[Libtirpc-devel] Build problems after changes to RPC database support

[Markus M. <mm...@br...>]

Hi all,

We have come across a build issue with libtirpc-1.3.7 that was not
present with libtirpc-1.3.6. It happens only with a certain toolchain
/ libc combination and is due to the introduction of the
"--enable-rpcdb" configure option.[1]

The build is now failing with our LLVM + musl-libc toolchain. As far
as I can tell, the issue isn't related to LLVM but caused by the fact
that musl-libc does not provide getrpcbyname() and getrpcbynumber().
Specifically, rpcbind will fail to link with the new libtirpc. The
error looks like this:

gcc  -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64
-O2 -g0  -I/local/users/jenkins/workspace.o/buildroot_linux-5.4_llvm/output/stbllvm-13.0-0.6/arm/host/bin/../arm-buildroot-linux-musleabihf/sysroot/usr/include/tirpc
  -o rpcinfo src/rpcinfo.o
-L/local/users/jenkins/workspace.o/buildroot_linux-5.4_llvm/output/stbllvm-13.0-0.6/arm/host/bin/../arm-buildroot-linux-musleabihf/sysroot/usr/lib
-ltirpc
ld.lld: error: undefined symbol: getrpcbynumber
>>> referenced by security.c
>>>               src/security.o:(logit)
clang-13: error: linker command failed with exit code 1 (use -v to see
invocation)
make[2]: *** [Makefile:534: rpcbind] Error 1
make[2]: *** Waiting for unfinished jobs....
ld.lld: error: undefined symbol: getrpcbynumber
>>> referenced by rpcinfo.c
>>>               src/rpcinfo.o:(main)
>>> referenced by rpcinfo.c
>>>               src/rpcinfo.o:(main)
>>> referenced by rpcinfo.c
>>>               src/rpcinfo.o:(pmapdump)
>>> referenced 3 more times

ld.lld: error: undefined symbol: getrpcbyname
>>> referenced by rpcinfo.c
>>>               src/rpcinfo.o:(getprognum)
clang-13: error: linker command failed with exit code 1 (use -v to see
invocation)

So, while the build error happens with rpcbind, the root cause lies
with libtirpc.

With libtirpc-1.3.6, this was not an issue, because it would detect at
configure time that it needed to provide getrpcbyname() & co, and so
rpcbind would link just fine. Configure in 1.3.7 is still detecting
correctly that getrpcbyname() & co. are missing, but it will do
nothing about it unless "--enable-rpcdb" is given on the command line.
Without "--enable-rpcdb", getrpcent.c will not be compiled in,
regardless of configure's findings, so the missing functions will not
be provided, and the link stage of rpcbind will fail.

Of course, this can be worked around by specifying "--enable-rpcdb"
when calling "configure" for libtirpc, but it is still a regression.
The build of another tool (rpcbind), which has remained unchanged,
suddenly fails, because libtirpc was upgraded from 1.3.6 to 1.3.7.

If I understand the commit message correctly, the change in question
was introduced to fix linker issues, specifically with LLD. Curiously,
we are now *experiencing* linker issues with LLD as a result of this
change (not that it is LLD's fault at all in this case).

Is there a way the original goal of the fix can be preserved while at
the same time also not causing the kind of linker problem we are
seeing? Would it work to have "configure" enable getrpcent.c whenever
it detects getrpcbyname() & co are missing without requiring a flag
given by the user? It doesn't seem right that it correctly detects the
functions are missing, but still won't include them, even though it
could.

Please let me know if there are some things I should try out or some
experiments I should run.

Thanks,
-Markus

[1] https://git.linux-nfs.org/?p=steved/libtirpc.git;a=commitdiff;h=7cea8ad66aec