On (03/04/18 10:43), Steven Rostedt wrote:
> On Sun, 04 Mar 2018 23:08:23 +0800
> "Qixuan.Wu" <qixuan...@linux.alibaba.com> wrote:
>
> > Suppose there is one scenario that the system has 100 CPU(0~99). While CPU
> > 0 is
> > calling slow console, CPU 1~99 are calling printk at the same time. And
> > suppose
> > CPU 1 will be waiter, as per the patch, 2~99 will return directly. After
> > CPU 0 finish
> > it's log to console, it will return when it finds CPU 1 are waiting. Then
> > CPU 1 need
> > flush all logs of CPU(1~99) to the console, which may cause softlockup or
> > rcu
> > stall. Above scenario is very unusual and it's very unlikely to happen.
>
> Yes, people keep bringing up this scenario.
Yeah.
> It would require a single burst of printks to all CPUs.
That's one possibility. The other one is - console_sem locked by a
preemptible context which gets scheduled out.
> And then no more printks after that. The last one will end up printing
> the entire buffer out the slow console. The thing is, this is a bounded
> time, and no printk will print more than one full buffer worth.
It can print more than "one full buffer worth". In theory and on practice.
> If this is a worry, then set the timeouts for the lockup detection to
> be longer than the time it takes to print one full buffer with the
> slowest console.
I see your point.
But I still think that it makes sense to change that "print it all" approach.
With more clear/explicit watchdog-dependent limits - we do direct printk for
1/2 (or 2/3) of a current watchdog threshold value and offload if there is
more stuff in the logbuf. Implicit "logbuf size * console throughput" is
harder to understand. Disabling watchdog because of printk is a bit too much
of a compromise, probably.
IOW, is logbuf worth of messages so critically important after all that we
are ready to jeopardize the system stability?
-ss
