Package ordering in rpm-4.0.1 and later

The package ordering algorithm in rpm-4.0.1 has changed.

Here's a simple test to illustrate the need for the change (from bugzilla #12327):

Assume the minimal 7.0 package manifest in /tmp/list

	/mnt/redhat/comps/dist/7.0/sparc/bash-2.04-11.sparc.rpm
	/mnt/redhat/comps/dist/7.0.2/sparc/glibc-2.1.94-1.sparc.rpm
	/mnt/redhat/comps/dist/7.0/sparc/mktemp-1.5-5.sparc.rpm
	/mnt/redhat/comps/dist/7.0/noarch/basesystem-7.0-2.noarch.rpm
	/mnt/redhat/comps/dist/7.0/noarch/setup-2.3.4-1.noarch.rpm
	/mnt/redhat/comps/dist/7.0/noarch/filesystem-2.0.7-1.noarch.rpm
	/mnt/redhat/comps/dist/7.0/sparc/libtermcap-2.0.8-25.sparc.rpm
	/mnt/redhat/comps/dist/7.0/noarch/termcap-11.0.1-3.noarch.rpm

with database initialization as

	mkdir -p /tmp/ROOT/var/lib/rpm
	rpm --initdb /tmp/ROOT/var/lib/rpm

This command "works"

	rpm -Uvh -r /tmp/ROOT `cat /tmp/list`

while this command

	rpm -Uvh -r /tmp/ROOT `tac /tmp/list`

fails with

	loop in prerequisite chain: libtermcap bash libtermcap

Note:

The 2nd upgrade reverse orders the packages in the manifest.

The problem is that the previous ordering algorithm, basically a very clever implementation of tsort, was sensitive to initial conditions, and the first command "happens" to snip a loop, while the second does not.

The current ordering algorithm is exactly tsort from Knuth V1, with one further twist. Since the only way out of a dependency loop is to snip the loop somewhere, rpm uses hints from Requires: dependencies to distinguish co-requisite (these are not needed to install, only to use, a package) from pre-requisite (these are guaranteed to be installed before the package that includes the dependency) relations.

There is now syntax in spec files to explicitly specify the source of a Requires: dependency. If, for example, you use grep in post, then you as a packager would normally add

	PreReq: grep

in order to insure that grep was installed before attempted use by the postun scriptlet.

Now the same dependency can be expressed more precisely as

	Requires(post): grep

For completeness, here's the complete set of tokens that may be added to Requires: as in the example above:

    "interp",         RPMSENSE_INTERP
    "prereq",         RPMSENSE_PREREQ
    "preun",          RPMSENSE_SCRIPT_PREUN
    "pre",            RPMSENSE_SCRIPT_PRE
    "postun",         RPMSENSE_SCRIPT_POSTUN
    "post",           RPMSENSE_SCRIPT_POST
    "rpmlib",         RPMSENSE_RPMLIB
    "verify",         RPMSENSE_SCRIPT_VERIFY

Ditto BuildRequires:

    "prep",           RPMSENSE_SCRIPT_PREP
    "build",          RPMSENSE_SCRIPT_BUILD
    "install",        RPMSENSE_SCRIPT_INSTALL
    "clean",          RPMSENSE_SCRIPT_CLEAN

but let's not go there (yet).

For giggles, you can also do stuff like

	Requires(pre,post): /bin/sh

By marking dependencies more precisely, rpm can distinguish between an upgrade context (like the use of grep in post above) and an installed context (like the autogenerated Requires: in a package that includes a script with #!/bin/sh), and that permits rpm to differentiate pre-requisites from co-requisites while doing package ordering.

Here's what cures the libtermcap <-> bash loop:

	Requires(postun): /bin/sh

which, since the dependency is clearly not useful or necessary in determining install ordering, is safely ignored.

One of the side effects of changing the package install ordering, is that there are a handful of new loops that are detected. Here's what I found looking at supported Red Hat releases:

    ghostscript-fonts	ghostscript
    /* 7.0 only */
    pango-gtkbeta-devel	pango-gtkbeta
    XFree86		Mesa
    compat-glibc	db2
    compat-glibc	db1
    pam			initscripts
    kernel		initscripts
    initscripts		sysklogd
    /* 6.2 */
    egcs-c++		libstdc++
    /* 6.1 */
    pilot-link-devel	pilot-link
    /* 5.2 */
    pam			pamconfig

Why are there new loops? Because tsort is trying to use all of the dependency relations for ordering, while the previous tsort ignored all Requires: from added packages.

Except for the "well known" libtermcap <-> bash loop (which is just wrong), all of the other dependencies are simply not needed in an upgrade context to perform package ordering. Please note that all of the known to cause loop dependencies listed above are, for now, explicitly ignored when determining package install ordering.

So what does this all mean? Basically not much, unless you find yourself trying to specify dependencies amongst a set of packages correctly and happen to create a dependency loop.

And, before you start adding the new-fangled syntax to packages, please remember that rpm will almost certainly be auto-generating fine-grained dependencies for post et al scriptlets pretty soon. Truly, rpm needs to make packaging easier, not provide Yet More Complicated Syntax in spec files.

With thanks to Ken Estes for doing the implementation in bash2 that makes it possible to auto-generate scriptlet dependencies, blame me for the long, slow deployment.

---