The Linux Scheduler

Linux can only do user preemption. Linus Torvalds, it seems, doesn't believe in kernel preemption. That's not as bad as it may seem; all is fine for semaphores. Critical sections protected by semaphores can be preempted at any time, as every contention will end in a schedule and there can't be any deadlock. However, critical sections protected by fast spin locks or by hand locks cannot be preempted unless we block the timer IRQ. So all spin locks should be IRQ-safe. Also, by avoiding kernel preemption, the kernel becomes more robust and simpler, since a lot of complicated code can be saved this way.

By the way, there are tools to monitor the scheduler latency in order to allow the interested hacker to catch potential code sections that need conditional schedules.

Linux, due to its GPL nature, allows us to do things faster than others, because you can adapt and recompile your own kernel instead of using a standard fit-all kernel. For example, in some popular proprietary operating systems such as Solaris 7, many code sections have been packaged within spin_lock and spin_unlock to make the same code work well in both UP and SMP systems. While vendors of these commercial operating systems tout this as a clear advantage for heavy SMP systems, these locks actually bog down UP systems and simple SMP machines, because the same binary driver must work on both SMP and UP kernels. A spin_unlock is one locked ASM instruction:

#define spin_unlock_string \
        "lock ; btrl $0,%0"

and calling such clear-bit instruction through a function pointer is complete overkill.

Linux, on the other hand, has in-line code sections for UP and SMP systems, adapting to the machine it is running on. Therefore, if only a UniProcessor system is hosting the kernel, no time is lost locking a code section that doesn't need it.

Last but not least, as we saw above, the goodness function makes the SMP scheduler in Linux very clever. Having a clever SMP scheduler is critical for performance. If the scheduler is not SMP-aware, OS theory teaches us that the performance on an SMP machine can be even worse than on a UP machine. (This was happening in 2.2.x without the new heuristics, but has now been fixed.)

That's it for this month. I hope you gained some deeper understanding of how Linux schedules tasks on UP and SMP systems.

Resources