BUG #16577: Segfault on altering a table located in a dropped tablespace

On Sun, Aug 09, 2020 at 11:00:01AM +0000, PG Bug reporting form wrote:

When executing the following query (a modified excerpt from the tablespace
regression test):
CREATE TABLESPACE regress_tblspace LOCATION '@testtablespace@';
CREATE TABLE test_default_tab_p(id bigint, val bigint)
PARTITION BY LIST (id) TABLESPACE regress_tblspace;
CREATE INDEX test_index2 on test_default_tab_p (val) TABLESPACE
regress_tblspace;
DROP TABLESPACE regress_tblspace;
\d+ test_default_tab_p;
ALTER TABLE test_default_tab_p ALTER val TYPE bigint;

Thanks Alexander for the report. Interesting case indeed.
For a normal table we would complain that the tablespace is not empty
when attempting to drop the tablespace. But here we have only one
partitioned table still holding references to the tablespace. Things
get even more spicy with stuff like that, once you try to play with
the orphaned tablespace reference:
CREATE TABLESPACE popo location '/tmp/popo';
CREATE TABLE parent_tab (a int) partition by list (a) tablespace popo;
DROP TABLESPACE popo;
CREATE TABLE child_tab partition of parent_tab for values in (1);
ERROR: 58P01: could not create directory
"pg_tblspc/24587/PG_12_201909212/16384": No such file or directory
LOCATION: TablespaceCreateDbspace, tablespace.c:161

The issue with indexes is present since 11, but we have more as
reltablespace gets also set for partitioned tables since 12.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

Thanks Alexander for the report. Interesting case indeed.
For a normal table we would complain that the tablespace is not empty
when attempting to drop the tablespace. But here we have only one
partitioned table still holding references to the tablespace.

Yeah, this seems like a mess. DROP TABLESPACE supposes that it can
drop the tablespace if there are no physical files in it, and it's
really hard to see how it could test any more carefully given that
it cannot see what is in pg_class of other databases.

Offhand it seems like we could either

1. Start creating an empty physical file for each partitioned table
or index.

2. Forget the idea that a partitioned table/index has an associated
tablespace.

Neither of these are terribly attractive. But I notice that we
already backed off the idea that this is a thing to some extent:

regression=# CREATE TABLE test_default_tab_p(id bigint, val bigint)
PARTITION BY LIST (id) TABLESPACE pg_default;
ERROR: cannot specify default tablespace for partitioned relations

I'm a bit inclined to think that this "feature" is sufficiently
broken that we should just drop it.

regards, tom lane

On 2020-Aug-09, Tom Lane wrote:

Michael Paquier <michael@paquier.xyz> writes:

Thanks Alexander for the report. Interesting case indeed.
For a normal table we would complain that the tablespace is not empty
when attempting to drop the tablespace. But here we have only one
partitioned table still holding references to the tablespace.

Ah, so it turns out that the physical files were necessary after all.
Maybe the solution to this problem is indeed to have them. It means
partly reverting this commit:

commit 807ae415c54628ade937cb209f0fc9913e6b0cf5
Author: Alvaro Herrera <alvherre@alvh.no-ip.org>
AuthorDate: Fri Jan 4 14:51:17 2019 -0300
CommitDate: Fri Jan 4 14:51:17 2019 -0300

Don't create relfilenode for relations without storage

Some relation kinds had relfilenode set to some non-zero value, but
apparently the actual files did not really exist because creation was
prevented elsewhere. Get rid of the phony pg_class.relfilenode values.

Catversion bumped, but only because the sanity_test check will fail if
run in a system initdb'd with the previous version.

Reviewed-by: Kyotaro HORIGUCHI, Michael Paquier
Discussion: /messages/by-id/20181206215552.fm2ypuxq6nhpwjuc@alvherre.pgsql

But also fixing whatever *other* code was preventing creating of the
filenodes.

As for the crash at hand, it seems it can be solved easily by making
ruleutils avoid trying to dereference a null pointer, as in the attached
patch. This seems necessary for branches 12 and 13, which have the
above commit, but not for 10/11 (which do have relfilenodes) nor master,
where we can revert it.

--
�lvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Alvaro Herrera <alvherre@2ndquadrant.com> writes:

On 2020-Aug-09, Tom Lane wrote:

Michael Paquier <michael@paquier.xyz> writes:

For a normal table we would complain that the tablespace is not empty
when attempting to drop the tablespace. But here we have only one
partitioned table still holding references to the tablespace.

Ah, so it turns out that the physical files were necessary after all.
Maybe the solution to this problem is indeed to have them. It means
partly reverting this commit:

I think actually the hardest part will be figuring out what is the
conversion path, e.g. will pg_upgrade have to know about this.

One point that strikes me is that that will put us in a place where
"does this relation have storage" is not a binary choice. The possible
answers will be "yes", "no", or "has an empty stub file". If we try
to take shortcuts rather than handling that honestly, we'll be in for
more bugs.

As for the crash at hand, it seems it can be solved easily by making
ruleutils avoid trying to dereference a null pointer, as in the attached
patch.

Meh. I have a feeling that that's just the tip of the iceberg of
things that will go wrong in this scenario. I'm not sure how much
band-aid code we want to expend on the case --- after all, tablespace
create/drop is a superuser-only activity, so people aren't doing it
carelessly (one hopes).

regards, tom lane

On Mon, Aug 10, 2020 at 02:59:26PM -0400, Tom Lane wrote:

I think actually the hardest part will be figuring out what is the
conversion path, e.g. will pg_upgrade have to know about this.

One point that strikes me is that that will put us in a place where
"does this relation have storage" is not a binary choice. The possible
answers will be "yes", "no", or "has an empty stub file". If we try
to take shortcuts rather than handling that honestly, we'll be in for
more bugs.

Hmm. Creating a file for partitioned table would be a completely new
thing as well. heap_create() has never created a file for partitioned
tables since 10 so this could open to a new class of bugs.

As for the crash at hand, it seems it can be solved easily by making
ruleutils avoid trying to dereference a null pointer, as in the attached
patch.

Meh. I have a feeling that that's just the tip of the iceberg of
things that will go wrong in this scenario. I'm not sure how much
band-aid code we want to expend on the case --- after all, tablespace
create/drop is a superuser-only activity, so people aren't doing it
carelessly (one hopes).

+1.  I would suspect that it is not the only code path that would run
into issues.
--
Michael

On Tue, Aug 11, 2020 at 04:04:07PM +0900, Michael Paquier wrote:

Hmm. Creating a file for partitioned table would be a completely new
thing as well. heap_create() has never created a file for partitioned
tables since 10 so this could open to a new class of bugs.

This thread has stalled for a couple of weeks now, and I would tend to
take the path where we'd basically revert 8725958 and ca41030. That's
too late for v13 to do anything about that. But not for 14. Any
opinions?
--
Michael

On 2020-Sep-08, Michael Paquier wrote:

On Tue, Aug 11, 2020 at 04:04:07PM +0900, Michael Paquier wrote:

Hmm. Creating a file for partitioned table would be a completely new
thing as well. heap_create() has never created a file for partitioned
tables since 10 so this could open to a new class of bugs.

This thread has stalled for a couple of weeks now, and I would tend to
take the path where we'd basically revert 8725958 and ca41030. That's
too late for v13 to do anything about that. But not for 14. Any
opinions?

Well, naturally I oppose this idea.

--
�lvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Tue, Sep 08, 2020 at 06:37:29PM -0300, Alvaro Herrera wrote:

On 2020-Sep-08, Michael Paquier wrote:

On Tue, Aug 11, 2020 at 04:04:07PM +0900, Michael Paquier wrote:

Hmm. Creating a file for partitioned table would be a completely new
thing as well. heap_create() has never created a file for partitioned
tables since 10 so this could open to a new class of bugs.

This thread has stalled for a couple of weeks now, and I would tend to
take the path where we'd basically revert 8725958 and ca41030. That's
too late for v13 to do anything about that. But not for 14. Any
opinions?

Well, naturally I oppose this idea.

Would it actually solve the issue? ISTM we'd still have to expect cases
with partitioned tables without storage, so presumably we'd have to do
something else ...

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Wed, Sep 09, 2020 at 01:55:53AM +0200, Tomas Vondra wrote:

Would it actually solve the issue? ISTM we'd still have to expect cases
with partitioned tables without storage, so presumably we'd have to do
something else ...

I am not sure what you mean here. If we don't keep anymore references
to tablespace OIDs in pg_class for partitioned tables, meaning that we
don't leave anything dangling if the tablespace is dropped without any
files in its location, how could that be a problem?
--
Michael

On 2020-Sep-09, Tomas Vondra wrote:

On Tue, Sep 08, 2020 at 06:37:29PM -0300, Alvaro Herrera wrote:

On 2020-Sep-08, Michael Paquier wrote:

On Tue, Aug 11, 2020 at 04:04:07PM +0900, Michael Paquier wrote:

Hmm. Creating a file for partitioned table would be a completely new
thing as well. heap_create() has never created a file for partitioned
tables since 10 so this could open to a new class of bugs.

This thread has stalled for a couple of weeks now, and I would tend to
take the path where we'd basically revert 8725958 and ca41030. That's
too late for v13 to do anything about that. But not for 14. Any
opinions?

Well, naturally I oppose this idea.

Would it actually solve the issue? ISTM we'd still have to expect cases
with partitioned tables without storage, so presumably we'd have to do
something else ...

It just dawned on me that a way to fix this is to use a pg_shdepend
entry to protect the tablespace from being dropped.

On Thu, Oct 15, 2020 at 12:19:59PM -0300, Alvaro Herrera wrote:

It just dawned on me that a way to fix this is to use a pg_shdepend
entry to protect the tablespace from being dropped.

Haven't thought of that approach, good idea! That would not be
backpatchable but that would be a solution that does not require
creating files where we don't need them. Did you begin to look at
that?
--
Michael

On 2020-Oct-28, Michael Paquier wrote:

On Thu, Oct 15, 2020 at 12:19:59PM -0300, Alvaro Herrera wrote:

It just dawned on me that a way to fix this is to use a pg_shdepend
entry to protect the tablespace from being dropped.

Haven't thought of that approach, good idea! That would not be
backpatchable but that would be a solution that does not require
creating files where we don't need them. Did you begin to look at
that?

I haven't started on this one yet, but I intend to do so shortly.

Strictly speaking, we can still introduce a new category of pg_shdepend
entries in back branches; it won't break anything that works today. And
while it won't fix the problem on existing partitioned tables, it is
possible to have users run a query on each database to create any rows
needed. The only *actual* incompatibility is that once you upgrade and
create the rows, an older server of the same major may misbehave when
does rows are present. But do we ever see users going back to previous
minors? I think this isn't really a terrible problem in practice. And
even if it is, users can work around it by deleting the offending rows.

What do you think?

Alvaro Herrera <alvherre@alvh.no-ip.org> writes:

On 2020-Oct-28, Michael Paquier wrote:

Haven't thought of that approach, good idea! That would not be
backpatchable but that would be a solution that does not require
creating files where we don't need them. Did you begin to look at
that?

I haven't started on this one yet, but I intend to do so shortly.

Strictly speaking, we can still introduce a new category of pg_shdepend
entries in back branches; it won't break anything that works today.

Yeah, as long as the patched version won't actively fail when those
pg_shdepend entries are missing, I don't think a backpatch is too
hazardous.  It might be worth checking that the extra entries don't
create huge problems if one does downgrade after some of them exist
--- but my feeling for how that mechanism works is that it'd Just
Work, and indeed provide the missing DROP protection even without
explicit action by the backend.

I would not be too excited about offering instructions for people
to manually add/remove the dependency entries. The amount of
value added, versus the risks of bollixing things completely,
doesn't sound like a good tradeoff.

regards, tom lane

On Wed, Oct 28, 2020 at 09:59:24AM -0400, Tom Lane wrote:

Alvaro Herrera <alvherre@alvh.no-ip.org> writes:

I haven't started on this one yet, but I intend to do so shortly.

Cool. Glad to hear that.

Strictly speaking, we can still introduce a new category of pg_shdepend
entries in back branches; it won't break anything that works today.

Yeah, as long as the patched version won't actively fail when those
pg_shdepend entries are missing, I don't think a backpatch is too
hazardous.  It might be worth checking that the extra entries don't
create huge problems if one does downgrade after some of them exist
--- but my feeling for how that mechanism works is that it'd Just
Work, and indeed provide the missing DROP protection even without
explicit action by the backend.

Good point. Even with that, we have never considered the case of a
downgrade as something actually supported, right? We had in the past
for example bug fixes that involve slight tweaks in the WAL records
that are upward-compatible, changing the way these get interpreted
when replayed, but it could be possible to finish with logical
corruptions or replay failures if a downgraded version replays a
slightly-modified record generated by a newer version, no?
--
Michael

On 2020-Oct-15, Alvaro Herrera wrote:

It just dawned on me that a way to fix this is to use a pg_shdepend
entry to protect the tablespace from being dropped.

Here's a proposed patch for this.

--
�lvaro Herrera

On Mon, Jan 11, 2021 at 08:42:44PM -0300, Alvaro Herrera wrote:

Here's a proposed patch for this.

Thanks!

+       /*
+        * If a tablespace is specified, removal of that tablespace is normally
+        * protected by the existence of a physical file; but for relations with
+        * no files, add a pg_shdepend entry to account for that.
+        */
+       if (!create_storage && reltablespace != InvalidOid)
On HEAD, could we consider using this dependency link even for
relations that have physical files, and remove the physical file
check?  Could that make the dependency handling cleaner?
--
Michael

On 2021-Jan-12, Michael Paquier wrote:

+       /*
+        * If a tablespace is specified, removal of that tablespace is normally
+        * protected by the existence of a physical file; but for relations with
+        * no files, add a pg_shdepend entry to account for that.
+        */
+       if (!create_storage && reltablespace != InvalidOid)

On HEAD, could we consider using this dependency link even for
relations that have physical files, and remove the physical file
check? Could that make the dependency handling cleaner?

That would bloat the catalog with a lot of entries for stuff that can be
detected with the current method. Did you notice that the code is
removing an "#ifdef NOT_USED" line to enable existing code? Well, when
I wrote this code in 2005 (59d1b3d99e69) it was doing things as you
suggest, but in the end we decided that it wasn't necessary so it was
taken out.

--
�lvaro Herrera 39�49'30"S 73�17'W
"Someone said that it is at least an order of magnitude more work to do
production software than a prototype. I think he is wrong by at least
an order of magnitude." (Brian Kernighan)

On Tue, Jan 12, 2021 at 12:08:04PM -0300, Alvaro Herrera wrote:

That would bloat the catalog with a lot of entries for stuff that can be
detected with the current method. Did you notice that the code is
removing an "#ifdef NOT_USED" line to enable existing code? Well, when
I wrote this code in 2005 (59d1b3d99e69) it was doing things as you
suggest, but in the end we decided that it wasn't necessary so it was
taken out.

I have not found a thread discussing that around the date of this
commit, but I'll take your word on that.

I just have one small comment.

+   if (!create_storage && reltablespace != InvalidOid)
+       recordDependencyOnTablespace(RelationRelationId, relid,
+                                    reltablespace);
For now we assume that this code path is taken only for partitioned
tables or indexes per the logic in heap_create().  Perhaps it would be
better to add to this code path, or to recordDependencyOnTablespace()
an assertion to check that only the supported relkinds register this
dependency?  If a new relkind is added, it would be easy to miss that
this shared dependency may need to be supported.
--
Michael

On 2021-Jan-13, Michael Paquier wrote:

On Tue, Jan 12, 2021 at 12:08:04PM -0300, Alvaro Herrera wrote:

That would bloat the catalog with a lot of entries for stuff that can be
detected with the current method. Did you notice that the code is
removing an "#ifdef NOT_USED" line to enable existing code? Well, when
I wrote this code in 2005 (59d1b3d99e69) it was doing things as you
suggest, but in the end we decided that it wasn't necessary so it was
taken out.

I have not found a thread discussing that around the date of this
commit, but I'll take your word on that.

I bet you didn't search pgsql-patches ;-)
/messages/by-id/20050703051522.GA13207@surnet.cl

I just have one small comment.

+   if (!create_storage && reltablespace != InvalidOid)
+       recordDependencyOnTablespace(RelationRelationId, relid,
+                                    reltablespace);
For now we assume that this code path is taken only for partitioned
tables or indexes per the logic in heap_create().  Perhaps it would be
better to add to this code path, or to recordDependencyOnTablespace()
an assertion to check that only the supported relkinds register this
dependency?  If a new relkind is added, it would be easy to miss that
this shared dependency may need to be supported.

Hmm ... the intent here is that if there is no storage, but a tablespace
is specified, then a dependency protects. This should be agnostic to
relkind considerations. I had first written the new symbol as
SHARED_DEPENDENCY_PARTITIONED_TABLE but then I realized the error of my
ways :-)

--
�lvaro Herrera Valdivia, Chile
Tom: There seems to be something broken here.
Teodor: I'm in sackcloth and ashes... Fixed.
http://archives.postgresql.org/message-id/482D1632.8010507@sigaev.ru