fix buffer refcount races v2

[Christoph Hellwig <hch@lst.de>]

Hi all,

this series fixes two races in buffer refcount handling, that I've
so far not actually seen in real life, but only found while reading
through the buffer cache code to understand some of the stranger looking
locking decisions.

One can cause a buffer about to be freed to be returned from
xfs_buf_insert.  I think this one is extremely unlikely to be hit,
as it requires the buffer to not be present for the initial lookup,
but already being evicted when trying to add the new buffer.  But
at least the fix is trivial.

The second causes buffer lookups to be missed when moving to the LRU.
This might actually be able to trigger the first one, but otherwise
just means we're doing a pass through insert which will find it.
For pure lookups using xfs_buf_incore it could cause us to miss buffer
invalidation.  The fix for that is bigger and has bigger implications
because it not requires all b_hold increments to be done under d_lock.
This causes more contention, but as releasing the buffer always takes
the lock it can't be too horrible.  I also have a only minimally
tested series to switch it over to a lockref here:

    http://git.infradead.org/?p=users/hch/xfs.git;a=shortlog;h=refs/heads/xfs-buffer-locking

Changes since v1:
 - don't move b_lock


Diffstat:
 xfs_buf.c   |   94 ++++++++++++++++++++++++++++++------------------------------
 xfs_buf.h   |    2 -
 xfs_trace.h |   10 +++---
 3 files changed, 54 insertions(+), 52 deletions(-)