Re: [PATCH 00/10] printk: Avoid deadlock in NMI + vprintk_emit() cleanup

From: "long.wanglong" <long.wanglong@huawei.com>
To: Jan Kara <jack@suse.cz>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Petr Mladek <pmladek@suse.cz>,
	Frederic Weisbecker <fweisbec@gmail.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Dave Anderson <anderson@redhat.com>,
	"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>,
	Kay Sievers <kay@vrfy.org>, "Jiri Kosina" <jkosina@suse.cz>,
	Michal Hocko <mhocko@suse.cz>, <linux-kernel@vger.kernel.org>,
	<peifeiyue@huawei.com>, <dzickus@redhat.com>,
	<morgan.wang@huawei.com>, <sasha.levin@oracle.com>,
	<xuhanbing@huawei.com>
Subject: Re: [PATCH 00/10] printk: Avoid deadlock in NMI + vprintk_emit() cleanup
Date: Tue, 2 Jun 2015 17:46:28 +0800	[thread overview]
Message-ID: <556D7B74.2070007@huawei.com> (raw)
In-Reply-To: <20150601130616.GE20288@quack.suse.cz>

On 2015/6/1 21:06, Jan Kara wrote:
> On Fri 29-05-15 13:50:45, Andrew Morton wrote:
>> On Mon, 25 May 2015 14:46:23 +0200 Petr Mladek <pmladek@suse.cz> wrote:
>>
>>> The main source of deadlocks caused by printk() in NMI context has been
>>> solved by the commit a9edc88093287 ("x86/nmi: Perform a safe NMI stack
>>> trace on all CPUs").
>>>
>>> But there are still few warnings printed in the NMI code that could
>>> case a deadlock. For example, see the freeze discussed at
>>> https://lkml.org/lkml/2015/5/20/481
>>
>> I'm not (yet) convinced that we want the entire patchset btw.  Do we
>> really want to try to semi-support printk from NMI?  With a rather
>> nasty set of hacks?
>>
>> Why not just delete the offending printks?
>   And what about WARN_ONs and BUG_ONs? Delete as well? Or just don't print
> anything when we are in NMI? I agree that NMI is so problematic context
> that restricting printk there makes some sence. OTOH propagating
> information from NMI to user is useful as well so I'm somewhat undecided.
> 
> 								Honza
> 

I think that delete all printks in NMI context is not a good solution.
because some information is very useful to user.

The commit a9edc88093287 ("x86/nmi: Perform a safe NMI stack trace on all CPUs")
solved the deadlock problem only in arch_trigger_all_cpu_backtrace_handler() by
replacing the printk to a special print function (nmi_vprintk). But all the
other NMI handlers still use the original printk.

How about replacing printk function earlier? we can replace printk function
before we calling default_do_nmi(arch/x86/kernel/nmi.c) and replace back
after calling.

Is it a feasible solution? or does it introduce other problems?

Best Regards
Wang Long