I decided today that I could do with a day without sightseeing and it was time for some good oldfashioned hacking. I’ve been wanting to graph the downloads from the FreeBSD mirror I’m administrating for some time, but never found the time. Today was a good day to do something about that.

The most significant, and easy to quantify, data is the number of installation disc downloads, so I wrote a quick and dirty logfile parser to filter out the disc1 isos. Quite a large number of downloads are aborted or cancelled halfway, so they needed to be filtered out. There still seems to be an issue with the data from some 10 weeks back seemed really glad to graph data per second and my measurements are only once a day, they were a factor 86400 off. Changing datatypes from absolute to gauge fixed this, so I’m happy to let this run while I’m off to Sydney at 5am tomorrow.

After removing CVSROOT/modules, all hell broke loose. A few committer still used it as a shorthand for checking out ports and saving a few keystrokes. Edwin Groothuis offered to write a script that would update the file once a day and generate it from INDEX. This will keep everybody happy, as the file is still there to use, but it no longer needs to be manually maintained, and the code does not need to be added again to the addport/rmport scripts after I edited them. This should have been done years ago, but there you have it. Thanks a lot to Edwin for writing the automation script!

After years of frustration voiced by many committers over time, Pav finally put the word out to remove CVSROOT/modules. It only adds a shorthand for checking out ports, so only the port name needs to be typed and not the full path, i.e. ports/$category/$portname. However, this can be easily achieved by a short shellscript (for other committers, see ~erwin/bin/coport on freefall for an example), and it takes quite an effort to maintain the modules file, as you can guess from its revision number of 1.20060. I would know, for some time I’ve been running a nightly script to check if all entries are done right and/or are missing and informing committers if anything is amiss. Thus, after considering the cost vs. benefit, I have emptied the file and edited addport/rmport to no longer edit it. Adding a new port was never this easy.

Oh, and it’s not every day you got to commit to a file and remove over 18.000 lines from it. Yay!

So far, I’ve had a wonderful day. It started by trying to figure out why several hundred of packages are failing on i386-6. After a long time trying to find a general pattern, I boiled it down to xorg-libraries. It turns out that when Kris did some spring cleaning and removed the /usr/X11R6 directory from bindist, it is now a symlink created during pre-install in xorg-libraries. What happens next on deinstall is that pkg-install sees the symlink and edits several /etc files to remove the mention of X11R6 there so those scripts are not run twice. As the removal of X11R6 in /etc was not MFCed to RELENG_6, this means that those files are changes and thus show up as leftovers on pointyhat.

So far so good. Now, how to fix this:

1. MFC the changes to /etc. This may break existing installations, but only if src is updated and not ports.
2. Readd the directory to bindist on 6 so the symlink isn’t created and pkg-install does not edit /etc. Nasty hack.
3. Don’t run pkg-install during PACKAGE_BUILDING. This is already set in Makefile, but not in the package. And now comes the fun part. Moving this to the package will cause problems as PACKAGE_BUILDING is set during sysinstall, so installing xorg packages from sysinstall will leave X11R6 behind in /etc.

Stay tuned for fixes.