Re: [RFC] irq: Skip printing irq when desc->action is null even if any_count is not zero

[Neil Horman <nhorman@tuxdriver.com>]

On Wed, Jan 22, 2020 at 01:42:48PM +0100, Thomas Gleixner wrote:
> Chao,
> 
> l00520965 <liuchao173@huawei.com> writes:
> 
> > When desc->action is empty, there is no need to print out the irq and its'
> > count in each cpu. The desc is not alloced in request_irq or freed in
> > free_irq.
> 
> request/free_irq() never allocate/free irq descriptors. 
> 
> > So some PCI devices, such as rtl8139, uses request_irq and free_irq,
> 
> All PCI devices use some variant of request_irq()/free_irq(). The
> interrupt descriptors are allocated by the underlying PCI
> machinery. They are only allocated/freed when the device driver is
> loaded/removed.
> 
> And this property exists for _ALL_ interrupts independent of PCI.
> 
> > which only modify the action of desc. So /proc/interrupts could be
> > like this:
> 
> I think you want to explain:
> 
>   If an interrupt is released via free_irq() without removing the
>   underlying irq descriptor, the interrupt count of the irq descriptor
>   is not reset. /proc/interrupt shows such interrupts with an empty
>   action handler name:
>   
> >            CPU0       CPU1
> >  38:         46          0     GICv3  36 Level     ehci_hcd:usb1
> >  39:         66          0     GICv3  37 Level
> 
>   irqbalance fails to detect that this interrupt is not longer in use
>   and parses the last word in the line 'Level' as the action handler
>   name.
> 
> > Irqbalance gets the list of interrupts according to /proc/interrupts. In
> > this case, irqbalance does not remove the interrupt from the balance list,
> > and the last string in this line,which is Level, is used as irq_name.
> 
> Right, this is historic behaviour and I don't know how irqbalance dealt
> with that in the past 20+ years. At least I haven't seen any complaints.
> 
> I'm not opposed to suppress the output, but I really want the opinion of
> the irqbalance maintainers on that.
> 
Actually, irqbalance ignores the trailing irq name (or it should at
least), so you should be able to drop that portion of /proc/irqbalance,
though I cant speak for any other users of it.

> > Or we can clear desc->kstat_irqs in each cpu in free_irq when
> > desc->action is null?
> 
> No, we can't. The historic behaviour is that the total interrupt count
> for a device is maintained independent of the number of
> request/free_irq() pairs.
> 
> > Signed-off-by: LiuChao <liuchao173@huawei.com>
> > Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
> 
> I really can't remember that I have reviewed this patch already. Please
> don't add tags which claim that some one has reviewed or acked your
> patch unless you really got that Reviewed-by or Acked-by from that
> person.
> 
> Thanks,
> 
>         tglx
>