GoogleCodeExporter
commented
8 years ago It turns out this is a problem with the endianness check of the
configure.ac script. The trouble is as follows:
If the user specifies a custom location for libevent (using
--with-libevent=), this custom location is not added to the
LD_LIBRARY_PATH for the configure process. This means that
programs compiled by the configure script will not run correctly.
In the case of the endianness check (AC_C_ENDIAN), a program is
compiled which will have exit status 0 for big endian machines
and 1 for little endian. However, the AC_C_ENDIAN only checks if
the exit status was 0 (success) or non-0 (failure). In the case
of a program which fails to run due to a shared library not being
found, the program exits with a non-1 status. However, this error
is swallowed and the machine is incorrectly assumed to be
little-endian.
This bad define (ENDIAN_LITTLE on a big-endian machine) leads to
the problem descibed above, of some keys being stored and others
not.
Note that this problem will only be revealed on big-endian
machines (e.g. Sparc) because if AC_C_ENDIAN fails on
little-endian machines (e.g. Intel) the problem will be hidden by
the fact that the endianness of the machine is set correctly.
Note also that I suggested above that the problem was due to a
changed libevent. This was in fact incorrect. The reason that I
was getting correct complies with libevent 1.2 but not other
versions was because libevent 1.2 was installed in
/usr/local/lib. So if I used --with-libevent=... to compile with
libevent 1.2a, the binary generated by the AC_C_ENDIAN check was
linked with libevent-1.2a.so, which was found in /usr/local/lib,
leading to a correct result, leading to ENDIAN_BIG being defined.
If memcached was compiled with a libevent version other than
1.2a, the shared library was not located when the AC_C_ENDIAN
check was run, leading to a ENDIAN_LITTLE being defined, and a
bad memcached binary.
Attached is a patch which does two things:
1) sets LD_LIBRARY_PATH to include the dir passed in as --with-libevent;
2) errors if the endian check binary exits with an exit status
other than 0 or 1.
I believe that it would be work checking all instances of
AC_RUN_IFELSE to ensure that exit statuses > 1 are not assumed to
be equivalent to exit status 1.
After applying this patch, running autoconf, and rebuilding, the
problem should be fixed.
Original comment by ehetzner@gmail.com on 29 Apr 2010 at 1:16
Original issue reported on code.google.com by
ehetzner@gmail.comon 25 Jan 2010 at 11:47Attachments: