Add support for DEVNAME:0.0 style hardware based addressing

Hi all,

With the recent serial core changes, we can now add DEVNAME:0.0 style addressing for the serial ports. When using DEVNAME:0.0 naming, we don't need to care which ttyS instance number is allocated depending on HSUART settings or if the devicetree has added aliases for all the ports.

We also prepare the serial core to handle the ttyS related quirks done in console_setup() to prepare things for eventually dropping the parsing from console_setup(). This can only happen after further changes to register_console().

Regards,

Tony

Changes since v3:
- Fix handling of brl_options by saving them too in console_opt_save()

- Drop changes to remove console_setup() parsing, further changes to register_console() are still needed before can leave out the early parsing

- Added a patch for serial8250_isa_init_preferred_console(), otherwise the console gets initialized later for x86 when the console_setup() parsing is dropped as pointed out by Petr

- Initialize __free() variables to NULL as noted by Dan

- Return handled in console_setup() if saving console options fails

- Simplify serial_base_add_preferred_console() and add a helper for serial_base_add_one_prefcon() as suggested by Andy

- Header include changes to kernel/printk/conopt.c as suggested by Andy

Changes since v2:

- Console name got constified and already applied as suggested by Ilpo and Andy

- Add printk/conopt.c to save console command line options

- Add a patch to drop old console_setup() character device name parsing

- Use cleanup.h to simplify freeing as suggested by Andy

- Use types.h instead of kernel.h as suggested by Andy

- Use strcspn() as suggested by Andy

- Various coding improvments suggested by Andy

Changes since v1:

- Constify printk add_preferred_console() as suggested by Jiri

- Use proper kernel command line helpers for parsing console as suggested by Jiri

- Update description for HSUART based on Andy's comments

- Standardize on DEVNAME:0.0 style naming as suggested by Andy

- Added missing put_device() calls paired with device_find_child()

Tony Lindgren (4):
printk: Save console options for add_preferred_console_match()
serial: core: Add support for DEVNAME:0.0 style naming for kernel
console
serial: core: Handle serial console options
serial: 8250: Add preferred console in serial8250_isa_init_ports()

 drivers/tty/serial/8250/8250_core.c  |  32 +++++++
 drivers/tty/serial/serial_base.h     |  14 +++
 drivers/tty/serial/serial_base_bus.c | 115 ++++++++++++++++++++++++
 drivers/tty/serial/serial_core.c     |   4 +
 include/linux/printk.h               |   3 +
 kernel/printk/Makefile               |   2 +-
 kernel/printk/conopt.c               | 128 +++++++++++++++++++++++++++
 kernel/printk/console_cmdline.h      |   6 ++
 kernel/printk/printk.c               |  14 ++-
 9 files changed, 314 insertions(+), 4 deletions(-)
 create mode 100644 kernel/printk/conopt.c

2.43.0

Driver subsystems may need to translate the preferred console name to the character device name used. We already do some of this in console_setup() with a few hardcoded names, but that does not scale well.

The console options are parsed early in console_setup(), and the consoles are added with __add_preferred_console(). At this point we don't know much about the character device names and device drivers getting probed.

To allow drivers subsystems to set up a preferred console, let's save the kernel command line console options. To add a preferred console from a driver subsystem with optional character device name translation, let's add a new function add_preferred_console_match().

This allows the serial core layer to support console=DEVNAME:0.0 style hardware based addressing in addition to the current console=ttyS0 style naming. And we can start moving console_setup() character device parsing to the driver subsystem specific code.

We use a separate array from the console_cmdline array as the character device name and index may be unknown at the console_setup() time. And eventually there's no need to call __add_preferred_console() until the character device name and index are known.

Adding the console name in addition to the character device name, and a flag for an added console, could be added to the struct console_cmdline. And the console_cmdline array handling could be modified accordingly. But that complicates things compared saving the console options, and then adding the consoles when the subsystems handling the consoles are ready.

---
 include/linux/printk.h          |   3 +
 kernel/printk/Makefile          |   2 +-
 kernel/printk/conopt.c          | 128 ++++++++++++++++++++++++++++++++
 kernel/printk/console_cmdline.h |   6 ++
 kernel/printk/printk.c          |  14 +++-
 5 files changed, 149 insertions(+), 4 deletions(-)
 create mode 100644 kernel/printk/conopt.c

@@ -60,6 +60,9 @@ static inline const char *printk_skip_headers(const char *buffer)
 #define CONSOLE_LOGLEVEL_DEFAULT CONFIG_CONSOLE_LOGLEVEL_DEFAULT
 #define CONSOLE_LOGLEVEL_QUIET	 CONFIG_CONSOLE_LOGLEVEL_QUIET
 
+int add_preferred_console_match(const char *match, const char *name,
+				const short idx);
+
 extern int console_printk[];
 
 #define console_loglevel (console_printk[0])

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: GPL-2.0-only
-obj-y	= printk.o
+obj-y	= printk.o conopt.o
 obj-$(CONFIG_PRINTK)	+= printk_safe.o nbcon.o
 obj-$(CONFIG_A11Y_BRAILLE_CONSOLE)	+= braille.o
 obj-$(CONFIG_PRINTK_INDEX)	+= index.o

@@ -2,6 +2,12 @@
 #ifndef _CONSOLE_CMDLINE_H
 #define _CONSOLE_CMDLINE_H
 
+#define MAX_CMDLINECONSOLES 8
+
+int console_opt_save(const char *str, const char *brl_opt);
+int console_opt_add_preferred_console(const char *name, const short idx,
+				      char *options, char *brl_options);
+
 struct console_cmdline
 {
 	char	name[16];			/* Name of the driver	    */

@@ -360,9 +360,6 @@ static int console_locked;
 /*
  *	Array of consoles built from command line options (console=)
  */
-
-#define MAX_CMDLINECONSOLES 8
-
 static struct console_cmdline console_cmdline[MAX_CMDLINECONSOLES];
 
 static int preferred_console = -1;

@@ -2458,6 +2455,10 @@ static int __init console_setup(char *str)
 	if (_braille_console_setup(&str, &brl_options))
 		return 1;
 
+	/* Save the console for driver subsystem use */
+	if (console_opt_save(str, brl_options))
+		return 1;
+
 	/*
 	 * Decode str into name, index, options.
 	 */

@@ -2488,6 +2489,13 @@ static int __init console_setup(char *str)
 }
 __setup("console=", console_setup);
 
+/* Only called from add_preferred_console_match() */
+int console_opt_add_preferred_console(const char *name, const short idx,
+				      char *options, char *brl_options)
+{
+	return __add_preferred_console(name, idx, options, brl_options, true);
+}
+
 /**
  * add_preferred_console - add a device to the list of preferred consoles.
  * @name: device name

2.43.0

We can now add hardware based addressing for serial ports. Starting with commit 84a9582fd203 ("serial: core: Start managing serial controllers to enable runtime PM"), and all the related fixes to this commit, the serial core now knows to which serial port controller the ports are connected.

The serial ports can be addressed with DEVNAME:0.0 style naming. The names are something like 00:04:0.0 for a serial port on qemu, and something like 2800000.serial:0.0 on platform device using systems like ARM64 for example.

The DEVNAME is the unique serial port hardware controller device name, AKA the name for port->dev. The 0.0 are the serial core controller id and port id.

Typically 0.0 are used for each controller and port instance unless the serial port hardware controller has multiple controllers or ports.

Using DEVNAME:0.0 style naming actually solves two long term issues for addressing the serial ports:

1. According to Andy Shevchenko, using DEVNAME:0.0 style naming fixes an issue where depending on the BIOS settings, the kernel serial port ttyS instance number may change if HSUART is enabled

2. Device tree using architectures no longer necessarily need to specify aliases to find a specific serial port, and we can just allocate the ttyS instance numbers dynamically in whatever probe order

To do this, let's match the hardware addressing style console name to the character device name used, and add a preferred console using the character device name.

---
 drivers/tty/serial/serial_base.h     | 14 ++++++++
 drivers/tty/serial/serial_base_bus.c | 48 ++++++++++++++++++++++++++++
 drivers/tty/serial/serial_core.c     |  4 +++
 3 files changed, 66 insertions(+)

@@ -45,3 +45,17 @@ void serial_ctrl_unregister_port(struct uart_driver *drv, struct uart_port *port
 
 int serial_core_register_port(struct uart_driver *drv, struct uart_port *port);
 void serial_core_unregister_port(struct uart_driver *drv, struct uart_port *port);
+
+#ifdef CONFIG_SERIAL_CORE_CONSOLE
+
+int serial_base_add_preferred_console(struct uart_driver *drv,
+				      struct uart_port *port);
+#else
+static inline
+int serial_base_add_preferred_console(struct uart_driver *drv,
+				      struct uart_port *port)
+{
+	return 0;
+}
+
+#endif

@@ -3359,6 +3359,10 @@ int serial_core_register_port(struct uart_driver *drv, struct uart_port *port)
 	if (ret)
 		goto err_unregister_ctrl_dev;
 
+	ret = serial_base_add_preferred_console(drv, port);
+	if (ret)
+		goto err_unregister_port_dev;
+
 	ret = serial_core_add_one_port(drv, port);
 	if (ret)
 		goto err_unregister_port_dev;

2.43.0

In order to start moving the serial console quirks out of console_setup(), let's add parsing for the quirks to the serial core layer. We can use add_preferred_console_match() to handle the quirks.

At this point we can't drop the quirks from console_setup() because it would confuse add_preferred_console(). And try_enable_default_console() would get called before try_enable_preferred_console().

Note that eventually we may want to set up driver specific console quirk handling for the serial port device drivers to use. But we need to figure out which driver(s) need to call the quirk. So for now, we just handle the sparc quirk directly.

---
 drivers/tty/serial/serial_base_bus.c | 67 ++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)

@@ -207,6 +207,43 @@ void serial_base_port_device_remove(struct serial_port_device *port_dev)
 
 #ifdef CONFIG_SERIAL_CORE_CONSOLE
 
+#ifdef __sparc__
+
+/* Handle Sparc ttya and ttyb options as done in console_setup() */
+static int serial_base_add_sparc_console(struct uart_driver *drv,
+					 struct uart_port *port)
+{
+	const char *name = NULL;
+	int ret;
+
+	switch (port->line) {
+	case 0:
+		name = "ttya";
+		break;
+	case 1:
+		name = "ttyb";
+		break;
+	default:
+		return 0;
+	}
+
+	ret = add_preferred_console_match(name, drv->dev_name, port->line);
+	if (ret && ret != -ENOENT)
+		return ret;
+
+	return 0;
+}
+
+#else
+
+static inline int serial_base_add_sparc_console(struct uart_driver *drv,
+						struct uart_port *port)
+{
+	return 0;
+}
+
+#endif
+
 static int serial_base_add_one_prefcon(const char *match, const char *dev_name,
 				       int port_id)
 {

@@ -240,12 +277,42 @@ int serial_base_add_preferred_console(struct uart_driver *drv,
 				      struct uart_port *port)
 {
 	const char *port_match __free(kfree);
+	const char *char_match __free(kfree) = NULL;
+	const char *nmbr_match __free(kfree) = NULL;
+	int ret;
 
 	port_match = kasprintf(GFP_KERNEL, "%s:%i.%i", dev_name(port->dev),
 			       port->ctrl_id, port->port_id);
 	if (!port_match)
 		return -ENOMEM;
 
+	char_match = kasprintf(GFP_KERNEL, "%s%i", drv->dev_name, port->line);
+	if (!char_match)
+		return -ENOMEM;
+
+	/* Handle ttyS specific options */
+	if (!strncmp(drv->dev_name, "ttyS", 4)) {
+		/* No name, just a number */
+		nmbr_match = kasprintf(GFP_KERNEL, "%i", port->line);
+		if (!nmbr_match)
+			return -ENODEV;
+
+		ret = serial_base_add_one_prefcon(nmbr_match, drv->dev_name,
+						  port->line);
+		if (ret)
+			return ret;
+
+		/* Sparc ttya and ttyb */
+		ret = serial_base_add_sparc_console(drv, port);
+		if (ret)
+			return ret;
+	}
+
+	/* Handle the traditional character device name style console=ttyS0 */
+	ret = serial_base_add_one_prefcon(char_match, drv->dev_name, port->line);
+	if (ret)
+		return ret;
+
 	/* Translate a hardware addressing style console=DEVNAME:0.0 */
 	return serial_base_add_one_prefcon(port_match, drv->dev_name, port->line);
 }

2.43.0

...

	return serial_base_add_one_prefcon(...);

...

With Best Regards,
Andy Shevchenko

Prepare 8250 isa ports to drop kernel command line serial console handling from console_setup().

We need to set the preferred console in serial8250_isa_init_ports(). Otherwise the console gets initialized only later on when the hardware specific driver takes over, and console_setup() is no longer handling the ttyS related quirks.

Note that this mostly affects x86 as this happens based on define SERIAL_PORT_DFNS.

---
 drivers/tty/serial/8250/8250_core.c | 32 +++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

@@ -15,6 +15,7 @@
  */
 
 #include <linux/acpi.h>
+#include <linux/cleanup.h>
 #include <linux/module.h>
 #include <linux/moduleparam.h>
 #include <linux/ioport.h>

@@ -517,6 +518,35 @@ static struct uart_8250_port *serial8250_setup_port(int index)
 	return up;
 }
 
+#ifdef CONFIG_SERIAL_8250_CONSOLE
+
+/*
+ * There is no struct device at this point, so let's not try to use
+ * serial_base_add_preferred_console().
+ */
+static void __init serial8250_isa_init_preferred_console(int idx)
+{
+	const char *name __free(kfree);
+	int ret;
+
+	name = kasprintf(GFP_KERNEL, "%s%i", serial8250_reg.dev_name, idx);
+	ret = add_preferred_console_match(name, serial8250_reg.dev_name, idx);
+	if (!ret || ret == -ENOENT)
+		return;
+
+	pr_err("Could not add preferred console for %s idx %i\n",
+	       serial8250_reg.dev_name, idx);
+}
+
+#else
+
+static inline void serial8250_isa_init_preferred_console(struct uart_port *port,
+							 int idx)
+{
+}
+
+#endif
+
 static void __init serial8250_isa_init_ports(void)
 {
 	struct uart_8250_port *up;

@@ -563,6 +593,8 @@ static void __init serial8250_isa_init_ports(void)
 		port->irqflags |= irqflag;
 		if (serial8250_isa_config != NULL)
 			serial8250_isa_config(i, &up->port, &up->capabilities);
+
+		serial8250_isa_init_preferred_console(i);
 	}
 }
 

2.43.0

ISA

...

	ret = serial_base_add_one_prefcon(...);

With Best Regards,
Andy Shevchenko

...

And, btw, you can simply reuse name here

	pr_err("Could not add preferred console for %s\n", name);

Would it work?

With Best Regards,
Andy Shevchenko

Oops

Yup that should work even before struct device. Will fix the other places too you noticed.

Thanks,

Tony

Petr FYI, so for dropping the console_setup() parsing, below is a hack patch to see what goes wrong in register_console() if you have some ideas on how to handle this.

We end up with the console device backed up seria8250 instead of ttyS0, and earlycon won't get properly disabled. And of course other consoles beyond ttyS need to be also considered.

Regards,

Tony

8< ----------------------

@@ -2438,9 +2438,7 @@ __setup("console_msg_format=", console_msg_format_setup);
  */
 static int __init console_setup(char *str)
 {
-	char buf[sizeof(console_cmdline[0].name) + 4]; /* 4 for "ttyS" */
-	char *s, *options, *brl_options = NULL;
-	int idx;
+	char *brl_options = NULL;
 
 	/*
 	 * console="" or console=null have been suggested as a way to

@@ -2459,32 +2457,9 @@ static int __init console_setup(char *str)
 	if (console_opt_save(str, brl_options))
 		return 1;
 
-	/*
-	 * Decode str into name, index, options.
-	 */
-	if (str[0] >= '0' && str[0] <= '9') {
-		strcpy(buf, "ttyS");
-		strncpy(buf + 4, str, sizeof(buf) - 5);
-	} else {
-		strncpy(buf, str, sizeof(buf) - 1);
-	}
-	buf[sizeof(buf) - 1] = 0;
-	options = strchr(str, ',');
-	if (options)
-		*(options++) = 0;
-#ifdef __sparc__
-	if (!strcmp(str, "ttya"))
-		strcpy(buf, "ttyS0");
-	if (!strcmp(str, "ttyb"))
-		strcpy(buf, "ttyS1");
-#endif
-	for (s = buf; *s; s++)
-		if (isdigit(*s) || *s == ',')
-			break;
-	idx = simple_strtoul(s, NULL, 10);
-	*s = 0;
+	/* Indicate register_console() a console was specified */
+	console_set_on_cmdline = 1;
 
-	__add_preferred_console(buf, idx, options, brl_options, true);
 	return 1;
 }
 __setup("console=", console_setup);

@@ -3476,7 +3451,7 @@ void register_console(struct console *newcon)
 	 * Note that a console with tty binding will have CON_CONSDEV
 	 * flag set and will be first in the list.
 	 */
-	if (preferred_console < 0) {
+	if (preferred_console < 0 && !console_set_on_cmdline) {
 		if (hlist_empty(&console_list) || !console_first()->device ||
 		    console_first()->flags & CON_BOOT) {
 			try_enable_default_console(newcon);

Hmm so the following extra patch seems to fix the issues based on light testing. But is it safe to assume that if CON_PRINTBUFFER is set we can disable the bootconsole?

I also noticed that the bootconsole not getting disabled issue is there also for console=DEVNAME:0.0 usage even before we start dropping handling in console_setup().

Regards,

Tony

8< ----------------------

@@ -3549,7 +3549,8 @@ void register_console(struct console *newcon)
 	 */
 	con_printk(KERN_INFO, newcon, "enabled\n");
 	if (bootcon_registered &&
-	    ((newcon->flags & (CON_CONSDEV | CON_BOOT)) == CON_CONSDEV) &&
+	    !(newcon->flags & CON_BOOT) &&
+	    (newcon->flags & (CON_CONSDEV | CON_PRINTBUFFER)) &&
 	    !keep_bootcon) {
 		struct hlist_node *tmp;
 

2.43.0

OK so no need for the CON_PRINTBUFFER change, it's wrong. I found a few bugs causing this issue and a lot of other confusion while testing:

- In console_setup(), a DEVNAME:0.0 style console can get added with the IO address turned into a ttyS console with some crazy index :) So we need to bail out early on consoles with ':' in the name.

- The brl_opts can be empty or NULL, but we need to pass NULL to __add_preferred_console() to get CON_CONSDEV flag set for DEVNAME:0.0 console. Otherwise the preferred_console won't get set and the boot console won't get disabled.

- The console_set_on_cmdline flag needs to be set if console_setup() does not call __add_preferred_console() for DEVNAME:0.0 style console as otherwise try_enable_default_console() may get called before the console handling driver has added the preferred console.

I think with these the remaining issues are sorted out :) I'll post a v5 set with as RFC as it's getting close to the merge window.

Regards,

Tony