On (19/11/11 10:12), Petr Mladek wrote:
[..]

quoted

I do recall that we talked about per-CPU printk state bit which would
start/end "just print it" section. We probably can extend it to "just
log_store" type of functionality. Doesn't look like a very bad idea.

The problem with per-CPU printk is that we would need to disable
interrupts.

Or disable preemption and have loglevel per-CPU and per-context.
preempt_count can navigate us to the right context loglevel on
particular CPU. I'm talking here only about backtrace (error)
reporting contexts. Those can be atomic perfectly fine.

I posted a silly code snippet.

[..]

But yes, this per-code-section loglevel is problematic. The feedback
against the patchset shows that people want it also the other way.
I mean to keep pr_debug() as pr_debug().

Hmm. Right.

A solution might be to use the per-code-section loglevel only instead
of some special loglevel.

So maybe we can "overwrite" only KERN_DEFAULT loglevels?
We certainly should not mess with SCHED or with anything in between
EMERG and ERR.

The explicitly passed loglevel makes me feel more confident that
all needed printk() calls were updated. But it might be a false
feeling. I do not really have any strong preference.

I'm not like really objecting, just trying to explore some other
options.

	-ss