Re: [PATCH 3/4] printk: Defer printing to irq work when we printedtoo much

[Andrew Morton]

On Fri, 8 Nov 2013 11:21:13 +0100 Jan Kara <jack@xxxxxxx> wrote:

> On Fri 08-11-13 00:46:49, Frederic Weisbecker wrote:
> > On Thu, Nov 07, 2013 at 06:37:17PM -0500, Steven Rostedt wrote:
> > > On Fri, 8 Nov 2013 00:21:51 +0100
> > > Frederic Weisbecker <fweisbec@xxxxxxxxx> wrote:
> > > > 
> > > > Offloading to a workqueue would be perhaps better, and writing to the serial
> > > > console could then be done with interrupts enabled, preemptible context, etc...
> > > 
> > > Oh God no ;-)  Adding workqueue logic into printk just spells a
> > > nightmare of much more complexity for a critical kernel infrastructure.
> > 
> > But yeah that's scary, that means workqueues itself can't printk that safely.
> > So, you're right after all.
>   Yeah, we've been there (that was actually my initial proposal). But
> Andrew and Steven (rightfully) objected and suggested irq_work should be
> used instead.

I still hate the patchset and so does everyone else, including you ;)
There must be something smarter we can do.  Let's start by restating
the problem:

CPU A is in printk, emitting log_buf characters to a slow device. 
Meanwhile other CPUs come into printk(), see that the system is busy,
dump their load into log_buf then scram, leaving CPU A to do even more
work.

Correct so far?

If so, what is the role of local_irq_disabled() in this?  Did CPU A
call printk() with local interrupts disabled, or is printk (or the
console driver) causing the irqs-off condition?  Where and why is this
IRQ disablement happening?


Could we fix this problem by not permitting CPUs B, C and D to DoS CPU
A?  When CPU B comes into printk() and sees that printk is busy, make
CPU A hand over to CPU B and let CPU A get out of there?


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/