Hello,
  This locking patch reminds me of an earlier version which dead locked due to 
the routine being called from two different locations :
One call site required locking send_lock, while the other call site already 
held the send_lock, hence the dead lock....
Is this patch different?
How much IPoIB_CM testing has been accomplished with this patch installed?

thanks,

stan.


From: [email protected] 
[mailto:[email protected]] On Behalf Of Alex Naslednikov
Sent: Thursday, April 28, 2011 2:44 AM
To: [email protected]
Subject: [ofw] [patch][IPoIB_NDIS6_CM] Take a send lock when join_mcast_group 
is not called directly from a send path

Function ipoib_endpt_queue requires send lock to be held.
ipoib_port_join_mcast  calls to __endpt_mgr_ref that calls to ipoib_endpt_queue 
, but ipoib_port_join_mcast isn't always send lock protected.
Signed off by: Alexander Naslednikov (xalex at mellanox.co.il)

Index: B:/users/xalex/MLNX_VPI_trunk/ulp/ipoib_NDIS6_CM/kernel/ipoib_adapter.cpp
===================================================================
--- B:/users/xalex/MLNX_VPI_trunk/ulp/ipoib_NDIS6_CM/kernel/ipoib_adapter.cpp   
(revision 7799)
+++ B:/users/xalex/MLNX_VPI_trunk/ulp/ipoib_NDIS6_CM/kernel/ipoib_adapter.cpp   
             (revision 7800)
@@ -987,6 +987,8 @@
                                }

                                /* Add new entries */
+                             /* The lock is important here cause 
ipoib_endpt_queue will be called later and it requires lock to be held */
+                             cl_spinlock_acquire( &p_port->send_lock );
                                for( i = 0; i < num_macs; i++ )
                                {
                                                for( j = 0; j < 
p_adapter->mcast_array_size; j++ )
@@ -1009,6 +1011,7 @@
                                                                
ipoib_port_join_mcast( p_port, p_mac_array[i], IB_MC_REC_STATE_FULL_MEMBER );
                                                }
                                }
+                             cl_spinlock_release( &p_port->send_lock );
                }

                /* Copy the MAC array. */
@@ -1316,6 +1319,8 @@
                                                }
                                }
                                /* Join all programmed multicast groups. */
+                             /* The lock is important here cause 
ipoib_endpt_queue will be called later and it requires lock to be held */
+                             cl_spinlock_acquire( 
&p_adapter->p_port->send_lock );
                                for( i = 0; i < p_adapter->mcast_array_size; 
i++ )
                                {
                                                IPOIB_PRINT_EXIT( 
TRACE_LEVEL_INFORMATION, IPOIB_DBG_MCAST,
@@ -1324,7 +1329,8 @@
                                                                                
                                                   p_adapter->mcast_array[i],
                                                                                
                                                   IB_MC_REC_STATE_FULL_MEMBER 
);
                                }
-
+                             cl_spinlock_release( 
&p_adapter->p_port->send_lock );
+
                                /* Register all existing addresses. */
                                ipoib_reg_addrs( p_adapter );


Alexander (XaleX) Naslednikov
SW Networking Team
Mellanox Technologies

_______________________________________________
ofw mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw

Reply via email to