On Fri, Jul 17, 2020 at 08:01:56AM +1000, Dave Chinner wrote:
> On Thu, Jul 16, 2020 at 08:18:49AM -0400, Brian Foster wrote:
> > Rewrite the macro as a static inline helper to clean up the logic
> > and have one less macro.
> >
> > Signed-off-by: Brian Foster <bfoster@xxxxxxxxxx>
> > ---
> > fs/xfs/libxfs/xfs_trans_space.h | 24 ++++++++++++++++--------
> > fs/xfs/xfs_inode.c | 4 ++--
> > fs/xfs/xfs_symlink.c | 2 +-
> > 3 files changed, 19 insertions(+), 11 deletions(-)
> >
> > diff --git a/fs/xfs/libxfs/xfs_trans_space.h b/fs/xfs/libxfs/xfs_trans_space.h
> > index c6df01a2a158..d08dfc8795c3 100644
> > --- a/fs/xfs/libxfs/xfs_trans_space.h
> > +++ b/fs/xfs/libxfs/xfs_trans_space.h
> > @@ -55,10 +55,18 @@
> > XFS_DIRENTER_MAX_SPLIT(mp,nl))
> > #define XFS_DIRREMOVE_SPACE_RES(mp) \
> > XFS_DAREMOVE_SPACE_RES(mp, XFS_DATA_FORK)
> > -#define XFS_IALLOC_SPACE_RES(mp) \
> > - (M_IGEO(mp)->ialloc_blks + \
> > - ((xfs_sb_version_hasfinobt(&mp->m_sb) ? 2 : 1) * \
> > - (M_IGEO(mp)->inobt_maxlevels - 1)))
> > +
> > +static inline int
> > +xfs_ialloc_space_res(
> > + struct xfs_mount *mp)
> > +{
> > + int res = M_IGEO(mp)->ialloc_blks;
> > +
> > + res += M_IGEO(mp)->inobt_maxlevels - 1;
> > + if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > + res += M_IGEO(mp)->inobt_maxlevels - 1;
> > + return res;
> > +}
>
> This misses the point I made. i.e. that the space reservation is
> constant and never changes, yet we calculate it -twice- per inode
> create. That means we can be calculating it hundreds of thousands of
> times a second instead of just reading a variable that is likely hot
> in cache.
>
Partly.. I mentioned in my earlier reply that the geometry structure
doesn't seem like the right place to stuff transaction reservation
related values. An alternative I mentioned would be a new (or update to
the existing) structure that similarly precalculates log reservations,
but I wasn't going to go do that without some discussion/feedback on the
approach. Replacing the macro seemed broadly acceptable, so I sent that
patch as an incremental improvement and to keep the bug fix isolated.
Also, while stuffing a new value in a pre-existing structure might lend
itself to a one-off patch, I'm not sure that creating a new data
structure does lest it fail to justify the existence of the structure
itself. Therefore, it might be better to create a small series to
convert over several values to start such a structure and perhaps do the
rest over time to reduce the churn..
> IOWs, if we are going to improve this code, it should to be moved to
> a pre-calculated, read-only, per-mount variable so the repeated
> calculation goes away entirely.
>
I figured replacing the macro was an incremental improvement independent
from how precalculated transaction block reservations might be
structured. TBH, I don't mind the macros as much as others seem to, so
feel free to defer or discard this patch altogether..
> Then the macro/function goes away entirely an is replaced simply
> by mp->m_ialloc_space_res or M_IGEO(mp)->alloc_space_res....
>
Feel free to comment on my feedback on this in the previous reply..
Brian
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@xxxxxxxxxxxxx
>
