Re: [PATCH] tracing/function-graph-tracer: tracing only syscalls mode

[Ingo Molnar <mingo@elte.hu>]

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

> Impact: make more easy the syscalls tracing
> 
> This patch extends to syscalls the features which let one to trace
> only one ore more function (and those they call).
> This way we can get rid of the workqueues, kernel threads, softirqs from
> ksoftirqd and only have the syscall path on the trace.
> 
> Note that hardirq that interrupt the syscalls are still traced. But we could
> add an option to disable the interrupt tracing as well in the future.
> 
> To use this new feature:
> 
> echo function_graph > /debugfs/tracing/current_tracer
> echo syscalls > /debugfs/tracing/set_graph_function

nice. I think a key item is missing though:

> @@ -1397,6 +1398,9 @@ asmregparm long syscall_trace_enter(struct pt_regs *regs)
>  {
>  	long ret = 0;
>  
> +	/* Tell the function graph tracer we are entering a system call */
> +	ftrace_graph_syscall_enter();

We could actually trace the _syscalls_ themselves, a'la strace:

    munmap(0xb7ff3000, 4096)                = 0

So instead of the current ftrace output:

  0)   cc1-30212    |               |  sys32_mmap2() {
  0)   cc1-30212    |               |    down_write() {
  0)   cc1-30212    |   0.272 us    |      _spin_lock_irq();
  0)   cc1-30212    |   0.969 us    |    }
  0)   cc1-30212    |               |    do_mmap_pgoff() {

We could get something like:

  0)   cc1-30212    |               |  sys32_mmap2(0xb7ff3000, 4096) {
  0)   cc1-30212    |               |    down_write() {
  0)   cc1-30212    |   0.272 us    |      _spin_lock_irq();
  0)   cc1-30212    |   0.969 us    |    }
  0)   cc1-30212    |               |    do_mmap_pgoff() {
  [...]
  0)   cc1-30212    | + 22.537 us   |  } = 0


Or maybe as separate trace entries:

  0)   cc1-30212    |               |> sys32_mmap2 [0xb7ff3000, 4096]
  0)   cc1-30212    |               |  sys32_mmap2 {
  0)   cc1-30212    |               |    down_write() {
  0)   cc1-30212    |   0.272 us    |      _spin_lock_irq();
  0)   cc1-30212    |   0.969 us    |    }
  0)   cc1-30212    |               |    do_mmap_pgoff() {
  [...]
  0)   cc1-30212    | + 22.537 us   |  }
  0)   cc1-30212    |               |< sys32_mmap2 => 0

For that we'd have to pass in something like the syscall function address 
(sys_call_table[regs->ax]), and the up to 6 parameters 
[regs->bx,cx,dx,si,di,bp].

(and initially we could just print all of them i guess, instead of a 
variable-width thing)

We wouldnt do smart decoding of syscall arguments normally - just print 
the raw arguments with no decoding. (like  strace -e raw=all)

How does this sound?

	Ingo