| From: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
|---|---|
| To: | pgsql-committers(at)lists(dot)postgresql(dot)org |
| Subject: | pgsql: Don't pfree a Path that belongs to a different parent relation. |
| Date: | 2026-11-21 19:08:07 |
| Message-ID: | E1xUqRS-00000000fOr-1hGg@gemulon.postgresql.org |
| Views: | Whole Thread | Raw Message | Download mbox | Resend email |
| Thread: | |
| Lists: | pgsql-committers |
Don't pfree a Path that belongs to a different parent relation.
add_path and add_partial_path pfree a dominated or rejected path
on the assumption, documented in add_path's header comment, that
"Paths of this rel cannot yet be referenced from any other rel".
At least two existing callers violate that: grouping_planner copies
the uppermost scan/join rel's paths to the query's final_rel, and
create_ordered_paths passes an already-sorted input_rel path
straight through to the ordered_rel. Most of the time we got away
with that, because the source relation's path list would not get
consulted again. But at least the grouping_planner case is a live
bug: since add_path and add_partial_path have somewhat different
comparison rules, it's possible for add_path to accept a Gather
path while add_partial_path rejects the corresponding partial path.
Then pfree'ing the partial path leaves a dangling pointer in the
Gather path, leading to crashes or strange planner failures if
the Gather path is selected for use. The create_ordered_paths
case has been reported to cause unexpected failures in debugging
situations, when one tries to dump the pathlist of the source rel.
To fix, refrain from pfree'ing a rejected path if its parent link
shows it was originally made for another relation. This almost
certainly indicates that it is still referenced in the other
rel's pathlist or partial_pathlist. Even if that list won't be
consulted again in normal planning, we want it to stay valid for
debugging purposes, so that this rule is unlikely to spare any
path that could safely be recycled.
There was discussion of solving this problem in a more principled
fashion by reference-counting Paths. But that would be a very
invasive (and not back-patchable) change, would add cycles that
we'd probably not win back, and it's not at all clear that it'd
make the planner any more robust anyway. We'd be replacing the
existing maybe-this-isn't-safe-to-pfree hazard with a hazard of
failing to update refcounts where needed, and that hazard would
be spread across much of the planner instead of being localized.
We have no test case that reliably exposes the grouping_planner bug
across branches, which is not so surprising because a case where
add_path and add_partial_path disagree on the merits of corresponding
paths is necessarily an edge case (probably involving fuzzily-equal
costs). So the behavior is exceedingly subject to change, and there's
little reason to hope that a test case that exposes the error today
would still do so in future versions. Hence, no test case in this
patch.
Bug: #19924
Reported-by: Jeevan Chalke <jeevan(dot)chalke(at)enterprisedb(dot)com>
Reported-by: Zheng Hacker <hackerzheng666(at)gmail(dot)com>
Author: Jeevan Chalke <jeevan(dot)chalke(at)enterprisedb(dot)com>
Reviewed-by: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Discussion: https://postgr.es/m/CAM2+6=XaYtMGwh27O7oDTfAw7kV4VdUVbxC2RVzWtKz_ysekKw@mail.gmail.com
Discussion: https://postgr.es/m/19924-8aecd924385fafcc@postgresql.org
Backpatch-through: 15
Branch
------
REL_15_STABLE
Details
-------
https://git.postgresql.org/pg/commitdiff/bf94082c101d077d55e436f1f9dc83a6d77a572b
Modified Files
--------------
src/backend/optimizer/util/pathnode.c | 46 ++++++++++++++++++++++++++---------
1 file changed, 34 insertions(+), 12 deletions(-)
| From | Date | Subject | |
|---|---|---|---|
| Previous Message | Fujii Masao | 2026-11-21 14:22:07 | pgsql: pg_basebackup: Fix crashes on malformed server responses |