[Tarantool-patches] [PATCH v15 11/11] box/journal: redesign journal operations

[Konstantin Osipov]

* Cyrill Gorcunov <gorcunov at gmail.com> [20/03/20 13:34]:
> > >  
> > > -	if (txn_write_to_wal(req) != 0)
> > > +	fiber_set_txn(fiber(), NULL);
> > > +	if (journal_write(req) != 0) {
> > > +		fiber_set_txn(fiber(), txn);
> > 
> > I wonder why do you need to clear/set txn in txn_commit()?

Forgive me for being really painful about it, but why not use
different complete callbacks for sync and async wal writes?-)
Under the hood they will still call txn_complete(), but one will
assert, and another will not? 

> 
> Because async write engine implies that fiber's txn is dropped.
> In particular
> 
> void
> txn_complete_async(struct journal_entry *entry)
> {
> 	struct txn *txn = entry->complete_data;
> 	txn->signature = entry->res;
> 	/*
> 	 * Some commit/rollback triggers require for in_txn fiber
> 	 * variable to be set so restore it for the time triggers
> 	 * are in progress.
> 	 */
> -->	assert(in_txn() == NULL);
> 	fiber_set_txn(fiber(), txn);
> 	txn_complete(txn);
> 	fiber_set_txn(fiber(), NULL);
> }
> 
> Thus to not use txn.h inside journal/wall I clear it before
> the write and restore it back in case of error because rollback
> needs txn bound to a fiber.
> 
> > > +		txn_rollback(txn);
> > > +		txn_free(txn);
> > > +
> > > +		diag_set(ClientError, ER_WAL_IO);
> > > +		diag_log();
> > >  		return -1;
> > > +	}
> > >  
> > > -	/*
> > > -	 * In case of non-yielding journal the transaction could already
> > > -	 * be done and there is nothing to wait in such cases.
> > > -	 */
> > >  	if (!txn_has_flag(txn, TXN_IS_DONE)) {
> > > -		bool cancellable = fiber_set_cancellable(false);
> > > -		fiber_yield();
> > > -		fiber_set_cancellable(cancellable);
> > > +		txn->signature = req->res;
> > > +		txn_complete(txn);
> > > +		fiber_set_txn(fiber(), NULL);
> > 
> > Seems you do it twice ?
> 
> If you mean the txn_complete call then no, in async write
> we call the completion wich sets up TXN_IS_DONE flag, in
> turn sync write (without wal) doesn't call the completion
> and this flag is clear.
> 
> Or you meant something different?

I mean fiber_set_txn() is done twice, if I am not mistaken.


-- 
Konstantin Osipov, Moscow, Russia