Stan, We found a problem at this patch and it should not be applied because of the lock is always taken at upper layer. I apologize for the bad patch
From: Smith, Stan [mailto:[email protected]] Sent: Friday, November 26, 2010 8:06 PM To: Alex Naslednikov; [email protected] Subject: RE: [ofw] [Patch][IPoIB_NDIS6_CM] Take a lock when calling __endpt_mgr_insert () Hello, Although your patch is correct, it brings to light a minor issue. In two places [__recv_get_endpts() & __recv_arp()] calls to 'status = __endpt_mgr_insert()', the port lock is not released immediately after the insert call. Instead the lock is held until after the 'status' is checked, thus resulting in a extra call to port unlock if the status != 0. The code could be cleaned up by releasing the port lock immediately after the __endpt_mgr_insert() call as shown in your patch. stan. ________________________________ From: [email protected] [mailto:[email protected]] On Behalf Of Alex Naslednikov Sent: Thursday, November 25, 2010 1:54 AM To: [email protected] Subject: [ofw] [Patch][IPoIB_NDIS6_CM] Take a lock when calling __endpt_mgr_insert () __endpt_mgr_insert () works with CL objects like queues that are not thread-safe. In order to protect on these objects, one need to take a lock Signed-off by: Alexander Naslednikov (xalex at mellanox.co.il) Index: ulp/ipoib_NDIS6_CM/kernel/ipoib_port.cpp =================================================================== --- ulp/ipoib_NDIS6_CM/kernel/ipoib_port.cpp (revision 2999) +++ ulp/ipoib_NDIS6_CM/kernel/ipoib_port.cpp (working copy) @@ -7242,9 +7242,11 @@ /* __endpt_mgr_insert expects *one* reference to be held. */ cl_atomic_inc( &p_port->endpt_rdr ); + cl_obj_lock( &p_port->obj ); status = __endpt_mgr_insert( p_port, p_port->p_adapter->params.conf_mac, p_endpt ); + cl_obj_unlock( &p_port->obj ); cl_atomic_dec( &p_port->endpt_rdr ); if( status != IB_SUCCESS ) {
_______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
