Lately I faced the requirement to create a backport of net-snmp 5.7.2, which is part of CentOS 7, to a largely CentOS 5 and 6 based infrastructure. Due to some prior adventures in the rpm world I already knew about Mock (the Fedora equivalent to pbuilder).

The easy part was the backporting to CentOS 6. Here the major objectiv is ripping out the newly added systemd support. That boils down to just ripping out a patch, removing a configure flag, some %post, %preun, %postun script snippets and the corresponding %install magic from the spec file. Builds fine but leaves us with the issue that the init script ships in a package net-snmp-sysvinit but the whole chkconfig registration magic is gone.

I put that aside for a moment and moved on to CentOS 5. It did not even start to build.

  1. There is no tcp_wrappers-devel in CentOS 5.
  2. The perl libraries are in the main perl package and not in perl-devel.
  3. lm_sensors-devel is not at version 3.

I lowered the versioned build-dep on lm_sensors, disabled via a spec file toggle the tcp-wrapper support and changed the perl build-dep. This time it failed in the configure script due to DTLS support issues with the ancient openssl in CentOS 5. Decided to drop that one as well from the list of configure flags since we do not use it. Tried to build it again, now it failed somewhere down the road post compilation when assembling the package. Not cool.

At that time we pushed a few other ideas back and forth. We found an isolated patch for net-snmp 5.6 that could fix our most pressing issue and someone had a look if that could be backported to CentOS 5 easily. Nope, internal structure of net-snmp changed a lot in between. So back to backporting.

After staring long enough at the build log I could understand that the variable %{RPM_BUILD_ROOT} is empty in the CentOS 5 world. And indeed the BuildRoot stanza is also gone from the spec file. So I reintroduced that one and later even found documentation in the Fedora Wiki about this issue. If you know it it's easy to work around it, so I just defined it by hand:

BuildRoot: %{_tmppath}/%{name}-%{version}-%{release}-root
%global RPM_BUILD_ROOT %{buildroot}

About ten minutes later I had a set of rpm files for CentOS 5.

Tried to install it, failed due to a depedency on libmysqlclient15 which is part of the mysql package in CentOS 5. Yes, eight years back RedHat did not split off the lib packages and that would mean for us to install a mysql database on every system. So no, we won't do it and we just removed the mysql support.

Left with only the init script setup missing I tried to be clever and failed miserable. Be warned, this would've also been a bad idea in the Debian world.

Since we have the init script in a seperate binary package I thought I could just use something like

Requires: net-snmp-sysvinit

in the main net-snmp package and re-add the old %post, %preun and %postun magic. Of course that failed at install time for a good reason.

In the second (desperate) try to work around it I tried something similar to "PreDepends"

Requires(post): net-snmp-sysvinit 

That works at install time but still breaks on a removal/update.

So I think I did the right thing in the end (after reading some more documentation) and just added the scripts to the net-snmp-sysvinit package. It's just about adding the package name to the %post macro invocation.

My initial plan included the wish to have one spec file that is usable to build on CentOS 5 and 6 but now I'm left with a number of special cases for CentOS 5.

I already knew about the %if %endif conditional macro you can use in rpm spec files, so what was missing was a proper variable to base my decisions on. Every CentOS release ships with a "/etc/rpm/" directory that includes several macro file. Within those one can find the %{rhel} variable that defines an int we can use.

%if %{rhel} <= 5
BuildRequires: perl
BuildRequires: perl-devel

And finally I could build from one spec file a backport for CentOS 5 and 6. My diff against the CentOS 7 spec is at

Drawbacks of the resulting packages:

  1. No tcp-wrapper support, we don't care but that should be just another package name issue for the conditional Requires, simililar to the perl isssue.
  2. No mysql support. We do not care, one could change that and/or make it conditional or build against newer mysql releases where the -lib package is split.
  3. No DTLS support on CentOS 5. We do not care and that is not easy to fix.
  4. You've to install the net-snmp-sysvinit script to get back the usual setup you'd expect.
  5. We ripped out the systemd support, I did not investigate yet if it's possible to keep that included and just not enable it.

By now CentOS 5 is old. It's painful to backport to it. But it's still possible. Another strategy someone tried was packaging the current net-snmp upstream release based on the CentOS 5 spec file. That would've meant throwing away all 106 patches RedHat includes in that package at the moment and to depend on new upstream releases for your own security support. In case we start using CentOS 7 we would also diverge on the version of snmpd we use. So we decided to stick with the RedHat sources and keep the version consistent on our infrastructure even if we can not keep feature parity.

After all the issues one encounters while backporting packages in the rpm world are similar to those in the dpkg world. Still it feels different if you're not familar with the toolchain. For example

  1. If you build a src.rpm with mock you've to copy it out of the result directory before you rebuild it with mock to create the binary package. Mock deletes everything from the result directory when starting a new build. So I guess parallel builds for the same chroot will not work.
  2. Depedencies on files instead of packages are a strange thing.
  3. How on earth do you maintain a patchset of over a hundred patches without something like quilt? I guess you always have to apply them in a for loop on your source directory and then you create a copy of that to implement your modifications and create another patch. I'm pretty sure eight years ago Fedora was far away from using git so I would rule out patch branches in a VCS for the moment. Not sure what they invented inside RedHat to keep track of it.