I must add a general comment: If there is a bluescreen there is a bug in a driver. One can not trust an application to be written correctly in order not to receive blue screens.
The way that the user and kernel code are communicating is through a device. The kernel gets a notification when the application dyes and he should clean whatever the program has not cleaned. Thanks Tzachi > -----Original Message----- > From: [email protected] > [mailto:[email protected]] On Behalf Of Usha > Srinivasan > Sent: Wednesday, February 03, 2010 9:38 PM > To: Sean Hefty > Cc: 'ofw_list' > Subject: Re: [ofw] ctrl-c of appl > > Thanks, Sean. > > Yes, I'm running ib_send_bw on Windows. If I kill the > application before it has received a connection from a > server, the ctrl-c works fine and see the destroy_obj calls > that result in mr being deregistered. > > However, once the application posts recvs & sends and is > waiting in poll-cq, the ctrl-c does not cleanup properly. > > Usha > > -----Original Message----- > From: Sean Hefty [mailto:[email protected]] > Sent: Wednesday, February 03, 2010 1:51 PM > To: Usha Srinivasan > Cc: 'ofw_list' > Subject: RE: ctrl-c of appl > > copying ofw_list, assuming this is a windows issue > > >I am running ib-send-bw and normally I can control C it at various > >points and everything gets torn down, for example, deregister-mr is > >called. But, when it is sitting in poll-cq, if I do a > control C, deregister-mr does not get called > >and I get a bug check saying 'blah' pages are still mapped. > Is this a known > >problem? What's an hca driver to do if umcloseca is called but > >resources haven't been destroyed/deregistered? > > I'm guessing that this is not a known problem. crtl-c should > always work to kill the app and cleanup all kernel resources. > > I'm guessing that you're running on windows. Is that > correct? If so, then ib_send_bw runs over ibal, which sounds > like it's not cleaning up properly. > > IMO, an HCA should assume that all resources associated with > an HCA will have been released before closeca is called. It > could use reference checks to verify this and return an error > on the close, or assert that the reference is 0 when close is > called. (This is almost certainly an error in the calling > program, rather than something that can be handled at run time.) > > - Sean > > _______________________________________________ > ofw mailing list > [email protected] > http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw > _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
