On Monday 31 August 2020 9:59:08 PM IST Darrick J. Wong wrote:
> On Thu, Aug 20, 2020 at 11:13:47AM +0530, Chandan Babu R wrote:
> > Moving an extent to data fork can cause a sub-interval of an existing
> > extent to be unmapped. This will increase extent count by 1. Mapping in
> > the new extent can increase the extent count by 1 again i.e.
> > | Old extent | New extent | Old extent |
> > Hence number of extents increases by 2.
> >
> > Signed-off-by: Chandan Babu R <chandanrlinux@xxxxxxxxx>
> > ---
> > fs/xfs/libxfs/xfs_inode_fork.h | 9 ++++++++-
> > fs/xfs/xfs_reflink.c | 5 +++++
> > 2 files changed, 13 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/xfs/libxfs/xfs_inode_fork.h b/fs/xfs/libxfs/xfs_inode_fork.h
> > index d0e49b015b62..850d53162545 100644
> > --- a/fs/xfs/libxfs/xfs_inode_fork.h
> > +++ b/fs/xfs/libxfs/xfs_inode_fork.h
> > @@ -78,7 +78,14 @@ struct xfs_ifork {
> > * split into two extents causing extent count to increase by 1.
> > */
> > #define XFS_IEXT_INSERT_HOLE_CNT (1)
> > -
> > +/*
> > + * Moving an extent to data fork can cause a sub-interval of an existing extent
> > + * to be unmapped. This will increase extent count by 1. Mapping in the new
> > + * extent can increase the extent count by 1 again i.e.
> > + * | Old extent | New extent | Old extent |
> > + * Hence number of extents increases by 2.
> > + */
> > +#define XFS_IEXT_REFLINK_END_COW_CNT (2)
> >
> > /*
> > * Fork handling.
> > diff --git a/fs/xfs/xfs_reflink.c b/fs/xfs/xfs_reflink.c
> > index aac83f9d6107..c1d2a741e1af 100644
> > --- a/fs/xfs/xfs_reflink.c
> > +++ b/fs/xfs/xfs_reflink.c
> > @@ -628,6 +628,11 @@ xfs_reflink_end_cow_extent(
> > xfs_ilock(ip, XFS_ILOCK_EXCL);
> > xfs_trans_ijoin(tp, ip, 0);
> >
> > + error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
> > + XFS_IEXT_REFLINK_END_COW_CNT);
> > + if (error)
> > + goto out_cancel;
>
> What happens if we fail here? I think for buffered writes this means
> that writeback fails and we store an EIO in the address space for
> eventual return via fsync()? And for a direct write this means that
> EIO gets sent back to the caller, right?
>
Yes, you are right about that.
> Assuming I understood that correctly, I think this is a reasonable
> enough place to check for overflows, and hence
>
> Reviewed-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
>
> It would be nicer to check this kind of thing at write() time to put all
> the EFBIG errors up front, but I don't think you can do that without
> tracking extent count "reservations" incore.
>
> --D
>
> > +
> > /*
> > * In case of racing, overlapping AIO writes no COW extents might be
> > * left by the time I/O completes for the loser of the race. In that
>
--
chandan
