Re: Rome university

Hi,

We're a team from Rome University (Italy) and we are working on an

hacking of PostgreSQL MVCC. The basic idea is to have multiple

instances

of a same user transaction concurrently executing against the DB in

order to achieve fault tolerance. We do not want to bother you with the

details of our idea, but of course we'd be delighted to provide you

additional information wheter you were interested or even just curious!

We're writing to you since we're studying the current MVCC

implementation and we're stuck with a doubt that's driving us crazy

since a couple of weeks so we'd be extremely grateful to anybody who

could enlighten us.

Our amlethic question is:

tqual.c, HeapTupleSatisfies* (Self,Now etc). Let's focus on

HeapTupleSatisfiesItself for simplicity. The piece of code that's

unclear is the following:

144: else if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmin(tuple)))

145: {

146: if (tuple->t_infomask & HEAP_XMAX_INVALID) /* xid invalid */

147: return true;

148:

149: if (tuple->t_infomask & HEAP_IS_LOCKED) /* not deleter */

150: return true;

151:

152: Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));

153:

154: /* deleting subtransaction aborted? */

155: if (TransactionIdDidAbort(HeapTupleHeaderGetXmax(tuple)))

156: {

157: tuple->t_infomask |= HEAP_XMAX_INVALID;

158: SetBufferCommitInfoNeedsSave(buffer);

159: return true;

160: }

161:

162: Assert(TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmax(tuple)));

163:

164: return false;

165: }

and the doubt is the following: how is it possible that -line 144- Xmin

is the current transaction ( i.e. it has created this tuple, it is

holding an exclusive lock on it since it has not committed yet) and

that

-line 149- there is a different (?) transaction that is also locking

the

tuple (HEAP_IS_LOCKED=(HEAP_XMAX_EXCL_LOCK||HEAP_XMAX_SHARED_LOCK) )?

Unless we are missing something, this situation is possible exclusively

in case the XMAX transaction is a subtransaction of XMIN, which can

access the tuple despite the exclusive lock held by XMIN. This seems

correct according to the comment in line 154, which refers to a

"subtransaction".

Are we understanding correctly what this code is doing and the related

underlying MVCC mechanisms?

Chiacchiera con i tuoi amici in tempo reale!
http://it.yahoo.com/mail_it/foot/*http://it.messenger.yahoo.com

letizia leo wrote:

> and the doubt is the following: how is it possible that -line 144- Xmin
> is the current transaction ( i.e. it has created this tuple, it is
> holding an exclusive lock on it since it has not committed yet) and
> that
> -line 149- there is a different (?) transaction that is also locking
> the
> tuple (HEAP_IS_LOCKED=(HEAP_XMAX_EXCL_LOCK||HEAP_XMAX_SHARED_LOCK) )?
> Unless we are missing something, this situation is possible exclusively
> in case the XMAX transaction is a subtransaction of XMIN, which can
> access the tuple despite the exclusive lock held by XMIN.

Exactly.

> This seems correct according to the comment in line 154, which refers
> to a "subtransaction".

It is correct.

> Are we understanding correctly what this code is doing and the related
> underlying MVCC mechanisms?

Can't say, but that part at least you got right :-)

--
Alvaro Herrera http://www.CommandPrompt.com/
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

On Tue, May 02, 2006 at 04:31:14PM +0200, letizia leo wrote:

<snip>
>
> and the doubt is the following: how is it possible that -line 144- Xmin
> is the current transaction ( i.e. it has created this tuple, it is
> holding an exclusive lock on it since it has not committed yet) and
> that
> -line 149- there is a different (?) transaction that is also locking
> the
> tuple (HEAP_IS_LOCKED=(HEAP_XMAX_EXCL_LOCK||HEAP_XMAX_SHARED_LOCK) )?

Are you considering READ COMMITTED access mode? There you can see
tuples added by commands that have not really committed.

> Unless we are missing something, this situation is possible exclusively
> in case the XMAX transaction is a subtransaction of XMIN, which can
> access the tuple despite the exclusive lock held by XMIN. This seems
> correct according to the comment in line 154, which refers to a
> "subtransaction".

I don't know much about this code, but at the very least it could be
there to check that what is in the xmax field is actually a real
transaction value and not a locking transaction... Future-proofing?

Hope this helps,
--
Martijn van Oosterhout <kleptog(at)svana(dot)org> http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.