Re: [RFC][PATCH v2 1/2] printk: Make printk() completely async
From: Sergey Senozhatsky
Date: Sat Mar 12 2016 - 00:03:36 EST
On (03/11/16 12:22), Tejun Heo wrote:
> On Tue, Mar 08, 2016 at 07:21:52PM +0900, Sergey Senozhatsky wrote:
> > I'd personally prefer to go with the "less dependency" option -- a dedicated
> > kthread, I think. mostly for the sake of simplicity. I agree with the point
> > that console_unlock() has unpredictable execution time, and in general case
> > we would have a busy kworker (or sleeping in console_lock() or doing
> > cond_resched()) and an idle extra WQ_RESCUER kthread, with activation rules
> > that don't depend on printk. printk with dedicated printk-kthread seems
> > easier to control. how does it sound?
> I don't think it makes sense to avoid workqueue for execution latency.
> The only case which can matter is the rescuer case and as I wrote
> before the system is already in an extremely high latency mode by the
> time rescuer is needed, so it's unlikely to make noticeable
> However, I agree that using kthread is a good idea here just to reduce
> the amount of dependency as prink working even during complex failures
> is important. workqueue itself is fairly complex and it also requires
> timer and task creation to work correctly for proper operation.
> That's a lot of extra dependency.
I agree that, in some cases (if not in most of them) the "value" of printk()
output is inversely proportional to the system health -- the worst the state,
the more attention people pay to printk() output; so a simpler solution here
gives more confidence.