[PATCH] xfs_repair: fix totally broken unit conversion in directory invalidation

[PATCH] xfs_repair: fix totally broken unit conversion in directory invalidation

[Darrick J. Wong]

From: Darrick J. Wong <darrick.wong@oracle.com>

Your humble author forgot that xfs_dablk_t has the same units as
xfs_fileoff_t, and totally screwed up the directory buffer invalidation
loop in dir_binval.  Not only is there an off-by-one error in the loop
conditional, but the unit conversions are wrong.

Fix all this stupidity by adding a for loop macro to take care of these
details for us so that everyone can iterate all logical directory blocks
(xfs_dir2_db_t) that start within a given bmbt record.

The pre-5.5 xfs_da_get_buf implementation mostly hides the off-by-one
error because dir_binval turns on "don't complain if no mapping" mode,
but on dirblocksize > fsblocksize filesystems the incorrect units can
cause us to miss invalidating some blocks, which can lead to other
buffer cache errors later.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
---
 libxfs/xfs_dir2.h |   10 ++++++++++
 repair/phase6.c   |    8 ++------
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/libxfs/xfs_dir2.h b/libxfs/xfs_dir2.h
index f5424477..13221243 100644
--- a/libxfs/xfs_dir2.h
+++ b/libxfs/xfs_dir2.h
@@ -308,6 +308,16 @@ xfs_dir2_leaf_tail_p(struct xfs_da_geometry *geo, struct xfs_dir2_leaf *lp)
 		  sizeof(struct xfs_dir2_leaf_tail));
 }
 
+/*
+ * For a given dir/attr geometry and extent mapping record, walk every file
+ * offset block (xfs_dablk_t) in the mapping that corresponds to the start
+ * of a logical directory block (xfs_dir2_db_t).
+ */
+#define for_each_xfs_bmap_dabno(geo, irec, dabno) \
+	for ((dabno) = round_up((irec)->br_startoff, (geo)->fsbcount); \
+	     (dabno) < (irec)->br_startoff + (irec)->br_blockcount; \
+	     (dabno) += (geo)->fsbcount)
+
 /*
  * The Linux API doesn't pass down the total size of the buffer
  * we read into down to the filesystem.  With the filldir concept
diff --git a/repair/phase6.c b/repair/phase6.c
index 91d208a6..a4dd3188 100644
--- a/repair/phase6.c
+++ b/repair/phase6.c
@@ -1276,7 +1276,7 @@ dir_binval(
 	struct xfs_ifork	*ifp;
 	struct xfs_da_geometry	*geo;
 	struct xfs_buf		*bp;
-	xfs_dablk_t		dabno, end_dabno;
+	xfs_dablk_t		dabno;
 	int			error = 0;
 
 	if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
@@ -1286,11 +1286,7 @@ dir_binval(
 	geo = tp->t_mountp->m_dir_geo;
 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
 	for_each_xfs_iext(ifp, &icur, &rec) {
-		dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
-				geo->fsbcount - 1);
-		end_dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
-				rec.br_blockcount);
-		for (; dabno <= end_dabno; dabno += geo->fsbcount) {
+		for_each_xfs_bmap_dabno(geo, &rec, dabno) {
 			bp = NULL;
 			error = -libxfs_da_get_buf(tp, ip, dabno, -2, &bp,
 					whichfork);

Re: [PATCH] xfs_repair: fix totally broken unit conversion in directory invalidation

[Darrick J. Wong]

On Tue, Dec 17, 2019 at 08:24:02PM -0800, Darrick J. Wong wrote:
> From: Darrick J. Wong <darrick.wong@oracle.com>
> 
> Your humble author forgot that xfs_dablk_t has the same units as
> xfs_fileoff_t, and totally screwed up the directory buffer invalidation
> loop in dir_binval.  Not only is there an off-by-one error in the loop
> conditional, but the unit conversions are wrong.
> 
> Fix all this stupidity by adding a for loop macro to take care of these
> details for us so that everyone can iterate all logical directory blocks
> (xfs_dir2_db_t) that start within a given bmbt record.
> 
> The pre-5.5 xfs_da_get_buf implementation mostly hides the off-by-one
> error because dir_binval turns on "don't complain if no mapping" mode,
> but on dirblocksize > fsblocksize filesystems the incorrect units can
> cause us to miss invalidating some blocks, which can lead to other
> buffer cache errors later.
> 
> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

Fixes: f9c559f4e4fb4 ("xfs_repair: invalidate dirty dir buffers when we zap a  directory")

--D

> ---
>  libxfs/xfs_dir2.h |   10 ++++++++++
>  repair/phase6.c   |    8 ++------
>  2 files changed, 12 insertions(+), 6 deletions(-)
> 
> diff --git a/libxfs/xfs_dir2.h b/libxfs/xfs_dir2.h
> index f5424477..13221243 100644
> --- a/libxfs/xfs_dir2.h
> +++ b/libxfs/xfs_dir2.h
> @@ -308,6 +308,16 @@ xfs_dir2_leaf_tail_p(struct xfs_da_geometry *geo, struct xfs_dir2_leaf *lp)
>  		  sizeof(struct xfs_dir2_leaf_tail));
>  }
>  
> +/*
> + * For a given dir/attr geometry and extent mapping record, walk every file
> + * offset block (xfs_dablk_t) in the mapping that corresponds to the start
> + * of a logical directory block (xfs_dir2_db_t).
> + */
> +#define for_each_xfs_bmap_dabno(geo, irec, dabno) \
> +	for ((dabno) = round_up((irec)->br_startoff, (geo)->fsbcount); \
> +	     (dabno) < (irec)->br_startoff + (irec)->br_blockcount; \
> +	     (dabno) += (geo)->fsbcount)
> +
>  /*
>   * The Linux API doesn't pass down the total size of the buffer
>   * we read into down to the filesystem.  With the filldir concept
> diff --git a/repair/phase6.c b/repair/phase6.c
> index 91d208a6..a4dd3188 100644
> --- a/repair/phase6.c
> +++ b/repair/phase6.c
> @@ -1276,7 +1276,7 @@ dir_binval(
>  	struct xfs_ifork	*ifp;
>  	struct xfs_da_geometry	*geo;
>  	struct xfs_buf		*bp;
> -	xfs_dablk_t		dabno, end_dabno;
> +	xfs_dablk_t		dabno;
>  	int			error = 0;
>  
>  	if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
> @@ -1286,11 +1286,7 @@ dir_binval(
>  	geo = tp->t_mountp->m_dir_geo;
>  	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
>  	for_each_xfs_iext(ifp, &icur, &rec) {
> -		dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> -				geo->fsbcount - 1);
> -		end_dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> -				rec.br_blockcount);
> -		for (; dabno <= end_dabno; dabno += geo->fsbcount) {
> +		for_each_xfs_bmap_dabno(geo, &rec, dabno) {
>  			bp = NULL;
>  			error = -libxfs_da_get_buf(tp, ip, dabno, -2, &bp,
>  					whichfork);

Re: [PATCH] xfs_repair: fix totally broken unit conversion in directory invalidation

[Christoph Hellwig]

On Tue, Dec 17, 2019 at 08:24:02PM -0800, Darrick J. Wong wrote:
> From: Darrick J. Wong <darrick.wong@oracle.com>
> 
> Your humble author forgot that xfs_dablk_t has the same units as
> xfs_fileoff_t, and totally screwed up the directory buffer invalidation
> loop in dir_binval.  Not only is there an off-by-one error in the loop
> conditional, but the unit conversions are wrong.

Can we kill off xfs_dablk_t?  I found the concept very, very confusing
when touching the dir code.

> --- a/libxfs/xfs_dir2.h
> +++ b/libxfs/xfs_dir2.h
> @@ -308,6 +308,16 @@ xfs_dir2_leaf_tail_p(struct xfs_da_geometry *geo, struct xfs_dir2_leaf *lp)
>  		  sizeof(struct xfs_dir2_leaf_tail));
>  }
>  
> +/*
> + * For a given dir/attr geometry and extent mapping record, walk every file
> + * offset block (xfs_dablk_t) in the mapping that corresponds to the start
> + * of a logical directory block (xfs_dir2_db_t).
> + */
> +#define for_each_xfs_bmap_dabno(geo, irec, dabno) \
> +	for ((dabno) = round_up((irec)->br_startoff, (geo)->fsbcount); \
> +	     (dabno) < (irec)->br_startoff + (irec)->br_blockcount; \
> +	     (dabno) += (geo)->fsbcount)

I think not having the magic for macro would be cleaner..

> +	xfs_dablk_t		dabno;
>  	int			error = 0;
>  
>  	if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
> @@ -1286,11 +1286,7 @@ dir_binval(
>  	geo = tp->t_mountp->m_dir_geo;
>  	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
>  	for_each_xfs_iext(ifp, &icur, &rec) {
> -		dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> -				geo->fsbcount - 1);
> -		end_dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> -				rec.br_blockcount);
> -		for (; dabno <= end_dabno; dabno += geo->fsbcount) {
> +		for_each_xfs_bmap_dabno(geo, &rec, dabno) {
>  			bp = NULL;
>  			error = -libxfs_da_get_buf(tp, ip, dabno, -2, &bp,
>  					whichfork);

But either way, the fix looks good.

Re: [PATCH] xfs_repair: fix totally broken unit conversion in directory invalidation

[Darrick J. Wong]

On Tue, Dec 24, 2019 at 12:38:43AM -0800, Christoph Hellwig wrote:
> On Tue, Dec 17, 2019 at 08:24:02PM -0800, Darrick J. Wong wrote:
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > Your humble author forgot that xfs_dablk_t has the same units as
> > xfs_fileoff_t, and totally screwed up the directory buffer invalidation
> > loop in dir_binval.  Not only is there an off-by-one error in the loop
> > conditional, but the unit conversions are wrong.
> 
> Can we kill off xfs_dablk_t?  I found the concept very, very confusing
> when touching the dir code.

I personally wouldn't mind if that happened (as a separate patch).

> > --- a/libxfs/xfs_dir2.h
> > +++ b/libxfs/xfs_dir2.h
> > @@ -308,6 +308,16 @@ xfs_dir2_leaf_tail_p(struct xfs_da_geometry *geo, struct xfs_dir2_leaf *lp)
> >  		  sizeof(struct xfs_dir2_leaf_tail));
> >  }
> >  
> > +/*
> > + * For a given dir/attr geometry and extent mapping record, walk every file
> > + * offset block (xfs_dablk_t) in the mapping that corresponds to the start
> > + * of a logical directory block (xfs_dir2_db_t).
> > + */
> > +#define for_each_xfs_bmap_dabno(geo, irec, dabno) \
> > +	for ((dabno) = round_up((irec)->br_startoff, (geo)->fsbcount); \
> > +	     (dabno) < (irec)->br_startoff + (irec)->br_blockcount; \
> > +	     (dabno) += (geo)->fsbcount)
> 
> I think not having the magic for macro would be cleaner..

Eh, yeah, we don't really need it...

--D

> 
> > +	xfs_dablk_t		dabno;
> >  	int			error = 0;
> >  
> >  	if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
> > @@ -1286,11 +1286,7 @@ dir_binval(
> >  	geo = tp->t_mountp->m_dir_geo;
> >  	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
> >  	for_each_xfs_iext(ifp, &icur, &rec) {
> > -		dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> > -				geo->fsbcount - 1);
> > -		end_dabno = xfs_dir2_db_to_da(geo, rec.br_startoff +
> > -				rec.br_blockcount);
> > -		for (; dabno <= end_dabno; dabno += geo->fsbcount) {
> > +		for_each_xfs_bmap_dabno(geo, &rec, dabno) {
> >  			bp = NULL;
> >  			error = -libxfs_da_get_buf(tp, ip, dabno, -2, &bp,
> >  					whichfork);
> 
> But either way, the fix looks good.