Date: Thu, 1 Feb 2024 15:28:28 -0800
From: Noah Misch <noah@leadboat.com>
To: pgsql-bugs@postgresql.org, dilipbalaut@gmail.com
Cc: robertmhaas@gmail.com
Subject: Re: STRATEGY=WAL_LOG missing checkpoint interlocks and sync
Message-ID: <20240201232828.b1.nmisch@google.com>
References: <20240130195003.0a.nmisch@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240130195003.0a.nmisch@google.com>
User-Agent: Mutt/2.2.12 (2023-09-09)
Archived-At: 
 <https://www.postgresql.org/message-id/20240201232828.b1.nmisch%40google.com>
Precedence: bulk

On Tue, Jan 30, 2024 at 11:50:03AM -0800, Noah Misch wrote:
> (1) CreateDirAndVersionFile() uses a code sequence: XLogInsert(), XLogFlush(),
> write(PG_VERSION).  This is missing the interlock with checkpoints, which
> pd_lsn usually secures.  This breaks when one takes a base backup between the
> XLogInsert() and the write().  The transam/README "action first and then write
> a WAL entry" strategy works well.  createdb_failure_callback() will remove the
> whole directory on failure.
> 
> (2) RelationMapCopy() reasons that it doesn't need RelationMappingLock.
> CheckPointRelationMap() relies on RelationMappingLock to avoid completing a
> checkpoint between the WAL and the sync.  A base backup taken in the middle of
> write_relmap_file() has no relmap after recovery.
> 
> (3) CreateDirAndVersionFile() populates the PG_VERSION file without syncing
> it.  An OS crash after the next checkpoint may leave PG_VERSION empty or
> missing.  Let's fsync the file immediately.  I used LazyFS to confirm the
> defect and fix.  Though I was skeptical about adding a wait event in back
> branches, I plan to do so anyway.  Commit 368ffde got away with it, and we do
> have this pattern of the sync event being separate from the write event.

Pushed at commit 0b6517a.