Re: [PATCH 012/119] xfs: during btree split, save new block key & ptr for future insertion

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Jun 21, 2016 at 09:00:45AM -0400, Brian Foster wrote:
> On Thu, Jun 16, 2016 at 06:19:08PM -0700, Darrick J. Wong wrote:
> > When a btree block has to be split, we pass the new block's ptr from
> > xfs_btree_split() back to xfs_btree_insert() via a pointer parameter;
> > however, we pass the block's key through the cursor's record.  It is a
> > little weird to "initialize" a record from a key since the non-key
> > attributes will have garbage values.
> > 
> > When we go to add support for interval queries, we have to be able to
> > pass the lowest and highest keys accessible via a pointer.  There's no
> > clean way to pass this back through the cursor's record field.
> > Therefore, pass the key directly back to xfs_btree_insert() the same
> > way that we pass the btree_ptr.
> > 
> > As a bonus, we no longer need init_rec_from_key and can drop it from the
> > codebase.
> > 
> > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > ---
> >  fs/xfs/libxfs/xfs_alloc_btree.c  |   12 ----------
> >  fs/xfs/libxfs/xfs_bmap_btree.c   |   12 ----------
> >  fs/xfs/libxfs/xfs_btree.c        |   44 +++++++++++++++++++-------------------
> >  fs/xfs/libxfs/xfs_btree.h        |    2 --
> >  fs/xfs/libxfs/xfs_ialloc_btree.c |   10 ---------
> >  5 files changed, 22 insertions(+), 58 deletions(-)
> > 
> > 
> ...
> > diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
> > index 046fbcf..a096539 100644
> > --- a/fs/xfs/libxfs/xfs_btree.c
> > +++ b/fs/xfs/libxfs/xfs_btree.c
> ...
> > @@ -2929,16 +2927,16 @@ xfs_btree_insrec(
> >  	struct xfs_btree_cur	*cur,	/* btree cursor */
> >  	int			level,	/* level to insert record at */
> >  	union xfs_btree_ptr	*ptrp,	/* i/o: block number inserted */
> > -	union xfs_btree_rec	*recp,	/* i/o: record data inserted */
> > +	union xfs_btree_key	*key,	/* i/o: block key for ptrp */
> >  	struct xfs_btree_cur	**curp,	/* output: new cursor replacing cur */
> >  	int			*stat)	/* success/failure */
> >  {
> >  	struct xfs_btree_block	*block;	/* btree block */
> >  	struct xfs_buf		*bp;	/* buffer for block */
> > -	union xfs_btree_key	key;	/* btree key */
> >  	union xfs_btree_ptr	nptr;	/* new block ptr */
> >  	struct xfs_btree_cur	*ncur;	/* new btree cursor */
> > -	union xfs_btree_rec	nrec;	/* new record count */
> > +	union xfs_btree_key	nkey;	/* new block key */
> > +	union xfs_btree_rec	rec;	/* record to insert */
> >  	int			optr;	/* old key/record index */
> >  	int			ptr;	/* key/record index */
> >  	int			numrecs;/* number of records */
> > @@ -2947,8 +2945,14 @@ xfs_btree_insrec(
> >  	int			i;
> >  #endif
> >  
> > +	/* Make a key out of the record data to be inserted, and save it. */
> > +	if (level == 0) {
> > +		cur->bc_ops->init_rec_from_cur(cur, &rec);
> > +		cur->bc_ops->init_key_from_rec(key, &rec);
> > +	}
> 
> The level == 0 check looks a bit hacky to me. IOW, I think it's cleaner
> that the key is initialized once in the caller rather than check for a
> particular iteration down in xfs_btree_insrec(). That said,
> xfs_btree_insrec() still needs rec initialized in the level == 0 case.
> 
> I wonder if we could create an inline xfs_btree_init_key_from_cur()
> helper to combine the above calls, invoke it once in xfs_btree_insert(),
> then push down the ->init_rec_from_cur() calls to the contexts further
> down in this function where rec is actually required. There are only two
> and one of them is DEBUG code. Thoughts?

How about I make btree_insert set both &key and &rec at the start and
pass them both into btree_insrec?  That would eliminate the hacky check
above and fix the dummy tracing hook too, in case it ever does anything.

> 
> > +
> >  	XFS_BTREE_TRACE_CURSOR(cur, XBT_ENTRY);
> > -	XFS_BTREE_TRACE_ARGIPR(cur, level, *ptrp, recp);
> > +	XFS_BTREE_TRACE_ARGIPR(cur, level, *ptrp, &rec);
> >  
> 
> So these look like unimplemented dummy tracing hooks. It sounds like
> previously rec could have a junk value after a btree split, but now it
> looks like rec is junk for every non-zero level. Kind of annoying, I
> wonder if we can just kill these.. :/

<shrug> I have no opinion either way. :)

--D
> 
> Brian
> 
> >  	ncur = NULL;
> >  
> > @@ -2973,9 +2977,6 @@ xfs_btree_insrec(
> >  		return 0;
> >  	}
> >  
> > -	/* Make a key out of the record data to be inserted, and save it. */
> > -	cur->bc_ops->init_key_from_rec(&key, recp);
> > -
> >  	optr = ptr;
> >  
> >  	XFS_BTREE_STATS_INC(cur, insrec);
> > @@ -2992,10 +2993,10 @@ xfs_btree_insrec(
> >  	/* Check that the new entry is being inserted in the right place. */
> >  	if (ptr <= numrecs) {
> >  		if (level == 0) {
> > -			ASSERT(cur->bc_ops->recs_inorder(cur, recp,
> > +			ASSERT(cur->bc_ops->recs_inorder(cur, &rec,
> >  				xfs_btree_rec_addr(cur, ptr, block)));
> >  		} else {
> > -			ASSERT(cur->bc_ops->keys_inorder(cur, &key,
> > +			ASSERT(cur->bc_ops->keys_inorder(cur, key,
> >  				xfs_btree_key_addr(cur, ptr, block)));
> >  		}
> >  	}
> > @@ -3008,7 +3009,7 @@ xfs_btree_insrec(
> >  	xfs_btree_set_ptr_null(cur, &nptr);
> >  	if (numrecs == cur->bc_ops->get_maxrecs(cur, level)) {
> >  		error = xfs_btree_make_block_unfull(cur, level, numrecs,
> > -					&optr, &ptr, &nptr, &ncur, &nrec, stat);
> > +					&optr, &ptr, &nptr, &ncur, &nkey, stat);
> >  		if (error || *stat == 0)
> >  			goto error0;
> >  	}
> > @@ -3058,7 +3059,7 @@ xfs_btree_insrec(
> >  #endif
> >  
> >  		/* Now put the new data in, bump numrecs and log it. */
> > -		xfs_btree_copy_keys(cur, kp, &key, 1);
> > +		xfs_btree_copy_keys(cur, kp, key, 1);
> >  		xfs_btree_copy_ptrs(cur, pp, ptrp, 1);
> >  		numrecs++;
> >  		xfs_btree_set_numrecs(block, numrecs);
> > @@ -3079,7 +3080,7 @@ xfs_btree_insrec(
> >  		xfs_btree_shift_recs(cur, rp, 1, numrecs - ptr + 1);
> >  
> >  		/* Now put the new data in, bump numrecs and log it. */
> > -		xfs_btree_copy_recs(cur, rp, recp, 1);
> > +		xfs_btree_copy_recs(cur, rp, &rec, 1);
> >  		xfs_btree_set_numrecs(block, ++numrecs);
> >  		xfs_btree_log_recs(cur, bp, ptr, numrecs);
> >  #ifdef DEBUG
> > @@ -3095,7 +3096,7 @@ xfs_btree_insrec(
> >  
> >  	/* If we inserted at the start of a block, update the parents' keys. */
> >  	if (optr == 1) {
> > -		error = xfs_btree_updkey(cur, &key, level + 1);
> > +		error = xfs_btree_updkey(cur, key, level + 1);
> >  		if (error)
> >  			goto error0;
> >  	}
> > @@ -3105,7 +3106,7 @@ xfs_btree_insrec(
> >  	 * we are at the far right edge of the tree, update it.
> >  	 */
> >  	if (xfs_btree_is_lastrec(cur, block, level)) {
> > -		cur->bc_ops->update_lastrec(cur, block, recp,
> > +		cur->bc_ops->update_lastrec(cur, block, &rec,
> >  					    ptr, LASTREC_INSREC);
> >  	}
> >  
> > @@ -3115,7 +3116,7 @@ xfs_btree_insrec(
> >  	 */
> >  	*ptrp = nptr;
> >  	if (!xfs_btree_ptr_is_null(cur, &nptr)) {
> > -		*recp = nrec;
> > +		*key = nkey;
> >  		*curp = ncur;
> >  	}
> >  
> > @@ -3146,14 +3147,13 @@ xfs_btree_insert(
> >  	union xfs_btree_ptr	nptr;	/* new block number (split result) */
> >  	struct xfs_btree_cur	*ncur;	/* new cursor (split result) */
> >  	struct xfs_btree_cur	*pcur;	/* previous level's cursor */
> > -	union xfs_btree_rec	rec;	/* record to insert */
> > +	union xfs_btree_key	key;	/* key of block to insert */
> >  
> >  	level = 0;
> >  	ncur = NULL;
> >  	pcur = cur;
> >  
> >  	xfs_btree_set_ptr_null(cur, &nptr);
> > -	cur->bc_ops->init_rec_from_cur(cur, &rec);
> >  
> >  	/*
> >  	 * Loop going up the tree, starting at the leaf level.
> > @@ -3165,7 +3165,7 @@ xfs_btree_insert(
> >  		 * Insert nrec/nptr into this level of the tree.
> >  		 * Note if we fail, nptr will be null.
> >  		 */
> > -		error = xfs_btree_insrec(pcur, level, &nptr, &rec, &ncur, &i);
> > +		error = xfs_btree_insrec(pcur, level, &nptr, &key, &ncur, &i);
> >  		if (error) {
> >  			if (pcur != cur)
> >  				xfs_btree_del_cursor(pcur, XFS_BTREE_ERROR);
> > diff --git a/fs/xfs/libxfs/xfs_btree.h b/fs/xfs/libxfs/xfs_btree.h
> > index b955e5d..b99c018 100644
> > --- a/fs/xfs/libxfs/xfs_btree.h
> > +++ b/fs/xfs/libxfs/xfs_btree.h
> > @@ -158,8 +158,6 @@ struct xfs_btree_ops {
> >  	/* init values of btree structures */
> >  	void	(*init_key_from_rec)(union xfs_btree_key *key,
> >  				     union xfs_btree_rec *rec);
> > -	void	(*init_rec_from_key)(union xfs_btree_key *key,
> > -				     union xfs_btree_rec *rec);
> >  	void	(*init_rec_from_cur)(struct xfs_btree_cur *cur,
> >  				     union xfs_btree_rec *rec);
> >  	void	(*init_ptr_from_cur)(struct xfs_btree_cur *cur,
> > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c
> > index 89c21d7..88da2ad 100644
> > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c
> > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c
> > @@ -146,14 +146,6 @@ xfs_inobt_init_key_from_rec(
> >  }
> >  
> >  STATIC void
> > -xfs_inobt_init_rec_from_key(
> > -	union xfs_btree_key	*key,
> > -	union xfs_btree_rec	*rec)
> > -{
> > -	rec->inobt.ir_startino = key->inobt.ir_startino;
> > -}
> > -
> > -STATIC void
> >  xfs_inobt_init_rec_from_cur(
> >  	struct xfs_btree_cur	*cur,
> >  	union xfs_btree_rec	*rec)
> > @@ -314,7 +306,6 @@ static const struct xfs_btree_ops xfs_inobt_ops = {
> >  	.get_minrecs		= xfs_inobt_get_minrecs,
> >  	.get_maxrecs		= xfs_inobt_get_maxrecs,
> >  	.init_key_from_rec	= xfs_inobt_init_key_from_rec,
> > -	.init_rec_from_key	= xfs_inobt_init_rec_from_key,
> >  	.init_rec_from_cur	= xfs_inobt_init_rec_from_cur,
> >  	.init_ptr_from_cur	= xfs_inobt_init_ptr_from_cur,
> >  	.key_diff		= xfs_inobt_key_diff,
> > @@ -336,7 +327,6 @@ static const struct xfs_btree_ops xfs_finobt_ops = {
> >  	.get_minrecs		= xfs_inobt_get_minrecs,
> >  	.get_maxrecs		= xfs_inobt_get_maxrecs,
> >  	.init_key_from_rec	= xfs_inobt_init_key_from_rec,
> > -	.init_rec_from_key	= xfs_inobt_init_rec_from_key,
> >  	.init_rec_from_cur	= xfs_inobt_init_rec_from_cur,
> >  	.init_ptr_from_cur	= xfs_finobt_init_ptr_from_cur,
> >  	.key_diff		= xfs_inobt_key_diff,
> > 
> > _______________________________________________
> > xfs mailing list
> > xfs@oss.sgi.com
> > http://oss.sgi.com/mailman/listinfo/xfs