printk in kernel.function("__switch_to") probe crashes machine

[Jim Keniston]

On Wed, 2007-04-25 at 16:17 -0700, Dave Nomura wrote:
> 3. I have written a kernel function probe in __switch_to and find that I 
> can define it as:
> probe kernel.function("__switch_to")
> {
>   print("calling ITrace_switch\n");
>   // ITrace_switch()
> }
> If I run this script I get:
> Running sudo /usr/bin/staprun -u dcn -d 3614 -c 
> "/home/dcn/itrace/ITrace_test/hanoi/hanoi64 -2" /tmp/stapAiagsM/pitrace.ko
> start itrace
> calling ITrace_switch
> ....
> calling ITrace_switch
> end itrace
> and this works just as expected.
> However, if I uncomment ITrace_switch() call in my __switch_to probe, 
> and define it as:
> function ITrace_switch()
> %{
>   printk("ITrace_switch called\n");
> %}
> this crashes the machine.
> 
> I'm a SystemTap novice, and kernel novice so I'm not sure if there is 
> some obvious reason why this "printk" should cause the system to crash.  
> There are no messages on the console.
> 
> If there is a simple explanation I guess it might help me avoid doing 
> this sort of thing in the future.  For now, I'll just avoid using printk 
> in this scenario.

We advise against using printk in SystemTap handlers.  We know of some
functions that, if kprobed with handlers that call printk, can deadlock.
See bugzilla #3541.  I wouldn't think that __switch_to() would be in
that group (since printk() doesn't block and so shouldn't call
__switch_to()), but I could be wrong.

Jim

> 
> Here is the machine info:
> Linux elm3b158 2.6.16.43-0.5-ppc64 #1 SMP Mon Mar 19 08:28:24 UTC 2007 
> ppc64 ppc64 ppc64 GNU/Linux
> SUSE Linux Enterprise Server 10 SP1 RC1 (ppc)
>