Mutual Exclusion for Critical Sections of Code

A principle of synchronization regulates the orderly flow of data into and out of structures and prevents resource contention. Thus, in an MP system, thread A executing on Processor 1 must not contend with thread B executing on Processor 2.

Figure 1-4 Synchronization

Three kinds of critical sections within the kernel require mutual exclusion:

In a uniprocessor environment these contentions were easily dealt with: Mutual exclusion was implemented for two interrupt service routines or an interrupt service routine and a thread of control by raising spl levels to the highest priority interrupt service routine. To ensure mutual exclusion between threads of control, no thread could be preempted while running in kernel mode.

These protection mechanisms are inadequate for an MP environment. The spl routines were local in nature and affected only the interrupt protection level of the calling CPU. Waiting for the current process to reach a safe point, sleep, or exit the kernel failed to give the desired parallelism and made for long, non-preemptable critical durations.

In HP-UX, kernel data structures are protected with software semaphores, locks, and synchronization primitives. Kernel data structures are then divided into sets, with a semaphore or lock guarding each set. The granularity of the semaphores and locks are empirically determined to minimize blocking of threads of control on the semaphores.