Re: [RFC][PATCH 1/2] tracing/ftrace: syscall tracing infrastructure

[Ingo Molnar <mingo@elte.hu>]

* Frederic Weisbecker <fweisbec@gmail.com> wrote:

> +static const struct syscall_trace_entry syscall_trace_entries[] = {
> +	/* For open, the first argument is a string, hence the given mask */
> +	[SYSCALL_TRACE_OPEN]	= SYS_TRACE_ENTRY(open, 3, 0x1),
> +	[SYSCALL_TRACE_CLOSE]	= SYS_TRACE_ENTRY(close, 1, 0),
> +	[SYSCALL_TRACE_READ]	= SYS_TRACE_ENTRY(read, 3, 0),
> +	[SYSCALL_TRACE_WRITE]	= SYS_TRACE_ENTRY(read, 3, 0),
> +};

s/read/write in the last entry i guess.

But i think the whole concept of duplicating the syscall table 
is the wrong way around.

Note that we dont have to build this information at all - in 
2.6.29-rc1 all syscalls got wrapper macros:

 SYSCALL_DEFINE1(nice, int, increment)
 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
                 unsigned long __user *, user_mask_ptr)

We also have the syscall table itself.

So what we can do: by changing the SYSCALL_DEFINEX() macros we 
can emit the following information into a table:

  (syscall_fn_address, syscall_name_string, #of arguments, array 
   of argument names and type sizeof()s)

then during bootup we can match up the sys_call_table[] to the 
secondary table we built, so that we can order the secondary 
table based on syscall NR. 99% of the syscalls will match up 
just fine.

	Ingo