Skip to contentSkip to conversation

bus_add_device() losing an error return from the probe() method

PATCH Applied Rene Herman 7 participants 44 messages Mar 24, 2006 – Apr 6, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi Greg, Takashi.

ALSA moved all ISA drivers over to the platform_driver interface in 2.6.16, using this code structure in the module_inits:

cards = 0;

   for (i = 0; i < SNDRV_CARDS; i++) {
	  struct platform_device *device;
	  device = platform_device_register_simple(
			SND_FOO_DRIVER, i, NULL, 0);
	  if (IS_ERR(device)) {
		  err = PTR_ERR(device);
		  goto errout;
	  }
	  devices[i] = device;
	  cards++;
   }
   if (!cards) {
	  printk(KERN_ERR "FOO soundcard not found or device busy\n");
	  err = -ENODEV;
	  goto errout;

}
return 0;
errout:
snd_foo_unregister_all();
return err;

Unfortunately, the snd_foo_unregister_all() part here is unreachable under normal circumstances, since platform_device_register_simple() returns !IS_ERR, regardless of what the driver probe method returned. The driver then never fails to load, even when no cards were found.

An error return from the driver probe() method is carried up through device_attach, but is then dropped on the floor in bus_add_device(). If I apply the attached patch, things work as I (and ALSA it seems) expect. Is it correct?

(the printk still isn't reached, but see next message for that).

Rene.

drivers/base/bus.c +9 -7
Index: local/drivers/base/bus.c
===================================================================
@@ -363,19 +363,21 @@ static void device_remove_attrs(struct b
 int bus_add_device(struct device * dev)
 {
 	struct bus_type * bus = get_bus(dev->bus);
-	int error = 0;
+	int error;
 
 	if (bus) {
 		pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-		device_attach(dev);
+		error = device_attach(dev);
+		if (error < 0)
+			return error;
 		klist_add_tail(&dev->knode_bus, &bus->klist_devices);
 		error = device_add_attrs(bus, dev);
-		if (!error) {
-			sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
-			sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
-		}
+		if (error)
+			return error;
+		sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+		sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
 	}
-	return error;
+	return 0;
 }
 
 /**
#1 lore raw
Quoted 23 lines · Rene Herman

===================================================================
--- local.orig/drivers/base/bus.c 2006-02-27 19:22:08.000000000 +0100
+++ local/drivers/base/bus.c 2006-03-24 04:27:02.000000000 +0100
@@ -363,19 +363,21 @@ static void device_remove_attrs(struct b
int bus_add_device(struct device * dev)
{
struct bus_type * bus = get_bus(dev->bus);
- int error = 0;
+ int error;

if (bus) {
pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
- device_attach(dev);
+ error = device_attach(dev);
+ if (error < 0)
+ return error;
klist_add_tail(&dev->knode_bus, &bus->klist_devices); error = device_add_attrs(bus, dev);
- if (!error) {
- sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
- sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
- }

Quoted 9 lines

+ if (error)
+ return error;
+ sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+ sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
}
- return error;
+ return 0;
}

Looks sane, but please don't sprinkle `return' statements all over a function in this manner.

drivers/base/bus.c renamed from drivers/base/bus.c~bus_add_device-losing-an-error-return-from-the-probe-method +8 -5
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
 	if (bus) {
 		pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-		device_attach(dev);
+		error = device_attach(dev);
+		if (error < 0)
+			goto out;
 		klist_add_tail(&dev->knode_bus, &bus->klist_devices);
 		error = device_add_attrs(bus, dev);
-		if (!error) {
-			sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
-			sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
-		}
+		if (error)
+			goto out;
+		sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+		sysfs_create_link(&dev->kobj,&dev->bus->subsys.kset.kobj,"bus");
 	}
+out:
 	return error;
 }
 

It's a little surprising that this function returns "OK" if bus==NULL.

Note that sysfs_create_link() can fail too. This was one optimistic function.

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#2 lore raw
In reply to Rene Herman
Quoted 23 lines · Rene Herman

===================================================================
--- local.orig/drivers/base/bus.c 2006-02-27 19:22:08.000000000 +0100
+++ local/drivers/base/bus.c 2006-03-24 04:27:02.000000000 +0100
@@ -363,19 +363,21 @@ static void device_remove_attrs(struct b
int bus_add_device(struct device * dev)
{
struct bus_type * bus = get_bus(dev->bus);
- int error = 0;
+ int error;

if (bus) {
pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
- device_attach(dev);
+ error = device_attach(dev);
+ if (error < 0)
+ return error;
klist_add_tail(&dev->knode_bus, &bus->klist_devices); error = device_add_attrs(bus, dev);
- if (!error) {
- sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
- sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
- }

Quoted 9 lines

+ if (error)
+ return error;
+ sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+ sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
}
- return error;
+ return 0;
}

Looks sane, but please don't sprinkle `return' statements all over a function in this manner.

drivers/base/bus.c renamed from drivers/base/bus.c~bus_add_device-losing-an-error-return-from-the-probe-method +8 -5
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
 	if (bus) {
 		pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-		device_attach(dev);
+		error = device_attach(dev);
+		if (error < 0)
+			goto out;
 		klist_add_tail(&dev->knode_bus, &bus->klist_devices);
 		error = device_add_attrs(bus, dev);
-		if (!error) {
-			sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
-			sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
-		}
+		if (error)
+			goto out;
+		sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+		sysfs_create_link(&dev->kobj,&dev->bus->subsys.kset.kobj,"bus");
 	}
+out:
 	return error;
 }
 

It's a little surprising that this function returns "OK" if bus==NULL.

Note that sysfs_create_link() can fail too. This was one optimistic function.

#3 lore raw
Quoted 1 line · Andrew Morton

Looks sane, but please don't sprinkle `return' statements all over a function in this manner.

I actually prefer the multiple returns. You then don't have to "visually scroll down" to the label to see what would happen when reading the code. Even when there's common code before the return, I've never seen GCC not optimise that to the goto form itself. You obviously the boss though.

Quoted 4 lines

It's a little surprising that this function returns "OK" if bus==NULL.

Note that sysfs_create_link() can fail too. This was one optimistic function.

I assume that Greg hasn't commented yet since he's busy rewriting it all, so that'll be okay :-)

Rene.

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#4 lore raw
In reply to Andrew Morton
Quoted 1 line · Andrew Morton

Looks sane, but please don't sprinkle `return' statements all over a function in this manner.

I actually prefer the multiple returns. You then don't have to "visually scroll down" to the label to see what would happen when reading the code. Even when there's common code before the return, I've never seen GCC not optimise that to the goto form itself. You obviously the boss though.

Quoted 4 lines

It's a little surprising that this function returns "OK" if bus==NULL.

Note that sysfs_create_link() can fail too. This was one optimistic function.

I assume that Greg hasn't commented yet since he's busy rewriting it all, so that'll be okay :-)

Rene.

#5 lore raw

patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman

This is a note to let you know that I've just added the patch titled

Subject: bus_add_device() losing an error return from the probe() method

to my gregkh-2.6 tree. Its filename is

bus_add_device-losing-an-error-return-from-the-probe-method.patch

This tree can be found at
http://www.kernel.org/pub/linux/kernel/people/gregkh/gregkh-2.6/patches/

Patches currently in gregkh-2.6 which might be from [email protected] are

driver/bus_add_device-losing-an-error-return-from-the-probe-method.patch

Quoted 1 line

From [email protected] Thu Mar 23 21:32:00 2006

Message-ID: <[email protected]>
Date: Fri, 24 Mar 2006 06:32:57 +0100
From: Rene Herman <[email protected]>
To: Greg Kroah-Hartman <[email protected]>
Cc: Takashi Iwai <[email protected]>, ALSA devel <[email protected]>, Linux Kernel <[email protected]>
Subject: bus_add_device() losing an error return from the probe() method

ALSA moved all ISA drivers over to the platform_driver interface in 2.6.16, using this code structure in the module_inits:

    cards = 0;
    for (i = 0; i < SNDRV_CARDS; i++) {
	  struct platform_device *device;
	  device = platform_device_register_simple(
			SND_FOO_DRIVER, i, NULL, 0);
	  if (IS_ERR(device)) {
		  err = PTR_ERR(device);
		  goto errout;
	  }
	  devices[i] = device;
	  cards++;
    }
    if (!cards) {
	  printk(KERN_ERR "FOO soundcard not found or device busy\n");
	  err = -ENODEV;
	  goto errout;
    }
    return 0;
errout:
    snd_foo_unregister_all();
    return err;

Unfortunately, the snd_foo_unregister_all() part here is unreachable under normal circumstances, since platform_device_register_simple() returns !IS_ERR, regardless of what the driver probe method returned. The driver then never fails to load, even when no cards were found.

An error return from the driver probe() method is carried up through device_attach, but is then dropped on the floor in bus_add_device(). If I apply the attached patch, things work as I (and ALSA it seems) expect.

From: Rene Herman <[email protected]>

---
 drivers/base/bus.c |   13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
drivers/base/bus.c +8 -5
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
 	if (bus) {
 		pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-		device_attach(dev);
+		error = device_attach(dev);
+		if (error < 0)
+			goto exit;
 		klist_add_tail(&dev->knode_bus, &bus->klist_devices);
 		error = device_add_attrs(bus, dev);
-		if (!error) {
-			sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
-			sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
-		}
+		if (error)
+			goto exit;
+		sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+		sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
 	}
+exit:
 	return error;
 }
 

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#6 lore raw

patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman

This is a note to let you know that I've just added the patch titled

Subject: bus_add_device() losing an error return from the probe() method

to my gregkh-2.6 tree. Its filename is

bus_add_device-losing-an-error-return-from-the-probe-method.patch

This tree can be found at
http://www.kernel.org/pub/linux/kernel/people/gregkh/gregkh-2.6/patches/

Patches currently in gregkh-2.6 which might be from [email protected] are

driver/bus_add_device-losing-an-error-return-from-the-probe-method.patch

Quoted 1 line

From [email protected] Thu Mar 23 21:32:00 2006

Message-ID: <[email protected]>
Date: Fri, 24 Mar 2006 06:32:57 +0100
From: Rene Herman <[email protected]>
To: Greg Kroah-Hartman <[email protected]>
Cc: Takashi Iwai <[email protected]>, ALSA devel <[email protected]>, Linux Kernel <[email protected]>
Subject: bus_add_device() losing an error return from the probe() method

ALSA moved all ISA drivers over to the platform_driver interface in 2.6.16, using this code structure in the module_inits:

    cards = 0;
    for (i = 0; i < SNDRV_CARDS; i++) {
	  struct platform_device *device;
	  device = platform_device_register_simple(
			SND_FOO_DRIVER, i, NULL, 0);
	  if (IS_ERR(device)) {
		  err = PTR_ERR(device);
		  goto errout;
	  }
	  devices[i] = device;
	  cards++;
    }
    if (!cards) {
	  printk(KERN_ERR "FOO soundcard not found or device busy\n");
	  err = -ENODEV;
	  goto errout;
    }
    return 0;
errout:
    snd_foo_unregister_all();
    return err;

Unfortunately, the snd_foo_unregister_all() part here is unreachable under normal circumstances, since platform_device_register_simple() returns !IS_ERR, regardless of what the driver probe method returned. The driver then never fails to load, even when no cards were found.

An error return from the driver probe() method is carried up through device_attach, but is then dropped on the floor in bus_add_device(). If I apply the attached patch, things work as I (and ALSA it seems) expect.

From: Rene Herman <[email protected]>

---
 drivers/base/bus.c |   13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
drivers/base/bus.c +8 -5
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
 	if (bus) {
 		pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-		device_attach(dev);
+		error = device_attach(dev);
+		if (error < 0)
+			goto exit;
 		klist_add_tail(&dev->knode_bus, &bus->klist_devices);
 		error = device_add_attrs(bus, dev);
-		if (!error) {
-			sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
-			sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
-		}
+		if (error)
+			goto exit;
+		sysfs_create_link(&bus->devices.kobj, &dev->kobj, dev->bus_id);
+		sysfs_create_link(&dev->kobj, &dev->bus->subsys.kset.kobj, "bus");
 	}
+exit:
 	return error;
 }
 
#7 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

drivers/base/bus.c int bus_add_device(struct device * dev) · line 377

Quoted 11 lines · On 4/4/06, [email protected] <[email protected]> wrote:
drivers/base/bus.c +3 -1
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
       if (bus) {
               pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-               device_attach(dev);
+               error = device_attach(dev);
+               if (error < 0)
+                       goto exit;

I do not believe that this is correct. The fact that _some_ driver failed to attach to a device does not necessarily mean that device itself does not exist. While this assuption might work for platform devices it won't work for other busses.

As fasr as ALSA goes, if they want to abort loading module if device is not present they will have to separate device probing from final driver binding.

--
Dmitry

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid0944&bid$1720&dat1642

#8 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to [email protected]

drivers/base/bus.c int bus_add_device(struct device * dev) · line 377

Quoted 11 lines · On 4/4/06, [email protected] <[email protected]> wrote:
drivers/base/bus.c +3 -1
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
       if (bus) {
               pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-               device_attach(dev);
+               error = device_attach(dev);
+               if (error < 0)
+                       goto exit;

I do not believe that this is correct. The fact that _some_ driver failed to attach to a device does not necessarily mean that device itself does not exist. While this assuption might work for platform devices it won't work for other busses.

As fasr as ALSA goes, if they want to abort loading module if device is not present they will have to separate device probing from final driver binding.

--
Dmitry

#9 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 6 lines · Dmitry Torokhov
Quoted 5 lines · On 4/4/06, [email protected] <[email protected]> wrote:
drivers/base/bus.c +0 -0
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
Quoted 11 lines

drivers/base/bus.c int bus_add_device(struct device * dev) · line 377

Quoted 6 lines
drivers/base/bus.c +3 -1
       if (bus) {
               pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-               device_attach(dev);
+               error = device_attach(dev);
+               if (error < 0)
+                       goto exit;

I do not believe that this is correct. The fact that _some_ driver failed to attach to a device does not necessarily mean that device itself does not exist. While this assuption might work for platform devices it won't work for other busses.

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.
- if the probe() function returns 0, the device is bound to the driver, and it returns 0. Hm, looks like we continue to loop here too, we could probably stop now that we have bound a driver to the device.

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Does that help?

thanks,

greg k-h

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#10 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 6 lines · Dmitry Torokhov
Quoted 5 lines · On 4/4/06, [email protected] <[email protected]> wrote:
drivers/base/bus.c +0 -0
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
Quoted 11 lines

drivers/base/bus.c int bus_add_device(struct device * dev) · line 377

Quoted 6 lines
drivers/base/bus.c +3 -1
       if (bus) {
               pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-               device_attach(dev);
+               error = device_attach(dev);
+               if (error < 0)
+                       goto exit;

I do not believe that this is correct. The fact that _some_ driver failed to attach to a device does not necessarily mean that device itself does not exist. While this assuption might work for platform devices it won't work for other busses.

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.
- if the probe() function returns 0, the device is bound to the driver, and it returns 0. Hm, looks like we continue to loop here too, we could probably stop now that we have bound a driver to the device.

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Does that help?

thanks,

greg k-h

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

#11 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 38 lines · Greg KH
Quoted 6 lines · Dmitry Torokhov
Quoted 5 lines · On 4/4/06, [email protected] <[email protected]> wrote:
drivers/base/bus.c +0 -0
@@ -372,14 +372,17 @@ int bus_add_device(struct device * dev)
 
Quoted 11 lines

drivers/base/bus.c int bus_add_device(struct device * dev) · line 377

Quoted 6 lines
drivers/base/bus.c +3 -1
       if (bus) {
               pr_debug("bus %s: add device %s\n", bus->name, dev->bus_id);
-               device_attach(dev);
+               error = device_attach(dev);
+               if (error < 0)
+                       goto exit;

I do not believe that this is correct. The fact that _some_ driver failed to attach to a device does not necessarily mean that device itself does not exist. While this assuption might work for platform devices it won't work for other busses.

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.
- if the probe() function returns 0, the device is bound to the driver, and it returns 0. Hm, looks like we continue to loop here too, we could probably stop now that we have bound a driver to the device.

Quoted 3 lines

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Yup, things still seem to work properly for me. The patch will show up in the next -mm for others to beat on...

thanks,

greg k-h

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#12 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 7 lines · Greg KH
Quoted 4 lines

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Yup, things still seem to work properly for me. The patch will show up in the next -mm for others to beat on...

It was in 2.6.16-mm2 and 2.6.17-rc1-mm1.

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#13 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 7 lines · Greg KH
Quoted 4 lines

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Yup, things still seem to work properly for me. The patch will show up in the next -mm for others to beat on...

It was in 2.6.16-mm2 and 2.6.17-rc1-mm1.

#14 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 16 lines · Greg KH

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

Quoted 4 lines

- if the probe() function returns 0, the device is bound to the driver, and it returns 0. Hm, looks like we continue to loop here too, we could probably stop now that we have bound a driver to the device.

Could it be that logic is simply reversed?

Quoted 6 lines

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Does that help?

Quoted 8 lines

thanks,

greg k-h

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

--
Dmitry

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid0944&bid$1720&dat1642

#15 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 16 lines · Greg KH

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

Quoted 4 lines

- if the probe() function returns 0, the device is bound to the driver, and it returns 0. Hm, looks like we continue to loop here too, we could probably stop now that we have bound a driver to the device.

Could it be that logic is simply reversed?

Quoted 6 lines

So, I'm pretty sure that this is safe and should work just fine. To be sure, let me go reboot my box with this change on it after I finish this email :)

Does that help?

Quoted 8 lines

thanks,

greg k-h

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

--
Dmitry

#16 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 18 lines · Dmitry Torokhov
Quoted 16 lines · Greg KH

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.

Quoted 4 lines

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

The device is still there.

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

If I changed the patch to do that, would it be acceptable to you?

thanks,

greg k-h

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#17 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 18 lines · Dmitry Torokhov
Quoted 16 lines · Greg KH

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.

Quoted 4 lines

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

The device is still there.

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

If I changed the patch to do that, would it be acceptable to you?

thanks,

greg k-h

#18 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 24 lines · Greg KH
Quoted 18 lines · Dmitry Torokhov
Quoted 16 lines · Greg KH

Hm, no, I unwound this mess, and found the following:

- bus_add_device() calls device_attach()
- device_attach() calls bus_for_each_drv() for every driver on the bus
- bus_for_each_drv() walks all drivers on the bus and calls __device_attach() for every individual driver
- __device_attach() calls driver_probe_device() for that driver and device
- driver_probe_device() calls down to the probe() function for the driver, passing it that driver, if match() for the bus matches this device.
- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers
- if the probe() function returns any other error code, it is propagated up, all the way back to bus_add_device.

Quoted 4 lines

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

Quoted 7 lines

The device is still there.

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

I don't think so because device creation did not fail. Otherwise how would you as a caller of device_register() distinguish between the following 2 scenarios:

- you got -ENOMEM (or other error code) because device creation indeed failed;
- you got -ENOMEM because some odd driver could not allocate 4MB of memory.

IOW you trying to propagate driver error to device creation code...

Also result of device_register() should not depend on whether driver_register() was called earlier or not.

Signature · 8 lines
Dmitry


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642
#19 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 15 lines · Dmitry Torokhov
Quoted 4 lines · Greg KH

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

I don't think so because device creation did not fail. Otherwise how would you as a caller of device_register() distinguish between the following 2 scenarios:

- you got -ENOMEM (or other error code) because device creation indeed failed;
- you got -ENOMEM because some odd driver could not allocate 4MB of memory.

Quoted 4 lines

IOW you trying to propagate driver error to device creation code...

Also result of device_register() should not depend on whether driver_register() was called earlier or not.

Indeed. Greg - this patch is bogus.

When device_register() returns a negative number, many device drivers assume (correctly) that the device was _not_ registered and they will take whatever cleanup action is necessary to free the memory associated with that device, believing that sysfs never used it.

If we start propagating driver probe error codes up through device_register() we then have a big headache. Is this -ENOMEM error a result of not being able to register the struct device, or is it because something failed in the driver.

This will also impact hotplug and PCMCIA. For example, you have a device driver loaded, but are short on memory, and you insert a PCMCIA device. That creates a struct device. Let's say that the driver fails to allocate enough memory, and returns -ENOMEM. With this patch, that'll kill off the struct device. Now, when the device is pulled, we could end up calling driver_unregister on an unregistered struct device.

The expectations by current subsystems is that struct device registration does _NOT_ depend on the currently loaded set of device drivers not returning errors.

The thing is that this is all corner case stuff - you can apply this patch and apparantly have a working system. It's the corner cases where some driver returns a failure which are the problem case.

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

There is also no 1:1 relationship between platform device drivers and platform devices. You can have multiple platform devices driven by one platform device driver. See the plethora of platform devices in the ARM sub-tree. Or serial sub-tree.

So, all round this patch seems wrong.

Signature · 10 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 Serial core


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642
#20 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 15 lines · Dmitry Torokhov
Quoted 4 lines · Greg KH

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

I don't think so because device creation did not fail. Otherwise how would you as a caller of device_register() distinguish between the following 2 scenarios:

- you got -ENOMEM (or other error code) because device creation indeed failed;
- you got -ENOMEM because some odd driver could not allocate 4MB of memory.

Quoted 4 lines

IOW you trying to propagate driver error to device creation code...

Also result of device_register() should not depend on whether driver_register() was called earlier or not.

Indeed. Greg - this patch is bogus.

When device_register() returns a negative number, many device drivers assume (correctly) that the device was _not_ registered and they will take whatever cleanup action is necessary to free the memory associated with that device, believing that sysfs never used it.

If we start propagating driver probe error codes up through device_register() we then have a big headache. Is this -ENOMEM error a result of not being able to register the struct device, or is it because something failed in the driver.

This will also impact hotplug and PCMCIA. For example, you have a device driver loaded, but are short on memory, and you insert a PCMCIA device. That creates a struct device. Let's say that the driver fails to allocate enough memory, and returns -ENOMEM. With this patch, that'll kill off the struct device. Now, when the device is pulled, we could end up calling driver_unregister on an unregistered struct device.

The expectations by current subsystems is that struct device registration does _NOT_ depend on the currently loaded set of device drivers not returning errors.

The thing is that this is all corner case stuff - you can apply this patch and apparantly have a working system. It's the corner cases where some driver returns a failure which are the problem case.

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

There is also no 1:1 relationship between platform device drivers and platform devices. You can have multiple platform devices driven by one platform device driver. See the plethora of platform devices in the ARM sub-tree. Or serial sub-tree.

So, all round this patch seems wrong.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 Serial core
#21 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 16 lines · Russell King
Quoted 15 lines · Dmitry Torokhov
Quoted 4 lines · Greg KH

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

I don't think so because device creation did not fail. Otherwise how would you as a caller of device_register() distinguish between the following 2 scenarios:

- you got -ENOMEM (or other error code) because device creation indeed failed;
- you got -ENOMEM because some odd driver could not allocate 4MB of memory.

Quoted 6 lines
Quoted 4 lines

IOW you trying to propagate driver error to device creation code...

Also result of device_register() should not depend on whether driver_register() was called earlier or not.

Indeed. Greg - this patch is bogus.

You and Dmitry are correct. I'm dropping this patch. The ISA drivers just need to get used to the proper way to use the driver model (they do not know if they have been bound to a device when the driver is loaded, no big deal).

If there's other issues with platform devices still, without this patch applied, please let me know.

thanks to you all for pointing out the real issues here.

greg k-h

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#22 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Russell King
Quoted 16 lines · Russell King
Quoted 15 lines · Dmitry Torokhov
Quoted 4 lines · Greg KH

Ah, I see what you are saying now. Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems. But then, we should still propagate the error back up, right?

I don't think so because device creation did not fail. Otherwise how would you as a caller of device_register() distinguish between the following 2 scenarios:

- you got -ENOMEM (or other error code) because device creation indeed failed;
- you got -ENOMEM because some odd driver could not allocate 4MB of memory.

Quoted 6 lines
Quoted 4 lines

IOW you trying to propagate driver error to device creation code...

Also result of device_register() should not depend on whether driver_register() was called earlier or not.

Indeed. Greg - this patch is bogus.

You and Dmitry are correct. I'm dropping this patch. The ISA drivers just need to get used to the proper way to use the driver model (they do not know if they have been bound to a device when the driver is loaded, no big deal).

If there's other issues with platform devices still, without this patch applied, please let me know.

thanks to you all for pointing out the real issues here.

greg k-h

#23 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 10 lines · Greg KH
Quoted 4 lines

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

The device is still there.

Ah, I see what you are saying now.  Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems.

Not only that, but device will be killed as well - if bus_add_device() fails device_add() branches into BusError which leads to kobject_del().

Signature · 8 lines
Dmitry


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid0944&bid$1720&dat1642
#24 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 10 lines · Greg KH
Quoted 4 lines

But why do we do that? probe() failing is driver's problem. The device is still there and should still be presented in sysfs. I don't think that we should stop if probe() fails - maybe next driver manages to bind itself.

The device is still there.

Ah, I see what you are saying now.  Yeah, we should still add the default attributes for the bus and create the bus link even if some random driver had problems.

Not only that, but device will be killed as well - if bus_add_device() fails device_add() branches into BusError which leads to kobject_del().

Signature · 1 line
Dmitry
#25 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 7 lines · Greg KH

- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

Is the "this" in "require this" referring to (-ENODEV or -ENXIO) or to -ENXIO alone and do you consider the -ENODEV behaviour correct?

ALSA wants platform_device_register_simple() to return an IS_ERR() on driver probe error and the submitted patch makes it do so for all the other errors but ALSA likes to propagate errors up as far as possible, and currently its probe() methods can return either...

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

As said before, if the behaviour makes sense for other busses, maybe propagating errors up should be dependent on a flags value somewhere that a platform-driver sets?

If platform_device_register_simple() never returns an IS_ERR() when the device is not found that means it's not a useful interface for hardware that needs to be probed for at the very least. ALSA would need to do something like, just before returning a succesfull return from the probe() method, set a global flag that the platform_device that is about to be registered is actually representing something, and freeing all platform_devices for which the flag is _not_ set again after this.

Which ofcourse means this is not at all useful. It's just working around the driver model then...

Rene.

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#26 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Greg KH
Quoted 7 lines · Greg KH

- if that probe() function returns -ENODEV or -ENXIO[1] then the error is ignored and 0 is returned, causing the loop to continue to try more drivers

[1] - stupid agp drivers (or some other video drivers) require this. I need to go fix them up so they don't do this, if they haven't been fixed already...

Is the "this" in "require this" referring to (-ENODEV or -ENXIO) or to -ENXIO alone and do you consider the -ENODEV behaviour correct?

ALSA wants platform_device_register_simple() to return an IS_ERR() on driver probe error and the submitted patch makes it do so for all the other errors but ALSA likes to propagate errors up as far as possible, and currently its probe() methods can return either...

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

As said before, if the behaviour makes sense for other busses, maybe propagating errors up should be dependent on a flags value somewhere that a platform-driver sets?

If platform_device_register_simple() never returns an IS_ERR() when the device is not found that means it's not a useful interface for hardware that needs to be probed for at the very least. ALSA would need to do something like, just before returning a succesfull return from the probe() method, set a global flag that the platform_device that is about to be registered is actually representing something, and freeing all platform_devices for which the flag is _not_ set again after this.

Which ofcourse means this is not at all useful. It's just working around the driver model then...

Rene.

#27 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 5 lines · Rene Herman

As said before, if the behaviour makes sense for other busses, maybe propagating errors up should be dependent on a flags value somewhere that a platform-driver sets?

If platform_device_register_simple() never returns an IS_ERR() when the device is not found that means it's not a useful interface for hardware that needs to be probed for at the very least. ALSA would need to do something like, just before returning a succesfull return from the probe() method, set a global flag that the platform_device that is about to be registered is actually representing something, and freeing all platform_devices for which the flag is _not_ set again after this.

Which ofcourse means this is not at all useful. It's just working around the driver model then...

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though. Having a flag somewhere in struct device_driver that a platform_driver would set and which would tell the driver model to honour the driver probe() method return seems the cleanest approach for all hardware where only the driver can now if the device is really present or not.

Ofcourse, I'm also still fine with just propagating the error up always (and if can be doing something about the ignore of -ENODEV / -ENOXIO)...

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#28 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman
Quoted 5 lines · Rene Herman

As said before, if the behaviour makes sense for other busses, maybe propagating errors up should be dependent on a flags value somewhere that a platform-driver sets?

If platform_device_register_simple() never returns an IS_ERR() when the device is not found that means it's not a useful interface for hardware that needs to be probed for at the very least. ALSA would need to do something like, just before returning a succesfull return from the probe() method, set a global flag that the platform_device that is about to be registered is actually representing something, and freeing all platform_devices for which the flag is _not_ set again after this.

Which ofcourse means this is not at all useful. It's just working around the driver model then...

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though. Having a flag somewhere in struct device_driver that a platform_driver would set and which would tell the driver model to honour the driver probe() method return seems the cleanest approach for all hardware where only the driver can now if the device is really present or not.

Ofcourse, I'm also still fine with just propagating the error up always (and if can be doing something about the ignore of -ENODEV / -ENOXIO)...

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#29 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 14 lines · Rene Herman
Quoted 12 lines · Rene Herman

As said before, if the behaviour makes sense for other busses, maybe propagating errors up should be dependent on a flags value somewhere that a platform-driver sets?

If platform_device_register_simple() never returns an IS_ERR() when the device is not found that means it's not a useful interface for hardware that needs to be probed for at the very least. ALSA would need to do something like, just before returning a succesfull return from the probe() method, set a global flag that the platform_device that is about to be registered is actually representing something, and freeing all platform_devices for which the flag is _not_ set again after this.

Quoted 6 lines
Quoted 2 lines

Which ofcourse means this is not at all useful. It's just working around the driver model then...

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though.

Plus it broke all the drivers that create platform devices before registering drivers or the ones simply not using private data. Given that some arches have a means to separate device creation from driver probing (see pcspkr on PPC for exaple) I don;t think this is acceptable.

Signature · 8 lines
Dmitry


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642
#30 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 4 lines · Dmitry Torokhov
Quoted 1 line

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though.

Plus it broke all the drivers that create platform devices before registering drivers or the ones simply not using private data.

No, this was just a suggestion for an ALSA specific workaround. I was suggesting ALSA drivers could do this.

Rene.

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642

#31 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 4 lines · Dmitry Torokhov
Quoted 1 line

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though.

Plus it broke all the drivers that create platform devices before registering drivers or the ones simply not using private data.

No, this was just a suggestion for an ALSA specific workaround. I was suggesting ALSA drivers could do this.

Rene.

#32 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 6 lines · Rene Herman
Quoted 4 lines · Dmitry Torokhov
Quoted 3 lines

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though.

Quoted 6 lines
Quoted 2 lines

Plus it broke all the drivers that create platform devices before registering drivers or the ones simply not using private data.

No, this was just a suggestion for an ALSA specific workaround. I was suggesting ALSA drivers could do this.

Yes, I am sorry - I misread the code snipped as if it was in driver core, not in ALSA itself.

--
Dmitry

#33 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman
Quoted 6 lines · Rene Herman
Quoted 4 lines · Dmitry Torokhov
Quoted 3 lines

Well, we could in fact hang an unregister off device->private_data as per attached example. Wouldn't be _excessively_ ugly. Still sucks though.

Quoted 6 lines
Quoted 2 lines

Plus it broke all the drivers that create platform devices before registering drivers or the ones simply not using private data.

No, this was just a suggestion for an ALSA specific workaround. I was suggesting ALSA drivers could do this.

Yes, I am sorry - I misread the code snipped as if it was in driver core, not in ALSA itself.

--
Dmitry

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid0944&bid$1720&dat1642

#34 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method. It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

Signature · 1 line
Dmitry
#35 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 11 lines · Dmitry Torokhov
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Takashi, how would you feel about a setup for ALSA going like:

static struct platform_device *devices[SNDRV_CARDS];

static int __devinit snd_foo_probe(int i)

{
	struct snd_card *card;

        if (port[i] == SNDRV_AUTO_PORT) {
                snd_printk(KERN_ERR "specify port\n");
		return -EINVAL;
        }

	card = snd_card_new(index[i], id[i], THIS_MODULE, 0);
        if (!card) {
                snd_printk(KERN_ERR "could not create card\n");
                return -EWHATEVER;
        }

	/* all the other things the probe method does */

        if (snd_card_register(card) < 0) {
                snd_printk(KERN_ERR "could not register card\n");
                return -EWHATEVERSNDCARDREGISTERRETURNED;
        }

        devices[i] = platform_device_register_simple(NAME, i, NULL, 0);
        if (IS_ERR(devices[i])) {
                snd_printk(KERN_ERR "could not register device\n");
		snd_card_free(card);
                return PTR_ERR(devices[i]);
        }

        platform_set_drvdata(devices[i], card);
        return 0;
}

Then deleting the .probe method from the foo_platform_driver struct and in the module init simply do a manual probe, using:

        if (platform_driver_register(&snd_foo_driver) < 0) {
                snd_printk(KERN_ERR "could not register driver\n");
                return -EARGHH;
        }

        for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
                if (enable[i] && snd_foo_probe(i) >= 0)
                        cards++

        if (!cards) {
                printk(KERN_ER "foo not found or device busy\n");
                platform_driver_unregister(&snd_foo_driver);
                return -ENODEV;
        }

Also attached as a patch against adlib.c (it's the simplest driver by far, so using that one as an example. it's present in 2.6.17-rc1).

The platform_device registration could also stay in the init loop, but since we want the free the card again if it fails same as with all the other failures it's best off at the end of probe() I feel. You introduced all these nonpnp_probe() methods for platform devices so we do have a nice place to put this...

Would you feel this would be an okay setup, assuming we're not getting that .discover method?

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#36 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 11 lines · Dmitry Torokhov
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Takashi, how would you feel about a setup for ALSA going like:

static struct platform_device *devices[SNDRV_CARDS];

static int __devinit snd_foo_probe(int i)

{
	struct snd_card *card;

        if (port[i] == SNDRV_AUTO_PORT) {
                snd_printk(KERN_ERR "specify port\n");
		return -EINVAL;
        }

	card = snd_card_new(index[i], id[i], THIS_MODULE, 0);
        if (!card) {
                snd_printk(KERN_ERR "could not create card\n");
                return -EWHATEVER;
        }

	/* all the other things the probe method does */

        if (snd_card_register(card) < 0) {
                snd_printk(KERN_ERR "could not register card\n");
                return -EWHATEVERSNDCARDREGISTERRETURNED;
        }

        devices[i] = platform_device_register_simple(NAME, i, NULL, 0);
        if (IS_ERR(devices[i])) {
                snd_printk(KERN_ERR "could not register device\n");
		snd_card_free(card);
                return PTR_ERR(devices[i]);
        }

        platform_set_drvdata(devices[i], card);
        return 0;
}

Then deleting the .probe method from the foo_platform_driver struct and in the module init simply do a manual probe, using:

        if (platform_driver_register(&snd_foo_driver) < 0) {
                snd_printk(KERN_ERR "could not register driver\n");
                return -EARGHH;
        }

        for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
                if (enable[i] && snd_foo_probe(i) >= 0)
                        cards++

        if (!cards) {
                printk(KERN_ER "foo not found or device busy\n");
                platform_driver_unregister(&snd_foo_driver);
                return -ENODEV;
        }

Also attached as a patch against adlib.c (it's the simplest driver by far, so using that one as an example. it's present in 2.6.17-rc1).

The platform_device registration could also stay in the init loop, but since we want the free the card again if it fails same as with all the other failures it's best off at the end of probe() I feel. You introduced all these nonpnp_probe() methods for platform devices so we do have a nice place to put this...

Would you feel this would be an okay setup, assuming we're not getting that .discover method?

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#37 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 29 lines · Rene Herman
Quoted 11 lines · Dmitry Torokhov
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

Quoted 9 lines

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

You need to let go of the model that driver that drives hardware also do the device discovery and it will all fall into place. While it may be contained in the same source module it is 2 different code chunks (for the lack of the better word) and we should not mix them together.

Quoted 7 lines
Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Quoted 8 lines

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Unfortunately it breaks manual driver binding/unbinding through sysfs so I don't think it is a good long-term solution.

--
Dmitry

-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language that extends applications into web and mobile media. Attend the live webcast and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid0944&bid$1720&dat1642

#38 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman
Quoted 29 lines · Rene Herman
Quoted 11 lines · Dmitry Torokhov
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

Quoted 9 lines

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

You need to let go of the model that driver that drives hardware also do the device discovery and it will all fall into place. While it may be contained in the same source module it is 2 different code chunks (for the lack of the better word) and we should not mix them together.

Quoted 7 lines
Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Quoted 8 lines

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Unfortunately it breaks manual driver binding/unbinding through sysfs so I don't think it is a good long-term solution.

--
Dmitry

#39 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

Quoted 2 lines · Dmitry Torokhov

You need to let go of the model that driver that drives hardware also do the device discovery and it will all fall into place.

It's not possible to sanely let go of that model. ISA discovery involves (as an example) poking at the same I/O ports as using does, meaning you will have to share a request_region() over discovery and use for one -- you can't decouple that due to obvious races.

Quoted 8 lines
Quoted 5 lines

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Unfortunately it breaks manual driver binding/unbinding through sysfs so I don't think it is a good long-term solution.

Yes. I don't see a significantly cleaner solution then than the slightly hackish "using drvdata as a private success flag" that I posted before. Example patch versus snd_adlib attached again. This seems to work well.

Takashi: do you agree? If the probe() method return is not going to be propagated up, there are few other options.

(Continuing the loop on IS_ERR(device) is then also a bit questionable again as the IS_ERR then signifies an eror in the bowels of the device model, but I feel it's still the correct thing to do)

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#40 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov
Quoted 2 lines · Dmitry Torokhov

You need to let go of the model that driver that drives hardware also do the device discovery and it will all fall into place.

It's not possible to sanely let go of that model. ISA discovery involves (as an example) poking at the same I/O ports as using does, meaning you will have to share a request_region() over discovery and use for one -- you can't decouple that due to obvious races.

Quoted 8 lines
Quoted 5 lines

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Unfortunately it breaks manual driver binding/unbinding through sysfs so I don't think it is a good long-term solution.

Yes. I don't see a significantly cleaner solution then than the slightly hackish "using drvdata as a private success flag" that I posted before. Example patch versus snd_adlib attached again. This seems to work well.

Takashi: do you agree? If the probe() method return is not going to be propagated up, there are few other options.

(Continuing the loop on IS_ERR(device) is then also a bit questionable again as the IS_ERR then signifies an eror in the bowels of the device model, but I feel it's still the correct thing to do)

Rene.

sound/isa/adlib.c +7 -4
Index: local/sound/isa/adlib.c
===================================================================
@@ -43,8 +43,7 @@ static int __devinit snd_adlib_probe(str
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
-	int error;
-	int i = device->id;
+	int error, i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -95,8 +94,7 @@ static int __devinit snd_adlib_probe(str
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -134,6 +132,11 @@ static int __init alsa_card_adlib_init(v
 		if (IS_ERR(device))
 			continue;
 
+		if (!platform_get_drvdata(device)) {
+			platform_device_unregister(device);
+			continue;
+		}
+
 		devices[i] = device;
 		cards++;
 	}
#41 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov

Dmitry Torokhov wrote:

[ crap, attached the wrong patch, this time with correct attachment ]

Quoted 11 lines · Rene Herman
Quoted 7 lines

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Takashi, how would you feel about a setup for ALSA going like:

static struct platform_device *devices[SNDRV_CARDS];

static int __devinit snd_foo_probe(int i)

{
	struct snd_card *card;

         if (port[i] == SNDRV_AUTO_PORT) {
                 snd_printk(KERN_ERR "specify port\n");
		return -EINVAL;
         }

	 card = snd_card_new(index[i], id[i], THIS_MODULE, 0);
         if (!card) {
                 snd_printk(KERN_ERR "could not create card\n");
                 return -EWHATEVER;
         }

	 /* all the other things the probe method does */

         if (snd_card_register(card) < 0) {
                 snd_printk(KERN_ERR "could not register card\n");
                 return -EWHATEVERSNDCARDREGISTERRETURNED;
         }

         devices[i] = platform_device_register_simple(NAME, i, NULL, 0);
         if (IS_ERR(devices[i])) {
                 snd_printk(KERN_ERR "could not register device\n");
		snd_card_free(card);
                 return PTR_ERR(devices[i]);
         }

         platform_set_drvdata(devices[i], card);
         return 0;
}

Then deleting the .probe method from the foo_platform_driver struct and in the module init simply do a manual probe, using:

         if (platform_driver_register(&snd_foo_driver) < 0) {
                 snd_printk(KERN_ERR "could not register driver\n");
                 return -EARGHH;
         }

         for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
                 if (enable[i] && snd_foo_probe(i) >= 0)
                         cards++

         if (!cards) {
                 printk(KERN_ER "foo not found or device busy\n");
                 platform_driver_unregister(&snd_foo_driver);
                 return -ENODEV;
         }

Also attached as a patch against adlib.c (it's the simplest driver by far, so using that one as an example. it's present in 2.6.17-rc1).

The platform_device registration could also stay in the init loop, but since we want the free the card again if it fails same as with all the other failures it's best off at the end of probe() I feel. You introduced all these nonpnp_probe() methods for platform devices so we do have a nice place to put this...

Would you feel this would be an okay setup, assuming we're not getting that .discover method?

Rene.

sound/isa/adlib.c +17 -23
Index: local/sound/isa/adlib.c
===================================================================
@@ -38,13 +38,12 @@ static void snd_adlib_free(struct snd_ca
 	release_and_free_resource(card->private_data);
 }
 
-static int __devinit snd_adlib_probe(struct platform_device *device)
+static int __devinit snd_adlib_probe(int i)
 {
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
 	int error;
-	int i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -91,12 +90,18 @@ static int __devinit snd_adlib_probe(str
 		goto out1;
 	}
 
-	platform_set_drvdata(device, card);
+	devices[i] = platform_device_register_simple(DRV_NAME, i, NULL, 0);
+	if (IS_ERR(devices[i])) {
+		snd_printk(KERN_ERR DRV_NAME ": could not register device\n");
+		error = PTR_ERR(devices[i]);
+		goto out1;
+	}
+
+	platform_set_drvdata(devices[i], card);
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -107,9 +112,7 @@ static int __devexit snd_adlib_remove(st
 }
 
 static struct platform_driver snd_adlib_driver = {
-	.probe		= snd_adlib_probe,
 	.remove		= __devexit_p(snd_adlib_remove),
-
 	.driver		= {
 		.name	= DRV_NAME
 	}
@@ -117,26 +120,17 @@ static struct platform_driver snd_adlib_
 
 static int __init alsa_card_adlib_init(void)
 {
-	int i, cards;
+	int error, cards, i;
 
-	if (platform_driver_register(&snd_adlib_driver) < 0) {
+	error = platform_driver_register(&snd_adlib_driver);
+	if (error < 0) {
 		snd_printk(KERN_ERR DRV_NAME ": could not register driver\n");
-		return -ENODEV;
+		return error;
 	}
 
-	for (cards = 0, i = 0; i < SNDRV_CARDS; i++) {
-		struct platform_device *device;
-
-		if (!enable[i])
-			continue;
-
-		device = platform_device_register_simple(DRV_NAME, i, NULL, 0);
-		if (IS_ERR(device))
-			continue;
-
-		devices[i] = device;
-		cards++;
-	}
+	for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
+		if (enable[i] && snd_adlib_probe(i) >= 0)
+			cards++
 
 	if (!cards) {
 #ifdef MODULE
#42 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Dmitry Torokhov

Dmitry Torokhov wrote:

[ crap, attached the wrong patch, this time with correct attachment ]

Quoted 11 lines · Rene Herman
Quoted 7 lines

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method.

Right...

Quoted 3 lines

It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

How do you feel about the flag that I've been proposing that a driver that needs to probe for its hardware could set and that says "if we return an error (or specifically -ENODEV I guess) the hardware's reallyreally not present and the device should not register"?

Greg, and how do you feel about such a flag?

As an alternative to the flag, how would either of you feel about either

1) adding a .discover method to struct device_driver that noone other than drivers for this old non generically discoverable hardware would set but which would make a registration fail if set and returning < 0?

2) adding that method only to platform_driver?

3) ... and after a few months when people aren't paying attention anymore rename .probe to .init and .discover back to .probe? ;-)

Quoted 4 lines · you

Note also that this patch will not do what the ALSA folk want - they want to know if the device was found or whether the probe returned -ENODEV. We knock -ENODEV and -ENXIO to zero in driver_probe_device(), so they won't even see that.

Yes, I know about the -ENODEV / -ENXIO thing. I asked for comment on that as well, but haven't gotten any.

Anyways, the additional method would, I feel, be the conceptually cleanest approach. Practically speaking though, simply doing a manual probe and only calling platform_register() after everything is found to be present and accounted for is not much of a problem either.

Takashi, how would you feel about a setup for ALSA going like:

static struct platform_device *devices[SNDRV_CARDS];

static int __devinit snd_foo_probe(int i)

{
	struct snd_card *card;

         if (port[i] == SNDRV_AUTO_PORT) {
                 snd_printk(KERN_ERR "specify port\n");
		return -EINVAL;
         }

	 card = snd_card_new(index[i], id[i], THIS_MODULE, 0);
         if (!card) {
                 snd_printk(KERN_ERR "could not create card\n");
                 return -EWHATEVER;
         }

	 /* all the other things the probe method does */

         if (snd_card_register(card) < 0) {
                 snd_printk(KERN_ERR "could not register card\n");
                 return -EWHATEVERSNDCARDREGISTERRETURNED;
         }

         devices[i] = platform_device_register_simple(NAME, i, NULL, 0);
         if (IS_ERR(devices[i])) {
                 snd_printk(KERN_ERR "could not register device\n");
		snd_card_free(card);
                 return PTR_ERR(devices[i]);
         }

         platform_set_drvdata(devices[i], card);
         return 0;
}

Then deleting the .probe method from the foo_platform_driver struct and in the module init simply do a manual probe, using:

         if (platform_driver_register(&snd_foo_driver) < 0) {
                 snd_printk(KERN_ERR "could not register driver\n");
                 return -EARGHH;
         }

         for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
                 if (enable[i] && snd_foo_probe(i) >= 0)
                         cards++

         if (!cards) {
                 printk(KERN_ER "foo not found or device busy\n");
                 platform_driver_unregister(&snd_foo_driver);
                 return -ENODEV;
         }

Also attached as a patch against adlib.c (it's the simplest driver by far, so using that one as an example. it's present in 2.6.17-rc1).

The platform_device registration could also stay in the init loop, but since we want the free the card again if it fails same as with all the other failures it's best off at the end of probe() I feel. You introduced all these nonpnp_probe() methods for platform devices so we do have a nice place to put this...

Would you feel this would be an okay setup, assuming we're not getting that .discover method?

Rene.

sound/isa/adlib.c +17 -23
Index: local/sound/isa/adlib.c
===================================================================
@@ -38,13 +38,12 @@ static void snd_adlib_free(struct snd_ca
 	release_and_free_resource(card->private_data);
 }
 
-static int __devinit snd_adlib_probe(struct platform_device *device)
+static int __devinit snd_adlib_probe(int i)
 {
 	struct snd_card *card;
 	struct snd_opl3 *opl3;
 
 	int error;
-	int i = device->id;
 
 	if (port[i] == SNDRV_AUTO_PORT) {
 		snd_printk(KERN_ERR DRV_NAME ": please specify port\n");
@@ -91,12 +90,18 @@ static int __devinit snd_adlib_probe(str
 		goto out1;
 	}
 
-	platform_set_drvdata(device, card);
+	devices[i] = platform_device_register_simple(DRV_NAME, i, NULL, 0);
+	if (IS_ERR(devices[i])) {
+		snd_printk(KERN_ERR DRV_NAME ": could not register device\n");
+		error = PTR_ERR(devices[i]);
+		goto out1;
+	}
+
+	platform_set_drvdata(devices[i], card);
 	return 0;
 
 out1:	snd_card_free(card);
- out0:	error = -EINVAL; /* FIXME: should be the original error code */
-	return error;
+out0:	return error;
 }
 
 static int __devexit snd_adlib_remove(struct platform_device *device)
@@ -107,9 +112,7 @@ static int __devexit snd_adlib_remove(st
 }
 
 static struct platform_driver snd_adlib_driver = {
-	.probe		= snd_adlib_probe,
 	.remove		= __devexit_p(snd_adlib_remove),
-
 	.driver		= {
 		.name	= DRV_NAME
 	}
@@ -117,26 +120,17 @@ static struct platform_driver snd_adlib_
 
 static int __init alsa_card_adlib_init(void)
 {
-	int i, cards;
+	int error, cards, i;
 
-	if (platform_driver_register(&snd_adlib_driver) < 0) {
+	error = platform_driver_register(&snd_adlib_driver);
+	if (error < 0) {
 		snd_printk(KERN_ERR DRV_NAME ": could not register driver\n");
-		return -ENODEV;
+		return error;
 	}
 
-	for (cards = 0, i = 0; i < SNDRV_CARDS; i++) {
-		struct platform_device *device;
-
-		if (!enable[i])
-			continue;
-
-		device = platform_device_register_simple(DRV_NAME, i, NULL, 0);
-		if (IS_ERR(device))
-			continue;
-
-		devices[i] = device;
-		cards++;
-	}
+	for (cards = 0, i = 0; i < SNDRV_CARDS; i++)
+		if (enable[i] && snd_adlib_probe(i) >= 0)
+			cards++
 
 	if (!cards) {
 #ifdef MODULE
#43 lore raw

Re: patch bus_add_device-losing-an-error-return-from-the-probe-method.patch added to gregkh-2.6 tree

In reply to Rene Herman
Quoted 6 lines · Rene Herman

To Dmitry, I see you saying "probe() failing is driver's problem. The device is still there and should still be presented in sysfs.". No, at least in the case of these platform drivers (or at least these old ISA cards using the platform driver interface), a -ENODEV return from the probe() method would mean the device is _not_ present (or not found at least). NODEV.

Or you could separate device probing code from driver->probe(). BTW I think that ->probe() is not the best name for that method. It really is supposed to allocate resources and initialize the device so that it is ready to be used, not to verify that device is present. The code that created device shoudl've done that.

Signature · 8 lines
Dmitry


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642