dtrace discuss - Aug 2010 - USDT probes

Hi,
I''m posting a question hoping someone will know the answer off hand
thereby reducing my search time. :-)

With USDT probes, the tracepoint is only installed by libdtrace itself, never by
the drti ioctl. So whenever I run a program with an USDT probe, no tracepoint is
installed. Only after I run the dtrace command the tracepoint is actually
installed on the victim process.

My question is about how Solaris discovers the correct instruction pointer (PC)
on the target victim. The DOF only has information relative to the offset inside
the function where the probe is defined and the address of the function in the
_relocatable_ object.

So, whenever I''m building a program with an usdt probe I get:

libdtrace DEBUG: defined probe  database:::query-start main() +0x13 (main)

When I run ''dtrace -n ''database*:::'''' the
fasttrap module inserts the probe at the 0x13 offset, not the actually vaddr of
main() + offset. I added a debug printf in fasttrap_tracepoint_enable():

installing tracepoint on pid 1288 pc 0x13

Of course this doesn''t work because main() is actually at 0x4011d0.

To overcome this I thought that dtrace(1) would try to grab the process and
discover the vaddr of main. But from I can see dtrace(1) never grabs the
process.

How does DTrace work for this case on Solaris? Where''s the magic for
installing the trap instruction at 0x4011d0 + 0x13 ?

Thanks,
--
Rui Paulo

Hey Rui,

There are two relocations that take place. The provider.o file (generated by
dtrace -G) contains relocations that are resolved with the final link pass
occurs. We record the function offset (libdtrace) and then the linker adds in
the base address of the function. The next relocation takes place when we
ioctl(2) the DOF to the kernel. There, we provide the base address of the
object.

  tracepoint address = base address of object + function offset + offset into
function
                              drti.o                ld(1)         dtrace -G /
libdtrace

Adam

On Aug 10, 2010, at 8:46 AM, Rui Paulo wrote:
> Hi,
> I''m posting a question hoping someone will know the answer off
hand thereby reducing my search time. :-)
> 
> With USDT probes, the tracepoint is only installed by libdtrace itself,
never by the drti ioctl. So whenever I run a program with an USDT probe, no
tracepoint is installed. Only after I run the dtrace command the tracepoint is
actually installed on the victim process.
> 
> My question is about how Solaris discovers the correct instruction pointer
(PC) on the target victim. The DOF only has information relative to the offset
inside the function where the probe is defined and the address of the function
in the _relocatable_ object.
> 
> So, whenever I''m building a program with an usdt probe I get:
> 
> libdtrace DEBUG: defined probe  database:::query-start main() +0x13 (main)
> 
> When I run ''dtrace -n ''database*:::''''
the fasttrap module inserts the probe at the 0x13 offset, not the actually vaddr
of main() + offset. I added a debug printf in fasttrap_tracepoint_enable():
> 
> installing tracepoint on pid 1288 pc 0x13
> 
> Of course this doesn''t work because main() is actually at
0x4011d0.
> 
> To overcome this I thought that dtrace(1) would try to grab the process and
discover the vaddr of main. But from I can see dtrace(1) never grabs the
process.
> 
> How does DTrace work for this case on Solaris? Where''s the magic
for installing the trap instruction at 0x4011d0 + 0x13 ?
> 
> Thanks,
> --
> Rui Paulo
> 
> 
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org

--
Adam Leventhal, Fishworks                        http://blogs.sun.com/ahl

On 10 Aug 2010, at 19:29, Adam Leventhal wrote:
> Hey Rui,
> 
> There are two relocations that take place. The provider.o file (generated
by dtrace -G) contains relocations that are resolved with the final link pass
occurs. We record the function offset (libdtrace) and then the linker adds in
the base address of the function. The next relocation takes place when we
ioctl(2) the DOF to the kernel. There, we provide the base address of the
object.
> 
>  tracepoint address = base address of object + function offset + offset
into function
>                              drti.o                ld(1)         dtrace -G
/ libdtrace
Oh, apparently there was a problem with the base address of object.
I''ll try again.

Thanks,
--
Rui Paulo