On Fri, Jul 1, 2016 at 10:20 PM, Amit Kapila <amit.kapila16@gmail.com> wrote:
>>>> Right, something like that. I think Andres actually wants something
>>>> like this in 9.6, and I'm inclined to think it might be a good idea,
>>>> too. I think there should probably be a test for
>>>> all_visible_according_to_vm at the beginning of that so that we don't
>>>> add more visibility map checks for pages where we already know the VM
>>>> bit can't possibly be set.
>>>>
>>>> Other opinions on the concept or the patch?
>>>
>>> +1 on the idea.
>>>
>>> + PageClearAllVisible(page);
>>> + MarkBufferDirty(buf);
>>>
>>> What is the need to clear the Page level bit, if it is already
>>> cleared, doesn't '!all_frozen' indicate that?
>>
>> No, I don't think so. I think all_frozen indicates whether we think
>> that all tuples on the page qualify as fully frozen. I don't think it
>> tells us anything about whether PD_ALL_VISIBLE is set on the page.
>
> Then, can we decide to clear it on that basis? Isn't it possible that
> page is marked as all_visible, even if it contains frozen tuples?In
> the other nearby code (refer below part of code), we only clear the
> page level bit after ensuring it is set. Am I missing something?
>
> else if (PageIsAllVisible(page) && has_dead_tuples)
> {
> elog(WARNING, "page containing dead tuples is marked as all-visible in
> relation \"%s\" page %u",
> relname, blkno);
> PageClearAllVisible(page);
> MarkBufferDirty(buf);
> visibilitymap_clear(onerel, blkno, vmbuffer);
> }
So I'm a bit confused about what you are saying here. If the page is
marked all-frozen but actually isn't all-frozen, then we should clear
the all-frozen bit in the VM. The easiest way to do that is to clear
both bits in the VM plus the page-level bit, as done here, because we
don't presently have a way of clearing just one of the visibility map
bits.
Now, since the heap_lock_tuple issue requires us to introduce a new
method to clear all-visible without clearing all-frozen, we could
possibly use that here too, once we have it.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company