same kernel working with 1 userspace filesystem and not with another one (Android IceCreamSandwich and JellyBean)

[Josh Stone]

On 10/05/2012 04:03 PM, Turgis, Frederic wrote:
> Hi,
> 
> We are starting an investigation but maybe someone would have an idea.
> - kernels (K3.0 and K3.4) have an issue with Android JellyBean OS
>    * probing tracepoints works well (we monitor IRQs)
>    * probing functions does not work, it seems that nothing happens in
>      that case (we just do a print, we only get tracepoints prints)
>    * at test start, we get now "WARNING: build-id address 0 < base 0"
> 
> - same kernel with Android ICS OS works OK. Same kernel exactly so
> same build of script
> 
> - staprun/stapio are statically linked so would have in theory no
> link with userspace. So the difference between the 2 OSes would be
> insmod but not sure that makes sense. I guess we should try ICS
> insmod in JB.

It's not insmod - we use init_module() directly.

> Any clue ? How can userspace influence systemtap if tool is
> statically built ?

Perhaps through the many procfs, sysfs, and debugfs controls...

e.g. if you weren't getting stap to work at all, I would suspect debugfs
mounted permissions, as we've see that before.

Since you're getting tracepoints but not kprobes, that seems to indicate
it's really a problem just in kprobes.  I know for instance that on x86
there are issues with enabled ftrace (-mfentry hooks) and kprobes on the
same addresses.  So if any of the JB userspace enables ftrace function
tracing, that may cause your conflict.


Hope that helps,
Josh