Parallel Computer Architecture

5.5 Synchronization 339

to return whether the atomic sequence was executed successfully or not. One way to

provide this atomicity for user processes is to place the lock routine in the operating

system and access it through a system call, but this is expensive and leaves the ques-

tion of how the locks are supported by the operating system

itself.

Another option is

to utilize a hardware lock around the instruction sequence for the lock routine, but

this requires hardware locks and tends to be slow on modern processors.

An efficient, general-purpose solution to the lock problem is to support an atomic

read-modify-write instruction in the processor' ...