Hello Robert,
06.06.2024 19:36, Robert Haas wrote:
On Thu, Jun 6, 2024 at 6:00 AM Alexander Lakhin <exclus...@gmail.com> wrote:
Am I missing something or the the page buffer indeed lacks locking there?
I don't know, but if the locks are really missing now, I feel like the
first question is "which commit got rid of them?". It's a little hard
to believe that they've never been there and somehow nobody has
noticed.
Then again, maybe we have; see Noah's thread about in-place updates
breaking stuff and some of the surprising discoveries there. But it
seems worth investigating.
Yes, my last experiment with memcmp for the whole buffer was wrong,
given the comment above heapgettup_pagemode(). I think the correct
check would be:
ItemId lpp;
OffsetNumber lineoff;
+ItemIdData iid;
lineoff = scan->rs_vistuples[lineindex];
lpp = PageGetItemId(page, lineoff);
+iid = *((ItemIdData *)lpp);
+for (int i = 0; i < 1000; i++)
+Assert(memcmp(&iid, lpp, sizeof(iid)) == 0);
It significantly alleviates reproducing of the test failure for me.
Will try to bisect this anomaly tomorrow.
Best regards,
Alexander
