On Tue, Mar 23, 2021 at 10:08 AM Amul Sul <sulamul@gmail.com> wrote:
On Mon, Mar 22, 2021 at 3:03 PM Amit Langote <amitlangote09@gmail.com> wrote: > > On Mon, Mar 22, 2021 at 5:26 PM Amul Sul <sulamul@gmail.com> wrote: > > In heapam_relation_copy_for_cluster(), begin_heap_rewrite() sets > > rwstate->rs_new_rel->rd_smgr correctly but next line tuplesort_begin_cluster() > > get called which cause the system cache invalidation and due to CCA setting, > > wipe out rwstate->rs_new_rel->rd_smgr which wasn't restored for the subsequent > > operations and causes segmentation fault. > > > > By calling RelationOpenSmgr() before calling smgrimmedsync() in > > end_heap_rewrite() would fix the failure. Did the same in the attached patch. > > That makes sense. I see a few commits in the git history adding > RelationOpenSmgr() before a smgr* operation, whenever such a problem > would have been discovered: 4942ee656ac, afa8f1971ae, bf347c60bdd7, > for example. >
Thanks for the confirmation.
> I do wonder if there are still other smgr* operations in the source > code that are preceded by operations that would invalidate the > SMgrRelation that those smgr* operations would be called with. For > example, the smgrnblocks() in gistBuildCallback() may get done too > late than a corresponding RelationOpenSmgr() on the index relation. >
I did the check for gistBuildCallback() by adding Assert(index->rd_smgr) before smgrnblocks() with CCA setting and didn't see any problem there.
I think the easiest way to find that is to run a regression suite with CCA build, perhaps, there is no guarantee that regression will hit all smgr* operations, but that might hit most of them.
Sure, will give a regression run with CCA enabled.