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

[Cyrill Gorcunov]

On Fri, Mar 20, 2020 at 04:09:05PM +0300, Konstantin Osipov wrote:
> * Cyrill Gorcunov <gorcunov at gmail.com> [20/03/20 14:15]:
> > On Fri, Mar 20, 2020 at 01:58:42PM +0300, Konstantin Osipov wrote:
> > > * 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?
> > 
> > Hmm. If I remember correctly we've been planning to use callbacks
> > only for async writes. Actually I can introduce callback helper
> > for sync writes as well but this ruines the whole idea, no?
> 
> But aren't you using the same callback for sync and async now?

Yes, but this is only because we _have_ to use callbacks in
wal engine for both sync\async writes. The general architecture
is that - sync writes do _not_have_ to use callbacks.

The use of callback in wal is transparent to the caller. At
least I tried to make it so.

> And if you are not using callback for sync, why do you need to
> manipulate with txn in sync?
> 
> I'm lost, I accept it. 

Because of wal and async engine in it :(

Look the whole idea is the following:

 - journal_write_async always use write_async_cb
 - journal_write should not use async write (or
   it could but transparently)
 - journal_write can complete transaction by self,
   for this sake it tests for TXN_IS_DONE bit and
   doesn't call for txn_complete if bit is set.

You know, I think we're in good shape right now
and can cleanup the series on top maybe?

> > I can easily hide this bit test inside txn_complete itself and
> > for sync write there will be plain txn_complete call, like
> > 
> > txn_commit
> > 	...
> > 	journal_write();
> > 	...
> > 	txn_complete();
> 
> My point is simple: can we avoid the whole mess of clearing and
> restoring fiber txn for sync write calls? 

Letme think about it. But I would prefer to make it on top.