Oops calling sysfs_create_link() from pci_probe()

Oops calling sysfs_create_link() from pci_probe()

[Luben Tuikov]

Hi,

I'm calling

sysfs_create_link(&class->kobj,
		  &pcidev->driver->driver.kobj, "driver");

To create a link from a syfs directory of an object which I've
created with class_device_regsiter(), to point to the
driver directory of the pci driver.
This is effectively called at the bottom of the pci_driver->probe
function.

But I get this oops:
 printing eip:
c0229e7b
*pde = 00000000
Oops: 0000 [#1]
PREEMPT SMP 
Modules linked in: aic94xx sas_class
CPU:    0
EIP:    0060:[<c0229e7b>]    Not tainted VLI
EFLAGS: 00010296   (2.6.12-rc6) 
EIP is at kref_get+0xb/0x50
eax: 00000060   ebx: 00000060   ecx: 00000000   edx: 0000002c
esi: e09342cb   edi: df02a997   ebp: fffffff4   esp: d5a81e0c
ds: 007b   es: 007b   ss: 0068
Process insmod (pid: 3166, threadinfo=d5a80000 task=df6bc540)
Stack: c014653b dfff5080 000000d0 df02a98c 00000048 c02293aa 00000060 d35257ac 
       c0195d77 00000048 000000d0 c022948e decdc040 c0229460 d335d4b0 c2a0b530 
       00000000 decdc020 decdc014 c0195e0f c2a0b530 e09342c4 00000048 e0935180 
Call Trace:
 [<c014653b>] __kmalloc+0x9b/0xd0
 [<c02293aa>] kobject_get+0x1a/0x30
 [<c0195d77>] sysfs_add_link+0x77/0xd0
 [<c022948e>] kobject_put+0x1e/0x30
 [<c0229460>] kobject_release+0x0/0x10
 [<c0195e0f>] sysfs_create_link+0x3f/0x70
 [<e093419e>] sas_register_ha+0x10e/0x160 [sas_class]
 [<e0a6cf67>] asd_pci_probe+0x6b7/0x760 [aic94xx]
 [<c0235a62>] pci_device_probe_static+0x52/0x70
 [<c0235abc>] __pci_device_probe+0x3c/0x50
 [<c0235afc>] pci_device_probe+0x2c/0x50
 [<c026d40f>] driver_probe_device+0x2f/0x80
 [<c026d55c>] driver_attach+0x5c/0x90
 [<c026da8e>] bus_add_driver+0x9e/0xd0
 [<c0235dbd>] pci_register_driver+0x7d/0xa0
 [<e0814044>] aic94xx_init+0x44/0x58 [aic94xx]
 [<c01389a3>] sys_init_module+0x223/0x250
 [<c0102fb5>] syscall_call+0x7/0xb
Code: 92 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 8b 44 24 04 c7 00 01 00 00 00 c3 90 8d 74 26 00 83 ec 14 89 5c 24 10 8b 5c 24 18 <8b> 03 85 c0 74 0b f0 ff 03 8b 5c 24 10 83 c4 14 c3 c7 04 24 fa 
 

Which suggests that I cannot call this from inside pci_probe(),
but will have to "wait" to call it after pci_regsiter_driver()
returns, effectively after pci_populate_driver_dir() returns.

Is this correct assumption? Or can I call the syslink call
above in other ways?

Thanks,
	Luben

Re: Oops calling sysfs_create_link() from pci_probe()

[Greg KH]

On Thu, Jun 16, 2005 at 03:57:34PM -0400, Luben Tuikov wrote:
> Hi,
> 
> I'm calling
> 
> sysfs_create_link(&class->kobj,
> 		  &pcidev->driver->driver.kobj, "driver");
> 
> To create a link from a syfs directory of an object which I've
> created with class_device_regsiter(), to point to the
> driver directory of the pci driver.

Ick, why?  Shouldn't something like this be done in the driver core, and
not in the individual drivers?

> This is effectively called at the bottom of the pci_driver->probe
> function.
> 
> But I get this oops:
>  printing eip:
> c0229e7b
> *pde = 00000000
> Oops: 0000 [#1]
> PREEMPT SMP 
> Modules linked in: aic94xx sas_class
> CPU:    0
> EIP:    0060:[<c0229e7b>]    Not tainted VLI
> EFLAGS: 00010296   (2.6.12-rc6) 
> EIP is at kref_get+0xb/0x50

Looks like one of the kobjects that you are wanting to link is not fully
initialized and registered with sysfs.  Where are you getting that
"&class->kobj" from?

Have a pointer to your patch anywhere?

Also, try turning on kobject and driver core debugging, you should get a
lot of helpful information in your syslog right before this oops.

thanks,

greg k-h

Re: Oops calling sysfs_create_link() from pci_probe()

[Luben Tuikov]

On 06/17/05 01:55, Greg KH wrote:
> On Thu, Jun 16, 2005 at 03:57:34PM -0400, Luben Tuikov wrote:
> 
>>Hi,
>>
>>I'm calling
>>
>>sysfs_create_link(&class->kobj,
>>		  &pcidev->driver->driver.kobj, "driver");
>>
>>To create a link from a syfs directory of an object which I've
>>created with class_device_regsiter(), to point to the
>>driver directory of the pci driver.
> 
> 
> Ick, why?  Shouldn't something like this be done in the driver core, and
> not in the individual drivers?

Hi Greg,

I agree, it should be done in the driver core.  The LLDD is registering
with the SAS class (unfinished, incomplete) and this is what appears:
/sys/class/sas/
/sys/class/sas/ha0/
/sys/class/sas/ha0/device -> link to PCI device
/sys/class/sas/ha0/device_name

where device name is the SAS address attribute (RO).
I'd ideally like to have a link to the LLDD in there
as well:

/sys/class/sas/ha0/driver -> link to the driver

But this host adapter registration with the SAS class happens
at pci_probe time, *at pci_register_driver* time so it seems that this
is the reason for the oops. (as opposed to PCI hot plugging the controller)
 
> Looks like one of the kobjects that you are wanting to link is not fully
> initialized and registered with sysfs.  Where are you getting that
> "&class->kobj" from?

It is the kobj of "ha0" which was just registered with class_device_register().
I suspect since all this is called from pci_probe at module init, it is
failing for the 2nd kobj, the driver.

Would this imply that had pci_probe been called on a PCI hot plug event
(not at pci_driver_register() time) then that symlink would've succeeded?
(since the driver had been registered already)

If so, can we reconcile this somehow so that code executed in pci_probe
at time B, could've also been executed at time A, A < B?

> Have a pointer to your patch anywhere?

It is quite incomplete.  Let me have something substantial and I'll post it
and then we can figure it out.  For now that line is /* XXX it would be good... */

> Also, try turning on kobject and driver core debugging, you should get a
> lot of helpful information in your syslog right before this oops.

Thanks,
	Luben