Re: better atomics - v0.6

Oskari Saarenmaa <os@ohmu.fi>

From: Oskari Saarenmaa <os@ohmu.fi>
To: Andres Freund <andres@2ndquadrant.com>, Robert Haas <robertmhaas@gmail.com>, Amit Kapila <amit.kapila16@gmail.com>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, Peter Geoghegan <pg@heroku.com>, Pg Hackers <pgsql-hackers@postgresql.org>, Ants Aasma <ants@cybertec.at>
Date: 2014-09-23T10:50:28Z
Lists: pgsql-hackers

Attachments

23.09.2014, 00:01, Andres Freund kirjoitti:
> I've finally managed to incorporate (all the?) feedback I got for
> 0.5. Imo the current version looks pretty good.
>
> Most notable changes:
> * Lots of comment improvements
> * code moved out of storage/ into port/
> * separated the s_lock.h changes into its own commit
> * merged i386/amd64 into one file
> * fixed lots of the little details Amit noticed
> * fixed lots of XXX/FIXMEs
> * rebased to today's master
> * tested various gcc/msvc versions
> * extended the regression tests
> * ...
>
> The patches:
> 0001: The actual atomics API

I tried building PG on Solaris 10/Sparc using GCC 4.9.0 (buildfarm 
animal dingo) with this patch but regression tests failed due to:

/export/home/os/postgresql/src/test/regress/regress.so: symbol 
pg_write_barrier_impl: referenced symbol not found

which turns out to be caused by a leftover PG_ prefix in ifdefs for 
HAVE_GCC__ATOMIC_INT64_CAS.  Removing the PG_ prefix fixed the build and 
regression tests.  Attached a patch to strip the invalid prefix.

> 0002: Implement s_lock.h support ontop the atomics API. Existing
>        implementations continue to be used unless
>        FORCE_ATOMICS_BASED_SPINLOCKS is defined

Applied this and built PG with and without FORCE_ATOMICS_BASED_SPINLOCKS 
- both builds passed regression tests.

> 0003-0005: Not proposed for review here. Just included because code
>       actually using the atomics make testing them easier.

I'll look at these patches later.

/ Oskari

Commits

  1. Avoid the use of a separate spinlock to protect a LWLock's wait queue.

  2. Improve LWLock scalability.

  3. Reduce the number of semaphores used under --disable-spinlocks.

  4. Alter the configure script to fail immediately if the C compiler does not