Re: IRQ off latency of printk is very high

[Pavel Machek <pavel@ucw.cz>]

Hi!

> > Hmm, I see this at the beginning of the post-BK era (2.6.12-rc2):
> > 
> > 	        spin_lock_irqsave(&logbuf_lock, flags);
> > 		...
> > 		spin_unlock(&logbuf_lock);
> > 		call_console_drivers(_con_start, _log_end);
> > 		local_irq_restore(flags);
> > 
> 
> Well, I need to do some more research.  This must be in
> release_console_sem().  I was looking at vprintk, through
> the ages.  At 2.6.16, it looked like this:
> 
>         spin_lock_irqsave(&logbuf_lock, flags);
>         ...
>               spin_unlock_irqrestore(&logbuf_lock, flags);
>               console_may_schedule = 0;
>               release_console_sem();
> 
> but the irq restore has been moving around to different places
> in that function over the last few years.  I suspect that in the
> common case the irqsave in vprintk is the one that disables
> ints.
> 
> It appears that formerly interrupts were enabled in vprintk but
> re-disabled immediately upon entering release_console_sem().
> As it is now, they're held during formatting, buffering,
> and output, which seems excessive.
> 
> It seems draconian to drain the entire buffer with ints disabled.
> Is it possible to break this up and send out smaller chunks
> at a time?  Maybe by putting a chunk loop in release_console_sem()?

Well, I believe someone got

DDetetccctted ed 113223 HHzz CPUCPU

in his dmesg, and now we have this 'draconian' locking. How can we
prevent mangled messages without it?
							Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html