Scaling address-space operations on Linux with TSX

Abstract

Concurrent programming is important due to increasing core counts, but scalable concurrency control is difficult and error-prone to implement. Hardware Transactional Memory (HTM) addresses this problem by providing hardware support for concurrently executing arbitrary read-modify-write memory transactions. Intel released Transactional Synchronization eXtensions (TSX), a HTM implementation, in select processors to support scalable concurrency control. This thesis contributes a case study in applying TSX to the Linux virtual memory system, which currently serializes address-space operations with a lock. TSX should provide scalability by supporting concurrent address-space operations. Achieving scalability with TSX, however, turned out to be difficult due to transactional aborts. This thesis details how to identify and resolve abort problems, and it describes the necessary modifications to make address-space operations scale in Linux. This thesis also describes a new TLB shootdown algorithm, TxShootDown, which removes TLB shootdown from a transactional critical section while avoiding races due to concurrent address-space operations.