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

From: Yu Tu
Date: Wed Jan 05 2022 - 00:54:08 EST


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@xxxxxxxxxxx> 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@xxxxxxxxxxx> 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@xxxxxxxxxxx>
---
   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);