Queries on SystemTap - with focus on User Space debugging

[Manjusha Ajay]

Hi Frank,

I tried adding the probe kernel.function("send_sig") in my code to
trace my test program, but it so happened that this particular probe
was not hit during the execution of the process. Not sure why that
happened.

probe kernel.function("send_sig") {
     if ($sig == 11 && p && task_pid(p) == target()) {
            print_ubacktrace()
     }
}

Also got a warning when I ran the above script

> stap -c "./test_segfault' test.stp
WARNING:  read-only local variable 'p'  : identifier 'p' at test.stp:19:20
   source:             if ($sig == 11 && p && task_pid(p) == target()) {



% stap -L 'kernel.function("send_sig")'   gave me the following result
kernel.function("send_sig@kernel/signal.c:1169")

Iam using RHEL 5.4,  kernel release 2.6.18-164.el5


I had also tried adding probe kernel.function("send_signal") which was
hit, but then print_ubacktrace gave me the kernel stack trace whereas
i was expecting the stack trace of the user space application.
I got the above mentioned warning even in this case.


Manjusha.


-----Original Message-----
From: Frank Ch. Eigler
Sent:  24/05/2011 18:34:38
Subject:  Re: Queries on SystemTap - with focus on User Space debugging


manjusha.ajay wrote:

> [...]
> probe process("test_segfault").syscall
>  {
>  pid = pid()
>  if (pid == target())
>  {
>      if ($syscall == 11)

Syscall #11 may be execve or munmap; neither appears to have much to
do with SIGSEGV (which is signal #11).  We do not have a convenient
probe point in the tapset to commemorate signal delivery, but the
following ought to do:

% stap -L 'kernel.function("send_sig")'
kernel.function("send_sig@kernel/signal.c:1300") $sig:int $p:struct
task_struct* $priv:int

i.e.,

probe kernel.function("send_sig") {
     if ($sig == 11 && p && task_pid(p) == target()) {
        ....
     }
}


>      {
>      // Trying to stop the process to collect the
>      // required data when it receives SIGSEGV
>      stop_cmd = sprintf("kill -s STOP %d", pid)
>      system(stop_cmd)
>
>      print_ubacktrace()
>
>      // Resume the process
>      cont_cmd = sprintf ("kill -s CONT %d", pid)
>      system(cont_cmd)
>     }
>     }
>  }


This is more than necessary.  You neither need to kill -STOP or -CONT
around a print_ubacktrace().  Just print_ubacktrace().  The other
problem is that the system() tapset function does not run the given
command string immediately.  It enqueues it for execution *soon*.
This should explain why in your tests some of the kill jobs showed an
error with a missing process: it was gone by the time the enqueued
jobs got started.

- FChE