Re: [PATCH 1/1] system call notification with self_ptrace

[Pierre Morel <pmorel@linux.vnet.ibm.com>]

Hello Oleg,

You are right, the functionality can be implemented with the system call.
But it means we have the overhead of a system call just to clear two bits,
the TIF_SYSCALL_TRACE and the PTS_SELF.

On the other hand we have an overhead of one single "if" inside
the handle_signal() function.

We can do the same with fork and ptrace, yes, but with a very big 
overhead on each system call and this is why this patch is so usefull: 
because with this patch you sit inside the thread when analysing it and 
have a direct access to all data without the need of IPC, ptrace or any 
task switch.

I will provide a test program and plan to release a tracing tool based 
on it.
I think I can reduce the task struct modification by using just a bit 
like you suggest if nobody seen any problem with this.

best regards,

Pierre

Oleg Nesterov wrote:
> On 09/10, Pierre Morel wrote:
>   
>> Oleg Nesterov wrote:
>>     
>>> I still think this patch shouldn't change handle_signal().
>>>
>>> Once again. The signal handler for SIGSYS can first do
>>> sys_ptrace(PTRACE_SELF_OFF) (which is filtered out), and then use any
>>> other syscall, so this change is not needed, afaics.
>>>
>>>       
>> Yes it can but what if the application forget to do it?
>> It is a security so that the application do not bounce for ever.
>>     
>
> The (buggy) task can be killed, this has nothing to do with security.
>
> From the security pov, this case doesn't differ from, say,
>
> 	void sigh(int sig)
> 	{
> 		kill(getpid(), sig);
> 	}
>
> 	void main(void)
> 	{
> 		signal(SIGSYS, sigh);
> 		kill(getpid(), SIGSYS);
> 	}
>
> Or I missed something?
>
>   
>>> So, PTRACE_SELF_XXX disables the "normal" ptrace. Not sure this is good.
>>>
>>>       
>> I think that having two tracing system one over the other may be
>> quite difficult to handle.
>>     
>
> Yes I see.
>
> But... well, I think we need Roland's opinion. I must admit, I am a bit
> sceptical about this patch ;) I mean, I don't really understand why it
> is useful. We can do the same with fork() + ptrace(). Yes, in that
> case we need an "extra" context switch for any traced syscall. But,
> do you have any "real life" example to demonstrate that the user-space
> solution sucks? We can even use CLONE_MM to speedup the context switch.
>
> Pierre, don't get me wrong. I never used debuggers for myself, I will
> be happy to know I am wrong. I just don't understand.
>
>
> As for ->instrumentation. If you are going to remove PTS_INSTRUMENTED,
> we need only one bit. We could use PF_PTS_SELF, but ->flags is already
> "contended". Perhaps you can do something like
>
> 	--- include/linux/sched.h
> 	+++ include/linux/sched.h
> 	@@ -1088,6 +1088,7 @@ struct task_struct {
> 		/* ??? */
> 		unsigned int personality;
> 		unsigned did_exec:1;
> 	+	unsigned pts_self:1;
> 		pid_t pid;
> 		pid_t tgid;
> 	 
>
> Both did_exec and pts_self can only be changed by current, so it is
> safe to share the same word. This way we don't enlarge task_struct.
>
> Oleg.
>
>   


-- 
=============
Pierre Morel
RTOS and Embedded Linux