On Fri, Jan 6, 2017 at 11:32 PM, Amit Kapila <amit.kapil...@gmail.com> wrote:
> On Sat, Jan 7, 2017 at 2:33 AM, Robert Haas <robertmh...@gmail.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);
            _hash_relbuf(rel, buf);
        retain_pin = false;


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

Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to