Re: 2.6.21-rc7-mm1 BUG at kernel/sched-clock.c:175 init_sched_clock()

[Tejun Heo <htejun@gmail.com>]

Andrew Morton wrote:
>> No, that would break host probing.  The port is in frozen (controller
>> initialized and IRQs masked off) state, so it's not allowed to take
>> interrupt.
> 
> Sounds dodgy.  What happens if the IRQ line is shared with some other
> device?  We'll enter ahci_interrupt().  We've alread set up ->port_map and
> ->n_ports so we're wholly dependent upon that read from HOST_IRQ_STAT
> returning zeroes.

Shared IRQ is okay because ahci has pending IRQ register which is masked
by ->freeze().  Controllers which don't have such feature check qc
status to determine whether it's expecting IRQ, so it's okay for them
too.  So, AFAIK, libata is okay with shared IRQs.

>         /* sigh.  0xffffffff is a valid return from h/w */
> 
> if that happens we're dead, aren't we?

Yes, we are.  :-)

>>  If interrupt triggers at this point, it's low level driver
>> bug.  Berck, how reliably can you reproduce this problem?  Can you post
>> the result of 'lspci -nn'?
> 
> Berck doesn't apepar to be sharing that IRQ, so it'll be something else.

I'm pretty sure Berck's problem is caused by ->freeze() not clearing
pending IRQ bit.  We haven't been too strict about ->freeze().
Considering the above 0xffffffff case, I agree it's better to fix it in
the core instead of fixing each LLD.  Will brew up another patch.

Thanks.

-- 
tejun