When writing into preallocated regions there is a case where XFS can oops
or hang doing the unwritten extent conversion on I/O completion. It turns
out that the problem is related to the btree cursor being invalid.

When we do an insert into the tree, we may need to split blocks in the
tree. When we only split at the leaf level (i.e. level 0), everything
works just fine. However, if we have a multi-level split in the btreee,
the cursor passed to the insert function is no longer valid once the
insert is complete.

The leaf level split is handled correctly because all the operations at
level 0 are done using the original cursor, hence it is updated correctly.
However, when we need to update the next level up the tree, we don't use
that cursor - we use a cloned cursor that points to the index in the next
level up where we need to do the insert.

Hence if we need to split a second level, the changes to the tree are
reflected in the cloned cursor and not the original cursor. This
clone-and-move-up-a-level-on-split behaviour recurses all the way to the
top of the tree.

The complexity here is that these cloned cursors do not point to the
original index that was inserted - they point to the newly allocated block
(the right block) and the original cursor pointer to that level may still
point to the left block. Hence, without deep examination of the cloned
cursor and buffers, we cannot update the original cursor with the new path
from the cloned cursor.

In these cases the original cursor could be pointing to the wrong block(s)
and hence a subsequent modification to the tree using that cursor will
lead to corruption of the tree.

The crash case occurs when the tree changes height - we insert a new level
in the tree, and the cursor does not have a buffer in it's path for that
level. Hence any attempt to walk back up the cursor to the root block will
result in a null pointer dereference.

To make matters even more complex, the BMAP BT is rooted in an inode, so
we can have a change of height in the btree *without a root split*. That
is, if the root block in the inode is full when we split a leaf node, we
cannot fit the pointer to the new block in the root, so we allocate a new
block, migrate all the ptrs out of the inode into the new block and point
the inode root block at the newly allocated block. This changes the height
of the tree without a root split having occurred and hence invalidates the
path in the original cursor.

The patch below prevents xfs_bmbt_insert() from returning with an invalid
cursor by detecting the cases that invalidate the original cursor and
refresh it by do a lookup into the btree for the original index we were
inserting at.

Note that the INOBT, AGFBNO and AGFCNT btree implementations also have
this bug, but the cursor is currently always destroyed or revalidated
after an insert for those trees. Hence this patch only address the problem
in the BMBT code.

SGI-PV: 979339
SGI-Modid: xfs-linux-melb:xfs-kern:30701a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>