Re: _hash_addovflpage has a bug

On Mon, Jan 9, 2017 at 11:59 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Fri, Jan 6, 2017 at 11:32 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>> On Sat, Jan 7, 2017 at 2:33 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>>> It looks to to me like the recent hash index changes have left
>>> _hash_addovflpage slightly broken. I think that if that function
>>> reaches the point where it calls _hash_getbuf() to fetch the next page
>>> in the bucket chain, we also need to clear retain_pin. Otherwise,
>>> we'll erroneously think that we're supposed to retain a pin even when
>>> the current page is an overflow page rather than the primary bucket
>>> page, which is wrong.
>>>
>>
>> How? I think we are ensuring that we retain pin only if it is a
>> bucket page. The check is as below:
>> if ((pageopaque->hasho_flag & LH_BUCKET_PAGE) && retain_pin)
>
> Oh, right. So I guess there's no bug, but I still think that's pretty
> ugly. How about:
>
> if (retain_pin)
> LockBuffer(buf, BUFFER_LOCK_UNLOCK);
> else
> _hash_relbuf(rel, buf);
> retain_pin = false;
>
> instead?
>

Yeah, we can write code that way, but then it is better to rely just
on retain_pin variable in the function and add an Assert for bucket
page whenever we are retaining pin. How about doing something like
attached patch?

>>> A larger question, I suppose, is why this function wants to be certain
>>> of adding a new page even if someone else has already done so. It
>>> seems like it might be smarter for it to just return the newly added
>>> page to the caller and let the caller try to use it. _hash_doinsert
>>> has an Assert() that the tuple fits on the returned page, but that
>>> could be deleted. As it stands, if two backends try to insert a tuple
>>> into the same full page at the same time, both of them will add an
>>> overflow page and we'll end up with 2 overflow pages each containing 1
>>> tuple.
>>
>> I think it is because in the current algorithm we first get an
>> overflow page and then add it to the end. Now we can change it such
>> that first, we acquire the lock on the tail page, check if we still
>> need an overflow page, if so, then proceed, else return the already
>> added overflow page.
>
> For the WAL patch, this is all going to need to be atomic anyway,
> right? Like, you have to allocate the overflow page and add it to the
> bucket chain as a single logged operation?

Yes.

> Not sure exactly how that
> works out in terms of locking.
>

We have to change the locking order as mentioned above by me. This
change is already present in that patch, so maybe we add the check as
suggested by you along with that patch. Now, another thing we could
do is to extract those changes from WAL patch, but I am not sure if it
is worth the effort.

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com