LU-11089 obdclass: make key_set_version an atomic_t

LU-11089 obdclass: make key_set_version an atomic_t

As a first step to simplifying the locking in lu_object.c,
change key_set_version to an atomic_t.  This will mean
we don't need to hold a spinlock when incrementing.

It is clear that keys_fill() (and so lu_context_refill())
cannot race with itself as it holds no locks for the main
part of the work.  So updates to lc_version, and testing of
that value cannot need locking either.
So remove the locking from keys_fill() and lu_context_refill()
as there is no longer anything to protect.  The locking around
lu_keys_initing_cnt is preserved for now.

Also, don't increment when deregistering or quiescing a key.
key_set_version is *only* use to avoid filling new key values
if there have been no changes to the set of key.
Deregistering a key does not mean that we need to try filling
any new value, so the increment is pointless.

Finally, remove the refill loop in keys_fill().  If a key
is registered or revived while keys_fill() is running it must be safe
to ignore it just as we would if it was registered immediately
after keys_fill() ran.  The important thing is that the
keys_set_version stored in ctx->lc_version must be sampled
*before* those unseen keys were added.
So sample keys_set_version early.

Linux-commit : 0fbfbc5ad0f892cf4c5e087a4e7e67102b2289af

Change-Id: Ic6907561d0bb864b10e1c53fb3e5469d0c60f888
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-on: https://review.whamcloud.com/32711
Reviewed-by: Gu Zheng <gzheng@ddn.com>
Tested-by: Jenkins
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Alexey Lyashkov <c17817@cray.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>