5.0.7 problems with libsocket.so.2
Sagara
Recently I installed Maintenance Pack 3 on OpenServer 5.0.7.
Now some of my programs have stopped working and return the following
when I try to run them:
"dynamic linker: tethereal : UDK executable cannot load OpenServer
shared object: /usr/lib/libsocket.so.2"
It's not just a single program, multiple ones are having this issue. I
tried uninstalling the Maintenance Pack to restore older copies of
libsocket.so.2 but it didn't help.
Anyone run into this before or have any idea how I should start
tracking down/fixing this problem?
Thanks,
Sagara
Bela Lubkin
Where did the affected binaries come from? Are they in fact UDK
binaries (built with the UnixWare development system or with a GCC built
for UnixWare)?
I would guess they are OSR5 binaries which aren't properly marked as
such. MP3 made some subtle changes to how OSR5 and UDK binaries are
distinguished. As far as we know, the only binaries that would be
affected are ELF binaries built with a few very old builds of GCC for
OSR5 (I believe the affected GCC builds predate EGCS).
There is a tool which may help: `elfmark`, /udk/usr/ccs/bin/elfmark.
You can use this to change the dynamic linker's idea of what sort of
binary it's looking at. For instance:
# cp tethereal tethereal.test
# /udk/usr/ccs/bin/elfmark -t osr5 tethereal.test
# ./tethereal.test
Test it on a copy, in case it doesn't help or even makes things worse.
If it _does_ fix the problem, go ahead and `elfmark -t osr5` the real
binaries you're having trouble with.
Technical details:
Starting with OpenServer 5.0.0, OSR5 ELF binaries are marked with a
.note section with certain specific contents. Older versions of the
dynamic linker assume that any ELF binary with a .note section of the
proper _size_ is an OSR5 binary. But this could cause mis-recognition
of other binaries (.note is a valid section that could be used for any
number of purposes, and might legitimately have any size). The newer
dynamic linker in OSR507MP3 checks the actual _contents_ of the .note
section.
All known compilers which generate OSR5 ELF binaries generate the
correct .note section -- except a few ancient GCC builds. Those ancient
GCCs generate a .note section of the expected size, but incorrect
contents. So those are now mis-recognized as UDK binaries.
The `elfmark` utility writes a different stamp into the binary's ELF
headers. This is not a .note section but an actual "type of this
binary" field in the header. Marking a binary with `elfmark` makes the
decision unambiguous.
You might ask: so why don't OSR5 compilers write that stamp in the first
place? Well, I'm pretty sure I used to know why, but I don't remember
at the moment. I remember that there is some reason; it wasn't just
forgotten. But this is irrelevant anyway, since the problem is for
newer dynamic linkers (with their more sophisticated OSR5 vs. UDK binary
detection) to handle _old_ binaries correctly -- binaries produced by
compilers which manifestly don't do this stuff right. Binaries produced
by the OSR5 devsys don't need the ELF header mark since they have
correctly formed OSR5 .note sections.
>Bela<