Parallel Index Scan vs BTP_DELETED and BTP_HALF_DEAD

Hi hackers,

I heard a report of a 10.1 cluster hanging with several 'BtreePage'
wait_events showing in pg_stat_activity. The query plan involved
Parallel Index Only Scan, and the table is concurrently updated quite
heavily. I tried and failed to make a reproducer, but from the clues
available it seemed clear that somehow *all* participants in a
Parallel Index Scan must be waiting for someone else to advance the
scan. The report came with a back trace[1] that was the same in all 3
backends (leader + 2 workers), which I'll summarise here:

ConditionVariableSleep
_bt_parallel_seize
_bt_readnextpage
_bt_steppage
_bt_next
btgettuple
index_getnext_tid
IndexOnlyNext

I think _bt_steppage() called _bt_parallel_seize(), then it called
_bt_readnextpage() which I guess must have encountered a BTP_DELETED
or BTP_HALF_DEAD-marked page so didn't take this early break out of
the loop:

/* check for deleted page */
if (!P_IGNORE(opaque))
{
PredicateLockPage(rel, blkno,
scan->xs_snapshot);
/* see if there are any matches on this page */
/* note that this will clear moreRight
if we can stop */
if (_bt_readpage(scan, dir,
P_FIRSTDATAKEY(opaque)))
break;
}

... and then it called _bt_parallel_seize() itself, in violation of
the rule (by my reading of the code) that you must call
_bt_parallel_release() (via _bt_readpage()) or _bt_parallel_done()
after seizing the scan. If you call _bt_parallel_seize() again
without doing that first, you'll finish up waiting for yourself
forever. Does this theory make sense?

[1] http://dpaste.com/05PGJ4E

--
Thomas Munro
http://www.enterprisedb.com