Abstract

In this paper we present a new form of revocable lock that streamlines the construction of higher level concurrency abstractions such as atomic multi-word heap updates. The key idea is to expose revocation by displacing the previous lock holder’s execution to a safe address. This provides mutual exclusion without needing to block threads. This brings many simplifications, often removing the need for dynamic memory management and letting us strip operations from common-case execution paths. As well as streamlining algorithms’ design, our results show that the technique leads to improved performance and scalability across a range of levels of contention.