Horton

Re: StandbyRecoverPreparedTransactions recovers subtrans links incorrectly

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Nikhil Sontakke <nikhils@2ndquadrant.com>
Cc: Simon Riggs <simon@2ndquadrant.com>, Andres Freund <andres@anarazel.de>, PostgreSQL-development <pgsql-hackers@postgresql.org>
Date: 2017-04-26T14:28:46Z
Lists: pgsql-hackers 
Nikhil Sontakke <nikhils@2ndquadrant.com> writes:
> A SELECT query on the newly promoted master on any of the tables involved
> in the 2PC hangs. The hang is due to a loop in
> SubTransGetTopmostTransaction(). Due to incorrect linkages, we get a
> circular reference in parentxid <-> subxid inducing the infinite loop.

I'm inclined to propose that

(1) SubTransSetParent should contain an assert that 
	Assert(TransactionIdFollows(xid, parent));
There is a similar assertion near one of the call sites, but that's
obviously too far away from the action.

(2) SubTransGetTopmostTransaction ought to contain an actual runtime
test and ereport (not just Assert) that the parent XID it got from
pg_subtrans precedes the child XID that was looked up.  This would
protect us against similar infinite loops in the future.  Even without
bugs, such a loop could arise due to corrupted data.

			regards, tom lane

Commits

  1. Rework handling of subtransactions in 2PC recovery

  2. Fix order of arguments to SubTransSetParent().

  3. Speedup 2PC recovery by skipping two phase state files in normal path