Re: [PATCH 5/9] xfs_repair: rebuild inode btrees with bulk loader

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

On Thu, May 28, 2020 at 11:11:21AM -0400, Brian Foster wrote:
> On Tue, May 19, 2020 at 06:51:15PM -0700, Darrick J. Wong wrote:
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > Use the btree bulk loading functions to rebuild the inode btrees
> > and drop the open-coded implementation.
> > 
> > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > ---
> >  libxfs/libxfs_api_defs.h |    1 
> >  repair/phase5.c          |  642 +++++++++++++++++-----------------------------
> >  2 files changed, 240 insertions(+), 403 deletions(-)
> > 
> > 
> ...
> > diff --git a/repair/phase5.c b/repair/phase5.c
> > index e69b042c..38f30753 100644
> > --- a/repair/phase5.c
> > +++ b/repair/phase5.c
> ...
> > @@ -372,6 +376,11 @@ estimate_ag_bload_slack(
> >  		bload->node_slack = 0;
> >  }
> >  
> > +static inline void skip_rebuild(struct bt_rebuild *btr)
> > +{
> > +	memset(btr, 0, sizeof(struct bt_rebuild));
> > +}
> > +
> 
> Is there any functional purpose to this? It looks like the memset could
> be open-coded, but also seems like it could be elided if we just check
> hasfinobt() before using the pointer..?

Hm, yeah, ok, I'll go audit this.

> >  /* Initialize a btree rebuild context. */
> >  static void
> >  init_rebuild(
> > @@ -765,48 +774,38 @@ _("Error %d while writing cntbt btree for AG %u.\n"), error, agno);
> ...
> > +
> > +/* Copy one incore inode record into the inobt cursor. */
> > +static void
> > +get_inode_data(
> > +	struct xfs_btree_cur		*cur,
> > +	struct ino_tree_node		*ino_rec,
> > +	struct agi_stat			*agi_stat)
> > +{
> > +	struct xfs_inobt_rec_incore	*irec = &cur->bc_rec.i;
> > +	int				inocnt = 0;
> > +	int				finocnt = 0;
> > +	int				k;
> > +
> > +	irec->ir_startino = ino_rec->ino_startnum;
> > +	irec->ir_free = ino_rec->ir_free;
> > +
> > +	for (k = 0; k < sizeof(xfs_inofree_t) * NBBY; k++)  {
> > +		ASSERT(is_inode_confirmed(ino_rec, k));
> > +
> > +		if (is_inode_sparse(ino_rec, k))
> >  			continue;
> > -
> > -		nfinos += rec_nfinos;
> > -		ninos += rec_ninos;
> > -		num_recs++;
> > +		if (is_inode_free(ino_rec, k))
> > +			finocnt++;
> > +		inocnt++;
> >  	}
> >  
> > -	if (num_recs == 0) {
> > -		/*
> > -		 * easy corner-case -- no inode records
> > -		 */
> > -		lptr->num_blocks = 1;
> > -		lptr->modulo = 0;
> > -		lptr->num_recs_pb = 0;
> > -		lptr->num_recs_tot = 0;
> > -
> > -		btree_curs->num_levels = 1;
> > -		btree_curs->num_tot_blocks = btree_curs->num_free_blocks = 1;
> > -
> > -		setup_cursor(mp, agno, btree_curs);
> > +	irec->ir_count = inocnt;
> > +	irec->ir_freecount = finocnt;
> >  
> > -		return;
> > -	}
> > +	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
> > +		uint64_t		sparse;
> > +		int			spmask;
> > +		uint16_t		holemask;
> >  
> > -	blocks_allocated = lptr->num_blocks = howmany(num_recs,
> > -					XR_INOBT_BLOCK_MAXRECS(mp, 0));
> > -
> > -	lptr->modulo = num_recs % lptr->num_blocks;
> > -	lptr->num_recs_pb = num_recs / lptr->num_blocks;
> > -	lptr->num_recs_tot = num_recs;
> > -	level = 1;
> > -
> > -	if (lptr->num_blocks > 1)  {
> > -		for (; btree_curs->level[level-1].num_blocks > 1
> > -				&& level < XFS_BTREE_MAXLEVELS;
> > -				level++)  {
> > -			lptr = &btree_curs->level[level];
> > -			p_lptr = &btree_curs->level[level - 1];
> > -			lptr->num_blocks = howmany(p_lptr->num_blocks,
> > -				XR_INOBT_BLOCK_MAXRECS(mp, level));
> > -			lptr->modulo = p_lptr->num_blocks % lptr->num_blocks;
> > -			lptr->num_recs_pb = p_lptr->num_blocks
> > -					/ lptr->num_blocks;
> > -			lptr->num_recs_tot = p_lptr->num_blocks;
> > -
> > -			blocks_allocated += lptr->num_blocks;
> > +		/*
> > +		 * Convert the 64-bit in-core sparse inode state to the
> > +		 * 16-bit on-disk holemask.
> > +		 */
> > +		holemask = 0;
> > +		spmask = (1 << XFS_INODES_PER_HOLEMASK_BIT) - 1;
> > +		sparse = ino_rec->ir_sparse;
> > +		for (k = 0; k < XFS_INOBT_HOLEMASK_BITS; k++) {
> > +			if (sparse & spmask) {
> > +				ASSERT((sparse & spmask) == spmask);
> > +				holemask |= (1 << k);
> > +			} else
> > +				ASSERT((sparse & spmask) == 0);
> > +			sparse >>= XFS_INODES_PER_HOLEMASK_BIT;
> >  		}
> > +
> > +		irec->ir_holemask = holemask;
> > +	} else {
> > +		irec->ir_holemask = 0;
> >  	}
> > -	ASSERT(lptr->num_blocks == 1);
> > -	btree_curs->num_levels = level;
> >  
> > -	btree_curs->num_tot_blocks = btree_curs->num_free_blocks
> > -			= blocks_allocated;
> > +	if (!agi_stat)
> > +		return;
> >  
> > -	setup_cursor(mp, agno, btree_curs);
> > +	if (agi_stat->first_agino != NULLAGINO)
> > +		agi_stat->first_agino = ino_rec->ino_startnum;
> 
> This is initialized to NULLAGINO. When do we ever update it?

I don't understand your question.  The purpose of the first_agino code
is to set agi_newino to the first inode cluster in the AG (or NULLAGINO
if there are no inodes) so we initialize first_ino to NULLAGINO and if
we process any inode records, we'll update it to ir_startino of the
first inode record that we put in the btree.

> > +	agi_stat->freecount += finocnt;
> > +	agi_stat->count += inocnt;
> > +}
> >  
> > -	*num_inos = ninos;
> > -	*num_free_inos = nfinos;
> > +/* Grab one inobt record. */
> > +static int
> > +get_inobt_record(
> > +	struct xfs_btree_cur		*cur,
> > +	void				*priv)
> > +{
> > +	struct bt_rebuild		*rebuild = priv;
> >  
> > -	return;
> > +	get_inode_data(cur, rebuild->ino_rec, rebuild->agi_stat);
> > +	rebuild->ino_rec = next_ino_rec(rebuild->ino_rec);
> > +	return 0;
> >  }
> >  
> > +/* Rebuild a inobt btree. */
> >  static void
> > -prop_ino_cursor(xfs_mount_t *mp, xfs_agnumber_t agno, bt_status_t *btree_curs,
> > -	xfs_btnum_t btnum, xfs_agino_t startino, int level)
> > +build_inobt(
> > +	struct repair_ctx	*sc,
> > +	xfs_agnumber_t		agno,
> > +	struct bt_rebuild	*btr_ino,
> > +	struct agi_stat		*agi_stat)
> >  {
> > -	struct xfs_btree_block	*bt_hdr;
> > -	xfs_inobt_key_t		*bt_key;
> > -	xfs_inobt_ptr_t		*bt_ptr;
> > -	xfs_agblock_t		agbno;
> > -	bt_stat_level_t		*lptr;
> > -	const struct xfs_buf_ops *ops = btnum_to_ops(btnum);
> >  	int			error;
> >  
> > -	level++;
> > -
> > -	if (level >= btree_curs->num_levels)
> > -		return;
> > -
> > -	lptr = &btree_curs->level[level];
> > -	bt_hdr = XFS_BUF_TO_BLOCK(lptr->buf_p);
> > -
> > -	if (be16_to_cpu(bt_hdr->bb_numrecs) == 0)  {
> > -		/*
> > -		 * this only happens once to initialize the
> > -		 * first path up the left side of the tree
> > -		 * where the agbno's are already set up
> > -		 */
> > -		prop_ino_cursor(mp, agno, btree_curs, btnum, startino, level);
> > -	}
> > -
> > -	if (be16_to_cpu(bt_hdr->bb_numrecs) ==
> > -				lptr->num_recs_pb + (lptr->modulo > 0))  {
> > -		/*
> > -		 * write out current prev block, grab us a new block,
> > -		 * and set the rightsib pointer of current block
> > -		 */
> > -#ifdef XR_BLD_INO_TRACE
> > -		fprintf(stderr, " ino prop agbno %d ", lptr->prev_agbno);
> > -#endif
> > -		if (lptr->prev_agbno != NULLAGBLOCK)  {
> > -			ASSERT(lptr->prev_buf_p != NULL);
> > -			libxfs_buf_mark_dirty(lptr->prev_buf_p);
> > -			libxfs_buf_relse(lptr->prev_buf_p);
> > -		}
> > -		lptr->prev_agbno = lptr->agbno;;
> > -		lptr->prev_buf_p = lptr->buf_p;
> > -		agbno = get_next_blockaddr(agno, level, btree_curs);
> > -
> > -		bt_hdr->bb_u.s.bb_rightsib = cpu_to_be32(agbno);
> > -
> > -		error = -libxfs_buf_get(mp->m_dev,
> > -				XFS_AGB_TO_DADDR(mp, agno, agbno),
> > -				XFS_FSB_TO_BB(mp, 1), &lptr->buf_p);
> > -		if (error)
> > -			do_error(_("Cannot grab inode btree buffer, err=%d"),
> > -					error);
> > -		lptr->agbno = agbno;
> > +	btr_ino->bload.get_record = get_inobt_record;
> > +	btr_ino->bload.claim_block = rebuild_claim_block;
> > +	agi_stat->count = agi_stat->freecount = 0;
> 
> These are already initialized to zero by the caller. I suppose we might
> as well also move the ->first_agino init to where this is allocated.

TBH we don't need to (re)init agi_stat at all.  I'll drop these two
lines.

> Otherwise I mostly just have the same general feedback as for the
> previous patch (wrt to the get_record() logic and build_[f]inobt()
> duplication.

<nod> Will fix that too.

> Brian
> 
> > +	agi_stat->first_agino = NULLAGINO;
> > +	btr_ino->agi_stat = agi_stat;
> > +	btr_ino->ino_rec = findfirst_inode_rec(agno);
> > +
> > +	error = -libxfs_trans_alloc_empty(sc->mp, &sc->tp);
> > +	if (error)
> > +		do_error(
> > +_("Insufficient memory to construct inobt rebuild transaction.\n"));
> > +
> > +	/* Add all observed inobt records. */
> > +	error = -libxfs_btree_bload(btr_ino->cur, &btr_ino->bload, btr_ino);
> > +	if (error)
> > +		do_error(
> > +_("Error %d while creating inobt btree for AG %u.\n"), error, agno);
> > +
> > +	/* Since we're not writing the AGI yet, no need to commit the cursor */
> > +	libxfs_btree_del_cursor(btr_ino->cur, 0);
> > +	error = -libxfs_trans_commit(sc->tp);
> > +	if (error)
> > +		do_error(
> > +_("Error %d while writing inobt btree for AG %u.\n"), error, agno);
> > +	sc->tp = NULL;
> > +}
> >  
> > -		if (lptr->modulo)
> > -			lptr->modulo--;
> > +/* Grab one finobt record. */
> > +static int
> > +get_finobt_record(
> > +	struct xfs_btree_cur		*cur,
> > +	void				*priv)
> > +{
> > +	struct bt_rebuild		*rebuild = priv;
> >  
> > -		/*
> > -		 * initialize block header
> > -		 */
> > -		lptr->buf_p->b_ops = ops;
> > -		bt_hdr = XFS_BUF_TO_BLOCK(lptr->buf_p);
> > -		memset(bt_hdr, 0, mp->m_sb.sb_blocksize);
> > -		libxfs_btree_init_block(mp, lptr->buf_p, btnum,
> > -					level, 0, agno);
> > +	get_inode_data(cur, rebuild->ino_rec, NULL);
> > +	rebuild->ino_rec = next_free_ino_rec(rebuild->ino_rec);
> > +	return 0;
> > +}
> >  
> > -		bt_hdr->bb_u.s.bb_leftsib = cpu_to_be32(lptr->prev_agbno);
> > +/* Rebuild a finobt btree. */
> > +static void
> > +build_finobt(
> > +	struct repair_ctx	*sc,
> > +	xfs_agnumber_t		agno,
> > +	struct bt_rebuild	*btr_fino)
> > +{
> > +	int			error;
> >  
> > -		/*
> > -		 * propagate extent record for first extent in new block up
> > -		 */
> > -		prop_ino_cursor(mp, agno, btree_curs, btnum, startino, level);
> > -	}
> > -	/*
> > -	 * add inode info to current block
> > -	 */
> > -	be16_add_cpu(&bt_hdr->bb_numrecs, 1);
> > -
> > -	bt_key = XFS_INOBT_KEY_ADDR(mp, bt_hdr,
> > -				    be16_to_cpu(bt_hdr->bb_numrecs));
> > -	bt_ptr = XFS_INOBT_PTR_ADDR(mp, bt_hdr,
> > -				    be16_to_cpu(bt_hdr->bb_numrecs),
> > -				    M_IGEO(mp)->inobt_mxr[1]);
> > -
> > -	bt_key->ir_startino = cpu_to_be32(startino);
> > -	*bt_ptr = cpu_to_be32(btree_curs->level[level-1].agbno);
> > +	btr_fino->bload.get_record = get_finobt_record;
> > +	btr_fino->bload.claim_block = rebuild_claim_block;
> > +	btr_fino->ino_rec = findfirst_free_inode_rec(agno);
> > +
> > +	error = -libxfs_trans_alloc_empty(sc->mp, &sc->tp);
> > +	if (error)
> > +		do_error(
> > +_("Insufficient memory to construct finobt rebuild transaction.\n"));
> > +
> > +	/* Add all observed finobt records. */
> > +	error = -libxfs_btree_bload(btr_fino->cur, &btr_fino->bload, btr_fino);
> > +	if (error)
> > +		do_error(
> > +_("Error %d while creating finobt btree for AG %u.\n"), error, agno);
> > +
> > +	/* Since we're not writing the AGI yet, no need to commit the cursor */
> > +	libxfs_btree_del_cursor(btr_fino->cur, 0);
> > +	error = -libxfs_trans_commit(sc->tp);
> > +	if (error)
> > +		do_error(
> > +_("Error %d while writing finobt btree for AG %u.\n"), error, agno);
> > +	sc->tp = NULL;
> >  }
> >  
> >  /*
> >   * XXX: yet more code that can be shared with mkfs, growfs.
> >   */
> >  static void
> > -build_agi(xfs_mount_t *mp, xfs_agnumber_t agno, bt_status_t *btree_curs,
> > -		bt_status_t *finobt_curs, struct agi_stat *agi_stat)
> > +build_agi(
> > +	struct xfs_mount	*mp,
> > +	xfs_agnumber_t		agno,
> > +	struct bt_rebuild	*btr_ino,
> > +	struct bt_rebuild	*btr_fino,
> > +	struct agi_stat		*agi_stat)
> >  {
> > -	xfs_buf_t	*agi_buf;
> > -	xfs_agi_t	*agi;
> > -	int		i;
> > -	int		error;
> > +	struct xfs_buf		*agi_buf;
> > +	struct xfs_agi		*agi;
> > +	int			i;
> > +	int			error;
> >  
> >  	error = -libxfs_buf_get(mp->m_dev,
> >  			XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
> > @@ -1008,8 +1053,8 @@ build_agi(xfs_mount_t *mp, xfs_agnumber_t agno, bt_status_t *btree_curs,
> >  		agi->agi_length = cpu_to_be32(mp->m_sb.sb_dblocks -
> >  			(xfs_rfsblock_t) mp->m_sb.sb_agblocks * agno);
> >  	agi->agi_count = cpu_to_be32(agi_stat->count);
> > -	agi->agi_root = cpu_to_be32(btree_curs->root);
> > -	agi->agi_level = cpu_to_be32(btree_curs->num_levels);
> > +	agi->agi_root = cpu_to_be32(btr_ino->newbt.afake.af_root);
> > +	agi->agi_level = cpu_to_be32(btr_ino->newbt.afake.af_levels);
> >  	agi->agi_freecount = cpu_to_be32(agi_stat->freecount);
> >  	agi->agi_newino = cpu_to_be32(agi_stat->first_agino);
> >  	agi->agi_dirino = cpu_to_be32(NULLAGINO);
> > @@ -1021,203 +1066,16 @@ build_agi(xfs_mount_t *mp, xfs_agnumber_t agno, bt_status_t *btree_curs,
> >  		platform_uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
> >  
> >  	if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
> > -		agi->agi_free_root = cpu_to_be32(finobt_curs->root);
> > -		agi->agi_free_level = cpu_to_be32(finobt_curs->num_levels);
> > +		agi->agi_free_root =
> > +				cpu_to_be32(btr_fino->newbt.afake.af_root);
> > +		agi->agi_free_level =
> > +				cpu_to_be32(btr_fino->newbt.afake.af_levels);
> >  	}
> >  
> >  	libxfs_buf_mark_dirty(agi_buf);
> >  	libxfs_buf_relse(agi_buf);
> >  }
> >  
> > -/*
> > - * rebuilds an inode tree given a cursor.  We're lazy here and call
> > - * the routine that builds the agi
> > - */
> > -static void
> > -build_ino_tree(xfs_mount_t *mp, xfs_agnumber_t agno,
> > -		bt_status_t *btree_curs, xfs_btnum_t btnum,
> > -		struct agi_stat *agi_stat)
> > -{
> > -	xfs_agnumber_t		i;
> > -	xfs_agblock_t		j;
> > -	xfs_agblock_t		agbno;
> > -	xfs_agino_t		first_agino;
> > -	struct xfs_btree_block	*bt_hdr;
> > -	xfs_inobt_rec_t		*bt_rec;
> > -	ino_tree_node_t		*ino_rec;
> > -	bt_stat_level_t		*lptr;
> > -	const struct xfs_buf_ops *ops = btnum_to_ops(btnum);
> > -	xfs_agino_t		count = 0;
> > -	xfs_agino_t		freecount = 0;
> > -	int			inocnt;
> > -	uint8_t			finocnt;
> > -	int			k;
> > -	int			level = btree_curs->num_levels;
> > -	int			spmask;
> > -	uint64_t		sparse;
> > -	uint16_t		holemask;
> > -	int			error;
> > -
> > -	ASSERT(btnum == XFS_BTNUM_INO || btnum == XFS_BTNUM_FINO);
> > -
> > -	for (i = 0; i < level; i++)  {
> > -		lptr = &btree_curs->level[i];
> > -
> > -		agbno = get_next_blockaddr(agno, i, btree_curs);
> > -		error = -libxfs_buf_get(mp->m_dev,
> > -				XFS_AGB_TO_DADDR(mp, agno, agbno),
> > -				XFS_FSB_TO_BB(mp, 1), &lptr->buf_p);
> > -		if (error)
> > -			do_error(_("Cannot grab inode btree buffer, err=%d"),
> > -					error);
> > -
> > -		if (i == btree_curs->num_levels - 1)
> > -			btree_curs->root = agbno;
> > -
> > -		lptr->agbno = agbno;
> > -		lptr->prev_agbno = NULLAGBLOCK;
> > -		lptr->prev_buf_p = NULL;
> > -		/*
> > -		 * initialize block header
> > -		 */
> > -
> > -		lptr->buf_p->b_ops = ops;
> > -		bt_hdr = XFS_BUF_TO_BLOCK(lptr->buf_p);
> > -		memset(bt_hdr, 0, mp->m_sb.sb_blocksize);
> > -		libxfs_btree_init_block(mp, lptr->buf_p, btnum, i, 0, agno);
> > -	}
> > -
> > -	/*
> > -	 * run along leaf, setting up records.  as we have to switch
> > -	 * blocks, call the prop_ino_cursor routine to set up the new
> > -	 * pointers for the parent.  that can recurse up to the root
> > -	 * if required.  set the sibling pointers for leaf level here.
> > -	 */
> > -	if (btnum == XFS_BTNUM_FINO)
> > -		ino_rec = findfirst_free_inode_rec(agno);
> > -	else
> > -		ino_rec = findfirst_inode_rec(agno);
> > -
> > -	if (ino_rec != NULL)
> > -		first_agino = ino_rec->ino_startnum;
> > -	else
> > -		first_agino = NULLAGINO;
> > -
> > -	lptr = &btree_curs->level[0];
> > -
> > -	for (i = 0; i < lptr->num_blocks; i++)  {
> > -		/*
> > -		 * block initialization, lay in block header
> > -		 */
> > -		lptr->buf_p->b_ops = ops;
> > -		bt_hdr = XFS_BUF_TO_BLOCK(lptr->buf_p);
> > -		memset(bt_hdr, 0, mp->m_sb.sb_blocksize);
> > -		libxfs_btree_init_block(mp, lptr->buf_p, btnum, 0, 0, agno);
> > -
> > -		bt_hdr->bb_u.s.bb_leftsib = cpu_to_be32(lptr->prev_agbno);
> > -		bt_hdr->bb_numrecs = cpu_to_be16(lptr->num_recs_pb +
> > -							(lptr->modulo > 0));
> > -
> > -		if (lptr->modulo > 0)
> > -			lptr->modulo--;
> > -
> > -		if (lptr->num_recs_pb > 0)
> > -			prop_ino_cursor(mp, agno, btree_curs, btnum,
> > -					ino_rec->ino_startnum, 0);
> > -
> > -		bt_rec = (xfs_inobt_rec_t *)
> > -			  ((char *)bt_hdr + XFS_INOBT_BLOCK_LEN(mp));
> > -		for (j = 0; j < be16_to_cpu(bt_hdr->bb_numrecs); j++) {
> > -			ASSERT(ino_rec != NULL);
> > -			bt_rec[j].ir_startino =
> > -					cpu_to_be32(ino_rec->ino_startnum);
> > -			bt_rec[j].ir_free = cpu_to_be64(ino_rec->ir_free);
> > -
> > -			inocnt = finocnt = 0;
> > -			for (k = 0; k < sizeof(xfs_inofree_t)*NBBY; k++)  {
> > -				ASSERT(is_inode_confirmed(ino_rec, k));
> > -
> > -				if (is_inode_sparse(ino_rec, k))
> > -					continue;
> > -				if (is_inode_free(ino_rec, k))
> > -					finocnt++;
> > -				inocnt++;
> > -			}
> > -
> > -			/*
> > -			 * Set the freecount and check whether we need to update
> > -			 * the sparse format fields. Otherwise, skip to the next
> > -			 * record.
> > -			 */
> > -			inorec_set_freecount(mp, &bt_rec[j], finocnt);
> > -			if (!xfs_sb_version_hassparseinodes(&mp->m_sb))
> > -				goto nextrec;
> > -
> > -			/*
> > -			 * Convert the 64-bit in-core sparse inode state to the
> > -			 * 16-bit on-disk holemask.
> > -			 */
> > -			holemask = 0;
> > -			spmask = (1 << XFS_INODES_PER_HOLEMASK_BIT) - 1;
> > -			sparse = ino_rec->ir_sparse;
> > -			for (k = 0; k < XFS_INOBT_HOLEMASK_BITS; k++) {
> > -				if (sparse & spmask) {
> > -					ASSERT((sparse & spmask) == spmask);
> > -					holemask |= (1 << k);
> > -				} else
> > -					ASSERT((sparse & spmask) == 0);
> > -				sparse >>= XFS_INODES_PER_HOLEMASK_BIT;
> > -			}
> > -
> > -			bt_rec[j].ir_u.sp.ir_count = inocnt;
> > -			bt_rec[j].ir_u.sp.ir_holemask = cpu_to_be16(holemask);
> > -
> > -nextrec:
> > -			freecount += finocnt;
> > -			count += inocnt;
> > -
> > -			if (btnum == XFS_BTNUM_FINO)
> > -				ino_rec = next_free_ino_rec(ino_rec);
> > -			else
> > -				ino_rec = next_ino_rec(ino_rec);
> > -		}
> > -
> > -		if (ino_rec != NULL)  {
> > -			/*
> > -			 * get next leaf level block
> > -			 */
> > -			if (lptr->prev_buf_p != NULL)  {
> > -#ifdef XR_BLD_INO_TRACE
> > -				fprintf(stderr, "writing inobt agbno %u\n",
> > -					lptr->prev_agbno);
> > -#endif
> > -				ASSERT(lptr->prev_agbno != NULLAGBLOCK);
> > -				libxfs_buf_mark_dirty(lptr->prev_buf_p);
> > -				libxfs_buf_relse(lptr->prev_buf_p);
> > -			}
> > -			lptr->prev_buf_p = lptr->buf_p;
> > -			lptr->prev_agbno = lptr->agbno;
> > -			lptr->agbno = get_next_blockaddr(agno, 0, btree_curs);
> > -			bt_hdr->bb_u.s.bb_rightsib = cpu_to_be32(lptr->agbno);
> > -
> > -			error = -libxfs_buf_get(mp->m_dev,
> > -					XFS_AGB_TO_DADDR(mp, agno, lptr->agbno),
> > -					XFS_FSB_TO_BB(mp, 1),
> > -					&lptr->buf_p);
> > -			if (error)
> > -				do_error(
> > -	_("Cannot grab inode btree buffer, err=%d"),
> > -						error);
> > -		}
> > -	}
> > -
> > -	if (agi_stat) {
> > -		agi_stat->first_agino = first_agino;
> > -		agi_stat->count = count;
> > -		agi_stat->freecount = freecount;
> > -	}
> > -}
> > -
> >  /* rebuild the rmap tree */
> >  
> >  /*
> > @@ -2142,14 +2000,10 @@ phase5_func(
> >  {
> >  	struct repair_ctx	sc = { .mp = mp, };
> >  	struct agi_stat		agi_stat = {0,};
> > -	uint64_t		num_inos;
> > -	uint64_t		num_free_inos;
> > -	uint64_t		finobt_num_inos;
> > -	uint64_t		finobt_num_free_inos;
> >  	struct bt_rebuild	btr_bno;
> >  	struct bt_rebuild	btr_cnt;
> > -	bt_status_t		ino_btree_curs;
> > -	bt_status_t		fino_btree_curs;
> > +	struct bt_rebuild	btr_ino;
> > +	struct bt_rebuild	btr_fino;
> >  	bt_status_t		rmap_btree_curs;
> >  	bt_status_t		refcnt_btree_curs;
> >  	int			extra_blocks = 0;
> > @@ -2184,19 +2038,8 @@ _("unable to rebuild AG %u.  Not enough free space in on-disk AG.\n"),
> >  			agno);
> >  	}
> >  
> > -	/*
> > -	 * ok, now set up the btree cursors for the on-disk btrees (includes
> > -	 * pre-allocating all required blocks for the trees themselves)
> > -	 */
> > -	init_ino_cursor(mp, agno, &ino_btree_curs, &num_inos,
> > -			&num_free_inos, 0);
> > -
> > -	if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > -		init_ino_cursor(mp, agno, &fino_btree_curs, &finobt_num_inos,
> > -				&finobt_num_free_inos, 1);
> > -
> > -	sb_icount_ag[agno] += num_inos;
> > -	sb_ifree_ag[agno] += num_free_inos;
> > +	init_ino_cursors(&sc, agno, num_freeblocks, &sb_icount_ag[agno],
> > +			&sb_ifree_ag[agno], &btr_ino, &btr_fino);
> >  
> >  	/*
> >  	 * Set up the btree cursors for the on-disk rmap btrees, which includes
> > @@ -2287,34 +2130,27 @@ _("unable to rebuild AG %u.  Not enough free space in on-disk AG.\n"),
> >  			&rmap_btree_curs, &refcnt_btree_curs, lost_fsb);
> >  
> >  	/*
> > -	 * build inode allocation tree.
> > +	 * build inode allocation trees.
> >  	 */
> > -	build_ino_tree(mp, agno, &ino_btree_curs, XFS_BTNUM_INO, &agi_stat);
> > -	write_cursor(&ino_btree_curs);
> > -
> > -	/*
> > -	 * build free inode tree
> > -	 */
> > -	if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
> > -		build_ino_tree(mp, agno, &fino_btree_curs,
> > -				XFS_BTNUM_FINO, NULL);
> > -		write_cursor(&fino_btree_curs);
> > -	}
> > +	build_inobt(&sc, agno, &btr_ino, &agi_stat);
> > +	if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > +		build_finobt(&sc, agno, &btr_fino);
> >  
> >  	/* build the agi */
> > -	build_agi(mp, agno, &ino_btree_curs, &fino_btree_curs, &agi_stat);
> > +	build_agi(mp, agno, &btr_ino, &btr_fino, &agi_stat);
> >  
> >  	/*
> >  	 * tear down cursors
> >  	 */
> >  	finish_rebuild(mp, &btr_bno, lost_fsb);
> >  	finish_rebuild(mp, &btr_cnt, lost_fsb);
> > +	finish_rebuild(mp, &btr_ino, lost_fsb);
> > +	if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > +		finish_rebuild(mp, &btr_fino, lost_fsb);
> >  	if (xfs_sb_version_hasrmapbt(&mp->m_sb))
> >  		finish_cursor(&rmap_btree_curs);
> >  	if (xfs_sb_version_hasreflink(&mp->m_sb))
> >  		finish_cursor(&refcnt_btree_curs);
> > -	if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > -		finish_cursor(&fino_btree_curs);
> >  
> >  	/*
> >  	 * release the incore per-AG bno/bcnt trees so the extent nodes
> > 
>