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: pgsql-hackers@lists.postgresql.org
Date: 2020-06-04T19:06:45Z
Lists: pgsql-hackers 
Hi,

On 2020-06-04 14:50:40 -0400, Tom Lane wrote:
> Actually ... we could probably use this design with a uint32 counter
> as well, on machines where the 64-bit operations would be slow.

On skylake-x even a 32bit [i]div is still 26 cycles. That's more than an
atomic operation 18 cycles.


> 2. The computed completePasses value would go backwards.  I bet
> that wouldn't matter too much either, or at least we could teach
> BgBufferSync to cope.  (I notice the comments therein suggest that
> it is already designed to cope with completePasses wrapping around,
> so maybe nothing needs to be done.)

If we're not concerned about that, then we can remove the
atomic-inside-spinlock, I think. The only reason for that right now is
to avoid assuming a wrong pass number.

I don't think completePasses wrapping around is comparable in frequency
to wrapping around nextVictimBuffer. It's not really worth worrying
about bgwriter wrongly assuming it lapped the clock sweep once ever
UINT32_MAX * NBuffers ticks, but there being a risk every NBuffers seems
worth worrying about.

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.