Re: [RFC][PATCH 1/4] tracing: Add ref count to ftrace_event_call

From: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
To: Steven Rostedt <rostedt@goodmis.org>
Cc: linux-kernel@vger.kernel.org, Oleg Nesterov <oleg@redhat.com>,
	"zhangwei(Jovi)" <jovi.zhangwei@huawei.com>,
	Jiri Olsa <jolsa@redhat.com>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>,
	Arnaldo Carvalho de Melo <acme@ghostprotocols.net>,
	Srikar Dronamraju <srikar@linux.vnet.ibm.com>,
	Frederic Weisbecker <fweisbec@gmail.com>,
	Ingo Molnar <mingo@redhat.com>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: [RFC][PATCH 1/4] tracing: Add ref count to ftrace_event_call
Date: Thu, 04 Jul 2013 13:22:29 +0900	[thread overview]
Message-ID: <51D4F885.3040105@hitachi.com> (raw)
In-Reply-To: <20130704034038.501144114@goodmis.org>

(2013/07/04 12:33), Steven Rostedt wrote:
> diff --git a/kernel/trace/trace_events.c b/kernel/trace/trace_events.c
> index 7d85429..90cf243 100644
> --- a/kernel/trace/trace_events.c
> +++ b/kernel/trace/trace_events.c
> @@ -391,6 +391,28 @@ static void __get_system_dir(struct ftrace_subsystem_dir *dir)
>  	__get_system(dir->subsystem);
>  }
>  
> +static int ftrace_event_call_get(struct ftrace_event_call *call)
> +{
> +	int ret = 0;
> +
> +	mutex_lock(&event_mutex);
> +	if ((call->flags & TRACE_EVENT_FL_REF_MASK) == TRACE_EVENT_FL_REF_MAX - 1)
> +		ret = -EBUSY;
> +	else
> +		call->flags++;
> +	mutex_unlock(&event_mutex);
> +
> +	return ret;
> +}
> +
> +static void ftrace_event_call_put(struct ftrace_event_call *call)
> +{
> +	mutex_lock(&event_mutex);
> +	if (!WARN_ON_ONCE(!(call->flags & TRACE_EVENT_FL_REF_MASK)))
> +		call->flags--;
> +	mutex_unlock(&event_mutex);
> +}

Hmm, I might misunderstand, but it seems that there is a small unsafe
time slot.

> @@ -424,7 +446,15 @@ static int tracing_open_generic_file(struct inode *inode, struct file *filp)
>  
>  	ret = tracing_open_generic(inode, filp);
>  	if (ret < 0)
> -		trace_array_put(tr);
> +		goto fail;
> +
> +	ret = ftrace_event_call_get(file->event_call);
> +	if (ret < 0)
> +		goto fail;
> +
> +	return 0;
> + fail:
> +	trace_array_put(tr);
>  	return ret;
>  }
>  
> @@ -433,12 +463,40 @@ static int tracing_release_generic_file(struct inode *inode, struct file *filp)
>  	struct ftrace_event_file *file = inode->i_private;
>  	struct trace_array *tr = file->tr;
>  
> +	ftrace_event_call_put(file->event_call);
>  	trace_array_put(tr);
>  
>  	return 0;
>  }

Here, at first we get an event_file from inode->i_private, and then locks
event_mutex and get/put refcount. This should be safe if we get (find) the object
from some list of event_file under the mutex, but we just use inode->i_private.

This can cause a race as below scenario,

CPU0                              CPU1
open(kprobe_events)
  trace_remove_event_call()    open(enable)
    lock event_mutex             refer event_file
      event_remove()             wait for unlock event_mutex
        ...
        free event_file
    unlock event_mutex
                                 lock event_mutex
                                    add refcount of event_file->call (*)

So, at (*) point, the event_file is already freed. Thus there still be
an unsafe time slot. Or, did I miss something (possibly..)?

Thank you,

-- 
Masami HIRAMATSU
IT Management Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com