Automounters (16 Aug 2003)

(IV should stop falling over quite as often now. It looks like there's a bug in the sunrpc code in 2.4.20 kernels. The maintainer doesn't know why, but 2.4.21 fixed it.)

I'm rewriting a replacement user-land automounting daemon at the moment which I'll post here when it's kindof ready. The two current solutions (autofs and amd) are either too fragile (autofs) or too bloated and too fragile (amd). Sadly enough, amd came from Imperial CSG originally and has turned into a monster over time.

I'm a little limited in what I can do on my laptop however (I'm at home for the weekend) because it only has GCC 2.95 which lacks the nice C99 features that GCC3 supports. (Actually I'm a little hazy on which are C99 features and which are GCC extensions. But since the automounter is only ever going to run on Linux they are pretty much the same).

Lexical functions and C++ style variable decl placement are an absolute wonder (though I have found a couple of placement gotchas which might be because I was using a CVS GCC). One thing that I'm not so sure about is run-time array sizes.

This allows the size value of an array to be a run-time variable. I suppose this (with the more flexible decl placement) is a `cleaner' way of doing alloca. But it was certainly a surprise when I forgot a sizeof, thus making the array size run-time and (so it happened) negative. It certainly scared the crap out of gdb.

And speaking of gdb, it doesn't seem to understand much that GCC3 turns out. I guess I need to start CVS chasing gdb again (like I did when pthreads support was still going in).

2.6 Kernels

I got 2.6.0-test3 to boot. This is the first 2.5/2.6 kernel that has ever managed to boot on my system. Personally it doesn't seem too obscure to me, but the AIC7xxx and megaraid cards have always freaked out a 2.5/2.6 kernel until now. (Which was kindof a bummer since the only root device I had left without those was a floppy disk).

I'm not sure if it's really that much quicker than a fully patched (preempt etc) 2.4, but at least it saves me from patching every kernel version with XFS. The new sysfs seems a little weird. I guess it's still a -test quality release, but different devices don't seem to argee on how to format things like device numbers and the like.

Still, udev thinks it can simulate devfs in userspace via sysfs and /sbin/hotplug, which is quite neat. I guess I'll need a `real' /dev directory again however, since you can't mount udev at boot like you can with devfs.

Also, I want disklabel support added to grub and the kernel. You can quite happily put UUID= and LABEL= tags in /etc/fstab, but you still need to give `real' device numbers to the kernel and grub. Worse yet, the kernel device numbers depend on the order of Linux drivers loading and grub's depend on the order of BIOSes loading. Under 2.4 there is pretty much no good way (that I know of) to get grub device numbers.

I think that the kernel people would say that support for labelled roots should go in an initrd. And I would agree with them if I didn't hate initrd's so much for all the pain they have caused me. I guess I should to hack linux/init/*.c.

I should also post my grub -R patch here (one shot reboots).


The BBC are reporting (lack of inet connection - no link at the moment) that Microsoft has `neutralised' the MS Blaster worm. What they have actually done is move the domain so that they don't get flooded off the face of the planet which would have been just reward.

It's still crashing unpatched systems left right and centre because it's so badly written. Once again, the world has been saved by the abject cluelessness of black-hat-wannabe-kiddies.

And from the world of less clueless coloured-hat people; the latest Phrack is out. Phrack is very much worth reading. It has a fair few wordz with too many z's on the end, but the actual content is of a very high quality. I certainly want to play with the ELFsh program.

Oh, another item for the wish list, code that will take a dynamic ELF and make it static.

And I guess there are a fair few computers on the Niagara Mohawk grid that haven't been patched yet .

(I shouldn't smile about that actually. The UK has deregulated the grid recently too and investment has fallen to 0. Sure, prices are lower but people are asking where all that money came from yet.)

Disk. Who needs that?

tmpfs is a cool thing. You almost certainly have it built in if you're running a 2.[456] kernel. I have a box sitting in the DoC machine room that, at rc.sysinit time, copies a Gentoo stage3 into a tmpfs and chroots into it. It has no swap enabled so you can (and I did) pull the SCSI ribbon off the motherboard (live) and the machine doesn't even blink.

It will be logging to the disk at some point and all those multilog and syslog-ng processes will drop into disk wait, but that shouldn't won't the main function of the box. My main worry is that syslog-ng creates /dev/log as a UNIX SOCK_STREAM (I have yet to test this) in which case some stuff that syslogs will lock up too.

My code expects syslog to block and will carry on working (dropping log messages as needed), but I'm not sure about sshd and the like. The solution is to make /dev/log a SOCK_DGRAM in the syslog-ng source code I guess.