From: Tom Lane <tgl@sss.pgh.pa.us>
To: Robert Haas <robertmhaas@gmail.com>
cc: "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Subject: Re: shm_toc_lookup API
In-reply-to: 
 <CA+Tgmoa2tpKPO0+qAyHT_dLFhTfqTzkxzDfFE81VJbSdwPTkqQ@mail.gmail.com>
References: <9697.1496675981@sss.pgh.pa.us>
 <CA+TgmoZ_GAT21HLZsw=whXaj3uN_=bez420cRLP0149RPtvS_Q@mail.gmail.com>
 <16984.1496679541@sss.pgh.pa.us>
 <CA+Tgmoa2tpKPO0+qAyHT_dLFhTfqTzkxzDfFE81VJbSdwPTkqQ@mail.gmail.com>
Comments: In-reply-to Robert Haas <robertmhaas@gmail.com>
	message dated "Mon, 05 Jun 2017 13:10:44 -0400"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <28313.1496689030.1@sss.pgh.pa.us>
Date: Mon, 05 Jun 2017 14:57:10 -0400
Message-ID: <28314.1496689030@sss.pgh.pa.us>
Precedence: bulk
Sender: pgsql-hackers-owner@postgresql.org

Robert Haas <robertmhaas@gmail.com> writes:
> On Mon, Jun 5, 2017 at 12:19 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> In practice it probably can't fail even if 64-bit reads aren't atomic,
>> simply because we'll never have enough entries in a shm_toc to make the
>> high-order half ever change.  But that just begs the question why the
>> field is declared Size rather than int.  I think we should make it the
>> latter.

> Yeah.  I think a shm_toc with more than 2^10 entries would probably
> perform badly enough that somebody would rewrite this entire module,
> so we don't really need to worry about having more than 2^31.
> Changing to int (or uint32) seems fine.

Done with uint32.

>> I am also thinking that most of the shm_toc functions need to have the
>> toc pointers declared as "volatile *", but particularly shm_toc_lookup.

(actually, they do already use volatile pointers, except for shm_toc_lookup)

>> That read_barrier call might prevent the hardware from reordering
>> accesses, but I don't think it stops the compiler from doing so.

> If it doesn't prevent both the hardware and the compiler from
> reordering, it's broken.  See the comments for pg_read_barrier() in
> atomics.h.

Meh.  Without volatile, I think that the compiler would be within its
rights to elide the nentry local variable and re-fetch toc->toc_nentry
each time through the loop.  It'd be unlikely to do so, granted, but
I'm not convinced that pg_read_barrier() would prevent that.

However, as long as the write barrier in shm_toc_insert does what it's
supposed to, I think we'd be safe even if that happened.  So probably
it's a moot point.

			regards, tom lane


-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers