debug keys (namely '*') vs. NMI watchdog (and softirqs)

debug keys (namely '*') vs. NMI watchdog (and softirqs)

[Jan Beulich]

Processing particularly of '*' may take many seconds, and thus
interfere with the NMI watchdog (immediately terminal) as well as
handling of other softirqs, namely the time related ones (possibly
getting the system into bad state as we know from past
experience).

For the NMI watchdog, adding a Linux-like touch_nmi_watchdog()
would certainly help, but it would leave the softirq part
unaddressed. For that, in -unstable it would seem feasible to simply
sprinkle in calls to process_pending_softirqs() (e.g. between each
individual handler being called), but it's not immediately clear
whether the same would also hold for 4.0 and earlier, as tasklets
(and hence the handling of the non-IRQ key handlers) there get
processed in a softirq action handler. Would it be correct to have
yet another flavor of the wrappers around __do_softirq() for this
purpose, also filtering out TASKLET_SOFTIRQ?

Thanks, Jan

Re: debug keys (namely '*') vs. NMI watchdog (and softirqs)

[Keir Fraser]

On 14/12/2010 09:48, "Jan Beulich" <JBeulich@novell.com> wrote:

> Processing particularly of '*' may take many seconds, and thus
> interfere with the NMI watchdog (immediately terminal) as well as
> handling of other softirqs, namely the time related ones (possibly
> getting the system into bad state as we know from past
> experience).
> 
> For the NMI watchdog, adding a Linux-like touch_nmi_watchdog()
> would certainly help, but it would leave the softirq part
> unaddressed. For that, in -unstable it would seem feasible to simply
> sprinkle in calls to process_pending_softirqs() (e.g. between each
> individual handler being called), but it's not immediately clear
> whether the same would also hold for 4.0 and earlier, as tasklets
> (and hence the handling of the non-IRQ key handlers) there get
> processed in a softirq action handler. Would it be correct to have
> yet another flavor of the wrappers around __do_softirq() for this
> purpose, also filtering out TASKLET_SOFTIRQ?

For 4.0.2? Yes, that's probably best. As you say it's not needed for
xen-unstable since tasklets execute in vcpu context there.

 -- Keir