on Mon Aug 01 2011, Gordon Woodhull <gordon-AT-woodhull.com> wrote:
On Aug 1, 2011, at 10:49 AM, Dave Abrahams wrote:
on Mon Aug 01 2011, Gordon Woodhull <gordon-AT-woodhull.com> wrote:
All of the classic lock-free algorithms, including those that are the basis for the synchronization primitives we should usually use, have data races.
IIUC, accesses to atomics are not in and of themselves data races. But it doesn't surprise me a bit to learn (again) that lock-free algorithms are hard to think about.
This is the part of the theory that seems like a cop-out to me. You can declare certain variables to be atomic/for synchronization and then *by definition* it's not a data race? Excuse me? No. It's still a race, but it's one that's condoned because you've told the implementation that you want it to be there.
It's not really different than locking. If you want to write to shared data, you need some way of making it not-a-race. It's just that when the data structure is small enough (like an int) you can make it atomic instead of putting a lock around it. -- Dave Abrahams BoostPro Computing http://www.boostpro.com