Monthly Archive for August, 2013

RPM File Colors

What are RPM file colors?

When building a package rpm will tag each file within the package with a file color. Usually the file color will fit into one of four categories as described by Jeff Johnson here. These categories are:
  • 0 is unknown or other
  • 1 is Elf32
  • 2 is Elf64
  • 3 is (contains) both
So why does rpm do this? The short answer is "for multilib support". Basically so we can install both the 32bit and 64bit version of a package on the system and have some hopes of everything still working correctly.

An example of this would be a system that needed both the 32bit and 64bit version of glibc (chances are you have them both installed because some package has been slow to move to 64bit and only provides software compiled for 32bit). The problem with having both rpms installed is that both rpms provide some of the same files (i.e /sbin/ldconfig). Which one should rpm choose? This is where file colors come in.

When installing a file from a new package rpm will check to see if the file is already provided by another rpm and will then use file color to determine if the file should be replaced or left alone. The current behavior of rpm is to prefer 64bit over 32bit files. That means when both i686 and x86_64 glibc are installed, ldconfig should be 64bit. This can easily be checked:

[root@Cent64 ~]# rpm -q glibc glibc-2.12-1.107.el6_4.2.x86_64 [root@Cent64 ~]# rpm -q --filecolor glibc | grep /sbin/ldconfig /sbin/ldconfig 2 [root@Cent64 ~]# rpm -ivvh glibc-*.i686.rpm nss-softokn-freebl-*.i686.rpm ... D: fini 100755 1 ( 0, 0) 787476 /sbin/ldconfig;521a4bd1 skipcolor ... [root@Cent64 ~]# rpm -qs glibc-2.12-1.107.el6_4.2.x86_64 | grep /sbin/ldconfig normal /sbin/ldconfig [root@Cent64 ~]# rpm -qs glibc-2.12-1.107.el6_4.2.i686 | grep /sbin/ldconfig wrong color /sbin/ldconfig [root@Cent64 ~]# file /sbin/ldconfig /sbin/ldconfig: ELF 64-bit LSB executable,...

As expected the 64bit ldconfig was left in place as denoted by skipcolor in the output. It is also shown when querying the state of the i686 glibc that /sbin/ldconfig has the "wrong color".

RPM file colors gone bad.

Yes, file colors are a nice hack (that takes place in the background) to make multilib work. Most of the time it works just fine and the user is none the wiser. However, sometimes there are nasty little complexities that this can introduce.

HP provides the hp-health rpm to help monitor HP hardware. In hp-health- /sbin/hpasmcli was 32bit. In hp-health- /sbin/hpasmcli had changed to 64bit. This can be seen by querying the rpms:

[root@Cent64 ~]# rpm -qp --filecolor hp-health- | grep /sbin/hpasmcli /sbin/hpasmcli 1 [root@Cent64 ~]# rpm -qp --filecolor hp-health- | grep /sbin/hpasmcli /sbin/hpasmcli 2

Upgrading from the older rpm to the newer rpm has the expected behavior:

[root@Cent64 ~]# rpm -i hp-health- &>/dev/null [root@Cent64 ~]# file /sbin/hpasmcli /sbin/hpasmcli: ELF 32-bit LSB executable,... [root@Cent64 ~]# rpm -U hp-health- &>/dev/null [root@Cent64 ~]# file /sbin/hpasmcli /sbin/hpasmcli: ELF 64-bit LSB executable,...

The problem comes in, however, when you try to go the other way. I know its not that common but sometimes I do "upgrade" to the older version of a package:

[root@Cent64 ~]# rpm -Uvv --oldpackage hp-health- ... D: fini 100755 1 ( 0, 0) 179005 /sbin/hpasmcli;521a55ec skipcolor ... [root@Cent64 ~]# file /sbin/hpasmcli /sbin/hpasmcli: ELF 64-bit LSB executable,... [root@Cent64 ~]# rpm -qs hp-health | grep /sbin/hpasmcli wrong color /sbin/hpasmcli

Basically what happened here is the 64bit version of hpasmcli was left in place. This was not what I expected when I originally did the "downgrade" and was caused by the subtle behavior of file colors within rpm. The workaround for this is to just fully uninstall and reinstall the version that you desire.