Re: [Patch] Losing many samples when profiling multiple JVMs

Hi,

Sorry for the late response.  I was on vacation.

Sorry, I totally forgot about the ABI difference between big endian
and little endian on ppc....

I found the anon_remap test succeeded even with OProfile 1.0.0,
where the test must have failed.  I am attaching a fix patch to
the test case. I changed the matching string from "anon" to " anon ",
because "anon" can match not only " anon " but also the name of
the binary, "anon_remap".

The fixed test correctly failed with 1.0.0 but succeeded with the master
branch version.


Regards,
Rei Odaira


2015-07-16 15:42 GMT-05:00 Will Schmidt <[email protected]>:

> On Thu, 2015-07-16 at 10:43 -0400, William Cohen wrote:
> > On 07/16/2015 10:27 AM, Will Schmidt wrote:
> > > On Wed, 2015-07-15 at 15:03 -0400, William Cohen wrote:
> > >>
> > >> I tried the test on rhel7 ppc64 and it doesn't work because of the way
> > >> that calls operates.  The ppc64 generated code is treating the address
> > >> returned by
> > >> initialize_anon_func as a pointer to a descriptor rather than the
> > >> address that the branch should branch to.  The first 3 double word in
> > >> the generated anon_func are used as a descriptor and the branch goes
> > >> off to the wrong location
> > >> (
> https://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi-1.9.html#FUNC-DES
> )
> > >>
> > >> Is this code working on ppc64le? How is the code being compiled on
> > >> your machine? Or does it have a similar problem?  Any suggestions on
> > >> how to best address this?
> > >
> > >
> > > Note that the PPC64le platform (which uses the ELF ABIv2) does not use
> > > function descriptors, so it is possible the code actually just works on
> > > ppc64le.
> > > I've not actually tried this patch, but saw the comment go past, and
> > > thought I should chime in in case folks were off scratching their heads
> > > in confusion.. :-)
> >
> > Hi Will,
> >
> > The problem wasn't with oprofile data collection; it was with some
> > test code that was generating code and a pointer to that code.  On
> > ppc64be that pointer was be interpreted as a pointer to the function
> > descriptor and the beginning of the function was being interpreted as
> > fields of the function descriptor causing the processor to jump off to
> > the wrong location. I did find information about ELF ABIv2 saying it
> > didn't use function descriptors. Last night I added an ifdef to the
> > anon_remap.c test in the testsuite that should only be enable
> > generating function descriptor generation for powerpc64be. I tested
> > that
>
> Ok, yeah, that code change looks good.
>
>
> >  the code worked on powerpc64be and x86-64. It would be good for
> > someone with ppc64le access to verify that it still works for ppc64le.
> >
> >   git clone git://git.code.sf.net/p/oprofile/oprofile-tests
> >   cd oprofile-tests/testsuite
> >
> > as root:
> >
> >   runtest --tool=oprofile
> >
>
> Thanks for the clear steps above. :-)
>
> I ran the suite against an assortment of Fedora/RH/SuSE/Ubuntu LE
> installs and nothing pops out with obvious errors, with the exception
> that a few of the environments throw an error on an XML test, which I
> believe is unrelated to this change.
>
> I'll retest later with newer oprofile build in each of those
> environments, but overall I think the change is good.
>
> Thanks,
> -Will
>
> >
> > -Will Cohen
> >
>
>
>

------------------------------------------------------------------------------

_______________________________________________
oprofile-list mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/oprofile-list

diff --git a/testsuite/lib/op_workloads.exp b/testsuite/lib/op_workloads.exp
index 88676ba..661e280 100644
--- a/testsuite/lib/op_workloads.exp
+++ b/testsuite/lib/op_workloads.exp
@@ -59,5 +59,5 @@ set op_workload_table(anon_remap)                  \
       {workloads/anon_remap_src/anon_remap.c}      \
       {}                                           \
       {}                                           \
-      {anon}                                       \
+      { anon }                                       \
     }