[ofa-general] [PATCH] opensm/osm_sa_mcmember_record.c: fix mcgrp cleanup crash

[Sasha Khapyorsky]

When multiple MCMember leave requests are issued the request processors
may run concurrently and then the race between port from multicast group
removing and multicast group deletion (then it becomes empty) is
possible for different processors, as result we are getting double free
crash (reproducible under stress testing).

This patch fixes this by moving multicast group cleanup call under some
lock protected block as port removing code.

Signed-off-by: Sasha Khapyorsky <sashak at voltaire.com>
---
 opensm/opensm/osm_sa_mcmember_record.c |    4 +---
 1 files changed, 1 insertions(+), 3 deletions(-)

diff --git a/opensm/opensm/osm_sa_mcmember_record.c b/opensm/opensm/osm_sa_mcmember_record.c
index b580eb7..f291bf0 100644
--- a/opensm/opensm/osm_sa_mcmember_record.c
+++ b/opensm/opensm/osm_sa_mcmember_record.c
@@ -964,13 +964,11 @@ static void mcmr_rcv_leave_mgrp(IN osm_sa_t * sa, IN osm_madw_t * p_madw)
 	/* remove port and/or update join state */
 	osm_mgrp_remove_port(sa->p_subn, sa->p_log, p_mgrp, p_mcm_port,
 			     &mcmember_rec);
+	osm_mgrp_cleanup(sa->p_subn, p_mgrp);
 	CL_PLOCK_RELEASE(sa->p_lock);
 
 	mcmr_rcv_respond(sa, p_madw, &mcmember_rec);
 
-	CL_PLOCK_EXCL_ACQUIRE(sa->p_lock);
-	osm_mgrp_cleanup(sa->p_subn, p_mgrp);
-	CL_PLOCK_RELEASE(sa->p_lock);
 Exit:
 	OSM_LOG_EXIT(sa->p_log);
 }
-- 
1.6.5.rc1