[PATCH 0 of 3] Add support for dtrace compatible sdt probes on GNU/Linux

[Mark Wielaard]

On Tue, 2012-05-22 at 21:25 +1000, David Holmes wrote:
> On 22/05/2012 9:11 PM, Mark Wielaard wrote:
> >But really the SDT markers themselves
> > are independent from any program that consumes them (on Solaris you
> > would use dtrace, on GNU/Linux you can also use GDB or just list them
> > with readelf -n).
> 
> I need to understand this aspect a bit better.

SDT markers are defined by simple macros. e.g:
DTRACE_PROBE5(provider,probe,parm1,parm2,parm3,parm4,parm5)
Which is a marker for a named provider with a probe name and 5
arguments. Of course in the hotspot source code there are macros around
these macros, and some convenience functions. All the patches really do
is make sure that if a sys/sdt.h file is found on a GNU/Linux system it
is included and the macros are used. These macros expand to a little bit
of code that just puts those names and where the arguments can be found
in a little ELF notes section.

So with the patch applied and compiled on a system that has sys/sdt.h
you will get a libjvm.so that contains:

$ eu-readelf -n libjvm.so

Note section [29] '.note.stapsdt' of 53916 bytes at offset 0x99fac8:
  Owner          Data size  Type
  stapsdt               77  Version: 3
    PC: 0x35612f, Base: 0x7f2378, Semaphore: 0
    Provider: hotspot, Name: class__loaded, Args: '8@%r13 -4@%r14d 8@%rax 1@%r12b'
  stapsdt               76  Version: 3
    PC: 0x3561e5, Base: 0x7f2378, Semaphore: 0
    Provider: hotspot, Name: class__unloaded, Args: '8@%rbx -4@%r12d 8@%rax 1@$0'

[etc.]

That ELF notes section can then be used by programs like stap or gdb to
put breakpoints/probes on those locations and retrieve the arguments
found.

A more technical explanation with some more details I left out can be
found at:
http://sourceware.org/systemtap/wiki/UserSpaceProbeImplementation
Another high level overview from one of the GDB maintainers is here:
http://tromey.com/blog/?p=687

It works similar for SDT probes on Solaris and dtrace, though they use
different conventions for storing the addresses/names/arguments. The
main difference is that on solaris a different ELF section is used and
the arguments are assumed to follow function calling conventions, while
on gnu/linux they can just be "anywhere".

Hope that helps,

Mark