Horton

Re: Race condition in TransactionIdIsInProgress

Robert Haas <robertmhaas@gmail.com>

View thread View on postgr.es
From: Robert Haas <robertmhaas@gmail.com>
To: Konstantin Knizhnik <knizhnik@garret.ru>
Cc: PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2022-02-10T20:21:27Z
Lists: pgsql-hackers 
On Thu, Feb 10, 2022 at 11:46 AM Konstantin Knizhnik <knizhnik@garret.ru> wrote:
> Postgres first records state transaction in CLOG, then removes
> transaction from procarray and finally releases locks.
> But it can happen that transaction is marked as committed in CLOG,
> XMAX_COMMITTED bit is set in modified tuple but
> TransactionIdIsInProgress still returns true. As a result "t_xmin %u is
> uncommitted in tuple (%u,%u) to be updated in table"
> error is reported.

This is exactly why the HeapTupleSatisfiesFOO() functions all test
TransactionIdIsInProgress() first and TransactionIdDidCommit() only if
it returns false. I don't think it's really legal to test
TransactionIdDidCommit() without checking TransactionIdIsInProgress()
first. And I bet that's exactly what this code is doing...

-- 
Robert Haas
EDB: http://www.enterprisedb.com

Commits

  1. Fix visibility check when XID is committed in CLOG but not in procarray.

  2. Fix possible HOT corruption when RECENTLY_DEAD changes to DEAD while pruning.

  3. Make TransactionIdIsInProgress check transam.c's single-item XID status cache