Horton

Re: Add bump memory context type and use it for tuplesorts

David Rowley <dgrowleyml@gmail.com>

View thread View on postgr.es
From: David Rowley <dgrowleyml@gmail.com>
To: Nathan Bossart <nathandbossart@gmail.com>
Cc: PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>
Date: 2024-02-20T09:46:08Z
Lists: pgsql-hackers 
On Wed, 26 Jul 2023 at 12:11, Nathan Bossart <nathandbossart@gmail.com> wrote:
> I think it'd be okay to steal those bits.  AFAICT it'd complicate the
> macros in memutils_memorychunk.h a bit, but that doesn't seem like such a
> terrible price to pay to allow us to keep avoiding the glibc bit patterns.

I've not adjusted anything here and I've kept the patch using the
>128KB glibc bit pattern.  I think it was a good idea to make our
lives easier if someone came to us with a bug report, but it's not
like the reserved patterns are guaranteed to cover all malloc
implementations.  What's there is just to cover the likely candidates.
I'd like to avoid adding any bit shift instructions in the code that
decodes the hdrmask.

> > +     if (base->sortopt & TUPLESORT_ALLOWBOUNDED)
> > +             tuplen = GetMemoryChunkSpace(tuple);
> > +     else
> > +             tuplen = MAXALIGN(tuple->t_len);
>
> nitpick: I see this repeated in a few places, and I wonder if it might
> deserve a comment.

I ended up adding a macro and a comment in each location that does this.

> I haven't had a chance to try out your benchmark, but I'm hoping to do so
> in the near future.

Great. It would be good to get a 2nd opinion.

David

Commits

  1. Update mmgr's README to mention BumpContext

  2. Push dedicated BumpBlocks to the tail of the blocks list

  3. Improve test coverage in bump.c

  4. Fix incorrect KeeperBlock macro in bump.c

  5. Use bump memory context for tuplesorts

  6. Introduce a bump memory allocator

  7. Enlarge bit-space for MemoryContextMethodID