The standard UNIX kernel is a non-preemptive kernel; it does not allow a user process to preempt a process executing in kernel mode. Once a running process issues a system call and enters kernel mode, preemptive context switches are disabled until the system call is completed. Although there are context switches, a system call may take an arbitrarily long time to execute without voluntarily giving up the processor. During that time, the process that made the system call may delay the execution of a higher-priority, run-able, real-time process.
The maximum process preemption latency for a non-preemptive kernel is the maximum amount of time it can take for the running, kernel-mode process to switch out of kernel mode back into user mode and then be preempted by a higher-priority process. Under these conditions it is not unusual for worst-case preemption to take seconds, which is clearly unacceptable for many real-time applications.
