Horton

Re: Atomic operations within spinlocks

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Robert Haas <robertmhaas@gmail.com>, pgsql-hackers@lists.postgresql.org
Date: 2020-06-06T03:32:32Z
Lists: pgsql-hackers 
Hi,

On 2020-06-05 22:52:47 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2020-06-05 21:01:56 -0400, Tom Lane wrote:
> >> 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.
>
> > Indeed. And any new architecture one would port PG to would have good
> > enough compiler intrinsics to make that trivial. I still think it'd make
> > sense to have a fallback implementation using compiler intrinsics...
>
> > And I think we should just require 32bit atomics at the same time. Would
> > probably kill gaur though.
>
> Not only gaur.  A quick buildfarm survey finds these active members
> reporting not having 32-bit atomics:

Hm, I don't think that's the right test. We have bespoke code to support
most of these, I think:


>  anole         | 2020-06-05 11:20:17 | pgac_cv_gcc_atomic_int32_cas=no

Has support via acc specific intrinsics.


>  chipmunk      | 2020-05-29 22:27:56 | pgac_cv_gcc_atomic_int32_cas=no

Doesn't have support for __atomic, but does have support for 32bit
__sync.


>  gharial       | 2020-06-05 12:41:14 | pgac_cv_gcc_atomic_int32_cas=no

__sync support for both 32 and 64 bit.


>  curculio      | 2020-06-05 22:30:06 | pgac_cv_gcc_atomic_int32_cas=no
>  frogfish      | 2020-05-31 13:00:25 | pgac_cv_gcc_atomic_int32_cas=no

__sync support for both 32 and 64 bit.


>  mandrill      | 2020-06-05 09:20:03 | pgac_cv_gcc_atomic_int32_cas=no

__sync support for 32, as well as as inline asm for 32bit atomics
(although we might be able to add 64 bit).


>  hornet        | 2020-06-05 09:11:26 | pgac_cv_gcc_atomic_int32_cas=no
>  hoverfly      | 2020-06-05 22:06:14 | pgac_cv_gcc_atomic_int32_cas=no

__sync support for both 32 and 64 bit, and we have open coded ppc asm.


>  locust        | 2020-06-05 10:14:29 | pgac_cv_gcc_atomic_int32_cas=no
>  prairiedog    | 2020-06-05 09:55:49 | pgac_cv_gcc_atomic_int32_cas=no

Wee, these don't have __sync? But I think it should be able to use the
asm ppc implementation for 32 bit atomics.



>  gaur          | 2020-05-19 13:33:25 | pgac_cv_gcc_atomic_int32_cas=no

As far as I understand pa-risc doesn't have any atomic instructions
except for TAS.


So I think gaur is really the only one that'd drop.



> It looks to me like this is mostly about compiler support not the
> hardware; that doesn't make it not a problem, though.  (I also
> remain skeptical about the quality of the compiler intrinsics
> on non-mainstream hardware.)

I think that's fair enough for really old platforms, but at least for
gcc / clang I don't think it's a huge concern for newer ones. Even if
not mainstream. For gcc/clang the intrinsics basically back the
C11/C++11 "language level" atomics support. And those are extremely
widely used these days.

Greetings,

Andres Freund

Commits

  1. Convert SpinLock* macros to static inline functions.

  2. Clean up includes of s_lock.h.

  3. Fix deadlock danger when atomic ops are done under spinlock.

  4. Add basic spinlock tests to regression tests.

  5. spinlock emulation: Fix bug when more than INT_MAX spinlocks are initialized.

  6. Avoid potential spinlock in a signal handler as part of global barriers.

  7. Make pg_stat_wal_receiver consistent with the WAL receiver's shmem info

  8. Improve 64bit atomics support.