Re: [PATCH 2/2] xfs_repair: correctly account for free space btree shrinks when fixing freelist

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

On Fri, Feb 01, 2019 at 12:39:16PM -0600, Eric Sandeen wrote:
> On 1/30/19 11:50 AM, Darrick J. Wong wrote:
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > When we fix the freelist at the end of build_agf_agfl in phase 5 of
> > repair, we need to create incore rmap records for the blocks that get
> > added to the AGFL.  We can't let the regular freelist fixing code use
> > the regular on-disk rmapbt update code because the rmapbt isn't fully
> > set up yet.
> > 
> > Unfortunately, the original code fails to account for the fact that the
> > free space btrees can shrink when we allocate blocks to fix the
> > freelist; those blocks are also put on the freelist, but there are
> > already incore rmaps for all the free space btree blocks.  We must not
> > create (redundant) incore rmaps for those blocks.  If we do, repair
> > fails with a complaint that rebuilding the rmapbt failed during phase 5.
> > xfs/137 on a 1k block size occasionally triggers this bug.
> > 
> > To fix the problem, construct a bitmap of all OWN_AG blocks that we know
> > about before traversing the AGFL, and only create new incore rmaps for
> > those AGFL blocks that are not already tracked in the bitmap.
> > 
> > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > ---
> >  repair/rmap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++----------
> >  1 file changed, 44 insertions(+), 10 deletions(-)
> > 
> > 
> > diff --git a/repair/rmap.c b/repair/rmap.c
> > index c5a86646..3bb05065 100644
> > --- a/repair/rmap.c
> > +++ b/repair/rmap.c
> > @@ -12,6 +12,7 @@
> >  #include "dinode.h"
> >  #include "slab.h"
> >  #include "rmap.h"
> > +#include "bitmap.h"
> >  
> >  #undef RMAP_DEBUG
> >  
> > @@ -450,6 +451,8 @@ rmap_store_ag_btree_rec(
> >  	struct xfs_buf		*agflbp = NULL;
> >  	struct xfs_trans	*tp;
> >  	__be32			*agfl_bno, *b;
> > +	struct xfs_ag_rmap	*ag_rmap = &ag_rmaps[agno];
> > +	struct bitmap		*own_ag_bitmap = NULL;
> >  	int			error = 0;
> >  	struct xfs_owner_info	oinfo;
> >  
> > @@ -457,9 +460,8 @@ rmap_store_ag_btree_rec(
> >  		return 0;
> >  
> >  	/* Release the ar_rmaps; they were put into the rmapbt during p5. */
> > -	free_slab(&ag_rmaps[agno].ar_rmaps);
> > -	error = init_slab(&ag_rmaps[agno].ar_rmaps,
> > -				  sizeof(struct xfs_rmap_irec));
> > +	free_slab(&ag_rmap->ar_rmaps);
> > +	error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec));
> >  	if (error)
> >  		goto err;
> >  
> > @@ -479,19 +481,50 @@ rmap_store_ag_btree_rec(
> >  	 * rmap, we only need to add rmap records for AGFL blocks past
> >  	 * that point in the AGFL because those blocks are a result of a
> >  	 * no-rmap no-shrink freelist fixup that we did earlier.
> > +	 *
> > +	 * However, some blocks end up on the AGFL because the free space
> > +	 * btrees shed blocks as a result of allocating space to fix the
> > +	 * freelist.  We already created in-core rmap records for the free
> > +	 * space btree blocks, so we must be careful not to create those
> > +	 * records again.  Create a bitmap of already-recorded OWN_AG rmaps.
> >  	 */
> > +	error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur);
> > +	if (error)
> > +		goto err;
> > +	if (!bitmap_init(&own_ag_bitmap)) {
> > +		error = -ENOMEM;
> > +		goto err;
> 
> I think that this ^^
> 
> > +	}
> > +	while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) {
> > +		if (rm_rec->rm_owner != XFS_RMAP_OWN_AG)
> > +			continue;
> > +		if (!bitmap_set(own_ag_bitmap, rm_rec->rm_startblock,
> > +					rm_rec->rm_blockcount)) {
> > +			error = EFSCORRUPTED;
> > +			goto err;
> 
> and this need to free up rm_cur to be tidy, right?

Yep, they do need to be 'goto err_slab', thanks for catching that.

--D

> -Eric