Re: SCSI aic7xxx driver: Initialization Failure over a kdump reboot

[Marcelo Tosatti <marcelo.tosatti@cyclades.com>]

On Sun, Jan 30, 2005 at 07:27:26PM +0100, Arjan van de Ven wrote:
> On Sun, 2005-01-30 at 13:27 -0200, Marcelo Tosatti wrote:
> > On Fri, Jan 07, 2005 at 04:38:32AM +0000, Matthew Wilcox wrote:
> > > On Thu, Jan 06, 2005 at 11:53:27PM -0500, Lukasz Kosewski wrote:
> > > > I have an idea of something I might do for 2.6.11, but I doubt anyone
> > > > will actually agree with it.  Say we keep a counter of how many times
> > > > interrupt x has been fired off since the last timer interrupt
> > > > (obviously, a timer interrupt resets the counter).  Then we can pick an
> > > > arbitrary threshold for masking out this interrupt until another device
> > > > actually pines for it.
> > > > 
> > > > Or something.  The point is, we need a general solution to the problem,
> > > > not poking about in every single driver trying to tie it down.
> > > 
> > > Something like note_interrupt() in kernel/irq/spurious.c?
> > 
> > BTW I wonder if its feasible to add an interface on top of kernel/irq/spurious.c for 
> > notifying drivers about interrupts storms, so they can take appropriate action 
> > (try to reset the device).
> 
> the problem is... the driver just denied it was their irq (at least in
> all the common cases)... 

Hum, drivers should, at least in theory, be able to return IRQ_NONE if interrupts
can't be handled.

So is 8390 a special case? 

drivers/net/8390.c 
irqreturn_t ei_interrupt(int irq, void *dev_id, struct pt_regs * regs)
{
...

        }
        spin_unlock(&ei_local->page_lock);
        return IRQ_RETVAL(nr_serviced > 0);
}


The "workaround" looks like (at the end of ei_interrupt):

        if (!nr_serviced)
                interrupt_cnt++;
        else
                interrupt_cnt = 0;
                                                                                
        if (interrupt_cnt > 256) {
                ei_status.reset_8390(dev);
                interrupt_cnt = 0;
        }


One could argue that it is a hardware problem...