On Thu, Dec 26, 2013 at 4:32 PM, Daniel Micay <danielmi...@gmail.com> wrote: > On Thu, Dec 26, 2013 at 4:19 PM, comex <com...@gmail.com> wrote: >> On Thu, Dec 26, 2013 at 8:48 PM, Daniel Micay <danielmi...@gmail.com> wrote: >>> Acquiring and releasing a mutex when there's low contention takes two >>> atomic operations. Switching to a linked data structure adds far more >>> overhead than an extra atomic operation. >>> [...] >>> A simple array-backed circular buffer has a tiny lock scope, so it >>> won't ever have high contention. In a less simple case, a mutex will >>> spin for a bit before falling back to a `futex` system call. >> >> Fair enough, but why do you assume a lock-free queue is a linked data >> structure? A simple (bounded) array-backed circular buffer can be >> lock-free too and then there are zero atomic operations. > > Lock-free data structures use atomic operations... you're doing at > least one atomic operation or you won't have consistency between > threads. An atomic operation will result in cache synchronization and > stall the CPU pipeline. On an architecture with a weaker memory model > (not x86), it might not have to stall the pipeline as much if you use > a weaker ordering.
Anyway, I think there's a misunderstanding in this thread of what locking is. A blocking queue is by definition not lock-free/wait-free, since the wait is known as a lock. It can be done via a spin lock, a system call like `futex` or the mix used by Linux mutexes but it's a lock by any other name. _______________________________________________ Rust-dev mailing list Rust-dev@mozilla.org https://mail.mozilla.org/listinfo/rust-dev
