xfs
[Top] [All Lists]

Re: [PATCH 4/6] xfs: allow reusing busy extents where safe

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: [PATCH 4/6] xfs: allow reusing busy extents where safe
From: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Date: Wed, 23 Mar 2011 08:26:34 -0400
Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
In-reply-to: <20110323002024.GF15270@dastard>
References: <20110322195550.260682574@xxxxxxxxxxxxxxxxxxxxxx> <20110322200137.837735220@xxxxxxxxxxxxxxxxxxxxxx> <20110323002024.GF15270@dastard>
User-agent: Mutt/1.5.21 (2010-09-15)
On Wed, Mar 23, 2011 at 11:20:24AM +1100, Dave Chinner wrote:
> >             overlap = xfs_alloc_busy_try_reuse(pag, busyp,
> >                                                fbno, fbno + flen);
> > -           if (overlap) {
> > +           if (overlap == -1 || (overlap && userdata)) {
> >                     spin_unlock(&pag->pagb_lock);
> >                     xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
> >                     goto restart;
> 
> Ok, so the only time we'll do a log force now is on an complete
> overlap or a partial userdata overlap?

We do it if we would have to split the busy extent, or reallocate
metadata to userdata.

> > +           if (!args->userdata ||
> > +               (busyp->flags & XFS_ALLOC_BUSY_USERDATA)) {
> > +                   int overlap;
> > +
> > +                   overlap = xfs_alloc_busy_try_reuse(args->pag, busyp,
> > +                                                      fbno, fbno + flen);
> > +                   if (unlikely(overlap == -1)) {
> > +                           spin_unlock(&args->pag->pagb_lock);
> > +                           xfs_log_force(args->mp, XFS_LOG_SYNC);
> > +                           goto restart;
> > +                   }
> 
> Hmmmm - I'm not so sure we can reuse overlapped data extents for
> data allocations without a log force at all as there is no guarantee
> that the data will not be overwritten before the original free
> transaction is on disk.
> 
> That is, recovery may not replay the original data extent free
> transaction or the new allocation transaction, but there is nothing
> stopping us from having written the new data into the extent before
> the crash occurred, especially as delayed allocation places the
> allocation very close the data IO issue. e.g.:
> 
>       thread X                thread Y
>       free data extent ABC
>                               allocate data extent BCD
>                               partial overlap, no log force
>                               issue data IO
>                               .....
> 
>                <crash>
> 
> That leads to corruption of the data in the original file because
> neither transaction is written to disk, but new data is....

You're right.  It's fine from the transaction point of view, but
it could leave wrong data in a file.  I'll disallow this case, too.

<Prev in Thread] Current Thread [Next in Thread>