Re: [PATCH 05/15] xfs: support dynamic btree cursor heights

[Dave Chinner <david@fromorbit.com>]

On Wed, Oct 13, 2021 at 09:52:18AM -0700, Darrick J. Wong wrote:
> On Wed, Oct 13, 2021 at 04:31:22PM +1100, Dave Chinner wrote:
> > On Tue, Oct 12, 2021 at 04:33:01PM -0700, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <djwong@kernel.org>
> > > 
> > > Split out the btree level information into a separate struct and put it
> > > at the end of the cursor structure as a VLA.  The realtime rmap btree
> > > (which is rooted in an inode) will require the ability to support many
> > > more levels than a per-AG btree cursor, which means that we're going to
> > > create two btree cursor caches to conserve memory for the more common
> > > case.
> > > 
> > > Signed-off-by: Darrick J. Wong <djwong@kernel.org>
> > > Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
> > > Reviewed-by: Christoph Hellwig <hch@lst.de>
> > > ---
> > >  fs/xfs/libxfs/xfs_alloc.c |    6 +-
> > >  fs/xfs/libxfs/xfs_bmap.c  |   10 +--
> > >  fs/xfs/libxfs/xfs_btree.c |  168 +++++++++++++++++++++++----------------------
> > >  fs/xfs/libxfs/xfs_btree.h |   28 ++++++--
> > >  fs/xfs/scrub/bitmap.c     |   22 +++---
> > >  fs/xfs/scrub/bmap.c       |    2 -
> > >  fs/xfs/scrub/btree.c      |   47 +++++++------
> > >  fs/xfs/scrub/trace.c      |    7 +-
> > >  fs/xfs/scrub/trace.h      |   10 +--
> > >  fs/xfs/xfs_super.c        |    2 -
> > >  fs/xfs/xfs_trace.h        |    2 -
> > >  11 files changed, 164 insertions(+), 140 deletions(-)
> > 
> > Hmmm - subject of the patch doesn't really match the changes being
> > made - there's nothing here that makes the btree cursor heights
> > dynamic. It's just a structure layout change...
> 
> "xfs: prepare xfs_btree_cur for dynamic cursor heights" ?

*nod*

> > > @@ -922,11 +922,11 @@ xfs_btree_readahead(
> > >  	    (lev == cur->bc_nlevels - 1))
> > >  		return 0;
> > >  
> > > -	if ((cur->bc_ra[lev] | lr) == cur->bc_ra[lev])
> > > +	if ((cur->bc_levels[lev].ra | lr) == cur->bc_levels[lev].ra)
> > >  		return 0;
> > 
> > That's whacky logic. Surely that's just:
> > 
> > 	if (cur->bc_levels[lev].ra & lr)
> > 		return 0;
> 
> This is an early-exit test, which means the careful check is necessary.
> 
> If (some day) someone calls this function with (LEFTRA|RIGHTRA) to
> readahead both siblings on a btree level where one sibling has been ra'd
> but not the other, we must avoid taking the branch.

Which I didn't see any callers do, so I ignored that possibility.
Regardless, it's the use of "|" to do an additive mask match that
makes it look wierd. i.e.  the normal way of writing a multi-biti
mask match is to apply the mask and check that the returned value
matches the mask, like so:

	if ((cur->bc_levels[lev].ra & lr) == lr)
		return 0;

Really, though, this was just another "ObHuh" comment, and you don't
need to "fix" it now...

> > > @@ -242,8 +250,17 @@ struct xfs_btree_cur
> > >  		struct xfs_btree_cur_ag	bc_ag;
> > >  		struct xfs_btree_cur_ino bc_ino;
> > >  	};
> > > +
> > > +	/* Must be at the end of the struct! */
> > > +	struct xfs_btree_level	bc_levels[];
> > >  };
> > >  
> > > +static inline size_t
> > > +xfs_btree_cur_sizeof(unsigned int nlevels)
> > > +{
> > > +	return struct_size((struct xfs_btree_cur *)NULL, bc_levels, nlevels);
> > > +}
> > 
> > Ooooh, yeah, we really need comments explaining how many btree
> > levels these VLAs are tracking, because this one doesn't have a "-
> > 1" in it like the previous one I commented on....
> 
> /*
>  * Compute the size of a btree cursor that can handle a btree of a given
>  * height.  The bc_levels array handles node and leaf blocks, so its
>  * size is exactly nlevels.
>  */

Nice. Thanks!

> > > diff --git a/fs/xfs/scrub/trace.c b/fs/xfs/scrub/trace.c
> > > index c0ef53fe6611..816dfc8e5a80 100644
> > > --- a/fs/xfs/scrub/trace.c
> > > +++ b/fs/xfs/scrub/trace.c
> > > @@ -21,10 +21,11 @@ xchk_btree_cur_fsbno(
> > >  	struct xfs_btree_cur	*cur,
> > >  	int			level)
> > >  {
> > > -	if (level < cur->bc_nlevels && cur->bc_bufs[level])
> > > +	if (level < cur->bc_nlevels && cur->bc_levels[level].bp)
> > >  		return XFS_DADDR_TO_FSB(cur->bc_mp,
> > > -				xfs_buf_daddr(cur->bc_bufs[level]));
> > > -	if (level == cur->bc_nlevels - 1 && cur->bc_flags & XFS_BTREE_LONG_PTRS)
> > > +				xfs_buf_daddr(cur->bc_levels[level].bp));
> > > +	else if (level == cur->bc_nlevels - 1 &&
> > > +		 cur->bc_flags & XFS_BTREE_LONG_PTRS)
> > 
> > No need for an else there as the first if () clause returns.
> > Also, needs more () around that "a & b" second line.
> 
> TBH I think we check the wrong flag, and that last bit should be:
> 
> 	if (level == cur->bc_nlevels - 1 &&
> 	    (cur->bc_flags & XFS_BTREE_ROOT_IN_INODE))
> 		return XFS_INO_TO_FSB(cur->bc_mp, cur->bc_ino.ip->i_ino);
> 
> 	return NULLFSBLOCK;

Yup, true, long ptrs and inodes are currently interchangable so it
works, but that's a landmine waiting to pounce....

> But for now I'll stick to the straight replacement and tack on another
> patch to fix that.

*nod*.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com