/dev/dump

It seems that I often run into the same problems over and over again, and I see many device drivers which often suffer from the same problem. Here are some strategies I use in my own coding -- maybe someone else will find them useful. To my knowledge they have not been documented elsewhere. KISS. Start with as few locks as you need to get the job done. Only add locks when performance analysis shows you need them. I usually start with an assumption that I'll need one lock per device instance, and possibly a single global lock. (If you don't have global state, you can elide the global lock.) For some kinds of drivers (NICs in particular) I introduce a lock for each major traffic direction (so one lock for rx, and one for tx) per instance. Global state is evil. Global state requires global locks. Global locks often introduce lock ordering problems, and can also more likely to be contended in systems with lots of devices. Thou shalt not sleep in STREAMs entry put (9E) and

Bryan Cantrill has just posted one of the more excellent blog posts I've seen this year -- he manages to put to words some of my heretofore unvoiced doubts about transactional memory. I'm a big fan of KISS, and both TM and multilevel threading seem to fail on that front.

In another update, I've updated fdc (the floppy disk controller) so that it supports suspend/resume even if a floppy drive is present. There is one caveat though -- namely that the drive must be empty. Still, its far easier to pull a disk out of a drive (if you even have one -- most floppy drives these days probably see very little, if any, use), than to disconnect the drive entirely. You can download the file from the device drivers download page (look for a file named fdc-2008-11-01.tar.gz). A webrev of the changes is also posted. For the curious, the reason for the above caveat is best stated by the following explanation extracted from a comment in the code: Bad, bad, bad things will happen if someone changes the disk in the drive while it is mounted and the system is suspended. We have no way to detect that. (Undetected filesystem corruption. Its akin to changing the boot disk while the system is suspended. Don't do it!) So we refuse to suspend if there is media p

I've updated iprb ( Intel Pro/100 Ethernet driver) to support suspend/resume and quiesce (fast reboot). I've also made some other changes, such as updating to the latest microcode from Intel, fixing some potential races, and improving the internal implementation of the statistics routine. I'd like folks to try it out on their Intel or AMD platforms. The tarball can be downloaded from here (look for iprb-2008-11-01.tar.gz). I still want to get this open sourced, but I haven't got approval from the lawyers yet. But here are the binary bits in case anyone is waiting for them. (I'm not planning on porting this driver to SPARC. There are too many places in the code that would have to be changed to use endian-safe calls to ddi_get, and I think there is probably very little demand for a SPARC version of the driver. If you are one of the people who do want SPARC support for iprb , please let me know. Enjoy!