[0/8,v2] Cleanup how we handle root refs, part 2
mbox series

Message ID 20200117135238.41601-1-josef@toxicpanda.com
Headers show
Series
  • Cleanup how we handle root refs, part 2
Related show

Message

Josef Bacik Jan. 17, 2020, 1:52 p.m. UTC
v1->v2:
- Rebased onto latest misc-next.
- Fixed a problem where we were dropping the ino_cache_inode in the final put of
  the root, which is braindead since it was holding a ref on that root.

--------------- Original email -------------------
This series depends on part 1 and is the final bit of cleaning up how we handle
root refs.  Historically the ref counting has been just about kfree'ing the root
itself and not actually cleaning up the root and free'ing it.  This makes it
sort of awkward to handle the lifetime of a root, as we will just free a bunch
of things related to the root, but then not free the actual root until we drop
the last reference.

This patch series brings the actual cleanup of the root inside the ref
accounting for the root.  In addition to that we also now hold refs on the root
for all of the various users of the root in order to make the lifetime more
coherent.  Previously if you looked up a root we just held it in memory until
unmount and then we had to do two put's to clear it out.  Now we hold refs to
the root when we open an inode in the root, so we could get rid of this extra
ref holding.  Now we hold refs for discrete operations, like putting it in the
radix tree, adding to the dead roots list, and of course opening an inode.

This final piece allows us to remove the subvol_srcu, which was the reason for
all of this work.  I fixed the original deadlock, but it's use was sporadic and
inconsistent, which is a recipe for trouble.

Unfortunately I was not able to kill the per-root inode rb tree yet.  We use it
in relocation to drop the extent cache for extents we are relocating, and
there's not a very clear way to accomplish that without the inode rbtree right
now.  However this work means that once I've fixed that problem we can delete
the inode rbtree completely.  Thanks,

Josef