Re: Delayed allocation and page_lock vs transaction start ordering

[Mingming Cao <cmm@us.ibm.com>]

On Fri, 2008-04-18 at 12:54 -0600, Andreas Dilger wrote:
> On Apr 16, 2008  11:38 +0200, Jan Kara wrote:
> > On Tue 15-04-08 11:08:52, Mingming Cao wrote:
> > > I guess this reserve locking ordering allows support writepages() for
> > > ext3/4? What other the benefits?
> >
> >   Yes, that is one advantage. The other one (which I care about the most)
> > is that transaction commit code can take page_lock in the new locking order
> > which is necessary for the new ordered mode rewrite.
> 
> My understanding is that the main reason for the ordered mode rewrite is
> specifically to allow delalloc to still support ordered mode semantics.
> If the lock ordering is changed, and the jbd ordered mode is changed, but
> we don't support that with delalloc then we will have made a lot of changes
> (and likely introduced some bugs) with little benefit.
> 
> My apologies in advance if I misunderstand, and delalloc will be supported
> with these changes.
> 
I agrees with you that if we rewrite a new ordered mode(separate from
this one), we should make it possible to work with delalloc. 

Just want to clarify that the inversing locking patch proposed here
could work delalloc(just the lock ordering. I have updated delalloc to
work for the inversed locking. Just FYI I have merged the inverse
locking patches for ext4 to the unstable ext4 patche queue to see how it
goes.


Mingming