Concurrent garbage collection involves several operations such as read/write barrier code, stack scanning, and object copy. These must be atomic, i.e. non-preemptible, otherwise they will lead to inconsistency of data due to race conditions. The latter two involve considerable blocking times. The garbage collector must ensure a consistent root set. In incremental GC, this requires the use of barriers, to enforce the consistency of marking.
The start point of tracing is the root set, which is known to be directly accessible. It consists of references to global variables (or static in Java), references that are local to a thread, references in thread runtime stack, and thread-local CPU registers.
To make root scanning preemptible (useful in real-time systems), the premise is that atomicity during root scanning is only necessary with respect to the thread whose stack is currently being scanned, because a thread can modify only its own stack. So the system can delegate local root scans to mutator threads. This results in more efficient code, reducing the disruptiveness of garbage collection.