On Sat, Apr 1, 2017 at 1:08 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Mon, Mar 27, 2017 at 5:39 AM, Ashutosh Sharma <ashu.coek88@gmail.com> wrote:
>> Thanks for reporting this problem. Could you please let me know on for
>> how long did you run sqlsmith to get this crash. However, I have found
>> the reason for this crash. This is basically happening when trying to
>> retrieve the tuples using cursor. Basically the current hash index
>> scan work tuple-at-a-time which means once it finds tuple on page, it
>> releases lock from the page but keeps pin on it and finally returns
>> the tuple. When the requested number of tuples are processed there is
>> no lock on the page that was being scanned but yes there is a pin on
>> it. Finally, when trying to close a cursor at the end of scan, if any
>> killed tuples has been identified we try to first mark these items as
>> dead with the help of _hash_kill_items(). But, since we only have pin
>> on this page, the assert check 'LWLockHeldByMe()' fails.
>
> Instead of adding bool haveLock to _hash_kill_items, how about just
> having the caller acquire the lock before calling the function and
> release it afterwards? Since the acquire is at the beginning of the
> function and the release at the end, there seems to be no point in
> putting the acquire/release inside the function rather than in the
> caller.
Well, That is another option but the main idea was to be inline with
the btree code. Moreover, I am not sure if acquiring lwlock inside
hashendscan (basically the place where we are trying to close down the
things) would look good.
--
With Regards,
Ashutosh Sharma
EnterpriseDB:http://www.enterprisedb.com
