staging: lustre: ldlm: Do not use cbpending for group locks

staging: lustre: ldlm: Do not use cbpending for group locks

Currently, the CBPENDING flag is set on group locks when
the osc lock above them is released (osc_cancel_base).

This results in a situation where a new group lock request
on a resource does not match an existing group lock because
LDLM_FL_CBPENDING is set on the existing lock.

So two group locks are granted on the same resource, which
is not valid, since a given client can only have one group
lock on a particular resource.

Since group locks are manually released and not called back
like other LDLM locks, the CBPENDING flag doesn't make
sense. Since they must be manually released, they also
cannot go in the LDLM LRU cache and must be fully released
immediately once they are no longer in use.

This was previously accomplished by setting CBPENDING when
the corresponding osc lock is released, but as noted above,
this prevents the group lock matching some future lock
requests.

This patch uses the fact that group locks have an l_writers
reference which they keep until they are manually released,
so we remove them when they have no more reader or writer
references, without checking cbpending.

Signed-off-by: Patrick Farrell <paf@cray.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6368
Reviewed-on: http://review.whamcloud.com/14093
Reviewed-by: frank zago <fzago@cray.com>
Reviewed-by: Jinshan Xiong <jinshan.xiong@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>