What's a "rpikernelhack"?

jimrh · November 15, 2021, 11:43pm

When running updates on Charlie, one of the things that happens is a whole bunch of “rpikernelhack” and “diversion” statements.

I figured it had to do with trying to fit the square peg of a FAT file system into the round hole of Linux-style filesystems.

Fussing around today, I found the canonical answer over at the Raspberry Pi Stack Exchange forum, written by a gentleman named Peter Green, who claims to have developed this “hack” to allow the Raspberry Pi to install updates.

He wrote:

“rpikernelhack” is a fake package name and a directory name used as part of a hack (in the sense of a dirty but expedient solution to a problem) to work around the fact that the Raspberry Pi foundation decided to make /boot a fat32 partition and dpkg does not get on well with fat32. I was the one who initially came up with the idea, though it was refined later by others.

dpkg will install new files onto a fat32 partition (spewing some warnings along the way), but if it tries to update an existing file on a fat32 partition it will fail (iirc it tries to make a backup of the old file by creating a hardlink and fat32 doesn’t support hardlinks).

When people (including me) started trying to make deb packages of Pi kernels and firmware they ran into this problem, a package would install initially but trying to upgrade it would fail, ouch.

My workaround was to (ab)use the “diversion” feature in dpkg. This feature was intended to allow files to be diverted so they could be replaced with either locally modified versions or versions from another package, but I was able to use it from the maintainer scripts in such a way that dpkg would perform its installation tasks on a Linux partition and then move the file to its final location at the end.

Diversions require you to either specify a “package name” or “local”. If you specify a package name then the diversion will affect files owned by all packages except the one you specify (the intent here is to allow a package to divert a file owned by another package and then install its own versions). I also needed a directory to divert the files to.

Using the name of the kernel package being installed would have rendered the hack ineffective. Using “local” also seemed wrong, since that is supposed to be reserved for use by the local sysadmin. So I needed a fake package name that was unlikely to conflict with anything. I came up with “rpikernelhack”, I also used this same string for the directory name.

An interesting insight!

What say ye?

cyclicalobsessive · November 16, 2021, 3:50am

Thank you for solving this years old mystery; seeing the word “hack” scrolling by was always discomforting.

I had presumed it was to allow updating the kernel while the kernel was running, but could never figure out why the kernel on the disk could not be updated since the running kernel was in memory. I further presumed it might be to not actually update the disk kernel until it was absolutely clear the update had succeeded.

Only to find out the hack is related to that ancient 8-inch floppy drive my buddies and I were allowed to take home from HP around 1978 time frame for our “g-job” 8080 computers. HP supported a personalized education program with the latest Intel and Motorola computers. We had taped up all the printed circuit cards, including a disk drive interface card and figured out how to read and write the File Allocation Table.

Some technologies seem to survive long after they should.

Mystery solved - wow. Thanks.

jimrh · November 16, 2021, 10:10am

If you really want to be picky, any disk file system is FAT-like in it’s structure, since every disk file system needs to know which parts are allocated and which are free.

The FAT filesystem uses the File Allocation Table as the volume bitmap and directories are linked lists of files and/or other directories.

The various Linix filesystems also have a volume bitmap that is stored in each superblock if I remember correctly.

The difference between these filesystems and FAT is in its complexity and error recovery mechanisms.

FAT, (and FAT-32), are popular because they are, (comparitively) simple. Since a file in a FAT filesystem is represented by a linked list of sectors, it’s easy to traverse.

cyclicalobsessive · November 16, 2021, 2:47pm

Yes, simple for a serial logic processor. I have rarely been able to imagine or understand structures that simplify multi-processing.

I am simply a serial logic bot in a parallel event world.