Re: [PATCH] Identify LWLocks in tracepoints

On Tue, Apr 13, 2021 at 10:42 PM Craig Ringer
<craig(dot)ringer(at)enterprisedb(dot)com> wrote:
> I'd really love it if a committer could add an explanatory comment or
> two in the area though. I'm happy to draft a comment patch if anyone
> agrees my suggestion is sensible. The key things I needed to know when
> studying the code were:
>
> * A LWLock* is always part of a tranche, but locks within a given
> tranche are not necessarily co-located in memory, cross referenced or
> associated in any way.
> * A LWLock tranche does not track how many LWLocks are in the tranche
> or where they are in memory. It only groups up LWLocks into categories
> and maps the tranche ID to a name.
> * Not all LWLocks are part of the main LWLock array; others can be
> embedded in shmem structs elsewhere, including in DSM segments.
> * LWLocks in DSM segments may not have the same address between
> different backends, because the DSM base address can vary, so a
> LWLock* cannot be reliably compared between backends unless you know
> it's in the main LWLock array or in static shmem.
>
> Having that info in lwlock.c near the tranche management code or the
> tranche and main lwlock arrays would've been very handy.

I'm willing to review a comment patch along those lines.

> I'm actually inclined to revise the patch I sent in order to *remove*
> the LWLock name from the tracepoint argument. At least for the
> fast-path tracepoints on start-of-acquire and end-of-acquire. I think
> it's probably OK to report it in the lock wait tracepoints, which is
> where it's most important to have anyway. So the tracepoint will
> always report the LWLock* and tranche ID, but it won't report the
> tranche name for the fast-path. I'll add trace events for tranche ID
> registration, so trace tools can either remember the tranche ID->name
> mappings from there, or capture them from lock wait events and
> remember them.
>
> Reasonable? That way we retain the most important trace functionality,
> but we reduce the overheads.

Reducing the overheads is good, but I have no opinion on what's
important for people doing tracing, because I am not one of those
people.

--
Robert Haas
EDB: http://www.enterprisedb.com