Re: [PATCH 1/4] xfs: introduce fake roots for ag-rooted btrees

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Thu, Mar 05, 2020 at 10:53:35AM +1100, Dave Chinner wrote:
> On Wed, Mar 04, 2020 at 03:34:59PM -0800, Darrick J. Wong wrote:
> > On Wed, Mar 04, 2020 at 01:21:03PM -0500, Brian Foster wrote:
> > > On Tue, Mar 03, 2020 at 07:28:28PM -0800, Darrick J. Wong wrote:
> > > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > > 
> > > > Create an in-core fake root for AG-rooted btree types so that callers
> > > > can generate a whole new btree using the upcoming btree bulk load
> > > > function without making the new tree accessible from the rest of the
> > > > filesystem.  It is up to the individual btree type to provide a function
> > > > to create a staged cursor (presumably with the appropriate callouts to
> > > > update the fakeroot) and then commit the staged root back into the
> > > > filesystem.
> > > > 
> > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > > > ---
> > > 
> > > The code all seems reasonable, mostly infrastructure. Just a few high
> > > level comments..
> > > 
> > > It would be helpful if the commit log (or code comments) explained more
> > > about the callouts that are replaced for a staging tree (and why).
> > 
> > Ok.  I have two block comments to add.
> > 
> > > >  fs/xfs/libxfs/xfs_btree.c |  117 +++++++++++++++++++++++++++++++++++++++++++++
> > > >  fs/xfs/libxfs/xfs_btree.h |   42 ++++++++++++++--
> > > >  fs/xfs/xfs_trace.h        |   28 +++++++++++
> > > >  3 files changed, 182 insertions(+), 5 deletions(-)
> > > > 
> > > > 
> > > > diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
> > > > index e6f898bf3174..9a7c1a4d0423 100644
> > > > --- a/fs/xfs/libxfs/xfs_btree.c
> > > > +++ b/fs/xfs/libxfs/xfs_btree.c
> > > > @@ -382,6 +382,8 @@ xfs_btree_del_cursor(
> > > >  	/*
> > > >  	 * Free the cursor.
> > > >  	 */
> > > > +	if (unlikely(cur->bc_flags & XFS_BTREE_STAGING))
> > > > +		kmem_free((void *)cur->bc_ops);
> > > >  	kmem_cache_free(xfs_btree_cur_zone, cur);
> > > >  }
> > > >  
> > > > @@ -4908,3 +4910,118 @@ xfs_btree_has_more_records(
> > > >  	else
> > > >  		return block->bb_u.s.bb_rightsib != cpu_to_be32(NULLAGBLOCK);
> > > >  }
> > 
> > Add here a new comment:
> > 
> > /*
> >  * Staging Cursors and Fake Roots for Btrees
> >  * =========================================
> >  *
> >  * A staging btree cursor is a special type of btree cursor that callers
> >  * must use to construct a new btree index using the btree bulk loader
> >  * code.  The bulk loading code uses the staging btree cursor to
> >  * abstract the details of initializing new btree blocks and filling
> >  * them with records or key/ptr pairs.  Regular btree operations (e.g.
> >  * queries and modifications) are not supported with staging cursors,
> >  * and callers must not invoke them.
> >  *
> >  * Fake root structures contain all the information about a btree that
> >  * is under construction by the bulk loading code.  Staging btree
> >  * cursors point to fake root structures instead of the usual AG header
> >  * or inode structure.
> >  *
> >  * Callers are expected to initialize a fake root structure and pass it
> >  * into the _stage_cursor function for a specific btree type.  When bulk
> >  * loading is complete, callers should call the _commit_staged_btree
> >  * function for that specific btree type to commit the new btree into
> >  * the filesystem.
> >  */
> > 
> > 
> > > > +
> > > > +/* We don't allow staging cursors to be duplicated. */
> > 
> > /*
> >  * Don't allow staging cursors to be duplicated because they're supposed
> >  * to be kept private to a single thread.
> >  */
> 
> private to a single -thread- or a single -btree modification
> context-?

Private to a single btree rebuilding context, really. :)

ATM btree rebuilding contexts are single-threaded.

--D

> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com