Re: BUG #18129: GiST index produces incorrect query results

On 24/09/2023 01:06, Heikki Linnakangas wrote:
> On 22/09/2023 09:00, PG Bug reporting form wrote:
>> First bad commit is 9155580fd, on which three different outcomes are
>> possible: test passes/test fails/assertion fails.
>> Assertion failures, that we had observed after 9155580fd, were fixed by
>> a7ee7c851 and 741b88435, but apparently there is another defect induced
>> by that gist implementation change.
> 
> This is indeed a similar bug to that fixed in 741b88435 and a7ee7c851.
> In those commits, we fixed the insertion code so that whenever we insert
> a downlink to the parent page and that causes the parent page to split,
> we invalidate the 'downlinkoffnum' we had memorized for the child.
> Because when the parent page is split, the downlink for the child likely
> moves.
> 
> However, the downlink can move even when we *don't* split the parent.
> Splitting a page not only inserts the new tuple for the new right page,
> but also updates the old downlink, because the key representing what is
> now the left page covers less keys than the original page. And updating
> a tuple is implemented as delete+insert. The delete moves existing
> tuples on the page! In other words, this code and comment in
> gistfinishsplit():
> 
>>     if (gistinserttuples(state, stack->parent, giststate,
>>                          tuples, 2,
>>                          stack->downlinkoffnum,
>>                          left->buf, right->buf,
>>                          true,    /* Unlock parent */
>>                          unlockbuf    /* Unlock stack->buffer if caller wants
>>                                      * that */
>>                          ))
>>     {
>>         /*
>>          * If the parent page was split, the downlink might have moved.
>>          */
>>         stack->downlinkoffnum = InvalidOffsetNumber;
>>     }
> 
> is wrong. We must invalidate 'downlinkoffnum' whether or not the parent
> page was split, because updating an existing tuple can also move other
> tuples on the page.
> 
> Attached is a set of patches to fix this:
> 
> 1. To catch the problem earlier - already during the index build - I
> modified getFindCorrectParent() to always find the parent the hard way
> by scanning the parent page, even if 'downlinkoffnum' is set. Once the
> downlink is found, it compares it with the 'downlinkoffnum', and PANICs
> if it wasn't where we expected. That defeats the point of memorizing the
> 'downlinkoffnum' in the first place, but it's a useful cross-check.
> 
> I don't intend to commit this, but it was useful during debugging.
> 
> 2. A straightforward fix for the bug.
> 
> 3. Looking at gistFindCorrectParent, when it steps right, it also
> doesn't update 'downlinkoffnum' of the parent page. Surely if we step to
> a different block altogether, the memorized downlink position of the
> previous block no longer applies.
> 
> Worryingly, I haven't been able to get a test failure caused by that.
> Maybe there are some reasons why we never need to split the parent after
> that, not sure. But it sure looks wrong.
> 
> 4. Given how many bugs we've already had with this, I'd like to make
> this more robust. Currently, we meticulously track if we have made any
> inserts that invalidate the memorized location of the downlink. Instead,
> gistFindCorrectParent() could treat the memorized location as just a
> hint, and always check if the downlink is still at the memorized
> location. If it is, great, and if it's not, find it the hard way. That
> makes it unnecessary to clear 'downlinkoffnum' at the right places. We
> still need the 'retry_from_parent' flag, though.
> 
> 5. Re-indent the code after previous commit. Separated for easier review.
> 
> I'd also like to add something to our regression test suite for better
> coverage of this. The intarray test case is very useful, but it'd be
> nice to have something in the main regression suite, as this is a
> general GiST issue, not related to intarray. I haven't spent any time on
> that yet.
I spent a while trying to create a test case for this that would not 
require expanding the test data so much, but no avail. I even tried to 
take the same data set, but instead of duplicating each element like you 
did, I appended a random number of integers to each array, but even that 
did not trigger the failure. That particular data set seems cursed; how 
did you stumble upon it?

I'm reluctant to just make all the arrays larger, as it makes the test 
2x slower. So instead of running all the commands over the larger data 
set, I added one more copy of the CREATE INDEX and the test queries to 
the end that uses the larger set.

See attached. It's squashed version of the previous patches I posted, 
with the test case. Barring objections, I'll commit this.


While I was playing with this, I ran into this unrelated issue:

CREATE EXTENSION pgcrypto; -- for gen_random_bytes
CREATE EXTENSION btree_gist;

CREATE TABLE byteatest (a bytea);
INSERT INTO byteatest SELECT (gen_random_bytes(1000) || 
gen_random_bytes(1000) || gen_random_bytes(1000)) FROM 
generate_series(1, 10) g;
CREATE INDEX ON byteatest USING GIST ( a );

Fails with:

ERROR:  failed to add item to index page in "byteatest_a_idx1"

If you make that value even larger, then it fails with a better error 
message:

ERROR:  index row requires 8208 bytes, maximum size is 8191

So the "failed to add item" error doesn't seem expected.

-- 
Heikki Linnakangas
Neon (https://neon.tech)