Currently the Linux kernel disables preemption inside spinlocks and interrupt handlers, and these code sections are not necessarily carefully written to be short and constant-time, so you can in theory get in a situation where all cores are running spinlock or interrupt handler code from low priority threads for arbitrarily long times, preventing higher priority realtime threads to run (which in practice means that whatever hardware you are controlling will be uncontrolled, and thus your self-driving car, rocket, plane or industrial machine might be destroyed).

The PREEMPT_RT patch solves this by reimplementing all kernel spinlocks as mutexes with priority inheritance that are fully preemptible and all interrupt handlers as preemptable threads.

The result is that if you have a bunch of real-time process/thread that do not make system calls, it's guaranteed that in a bounded amount time the highest priority set will be running.

If you do however make system calls (or even if you cause page faults because you didn't call mlockall() and accessed non-mlocked memory), with the exception of some system calls that provide better guarantees, then you can still have kernel spinlock contention with other threads that may hold the same spinlock for an unbounded amount of time, so the system is more limited than dedicated real-time OSes that are careful about this.