On Tue, Sep 20, 2016 at 09:27:57AM -0400, David Hill wrote: > On Tue, Sep 20, 2016 at 03:16:50PM +0200, Alexander Bluhm wrote: > > On Tue, Sep 20, 2016 at 08:21:55AM -0400, David Hill wrote: > > > With bluhm's r1.160 uipc_socket.c. > > > Here are the first ones that were detected. > > > > Thanks for the fast report. > > > > So fifo works around the socket layer. Better call soconnect2() > > instead of unp_connect2(). This adds the missing splsoftnet(). > > > > I think we should demand that socantsendmore() and socantrcvmore() > > in uipc_socket2.c should be called with splsoftnet(). > > Should both socantsendmore() and socantrcvmore() get a splsoftassert() > then, for now?
No, not more asserts please :-) The goal is to find one entry layer in the kernel where we can put the network lock. I is pretty much where we have the splsoftnet() now. The fifo, unix, pfkey and routing sockets don't work well with this idea now. Maybe I backout the splsoftasserts in the wakeup code if I cannot fix it fast. bluhm
