Re: WAL record CRC calculated incorrectly because of underlying buffer modification

11.05.2024 07:25, Thomas Munro wrote:
> On Sat, May 11, 2024 at 4:00 PM Alexander Lakhin <exclusion@gmail.com> wrote:
>> 11.05.2024 06:26, Thomas Munro wrote:
>>> Perhaps a no-image, no-change registered buffer should not be
>>> including an image, even for XLR_CHECK_CONSISTENCY?  It's actually
>>> useless for consistency checking too I guess, this issue aside,
>>> because it doesn't change anything so there is nothing to check.
>> Yes, I think something wrong is here. I've reduced the reproducer to:
> Does it reproduce if you do this?
>
> -               include_image = needs_backup || (info &
> XLR_CHECK_CONSISTENCY) != 0;
> +               include_image = needs_backup ||
> +                       ((info & XLR_CHECK_CONSISTENCY) != 0 &&
> +                        (regbuf->flags & REGBUF_NO_CHANGE) == 0);

No, it doesn't (at least with the latter, more targeted reproducer).

> Unfortunately the back branches don't have that new flag from 00d7fb5e
> so, even if this is the right direction (not sure, I don't understand
> this clean registered buffer trick) then ... but wait, why are there
> are no failures like this in the back branches (yet at least)?  Does
> your reproducer work for 16?  I wonder if something relevant changed
> recently, like f56a9def.  CC'ing Michael and Amit K for info.

Maybe it's hard to hit (autovacuum needs to process the index page in a
narrow time frame), but locally I could reproduce the issue even on
ac27c74de(~1 too) from 2018-09-06 (I tried several last commits touching
hash indexes, didn't dig deeper).

Best regards,
Alexander