This patch fixed a race between ext3_discard_reservation() and
ext3_try_to_allocate_with_rsv().
There is a window where ext3_discard_reservation will remove an already
unlinked reservation window node from the filesystem reservation tree:
It thinks the reservation is still linked in the filesystem reservation
tree, but it is actually temperately removed from the tree by
allocate_new_reservation() when it failed to make a new reservation from
the current group and try to make a new reservation from next block
group.
Here is how it could happen:
CPU 1
try to allocate a block in group1 with given reservation window my_rsv
ext3_try_to_allocate_with_rsv(group
----copy reservation window my_rsv into local rsv_copy
ext3_try_to_allocate(...rsv_copy)
----no free block in existing reservation window,
----need a new reservation window
spin_lock(&rsv_lock);
CPU 2
ext3_discard_reservation
if (!rsv_is_empty()
----this is true
spin_lock(&rsv_lock)
----waiting for thread 1
CPU 1:
allocate_new_reservation
failed to reserve blocks in this group
remove the window from the tree
rsv_window_remove(my_rsv)
----window node is unlinked from the tree here
return -1
spin_unlock(&rsv_lock)
ext3_try_to_allocate_with_rsv() failed in this group
group++
CPU 2
spin_lock(&rsv_lock) succeed
rsv_remove_window ()
---------------break, trying to remove a unlinked node from the tree
....
CPU 1:
ext3_try_to_allocate_with_rsv(group, my_rsv)
rsv_is_empty is true, need a new reservation window
spin_lock(&rsv_lock);
^--------------- spinning forever
We need to re-check whether the reservation window is still linked to
the tree after grab the rsv_lock spin lock in ext3_discard_reservation,
to prevent panic in rsv_remove_window->rb_erase.
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: Chris Wright <chrisw@osdl.org>