On Thu, 9 Nov 2017 09:56:35 +0900
Sergey Senozhatsky <sergey.senozhatsky.w...@gmail.com> wrote:

> Hello Steven,
> 
> On (11/08/17 09:29), Steven Rostedt wrote:
> > On Wed, 8 Nov 2017 14:19:55 +0900
> > Sergey Senozhatsky <sergey.senozhatsky.w...@gmail.com> wrote:
> >   
> > > the change goes further. I did express some of my concerns during the KS,
> > > I'll just bring them to the list.
> > > 
> > > 
> > > we now always shift printing from a save - scheduleable - context to
> > > a potentially unsafe one - atomic. by example:  
> > 
> > And vice versa. We are now likely to go from a unscheduleable context
> > to a schedule one, where before, that didn't exist.  
> 
> the existence of "and vice versa" is kinda alarming, isn't it? it's sort
> of "yes, we can break some things, but we also can improve some things."

Not really. Because the heuristic is that what calls printk will do the
printk.

> 
> > And my approach, makes it more likely that the task doing the printk
> > prints its own message, and less likely to print someone else's.
> >   
> > > 
> > > CPU0                      CPU1~CPU10      CPU11
> > > 
> > > console_lock()
> > > 
> > >                   printk();
> > > 
> > > console_unlock()                  IRQ
> > >  set console_owner                        printk()
> > >                                    sees console_owner
> > >                                    set console_waiter
> > >  sees console_waiter
> > >  break
> > >                                    console_unlock()
> > >                                    ^^^^ lockup [?]  
> > 
> > How?  
> 
> oh, yes, the missing part - assume CPU1~CPU10 did 5000 printk() calls,
> while console_sem was locked on CPU0. then we console_unlock() from CPU0
> and shortly after IRQ->printk() from CPU11 forcibly takes over, so now
> we are in console_unlock() from atomic, printing some 5000 messages.

I'd say remove those 5000 printks ;-)

-- Steve

Reply via email to