NeillC

Critical sections are unfair also. Could be differences in spin count you are seeing.

It's typical not to allow a second share acquire to proceed if we have an exclusive waiter: t1: share t2: exclusive so waits t3: share also waits

Thats how we arrive at the rule that shared acquires in anything, but a trivial system (no exclusive acquires) may not be compatible. So, contract #1 is typically not satisfied. Of course, this particular case is slightly different and so you might decide to support it.

The lock in unfair. We are unfair because you get better performance by not allowing the lock hold time to be extended by context swap. As a result, we always unconditionally release a lock when an exclusive acquire is release or the last shared guy releases. Then we go find people to wake. In this gap the lock can be stollen. The threads that steal all look like exclusive guys. My overall rule of thumb is that in the presence of any exclusive acquires you can never assume a shared acquire is compatible with any other shared acquire. A second shared acquire might wait for the first acquire to exit. We do this for example, so a stream of readers don't starve a writer. This is another case like that but somewhat less obvious. I have no idea if the owners will attempt to fix or take the view that you're trying to assume something we don't guarantee. I'll admit that this is a kind of strange case. You could obviously queue a wait block and the in progress waker will sort this out. Could well get a performance impact from that. I tried to explain this general rule of thumb to you this morning when you contacted me. I could not understand much of what you said.

Maybe my memory is faulty but I believe it was analyzed by Leslie Lamports stuff. Ofcourse your building a model of how it should work and you might have faults in that.

We use bit test and set as it causes us to get the cache line exclusive. This avoids using the prefetch write before fetching if we were to use CompareExchange. somebody was trying to talk to me about this stuff this morning but I didn't understand him. You can't expect a reader to always be compatible with other readers otherwise you livelock with a constant stream of readers. So we become incompatible. I am unsure if there is something beyond this.