Re: recent deadlock regression test failures

... BTW, one other minor coding suggestion for
GetSafeSnapshotBlockingPids(): it might be better to avoid doing so much
palloc work while holding the SerializableXactHashLock.  Even if it's
only held shared, I imagine that it's a contention bottleneck.  You
could avoid that by returning an array rather than a list; the array
could be preallocated of size MaxBackends before ever taking the lock.
That would be a little bit space-wasteful, but since it's only short-lived
storage it doesn't seem like much of a problem.

			regards, tom lane