Looks okay to me.

I think the mutex is still needed to reliably update deferred_count. E.g. this interleaving loses the count value.

T1: read and zero count, young_count =non zero
T2: read and zero count, young_count = 0
T1: read deferred_count
T2: read deferred_count
T1: write deferred_count + young_count
T2: write deferred_count + young_count

Well, that's why I suggested a swap and anadd :)

T1: read and zero count, young_count = non zero
T2: read and zero count, young_count = 0
T1: add young_count to deferred_count
T2: add 0 to deferred_count

The two sequential operationsprobably aren't noticeably faster than the locking though.