Re: [RFC PATCH v1 10/25] printk: redirect emit/store to new ringbuffer

From: Petr Mladek <pmladek@suse.com>
To: John Ogness <john.ogness@linutronix.de>
Cc: linux-kernel@vger.kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Daniel Wang <wonderfly@google.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Alan Cox <gnomes@lxorguk.ukuu.org.uk>,
	Jiri Slaby <jslaby@suse.com>, Peter Feiner <pfeiner@google.com>,
	linux-serial@vger.kernel.org,
	Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Subject: Re: [RFC PATCH v1 10/25] printk: redirect emit/store to new ringbuffer
Date: Fri, 22 Feb 2019 16:25:41 +0100	[thread overview]
Message-ID: <20190222152541.33xp2btltwcecxz7@pathway.suse.cz> (raw)
In-Reply-To: <87va1byia5.fsf@linutronix.de>

On Fri 2019-02-22 16:06:26, John Ogness wrote:
> On 2019-02-22, Petr Mladek <pmladek@suse.com> wrote:
> >>>> +	rbuf = prb_reserve(&h, &sprint_rb, PRINTK_SPRINT_MAX);
> >>>
> >>> The second ring buffer for temporary buffers is really interesting
> >>> idea.
> >>>
> >>> Well, it brings some questions. For example, how many users might
> >>> need a reservation in parallel. Or if the nested use might cause
> >>> some problems when we decide to use printk-specific ring buffer
> >>> implementation. I still have to think about it.
> >> 
> >> Keep in mind that it is only used by the writers, which have the
> >> prb_cpulock. Typically there would only be 2 max users: a non-NMI
> >> writer that was interrupted during the reserve/commit window and the
> >> interrupting NMI that does printk. The only exception would be if the
> >> printk-code code itself triggers a BUG_ON or WARN_ON within the
> >> reserve/commit window. Then you will have an additional user per
> >> recursion level.
> >
> > I am not sure it is worth to call the ring buffer machinery just
> > to handle 2-3 buffers.
> 
> It may be slightly overkill, but:
> 
> 1. We have the prb_cpulock at this point anyway, so it will be
>    fast. (Both ring buffers share the same prb_cpulock.)

I am still not persuaded that we really need the lock. The
implementation looks almost ready for a fully lockless
writers. But I might be wrong.

The lock might be fine when it makes the code easier and does
not bring any deadlocks.

> 2. Getting a safe buffer is just 1 line of code: prb_reserve()

The problem is how complicated code is hidden behind
this 1 line of code.

> 3. Why should we waste _any_ lines of code implementing the handling of
>    these special 3-4 buffers?

It might be worth if it makes the code more strighforward
and less prone to bugs.

> > Well, it might be just my mental block. We need to be really careful
> > to avoid infinite recursion when storing messages into the log
> > buffer.
> 
> The recursion works well. I inserted a triggerable BUG_ON() in
> vprintk_emit() _within_ the reserve/commit window and I see a clean
> backtrace on the emergency console.

Have you tested all possible error situations that might happen?
Testing helps a lot. But the real life often brings surprises.

Best Regards,
Petr