Re: inefficient loop in StandbyReleaseLockList()

Поиск
Список
Период
Сортировка
От Kyotaro Horiguchi
Тема Re: inefficient loop in StandbyReleaseLockList()
Дата
Msg-id 20211101.114019.1470121542110379535.horikyota.ntt@gmail.com
обсуждение исходный текст
Ответ на Re: inefficient loop in StandbyReleaseLockList()  (Tom Lane <tgl@sss.pgh.pa.us>)
Ответы Re: inefficient loop in StandbyReleaseLockList()  (Tom Lane <tgl@sss.pgh.pa.us>)
Список pgsql-hackers
At Sun, 31 Oct 2021 16:55:01 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote in 
> I wrote:
> > Pushed the patch as given.  I've not yet reviewed other list_delete_first
> > callers, but I'll take a look.  (I seem to remember that I did survey
> > them while writing 1cff1b95a, but I evidently missed that this code
> > could be dealing with a list long enough to be problematic.)
> 
> I looked at the remaining list_delete_first callers.
> 
> 1. Attached is a proposed patch to get rid of the calls in trgm_regexp.c.
> I'm not certain that those lists could get long enough to be a problem,
> given the existing complexity limits in that file (MAX_EXPANDED_STATES
> etc).  But I'm not certain they can't, either, and it's easy enough to
> fix along the same lines as in StandbyReleaseLockList.

I should be missing something, but at the added list_free() there's a
case where keysQueue has some elelments.  I think the remaining
elements are useless but AFAICS all the memory allocated there is
freed after createTrgmNFAInternal returnes, before the "next cycle"
comes. Do we need to add that list_free there?

> 2. I think we almost certainly have a problem in SyncPostCheckpoint.

Maybe we want list_delete_first_n() or such to remove the first n
elements in a list at once.

> 3. Is agg_refill_hash_table a problem?  Probably; we've seen cases
> with lots of batches.

I excluded it since I'm not sure it is in the pattern at a glance. I
would want to leave it alone, since changing the logic there seems
making things a bit complex and the gain by removing list_delete_first
doesn't look so large..

> 4. I'm a bit worried about the uses in access/gist/, but I don't know
> that code well enough to want to mess with it.  It's possible the
> list lengths are bounded by the index tree height, in which case it
> likely doesn't matter.  The logic in gistFindPath looks like a mess
> anyway since it's appending to both ends of the "fifo" list in different
> places (is that really necessary?).

From the other side, the elemnts are inserted by lcons, then removed
by list_delete_first.  It is the worst usage of the current list
implementation as a FIFO. Couldn't we construct and iterate over a
list in the reverse order?

> 5. Not sure about CopyMultiInsertInfoFlush ... how many buffers
> could we have there?

(I'm not sure..)

> 6. llvm_release_context may not have a long enough list to be a
> problem, but on the other hand, it looks easy to fix.

Agreed.

> 7. The list lengths in the parser and dependency.c, ruleutils.c,
> explain.c are bounded by subquery nesting depth or plan tree depth,
> so I doubt it's worth worrying about.

Agreed.

> 8. The uses in namespace.c don't seem like an issue either -- for
> instance, GetOverrideSearchPath can't iterate more than twice,
> and the overrideStack list shouldn't get very deep.

If we didn't need the resulting list I'm for changing it but actually
it is needed. So I think we won't get so much by changing the
function.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center



В списке pgsql-hackers по дате отправления:

Предыдущее
От: Andres Freund
Дата:
Сообщение: Use -fvisibility=hidden for shared libraries
Следующее
От: Michael Paquier
Дата:
Сообщение: Re: Opclass parameters of indexes lost after REINDEX CONCURRENTLY[