JBD commmit : Get list of all 'to be written' buffers in advance

JBD commmit : Get list of all 'to be written' buffers in advance

[Niraj Kulkarni]

Hi all,
          I am trying to implement a concept called Transactional Flash 
(http://www.usenix.org/event/osdi08/tech/full_papers/prabhakaran/prabhakaran_html/) 

          in which all writes are flushed to disk (in this case Flash) 
as a circular linked list of pages. No journaling mechanisms are used.
           For that I am trying to get list of all buffers to be written 
(metadata or data) before actually writing them. I thought of 2 
approaches :

1. Get conservative list of all buffers before flushing any of them 
(before commit phase 2) in journal_commit_transaction function.
      Problem : Metadata writes use temporary buffers which are 
allocated just before write.

2. Instead of submitting buffers straightaway, collect all of them ( 
data + metadata ) and flush them in end.
      Problem :  i.   Possibility of blocking after data buffers are  
collected but before metadata buffer collection.
                            Since data buffers are locked, this may lead 
to deadlock. (I am not sure if it is spinlock)

                       ii.   Not able to differentiate between metadata 
buffers and Descriptor buffers.


Can anybody please help me in this regard?

Niraj

Re: JBD commmit : Get list of all 'to be written' buffers in advance

[Jan Kara]

  Hello,

On Thu 12-05-11 19:52:21, Niraj Kulkarni wrote:
>          I am trying to implement a concept called Transactional
> Flash (http://www.usenix.org/event/osdi08/tech/full_papers/prabhakaran/prabhakaran_html/)
> 
>          in which all writes are flushed to disk (in this case
> Flash) as a circular linked list of pages. No journaling mechanisms
> are used.
>           For that I am trying to get list of all buffers to be
> written (metadata or data) before actually writing them. I thought
> of 2 approaches :
> 
> 1. Get conservative list of all buffers before flushing any of them
> (before commit phase 2) in journal_commit_transaction function.
>      Problem : Metadata writes use temporary buffers which are
> allocated just before write.
  Well, if I understand the transactional flash concept right, what you
should do is to replace code committing metadata to the journal by
something which writes metadata directly to final location. That should be
relatively easy modification of loops in journal_commit_transaction(). BTW,
you might want to use ext4/jbd2 instead of ext3/jbd as a base for your
experiments since ext3 is basically deprecated.

> 2. Instead of submitting buffers straightaway, collect all of them (
> data + metadata ) and flush them in end.
>      Problem :  i.   Possibility of blocking after data buffers are
> collected but before metadata buffer collection.
>                            Since data buffers are locked, this may
> lead to deadlock. (I am not sure if it is spinlock)
> 
>                       ii.   Not able to differentiate between
> metadata buffers and Descriptor buffers.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: JBD commmit : Get list of all 'to be written' buffers in advance

[Jan Kara]

On Wed 18-05-11 13:38:47, niraj kulkarni wrote:
> Thanks for the reply. That solves most of the problem.
> 
> So now all I need is to gather information of all submitted buffers in
> current transactions, and pass it through ioctl to lower layer.
> What I am currently doing is adding buffer  to a list instead of submitting
> them straightaway, and side by side constructing an array holding these
> buffer numbers. So i submit this buffer number array through ioctl, and then
> actually submit buffer.
> 
> This needs a few kmallocs while I am holding buffer_lock() on few buffers.
> So is it potentially deadlocking situation?
  It shouldn't be if you use proper gfp mask - like GFP_NOFS should be OK.

									Honza
> On Tue, May 17, 2011 at 11:39 PM, Jan Kara <jack@suse.cz> wrote:
> 
> >  Hello,
> >
> > On Thu 12-05-11 19:52:21, Niraj Kulkarni wrote:
> > >          I am trying to implement a concept called Transactional
> > > Flash (
> > http://www.usenix.org/event/osdi08/tech/full_papers/prabhakaran/prabhakaran_html/
> > )
> > >
> > >          in which all writes are flushed to disk (in this case
> > > Flash) as a circular linked list of pages. No journaling mechanisms
> > > are used.
> > >           For that I am trying to get list of all buffers to be
> > > written (metadata or data) before actually writing them. I thought
> > > of 2 approaches :
> > >
> > > 1. Get conservative list of all buffers before flushing any of them
> > > (before commit phase 2) in journal_commit_transaction function.
> > >      Problem : Metadata writes use temporary buffers which are
> > > allocated just before write.
> >   Well, if I understand the transactional flash concept right, what you
> > should do is to replace code committing metadata to the journal by
> > something which writes metadata directly to final location. That should be
> > relatively easy modification of loops in journal_commit_transaction(). BTW,
> > you might want to use ext4/jbd2 instead of ext3/jbd as a base for your
> > experiments since ext3 is basically deprecated.
> >
> > > 2. Instead of submitting buffers straightaway, collect all of them (
> > > data + metadata ) and flush them in end.
> > >      Problem :  i.   Possibility of blocking after data buffers are
> > > collected but before metadata buffer collection.
> > >                            Since data buffers are locked, this may
> > > lead to deadlock. (I am not sure if it is spinlock)
> > >
> > >                       ii.   Not able to differentiate between
> > > metadata buffers and Descriptor buffers.
> >
> >                                                                 Honza
> > --
> > Jan Kara <jack@suse.cz>
> > SUSE Labs, CR
> >
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR