If a physical serial port device driver uses arch_initcall() we fail to
probe the serial_base_bus devices and the serial port tx fails. This is
because as serial_base_bus uses module_initcall().

Let's fix the issue by changing serial_base_bus to use arch_initcall().

Let's also return an error if a driver attempts to call uart_add_one_port()
too early.

Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Closes: https://lore.kernel.org/linux-serial/20230601132012.GB14287@atomide.com/T/#m6a40440fc04d551d27b147da8602e065c982a115
Fixes: 84a9582fd203 ("serial: core: Start managing serial controllers to enable runtime PM")
Signed-off-by: Tony Lindgren <tony@atomide.com>
---
 drivers/tty/serial/serial_base_bus.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/drivers/tty/serial/serial_base_bus.c b/drivers/tty/serial/serial_base_bus.c
--- a/drivers/tty/serial/serial_base_bus.c
+++ b/drivers/tty/serial/serial_base_bus.c
@@ -17,6 +17,8 @@
 
 #include "serial_base.h"
 
+static bool serial_base_initialized;
+
 static int serial_base_match(struct device *dev, struct device_driver *drv)
 {
 	int len = strlen(drv->name);
@@ -48,6 +50,11 @@ static int serial_base_device_init(struct uart_port *port,
 				   void (*release)(struct device *dev),
 				   int id)
 {
+	if (!serial_base_initialized) {
+		dev_err(port->dev, "uart_add_one_port() called before arch_initcall()?\n");
+		return -EPROBE_DEFER;
+	}
+
 	device_initialize(dev);
 	dev->type = type;
 	dev->parent = parent_dev;
@@ -175,6 +182,8 @@ static int serial_base_init(void)
 	if (ret)
 		goto err_ctrl_exit;
 
+	serial_base_initialized = true;
+
 	return 0;
 
 err_ctrl_exit:
@@ -185,7 +194,7 @@ static int serial_base_init(void)
 
 	return ret;
 }
-module_init(serial_base_init);
+arch_initcall(serial_base_init);
 
 static void serial_base_exit(void)
 {
-- 
2.40.1

On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
> If a physical serial port device driver uses arch_initcall() we fail to
> probe the serial_base_bus devices and the serial port tx fails. This is
> because as serial_base_bus uses module_initcall().
> 
> Let's fix the issue by changing serial_base_bus to use arch_initcall().

This will only work if the linking order is such that this will always
come before the drivers.  Is that the case here?

thanks,

greg k-h

* Greg Kroah-Hartman <gregkh@linuxfoundation.org> [230601 14:21]:
> On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
> > If a physical serial port device driver uses arch_initcall() we fail to
> > probe the serial_base_bus devices and the serial port tx fails. This is
> > because as serial_base_bus uses module_initcall().
> > 
> > Let's fix the issue by changing serial_base_bus to use arch_initcall().
> 
> This will only work if the linking order is such that this will always
> come before the drivers.  Is that the case here?

I guess based on Makefile. And also if serial drivers are modules as we
export uart_add_one_port() from serial_base.ko. But yeah this is pretty
fragile potentially.

Hmm maybe we could keep module_init() and then also call serial_base_init()
on uart_add_one_port() path if not yet initialized?

Probably the module_init() should be still there for case when no serial
port device drivers are loaded and serial_base is unloaded..

Regards,

Tony

On Thu, Jun 01, 2023 at 06:07:06PM +0300, Tony Lindgren wrote:
> * Greg Kroah-Hartman <gregkh@linuxfoundation.org> [230601 14:21]:
> > On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
> > > If a physical serial port device driver uses arch_initcall() we fail to
> > > probe the serial_base_bus devices and the serial port tx fails. This is
> > > because as serial_base_bus uses module_initcall().
> > > 
> > > Let's fix the issue by changing serial_base_bus to use arch_initcall().
> > 
> > This will only work if the linking order is such that this will always
> > come before the drivers.  Is that the case here?
> 
> I guess based on Makefile. And also if serial drivers are modules as we
> export uart_add_one_port() from serial_base.ko. But yeah this is pretty
> fragile potentially.

It's fine, and normal, the Makefile is the ordering here so all is good.

> Hmm maybe we could keep module_init() and then also call serial_base_init()
> on uart_add_one_port() path if not yet initialized?
> 
> Probably the module_init() should be still there for case when no serial
> port device drivers are loaded and serial_base is unloaded..

I'll leave this as-is, it looks correct, and I'll queue it up now,
thanks.

greg k-h

On 01.06.2023 16:21, Greg Kroah-Hartman wrote:
> On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
>> If a physical serial port device driver uses arch_initcall() we fail to
>> probe the serial_base_bus devices and the serial port tx fails. This is
>> because as serial_base_bus uses module_initcall().
>>
>> Let's fix the issue by changing serial_base_bus to use arch_initcall().
> This will only work if the linking order is such that this will always
> come before the drivers.  Is that the case here?

Yes, serial_base_bus is linked as a second object, just after the 
serial_core. Device drivers come later.


Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland

* Marek Szyprowski <m.szyprowski@samsung.com> [230601 15:09]:
> On 01.06.2023 16:21, Greg Kroah-Hartman wrote:
> > On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
> >> If a physical serial port device driver uses arch_initcall() we fail to
> >> probe the serial_base_bus devices and the serial port tx fails. This is
> >> because as serial_base_bus uses module_initcall().
> >>
> >> Let's fix the issue by changing serial_base_bus to use arch_initcall().
> > This will only work if the linking order is such that this will always
> > come before the drivers.  Is that the case here?
> 
> Yes, serial_base_bus is linked as a second object, just after the 
> serial_core. Device drivers come later.

If we don't want to rely on the Makefile order here, we could do something
like the patch below.

Regards,

Tony

8< ---------------------
diff --git a/drivers/tty/serial/serial_base_bus.c b/drivers/tty/serial/serial_base_bus.c
--- a/drivers/tty/serial/serial_base_bus.c
+++ b/drivers/tty/serial/serial_base_bus.c
@@ -11,12 +11,18 @@
 #include <linux/container_of.h>
 #include <linux/device.h>
 #include <linux/module.h>
+#include <linux/mutex.h>
 #include <linux/serial_core.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 
 #include "serial_base.h"
 
+static bool serial_base_initialized;
+static DEFINE_MUTEX(serial_base_lock);
+
+static int serial_base_init(void);
+
 static int serial_base_match(struct device *dev, struct device_driver *drv)
 {
 	int len = strlen(drv->name);
@@ -48,6 +54,12 @@ static int serial_base_device_init(struct uart_port *port,
 				   void (*release)(struct device *dev),
 				   int id)
 {
+	/*
+	 * Initialize bus if not yet initialized. Some drivers are using
+	 * arch_initcall()
+	 */
+	serial_base_init();
+
 	device_initialize(dev);
 	dev->type = type;
 	dev->parent = parent_dev;
@@ -163,6 +175,14 @@ static int serial_base_init(void)
 {
 	int ret;
 
+	mutex_lock(&serial_base_lock);
+
+	/* Also called from serial_base_device_init() in some cases */
+	if (serial_base_initialized) {
+		ret = 0;
+		goto out_unlock;
+	}
+
 	ret = bus_register(&serial_base_bus_type);
 	if (ret)
 		return ret;
@@ -175,6 +195,10 @@ static int serial_base_init(void)
 	if (ret)
 		goto err_ctrl_exit;
 
+	serial_base_initialized = true;
+
+	mutex_unlock(&serial_base_lock);
+
 	return 0;
 
 err_ctrl_exit:
@@ -183,15 +207,21 @@ static int serial_base_init(void)
 err_bus_unregister:
 	bus_unregister(&serial_base_bus_type);
 
+out_unlock:
+	mutex_unlock(&serial_base_lock);
+
 	return ret;
 }
 module_init(serial_base_init);
 
 static void serial_base_exit(void)
 {
+	mutex_lock(&serial_base_lock);
 	serial_base_port_exit();
 	serial_base_ctrl_exit();
 	bus_unregister(&serial_base_bus_type);
+	serial_base_initialized = false;
+	mutex_unlock(&serial_base_lock);
 }
 module_exit(serial_base_exit);
 
-- 
2.40.1

On Thu, Jun 01, 2023 at 05:09:52PM +0200, Marek Szyprowski wrote:
> On 01.06.2023 16:21, Greg Kroah-Hartman wrote:
> > On Thu, Jun 01, 2023 at 05:14:44PM +0300, Tony Lindgren wrote:
> >> If a physical serial port device driver uses arch_initcall() we fail to
> >> probe the serial_base_bus devices and the serial port tx fails. This is
> >> because as serial_base_bus uses module_initcall().
> >>
> >> Let's fix the issue by changing serial_base_bus to use arch_initcall().
> > This will only work if the linking order is such that this will always
> > come before the drivers.  Is that the case here?
> 
> Yes, serial_base_bus is linked as a second object, just after the 
> serial_core. Device drivers come later.

Oh good, I guess it wouldn't work at all as the serial_core is needed by
all of those drivers first too, so this should work, thanks for
checking.

greg k-h