Re: [PATCH v3] xfs: reserve enough blocks to handle btree splits when remapping

[Christoph Hellwig <hch@infradead.org>]

On Mon, Apr 24, 2017 at 06:58:37PM -0700, Darrick J. Wong wrote:
> In xfs_reflink_end_cow, we erroneously reserve only enough blocks to
> handle adding 1 extent.  This is problematic if we fragment free space,
> have to do CoW, and then have to perform multiple bmap btree expansions.
> Furthermore, the BUI recovery routine doesn't reserve /any/ blocks to
> handle btree splits, so log recovery fails after our first error causes
> the filesystem to go down.
> 
> Therefore, refactor the transaction block reservation macros until we
> have a macro that works for our deferred (re)mapping activities, and fix
> both problems by using that macro.
> 
> With 1k blocks we can hit this fairly often in g/187 if the scratch fs
> is big enough.
> 
> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> ---
> v3: don't use the swap-extents-with-rmap block reservation for cow
> remapping; we should already have sufficient per-ag reservation
> v2: avoid 64-bit division when calculating block reservation
> ---
>  fs/xfs/libxfs/xfs_trans_space.h |   13 +++++++------
>  fs/xfs/xfs_bmap_item.c          |    7 ++++++-
>  fs/xfs/xfs_reflink.c            |   12 ++++++++++--
>  3 files changed, 23 insertions(+), 9 deletions(-)
> 
> diff --git a/fs/xfs/libxfs/xfs_trans_space.h b/fs/xfs/libxfs/xfs_trans_space.h
> index 7917f6e..0044e14 100644
> --- a/fs/xfs/libxfs/xfs_trans_space.h
> +++ b/fs/xfs/libxfs/xfs_trans_space.h
> @@ -23,6 +23,11 @@
>   */
>  #define XFS_MAX_CONTIG_RMAPS_PER_BLOCK(mp)    \
>  		(((mp)->m_rmap_mxr[0]) - ((mp)->m_rmap_mnr[0]))
> +#define XFS_RMAPADD_SPACE_RES(mp) ((mp)->m_rmap_maxlevels)
> +#define XFS_NRMAPADD_SPACE_RES(mp,b,w)\
> +	(((b + XFS_MAX_CONTIG_RMAPS_PER_BLOCK(mp) - 1) / \
> +	  XFS_MAX_CONTIG_RMAPS_PER_BLOCK(mp)) * \
> +	  XFS_RMAPADD_SPACE_RES(mp))

Comments, please!  (I know the existing defintions don't have any,
but that's bad enough to start with..)


>  	xfs_fsblock_t			firstfsb;
> +	unsigned int			resblks;
>  
>  	ASSERT(!test_bit(XFS_BUI_RECOVERED, &buip->bui_flags));
>  
> @@ -447,7 +450,9 @@ xfs_bui_recover(
>  		return -EIO;
>  	}
>  
> -	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
> +	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
> +	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks, 0,
> +			0, &tp);

Do we really need that resblks variable?

>  	if (error)
>  		return error;
>  	budp = xfs_trans_get_bud(tp, buip);
> diff --git a/fs/xfs/xfs_reflink.c b/fs/xfs/xfs_reflink.c
> index c0f3754..aab156a 100644
> --- a/fs/xfs/xfs_reflink.c
> +++ b/fs/xfs/xfs_reflink.c
> @@ -705,8 +705,16 @@ xfs_reflink_end_cow(
>  	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
>  	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
>  
> -	/* Start a rolling transaction to switch the mappings */
> -	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
> +	/*
> +	 * Start a rolling transaction to switch the mappings.  We're
> +	 * unlikely ever to have to remap 16T worth of single-block
> +	 * extents, so just cap the worst case extent count to 2^32-1.
> +	 * Stick a warning in just in case, and avoid 64-bit division.
> +	 */
> +	WARN_ON(end_fsb - offset_fsb + 1 > UINT_MAX);
> +	resblks = min_t(xfs_fileoff_t, UINT_MAX, end_fsb - offset_fsb + 1);

I don't like unlikely statements.  What prevents us from doing so?
I think the way Linux limits writes to 32-bits using MAX_RW_COUNT
does, but then the language should be more assertive here.  I think
we should also shut down the fs here if the assert fails - otherwise
we'll leave a partially converted region around.

A BUILD_BUG_ON on MAX_RW_COUNT might also be useful to make the whole
scheme even safer.