Horton

Re: Fixing a couple of buglets in how VACUUM sets visibility map bits

Peter Geoghegan <pg@bowt.ie>

View thread View on postgr.es
From: Peter Geoghegan <pg@bowt.ie>
To: Robert Haas <robertmhaas@gmail.com>
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2023-01-10T20:08:27Z
Lists: pgsql-hackers 
On Tue, Jan 10, 2023 at 11:47 AM Peter Geoghegan <pg@bowt.ie> wrote:
> In summary, I think that there is currently no way that we can have
> the VM (or the PD_ALL_VISIBLE flag) concurrently unset, while leaving
> the page all_frozen. It can happen and leave the page all_visible, but
> not all_frozen, due to these very fine details. (Assuming I haven't
> missed another path to the problem with aborted Multis or something,
> but looks like I haven't.)

Actually, FreezeMultiXactId() can fully remove an xmax that has some
member XIDs >= OldestXmin, provided FRM_NOOP processing isn't
possible, at least when no individual member is still running. Doesn't
have to involve transaction aborts at all.

Let me go try to break it that way...

-- 
Peter Geoghegan

Commits

  1. Tighten up VACUUM's approach to setting VM bits.

  2. Make lazy_vacuum_heap_rel match lazy_scan_heap.

  3. vacuumlazy.c: Tweak local variable name.

  4. vacuumlazy.c: Save get_database_name() in vacrel.

  5. Avoid special XID snapshotConflictHorizon values.

  6. Change the format of the VM fork to add a second bit per page.