Re: Second call for platform testing

Mark Knox <segfault(at)hardline(dot)org> writes:
>> Had a look at 7.2b3 and sadly it's failing several tests. I saw several
>> "ERROR: PGSTAT: Creation of DB hash table failed" which I haven't seen before.

That error is coming from the following ugly coding:

*dbhash = hash_create("Databases hash", PGSTAT_DB_HASH_SIZE, &hash_ctl,
HASH_ELEM | HASH_FUNCTION | mcxt_flags);
if (pgStatDBHash == NULL)
/* raise error */

AFAICT dbhash always points at the static variable pgStatDBHash, so the
code is not quite incorrect, though it's certainly trouble waiting to
happen as soon as someone changes things so that dbhash might point
elsewhere. What I'm wondering is if your compiler is missing the
potential for aliasing and is emitting code that loads pgStatDBHash
before the store through dbhash occurs. Does it help if you change
the second line (line 2094 in src/backend/postmaster/pgstat.c) to:

if (*dbhash == NULL)

I'm going to commit this change in CVS anyway, but I'm wondering if it
explains your problem or not.

regards, tom lane