Re: Can't cleanly unload driver

[Eric Fowler <eric.fowler@xxxxxxxxx>]

So, I finally figured this out, and for the benefit of web posterity and the search engines, I'm gonna lay it out:

I was using *both* the register_chrdev() and the cdev_*() functions on the same device. 

I guess you can't do that. I can't, anyway. Rubini's scull code uses both, and it looks like he uses both on the same device, although I bet he doesn't (haven't looked closely at the code). 

I switched to using register_chrdev()/unregister_chrdev() and all is happy now. 

Thanks to all who helped. 

On Mon, Dec 30, 2013 at 10:57 AM, Eric Fowler <eric.fowler@xxxxxxxxx> wrote:

Actually, I was working with a homegrown makefile. I made the change you recommended and am still having the problem. 

I am thinking now about something else as being the problem.

Is this code legitimate?

//// file scope 

struct cdev my_cdev; 

///inside init function

cdev_init(&my_cdev, ....); 

cdev_add(&my_cdev, ....); 

///exit fxn

cdev_del(&my_cdev); 

In other words, the memory for the cdev struct comes from static memory for the driver, not a call to cdev_alloc() or kmalloc(). 

On Mon, Dec 30, 2013 at 10:18 AM, Rajat Sharma <fs.rajat@xxxxxxxxx> wrote:

Hi Eric,

I have seen some errors with module reference counting with a nicely written code, but culprit for my case was a missing compilation flag -DMODULE which gives definition of THIS_MODULE, otherwise it is null e.g. for modules which are compiled in kernel, so they are never unloaded. Unless you have some customization done to Makefiles, this definition should be included, but its anyways good to double check and rule out this possibility.

-Rajat

On Mon, Dec 30, 2013 at 9:48 AM, Eric Fowler <eric.fowler@xxxxxxxxx> wrote:

Still working on this. Here is some dmesg spew:

[  514.245846] foobar: module verification failed: signature and/or required key missing - tainting kernel

[  514.245937] kobject: 'foobar' (f7f060c8): kobject_add_internal: parent: 'module', set: 'module'

[  514.245951] kobject: 'holders' (f5ff3d40): kobject_add_internal: parent: 'foobar', set: '<NULL>'

[  514.245981] kobject: 'notes' (f2d25f80): kobject_add_internal: parent: 'foobar', set: '<NULL>'

[  514.245987] kobject: 'foobar' (f7f060c8): kobject_uevent_env

[  514.245998] kobject: 'foobar' (f7f060c8): fill_kobj_path: path = '/module/foobar'

So it looks like kernel validation is failing. I have printk's in my init fxn that are never turning up in /var/log/messages, until, weirdly, AFTER I remove the device:

<insmod device>

Dec 30 09:43:03 localhost kernel: [  514.245846] foobar: module verification failed: signature and/or required key missing - tainting kernel

Dec 30 09:43:16 localhost fprintd[1085]: ** Message: No devices in use, exit

<rmmod device>

Dec 30 09:45:53 localhost kernel: [  514.249323] foobar: got device number 248, minor is 0   <<<<----THIS IS IN init() fxn

Dec 30 09:45:53 localhost kernel: [  684.102912] unregister_chrdev(248) called for foobar<7>[  684.102927] kobject: '(null)' (f7f06220): kobject_cleanup, parent   (null)

<insmod>

insmod: ERROR: could not insert module ./foobar.ko: Device or resource busy

On Fri, Dec 27, 2013 at 9:13 PM, <Valdis.Kletnieks@xxxxxx> wrote:

On Fri, 27 Dec 2013 19:33:50 -0800, Eric Fowler said:

> I suspect I am doing something wrong in the code with
> register/unregister_chrdev(), but I have been over that code a million
> times. It looks fine.
>
> Now:
> insmod the device, OK
> rmmod the device, OK
> Check /proc/devices , device # is present
> insmod the device again, fails with ERROR: could not insert module
> ./foobar.ko: Device or resource busy

It does smell like an unregister issue.  You may want to try adding
printk() calls to print out the return code from register and unregister.
I'm willing to bet that (a) the unegister is failing because somebody
still has a reference on the device, and (b) the second register call fails
because the device already exists, causing your module_init() to bail out.

The fun is that you may not have taken a reference on the device directly
yourself - you may have called some other get_foo() that ends up taking an
implicit reference under the covers, causing issues when you fail to call
put_foo() at the right place...

--

cc:NSA

_______________________________________________
Kernelnewbies mailing list
Kernelnewbies@xxxxxxxxxxxxxxxxx
http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies

--

cc:NSA

--

cc:NSA

_______________________________________________
Kernelnewbies mailing list
Kernelnewbies@xxxxxxxxxxxxxxxxx
http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies