failed unit test: sal_osl_file

I am trying to build master commit id 5c2c0e3, pulled 2012-05-08, on
ubuntu-natty (11.04), and it is failing in unit test sal_osl_file.

An excerpt from the terminal output from `make sal` is in attachment
20120509c.  I think the bad address passed to free() is just a
secondary error, right?  Is it of interest on its own account?


A backtrace from gdb is in attachment 20120509d.  I think I see the
following ...

(*) sal/inc/osl/file.hxx line 1645 sets Directory._pData to 0

    ( Directory::isOpen(), defined at line 1691 of the same file, and
    Directory::close(), particularly line 1712 of the same file,
    suggest that _pData equal NULL is the normal state for a closed
    Directory. )

(*) qa/osl/file/osl_File.cxx line 5562 applies that to testDirectory
    and in particular to testDirectory._pData.

(*) the same file, line 5563, frame 19 in the gdb backtrace, invokes
    Directory::getNextItem on testDirectory.

(*) sal/inc/osl/file.hxx line 1772, frame 18 in the gdb backtrace,
    passes that zero as parameter 1 to osl_getNextDirectoryItem.

(*) sal/osl/unx/file_misc.cxx line 321, frame 17 in the gdb backtrace,
    asserts that parameter 1 of osl_getNextDirectoryItem is non zero,
    and we are doomed to failure.


So, I ask ...

(*) Am I reading the code correctly?

(*) If so, how could this unit test have ever come close to working?

(*) What should I do?

(*) The presence of same-named data members _pData in related classes
    FileBase, Directory, and DirectoryItem is a normal programming
    convention.  I should just get over my confusion.  Right?

(*) Those three data members _pData are of three different types, each
    typedef'd to `void *`.  So the compiler will not diagnose the use
    of one where another was intended.  This is not something to be
    fixed.  Right?


Thank you for your help.
Terry.


_______________________________________________
LibreOffice mailing list
[hidden email]
http://lists.freedesktop.org/mailman/listinfo/libreoffice

On Wed, 2012-05-09 at 17:17 -0400, Terrence Enger wrote:
> I am trying to build master commit id 5c2c0e3, pulled 2012-05-08, on
> ubuntu-natty (11.04), and it is failing in unit test sal_osl_file.

You didn't flip from a system cppunit to an internal one without a
rebuild inbetween, or anything like that ?

C.

_______________________________________________
LibreOffice mailing list
[hidden email]
http://lists.freedesktop.org/mailman/listinfo/libreoffice

On 05/09/2012 11:17 PM, Terrence Enger wrote:
> I am trying to build master commit id 5c2c0e3, pulled 2012-05-08, on
> ubuntu-natty (11.04), and it is failing in unit test sal_osl_file.

I saw this exact same strange error with the
Linux-RHEL6-x86_64@14-with-check tinderbox recently, and assume it is
related to the GCC version used on that machine (RHEL 6 GCC 4.4.6) in
combination with the recent clean-up of how configure options affect the
-O level (switching a --enable-dbgutil build from -O0 to -O2,
potentially uncovering optimization-related issues).  I worked around it
by adding --enable-debug (so I'm back at -O0).

Stephan
_______________________________________________
LibreOffice mailing list
[hidden email]
http://lists.freedesktop.org/mailman/listinfo/libreoffice

On Thu, 2012-05-10 at 16:18 +0100, Caolán McNamara wrote:
> On Wed, 2012-05-09 at 17:17 -0400, Terrence Enger wrote:
> > I am trying to build master commit id 5c2c0e3, pulled 2012-05-08, on
> > ubuntu-natty (11.04), and it is failing in unit test sal_osl_file.
>
> You didn't flip from a system cppunit to an internal one without a
> rebuild inbetween, or anything like that ?


I think not.  But my system does have ubuntu-supplied
libcppunit.  Could a change in the build system since my
previous successful build (2012-04-01) have done this
silently?  My autogen.lastrun was last changed 2012-03-09,
and it shows no reference to cppunit:

    --disable-mozilla
    --enable-symbols
    --enable-dbgutil
    --enable-crashdump
    --disable-build-mozilla
    --without-system-postgresql
    --enable-python=internal

Moreover, the error looks the same after `make clean`.


Meanwhile, I applied a blunt instrument to the failing
testcase.  This lets the build progress to CUT sal_osl_module.
This unit test fails in SAL_INFO at module.cxx:321, which
leads to "free: invalid pointer" and SIGABRT.  One must
suspect a problem between the system-supplied stream
functions and the LO-supplied custom allocator, but to trace
through that is beyond my skill level.  (If I were smarter,
I would have given up sooner.  Sigh!)


gdb, of course, eats the return code from a failed unit
test.  So I can go on to see four more unit tests fail with
"free(): invalid pointer" after one of the "message" macros.



Thank you for your patience.
Terry.


_______________________________________________
LibreOffice mailing list
[hidden email]
http://lists.freedesktop.org/mailman/listinfo/libreoffice

On Fri, 2012-05-11 at 14:23 +0200, Stephan Bergmann wrote:
I have added --enable-debug to my configure options.  The build seems to
have passed the previous point of failure.

Thank you for your help.
Terry.


_______________________________________________
LibreOffice mailing list
[hidden email]
http://lists.freedesktop.org/mailman/listinfo/libreoffice