On (04/26/19 16:14), Petr Mladek wrote:
>
> Then I wonder why, for example, native_stop_other_cpus() waits
> 10ms at maximum after sending the NMIs. What is the state
> of the CPUs that miss this deadline?
Well, I saw a case when CPU was forcibly powered off (embedded),
for instance.
> > - But, more importantly, if that CPUB is in atomic context, then panic
> > CPUA will spin, waiting for that CPUB to handoff printing, before
> > panic CPU will even try to stop all CPUs.
> >
> > pr_emerg("Kernel panic - not syncing: %s\n", buf)
> >
> > is the point of 'synchronization' - panic CPU will wait for
> > current console owner.
>
> "Synchronization point" is too strong formulation.
But it is sort of synchronized. That's why console_owner patch
set solved the panic-printk deadlock which Google folks reported
a while ago.
> The console waiter logic is effective but it does not always
> work. The current console owner must be calling the console
> drivers.
>
> > Hmm, we might have a bit of a problem here, maybe.
>
> Hmm, the printk() might wait forever when NMI stopped
> the current console owner in the console driver code
> or with the logbuf_lock taken.
I guess this is why we re-init logbuf lock from panic,
however, we don't do anything with the console_owner.
> The console waiter logic might get solved by clearing
> the console_owner in console_flush_on_panic(). It can't
> be much worse, we already ignore console_lock() there, ...
Right.
[..]
> Anyway, do we really need to have length discussion about
> whether the locks are needed? They will not break anything.
I'm not objecting v5 nor your request to add that locking there.
I'm talking about different things.
-ss
