Re: fyi: scheduling while atomic dmesg output 3.12-rc1

["Arend van Spriel" <arend@broadcom.com>]

On 09/18/2013 03:57 PM, Hauke Mehrtens wrote:
> On 09/18/2013 11:19 AM, Arend van Spriel wrote:
>> On 09/17/2013 07:45 PM, Joe Perches wrote:
>>> <3>[   11.206312] BUG: scheduling while atomic:
>>> NetworkManager/866/0x00000200
>>
>> Thanks, Joe
>>
>> I got a report on this few days ago. It was introduced by bcma API
>> change and I already sent email to the committer of that change, ie.
>> Hauke Mehrtens. Hope it will be settled soon how to fix this.
>>
>> Gr. AvS
>
> Hi,
>
> I see four solutions for the problem:
>
> 1. convert the usleep_range(1000, 2000) into udelay(1000) in
> drivers/bcma/driver_pci.c
>
> 2. remove the call of bcma_core_pci_power_save() from bcma_core_pci_up()
> so that it does not get called by brcmsmac.
>
> 3. remove the call of bcma_core_pci_power_save() from bcma_core_pci_up()
> and move the call to somewhere out of the big spin lock.
>
> 4. convert the big brmcsmac spin lock into a mutex lock and use an
> additional spin lock for the parts where it is actually needed.
>
> For 3.12 I am for solution 1 or 2 and for the long term 3.13? I am for
> solution 4, but that needs bigger changes.

Agree. When looking into this I considered option 4 would be a bigger 
work, but I agree we should aim for that in the long term. For the short 
term I would say option 2 makes sense although I guess the power_save 
call is there for a reason. So I will also look if option 3 is doable.

Regards,
Arend