Fix space leaks of Michael-Scott queue and avoid the impossible

[polytypic]

This PR fixes the space leak issue of the previous implementation (i.e. issue #63). To properly fix the leaks, it is also necessary to avoid a similar leak through the tail, which could previously also retain a reference to a node containing a value. This also means that the optimization to allow the tail to fall further behind needs more care as multiple push operations could race to update the same fallen behind tail.

One approach taken in this fix is that instead of having head and tail point to a node, they actually point directly to the atomic within a node. This also conveniently avoids the impossible pattern match cases of the previous implementation.

Another approach taken in this fix is to make sure that the tail does not fall behind. After a successful update of the tail, the new implementation verifies that the tail is still the tail and, if not, tries to update the tail again. If the tail is allowed to fall behind, then it can cause a leak as then the tail could be left pointing to nodes that have already been removed from the queue.

[polytypic]

BTW, it seems (unsurprisingly) that avoiding false sharing with the MS queue has a very noticeable effect on performance.

The whole idea of the the MS queue is that the head and tail are two distinct shared memory locations and the push and pop operations only access one of those. Naïvely allocating the head and tail atomics leads directly to high probability of false sharing.

Here is a benchmark run where the implementation avoids false sharing:

Hand-written Lockfree.MSQueue: mean = 0.005300, sd = 0.000000 tp=56601255.915217

Without false sharing avoidance, the timings from the same benchmark vary highly and are sometimes more than twice as high and basically never as good as above.

To be clear, this PR does not include the change to avoid false sharing:

 let create () =
   let next = Atomic.make Nil in
-  { head = Atomic.make next; tail = Atomic.make next }
+  Multicore_magic.copy_as_padded
+    {
+      head = Multicore_magic.copy_as_padded @@ Atomic.make next;
+      tail = Multicore_magic.copy_as_padded @@ Atomic.make next;
+    }

[lyrm]

Thanks, that is a great fix ! I just finished reviewing it and everything looks good.

I will have a look at why the CI checks are failing tomorrow.

[polytypic]

I will have a look at why the CI checks are failing tomorrow.

CI says:

•  # ocamlfind: Package `threads' not found
  

Perhaps related to this?