MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII;
 format=flowed
Content-Transfer-Encoding: 7bit
Date: Thu, 28 Jan 2021 14:42:40 +0300
From: Alexey Kondratov <a.kondratov@postgrespro.ru>
To: Alvaro Herrera <alvherre@alvh.no-ip.org>
Cc: Michael Paquier <michael@paquier.xyz>, Justin Pryzby
 <pryzby@telsasoft.com>, Peter Eisentraut
 <peter.eisentraut@enterprisedb.com>, Masahiko Sawada
 <masahiko.sawada@2ndquadrant.com>, Steve Singer <steve@ssinger.info>,
 pgsql-hackers@lists.postgresql.org, Robert Haas <robertmhaas@gmail.com>,
 Alexander Korotkov <a.korotkov@postgrespro.ru>, Masahiko Sawada
 <sawada.mshk@gmail.com>, Jose Luis Tallon <jltallon@adv-solutions.net>
Subject: Re: Allow CLUSTER, VACUUM FULL and REINDEX to change tablespace on
 the fly
In-Reply-To: <20210127213658.GA736@alvherre.pgsql>
References: <20210127213658.GA736@alvherre.pgsql>
User-Agent: Roundcube Webmail/1.4.0
Message-ID: <55c6c1526b94d2caf54c33262fe00674@postgrespro.ru>
Precedence: bulk

On 2021-01-28 00:36, Alvaro Herrera wrote:
> On 2021-Jan-28, Alexey Kondratov wrote:
> 
>> I have read more about lock levels and ShareLock should prevent any 
>> kind of
>> physical modification of indexes. We already hold ShareLock doing
>> find_all_inheritors(), which is higher than ShareUpdateExclusiveLock, 
>> so
>> using ShareLock seems to be safe here, but I will look on it closer.
> 
> You can look at lock.c where LockConflicts[] is; that would tell you
> that ShareLock indeed conflicts with ShareUpdateExclusiveLock ... but 
> it
> does not conflict with itself!  So it would be possible to have more
> than one process doing this thing at the same time, which surely makes
> no sense.
> 

Thanks for the explanation and pointing me to the LockConflicts[]. This 
is a good reference.

> 
> I didn't look at the patch closely enough to understand why you're
> trying to do something like CLUSTER, VACUUM FULL or REINDEX without
> holding full AccessExclusiveLock on the relation.  But do keep in mind
> that once you hold a lock on a relation, trying to grab a weaker lock
> afterwards is pretty pointless.
> 

No, you are right, we are doing REINDEX with AccessExclusiveLock as it 
was before. This part is a more specific one. It only applies to 
partitioned indexes, which do not hold any data, so we do not reindex 
them directly, only their leafs. However, if we are doing a TABLESPACE 
change, we have to record it in their pg_class entry, so all future leaf 
partitions were created in the proper tablespace.

That way, we open partitioned index relation only for a reference, i.e. 
read-only, but modify pg_class entry under a proper lock 
(RowExclusiveLock). That's why I thought that ShareLock will be enough.

IIUC, 'ALTER TABLE ... SET TABLESPACE' uses AccessExclusiveLock even for 
relations with no storage, since AlterTableGetLockLevel() chooses it if 
AT_SetTableSpace is met. This is very similar to our case, so probably 
we should do the same?

Actually it is not completely clear for me why ShareUpdateExclusiveLock 
is sufficient for newly added SetRelationTableSpace() as Michael wrote 
in the comment.


Regards
-- 
Alexey Kondratov

Postgres Professional https://www.postgrespro.com
Russian Postgres Company