On 07.03.2018 16:13, Greg Kroah-Hartman wrote:
> On Wed, Mar 07, 2018 at 10:49:02AM +0100, Florian Schmaus wrote:
>> I triggerd the BUG_ON(), which was added in
>> f48f3febb2cbfd0f2ecee7690835ba745c1034a4, when booting a domU Xen
>> domain.
> 
> How?

Basically just by booting the domU kernel with the wmi-bmof driver
compiled in. After I deselected the wmi-bmof driver, the kernel booted
just fine.

I'd probably be able to re-create the faulty kernel and show you the
image and/or kernel config if you want.

>> Since there was no contextual information logged, I needed to
>> attach kgdb to determine the culprit (the wmi-bmof driver in my case).
>>
>> Signed-off-by: Florian Schmaus <flo@geekplace.eu>
>> ---
>>  drivers/base/driver.c | 6 +++++-
>>  1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/base/driver.c b/drivers/base/driver.c
>> index ba912558a510..55f8db41df2b 100644
>> --- a/drivers/base/driver.c
>> +++ b/drivers/base/driver.c
>> @@ -148,7 +148,11 @@ int driver_register(struct device_driver *drv)
>>  	int ret;
>>  	struct device_driver *other;
>>  
>> -	BUG_ON(!drv->bus->p);
>> +	if (!drv->bus->p) {
>> +		printk(KERN_ERR "Driver '%s' was unable to register bus_type\n",
>> +			   drv->name);
>> +		BUG();
>> +	}
> 
> We could just log the error and return with an error, that would be
> nicer, right?  Crashing the system isn't usually a good idea, even at
> boot time :)

It sure is, and I did consider it. I just wasn't sure which return value
to use in that case. Probably the best solution would be to use the
return value of bus_register(), which caused p to become NULL. Although
that possibly requires extending the bus_type struct by an 'int' storing
the return value (Or could the return value be mangled into the 'p'
pointer?).

I'd be happy to hear your thoughts on this.

BTW: Should I retain the 'unlikely' semantic of BUG_ON() in this case?

- Florian