Обсуждение: [9.5] possible fast path for pinning a page multiple times

Context:

A patch from a while ago was rejected:

http://www.postgresql.org/message-id/1369886097.23418.0.camel@jdavis

Most of the objection seemed to be that extra page pins might happen in
some circumstances, such as this one mentioned by Heikki:

http://www.postgresql.org/message-id/50FD11C5.1030700@vmware.com

That was a single-threaded case, but did represent additional pins being
acquired, which could add up to contention over the
BufMappingPartitionLock if there is other traffic on that lock
partition.

Idea:

Let's say we have a routine PinBufferTag, that's like PinBuffer but it
takes an additional BufferTag argument. When it locks the buffer header,
it would also compare the argument to the buffer's tag, and if they
don't match, return a status indicating that it's the wrong buffer and
don't pin it. In other words, it pins the buffer only if it's the right
one.

Then, we can just have a backend-local cache that maps BufferTag to
buffer ID. If it's missing an entry, or if the entry is wrong, then it
just proceeds with the normal BufferAlloc path. But if the cache holds
the right value, then we completely bypass the BufMappingPartitionLock
while getting the pin.

Before I do too much performance testing of this, is it a correct
approach? It seems too easy.

Regards,Jeff Davis

Jeff Davis <pgsql@j-davis.com> writes:
> Idea:

> Let's say we have a routine PinBufferTag, that's like PinBuffer but it
> takes an additional BufferTag argument. When it locks the buffer header,
> it would also compare the argument to the buffer's tag, and if they
> don't match, return a status indicating that it's the wrong buffer and
> don't pin it. In other words, it pins the buffer only if it's the right
> one.

> Then, we can just have a backend-local cache that maps BufferTag to
> buffer ID. If it's missing an entry, or if the entry is wrong, then it
> just proceeds with the normal BufferAlloc path. But if the cache holds
> the right value, then we completely bypass the BufMappingPartitionLock
> while getting the pin.

> Before I do too much performance testing of this, is it a correct
> approach? It seems too easy.

Don't see why it wouldn't work.  I think you'd need to test buffer status
not just the contents of the tag field, but that's easy enough.

Bigger issues are how large we'd want to allow the local cache to get
(and what to do to keep it within that) and how much performance we
really gain.  But those are tuning questions not correctness questions.
        regards, tom lane