Merge tag 'btree-ifork-records_2024-12-23' of https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux into for-next

xfs: refactor btrees to support records in inode root [v6.2 02/14]

Amend the btree code to support storing btree rcords in the inode root,
because the current bmbt code does not support this.

This has been running on the djcloud for months with no problems.  Enjoy!

Signed-off-by: "Darrick J. Wong" <djwong@kernel.org>

8 files changed, 445 insertions, 212 deletions

diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
index 5255f93bae31..0842577755f7 100644
--- a/fs/xfs/libxfs/xfs_bmap.c
+++ b/fs/xfs/libxfs/xfs_bmap.c
@@ -615,7 +615,7 @@ xfs_bmap_btree_to_extents(
 	xfs_trans_binval(tp, cbp);
 	if (cur->bc_levels[0].bp == cbp)
 		cur->bc_levels[0].bp = NULL;
-	xfs_iroot_realloc(ip, -1, whichfork);
+	xfs_bmap_broot_realloc(ip, whichfork, 0);
 	ASSERT(ifp->if_broot == NULL);
 	ifp->if_format = XFS_DINODE_FMT_EXTENTS;
 	*logflagsp |= XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
@@ -659,12 +659,11 @@ xfs_bmap_extents_to_btree(
 	 * Make space in the inode incore. This needs to be undone if we fail
 	 * to expand the root.
 	 */
-	xfs_iroot_realloc(ip, 1, whichfork);
+	block = xfs_bmap_broot_realloc(ip, whichfork, 1);
 
 	/*
 	 * Fill in the root.
 	 */
-	block = ifp->if_broot;
 	xfs_bmbt_init_block(ip, block, NULL, 1, 1);
 	/*
 	 * Need a cursor.  Can't allocate until bb_level is filled in.
@@ -746,7 +745,7 @@ xfs_bmap_extents_to_btree(
 out_unreserve_dquot:
 	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, -1L);
 out_root_realloc:
-	xfs_iroot_realloc(ip, -1, whichfork);
+	xfs_bmap_broot_realloc(ip, whichfork, 0);
 	ifp->if_format = XFS_DINODE_FMT_EXTENTS;
 	ASSERT(ifp->if_broot == NULL);
 	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
diff --git a/fs/xfs/libxfs/xfs_bmap_btree.c b/fs/xfs/libxfs/xfs_bmap_btree.c
index 3464be771f95..908d7b050e9c 100644
--- a/fs/xfs/libxfs/xfs_bmap_btree.c
+++ b/fs/xfs/libxfs/xfs_bmap_btree.c
@@ -516,6 +516,116 @@ xfs_bmbt_keys_contiguous(
 				 be64_to_cpu(key2->bmbt.br_startoff));
 }
 
+static inline void
+xfs_bmbt_move_ptrs(
+	struct xfs_mount	*mp,
+	struct xfs_btree_block	*broot,
+	short			old_size,
+	size_t			new_size,
+	unsigned int		numrecs)
+{
+	void			*dptr;
+	void			*sptr;
+
+	sptr = xfs_bmap_broot_ptr_addr(mp, broot, 1, old_size);
+	dptr = xfs_bmap_broot_ptr_addr(mp, broot, 1, new_size);
+	memmove(dptr, sptr, numrecs * sizeof(xfs_bmbt_ptr_t));
+}
+
+/*
+ * Reallocate the space for if_broot based on the number of records.  Move the
+ * records and pointers in if_broot to fit the new size.  When shrinking this
+ * will eliminate holes between the records and pointers created by the caller.
+ * When growing this will create holes to be filled in by the caller.
+ *
+ * The caller must not request to add more records than would fit in the
+ * on-disk inode root.  If the if_broot is currently NULL, then if we are
+ * adding records, one will be allocated.  The caller must also not request
+ * that the number of records go below zero, although it can go to zero.
+ *
+ * ip -- the inode whose if_broot area is changing
+ * whichfork -- which inode fork to change
+ * new_numrecs -- the new number of records requested for the if_broot array
+ *
+ * Returns the incore btree root block.
+ */
+struct xfs_btree_block *
+xfs_bmap_broot_realloc(
+	struct xfs_inode	*ip,
+	int			whichfork,
+	unsigned int		new_numrecs)
+{
+	struct xfs_mount	*mp = ip->i_mount;
+	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, whichfork);
+	struct xfs_btree_block	*broot;
+	unsigned int		new_size;
+	unsigned int		old_size = ifp->if_broot_bytes;
+
+	/*
+	 * Block mapping btrees do not support storing zero records; if this
+	 * happens, the fork is being changed to FMT_EXTENTS.  Free the broot
+	 * and get out.
+	 */
+	if (new_numrecs == 0)
+		return xfs_broot_realloc(ifp, 0);
+
+	new_size = xfs_bmap_broot_space_calc(mp, new_numrecs);
+
+	/* Handle the nop case quietly. */
+	if (new_size == old_size)
+		return ifp->if_broot;
+
+	if (new_size > old_size) {
+		unsigned int	old_numrecs;
+
+		/*
+		 * If there wasn't any memory allocated before, just
+		 * allocate it now and get out.
+		 */
+		if (old_size == 0)
+			return xfs_broot_realloc(ifp, new_size);
+
+		/*
+		 * If there is already an existing if_broot, then we need
+		 * to realloc() it and shift the pointers to their new
+		 * location.  The records don't change location because
+		 * they are kept butted up against the btree block header.
+		 */
+		old_numrecs = xfs_bmbt_maxrecs(mp, old_size, false);
+		broot = xfs_broot_realloc(ifp, new_size);
+		ASSERT(xfs_bmap_bmdr_space(broot) <=
+			xfs_inode_fork_size(ip, whichfork));
+		xfs_bmbt_move_ptrs(mp, broot, old_size, new_size, old_numrecs);
+		return broot;
+	}
+
+	/*
+	 * We're reducing, but not totally eliminating, numrecs.  In this case,
+	 * we are shrinking the if_broot buffer, so it must already exist.
+	 */
+	ASSERT(ifp->if_broot != NULL && old_size > 0 && new_size > 0);
+
+	/*
+	 * Shrink the btree root by moving the bmbt pointers, since they are
+	 * not butted up against the btree block header, then reallocating
+	 * broot.
+	 */
+	xfs_bmbt_move_ptrs(mp, ifp->if_broot, old_size, new_size, new_numrecs);
+	broot = xfs_broot_realloc(ifp, new_size);
+	ASSERT(xfs_bmap_bmdr_space(broot) <=
+	       xfs_inode_fork_size(ip, whichfork));
+	return broot;
+}
+
+static struct xfs_btree_block *
+xfs_bmbt_broot_realloc(
+	struct xfs_btree_cur	*cur,
+	unsigned int		new_numrecs)
+{
+	return xfs_bmap_broot_realloc(cur->bc_ino.ip, cur->bc_ino.whichfork,
+			new_numrecs);
+}
+
 const struct xfs_btree_ops xfs_bmbt_ops = {
 	.name			= "bmap",
 	.type			= XFS_BTREE_TYPE_INODE,
@@ -543,6 +653,7 @@ const struct xfs_btree_ops xfs_bmbt_ops = {
 	.keys_inorder		= xfs_bmbt_keys_inorder,
 	.recs_inorder		= xfs_bmbt_recs_inorder,
 	.keys_contiguous	= xfs_bmbt_keys_contiguous,
+	.broot_realloc		= xfs_bmbt_broot_realloc,
 };
 
 /*
diff --git a/fs/xfs/libxfs/xfs_bmap_btree.h b/fs/xfs/libxfs/xfs_bmap_btree.h
index 49a3bae3f6ec..b238d559ab03 100644
--- a/fs/xfs/libxfs/xfs_bmap_btree.h
+++ b/fs/xfs/libxfs/xfs_bmap_btree.h
@@ -198,4 +198,7 @@ xfs_bmap_bmdr_space(struct xfs_btree_block *bb)
 	return xfs_bmdr_space_calc(be16_to_cpu(bb->bb_numrecs));
 }
 
+struct xfs_btree_block *xfs_bmap_broot_realloc(struct xfs_inode *ip,
+		int whichfork, unsigned int new_numrecs);
+
 #endif	/* __XFS_BMAP_BTREE_H__ */
diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
index 68ee1c299c25..5ab201ef041e 100644
--- a/fs/xfs/libxfs/xfs_btree.c
+++ b/fs/xfs/libxfs/xfs_btree.c
@@ -1537,12 +1537,16 @@ xfs_btree_log_recs(
 	int			first,
 	int			last)
 {
+	if (!bp) {
+		xfs_trans_log_inode(cur->bc_tp, cur->bc_ino.ip,
+				xfs_ilog_fbroot(cur->bc_ino.whichfork));
+		return;
+	}
 
 	xfs_trans_buf_set_type(cur->bc_tp, bp, XFS_BLFT_BTREE_BUF);
 	xfs_trans_log_buf(cur->bc_tp, bp,
 			  xfs_btree_rec_offset(cur, first),
 			  xfs_btree_rec_offset(cur, last + 1) - 1);
-
 }
 
 /*
@@ -3078,6 +3082,131 @@ xfs_btree_split(
 #define xfs_btree_split	__xfs_btree_split
 #endif /* __KERNEL__ */
 
+/* Move the records from a root leaf block to a separate block. */
+STATIC void
+xfs_btree_promote_leaf_iroot(
+	struct xfs_btree_cur	*cur,
+	struct xfs_btree_block	*block,
+	struct xfs_buf		*cbp,
+	union xfs_btree_ptr	*cptr,
+	struct xfs_btree_block	*cblock)
+{
+	union xfs_btree_rec	*rp;
+	union xfs_btree_rec	*crp;
+	union xfs_btree_key	*kp;
+	union xfs_btree_ptr	*pp;
+	struct xfs_btree_block	*broot;
+	int			numrecs = xfs_btree_get_numrecs(block);
+
+	/* Copy the records from the leaf broot into the new child block. */
+	rp = xfs_btree_rec_addr(cur, 1, block);
+	crp = xfs_btree_rec_addr(cur, 1, cblock);
+	xfs_btree_copy_recs(cur, crp, rp, numrecs);
+
+	/*
+	 * Increment the tree height.
+	 *
+	 * Trickery here: The amount of memory that we need per record for the
+	 * ifork's btree root block may change when we convert the broot from a
+	 * leaf to a node block.  Free the existing leaf broot so that nobody
+	 * thinks we need to migrate node pointers when we realloc the broot
+	 * buffer after bumping nlevels.
+	 */
+	cur->bc_ops->broot_realloc(cur, 0);
+	cur->bc_nlevels++;
+	cur->bc_levels[1].ptr = 1;
+
+	/*
+	 * Allocate a new node broot and initialize it to point to the new
+	 * child block.
+	 */
+	broot = cur->bc_ops->broot_realloc(cur, 1);
+	xfs_btree_init_block(cur->bc_mp, broot, cur->bc_ops,
+			cur->bc_nlevels - 1, 1, cur->bc_ino.ip->i_ino);
+
+	pp = xfs_btree_ptr_addr(cur, 1, broot);
+	kp = xfs_btree_key_addr(cur, 1, broot);
+	xfs_btree_copy_ptrs(cur, pp, cptr, 1);
+	xfs_btree_get_keys(cur, cblock, kp);
+
+	/* Attach the new block to the cursor and log it. */
+	xfs_btree_setbuf(cur, 0, cbp);
+	xfs_btree_log_block(cur, cbp, XFS_BB_ALL_BITS);
+	xfs_btree_log_recs(cur, cbp, 1, numrecs);
+}
+
+/*
+ * Move the keys and pointers from a root block to a separate block.
+ *
+ * Since the keyptr size does not change, all we have to do is increase the
+ * tree height, copy the keyptrs to the new internal node (cblock), shrink
+ * the root, and copy the pointers there.
+ */
+STATIC int
+xfs_btree_promote_node_iroot(
+	struct xfs_btree_cur	*cur,
+	struct xfs_btree_block	*block,
+	int			level,
+	struct xfs_buf		*cbp,
+	union xfs_btree_ptr	*cptr,
+	struct xfs_btree_block	*cblock)
+{
+	union xfs_btree_key	*ckp;
+	union xfs_btree_key	*kp;
+	union xfs_btree_ptr	*cpp;
+	union xfs_btree_ptr	*pp;
+	int			i;
+	int			error;
+	int			numrecs = xfs_btree_get_numrecs(block);
+
+	/*
+	 * Increase tree height, adjusting the root block level to match.
+	 * We cannot change the root btree node size until we've copied the
+	 * block contents to the new child block.
+	 */
+	be16_add_cpu(&block->bb_level, 1);
+	cur->bc_nlevels++;
+	cur->bc_levels[level + 1].ptr = 1;
+
+	/*
+	 * Adjust the root btree record count, then copy the keys from the old
+	 * root to the new child block.
+	 */
+	xfs_btree_set_numrecs(block, 1);
+	kp = xfs_btree_key_addr(cur, 1, block);
+	ckp = xfs_btree_key_addr(cur, 1, cblock);
+	xfs_btree_copy_keys(cur, ckp, kp, numrecs);
+
+	/* Check the pointers and copy them to the new child block. */
+	pp = xfs_btree_ptr_addr(cur, 1, block);
+	cpp = xfs_btree_ptr_addr(cur, 1, cblock);
+	for (i = 0; i < numrecs; i++) {
+		error = xfs_btree_debug_check_ptr(cur, pp, i, level);
+		if (error)
+			return error;
+	}
+	xfs_btree_copy_ptrs(cur, cpp, pp, numrecs);
+
+	/*
+	 * Set the first keyptr to point to the new child block, then shrink
+	 * the memory buffer for the root block.
+	 */
+	error = xfs_btree_debug_check_ptr(cur, cptr, 0, level);
+	if (error)
+		return error;
+	xfs_btree_copy_ptrs(cur, pp, cptr, 1);
+	xfs_btree_get_keys(cur, cblock, kp);
+
+	cur->bc_ops->broot_realloc(cur, 1);
+
+	/* Attach the new block to the cursor and log it. */
+	xfs_btree_setbuf(cur, level, cbp);
+	xfs_btree_log_block(cur, cbp, XFS_BB_ALL_BITS);
+	xfs_btree_log_keys(cur, cbp, 1, numrecs);
+	xfs_btree_log_ptrs(cur, cbp, 1, numrecs);
+	return 0;
+}
+
 /*
  * Copy the old inode root contents into a real block and make the
  * broot point to it.
@@ -3091,14 +3220,10 @@ xfs_btree_new_iroot(
 	struct xfs_buf		*cbp;		/* buffer for cblock */
 	struct xfs_btree_block	*block;		/* btree block */
 	struct xfs_btree_block	*cblock;	/* child btree block */
-	union xfs_btree_key	*ckp;		/* child key pointer */
-	union xfs_btree_ptr	*cpp;		/* child ptr pointer */
-	union xfs_btree_key	*kp;		/* pointer to btree key */
-	union xfs_btree_ptr	*pp;		/* pointer to block addr */
+	union xfs_btree_ptr	aptr;
 	union xfs_btree_ptr	nptr;		/* new block addr */
 	int			level;		/* btree level */
 	int			error;		/* error return code */
-	int			i;		/* loop counter */
 
 	XFS_BTREE_STATS_INC(cur, newroot);
 
@@ -3107,10 +3232,15 @@ xfs_btree_new_iroot(
 	level = cur->bc_nlevels - 1;
 
 	block = xfs_btree_get_iroot(cur);
-	pp = xfs_btree_ptr_addr(cur, 1, block);
+	ASSERT(level > 0 || (cur->bc_ops->geom_flags & XFS_BTGEO_IROOT_RECORDS));
+	if (level > 0)
+		aptr = *xfs_btree_ptr_addr(cur, 1, block);
+	else
+		aptr.l = cpu_to_be64(XFS_INO_TO_FSB(cur->bc_mp,
+				cur->bc_ino.ip->i_ino));
 
 	/* Allocate the new block. If we can't do it, we're toast. Give up. */
-	error = xfs_btree_alloc_block(cur, pp, &nptr, stat);
+	error = xfs_btree_alloc_block(cur, &aptr, &nptr, stat);
 	if (error)
 		goto error0;
 	if (*stat == 0)
@@ -3136,47 +3266,16 @@ xfs_btree_new_iroot(
 			cblock->bb_u.s.bb_blkno = bno;
 	}
 
-	be16_add_cpu(&block->bb_level, 1);
-	xfs_btree_set_numrecs(block, 1);
-	cur->bc_nlevels++;
-	ASSERT(cur->bc_nlevels <= cur->bc_maxlevels);
-	cur->bc_levels[level + 1].ptr = 1;
-
-	kp = xfs_btree_key_addr(cur, 1, block);
-	ckp = xfs_btree_key_addr(cur, 1, cblock);
-	xfs_btree_copy_keys(cur, ckp, kp, xfs_btree_get_numrecs(cblock));
-
-	cpp = xfs_btree_ptr_addr(cur, 1, cblock);
-	for (i = 0; i < be16_to_cpu(cblock->bb_numrecs); i++) {
-		error = xfs_btree_debug_check_ptr(cur, pp, i, level);
+	if (level > 0) {
+		error = xfs_btree_promote_node_iroot(cur, block, level, cbp,
+				&nptr, cblock);
 		if (error)
 			goto error0;
+	} else {
+		xfs_btree_promote_leaf_iroot(cur, block, cbp, &nptr, cblock);
 	}
 
-	xfs_btree_copy_ptrs(cur, cpp, pp, xfs_btree_get_numrecs(cblock));
-
-	error = xfs_btree_debug_check_ptr(cur, &nptr, 0, level);
-	if (error)
-		goto error0;
-
-	xfs_btree_copy_ptrs(cur, pp, &nptr, 1);
-
-	xfs_iroot_realloc(cur->bc_ino.ip,
-			  1 - xfs_btree_get_numrecs(cblock),
-			  cur->bc_ino.whichfork);
-
-	xfs_btree_setbuf(cur, level, cbp);
-
-	/*
-	 * Do all this logging at the end so that
-	 * the root is at the right level.
-	 */
-	xfs_btree_log_block(cur, cbp, XFS_BB_ALL_BITS);
-	xfs_btree_log_keys(cur, cbp, 1, be16_to_cpu(cblock->bb_numrecs));
-	xfs_btree_log_ptrs(cur, cbp, 1, be16_to_cpu(cblock->bb_numrecs));
-
-	*logflags |=
-		XFS_ILOG_CORE | xfs_ilog_fbroot(cur->bc_ino.whichfork);
+	*logflags |= XFS_ILOG_CORE | xfs_ilog_fbroot(cur->bc_ino.whichfork);
 	*stat = 1;
 	return 0;
 error0:
@@ -3347,7 +3446,7 @@ xfs_btree_make_block_unfull(
 
 		if (numrecs < cur->bc_ops->get_dmaxrecs(cur, level)) {
 			/* A root block that can be made bigger. */
-			xfs_iroot_realloc(ip, 1, cur->bc_ino.whichfork);
+			cur->bc_ops->broot_realloc(cur, numrecs + 1);
 			*stat = 1;
 		} else {
 			/* A root block that needs replacing */
@@ -3693,6 +3792,97 @@ error0:
 	return error;
 }
 
+/* Move the records from a child leaf block to the root block. */
+STATIC void
+xfs_btree_demote_leaf_child(
+	struct xfs_btree_cur	*cur,
+	struct xfs_btree_block	*cblock,
+	int			numrecs)
+{
+	union xfs_btree_rec	*rp;
+	union xfs_btree_rec	*crp;
+	struct xfs_btree_block	*broot;
+
+	/*
+	 * Decrease the tree height.
+	 *
+	 * Trickery here: The amount of memory that we need per record for the
+	 * ifork's btree root block may change when we convert the broot from a
+	 * node to a leaf.  Free the old node broot so that we can get a fresh
+	 * leaf broot.
+	 */
+	cur->bc_ops->broot_realloc(cur, 0);
+	cur->bc_nlevels--;
+
+	/*
+	 * Allocate a new leaf broot and copy the records from the old child.
+	 * Detach the old child from the cursor.
+	 */
+	broot = cur->bc_ops->broot_realloc(cur, numrecs);
+	xfs_btree_init_block(cur->bc_mp, broot, cur->bc_ops, 0, numrecs,
+			cur->bc_ino.ip->i_ino);
+
+	rp = xfs_btree_rec_addr(cur, 1, broot);
+	crp = xfs_btree_rec_addr(cur, 1, cblock);
+	xfs_btree_copy_recs(cur, rp, crp, numrecs);
+
+	cur->bc_levels[0].bp = NULL;
+}
+
+/*
+ * Move the keyptrs from a child node block to the root block.
+ *
+ * Since the keyptr size does not change, all we have to do is increase the
+ * tree height, copy the keyptrs to the new internal node (cblock), shrink
+ * the root, and copy the pointers there.
+ */
+STATIC int
+xfs_btree_demote_node_child(
+	struct xfs_btree_cur	*cur,
+	struct xfs_btree_block	*cblock,
+	int			level,
+	int			numrecs)
+{
+	struct xfs_btree_block	*block;
+	union xfs_btree_key	*ckp;
+	union xfs_btree_key	*kp;
+	union xfs_btree_ptr	*cpp;
+	union xfs_btree_ptr	*pp;
+	int			i;
+	int			error;
+
+	/*
+	 * Adjust the root btree node size and the record count to match the
+	 * doomed child so that we can copy the keyptrs ahead of changing the
+	 * tree shape.
+	 */
+	block = cur->bc_ops->broot_realloc(cur, numrecs);
+
+	xfs_btree_set_numrecs(block, numrecs);
+	ASSERT(block->bb_numrecs == cblock->bb_numrecs);
+
+	/* Copy keys from the doomed block. */
+	kp = xfs_btree_key_addr(cur, 1, block);
+	ckp = xfs_btree_key_addr(cur, 1, cblock);
+	xfs_btree_copy_keys(cur, kp, ckp, numrecs);
+
+	/* Copy pointers from the doomed block. */
+	pp = xfs_btree_ptr_addr(cur, 1, block);
+	cpp = xfs_btree_ptr_addr(cur, 1, cblock);
+	for (i = 0; i < numrecs; i++) {
+		error = xfs_btree_debug_check_ptr(cur, cpp, i, level - 1);
+		if (error)
+			return error;
+	}
+	xfs_btree_copy_ptrs(cur, pp, cpp, numrecs);
+
+	/* Decrease tree height, adjusting the root block level to match. */
+	cur->bc_levels[level - 1].bp = NULL;
+	be16_add_cpu(&block->bb_level, -1);
+	cur->bc_nlevels--;
+	return 0;
+}
+
 /*
  * Try to merge a non-leaf block back into the inode root.
  *
@@ -3705,34 +3895,31 @@ STATIC int
 xfs_btree_kill_iroot(
 	struct xfs_btree_cur	*cur)
 {
-	int			whichfork = cur->bc_ino.whichfork;
 	struct xfs_inode	*ip = cur->bc_ino.ip;
-	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, whichfork);
 	struct xfs_btree_block	*block;
 	struct xfs_btree_block	*cblock;
-	union xfs_btree_key	*kp;
-	union xfs_btree_key	*ckp;
-	union xfs_btree_ptr	*pp;
-	union xfs_btree_ptr	*cpp;
 	struct xfs_buf		*cbp;
 	int			level;
-	int			index;
 	int			numrecs;
 	int			error;
 #ifdef DEBUG
 	union xfs_btree_ptr	ptr;
 #endif
-	int			i;
 
 	ASSERT(cur->bc_ops->type == XFS_BTREE_TYPE_INODE);
-	ASSERT(cur->bc_nlevels > 1);
+	ASSERT((cur->bc_ops->geom_flags & XFS_BTGEO_IROOT_RECORDS) ||
+	       cur->bc_nlevels > 1);
 
 	/*
 	 * Don't deal with the root block needs to be a leaf case.
 	 * We're just going to turn the thing back into extents anyway.
 	 */
 	level = cur->bc_nlevels - 1;
-	if (level == 1)
+	if (level == 1 && !(cur->bc_ops->geom_flags & XFS_BTGEO_IROOT_RECORDS))
+		goto out0;
+
+	/* If we're already a leaf, jump out. */
+	if (level == 0)
 		goto out0;
 
 	/*
@@ -3762,40 +3949,20 @@ xfs_btree_kill_iroot(
 	ASSERT(xfs_btree_ptr_is_null(cur, &ptr));
 #endif
 
-	index = numrecs - cur->bc_ops->get_maxrecs(cur, level);
-	if (index) {
-		xfs_iroot_realloc(cur->bc_ino.ip, index,
-				  cur->bc_ino.whichfork);
-		block = ifp->if_broot;
-	}
-
-	be16_add_cpu(&block->bb_numrecs, index);
-	ASSERT(block->bb_numrecs == cblock->bb_numrecs);
-
-	kp = xfs_btree_key_addr(cur, 1, block);
-	ckp = xfs_btree_key_addr(cur, 1, cblock);
-	xfs_btree_copy_keys(cur, kp, ckp, numrecs);
-
-	pp = xfs_btree_ptr_addr(cur, 1, block);
-	cpp = xfs_btree_ptr_addr(cur, 1, cblock);
-
-	for (i = 0; i < numrecs; i++) {
-		error = xfs_btree_debug_check_ptr(cur, cpp, i, level - 1);
+	if (level > 1) {
+		error = xfs_btree_demote_node_child(cur, cblock, level,
+				numrecs);
 		if (error)
 			return error;
-	}
-
-	xfs_btree_copy_ptrs(cur, pp, cpp, numrecs);
+	} else
+		xfs_btree_demote_leaf_child(cur, cblock, numrecs);
 
 	error = xfs_btree_free_block(cur, cbp);
 	if (error)
 		return error;
 
-	cur->bc_levels[level - 1].bp = NULL;
-	be16_add_cpu(&block->bb_level, -1);
 	xfs_trans_log_inode(cur->bc_tp, ip,
 		XFS_ILOG_CORE | xfs_ilog_fbroot(cur->bc_ino.whichfork));
-	cur->bc_nlevels--;
 out0:
 	return 0;
 }
@@ -3949,10 +4116,10 @@ xfs_btree_delrec(
 	/*
 	 * We're at the root level.  First, shrink the root block in-memory.
 	 * Try to get rid of the next level down.  If we can't then there's
-	 * nothing left to do.
+	 * nothing left to do.  numrecs was decremented above.
 	 */
 	if (xfs_btree_at_iroot(cur, level)) {
-		xfs_iroot_realloc(cur->bc_ino.ip, -1, cur->bc_ino.whichfork);
+		cur->bc_ops->broot_realloc(cur, numrecs);
 
 		error = xfs_btree_kill_iroot(cur);
 		if (error)
diff --git a/fs/xfs/libxfs/xfs_btree.h b/fs/xfs/libxfs/xfs_btree.h
index c5bff273cae2..3b8c2ccad908 100644
--- a/fs/xfs/libxfs/xfs_btree.h
+++ b/fs/xfs/libxfs/xfs_btree.h
@@ -213,11 +213,27 @@ struct xfs_btree_ops {
 			       const union xfs_btree_key *key1,
 			       const union xfs_btree_key *key2,
 			       const union xfs_btree_key *mask);
+
+	/*
+	 * Reallocate the space for if_broot to fit the number of records.
+	 * Move the records and pointers in if_broot to fit the new size.  When
+	 * shrinking this will eliminate holes between the records and pointers
+	 * created by the caller.  When growing this will create holes to be
+	 * filled in by the caller.
+	 *
+	 * The caller must not request to add more records than would fit in
+	 * the on-disk inode root.  If the if_broot is currently NULL, then if
+	 * we are adding records, one will be allocated.  The caller must also
+	 * not request that the number of records go below zero, although it
+	 * can go to zero.
+	 */
+	struct xfs_btree_block *(*broot_realloc)(struct xfs_btree_cur *cur,
+				unsigned int new_numrecs);
 };
 
 /* btree geometry flags */
 #define XFS_BTGEO_OVERLAPPING		(1U << 0) /* overlapping intervals */
-
+#define XFS_BTGEO_IROOT_RECORDS		(1U << 1) /* iroot can store records */
 
 union xfs_btree_irec {
 	struct xfs_alloc_rec_incore	a;
diff --git a/fs/xfs/libxfs/xfs_btree_staging.c b/fs/xfs/libxfs/xfs_btree_staging.c
index 694929703152..58c146b5c9d4 100644
--- a/fs/xfs/libxfs/xfs_btree_staging.c
+++ b/fs/xfs/libxfs/xfs_btree_staging.c
@@ -573,6 +573,7 @@ xfs_btree_bload_compute_geometry(
 	struct xfs_btree_bload	*bbl,
 	uint64_t		nr_records)
 {
+	const struct xfs_btree_ops *ops = cur->bc_ops;
 	uint64_t		nr_blocks = 0;
 	uint64_t		nr_this_level;
 
@@ -599,7 +600,7 @@ xfs_btree_bload_compute_geometry(
 		xfs_btree_bload_level_geometry(cur, bbl, level, nr_this_level,
 				&avg_per_block, &level_blocks, &dontcare64);
 
-		if (cur->bc_ops->type == XFS_BTREE_TYPE_INODE) {
+		if (ops->type == XFS_BTREE_TYPE_INODE) {
 			/*
 			 * If all the items we want to store at this level
 			 * would fit in the inode root block, then we have our
@@ -607,7 +608,9 @@ xfs_btree_bload_compute_geometry(
 			 *
 			 * Note that bmap btrees forbid records in the root.
 			 */
-			if (level != 0 && nr_this_level <= avg_per_block) {
+			if ((level != 0 ||
+			     (ops->geom_flags & XFS_BTGEO_IROOT_RECORDS)) &&
+			    nr_this_level <= avg_per_block) {
 				nr_blocks++;
 				break;
 			}
@@ -658,7 +661,7 @@ xfs_btree_bload_compute_geometry(
 		return -EOVERFLOW;
 
 	bbl->btree_height = cur->bc_nlevels;
-	if (cur->bc_ops->type == XFS_BTREE_TYPE_INODE)
+	if (ops->type == XFS_BTREE_TYPE_INODE)
 		bbl->nr_blocks = nr_blocks - 1;
 	else
 		bbl->nr_blocks = nr_blocks;
diff --git a/fs/xfs/libxfs/xfs_inode_fork.c b/fs/xfs/libxfs/xfs_inode_fork.c
index 1158ca48626b..60853bac289a 100644
--- a/fs/xfs/libxfs/xfs_inode_fork.c
+++ b/fs/xfs/libxfs/xfs_inode_fork.c
@@ -178,7 +178,7 @@ xfs_iformat_btree(
 	struct xfs_mount	*mp = ip->i_mount;
 	xfs_bmdr_block_t	*dfp;
 	struct xfs_ifork	*ifp;
-	/* REFERENCED */
+	struct xfs_btree_block	*broot;
 	int			nrecs;
 	int			size;
 	int			level;
@@ -211,16 +211,13 @@ xfs_iformat_btree(
 		return -EFSCORRUPTED;
 	}
 
-	ifp->if_broot_bytes = size;
-	ifp->if_broot = kmalloc(size,
-				GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
-	ASSERT(ifp->if_broot != NULL);
+	broot = xfs_broot_alloc(ifp, size);
 	/*
 	 * Copy and convert from the on-disk structure
 	 * to the in-memory structure.
 	 */
 	xfs_bmdr_to_bmbt(ip, dfp, XFS_DFORK_SIZE(dip, ip->i_mount, whichfork),
-			 ifp->if_broot, size);
+			 broot, size);
 
 	ifp->if_bytes = 0;
 	ifp->if_data = NULL;
@@ -363,135 +360,68 @@ xfs_iformat_attr_fork(
 }
 
 /*
- * Reallocate the space for if_broot based on the number of records
- * being added or deleted as indicated in rec_diff.  Move the records
- * and pointers in if_broot to fit the new size.  When shrinking this
- * will eliminate holes between the records and pointers created by
- * the caller.  When growing this will create holes to be filled in
- * by the caller.
- *
- * The caller must not request to add more records than would fit in
- * the on-disk inode root.  If the if_broot is currently NULL, then
- * if we are adding records, one will be allocated.  The caller must also
- * not request that the number of records go below zero, although
- * it can go to zero.
- *
- * ip -- the inode whose if_broot area is changing
- * ext_diff -- the change in the number of records, positive or negative,
- *	 requested for the if_broot array.
+ * Allocate the if_broot component of an inode fork so that it is @new_size
+ * bytes in size, using __GFP_NOLOCKDEP like all the other code that
+ * initializes a broot during inode load.  Returns if_broot.
  */
-void
-xfs_iroot_realloc(
-	xfs_inode_t		*ip,
-	int			rec_diff,
-	int			whichfork)
+struct xfs_btree_block *
+xfs_broot_alloc(
+	struct xfs_ifork	*ifp,
+	size_t			new_size)
 {
-	struct xfs_mount	*mp = ip->i_mount;
-	int			cur_max;
-	struct xfs_ifork	*ifp;
-	struct xfs_btree_block	*new_broot;
-	int			new_max;
-	size_t			new_size;
-	char			*np;
-	char			*op;
+	ASSERT(ifp->if_broot == NULL);
 
-	/*
-	 * Handle the degenerate case quietly.
-	 */
-	if (rec_diff == 0) {
-		return;
-	}
+	ifp->if_broot = kmalloc(new_size,
+				GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
+	ifp->if_broot_bytes = new_size;
+	return ifp->if_broot;
+}
 
-	ifp = xfs_ifork_ptr(ip, whichfork);
-	if (rec_diff > 0) {
-		/*
-		 * If there wasn't any memory allocated before, just
-		 * allocate it now and get out.
-		 */
-		if (ifp->if_broot_bytes == 0) {
-			new_size = xfs_bmap_broot_space_calc(mp, rec_diff);
-			ifp->if_broot = kmalloc(new_size,
-						GFP_KERNEL | __GFP_NOFAIL);
-			ifp->if_broot_bytes = (int)new_size;
-			return;
-		}
+/*
+ * Reallocate the if_broot component of an inode fork so that it is @new_size
+ * bytes in size.  Returns if_broot.
+ */
+struct xfs_btree_block *
+xfs_broot_realloc(
+	struct xfs_ifork	*ifp,
+	size_t			new_size)
+{
+	/* No size change?  No action needed. */
+	if (new_size == ifp->if_broot_bytes)
+		return ifp->if_broot;
 
-		/*
-		 * If there is already an existing if_broot, then we need
-		 * to realloc() it and shift the pointers to their new
-		 * location.  The records don't change location because
-		 * they are kept butted up against the btree block header.
-		 */
-		cur_max = xfs_bmbt_maxrecs(mp, ifp->if_broot_bytes, false);
-		new_max = cur_max + rec_diff;
-		new_size = xfs_bmap_broot_space_calc(mp, new_max);
-		ifp->if_broot = krealloc(ifp->if_broot, new_size,
-					 GFP_KERNEL | __GFP_NOFAIL);
-		op = (char *)xfs_bmap_broot_ptr_addr(mp, ifp->if_broot, 1,
-						     ifp->if_broot_bytes);
-		np = (char *)xfs_bmap_broot_ptr_addr(mp, ifp->if_broot, 1,
-						     (int)new_size);
-		ifp->if_broot_bytes = (int)new_size;
-		ASSERT(xfs_bmap_bmdr_space(ifp->if_broot) <=
-			xfs_inode_fork_size(ip, whichfork));
-		memmove(np, op, cur_max * (uint)sizeof(xfs_fsblock_t));
-		return;
+	/* New size is zero, free it. */
+	if (new_size == 0) {
+		ifp->if_broot_bytes = 0;
+		kfree(ifp->if_broot);
+		ifp->if_broot = NULL;
+		return NULL;
 	}
 
 	/*
-	 * rec_diff is less than 0.  In this case, we are shrinking the
-	 * if_broot buffer.  It must already exist.  If we go to zero
-	 * records, just get rid of the root and clear the status bit.
+	 * Shrinking the iroot means we allocate a new smaller object and copy
+	 * it.  We don't trust krealloc not to nop on realloc-down.
 	 */
-	ASSERT((ifp->if_broot != NULL) && (ifp->if_broot_bytes > 0));
-	cur_max = xfs_bmbt_maxrecs(mp, ifp->if_broot_bytes, false);
-	new_max = cur_max + rec_diff;
-	ASSERT(new_max >= 0);
-	if (new_max > 0)
-		new_size = xfs_bmap_broot_space_calc(mp, new_max);
-	else
-		new_size = 0;
-	if (new_size > 0) {
-		new_broot = kmalloc(new_size, GFP_KERNEL | __GFP_NOFAIL);
-		/*
-		 * First copy over the btree block header.
-		 */
-		memcpy(new_broot, ifp->if_broot,
-			xfs_bmbt_block_len(ip->i_mount));
-	} else {
-		new_broot = NULL;
+	if (ifp->if_broot_bytes > 0 && ifp->if_broot_bytes > new_size) {
+		struct xfs_btree_block	*old_broot = ifp->if_broot;
+
+		ifp->if_broot = kmalloc(new_size, GFP_KERNEL | __GFP_NOFAIL);
+		ifp->if_broot_bytes = new_size;
+		memcpy(ifp->if_broot, old_broot, new_size);
+		kfree(old_broot);
+		return ifp->if_broot;
 	}
 
 	/*
-	 * Only copy the keys and pointers if there are any.
+	 * Growing the iroot means we can krealloc.  This may get us the same
+	 * object.
 	 */
-	if (new_max > 0) {
-		/*
-		 * First copy the keys.
-		 */
-		op = (char *)xfs_bmbt_key_addr(mp, ifp->if_broot, 1);
-		np = (char *)xfs_bmbt_key_addr(mp, new_broot, 1);
-		memcpy(np, op, new_max * (uint)sizeof(xfs_bmbt_key_t));
-
-		/*
-		 * Then copy the pointers.
-		 */
-		op = (char *)xfs_bmap_broot_ptr_addr(mp, ifp->if_broot, 1,
-						     ifp->if_broot_bytes);
-		np = (char *)xfs_bmap_broot_ptr_addr(mp, new_broot, 1,
-						     (int)new_size);
-		memcpy(np, op, new_max * (uint)sizeof(xfs_fsblock_t));
-	}
-	kfree(ifp->if_broot);
-	ifp->if_broot = new_broot;
-	ifp->if_broot_bytes = (int)new_size;
-	if (ifp->if_broot)
-		ASSERT(xfs_bmap_bmdr_space(ifp->if_broot) <=
-			xfs_inode_fork_size(ip, whichfork));
-	return;
+	ifp->if_broot = krealloc(ifp->if_broot, new_size,
+			GFP_KERNEL | __GFP_NOFAIL);
+	ifp->if_broot_bytes = new_size;
+	return ifp->if_broot;
 }
 
-
 /*
  * This is called when the amount of space needed for if_data
  * is increased or decreased.  The change in size is indicated by
diff --git a/fs/xfs/libxfs/xfs_inode_fork.h b/fs/xfs/libxfs/xfs_inode_fork.h
index 2373d12fd474..69ed0919d60b 100644
--- a/fs/xfs/libxfs/xfs_inode_fork.h
+++ b/fs/xfs/libxfs/xfs_inode_fork.h
@@ -170,7 +170,11 @@ void		xfs_iflush_fork(struct xfs_inode *, struct xfs_dinode *,
 void		xfs_idestroy_fork(struct xfs_ifork *ifp);
 void *		xfs_idata_realloc(struct xfs_inode *ip, int64_t byte_diff,
 				int whichfork);
-void		xfs_iroot_realloc(struct xfs_inode *, int, int);
+struct xfs_btree_block *xfs_broot_alloc(struct xfs_ifork *ifp,
+				size_t new_size);
+struct xfs_btree_block *xfs_broot_realloc(struct xfs_ifork *ifp,
+				size_t new_size);
+
 int		xfs_iread_extents(struct xfs_trans *, struct xfs_inode *, int);
 int		xfs_iextents_copy(struct xfs_inode *, struct xfs_bmbt_rec *,
 				  int);