Re: softlockup interaction with slow consoles

[Ingo Molnar <mingo@redhat.com>]

On Mon, 20 Feb 2006, David S. Miller wrote:

> When register_console() runs, it dumps the entire log buffer to the
> console device with interrupts off.
> 
> If you're going over a slow console, this can easily take more than 10
> seconds especially on SMP with many cpus brought up.
> 
> This makes the softlock fire, which is terribly annoying :-)
> 
> I think the bug is in the console registry code, I think it should
> capture chunks of the existing console buffer into some local memory and
> push things piece by piece with interrupts enabled to the console
> driver(s).
> 
> Any better ideas?

is this a boot-time-only problem? In recent updates to the softlockup 
code in -mm:

 timer-irq-driven-soft-watchdog-cleanups.patch
 timer-irq-driven-soft-watchdog-percpu-race-fix.patch
 timer-irq-driven-soft-watchdog-percpu-fix.patch
 timer-irq-driven-soft-watchdog-boot-fix.patch

i changed soft lockup detection to be turned off during bootup. That
should work around any boot-time warnings.

(if this can happen on a booted up system then the real fix would indeed
be to split up register_console()'s workload - that would also make it
more preemption-friendly. But at first sight it looks quite complex to
do.)

	Ingo