Without Preempt-RT, the Linux kernel has multiple sections where preemption is disabled:

• Spinlock-protected regions
• Interrupt handlers
• Portions of the scheduler

Such sections introduce unbounded latency for real-time tasks that need to be scheduled on the CPU to meet their deadline.

Preempt-RT addresses this mainly by:

• Converting spinlocks into preemptible, sleeping locks implemented via rtmutex
• Turning interrupt handlers into preemptible kernel threads, which by default run under the SCHED_FIFO scheduling policy at priority 50 (MAX_RT_PRIO / 2), midway between the minimum and maximum real-time priorities
• Using high-resolution hardware timers with nanosecond precision in the kernel timer subsystem
• Applying priority inheritance to prevent priority inversion across lock chains

This results in a kernel where a real-time task can preempt almost any other kernel code path, thus getting it scheduled on the CPU before its deadline expires.

Worst-case wake-up latency on modern hardware is typically in the range of 20–50 microseconds, depending on the SoC, interrupt load, and configuration.