fail as a result. Then report this number in the generated email and
html summary. The goal is to help the pkgsrc crew focus our efforts
on the broken packages which have the largest impact. Thanks to
Christoph Badura for suggesting that I do this and Hubert Feyrer who
suggested an easier to read output format.
over several bulk builds on an alpha:
- At the start of a bulk build, 4 files are created to allow fast lookup of
various dependency tree things. These files are
.index == maps package name (foo-2.3) to directory (bar/foo)
.dependstree == contains the complete pkgsrc dependency tree in a tsort(1)
compatible format.
.depends == contains 1 line per package and lists all the build depends
for the package.
.supports == contains 1 line per package and lists all packages which
depend on this package.
- When a package fails to build, the list of all packages which depend upon
it is read from .supports. Each of those packages is immediately marked
as broken. This prevents us from trying to build those packages which can
have significant overhead if lots of other depends are installed before the
system notices the broken one.
In addition, the post-build postprocessing will now indicate that a package
is 'truely broken' (ie, bad PLIST, failed compilation) vs. a package which
is broken because it depends on a failed package. This assists in determining
where to focus our efforts in fixing broken packages.
- In the old approach, all packages are removed after each one is built. The
purpose was to a) conserve disk space, b) verify that all required dependencies
are in fact listed, and c) prevent conflicts. The problem was that often times
several packages in a row have similar depends. For example we might want to
build several perl packages in a row. In the old approach, we would install perl,
build the pkg, deinstall perl and continue with the next one.
In the new approach, when it is decided that a pkg is out of date and should be
rebuilt, the .depends file is used to obtain a list of pkgs we require. Then
and pkgs which are no longer needed are removed. This helps to minimize the total
number of pkg_add/pkg_delete's in a bulk build.
- Since the order of the build is controlled by the depends tree, all depends for a
given package will have been built by the time the pkg in question needs them.
If any of the depends failed to build, then the pkg which needs the failed one will
have been marked as broken. Given this, the complete depends list for a pkg is
read from .depends and the depends installed via pkg_add rather than relying on
recursive make calls to install the depends.
- while here, fix a few minor bugs
x - don't leave .make log files around when the build succeeds
x - make sure we refer to the correct report file in the email
x - use '.order' for the build order file instead of '.l'
x - use 'grep -c' instead of 'grep | wc -l'
- quote arguments to pkg_admin so the shell doesn't try to expand any *'s.
- be sure that all required binary packages exist. The previous version
made sure the required packages were not newer, but missed catching missing
required packages.
Still todo:
When examining required binary packages to see if they have changed, the test
only looks at the modification time of the binary package. It does not verify
that the required binary package is also up to date with respect to its
pkgsrc files.
