Feedback-ID: ia2694551:Fastmail
Date: Tue, 19 Nov 2024 09:45:43 +0100
From: Alvaro Herrera <alvherre@alvh.no-ip.org>
To: Will Mortensen <will@extrahop.com>
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>, 
	Yvonne Chen <yvonne@extrahop.com>
Subject: Re: README.tuplock and SHARE lock
Message-ID: <202411190845.qfykjp5l4ijx@alvherre.pgsql>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CAMpnoC6yEQ=c0Rdq-J7uRedrP7Zo9UMp6VZyP23QMT68n06cvA@mail.gmail.com>
Archived-At: <https://www.postgresql.org/message-id/202411190845.qfykjp5l4ijx%40alvherre.pgsql>
Precedence: bulk

On 2024-Nov-18, Will Mortensen wrote:

> README.tuplock says:
> 
> > There is one exception
> > here: since infomask space is limited, we do not provide a separate bit
> > for SELECT FOR SHARE, so we have to use the extended info in a MultiXact in
> > that case.  (The other cases, SELECT FOR UPDATE and SELECT FOR KEY SHARE, are
> > presumably more commonly used due to being the standards-mandated locking
> > mechanism, or heavily used by the RI code, so we want to provide fast paths
> > for those.)
> 
> But looking at the explanations of the infomask bits further down (as
> updated in commit cdbdc4382743fcfabb3437ea7c4d103adaa01324), as well
> as the actual code for locking a not-yet-locked tuple in
> compute_new_xmax_infomask(), this doesn't seem to be true. Was this an
> oversight?

Hmm, yeah, it seems you're correct about this being an oversight -- we
don't necessarily use a multixact if all we want to do is to store a FOR
SHARE lock.  Instead, what we do is mark the tuple with two lock bits,
per this bit in src/include/access/htup_details.h:

#define HEAP_XMAX_SHR_LOCK    (HEAP_XMAX_EXCL_LOCK | HEAP_XMAX_KEYSHR_LOCK)

This can be seen in a WAL_DEBUG build, when doing SELECT FOR SHARE of a
tuple does this:

2024-11-19 09:39:37.011 CET [65326] LOG:  INSERT @ 0/1B92EB0:  - Heap/LOCK: xmax: 744, off: 1, infobits: [LOCK_ONLY, EXCL_LOCK, KEYSHR_LOCK], flags: 0x00

Note that the infobits don't include anything about it being MULTI.
Contrast that with the case where the same tuple is locked by two
transactions simultaneously:

2024-11-19 09:40:45.582 CET [65326] LOG:  INSERT @ 0/1B93008:  - MultiXact/CREATE_ID: 1 offset 1 nmembers 2: 745 (sh) 746 (sh) 
2024-11-19 09:40:45.582 CET [65326] LOG:  INSERT @ 0/1B93040:  - Heap/LOCK: xmax: 1, off: 1, infobits: [IS_MULTI, LOCK_ONLY, EXCL_LOCK, KEYSHR_LOCK], flags: 0x00

Here we first see the MultiXact being created (with two transactions,
both using SHARE mode), then the tuple being locked with IS_MULTI.

-- 
Álvaro Herrera        Breisgau, Deutschland  —  https://www.EnterpriseDB.com/
Subversion to GIT: the shortest path to happiness I've ever heard of
                                                (Alexey Klyukin)