This is second time you call devm_clk_get() on these clocks. One
> instance has to go away
> 
I agree with what you say, and I will correct it.
>> +	register_clk81_div4 = clk_get_rate(clk_xtal) != clk_get_rate(clk_baud);
>> +
> 
> The above is a ugly way to distinguish the meson8 (32bit) SoC family
> from the rest. This definitely not the way to achieve it.
> 
> The right is the compatible data but here I think it is not
> necessary. The clock input should be fixed.
> 
>> +	port = &private_data->port;
>> +
>>   	port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
>>   	if (IS_ERR(port->membase))
>>   		return PTR_ERR(port->membase);
>>   
>> -	ret = meson_uart_probe_clocks(pdev, port);
>> -	if (ret)
>> -		return ret;
>> -
>>   	port->iotype = UPIO_MEM;
>>   	port->mapbase = res_mem->start;
>>   	port->mapsize = resource_size(res_mem);
>> @@ -735,6 +898,12 @@ static int meson_uart_probe(struct platform_device *pdev)
>>   	port->x_char = 0;
>>   	port->ops = &meson_uart_ops;
>>   	port->fifosize = 64;
>> +	port->uartclk = clk_get_rate(clk_baud);
>> +	port->private_data = private_data;
>> +
>> +	ret = meson_uart_probe_clocks(port, register_clk81_div4);
>> +	if (ret)
>> +		return ret;
>>   
>>   	meson_ports[pdev->id] = port;
>>   	platform_set_drvdata(pdev, port);
>> @@ -761,10 +930,42 @@ static int meson_uart_remove(struct platform_device *pdev)
>>   }
>>   
>>   static const struct of_device_id meson_uart_dt_match[] = {
>> -	{ .compatible = "amlogic,meson6-uart" },
>> -	{ .compatible = "amlogic,meson8-uart" },
>> -	{ .compatible = "amlogic,meson8b-uart" },
>> -	{ .compatible = "amlogic,meson-gx-uart" },
>> +	{
>> +		.compatible = "amlogic,meson6-uart",
>> +		.data = (void *)false,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson8-uart",
>> +		.data = (void *)false,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson8b-uart",
>> +		.data = (void *)false,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson-gxbb-uart",
>> +		.data = (void *)false,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson-gxl-uart",
>> +		.data = (void *)true,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson-g12a-uart",
>> +		.data = (void *)true,
>> +	},
>> +	{
>> +		.compatible = "amlogic,meson-s4-uart",
>> +		.data = (void *)true,
>> +	},
>> +	/*
>> +	 * deprecated, don't use anymore because it doesn't differentiate
>> +	 * between GXBB and GXL which have different revisions of the UART IP.
>> +	 */
>> +	{
>> +		.compatible = "amlogic,meson-gx-uart",
>> +		.data = (void *)false,
>> +	},
>>   	{ /* sentinel */ },
>>   };
>>   MODULE_DEVICE_TABLE(of, meson_uart_dt_match);
>
Jerome Brunet Jan. 4, 2022, 10:36 a.m. UTC | #12
On Tue 04 Jan 2022 at 17:57, Yu Tu <yu.tu@amlogic.com> wrote:

> Hi Jerome,
> 	Thank you very much for your reply.
>
> On 2022/1/3 20:40, Jerome Brunet wrote:
>> [ EXTERNAL EMAIL ]
>> 
>> On Thu 30 Dec 2021 at 18:21, Yu Tu <yu.tu@amlogic.com> wrote:
>> 
>>> Using the common Clock code to describe the UART baud rate clock makes
>>> it easier for the UART driver to be compatible with the baud rate
>>> requirements of the UART IP on different meson chips
>>>
>>> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
>>> ---
>>>   drivers/tty/serial/Kconfig      |   1 +
>>>   drivers/tty/serial/meson_uart.c | 311 ++++++++++++++++++++++++++------
>>>   2 files changed, 257 insertions(+), 55 deletions(-)
>>>
>>> diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
>>> index 780908d43557..32e238173036 100644
>>> --- a/drivers/tty/serial/Kconfig
>>> +++ b/drivers/tty/serial/Kconfig
>>> @@ -198,6 +198,7 @@ config SERIAL_KGDB_NMI
>>>   config SERIAL_MESON
>>>   	tristate "Meson serial port support"
>>>   	depends on ARCH_MESON
>>> +	depends on COMMON_CLK
>>>   	select SERIAL_CORE
>>>   	help
>>>   	  This enables the driver for the on-chip UARTs of the Amlogic
>>> diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
>>> index 99efe62a1507..07eb1f40aaaa 100644
>>> --- a/drivers/tty/serial/meson_uart.c
>>> +++ b/drivers/tty/serial/meson_uart.c
>>> @@ -6,6 +6,7 @@
>>>    */
>>>     #include <linux/clk.h>
>>> +#include <linux/clk-provider.h>
>>>   #include <linux/console.h>
>>>   #include <linux/delay.h>
>>>   #include <linux/init.h>
>>> @@ -65,9 +66,7 @@
>>>   #define AML_UART_RECV_IRQ(c)		((c) & 0xff)
>>>     /* AML_UART_REG5 bits */
>>> -#define AML_UART_BAUD_MASK		0x7fffff
>>>   #define AML_UART_BAUD_USE		BIT(23)
>>> -#define AML_UART_BAUD_XTAL		BIT(24)
>>>     #define AML_UART_PORT_NUM		12
>>>   #define AML_UART_PORT_OFFSET		6
>>> @@ -76,6 +75,21 @@
>>>   #define AML_UART_POLL_USEC		5
>>>   #define AML_UART_TIMEOUT_USEC		10000
>>>   +struct meson_uart_data {
>>> +	struct uart_port	port;
>>> +	struct clk		*pclk;
>>> +	struct clk		*baud_clk;
>>> +	struct clk_divider	baud_div;
>>> +	struct clk_mux		use_xtal_mux;
>>> +	struct clk_mux		xtal_clk_sel_mux;
>>> +	struct clk_mux		xtal2_clk_sel_mux;
>>> +	struct clk_fixed_factor	xtal_div2;
>>> +	struct clk_fixed_factor	xtal_div3;
>>> +	struct clk_fixed_factor	clk81_div4;
>> Keeping all these internal elements around is not useful since they are
>> registered using devm_
>> 
> I'm sorry. I don't know what you mean. That's exactly what you said, but
> what's wrong, I don't understand. Do you have any better suggestions, 
> please specify specific points, preferably give examples.

I'm saying that you don't need to keep reference to the internal
elements of the clock tree you have registered since devm_ will take
care of the removal later on. IOW, Once they are registered, the pointer
is never used again so you don't need it in the private data.


>>> +	bool			no_clk81_input;
>> What is this ?
>> 
> To distinguish between clK81 and XTAL.

... Yet, it is not used

>>> +	bool			has_xtal_clk_sel;
>>> +};
>>> +
>>>   static struct uart_driver meson_uart_driver;
>>>     static struct uart_port *meson_ports[AML_UART_PORT_NUM];
>>> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>>>   static int meson_uart_startup(struct uart_port *port)
>>>   {
>>>   	u32 val;
>>> -	int ret = 0;
>>> +	int ret;
>>>   -	val = readl(port->membase + AML_UART_CONTROL);
>>> -	val |= AML_UART_CLEAR_ERR;
>>> -	writel(val, port->membase + AML_UART_CONTROL);
>>> -	val &= ~AML_UART_CLEAR_ERR;
>>> -	writel(val, port->membase + AML_UART_CONTROL);
>>> +	meson_uart_reset(port);
>>>   +	val = readl(port->membase + AML_UART_CONTROL);
>>>   	val |= (AML_UART_RX_EN | AML_UART_TX_EN);
>>>   	writel(val, port->membase + AML_UART_CONTROL);
>>>   @@ -295,19 +306,17 @@ static int meson_uart_startup(struct uart_port
>>> *port)
>>>     static void meson_uart_change_speed(struct uart_port *port, unsigned
>>> long baud)
>>>   {
>>> +	struct meson_uart_data *private_data = port->private_data;
>>>   	u32 val;
>>>     	while (!meson_uart_tx_empty(port))
>>>   		cpu_relax();
>>>   -	if (port->uartclk == 24000000) {
>> This check shows that previous code assumed bit 24 was left untouched
>> Below you can see that GXBB and newer used the XTAL path while older
>> used the other
>> Your change makes this dynamic which is another "unexpected" change.
>> Please make the bit 24 mux RO to start with so the behavior remains unchanged
>> for older SoCs.
> I agree with what you say, and I will correct it.
>> 
>>> -		val = ((port->uartclk / 3) / baud) - 1;
>>> -		val |= AML_UART_BAUD_XTAL;
>>> -	} else {
>>> -		val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
>>> -	}
>>> +	val = readl(port->membase + AML_UART_REG5);
>>>   	val |= AML_UART_BAUD_USE;
>>>   	writel(val, port->membase + AML_UART_REG5);
>>> +
>>> +	clk_set_rate(private_data->baud_clk, baud);
>>>   }
>>>     static void meson_uart_set_termios(struct uart_port *port,
>>> @@ -397,11 +406,27 @@ static int meson_uart_verify_port(struct uart_port *port,
>>>     static void meson_uart_release_port(struct uart_port *port)
>>>   {
>>> -	/* nothing to do */
>>> +	struct meson_uart_data *private_data = port->private_data;
>>> +
>>> +	clk_disable_unprepare(private_data->baud_clk);
>>> +	clk_disable_unprepare(private_data->pclk);
>>>   }
>>>     static int meson_uart_request_port(struct uart_port *port)
>>>   {
>>> +	struct meson_uart_data *private_data = port->private_data;
>>> +	int ret;
>>> +
>>> +	ret = clk_prepare_enable(private_data->pclk);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	ret = clk_prepare_enable(private_data->baud_clk);
>>> +	if (ret) {
>>> +		clk_disable_unprepare(private_data->pclk);
>>> +		return ret;
>>> +	}
>>> +
>>>   	return 0;
>>>   }
>>>   @@ -629,56 +654,175 @@ static struct uart_driver meson_uart_driver = {
>>>   	.cons		= MESON_SERIAL_CONSOLE,
>>>   };
>>>   -static inline struct clk *meson_uart_probe_clock(struct device *dev,
>>> -						 const char *id)
>>> +static int meson_uart_register_clk(struct uart_port *port,
>>> +				   const char *name_suffix,
>>> +				   const struct clk_parent_data *parent_data,
>>> +				   unsigned int num_parents,
>>> +				   const struct clk_ops *ops,
>>> +				   struct clk_hw *hw)
>>>   {
>>> -	struct clk *clk = NULL;
>>> +	struct clk_init_data init = { };
>>> +	char clk_name[32];
>>>   	int ret;
>>>   -	clk = devm_clk_get(dev, id);
>>> -	if (IS_ERR(clk))
>>> -		return clk;
>>> +	snprintf(clk_name, sizeof(clk_name), "%s#%s", dev_name(port->dev),
>>> +		 name_suffix);
>>>   -	ret = clk_prepare_enable(clk);
>>> -	if (ret) {
>>> -		dev_err(dev, "couldn't enable clk\n");
>>> -		return ERR_PTR(ret);
>>> -	}
>>> +	init.name = clk_name;
>>> +	init.ops = ops;
>>> +	init.flags = CLK_SET_RATE_PARENT;
>>> +	init.parent_data = parent_data;
>>> +	init.num_parents = num_parents;
>>> +
>>> +	hw->init = &init;
>>>   -	devm_add_action_or_reset(dev,
>>> -			(void(*)(void *))clk_disable_unprepare,
>>> -			clk);
>>> +	ret = devm_clk_hw_register(port->dev, hw);
>>> +	if (ret)
>>> +		return dev_err_probe(port->dev, ret,
>>> +				     "Failed to register the '%s' clock\n",
>>> +				     clk_name);
>>>   -	return clk;
>>> +	return ret;
>>>   }
>>>   -static int meson_uart_probe_clocks(struct platform_device *pdev,
>>> -				   struct uart_port *port)
>>> -{
>>> -	struct clk *clk_xtal = NULL;
>>> -	struct clk *clk_pclk = NULL;
>>> -	struct clk *clk_baud = NULL;
>>> +static int meson_uart_probe_clocks(struct uart_port *port,
>>> +				   bool register_clk81_div4)
>>> +{
>>> +	struct meson_uart_data *private_data = port->private_data;
>>> +	struct clk_parent_data use_xtal_mux_parents[2] = {
>>> +		{ .index = -1, },
>>> +		{ .index = -1, },
>>> +	};
>>> +	struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
>>> +	struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
>>> +	struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
>>> +	struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
>>> +	struct clk_parent_data baud_div_parent = { };
>>> +	struct clk *clk_baud, *clk_xtal;
>>> +	int ret;
>>>   -	clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
>>> -	if (IS_ERR(clk_pclk))
>>> -		return PTR_ERR(clk_pclk);
>>> +	private_data->pclk = devm_clk_get(port->dev, "pclk");
>>> +	if (IS_ERR(private_data->pclk))
>>> +		return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
>>> +				     "Failed to get the 'pclk' clock\n");
>>> +
>>> +	clk_baud = devm_clk_get(port->dev, "baud");
>>> +	if (IS_ERR(clk_baud))
>>> +		return dev_err_probe(port->dev, PTR_ERR(clk_baud),
>>> +				     "Failed to get the 'baud' clock\n");
>>>   -	clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
>>> +	clk_xtal = devm_clk_get(port->dev, "xtal");
>>>   	if (IS_ERR(clk_xtal))
>>> -		return PTR_ERR(clk_xtal);
>>> +		return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
>>> +				     "Failed to get the 'xtal' clock\n");
>>> +
>>> +	private_data->xtal_div3.mult = 1;
>>> +	private_data->xtal_div3.div = 3;
>>> +	ret = meson_uart_register_clk(port, "xtal_div3", &xtal_div_parent,
>>> +				      1, &clk_fixed_factor_ops,
>>> +				      &private_data->xtal_div3.hw);
>>> +	if (ret)
>>> +		return ret;
>>>   -	clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
>>> -	if (IS_ERR(clk_baud))
>>> -		return PTR_ERR(clk_baud);
>>> +	if (register_clk81_div4) {
>> Clock dividers represent HW elements. The presence of an HW element
>> cannot
>> dependent on the current of the input clocks. There is no way this is right
>> 
>>> +		private_data->clk81_div4.mult = 1;
>>> +		private_data->clk81_div4.div = 4;
>>> +		ret = meson_uart_register_clk(port, "clk81_div4",
>>> +					      &clk81_div_parent, 1,
>>> +					      &clk_fixed_factor_ops,
>>> +					      &private_data->clk81_div4.hw);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;
>> This is only going to be used on meson8 and older. Worst case it should
>> depend
>> on this compatible but I don't think this would be right here.
>> IMO, this shows that the UART "baud" input was actually fed by a
>> derivation of "fclk_div4" instead of "clk_81" on these older SoCs.
>> So instead of registering what I suspect to be a fake element, the
>> meson8 clock controller driver and DT should be fixed a this should go away.
>> 
> Virtually all UART controllers are supported, but they are not used after
> Meson8. So the description is more reasonable in UART driver.

No it is not. It's not virtual, the support for meson8 is there in
mainline.

Unless you tell me there is an actual /4 divider in the UART block of
the meson8 and older and that is was removed from the newer SoC,
the bit above should be removed and the meson8 clock tree fixed.

I think it is more probable that the /4 was never there to begin with,
and that clk81/4 aka "fclk_div4", which is already available from the
clock controller, was actually routed on these older SoCs.

>>> +	}
>>>   -	port->uartclk = clk_get_rate(clk_baud);
>>> +	if (private_data->has_xtal_clk_sel) {
>>> +		private_data->xtal_div2.mult = 1;
>>> +		private_data->xtal_div2.div = 2;
>>> +		ret = meson_uart_register_clk(port, "xtal_div2",
>>> +					      &xtal_div_parent, 1,
>>> +					      &clk_fixed_factor_ops,
>>> +					      &private_data->xtal_div2.hw);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
>>> +		xtal_clk_sel_mux_parents[1].fw_name = "xtal";
>>> +
>>> +		private_data->xtal_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>>> +		private_data->xtal_clk_sel_mux.mask = 0x1;
>>> +		private_data->xtal_clk_sel_mux.shift = 26;
>>> +		private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>> +		ret = meson_uart_register_clk(port, "xtal_clk_sel",
>>> +					      xtal_clk_sel_mux_parents,
>>> +					      ARRAY_SIZE(xtal_clk_sel_mux_parents),
>>> +					      &clk_mux_ops,
>>> +					      &private_data->xtal_clk_sel_mux.hw);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		xtal2_clk_sel_mux_parents[0].hw = &private_data->xtal_clk_sel_mux.hw;
>>> +		xtal2_clk_sel_mux_parents[1].hw = &private_data->xtal_div2.hw;
>>> +
>>> +		private_data->xtal2_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>>> +		private_data->xtal2_clk_sel_mux.mask = 0x1;
>>> +		private_data->xtal2_clk_sel_mux.shift = 27;
>>> +		private_data->xtal2_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>> +		ret = meson_uart_register_clk(port, "xtal2_clk_sel",
>>> +					      xtal2_clk_sel_mux_parents,
>>> +					      ARRAY_SIZE(xtal2_clk_sel_mux_parents),
>>> +					      &clk_mux_ops,
>>> +					      &private_data->xtal2_clk_sel_mux.hw);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		use_xtal_mux_parents[1].hw = &private_data->xtal2_clk_sel_mux.hw;
>>> +	} else {
>>> +		use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;
>> Well the above is a bit over-complicated. If I summarize:
>> GXBB and older used a fixed divider of 3. Bits 26 and 27 read
>> 0 according to the documentation.
>> Chips after GXBB have 2 configurable divider of 2 and 3. bits 26 and 27
>> selects which of these dividers is used.
>> So the above could be replaced with a single divider covering bits 26 and
>> 27
>> with the following divider table { 2, 2, 1, 3 }. The divider should use
>> RO ops and not have CLK_SET_RATE_PARENT. It can be used for all chips
>> variant like this, including the older ones.
>> 
> As you say, the way you say it is complicated. I think you're complicating
> things.I don't understand what you're trying to achieve.
> Can you be more specific?

I'm proposing to replace the 4 elements above, paths, parent_table and
all with a single divider with a specific table. It accomplishes the same thing.

|--------+--------+-----------|
| Bit 27 | Bit 26 | Div Value |
|--------+--------+-----------|
|      0 |      0 |         3 |
|--------+--------+-----------|
|      0 |      1 |         1 |
|--------+--------+-----------|
|      1 |      0 |         2 |
|--------+--------+-----------|
|      1 |      1 |         2 |
|--------+--------+-----------|

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/clk-provider.h?h=v5.16-rc8#n602

something like:

static struct clk_div_table reg5_table[] = {
	{ 0, 3 },
        { 1, 1 },
        { 2, 2 },
        { 3, 2 },
};

I'm also pointing out that, since bit 26 and 27 reads 0 on the older
SoCs, the divider would work for these too, as long as the Ops is
read-only. So you don't even need to register different element
depending on the SoC which simplify things a bit more.

If you prefer to keep a fixed divider for the older ones, It's fine by
me. It gives a bit more work but it is closer to reality. The newer SoC
could still use the custom divider regardless.

Last, and as explained on other bits, that part of the clock path should
remain RO at first, to keep behavior stable with this change. You may
change that later on, in a dedicated patch describing the change.

>
>> The only information you need to carry if whether or not you want to
>> make this divider modifiable. This means using the dt data to store
>> clk_div_ops or clk_div_ro_ops pointer.
>> To avoid changing the behavior of the older platforms in this patch, I
>> would
>> suggest to make everything use clk_div_ro_ops first, and make another
>> patch to use clk_div_ops if necessary.
>> 
>>> +	}
>>> +
>>> +	private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
>>> +	private_data->use_xtal_mux.mask = 0x1;
>>> +	private_data->use_xtal_mux.shift = 24;
>>> +	private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>> +	ret = meson_uart_register_clk(port, "use_xtal", use_xtal_mux_parents,
>>> +				      ARRAY_SIZE(use_xtal_mux_parents),
>>> +				      &clk_mux_ops,
>> Use RO ops here to start with.
>> You can make this writable with another patch explicitly describing the change.
> I agree with what you say, and I will correct it.
>> 
>>> +				      &private_data->use_xtal_mux.hw);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	baud_div_parent.hw = &private_data->use_xtal_mux.hw;
>>> +
>>> +	private_data->baud_div.reg = port->membase + AML_UART_REG5;
>>> +	private_data->baud_div.shift = 0;
>>> +	private_data->baud_div.width = 23;
>>> +	private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
>>> +	ret = meson_uart_register_clk(port, "baud_div",
>>> +				      &baud_div_parent, 1,
>>> +				      &clk_divider_ops,
>>> +				      &private_data->baud_div.hw);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
>>> +						     &private_data->baud_div.hw,
>>> +						     "baud_rate");
>> There is a problem with this function in CCF. It will pin the driver to
>> itself, making it unremovable. It is an ongoing topic. For now, just use
>> "hw->clk" to get the struct *clk.
>> 
>>> +	if (IS_ERR(private_data->baud_clk))
>>> +		return dev_err_probe(port->dev,
>> I don't think anything can defer here, so dev_err_probe() is not
>> necessary I think
>> 
> I agree with what you say, and I will correct it.
>>> +				     PTR_ERR(private_data->baud_clk),
>>> +				     "Failed to request the 'baud_rate' clock\n");
>>>     	return 0;
>>>   }
>>>     static int meson_uart_probe(struct platform_device *pdev)
>>>   {
>>> +	struct meson_uart_data *private_data;
>>>   	struct resource *res_mem, *res_irq;
>>> +	struct clk *clk_baud, *clk_xtal;
>>> +	bool register_clk81_div4;
>>>   	struct uart_port *port;
>>>   	int ret = 0;
>>>   	int id = -1;
>>> @@ -711,18 +855,37 @@ static int meson_uart_probe(struct platform_device *pdev)
>>>   		return -EBUSY;
>>>   	}
>>>   -	port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port),
>>> GFP_KERNEL);
>>> -	if (!port)
>>> +	private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
>>> +				    GFP_KERNEL);
>>> +	if (!private_data)
>>>   		return -ENOMEM;
>>>   +	if (device_get_match_data(&pdev->dev))
>>> +		private_data->has_xtal_clk_sel = true;
>>> +
>>> +	private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
>>> +	if (IS_ERR(private_data->pclk))
>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
>>> +				     "Failed to get the 'pclk' clock\n");
>>> +
>>> +	clk_baud = devm_clk_get(&pdev->dev, "baud");
>>> +	if (IS_ERR(clk_baud))
>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
>>> +				     "Failed to get the 'baud' clock\n");
>>> +
>>> +	clk_xtal = devm_clk_get(&pdev->dev, "xtal");
>>> +	if (IS_ERR(clk_xtal))
>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
>>> +				     "Failed to get the 'xtal' clock\n");
>>> +
>> This is second time you call devm_clk_get() on these clocks. One
>> instance has to go away
>> 
> I agree with what you say, and I will correct it.
>>> +	register_clk81_div4 = clk_get_rate(clk_xtal) != clk_get_rate(clk_baud);
>>> +
>> The above is a ugly way to distinguish the meson8 (32bit) SoC family
>> from the rest. This definitely not the way to achieve it.
>> The right is the compatible data but here I think it is not
>> necessary. The clock input should be fixed.
>> 
>>> +	port = &private_data->port;
>>> +
>>>   	port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
>>>   	if (IS_ERR(port->membase))
>>>   		return PTR_ERR(port->membase);
>>>   -	ret = meson_uart_probe_clocks(pdev, port);
>>> -	if (ret)
>>> -		return ret;
>>> -
>>>   	port->iotype = UPIO_MEM;
>>>   	port->mapbase = res_mem->start;
>>>   	port->mapsize = resource_size(res_mem);
>>> @@ -735,6 +898,12 @@ static int meson_uart_probe(struct platform_device *pdev)
>>>   	port->x_char = 0;
>>>   	port->ops = &meson_uart_ops;
>>>   	port->fifosize = 64;
>>> +	port->uartclk = clk_get_rate(clk_baud);
>>> +	port->private_data = private_data;
>>> +
>>> +	ret = meson_uart_probe_clocks(port, register_clk81_div4);
>>> +	if (ret)
>>> +		return ret;
>>>     	meson_ports[pdev->id] = port;
>>>   	platform_set_drvdata(pdev, port);
>>> @@ -761,10 +930,42 @@ static int meson_uart_remove(struct platform_device *pdev)
>>>   }
>>>     static const struct of_device_id meson_uart_dt_match[] = {
>>> -	{ .compatible = "amlogic,meson6-uart" },
>>> -	{ .compatible = "amlogic,meson8-uart" },
>>> -	{ .compatible = "amlogic,meson8b-uart" },
>>> -	{ .compatible = "amlogic,meson-gx-uart" },
>>> +	{
>>> +		.compatible = "amlogic,meson6-uart",
>>> +		.data = (void *)false,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson8-uart",
>>> +		.data = (void *)false,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson8b-uart",
>>> +		.data = (void *)false,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson-gxbb-uart",
>>> +		.data = (void *)false,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson-gxl-uart",
>>> +		.data = (void *)true,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson-g12a-uart",
>>> +		.data = (void *)true,
>>> +	},
>>> +	{
>>> +		.compatible = "amlogic,meson-s4-uart",
>>> +		.data = (void *)true,
>>> +	},
>>> +	/*
>>> +	 * deprecated, don't use anymore because it doesn't differentiate
>>> +	 * between GXBB and GXL which have different revisions of the UART IP.
>>> +	 */
>>> +	{
>>> +		.compatible = "amlogic,meson-gx-uart",
>>> +		.data = (void *)false,
>>> +	},
>>>   	{ /* sentinel */ },
>>>   };
>>>   MODULE_DEVICE_TABLE(of, meson_uart_dt_match);
>>
Yu Tu Jan. 4, 2022, 2:35 p.m. UTC | #13
Hi Jerome,
	Thank you very much for your patient reply. I have learned a lot from it.

On 2022/1/4 18:36, Jerome Brunet wrote:
> [ EXTERNAL EMAIL ]
> 
> 
> On Tue 04 Jan 2022 at 17:57, Yu Tu <yu.tu@amlogic.com> wrote:
> 
>> Hi Jerome,
>> 	Thank you very much for your reply.
>>
>> On 2022/1/3 20:40, Jerome Brunet wrote:
>>> [ EXTERNAL EMAIL ]
>>>
>>> On Thu 30 Dec 2021 at 18:21, Yu Tu <yu.tu@amlogic.com> wrote:
>>>
>>>> Using the common Clock code to describe the UART baud rate clock makes
>>>> it easier for the UART driver to be compatible with the baud rate
>>>> requirements of the UART IP on different meson chips
>>>>
>>>> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
>>>> ---
>>>>    drivers/tty/serial/Kconfig      |   1 +
>>>>    drivers/tty/serial/meson_uart.c | 311 ++++++++++++++++++++++++++------
>>>>    2 files changed, 257 insertions(+), 55 deletions(-)
>>>>
>>>> diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
>>>> index 780908d43557..32e238173036 100644
>>>> --- a/drivers/tty/serial/Kconfig
>>>> +++ b/drivers/tty/serial/Kconfig
>>>> @@ -198,6 +198,7 @@ config SERIAL_KGDB_NMI
>>>>    config SERIAL_MESON
>>>>    	tristate "Meson serial port support"
>>>>    	depends on ARCH_MESON
>>>> +	depends on COMMON_CLK
>>>>    	select SERIAL_CORE
>>>>    	help
>>>>    	  This enables the driver for the on-chip UARTs of the Amlogic
>>>> diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
>>>> index 99efe62a1507..07eb1f40aaaa 100644
>>>> --- a/drivers/tty/serial/meson_uart.c
>>>> +++ b/drivers/tty/serial/meson_uart.c
>>>> @@ -6,6 +6,7 @@
>>>>     */
>>>>      #include <linux/clk.h>
>>>> +#include <linux/clk-provider.h>
>>>>    #include <linux/console.h>
>>>>    #include <linux/delay.h>
>>>>    #include <linux/init.h>
>>>> @@ -65,9 +66,7 @@
>>>>    #define AML_UART_RECV_IRQ(c)		((c) & 0xff)
>>>>      /* AML_UART_REG5 bits */
>>>> -#define AML_UART_BAUD_MASK		0x7fffff
>>>>    #define AML_UART_BAUD_USE		BIT(23)
>>>> -#define AML_UART_BAUD_XTAL		BIT(24)
>>>>      #define AML_UART_PORT_NUM		12
>>>>    #define AML_UART_PORT_OFFSET		6
>>>> @@ -76,6 +75,21 @@
>>>>    #define AML_UART_POLL_USEC		5
>>>>    #define AML_UART_TIMEOUT_USEC		10000
>>>>    +struct meson_uart_data {
>>>> +	struct uart_port	port;
>>>> +	struct clk		*pclk;
>>>> +	struct clk		*baud_clk;
>>>> +	struct clk_divider	baud_div;
>>>> +	struct clk_mux		use_xtal_mux;
>>>> +	struct clk_mux		xtal_clk_sel_mux;
>>>> +	struct clk_mux		xtal2_clk_sel_mux;
>>>> +	struct clk_fixed_factor	xtal_div2;
>>>> +	struct clk_fixed_factor	xtal_div3;
>>>> +	struct clk_fixed_factor	clk81_div4;
>>> Keeping all these internal elements around is not useful since they are
>>> registered using devm_
>>>
>> I'm sorry. I don't know what you mean. That's exactly what you said, but
>> what's wrong, I don't understand. Do you have any better suggestions,
>> please specify specific points, preferably give examples.
> 
> I'm saying that you don't need to keep reference to the internal
> elements of the clock tree you have registered since devm_ will take
> care of the removal later on. IOW, Once they are registered, the pointer
> is never used again so you don't need it in the private data.
> 
I understand you now. What you say is right. I will correct it.
> 
>>>> +	bool			no_clk81_input;
>>> What is this ?
>>>
>> To distinguish between clK81 and XTAL.
> 
> ... Yet, it is not used
> 
>>>> +	bool			has_xtal_clk_sel;
>>>> +};
>>>> +
>>>>    static struct uart_driver meson_uart_driver;
>>>>      static struct uart_port *meson_ports[AML_UART_PORT_NUM];
>>>> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port *port)
>>>>    static int meson_uart_startup(struct uart_port *port)
>>>>    {
>>>>    	u32 val;
>>>> -	int ret = 0;
>>>> +	int ret;
>>>>    -	val = readl(port->membase + AML_UART_CONTROL);
>>>> -	val |= AML_UART_CLEAR_ERR;
>>>> -	writel(val, port->membase + AML_UART_CONTROL);
>>>> -	val &= ~AML_UART_CLEAR_ERR;
>>>> -	writel(val, port->membase + AML_UART_CONTROL);
>>>> +	meson_uart_reset(port);
>>>>    +	val = readl(port->membase + AML_UART_CONTROL);
>>>>    	val |= (AML_UART_RX_EN | AML_UART_TX_EN);
>>>>    	writel(val, port->membase + AML_UART_CONTROL);
>>>>    @@ -295,19 +306,17 @@ static int meson_uart_startup(struct uart_port
>>>> *port)
>>>>      static void meson_uart_change_speed(struct uart_port *port, unsigned
>>>> long baud)
>>>>    {
>>>> +	struct meson_uart_data *private_data = port->private_data;
>>>>    	u32 val;
>>>>      	while (!meson_uart_tx_empty(port))
>>>>    		cpu_relax();
>>>>    -	if (port->uartclk == 24000000) {
>>> This check shows that previous code assumed bit 24 was left untouched
>>> Below you can see that GXBB and newer used the XTAL path while older
>>> used the other
>>> Your change makes this dynamic which is another "unexpected" change.
>>> Please make the bit 24 mux RO to start with so the behavior remains unchanged
>>> for older SoCs.
>> I agree with what you say, and I will correct it.
>>>
>>>> -		val = ((port->uartclk / 3) / baud) - 1;
>>>> -		val |= AML_UART_BAUD_XTAL;
>>>> -	} else {
>>>> -		val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
>>>> -	}
>>>> +	val = readl(port->membase + AML_UART_REG5);
>>>>    	val |= AML_UART_BAUD_USE;
>>>>    	writel(val, port->membase + AML_UART_REG5);
>>>> +
>>>> +	clk_set_rate(private_data->baud_clk, baud);
>>>>    }
>>>>      static void meson_uart_set_termios(struct uart_port *port,
>>>> @@ -397,11 +406,27 @@ static int meson_uart_verify_port(struct uart_port *port,
>>>>      static void meson_uart_release_port(struct uart_port *port)
>>>>    {
>>>> -	/* nothing to do */
>>>> +	struct meson_uart_data *private_data = port->private_data;
>>>> +
>>>> +	clk_disable_unprepare(private_data->baud_clk);
>>>> +	clk_disable_unprepare(private_data->pclk);
>>>>    }
>>>>      static int meson_uart_request_port(struct uart_port *port)
>>>>    {
>>>> +	struct meson_uart_data *private_data = port->private_data;
>>>> +	int ret;
>>>> +
>>>> +	ret = clk_prepare_enable(private_data->pclk);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	ret = clk_prepare_enable(private_data->baud_clk);
>>>> +	if (ret) {
>>>> +		clk_disable_unprepare(private_data->pclk);
>>>> +		return ret;
>>>> +	}
>>>> +
>>>>    	return 0;
>>>>    }
>>>>    @@ -629,56 +654,175 @@ static struct uart_driver meson_uart_driver = {
>>>>    	.cons		= MESON_SERIAL_CONSOLE,
>>>>    };
>>>>    -static inline struct clk *meson_uart_probe_clock(struct device *dev,
>>>> -						 const char *id)
>>>> +static int meson_uart_register_clk(struct uart_port *port,
>>>> +				   const char *name_suffix,
>>>> +				   const struct clk_parent_data *parent_data,
>>>> +				   unsigned int num_parents,
>>>> +				   const struct clk_ops *ops,
>>>> +				   struct clk_hw *hw)
>>>>    {
>>>> -	struct clk *clk = NULL;
>>>> +	struct clk_init_data init = { };
>>>> +	char clk_name[32];
>>>>    	int ret;
>>>>    -	clk = devm_clk_get(dev, id);
>>>> -	if (IS_ERR(clk))
>>>> -		return clk;
>>>> +	snprintf(clk_name, sizeof(clk_name), "%s#%s", dev_name(port->dev),
>>>> +		 name_suffix);
>>>>    -	ret = clk_prepare_enable(clk);
>>>> -	if (ret) {
>>>> -		dev_err(dev, "couldn't enable clk\n");
>>>> -		return ERR_PTR(ret);
>>>> -	}
>>>> +	init.name = clk_name;
>>>> +	init.ops = ops;
>>>> +	init.flags = CLK_SET_RATE_PARENT;
>>>> +	init.parent_data = parent_data;
>>>> +	init.num_parents = num_parents;
>>>> +
>>>> +	hw->init = &init;
>>>>    -	devm_add_action_or_reset(dev,
>>>> -			(void(*)(void *))clk_disable_unprepare,
>>>> -			clk);
>>>> +	ret = devm_clk_hw_register(port->dev, hw);
>>>> +	if (ret)
>>>> +		return dev_err_probe(port->dev, ret,
>>>> +				     "Failed to register the '%s' clock\n",
>>>> +				     clk_name);
>>>>    -	return clk;
>>>> +	return ret;
>>>>    }
>>>>    -static int meson_uart_probe_clocks(struct platform_device *pdev,
>>>> -				   struct uart_port *port)
>>>> -{
>>>> -	struct clk *clk_xtal = NULL;
>>>> -	struct clk *clk_pclk = NULL;
>>>> -	struct clk *clk_baud = NULL;
>>>> +static int meson_uart_probe_clocks(struct uart_port *port,
>>>> +				   bool register_clk81_div4)
>>>> +{
>>>> +	struct meson_uart_data *private_data = port->private_data;
>>>> +	struct clk_parent_data use_xtal_mux_parents[2] = {
>>>> +		{ .index = -1, },
>>>> +		{ .index = -1, },
>>>> +	};
>>>> +	struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
>>>> +	struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
>>>> +	struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
>>>> +	struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
>>>> +	struct clk_parent_data baud_div_parent = { };
>>>> +	struct clk *clk_baud, *clk_xtal;
>>>> +	int ret;
>>>>    -	clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
>>>> -	if (IS_ERR(clk_pclk))
>>>> -		return PTR_ERR(clk_pclk);
>>>> +	private_data->pclk = devm_clk_get(port->dev, "pclk");
>>>> +	if (IS_ERR(private_data->pclk))
>>>> +		return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
>>>> +				     "Failed to get the 'pclk' clock\n");
>>>> +
>>>> +	clk_baud = devm_clk_get(port->dev, "baud");
>>>> +	if (IS_ERR(clk_baud))
>>>> +		return dev_err_probe(port->dev, PTR_ERR(clk_baud),
>>>> +				     "Failed to get the 'baud' clock\n");
>>>>    -	clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
>>>> +	clk_xtal = devm_clk_get(port->dev, "xtal");
>>>>    	if (IS_ERR(clk_xtal))
>>>> -		return PTR_ERR(clk_xtal);
>>>> +		return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
>>>> +				     "Failed to get the 'xtal' clock\n");
>>>> +
>>>> +	private_data->xtal_div3.mult = 1;
>>>> +	private_data->xtal_div3.div = 3;
>>>> +	ret = meson_uart_register_clk(port, "xtal_div3", &xtal_div_parent,
>>>> +				      1, &clk_fixed_factor_ops,
>>>> +				      &private_data->xtal_div3.hw);
>>>> +	if (ret)
>>>> +		return ret;
>>>>    -	clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
>>>> -	if (IS_ERR(clk_baud))
>>>> -		return PTR_ERR(clk_baud);
>>>> +	if (register_clk81_div4) {
>>> Clock dividers represent HW elements. The presence of an HW element
>>> cannot
>>> dependent on the current of the input clocks. There is no way this is right
>>>
>>>> +		private_data->clk81_div4.mult = 1;
>>>> +		private_data->clk81_div4.div = 4;
>>>> +		ret = meson_uart_register_clk(port, "clk81_div4",
>>>> +					      &clk81_div_parent, 1,
>>>> +					      &clk_fixed_factor_ops,
>>>> +					      &private_data->clk81_div4.hw);
>>>> +		if (ret)
>>>> +			return ret;
>>>> +
>>>> +		use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;
>>> This is only going to be used on meson8 and older. Worst case it should
>>> depend
>>> on this compatible but I don't think this would be right here.
>>> IMO, this shows that the UART "baud" input was actually fed by a
>>> derivation of "fclk_div4" instead of "clk_81" on these older SoCs.
>>> So instead of registering what I suspect to be a fake element, the
>>> meson8 clock controller driver and DT should be fixed a this should go away.
>>>
>> Virtually all UART controllers are supported, but they are not used after
>> Meson8. So the description is more reasonable in UART driver.
> 
> No it is not. It's not virtual, the support for meson8 is there in
> mainline.
> 
> Unless you tell me there is an actual /4 divider in the UART block of
> the meson8 and older and that is was removed from the newer SoC,
> the bit above should be removed and the meson8 clock tree fixed.
> 
> I think it is more probable that the /4 was never there to begin with,
> and that clk81/4 aka "fclk_div4", which is already available from the
> clock controller, was actually routed on these older SoCs.
I will check with the chip design department if it is divided by 4. I'll 
get back to you.
> 
>>>> +	}
>>>>    -	port->uartclk = clk_get_rate(clk_baud);
>>>> +	if (private_data->has_xtal_clk_sel) {
>>>> +		private_data->xtal_div2.mult = 1;
>>>> +		private_data->xtal_div2.div = 2;
>>>> +		ret = meson_uart_register_clk(port, "xtal_div2",
>>>> +					      &xtal_div_parent, 1,
>>>> +					      &clk_fixed_factor_ops,
>>>> +					      &private_data->xtal_div2.hw);
>>>> +		if (ret)
>>>> +			return ret;
>>>> +
>>>> +		xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
>>>> +		xtal_clk_sel_mux_parents[1].fw_name = "xtal";
>>>> +
>>>> +		private_data->xtal_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>>>> +		private_data->xtal_clk_sel_mux.mask = 0x1;
>>>> +		private_data->xtal_clk_sel_mux.shift = 26;
>>>> +		private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>>> +		ret = meson_uart_register_clk(port, "xtal_clk_sel",
>>>> +					      xtal_clk_sel_mux_parents,
>>>> +					      ARRAY_SIZE(xtal_clk_sel_mux_parents),
>>>> +					      &clk_mux_ops,
>>>> +					      &private_data->xtal_clk_sel_mux.hw);
>>>> +		if (ret)
>>>> +			return ret;
>>>> +
>>>> +		xtal2_clk_sel_mux_parents[0].hw = &private_data->xtal_clk_sel_mux.hw;
>>>> +		xtal2_clk_sel_mux_parents[1].hw = &private_data->xtal_div2.hw;
>>>> +
>>>> +		private_data->xtal2_clk_sel_mux.reg = port->membase + AML_UART_REG5;
>>>> +		private_data->xtal2_clk_sel_mux.mask = 0x1;
>>>> +		private_data->xtal2_clk_sel_mux.shift = 27;
>>>> +		private_data->xtal2_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>>> +		ret = meson_uart_register_clk(port, "xtal2_clk_sel",
>>>> +					      xtal2_clk_sel_mux_parents,
>>>> +					      ARRAY_SIZE(xtal2_clk_sel_mux_parents),
>>>> +					      &clk_mux_ops,
>>>> +					      &private_data->xtal2_clk_sel_mux.hw);
>>>> +		if (ret)
>>>> +			return ret;
>>>> +
>>>> +		use_xtal_mux_parents[1].hw = &private_data->xtal2_clk_sel_mux.hw;
>>>> +	} else {
>>>> +		use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;
>>> Well the above is a bit over-complicated. If I summarize:
>>> GXBB and older used a fixed divider of 3. Bits 26 and 27 read
>>> 0 according to the documentation.
>>> Chips after GXBB have 2 configurable divider of 2 and 3. bits 26 and 27
>>> selects which of these dividers is used.
>>> So the above could be replaced with a single divider covering bits 26 and
>>> 27
>>> with the following divider table { 2, 2, 1, 3 }. The divider should use
>>> RO ops and not have CLK_SET_RATE_PARENT. It can be used for all chips
>>> variant like this, including the older ones.
>>>
>> As you say, the way you say it is complicated. I think you're complicating
>> things.I don't understand what you're trying to achieve.
>> Can you be more specific?
> 
> I'm proposing to replace the 4 elements above, paths, parent_table and
> all with a single divider with a specific table. It accomplishes the same thing.
> 
> |--------+--------+-----------|
> | Bit 27 | Bit 26 | Div Value |
> |--------+--------+-----------|
> |      0 |      0 |         3 |
> |--------+--------+-----------|
> |      0 |      1 |         1 |
> |--------+--------+-----------|
> |      1 |      0 |         2 |
> |--------+--------+-----------|
> |      1 |      1 |         2 |
> |--------+--------+-----------|
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/clk-provider.h?h=v5.16-rc8#n602
> 
> something like:
> 
> static struct clk_div_table reg5_table[] = {
> 	{ 0, 3 },
>          { 1, 1 },
>          { 2, 2 },
>          { 3, 2 },
> };
> 
> I'm also pointing out that, since bit 26 and 27 reads 0 on the older
> SoCs, the divider would work for these too, as long as the Ops is
> read-only. So you don't even need to register different element
> depending on the SoC which simplify things a bit more.
> 
> If you prefer to keep a fixed divider for the older ones, It's fine by
> me. It gives a bit more work but it is closer to reality. The newer SoC
> could still use the custom divider regardless.
> 
> Last, and as explained on other bits, that part of the clock path should
> remain RO at first, to keep behavior stable with this change. You may
> change that later on, in a dedicated patch describing the change.
> 
I understand what you mean, and i will try to realize it. I hope you can 
help review and give me your suggestions.
>>
>>> The only information you need to carry if whether or not you want to
>>> make this divider modifiable. This means using the dt data to store
>>> clk_div_ops or clk_div_ro_ops pointer.
>>> To avoid changing the behavior of the older platforms in this patch, I
>>> would
>>> suggest to make everything use clk_div_ro_ops first, and make another
>>> patch to use clk_div_ops if necessary.
>>>
>>>> +	}
>>>> +
>>>> +	private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
>>>> +	private_data->use_xtal_mux.mask = 0x1;
>>>> +	private_data->use_xtal_mux.shift = 24;
>>>> +	private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>>> +	ret = meson_uart_register_clk(port, "use_xtal", use_xtal_mux_parents,
>>>> +				      ARRAY_SIZE(use_xtal_mux_parents),
>>>> +				      &clk_mux_ops,
>>> Use RO ops here to start with.
>>> You can make this writable with another patch explicitly describing the change.
>> I agree with what you say, and I will correct it.
>>>
>>>> +				      &private_data->use_xtal_mux.hw);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	baud_div_parent.hw = &private_data->use_xtal_mux.hw;
>>>> +
>>>> +	private_data->baud_div.reg = port->membase + AML_UART_REG5;
>>>> +	private_data->baud_div.shift = 0;
>>>> +	private_data->baud_div.width = 23;
>>>> +	private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
>>>> +	ret = meson_uart_register_clk(port, "baud_div",
>>>> +				      &baud_div_parent, 1,
>>>> +				      &clk_divider_ops,
>>>> +				      &private_data->baud_div.hw);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
>>>> +						     &private_data->baud_div.hw,
>>>> +						     "baud_rate");
>>> There is a problem with this function in CCF. It will pin the driver to
>>> itself, making it unremovable. It is an ongoing topic. For now, just use
>>> "hw->clk" to get the struct *clk.
>>>
>>>> +	if (IS_ERR(private_data->baud_clk))
>>>> +		return dev_err_probe(port->dev,
>>> I don't think anything can defer here, so dev_err_probe() is not
>>> necessary I think
>>>
>> I agree with what you say, and I will correct it.
>>>> +				     PTR_ERR(private_data->baud_clk),
>>>> +				     "Failed to request the 'baud_rate' clock\n");
>>>>      	return 0;
>>>>    }
>>>>      static int meson_uart_probe(struct platform_device *pdev)
>>>>    {
>>>> +	struct meson_uart_data *private_data;
>>>>    	struct resource *res_mem, *res_irq;
>>>> +	struct clk *clk_baud, *clk_xtal;
>>>> +	bool register_clk81_div4;
>>>>    	struct uart_port *port;
>>>>    	int ret = 0;
>>>>    	int id = -1;
>>>> @@ -711,18 +855,37 @@ static int meson_uart_probe(struct platform_device *pdev)
>>>>    		return -EBUSY;
>>>>    	}
>>>>    -	port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port),
>>>> GFP_KERNEL);
>>>> -	if (!port)
>>>> +	private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
>>>> +				    GFP_KERNEL);
>>>> +	if (!private_data)
>>>>    		return -ENOMEM;
>>>>    +	if (device_get_match_data(&pdev->dev))
>>>> +		private_data->has_xtal_clk_sel = true;
>>>> +
>>>> +	private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
>>>> +	if (IS_ERR(private_data->pclk))
>>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
>>>> +				     "Failed to get the 'pclk' clock\n");
>>>> +
>>>> +	clk_baud = devm_clk_get(&pdev->dev, "baud");
>>>> +	if (IS_ERR(clk_baud))
>>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
>>>> +				     "Failed to get the 'baud' clock\n");
>>>> +
>>>> +	clk_xtal = devm_clk_get(&pdev->dev, "xtal");
>>>> +	if (IS_ERR(clk_xtal))
>>>> +		return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
>>>> +				     "Failed to get the 'xtal' clock\n");
>>>> +
>>> This is second time you call devm_clk_get() on these clocks. One
>>> instance has to go away
>>>
>> I agree with what you say, and I will correct it.
>>>> +	register_clk81_div4 = clk_get_rate(clk_xtal) != clk_get_rate(clk_baud);
>>>> +
>>> The above is a ugly way to distinguish the meson8 (32bit) SoC family
>>> from the rest. This definitely not the way to achieve it.
>>> The right is the compatible data but here I think it is not
>>> necessary. The clock input should be fixed.
>>>
>>>> +	port = &private_data->port;
>>>> +
>>>>    	port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
>>>>    	if (IS_ERR(port->membase))
>>>>    		return PTR_ERR(port->membase);
>>>>    -	ret = meson_uart_probe_clocks(pdev, port);
>>>> -	if (ret)
>>>> -		return ret;
>>>> -
>>>>    	port->iotype = UPIO_MEM;
>>>>    	port->mapbase = res_mem->start;
>>>>    	port->mapsize = resource_size(res_mem);
>>>> @@ -735,6 +898,12 @@ static int meson_uart_probe(struct platform_device *pdev)
>>>>    	port->x_char = 0;
>>>>    	port->ops = &meson_uart_ops;
>>>>    	port->fifosize = 64;
>>>> +	port->uartclk = clk_get_rate(clk_baud);
>>>> +	port->private_data = private_data;
>>>> +
>>>> +	ret = meson_uart_probe_clocks(port, register_clk81_div4);
>>>> +	if (ret)
>>>> +		return ret;
>>>>      	meson_ports[pdev->id] = port;
>>>>    	platform_set_drvdata(pdev, port);
>>>> @@ -761,10 +930,42 @@ static int meson_uart_remove(struct platform_device *pdev)
>>>>    }
>>>>      static const struct of_device_id meson_uart_dt_match[] = {
>>>> -	{ .compatible = "amlogic,meson6-uart" },
>>>> -	{ .compatible = "amlogic,meson8-uart" },
>>>> -	{ .compatible = "amlogic,meson8b-uart" },
>>>> -	{ .compatible = "amlogic,meson-gx-uart" },
>>>> +	{
>>>> +		.compatible = "amlogic,meson6-uart",
>>>> +		.data = (void *)false,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson8-uart",
>>>> +		.data = (void *)false,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson8b-uart",
>>>> +		.data = (void *)false,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson-gxbb-uart",
>>>> +		.data = (void *)false,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson-gxl-uart",
>>>> +		.data = (void *)true,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson-g12a-uart",
>>>> +		.data = (void *)true,
>>>> +	},
>>>> +	{
>>>> +		.compatible = "amlogic,meson-s4-uart",
>>>> +		.data = (void *)true,
>>>> +	},
>>>> +	/*
>>>> +	 * deprecated, don't use anymore because it doesn't differentiate
>>>> +	 * between GXBB and GXL which have different revisions of the UART IP.
>>>> +	 */
>>>> +	{
>>>> +		.compatible = "amlogic,meson-gx-uart",
>>>> +		.data = (void *)false,
>>>> +	},
>>>>    	{ /* sentinel */ },
>>>>    };
>>>>    MODULE_DEVICE_TABLE(of, meson_uart_dt_match);
>>>
>
Yu Tu Jan. 5, 2022, 5:53 a.m. UTC | #14
Hi Jerome,

On 2022/1/4 22:35, Yu Tu wrote:
> Hi Jerome,
>      Thank you very much for your patient reply. I have learned a lot 
> from it.
> 
> On 2022/1/4 18:36, Jerome Brunet wrote:
>> [ EXTERNAL EMAIL ]
>>
>>
>> On Tue 04 Jan 2022 at 17:57, Yu Tu <yu.tu@amlogic.com> wrote:
>>
>>> Hi Jerome,
>>>     Thank you very much for your reply.
>>>
>>> On 2022/1/3 20:40, Jerome Brunet wrote:
>>>> [ EXTERNAL EMAIL ]
>>>>
>>>> On Thu 30 Dec 2021 at 18:21, Yu Tu <yu.tu@amlogic.com> wrote:
>>>>
>>>>> Using the common Clock code to describe the UART baud rate clock makes
>>>>> it easier for the UART driver to be compatible with the baud rate
>>>>> requirements of the UART IP on different meson chips
>>>>>
>>>>> Signed-off-by: Yu Tu <yu.tu@amlogic.com>
>>>>> ---
>>>>>    drivers/tty/serial/Kconfig      |   1 +
>>>>>    drivers/tty/serial/meson_uart.c | 311 
>>>>> ++++++++++++++++++++++++++------
>>>>>    2 files changed, 257 insertions(+), 55 deletions(-)
>>>>>
>>>>> diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
>>>>> index 780908d43557..32e238173036 100644
>>>>> --- a/drivers/tty/serial/Kconfig
>>>>> +++ b/drivers/tty/serial/Kconfig
>>>>> @@ -198,6 +198,7 @@ config SERIAL_KGDB_NMI
>>>>>    config SERIAL_MESON
>>>>>        tristate "Meson serial port support"
>>>>>        depends on ARCH_MESON
>>>>> +    depends on COMMON_CLK
>>>>>        select SERIAL_CORE
>>>>>        help
>>>>>          This enables the driver for the on-chip UARTs of the Amlogic
>>>>> diff --git a/drivers/tty/serial/meson_uart.c 
>>>>> b/drivers/tty/serial/meson_uart.c
>>>>> index 99efe62a1507..07eb1f40aaaa 100644
>>>>> --- a/drivers/tty/serial/meson_uart.c
>>>>> +++ b/drivers/tty/serial/meson_uart.c
>>>>> @@ -6,6 +6,7 @@
>>>>>     */
>>>>>      #include <linux/clk.h>
>>>>> +#include <linux/clk-provider.h>
>>>>>    #include <linux/console.h>
>>>>>    #include <linux/delay.h>
>>>>>    #include <linux/init.h>
>>>>> @@ -65,9 +66,7 @@
>>>>>    #define AML_UART_RECV_IRQ(c)        ((c) & 0xff)
>>>>>      /* AML_UART_REG5 bits */
>>>>> -#define AML_UART_BAUD_MASK        0x7fffff
>>>>>    #define AML_UART_BAUD_USE        BIT(23)
>>>>> -#define AML_UART_BAUD_XTAL        BIT(24)
>>>>>      #define AML_UART_PORT_NUM        12
>>>>>    #define AML_UART_PORT_OFFSET        6
>>>>> @@ -76,6 +75,21 @@
>>>>>    #define AML_UART_POLL_USEC        5
>>>>>    #define AML_UART_TIMEOUT_USEC        10000
>>>>>    +struct meson_uart_data {
>>>>> +    struct uart_port    port;
>>>>> +    struct clk        *pclk;
>>>>> +    struct clk        *baud_clk;
>>>>> +    struct clk_divider    baud_div;
>>>>> +    struct clk_mux        use_xtal_mux;
>>>>> +    struct clk_mux        xtal_clk_sel_mux;
>>>>> +    struct clk_mux        xtal2_clk_sel_mux;
>>>>> +    struct clk_fixed_factor    xtal_div2;
>>>>> +    struct clk_fixed_factor    xtal_div3;
>>>>> +    struct clk_fixed_factor    clk81_div4;
>>>> Keeping all these internal elements around is not useful since they are
>>>> registered using devm_
>>>>
>>> I'm sorry. I don't know what you mean. That's exactly what you said, but
>>> what's wrong, I don't understand. Do you have any better suggestions,
>>> please specify specific points, preferably give examples.
>>
>> I'm saying that you don't need to keep reference to the internal
>> elements of the clock tree you have registered since devm_ will take
>> care of the removal later on. IOW, Once they are registered, the pointer
>> is never used again so you don't need it in the private data.
>>
> I understand you now. What you say is right. I will correct it.
>>
>>>>> +    bool            no_clk81_input;
>>>> What is this ?
>>>>
>>> To distinguish between clK81 and XTAL.
>>
>> ... Yet, it is not used
>>
>>>>> +    bool            has_xtal_clk_sel;
>>>>> +};
>>>>> +
>>>>>    static struct uart_driver meson_uart_driver;
>>>>>      static struct uart_port *meson_ports[AML_UART_PORT_NUM];
>>>>> @@ -270,14 +284,11 @@ static void meson_uart_reset(struct uart_port 
>>>>> *port)
>>>>>    static int meson_uart_startup(struct uart_port *port)
>>>>>    {
>>>>>        u32 val;
>>>>> -    int ret = 0;
>>>>> +    int ret;
>>>>>    -    val = readl(port->membase + AML_UART_CONTROL);
>>>>> -    val |= AML_UART_CLEAR_ERR;
>>>>> -    writel(val, port->membase + AML_UART_CONTROL);
>>>>> -    val &= ~AML_UART_CLEAR_ERR;
>>>>> -    writel(val, port->membase + AML_UART_CONTROL);
>>>>> +    meson_uart_reset(port);
>>>>>    +    val = readl(port->membase + AML_UART_CONTROL);
>>>>>        val |= (AML_UART_RX_EN | AML_UART_TX_EN);
>>>>>        writel(val, port->membase + AML_UART_CONTROL);
>>>>>    @@ -295,19 +306,17 @@ static int meson_uart_startup(struct 
>>>>> uart_port
>>>>> *port)
>>>>>      static void meson_uart_change_speed(struct uart_port *port, 
>>>>> unsigned
>>>>> long baud)
>>>>>    {
>>>>> +    struct meson_uart_data *private_data = port->private_data;
>>>>>        u32 val;
>>>>>          while (!meson_uart_tx_empty(port))
>>>>>            cpu_relax();
>>>>>    -    if (port->uartclk == 24000000) {
>>>> This check shows that previous code assumed bit 24 was left untouched
>>>> Below you can see that GXBB and newer used the XTAL path while older
>>>> used the other
>>>> Your change makes this dynamic which is another "unexpected" change.
>>>> Please make the bit 24 mux RO to start with so the behavior remains 
>>>> unchanged
>>>> for older SoCs.
>>> I agree with what you say, and I will correct it.
>>>>
>>>>> -        val = ((port->uartclk / 3) / baud) - 1;
>>>>> -        val |= AML_UART_BAUD_XTAL;
>>>>> -    } else {
>>>>> -        val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
>>>>> -    }
>>>>> +    val = readl(port->membase + AML_UART_REG5);
>>>>>        val |= AML_UART_BAUD_USE;
>>>>>        writel(val, port->membase + AML_UART_REG5);
>>>>> +
>>>>> +    clk_set_rate(private_data->baud_clk, baud);
>>>>>    }
>>>>>      static void meson_uart_set_termios(struct uart_port *port,
>>>>> @@ -397,11 +406,27 @@ static int meson_uart_verify_port(struct 
>>>>> uart_port *port,
>>>>>      static void meson_uart_release_port(struct uart_port *port)
>>>>>    {
>>>>> -    /* nothing to do */
>>>>> +    struct meson_uart_data *private_data = port->private_data;
>>>>> +
>>>>> +    clk_disable_unprepare(private_data->baud_clk);
>>>>> +    clk_disable_unprepare(private_data->pclk);
>>>>>    }
>>>>>      static int meson_uart_request_port(struct uart_port *port)
>>>>>    {
>>>>> +    struct meson_uart_data *private_data = port->private_data;
>>>>> +    int ret;
>>>>> +
>>>>> +    ret = clk_prepare_enable(private_data->pclk);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>> +
>>>>> +    ret = clk_prepare_enable(private_data->baud_clk);
>>>>> +    if (ret) {
>>>>> +        clk_disable_unprepare(private_data->pclk);
>>>>> +        return ret;
>>>>> +    }
>>>>> +
>>>>>        return 0;
>>>>>    }
>>>>>    @@ -629,56 +654,175 @@ static struct uart_driver 
>>>>> meson_uart_driver = {
>>>>>        .cons        = MESON_SERIAL_CONSOLE,
>>>>>    };
>>>>>    -static inline struct clk *meson_uart_probe_clock(struct device 
>>>>> *dev,
>>>>> -                         const char *id)
>>>>> +static int meson_uart_register_clk(struct uart_port *port,
>>>>> +                   const char *name_suffix,
>>>>> +                   const struct clk_parent_data *parent_data,
>>>>> +                   unsigned int num_parents,
>>>>> +                   const struct clk_ops *ops,
>>>>> +                   struct clk_hw *hw)
>>>>>    {
>>>>> -    struct clk *clk = NULL;
>>>>> +    struct clk_init_data init = { };
>>>>> +    char clk_name[32];
>>>>>        int ret;
>>>>>    -    clk = devm_clk_get(dev, id);
>>>>> -    if (IS_ERR(clk))
>>>>> -        return clk;
>>>>> +    snprintf(clk_name, sizeof(clk_name), "%s#%s", 
>>>>> dev_name(port->dev),
>>>>> +         name_suffix);
>>>>>    -    ret = clk_prepare_enable(clk);
>>>>> -    if (ret) {
>>>>> -        dev_err(dev, "couldn't enable clk\n");
>>>>> -        return ERR_PTR(ret);
>>>>> -    }
>>>>> +    init.name = clk_name;
>>>>> +    init.ops = ops;
>>>>> +    init.flags = CLK_SET_RATE_PARENT;
>>>>> +    init.parent_data = parent_data;
>>>>> +    init.num_parents = num_parents;
>>>>> +
>>>>> +    hw->init = &init;
>>>>>    -    devm_add_action_or_reset(dev,
>>>>> -            (void(*)(void *))clk_disable_unprepare,
>>>>> -            clk);
>>>>> +    ret = devm_clk_hw_register(port->dev, hw);
>>>>> +    if (ret)
>>>>> +        return dev_err_probe(port->dev, ret,
>>>>> +                     "Failed to register the '%s' clock\n",
>>>>> +                     clk_name);
>>>>>    -    return clk;
>>>>> +    return ret;
>>>>>    }
>>>>>    -static int meson_uart_probe_clocks(struct platform_device *pdev,
>>>>> -                   struct uart_port *port)
>>>>> -{
>>>>> -    struct clk *clk_xtal = NULL;
>>>>> -    struct clk *clk_pclk = NULL;
>>>>> -    struct clk *clk_baud = NULL;
>>>>> +static int meson_uart_probe_clocks(struct uart_port *port,
>>>>> +                   bool register_clk81_div4)
>>>>> +{
>>>>> +    struct meson_uart_data *private_data = port->private_data;
>>>>> +    struct clk_parent_data use_xtal_mux_parents[2] = {
>>>>> +        { .index = -1, },
>>>>> +        { .index = -1, },
>>>>> +    };
>>>>> +    struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
>>>>> +    struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
>>>>> +    struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
>>>>> +    struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
>>>>> +    struct clk_parent_data baud_div_parent = { };
>>>>> +    struct clk *clk_baud, *clk_xtal;
>>>>> +    int ret;
>>>>>    -    clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
>>>>> -    if (IS_ERR(clk_pclk))
>>>>> -        return PTR_ERR(clk_pclk);
>>>>> +    private_data->pclk = devm_clk_get(port->dev, "pclk");
>>>>> +    if (IS_ERR(private_data->pclk))
>>>>> +        return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
>>>>> +                     "Failed to get the 'pclk' clock\n");
>>>>> +
>>>>> +    clk_baud = devm_clk_get(port->dev, "baud");
>>>>> +    if (IS_ERR(clk_baud))
>>>>> +        return dev_err_probe(port->dev, PTR_ERR(clk_baud),
>>>>> +                     "Failed to get the 'baud' clock\n");
>>>>>    -    clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
>>>>> +    clk_xtal = devm_clk_get(port->dev, "xtal");
>>>>>        if (IS_ERR(clk_xtal))
>>>>> -        return PTR_ERR(clk_xtal);
>>>>> +        return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
>>>>> +                     "Failed to get the 'xtal' clock\n");
>>>>> +
>>>>> +    private_data->xtal_div3.mult = 1;
>>>>> +    private_data->xtal_div3.div = 3;
>>>>> +    ret = meson_uart_register_clk(port, "xtal_div3", 
>>>>> &xtal_div_parent,
>>>>> +                      1, &clk_fixed_factor_ops,
>>>>> +                      &private_data->xtal_div3.hw);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>>    -    clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
>>>>> -    if (IS_ERR(clk_baud))
>>>>> -        return PTR_ERR(clk_baud);
>>>>> +    if (register_clk81_div4) {
>>>> Clock dividers represent HW elements. The presence of an HW element
>>>> cannot
>>>> dependent on the current of the input clocks. There is no way this 
>>>> is right
>>>>
>>>>> +        private_data->clk81_div4.mult = 1;
>>>>> +        private_data->clk81_div4.div = 4;
>>>>> +        ret = meson_uart_register_clk(port, "clk81_div4",
>>>>> +                          &clk81_div_parent, 1,
>>>>> +                          &clk_fixed_factor_ops,
>>>>> +                          &private_data->clk81_div4.hw);
>>>>> +        if (ret)
>>>>> +            return ret;
>>>>> +
>>>>> +        use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;
>>>> This is only going to be used on meson8 and older. Worst case it should
>>>> depend
>>>> on this compatible but I don't think this would be right here.
>>>> IMO, this shows that the UART "baud" input was actually fed by a
>>>> derivation of "fclk_div4" instead of "clk_81" on these older SoCs.
>>>> So instead of registering what I suspect to be a fake element, the
>>>> meson8 clock controller driver and DT should be fixed a this should 
>>>> go away.
>>>>
>>> Virtually all UART controllers are supported, but they are not used 
>>> after
>>> Meson8. So the description is more reasonable in UART driver.
>>
>> No it is not. It's not virtual, the support for meson8 is there in
>> mainline.
>>
>> Unless you tell me there is an actual /4 divider in the UART block of
>> the meson8 and older and that is was removed from the newer SoC,
>> the bit above should be removed and the meson8 clock tree fixed.
>>
>> I think it is more probable that the /4 was never there to begin with,
>> and that clk81/4 aka "fclk_div4", which is already available from the
>> clock controller, was actually routed on these older SoCs.
> I will check with the chip design department if it is divided by 4. I'll 
> get back to you.
I confirmed with the chip design department that there is a 4 divider 
inside. So I'm going to use the divider 4 for the description, but only 
the older meson8 is in use, so I'm going to pass the parameter through 
compatible. Do you have any better suggestions?

>>
>>>>> +    }
>>>>>    -    port->uartclk = clk_get_rate(clk_baud);
>>>>> +    if (private_data->has_xtal_clk_sel) {
>>>>> +        private_data->xtal_div2.mult = 1;
>>>>> +        private_data->xtal_div2.div = 2;
>>>>> +        ret = meson_uart_register_clk(port, "xtal_div2",
>>>>> +                          &xtal_div_parent, 1,
>>>>> +                          &clk_fixed_factor_ops,
>>>>> +                          &private_data->xtal_div2.hw);
>>>>> +        if (ret)
>>>>> +            return ret;
>>>>> +
>>>>> +        xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
>>>>> +        xtal_clk_sel_mux_parents[1].fw_name = "xtal";
>>>>> +
>>>>> +        private_data->xtal_clk_sel_mux.reg = port->membase + 
>>>>> AML_UART_REG5;
>>>>> +        private_data->xtal_clk_sel_mux.mask = 0x1;
>>>>> +        private_data->xtal_clk_sel_mux.shift = 26;
>>>>> +        private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>>>> +        ret = meson_uart_register_clk(port, "xtal_clk_sel",
>>>>> +                          xtal_clk_sel_mux_parents,
>>>>> +                          ARRAY_SIZE(xtal_clk_sel_mux_parents),
>>>>> +                          &clk_mux_ops,
>>>>> +                          &private_data->xtal_clk_sel_mux.hw);
>>>>> +        if (ret)
>>>>> +            return ret;
>>>>> +
>>>>> +        xtal2_clk_sel_mux_parents[0].hw = 
>>>>> &private_data->xtal_clk_sel_mux.hw;
>>>>> +        xtal2_clk_sel_mux_parents[1].hw = 
>>>>> &private_data->xtal_div2.hw;
>>>>> +
>>>>> +        private_data->xtal2_clk_sel_mux.reg = port->membase + 
>>>>> AML_UART_REG5;
>>>>> +        private_data->xtal2_clk_sel_mux.mask = 0x1;
>>>>> +        private_data->xtal2_clk_sel_mux.shift = 27;
>>>>> +        private_data->xtal2_clk_sel_mux.flags = 
>>>>> CLK_MUX_ROUND_CLOSEST;
>>>>> +        ret = meson_uart_register_clk(port, "xtal2_clk_sel",
>>>>> +                          xtal2_clk_sel_mux_parents,
>>>>> +                          ARRAY_SIZE(xtal2_clk_sel_mux_parents),
>>>>> +                          &clk_mux_ops,
>>>>> +                          &private_data->xtal2_clk_sel_mux.hw);
>>>>> +        if (ret)
>>>>> +            return ret;
>>>>> +
>>>>> +        use_xtal_mux_parents[1].hw = 
>>>>> &private_data->xtal2_clk_sel_mux.hw;
>>>>> +    } else {
>>>>> +        use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;
>>>> Well the above is a bit over-complicated. If I summarize:
>>>> GXBB and older used a fixed divider of 3. Bits 26 and 27 read
>>>> 0 according to the documentation.
>>>> Chips after GXBB have 2 configurable divider of 2 and 3. bits 26 and 27
>>>> selects which of these dividers is used.
>>>> So the above could be replaced with a single divider covering bits 
>>>> 26 and
>>>> 27
>>>> with the following divider table { 2, 2, 1, 3 }. The divider should use
>>>> RO ops and not have CLK_SET_RATE_PARENT. It can be used for all chips
>>>> variant like this, including the older ones.
>>>>
>>> As you say, the way you say it is complicated. I think you're 
>>> complicating
>>> things.I don't understand what you're trying to achieve.
>>> Can you be more specific?
>>
>> I'm proposing to replace the 4 elements above, paths, parent_table and
>> all with a single divider with a specific table. It accomplishes the 
>> same thing.
>>
>> |--------+--------+-----------|
>> | Bit 27 | Bit 26 | Div Value |
>> |--------+--------+-----------|
>> |      0 |      0 |         3 |
>> |--------+--------+-----------|
>> |      0 |      1 |         1 |
>> |--------+--------+-----------|
>> |      1 |      0 |         2 |
>> |--------+--------+-----------|
>> |      1 |      1 |         2 |
>> |--------+--------+-----------|
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/clk-provider.h?h=v5.16-rc8#n602 
>>
>>
>> something like:
>>
>> static struct clk_div_table reg5_table[] = {
>>     { 0, 3 },
>>          { 1, 1 },
>>          { 2, 2 },
>>          { 3, 2 },
>> };
>>
>> I'm also pointing out that, since bit 26 and 27 reads 0 on the older
>> SoCs, the divider would work for these too, as long as the Ops is
>> read-only. So you don't even need to register different element
>> depending on the SoC which simplify things a bit more.
>>
>> If you prefer to keep a fixed divider for the older ones, It's fine by
>> me. It gives a bit more work but it is closer to reality. The newer SoC
>> could still use the custom divider regardless.
>>
>> Last, and as explained on other bits, that part of the clock path should
>> remain RO at first, to keep behavior stable with this change. You may
>> change that later on, in a dedicated patch describing the change.
>>
> I understand what you mean, and i will try to realize it. I hope you can 
> help review and give me your suggestions.
>>>
>>>> The only information you need to carry if whether or not you want to
>>>> make this divider modifiable. This means using the dt data to store
>>>> clk_div_ops or clk_div_ro_ops pointer.
>>>> To avoid changing the behavior of the older platforms in this patch, I
>>>> would
>>>> suggest to make everything use clk_div_ro_ops first, and make another
>>>> patch to use clk_div_ops if necessary.
>>>>
>>>>> +    }
>>>>> +
>>>>> +    private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
>>>>> +    private_data->use_xtal_mux.mask = 0x1;
>>>>> +    private_data->use_xtal_mux.shift = 24;
>>>>> +    private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
>>>>> +    ret = meson_uart_register_clk(port, "use_xtal", 
>>>>> use_xtal_mux_parents,
>>>>> +                      ARRAY_SIZE(use_xtal_mux_parents),
>>>>> +                      &clk_mux_ops,
>>>> Use RO ops here to start with.
>>>> You can make this writable with another patch explicitly describing 
>>>> the change.
>>> I agree with what you say, and I will correct it.
>>>>
>>>>> +                      &private_data->use_xtal_mux.hw);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>> +
>>>>> +    baud_div_parent.hw = &private_data->use_xtal_mux.hw;
>>>>> +
>>>>> +    private_data->baud_div.reg = port->membase + AML_UART_REG5;
>>>>> +    private_data->baud_div.shift = 0;
>>>>> +    private_data->baud_div.width = 23;
>>>>> +    private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
>>>>> +    ret = meson_uart_register_clk(port, "baud_div",
>>>>> +                      &baud_div_parent, 1,
>>>>> +                      &clk_divider_ops,
>>>>> +                      &private_data->baud_div.hw);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>> +
>>>>> +    private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
>>>>> +                             &private_data->baud_div.hw,
>>>>> +                             "baud_rate");
>>>> There is a problem with this function in CCF. It will pin the driver to
>>>> itself, making it unremovable. It is an ongoing topic. For now, just 
>>>> use
>>>> "hw->clk" to get the struct *clk.
>>>>
>>>>> +    if (IS_ERR(private_data->baud_clk))
>>>>> +        return dev_err_probe(port->dev,
>>>> I don't think anything can defer here, so dev_err_probe() is not
>>>> necessary I think
>>>>
>>> I agree with what you say, and I will correct it.
>>>>> +                     PTR_ERR(private_data->baud_clk),
>>>>> +                     "Failed to request the 'baud_rate' clock\n");
>>>>>          return 0;
>>>>>    }
>>>>>      static int meson_uart_probe(struct platform_device *pdev)
>>>>>    {
>>>>> +    struct meson_uart_data *private_data;
>>>>>        struct resource *res_mem, *res_irq;
>>>>> +    struct clk *clk_baud, *clk_xtal;
>>>>> +    bool register_clk81_div4;
>>>>>        struct uart_port *port;
>>>>>        int ret = 0;
>>>>>        int id = -1;
>>>>> @@ -711,18 +855,37 @@ static int meson_uart_probe(struct 
>>>>> platform_device *pdev)
>>>>>            return -EBUSY;
>>>>>        }
>>>>>    -    port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port),
>>>>> GFP_KERNEL);
>>>>> -    if (!port)
>>>>> +    private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
>>>>> +                    GFP_KERNEL);
>>>>> +    if (!private_data)
>>>>>            return -ENOMEM;
>>>>>    +    if (device_get_match_data(&pdev->dev))
>>>>> +        private_data->has_xtal_clk_sel = true;
>>>>> +
>>>>> +    private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
>>>>> +    if (IS_ERR(private_data->pclk))
>>>>> +        return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
>>>>> +                     "Failed to get the 'pclk' clock\n");
>>>>> +
>>>>> +    clk_baud = devm_clk_get(&pdev->dev, "baud");
>>>>> +    if (IS_ERR(clk_baud))
>>>>> +        return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
>>>>> +                     "Failed to get the 'baud' clock\n");
>>>>> +
>>>>> +    clk_xtal = devm_clk_get(&pdev->dev, "xtal");
>>>>> +    if (IS_ERR(clk_xtal))
>>>>> +        return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
>>>>> +                     "Failed to get the 'xtal' clock\n");
>>>>> +
>>>> This is second time you call devm_clk_get() on these clocks. One
>>>> instance has to go away
>>>>
>>> I agree with what you say, and I will correct it.
>>>>> +    register_clk81_div4 = clk_get_rate(clk_xtal) != 
>>>>> clk_get_rate(clk_baud);
>>>>> +
>>>> The above is a ugly way to distinguish the meson8 (32bit) SoC family
>>>> from the rest. This definitely not the way to achieve it.
>>>> The right is the compatible data but here I think it is not
>>>> necessary. The clock input should be fixed.
>>>>
>>>>> +    port = &private_data->port;
>>>>> +
>>>>>        port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
>>>>>        if (IS_ERR(port->membase))
>>>>>            return PTR_ERR(port->membase);
>>>>>    -    ret = meson_uart_probe_clocks(pdev, port);
>>>>> -    if (ret)
>>>>> -        return ret;
>>>>> -
>>>>>        port->iotype = UPIO_MEM;
>>>>>        port->mapbase = res_mem->start;
>>>>>        port->mapsize = resource_size(res_mem);
>>>>> @@ -735,6 +898,12 @@ static int meson_uart_probe(struct 
>>>>> platform_device *pdev)
>>>>>        port->x_char = 0;
>>>>>        port->ops = &meson_uart_ops;
>>>>>        port->fifosize = 64;
>>>>> +    port->uartclk = clk_get_rate(clk_baud);
>>>>> +    port->private_data = private_data;
>>>>> +
>>>>> +    ret = meson_uart_probe_clocks(port, register_clk81_div4);
>>>>> +    if (ret)
>>>>> +        return ret;
>>>>>          meson_ports[pdev->id] = port;
>>>>>        platform_set_drvdata(pdev, port);
>>>>> @@ -761,10 +930,42 @@ static int meson_uart_remove(struct 
>>>>> platform_device *pdev)
>>>>>    }
>>>>>      static const struct of_device_id meson_uart_dt_match[] = {
>>>>> -    { .compatible = "amlogic,meson6-uart" },
>>>>> -    { .compatible = "amlogic,meson8-uart" },
>>>>> -    { .compatible = "amlogic,meson8b-uart" },
>>>>> -    { .compatible = "amlogic,meson-gx-uart" },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson6-uart",
>>>>> +        .data = (void *)false,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson8-uart",
>>>>> +        .data = (void *)false,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson8b-uart",
>>>>> +        .data = (void *)false,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson-gxbb-uart",
>>>>> +        .data = (void *)false,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson-gxl-uart",
>>>>> +        .data = (void *)true,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson-g12a-uart",
>>>>> +        .data = (void *)true,
>>>>> +    },
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson-s4-uart",
>>>>> +        .data = (void *)true,
>>>>> +    },
>>>>> +    /*
>>>>> +     * deprecated, don't use anymore because it doesn't differentiate
>>>>> +     * between GXBB and GXL which have different revisions of the 
>>>>> UART IP.
>>>>> +     */
>>>>> +    {
>>>>> +        .compatible = "amlogic,meson-gx-uart",
>>>>> +        .data = (void *)false,
>>>>> +    },
>>>>>        { /* sentinel */ },
>>>>>    };
>>>>>    MODULE_DEVICE_TABLE(of, meson_uart_dt_match);
>>>>
>>
diff mbox series

Patch

diff --git a/drivers/tty/serial/Kconfig b/drivers/tty/serial/Kconfig
index 780908d43557..32e238173036 100644
--- a/drivers/tty/serial/Kconfig
+++ b/drivers/tty/serial/Kconfig
@@ -198,6 +198,7 @@  config SERIAL_KGDB_NMI
 config SERIAL_MESON
 	tristate "Meson serial port support"
 	depends on ARCH_MESON
+	depends on COMMON_CLK
 	select SERIAL_CORE
 	help
 	  This enables the driver for the on-chip UARTs of the Amlogic
diff --git a/drivers/tty/serial/meson_uart.c b/drivers/tty/serial/meson_uart.c
index 99efe62a1507..07eb1f40aaaa 100644
--- a/drivers/tty/serial/meson_uart.c
+++ b/drivers/tty/serial/meson_uart.c
@@ -6,6 +6,7 @@ 
  */
 
 #include <linux/clk.h>
+#include <linux/clk-provider.h>
 #include <linux/console.h>
 #include <linux/delay.h>
 #include <linux/init.h>
@@ -65,9 +66,7 @@ 
 #define AML_UART_RECV_IRQ(c)		((c) & 0xff)
 
 /* AML_UART_REG5 bits */
-#define AML_UART_BAUD_MASK		0x7fffff
 #define AML_UART_BAUD_USE		BIT(23)
-#define AML_UART_BAUD_XTAL		BIT(24)
 
 #define AML_UART_PORT_NUM		12
 #define AML_UART_PORT_OFFSET		6
@@ -76,6 +75,21 @@ 
 #define AML_UART_POLL_USEC		5
 #define AML_UART_TIMEOUT_USEC		10000
 
+struct meson_uart_data {
+	struct uart_port	port;
+	struct clk		*pclk;
+	struct clk		*baud_clk;
+	struct clk_divider	baud_div;
+	struct clk_mux		use_xtal_mux;
+	struct clk_mux		xtal_clk_sel_mux;
+	struct clk_mux		xtal2_clk_sel_mux;
+	struct clk_fixed_factor	xtal_div2;
+	struct clk_fixed_factor	xtal_div3;
+	struct clk_fixed_factor	clk81_div4;
+	bool			no_clk81_input;
+	bool			has_xtal_clk_sel;
+};
+
 static struct uart_driver meson_uart_driver;
 
 static struct uart_port *meson_ports[AML_UART_PORT_NUM];
@@ -270,14 +284,11 @@  static void meson_uart_reset(struct uart_port *port)
 static int meson_uart_startup(struct uart_port *port)
 {
 	u32 val;
-	int ret = 0;
+	int ret;
 
-	val = readl(port->membase + AML_UART_CONTROL);
-	val |= AML_UART_CLEAR_ERR;
-	writel(val, port->membase + AML_UART_CONTROL);
-	val &= ~AML_UART_CLEAR_ERR;
-	writel(val, port->membase + AML_UART_CONTROL);
+	meson_uart_reset(port);
 
+	val = readl(port->membase + AML_UART_CONTROL);
 	val |= (AML_UART_RX_EN | AML_UART_TX_EN);
 	writel(val, port->membase + AML_UART_CONTROL);
 
@@ -295,19 +306,17 @@  static int meson_uart_startup(struct uart_port *port)
 
 static void meson_uart_change_speed(struct uart_port *port, unsigned long baud)
 {
+	struct meson_uart_data *private_data = port->private_data;
 	u32 val;
 
 	while (!meson_uart_tx_empty(port))
 		cpu_relax();
 
-	if (port->uartclk == 24000000) {
-		val = ((port->uartclk / 3) / baud) - 1;
-		val |= AML_UART_BAUD_XTAL;
-	} else {
-		val = ((port->uartclk * 10 / (baud * 4) + 5) / 10) - 1;
-	}
+	val = readl(port->membase + AML_UART_REG5);
 	val |= AML_UART_BAUD_USE;
 	writel(val, port->membase + AML_UART_REG5);
+
+	clk_set_rate(private_data->baud_clk, baud);
 }
 
 static void meson_uart_set_termios(struct uart_port *port,
@@ -397,11 +406,27 @@  static int meson_uart_verify_port(struct uart_port *port,
 
 static void meson_uart_release_port(struct uart_port *port)
 {
-	/* nothing to do */
+	struct meson_uart_data *private_data = port->private_data;
+
+	clk_disable_unprepare(private_data->baud_clk);
+	clk_disable_unprepare(private_data->pclk);
 }
 
 static int meson_uart_request_port(struct uart_port *port)
 {
+	struct meson_uart_data *private_data = port->private_data;
+	int ret;
+
+	ret = clk_prepare_enable(private_data->pclk);
+	if (ret)
+		return ret;
+
+	ret = clk_prepare_enable(private_data->baud_clk);
+	if (ret) {
+		clk_disable_unprepare(private_data->pclk);
+		return ret;
+	}
+
 	return 0;
 }
 
@@ -629,56 +654,175 @@  static struct uart_driver meson_uart_driver = {
 	.cons		= MESON_SERIAL_CONSOLE,
 };
 
-static inline struct clk *meson_uart_probe_clock(struct device *dev,
-						 const char *id)
+static int meson_uart_register_clk(struct uart_port *port,
+				   const char *name_suffix,
+				   const struct clk_parent_data *parent_data,
+				   unsigned int num_parents,
+				   const struct clk_ops *ops,
+				   struct clk_hw *hw)
 {
-	struct clk *clk = NULL;
+	struct clk_init_data init = { };
+	char clk_name[32];
 	int ret;
 
-	clk = devm_clk_get(dev, id);
-	if (IS_ERR(clk))
-		return clk;
+	snprintf(clk_name, sizeof(clk_name), "%s#%s", dev_name(port->dev),
+		 name_suffix);
 
-	ret = clk_prepare_enable(clk);
-	if (ret) {
-		dev_err(dev, "couldn't enable clk\n");
-		return ERR_PTR(ret);
-	}
+	init.name = clk_name;
+	init.ops = ops;
+	init.flags = CLK_SET_RATE_PARENT;
+	init.parent_data = parent_data;
+	init.num_parents = num_parents;
+
+	hw->init = &init;
 
-	devm_add_action_or_reset(dev,
-			(void(*)(void *))clk_disable_unprepare,
-			clk);
+	ret = devm_clk_hw_register(port->dev, hw);
+	if (ret)
+		return dev_err_probe(port->dev, ret,
+				     "Failed to register the '%s' clock\n",
+				     clk_name);
 
-	return clk;
+	return ret;
 }
 
-static int meson_uart_probe_clocks(struct platform_device *pdev,
-				   struct uart_port *port)
-{
-	struct clk *clk_xtal = NULL;
-	struct clk *clk_pclk = NULL;
-	struct clk *clk_baud = NULL;
+static int meson_uart_probe_clocks(struct uart_port *port,
+				   bool register_clk81_div4)
+{
+	struct meson_uart_data *private_data = port->private_data;
+	struct clk_parent_data use_xtal_mux_parents[2] = {
+		{ .index = -1, },
+		{ .index = -1, },
+	};
+	struct clk_parent_data xtal_clk_sel_mux_parents[2] = { };
+	struct clk_parent_data xtal2_clk_sel_mux_parents[2] = { };
+	struct clk_parent_data xtal_div_parent = { .fw_name = "xtal", };
+	struct clk_parent_data clk81_div_parent = { .fw_name = "baud", };
+	struct clk_parent_data baud_div_parent = { };
+	struct clk *clk_baud, *clk_xtal;
+	int ret;
 
-	clk_pclk = meson_uart_probe_clock(&pdev->dev, "pclk");
-	if (IS_ERR(clk_pclk))
-		return PTR_ERR(clk_pclk);
+	private_data->pclk = devm_clk_get(port->dev, "pclk");
+	if (IS_ERR(private_data->pclk))
+		return dev_err_probe(port->dev, PTR_ERR(private_data->pclk),
+				     "Failed to get the 'pclk' clock\n");
+
+	clk_baud = devm_clk_get(port->dev, "baud");
+	if (IS_ERR(clk_baud))
+		return dev_err_probe(port->dev, PTR_ERR(clk_baud),
+				     "Failed to get the 'baud' clock\n");
 
-	clk_xtal = meson_uart_probe_clock(&pdev->dev, "xtal");
+	clk_xtal = devm_clk_get(port->dev, "xtal");
 	if (IS_ERR(clk_xtal))
-		return PTR_ERR(clk_xtal);
+		return dev_err_probe(port->dev, PTR_ERR(clk_xtal),
+				     "Failed to get the 'xtal' clock\n");
+
+	private_data->xtal_div3.mult = 1;
+	private_data->xtal_div3.div = 3;
+	ret = meson_uart_register_clk(port, "xtal_div3", &xtal_div_parent,
+				      1, &clk_fixed_factor_ops,
+				      &private_data->xtal_div3.hw);
+	if (ret)
+		return ret;
 
-	clk_baud = meson_uart_probe_clock(&pdev->dev, "baud");
-	if (IS_ERR(clk_baud))
-		return PTR_ERR(clk_baud);
+	if (register_clk81_div4) {
+		private_data->clk81_div4.mult = 1;
+		private_data->clk81_div4.div = 4;
+		ret = meson_uart_register_clk(port, "clk81_div4",
+					      &clk81_div_parent, 1,
+					      &clk_fixed_factor_ops,
+					      &private_data->clk81_div4.hw);
+		if (ret)
+			return ret;
+
+		use_xtal_mux_parents[0].hw = &private_data->clk81_div4.hw;
+	}
 
-	port->uartclk = clk_get_rate(clk_baud);
+	if (private_data->has_xtal_clk_sel) {
+		private_data->xtal_div2.mult = 1;
+		private_data->xtal_div2.div = 2;
+		ret = meson_uart_register_clk(port, "xtal_div2",
+					      &xtal_div_parent, 1,
+					      &clk_fixed_factor_ops,
+					      &private_data->xtal_div2.hw);
+		if (ret)
+			return ret;
+
+		xtal_clk_sel_mux_parents[0].hw = &private_data->xtal_div3.hw;
+		xtal_clk_sel_mux_parents[1].fw_name = "xtal";
+
+		private_data->xtal_clk_sel_mux.reg = port->membase + AML_UART_REG5;
+		private_data->xtal_clk_sel_mux.mask = 0x1;
+		private_data->xtal_clk_sel_mux.shift = 26;
+		private_data->xtal_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
+		ret = meson_uart_register_clk(port, "xtal_clk_sel",
+					      xtal_clk_sel_mux_parents,
+					      ARRAY_SIZE(xtal_clk_sel_mux_parents),
+					      &clk_mux_ops,
+					      &private_data->xtal_clk_sel_mux.hw);
+		if (ret)
+			return ret;
+
+		xtal2_clk_sel_mux_parents[0].hw = &private_data->xtal_clk_sel_mux.hw;
+		xtal2_clk_sel_mux_parents[1].hw = &private_data->xtal_div2.hw;
+
+		private_data->xtal2_clk_sel_mux.reg = port->membase + AML_UART_REG5;
+		private_data->xtal2_clk_sel_mux.mask = 0x1;
+		private_data->xtal2_clk_sel_mux.shift = 27;
+		private_data->xtal2_clk_sel_mux.flags = CLK_MUX_ROUND_CLOSEST;
+		ret = meson_uart_register_clk(port, "xtal2_clk_sel",
+					      xtal2_clk_sel_mux_parents,
+					      ARRAY_SIZE(xtal2_clk_sel_mux_parents),
+					      &clk_mux_ops,
+					      &private_data->xtal2_clk_sel_mux.hw);
+		if (ret)
+			return ret;
+
+		use_xtal_mux_parents[1].hw = &private_data->xtal2_clk_sel_mux.hw;
+	} else {
+		use_xtal_mux_parents[1].hw = &private_data->xtal_div3.hw;
+	}
+
+	private_data->use_xtal_mux.reg = port->membase + AML_UART_REG5;
+	private_data->use_xtal_mux.mask = 0x1;
+	private_data->use_xtal_mux.shift = 24;
+	private_data->use_xtal_mux.flags = CLK_MUX_ROUND_CLOSEST;
+	ret = meson_uart_register_clk(port, "use_xtal", use_xtal_mux_parents,
+				      ARRAY_SIZE(use_xtal_mux_parents),
+				      &clk_mux_ops,
+				      &private_data->use_xtal_mux.hw);
+	if (ret)
+		return ret;
+
+	baud_div_parent.hw = &private_data->use_xtal_mux.hw;
+
+	private_data->baud_div.reg = port->membase + AML_UART_REG5;
+	private_data->baud_div.shift = 0;
+	private_data->baud_div.width = 23;
+	private_data->baud_div.flags = CLK_DIVIDER_ROUND_CLOSEST;
+	ret = meson_uart_register_clk(port, "baud_div",
+				      &baud_div_parent, 1,
+				      &clk_divider_ops,
+				      &private_data->baud_div.hw);
+	if (ret)
+		return ret;
+
+	private_data->baud_clk = devm_clk_hw_get_clk(port->dev,
+						     &private_data->baud_div.hw,
+						     "baud_rate");
+	if (IS_ERR(private_data->baud_clk))
+		return dev_err_probe(port->dev,
+				     PTR_ERR(private_data->baud_clk),
+				     "Failed to request the 'baud_rate' clock\n");
 
 	return 0;
 }
 
 static int meson_uart_probe(struct platform_device *pdev)
 {
+	struct meson_uart_data *private_data;
 	struct resource *res_mem, *res_irq;
+	struct clk *clk_baud, *clk_xtal;
+	bool register_clk81_div4;
 	struct uart_port *port;
 	int ret = 0;
 	int id = -1;
@@ -711,18 +855,37 @@  static int meson_uart_probe(struct platform_device *pdev)
 		return -EBUSY;
 	}
 
-	port = devm_kzalloc(&pdev->dev, sizeof(struct uart_port), GFP_KERNEL);
-	if (!port)
+	private_data = devm_kzalloc(&pdev->dev, sizeof(*private_data),
+				    GFP_KERNEL);
+	if (!private_data)
 		return -ENOMEM;
 
+	if (device_get_match_data(&pdev->dev))
+		private_data->has_xtal_clk_sel = true;
+
+	private_data->pclk = devm_clk_get(&pdev->dev, "pclk");
+	if (IS_ERR(private_data->pclk))
+		return dev_err_probe(&pdev->dev, PTR_ERR(private_data->pclk),
+				     "Failed to get the 'pclk' clock\n");
+
+	clk_baud = devm_clk_get(&pdev->dev, "baud");
+	if (IS_ERR(clk_baud))
+		return dev_err_probe(&pdev->dev, PTR_ERR(clk_baud),
+				     "Failed to get the 'baud' clock\n");
+
+	clk_xtal = devm_clk_get(&pdev->dev, "xtal");
+	if (IS_ERR(clk_xtal))
+		return dev_err_probe(&pdev->dev, PTR_ERR(clk_xtal),
+				     "Failed to get the 'xtal' clock\n");
+
+	register_clk81_div4 = clk_get_rate(clk_xtal) != clk_get_rate(clk_baud);
+
+	port = &private_data->port;
+
 	port->membase = devm_ioremap_resource(&pdev->dev, res_mem);
 	if (IS_ERR(port->membase))
 		return PTR_ERR(port->membase);
 
-	ret = meson_uart_probe_clocks(pdev, port);
-	if (ret)
-		return ret;
-
 	port->iotype = UPIO_MEM;
 	port->mapbase = res_mem->start;
 	port->mapsize = resource_size(res_mem);
@@ -735,6 +898,12 @@  static int meson_uart_probe(struct platform_device *pdev)
 	port->x_char = 0;
 	port->ops = &meson_uart_ops;
 	port->fifosize = 64;
+	port->uartclk = clk_get_rate(clk_baud);
+	port->private_data = private_data;
+
+	ret = meson_uart_probe_clocks(port, register_clk81_div4);
+	if (ret)
+		return ret;
 
 	meson_ports[pdev->id] = port;
 	platform_set_drvdata(pdev, port);
@@ -761,10 +930,42 @@  static int meson_uart_remove(struct platform_device *pdev)
 }
 
 static const struct of_device_id meson_uart_dt_match[] = {
-	{ .compatible = "amlogic,meson6-uart" },
-	{ .compatible = "amlogic,meson8-uart" },
-	{ .compatible = "amlogic,meson8b-uart" },
-	{ .compatible = "amlogic,meson-gx-uart" },
+	{
+		.compatible = "amlogic,meson6-uart",
+		.data = (void *)false,
+	},
+	{
+		.compatible = "amlogic,meson8-uart",
+		.data = (void *)false,
+	},
+	{
+		.compatible = "amlogic,meson8b-uart",
+		.data = (void *)false,
+	},
+	{
+		.compatible = "amlogic,meson-gxbb-uart",
+		.data = (void *)false,
+	},
+	{
+		.compatible = "amlogic,meson-gxl-uart",
+		.data = (void *)true,
+	},
+	{
+		.compatible = "amlogic,meson-g12a-uart",
+		.data = (void *)true,
+	},
+	{
+		.compatible = "amlogic,meson-s4-uart",
+		.data = (void *)true,
+	},
+	/*
+	 * deprecated, don't use anymore because it doesn't differentiate
+	 * between GXBB and GXL which have different revisions of the UART IP.
+	 */
+	{
+		.compatible = "amlogic,meson-gx-uart",
+		.data = (void *)false,
+	},
 	{ /* sentinel */ },
 };
 MODULE_DEVICE_TABLE(of, meson_uart_dt_match);