fix(server): Reorder ExecuteAsync callback seqlock check

[dranikpg]

Currently we re-use transactions in the non-atomic multi mode and the callback in ExecuteAsync looks like the following:

mask = GetMask(shard)
seq_after = Add(seqlock, 0, release)

if seq_after == seq_old and mask & ARMED: Poll()

However it might happen that the transaction is already being re-used when the callback is run, so we should avoid accessing the mask and just quietly drop the callback.

I suggest accessing the mask only when checking the seqlock, which means that the wait barrier didn't finish and all state remained untouched.

seq_after = Load(seqlock, acquire)
if seq_after == seq_old and GetMask(shard) & ARMED: Poll()

I must admit I don't understand your "trick".

we use fetch_add with release trick to make sure that local_mask is loaded before we load seq_after

Why?

We could gain similar result with "atomic_thread_fence(acquire)"

Thats what I did and release-acquire is standard practice. But how is it similar? Its the opposite in fact

[romange]

Yeah, I do not know what I was thinking.