From: Tom Lane <tgl@sss.pgh.pa.us>
To: Andres Freund <andres@anarazel.de>
cc: Robert Haas <robertmhaas@gmail.com>, pgsql-hackers@lists.postgresql.org
Subject: Re: Atomic operations within spinlocks
In-reply-to: <20200606001926.zin5getvvhqppnm2@alap3.anarazel.de>
References: <1141819.1591208385@sss.pgh.pa.us>
 <20200603204542.reuladbmgmjqzoi6@alap3.anarazel.de>
 <1253593.1591293439@sss.pgh.pa.us>
 <20200604185525.wuyjtkluppj4eddr@alap3.anarazel.de>
 <1256701.1591297654@sss.pgh.pa.us>
 <20200605023302.g6v3ydozy5txifji@alap3.anarazel.de>
 <20200606001926.zin5getvvhqppnm2@alap3.anarazel.de>
Comments: In-reply-to Andres Freund <andres@anarazel.de>
	message dated "Fri, 05 Jun 2020 17:19:26 -0700"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <1642803.1591405316.1@sss.pgh.pa.us>
Date: Fri, 05 Jun 2020 21:01:56 -0400
Message-ID: <1642804.1591405316@sss.pgh.pa.us>
Precedence: bulk

Andres Freund <andres@anarazel.de> writes:
> I wrote a patch for this, and when I got around to to testing it, I
> found that our tests currently don't pass when using both
> --disable-spinlocks and --disable-atomics. Turns out to not be related
> to the issue above, but the global barrier support added in 13.
> That *reads* two 64 bit atomics in a signal handler. Which is normally
> fine, but not at all cool when atomics (or just 64 bit atomics) are
> backed by spinlocks. Because we can "self interrupt" while already
> holding the spinlock.

This is the sort of weird platform-specific problem that I'd prefer to
avoid by minimizing our expectations of what spinlocks can be used for.

> I'm not really sure what to do about that issue. The easisest thing
> would probably be to change the barrier generation to 32bit (which
> doesn't have to use locks for reads in any situation).

Yeah, I think we need a hard rule that you can't use a spinlock in
an interrupt handler --- which means no atomics that don't have
non-spinlock implementations on every platform.

At some point I think we'll have to give up --disable-spinlocks;
it's really of pretty marginal use (how often does anyone port PG
to a new CPU type?) and the number of weird interactions it adds
in this area seems like more than it's worth.  But of course
requiring 64-bit atomics is still a step too far.

> Randomly noticed while looking at the code:
> 	uint64		flagbit = UINT64CONST(1) << (uint64) type;

I'm surprised we didn't get any compiler warnings about that.

			regards, tom lane