Обсуждение: Fix possible 'unexpected data beyond EOF' on replica restart

On Tue, Dec 16, 2025 at 6:38 PM Anthonin Bonnefoy
<anthonin.bonnefoy@datadoghq.com> wrote:
>
>  [...]
> The patch fixes the issue by moving smgr_cached_nblocks updates in mdtruncate and only updating the cached value if
truncatewas successful. 
>

Thanks for detailed reproducible steps, I can see the reported issue
and proposed patch fixes the same. Patch looks good to me except
following changes in smgrtruncate():

-       /* Make the cached size is invalid if we encounter an error. */
-       reln->smgr_cached_nblocks[forknum[i]] = InvalidBlockNumber;
-
        smgrsw[reln->smgr_which].smgr_truncate(reln, forknum[i],
                                               old_nblocks[i], nblocks[i]);

The deleted code you moved to mdtruncate() should be kept where it
was. We cannot ensure that every extension using this interface will
adhere to that requirement what the comment describes. Furthermore,
an extension's routine might miss updating smgr_cached_nblocks.
To ensure that updates smgr_cached_nblocks properly, we should also
add an assertion after the call, like this:

        /*
         * Ensure that the local smgr_cached_nblocks value is updated.
         */
        Assert(reln->smgr_cached_nblocks[forknum[i]] != InvalidBlockNumber);

Regards,
Amul

On 17/12/2025 10:40, Anthonin Bonnefoy wrote:
> On Wed, Dec 17, 2025 at 8:26 AM Amul Sul <sulamul@gmail.com 
> <mailto:sulamul@gmail.com>> wrote:
> 
>> The deleted code you moved to mdtruncate() should be kept where it
>> was. We cannot ensure that every extension using this interface will
>> adhere to that requirement what the comment describes. 
> 
> Yeah, I've overlooked the case of extensions. I've moved back the 
> InvalidBlockNumber assignment.

The smgr interface isn't really an extension point, so we don't need to 
worry about that. (I wish it was, but that's a different story [1])

I don't think mdtruncate() should modify smgr_cached_nblocks, we should 
keep that in smgrtruncate(). smgr.c is responsible for all other updates 
of smgr_cached_nblocks, so doing it in mdtruncate() would be a layering 
violation.

I'm thinking that we should do the attached. Untested, and we should 
also add a comment to smgrtruncate() and mdtruncate() to explain how 
they behave if nblocks > curnblk.

I wonder if we should move the whole "if (nblocks > curnblk)" check and 
ereport() from mdtruncate() to smgrtruncate(). That logic doesn't really 
depend on anything specific to md.c. If you'd imagine a different smgr 
implementation, it'd need to just copy-paste that check. It's the 
caller's mistake if it passes nblocks > curnblk, when not in recovery. 
Then again, we do have other places in md.c too that behave differently 
when InRecovery.

What do you think?

> I wonder how critical it is to have an up to date value of 
> smgr_cached_nblocks after smgr_truncate. Leaving InvalidBlockNumber was 
> also an option as the next user will ask the kernel for the real size. 
> There are some functions like DropRelationBuffers which rely only on the 
> cached value, so it's probably safer to keep the same behaviour.

Leaving it invalid should work. But as the comment says, we might as 
well update the cached value since we have the value at hand. It's just 
that we were doing it wrong.

[1] 
https://www.postgresql.org/message-id/CAEze2WgMySu2suO_TLvFyGY3URa4mAx22WeoEicnK%3DPCNWEMrA%40mail.gmail.com

- Heikki

On Thu, Dec 18, 2025 at 3:18 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:
>
> The smgr interface isn't really an extension point, so we don't need to
> worry about that. (I wish it was, but that's a different story [1])

Thanks for the additional context.

> I don't think mdtruncate() should modify smgr_cached_nblocks, we should
> keep that in smgrtruncate(). smgr.c is responsible for all other updates
> of smgr_cached_nblocks, so doing it in mdtruncate() would be a layering
> violation.
>
> I'm thinking that we should do the attached. Untested, and we should
> also add a comment to smgrtruncate() and mdtruncate() to explain how
> they behave if nblocks > curnblk.

This appears to work. The cached size is correctly updated and my
script doesn't trigger the bug anymore. I've updated the patch with
this approach.

> I wonder if we should move the whole "if (nblocks > curnblk)" check and
> ereport() from mdtruncate() to smgrtruncate(). That logic doesn't really
> depend on anything specific to md.c. If you'd imagine a different smgr
> implementation, it'd need to just copy-paste that check. It's the
> caller's mistake if it passes nblocks > curnblk, when not in recovery.
> Then again, we do have other places in md.c too that behave differently
> when InRecovery.
>
> What do you think?

I imagine we will still want some sanity checks in mdtruncate(), or at
least an Assert to make sure the provided block values are correct.
This also assumes that we will only have one caller to mdtruncate(),
another caller will have to duplicate the check.

And in the context of a backpatch, the current approach has the
benefit of minimising the amount of change, so I am slightly partial
to keeping the check in mdtruncate().

Regards,
Anthonin Bonnefoy