simplehash: tb->sizemask = 0

Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:

I'm a bit puzzled by this code in SH_COMPUTE_PARAMETERS:

if (tb->size == SH_MAX_SIZE)
tb->sizemask = 0;
else
tb->sizemask = tb->size - 1;

Doesn't that mean that with SH_MAX_SIZE we end up with sizemask being 0
(i.e. no bits set)?

Yeah, which is very obviously broken: for one thing, the Asserts
in SH_NEXT/SH_PREV would surely go off.

(Why are those assertions, anyway, and not test-and-elog?
I do not think an assertion failure is a suitable way to
report "hash table full".)

I don't think we're building hash tables with 2^32 buckets, though.

What this proves is that nobody has ever tested the behavior at
SH_MAX_SIZE. I would suggest building a test version with that
set small enough to be conveniently reachable, and then exercising
the behavior as the limit is approached and reached.

regards, tom lane

On 2017-11-27 22:53:41 -0500, Tom Lane wrote:

Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:

I'm a bit puzzled by this code in SH_COMPUTE_PARAMETERS:

if (tb->size == SH_MAX_SIZE)
tb->sizemask = 0;
else
tb->sizemask = tb->size - 1;

Doesn't that mean that with SH_MAX_SIZE we end up with sizemask being 0
(i.e. no bits set)?

Yeah, which is very obviously broken: for one thing, the Asserts
in SH_NEXT/SH_PREV would surely go off.

That's obviously wrong. Not sure how that happened. I might have had it
as a shift at first?

(Why are those assertions, anyway, and not test-and-elog?
I do not think an assertion failure is a suitable way to
report "hash table full".)

There's a test and elog during insert. Adding actual branches into
SH_NEXT/SH_PREV seems like a bad idea.

Will test a fix.

Greetings,

Andres Freund

On 11/29/17 13:49, Andres Freund wrote:

On 2017-11-27 22:53:41 -0500, Tom Lane wrote:

Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:

I'm a bit puzzled by this code in SH_COMPUTE_PARAMETERS:

if (tb->size == SH_MAX_SIZE)
tb->sizemask = 0;
else
tb->sizemask = tb->size - 1;

Doesn't that mean that with SH_MAX_SIZE we end up with sizemask being 0
(i.e. no bits set)?

Yeah, which is very obviously broken: for one thing, the Asserts
in SH_NEXT/SH_PREV would surely go off.

That's obviously wrong. Not sure how that happened. I might have had it
as a shift at first?

(Why are those assertions, anyway, and not test-and-elog?
I do not think an assertion failure is a suitable way to
report "hash table full".)

There's a test and elog during insert. Adding actual branches into
SH_NEXT/SH_PREV seems like a bad idea.

Will test a fix.

I'll be happy to help test this fix when it's ready.

-- todd