On Thu, Nov 30, 2017 at 10:18 AM, Thomas Munro
<thomas.munro@enterprisedb.com> wrote:
> On Thu, Nov 30, 2017 at 9:34 AM, Alexander Voytsekhovskyy
> <young.inbox@gmail.com> wrote:
>> Thanks for helping, here is one more try
>>
>> #0 get_segment_by_index (area=area@entry=0x556026700be8, index=1) at
>> /build/postgresql-10-qAeTPy/postgresql-10-10.1/build/../src/backend/utils/mmgr/dsa.c:1736
>> #1 0x00005560252c2b90 in dsa_get_address (area=area@entry=0x556026700be8,
>> dp=dp@entry=1099511685280) at
>> /build/postgresql-10-qAeTPy/postgresql-10-10.1/build/../src/backend/utils/mmgr/dsa.c:945
>> #2 0x00005560250a2c2b in tbm_attach_shared_iterate
>> (dsa=dsa@entry=0x556026700be8, dp=1099511685280) at
>> /build/postgresql-10-qAeTPy/postgresql-10-10.1/build/../src/backend/nodes/tidbitmap.c:1503
>> #3 0x0000556025066c7b in BitmapHeapNext (node=node@entry=0x556026460710) at
>> /build/postgresql-10-qAeTPy/postgresql-10-10.1/build/../src/backend/executor/nodeBitmapHeapscan.c:176
>
> Thank you for the report and the back trace. I think this might be a
> manifestation of the problem I just described[1] on -hackers.
> Depending on the shape of a multi-Gather query plan and therefore the
> order of control flow, you might finish up using the DSA area that
> belongs to a different Gather node and then find that it goes away too
> soon. Investigating.
I haven't managed to reproduce this, but I was coincidentally
investigating a bug that appears to explain it. I think what happened
is that a background worker was first to execute BitmapHeapNext and
allocated a dsa_pointer, and then the leader process reached
BitmapHeapNext and called tbm_attach_shared_iterate which tried to
deference it, but it had es_query_dsa set to another gather node's DSA
area (whichever Gather most recently ran ExecInitParallelPlan). That
requires a certain order of execution and timing that I'm not sure how
to reach. I have posted a patch that should fix it over here:
https://www.postgresql.org/message-id/CAEepm%3D0Mv9BigJPpribGQhnHqVGYo2%2BkmzekGUVJJc9Y_ZVaYA%40mail.gmail.com
Are you able to provide a minimal reproducer, an anonymised partial
dump, or perhaps try out the patch on a copy of your database?
--
Thomas Munro
http://www.enterprisedb.com
