Re: SSI patch version 14

Jeff Davis wrote:
> On Sat, 2011-02-05 at 14:43 -0600, Kevin Grittner wrote:

>> In working on this I noticed the apparent need to move two calls
>> to PredicateLockTuple a little bit to keep them inside the buffer
>> lock. Without at least a share lock on the buffer, it seems that
>> here is a window where a read could miss the MVCC from a write and
>> the write could fail to see the predicate lock.

> What does PredicateLockTuple do that needs a share lock? Does a pin
> suffice?

If one process is reading a tuple and another is writing it (e.g.,
UPDATE or DELETE) the concern is that we need to be able to guarantee
that either the predicate lock appears in time for the writer to see
it on the call to CheckForSerializableConflictIn or the reader sees
the MVCC changes in CheckForSerializableConflictOut. It's OK if the
conflict is detected both ways, but if it's missed both ways then we
could have a false negative, allowing anomalies to slip through. It
didn't seem to me on review that acquiring the predicate lock after
releasing the shared buffer lock (and just having the pin) would be
enough to ensure that a write which followed that would see the
predicate lock.

reader has pin and shared lock
writer increments pin count
reader releases shared lock
writer acquires exclusive lock
writer checks predicate lock and fails to see one
reader adds predicate lock
we have a problem

-Kevin