Re: Interruptible sleeps (was Re: CommitFest 2009-07: Yay, Kevin! Thanks, reviewers!)

On 12/09/10 20:13, Jeff Davis wrote:
> On Sun, 2010-09-12 at 12:29 -0400, Tom Lane wrote:
>>> ... why throw an ERROR there if it can't happen (or
>>> indicates an inconsistent state when it does happen)?
>>
>> Are you suggesting that an Assert would be sufficient?
>
> I'm not too picky about whether it's Assert, ERROR, or PANIC (Asserts
> aren't available in production systems, which I assume is why elog was
> used);

Right, OwnLatch is a not performance-critical, so it's better to elog() 
IMHO.

> but we should be consistent and document that:
>   (a) it shouldn't happen
>   (b) that it's just a sanity check and we're ignoring the race

Would this be sufficient?

--- a/src/backend/port/unix_latch.c
+++ b/src/backend/port/unix_latch.c
@@ -156,6 +156,7 @@ OwnLatch(volatile Latch *latch)
  	if (selfpipe_readfd == -1)
  		initSelfPipe();

+	/* sanity check */
  	if (latch->owner_pid != 0)
  		elog(ERROR, "latch already owned");
  	latch->owner_pid = MyProcPid;

Or you want to suggest something better?

-- 
   Heikki Linnakangas
   EnterpriseDB   http://www.enterprisedb.com