Re: [PATCH] printk: drop in_nmi check from printk_safe_flush_on_panic()

[Petr Mladek]

On Wed 2018-05-30 19:00:37, Sergey Senozhatsky wrote:
> On (05/30/18 18:55), Sergey Senozhatsky wrote:
> > The thing is, we, in fact, already invoke panic() in printk_safe mode.
> > Sometimes.
> > 
> > Namely,
> > 
> >   nmi_panic() -> panic()
> > 
> > is invoked while we are in printk_nmi(), so all printk()-s go
> > to the per-CPU buffers. So, at least to some extent, panic()
> > in printk_safe context is not something never seen before. Just
> > saying.
> 
> Well, we have a PRINTK_NMI_DEFERRED_CONTEXT_MASK mode for
> printk_nmi(). May be we can [if need be] come up with the same trick
> for printk_safe_panic() mode. If logbuf spin_lock is unlocked, then
> we use the main logbuf, if it is locked, we redirect printk to per-CPU
> buffers and then flush it via printk_safe_flush_on_panic(), which will
> re-init (unlock) the logbuf.

All these checks are racy. Now, I believe that it might really prevent
a deadlock in some situations but it might also cause loosing messages
in other situations (never flushed buffer). I am sorry but I am still
unable to decide if it is worth the risk.

I would want to keep it as is until anyone comes with a more detailed
analyze or until we get some bug reports.

Best Regards,
Petr