[PATCH 26/29] fsdax: shift partition offset handling into the file systems

[Dan Williams]

On Tue, Nov 9, 2021 at 12:34 AM Christoph Hellwig <hch at lst.de> wrote:
>
> Remove the last user of ->bdev in dax.c by requiring the file system to
> pass in an address that already includes the DAX offset.  As part of the
> only set ->bdev or ->daxdev when actually required in the ->iomap_begin
> methods.

Changes look good except for what looks like an argument position
fixup needed for an xfs_bmbt_to_iomap() caller below...

>
> Signed-off-by: Christoph Hellwig <hch at lst.de>
> ---
>  fs/dax.c                 |  6 +-----
>  fs/erofs/data.c          | 11 ++++++++--
>  fs/erofs/internal.h      |  1 +
>  fs/ext2/inode.c          |  8 +++++--
>  fs/ext4/inode.c          | 16 +++++++++-----
>  fs/xfs/libxfs/xfs_bmap.c |  4 ++--
>  fs/xfs/xfs_aops.c        |  2 +-
>  fs/xfs/xfs_iomap.c       | 45 +++++++++++++++++++++++++---------------
>  fs/xfs/xfs_iomap.h       |  5 +++--
>  fs/xfs/xfs_pnfs.c        |  2 +-
>  10 files changed, 63 insertions(+), 37 deletions(-)
>
> diff --git a/fs/dax.c b/fs/dax.c
> index 0bd6cdcbacfc4..2c13c681edf09 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -711,11 +711,7 @@ int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
>
>  static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
>  {
> -       phys_addr_t paddr = iomap->addr + (pos & PAGE_MASK) - iomap->offset;
> -
> -       if (iomap->bdev)
> -               paddr += (get_start_sect(iomap->bdev) << SECTOR_SHIFT);
> -       return PHYS_PFN(paddr);
> +       return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
>  }
>
>  static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
> diff --git a/fs/erofs/data.c b/fs/erofs/data.c
> index 0e35ef3f9f3d7..9b1bb177ce303 100644
> --- a/fs/erofs/data.c
> +++ b/fs/erofs/data.c
[..]
               }
> @@ -215,9 +218,13 @@ static int erofs_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
>         if (ret)
>                 return ret;
>
> -       iomap->bdev = mdev.m_bdev;
> -       iomap->dax_dev = mdev.m_daxdev;
>         iomap->offset = map.m_la;
> +       if (flags & IOMAP_DAX) {
> +               iomap->dax_dev = mdev.m_daxdev;
> +               iomap->offset += mdev.m_dax_part_off;
> +       } else {
> +               iomap->bdev = mdev.m_bdev;
> +       }

Ah, that's what IOMAP_DAX is for, to stop making iomap carry bdev
details unnecessarily.

[..]
> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> index 704292c6ce0c7..74dbf1fd99d39 100644
> --- a/fs/xfs/xfs_iomap.c
> +++ b/fs/xfs/xfs_iomap.c
> @@ -54,7 +54,8 @@ xfs_bmbt_to_iomap(
>         struct xfs_inode        *ip,
>         struct iomap            *iomap,
>         struct xfs_bmbt_irec    *imap,
> -       u16                     flags)
> +       unsigned int            flags,
> +       u16                     iomap_flags)

It would be nice if the compiler could help with making sure that
right 'flags' values are passed to the right 'flags' parameter, but I
can't think of

[..]
> @@ -1053,23 +1061,24 @@ xfs_buffered_write_iomap_begin(
>          */
>         xfs_iunlock(ip, XFS_ILOCK_EXCL);
>         trace_xfs_iomap_alloc(ip, offset, count, allocfork, &imap);
> -       return xfs_bmbt_to_iomap(ip, iomap, &imap, IOMAP_F_NEW);
> +       return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_NEW);
>
>  found_imap:
>         xfs_iunlock(ip, XFS_ILOCK_EXCL);
> -       return xfs_bmbt_to_iomap(ip, iomap, &imap, 0);
> +       return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0);

The iomap flags are supposed to be the last argument, right?