Brizzled

... wherein I bloviate discursively

Brian Clapper, bmc@clapper.org

War Stories: Deadlocks and Semaphores

| Comments

At work a couple of days ago, we had a lively discussion about deadlocks and deadlock avoidance, as it relates to our product. The discussion somehow reminded me of my first job, and similar problems we sometimes encountered there.

After graduating from Temple University in 1983, with a degree in computer science, I took a job with the Philadelphia Stock Exchange, working in the small group that built and maintained its online trading system. The system was written in an extended and robust dialect of the Pascal programming language, and it ran on Honeywell DPS-6/90 minicomputers.

This system used two locking mechanisms to protect access to critical regions of code. The first locking mechanism was an explicit system lock (a system semaphore)), and the second was a priority switch. We had three classes of processes. Each class ran at a different priority; within a class, all processes ran at the same priority. If you had to lock a resource (e.g., a queue) that could be used across process classes, you had to use the more expensive system semaphore. If you simply wanted to lock a region of memory that was shared by processes within a class, you made a significantly cheaper system call that temporarily raised the process’s priority, ensuring that none of its brethren could preempt it until it finished; at the end of the critical region, you’d make another call to reset the process’s priority.

Obviously, attempting to put something on a queue (thus using a semaphore) within a “raised priority” region was a strict no-no, because it could result in a deadlock. Naturally, there were occasional bugs, usually in code written in haste under deadline pressure, where that happened. Having a real-time stock trading system grind to a halt because of a hard-to-reproduce deadlock (or, for that matter, because of any bug) was no fun. As we sat in the computer room at the console, desperately trying to diagnose the problem (not always easy to do), Stock Exchange VPs would start to trickle in. A row of executives with crossed arms and tapping feet is not conducive to solving a deadlock problem.

Comments