[QUESTION/sysfs] strange refcounting

[QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

Hello folks,

I'm writing a simple device driver and want to expose some of its 
attributes to userspace via sysfs.

As usually, I have main device description structure "struct 
mydev_info". I've embedded a struct device object there. What I do is:

struct mydev_info mydev
{
	struct device *dev;
	... bla bla bla ...
} mydev;


mydev->dev=kzalloc(sizeof(struct device), GFP_KERNEL);
mydev->dev->bus_id = "mydev";
mydev->dev->release = mydev_release;
err = device_register(&mydev->dev);

Then, I see /sys/devices/mydev/ in sysfs. I open pre-defined 
/sys/devices/mydev/power/state in userspace and don't close it.

Then I run lsmod, and see zero refcount to my module. Well, I run rmmod 
mymod, module is unloaded.

Then I close /sys/devices/mydev/power/state, and enjoy segfault.

I thought sysfs subsystem have to increase module refcount when one 
opens its sysfs files. Well, there is a release function, but it is also 
unloaded with the module.

May be there is a problem because of I have mydev->dev->parent == NULL, 
mydev->dev->bus == NULL, mydev->dev->driver == NULL? But I really don't 
have any bus, any parent and I don't want to introduce struct 
device_driver ...

Kernel is 2.6.15.1.

Although this is my first meet with sysfs, this looks strange.

Thanks.

-- 
Best regards, Artem B. Bityutskiy
Oktet Labs (St. Petersburg), Software Engineer.
+78124286709 (office) +79112449030 (mobile)
E-mail: dedekind@oktetlabs.ru, web: http://www.oktetlabs.ru

Re: [QUESTION/sysfs] strange refcounting

[Greg KH]

On Fri, Feb 03, 2006 at 05:16:22PM +0300, Artem B. Bityutskiy wrote:
> Hello folks,
> 
> I'm writing a simple device driver and want to expose some of its 
> attributes to userspace via sysfs.
> 
> As usually, I have main device description structure "struct 
> mydev_info". I've embedded a struct device object there. What I do is:
> 
> struct mydev_info mydev
> {
> 	struct device *dev;

First off, this should not be a pointer, but rather:
	struct device dev;

That properly embedds the struct device into your object.

> 	... bla bla bla ...
> } mydev;
> 
> 
> mydev->dev=kzalloc(sizeof(struct device), GFP_KERNEL);
> mydev->dev->bus_id = "mydev";
> mydev->dev->release = mydev_release;
> err = device_register(&mydev->dev);

What type of bus does this device live on?  You should not be calling
device_register() on your own directly.  Either use a bus, and be a
device of it, or use the platform_device() interface.

> Then, I see /sys/devices/mydev/ in sysfs. I open pre-defined 
> /sys/devices/mydev/power/state in userspace and don't close it.
> 
> Then I run lsmod, and see zero refcount to my module. Well, I run rmmod 
> mymod, module is unloaded.

Yup.

> Then I close /sys/devices/mydev/power/state, and enjoy segfault.

What is the backtrace?

> I thought sysfs subsystem have to increase module refcount when one 
> opens its sysfs files. Well, there is a release function, but it is also 
> unloaded with the module.

Again, register with a bus or use the platform_device() interface, and
this should work properly.

> May be there is a problem because of I have mydev->dev->parent == NULL, 
> mydev->dev->bus == NULL, mydev->dev->driver == NULL? But I really don't 
> have any bus, any parent and I don't want to introduce struct 
> device_driver ...

Yes, you kind of need all of that :)

Make the above changes and let us know if that helps things.

thanks,

greg k-h

[QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

Hello folks, 

I'm writing a simple device driver and want to expose some of its
attributes to userspace via sysfs. 

As usually, I have main device description structure "struct
mydev_info". I've embedded a struct device object there. What I do is: 

struct mydev_info mydev 
{ 
    struct device *dev; 
    ... bla bla bla ... 
} mydev; 


mydev->dev=kzalloc(sizeof(struct device), GFP_KERNEL); 
mydev->dev->bus_id = "mydev"; 
mydev->dev->release = mydev_release; 
err = device_register(&mydev->dev); 

Then, I see /sys/devices/mydev/ in sysfs. I open
pre-defined /sys/devices/mydev/power/state in userspace and don't close it. 

Then I run lsmod, and see zero refcount to my module. Well, I run rmmod
mymod, module is unloaded. 

Then I close /sys/devices/mydev/power/state, and enjoy segfault. 

I thought sysfs subsystem have to increase module refcount when one
opens its sysfs files. Well, there is a release function, but it is also
unloaded with the module. 

May be there is a problem because of I have mydev->dev->parent == NULL,
mydev->dev->bus == NULL, mydev->dev->driver == NULL? But I really don't
have any bus, any parent and I don't want to introduce struct
device_driver ... 

Kernel is 2.6.15.1. 

Although this is my first meet with sysfs, this looks strange. 

Thanks.

-- 
Best Regards,
Artem B. Bityuckiy,
St.-Petersburg, Russia.

Re: [QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

Artem B. Bityutskiy wrote:
> Then, I see /sys/devices/mydev/ in sysfs. I open
> pre-defined /sys/devices/mydev/power/state in userspace and don't close it. 
> 
> Then I run lsmod, and see zero refcount to my module. Well, I run rmmod
> mymod, module is unloaded. 
> 
> Then I close /sys/devices/mydev/power/state, and enjoy segfault. 
> 
I actually forgot to formulate my question: why module's refcount is not 
increased when somebody opens a sysfs file which belongs to this module? 
How to withstan to an unexpected module unload?

Thanks.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

Artem B. Bityutskiy wrote:
> I actually forgot to formulate my question: why module's refcount is not 
> increased when somebody opens a sysfs file which belongs to this module? 
> How to withstan to an unexpected module unload?
> 
> Thanks.
I see this code drivers/base/core.c, device_add().

if (dev->driver)
	dev->uevent_attr.attr.owner = dev->driver->owner;

I assume it is expected that I must have a driver structure. But I 
don't. Why do I have to?

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

As &struct device structure has no @owner field, and corresponding 
functions rely on the @owner field at &struct device_driver, I conclude 
that I cannot use &struct device objects without bus and device driver 
objects, just by design.

On the other hand, 'device_register()' accepts &struct device objects 
with NULL-filled @bus and @driver fields perfectly fine, does not 
complain, does not return any error, and I even see corresponding 
entries at /sys/devices/. But there is a refcounting problem described 
at my first mail.

This is obviously a confusing discrepancy. Sysfs has to either reject 
bus-less and driver-less &struct device objects or deal with them 
correctly. The latter is impossible due to lack of an @owner field in 
&struct device.

In connection with this, I have a question. There is a whole bunch of 
drivers which do not directly relate to hardware devices, but which 
still want to expose their parameters via sysfs. For example, this could 
be a filesystem, LVM, a compression layer on top of a file system of a 
block device, whatever. These are "virtual" devices and they are not 
physically connected to any bus. How should they deal with sysfs?

I see there is the "class" stuff in sysfs, but it seems that it is far 
not as flexible as the "device, driver, and bus" stuff, because I cannot 
create many nested layers within classes. I can create a class, which 
goes to /sys/class/, and devices within this class, which go to 
/sys/class/myclass/mydev/. But I cannot create a class, devices within 
that class, and daughter devices within them, like:

/sys/class/myclass/
            |-- mydev1/
                | -- doughterdev1/
                | -- doughterdev1/
                | -- ...
            |-- mydev2/
            |-- mydev3/
            |-- ...

Please, comment this.

Thanks.

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.

Re: [QUESTION/sysfs] strange refcounting

[Artem B. Bityutskiy]

Artem B. Bityutskiy wrote:
> In connection with this, I have a question. There is a whole bunch of 
> drivers which do not directly relate to hardware devices, but which 
> still want to expose their parameters via sysfs. For example, this could 
> be a filesystem, LVM, a compression layer on top of a file system of a 
> block device, whatever. These are "virtual" devices and they are not 
> physically connected to any bus. How should they deal with sysfs?

For some reasons I missed Greg's reply to my first message and was 
talking to myself. To end this thread, the answer to this question is: 
"include/sysfs.h".

-- 
Best Regards,
Artem B. Bityutskiy,
St.-Petersburg, Russia.