On Nov 23 14:33, Corinna Vinschen wrote: > On Nov 23 14:10, Corinna Vinschen wrote: > > On Nov 23 12:36, Corinna Vinschen wrote: > > > On Nov 19 21:06, Lavrentiev, Anton (NIH/NLM/NCBI) [C] wrote: > > > > Hello again, > > > > > > > > I can now positively confirm the race condition in cygserver w.r.t. the > > > > named > > > > pipe used to serialize SYSV requests through the server. The race is > > > > due to > > > > that transport_layer_pipes::accept() (bool *const recoverable) (file: > > > > transport_pipes.cc) does actually _create_ the pipe when pipe_instance > > > > == 0 > > > > (ironically, transport_layer_pipes::listen() does not create any OS > > > > primitives > > > > at all!). > > > > > > > > This means that under heavy load, cygserver threads may all end up > > > > processing > > > > their requests and closing all instances of the pipe (bringing > > > > pipe_instance == 0) > > > > yet not being able to get to the point of accepting new request (that > > > > is, to > > > > re-create the pipe). For the client (user process), this looks like > > > > the pipe > > > > does not exist (during that very tiny period of time), and the > > > > following message > > > > gets printed: > > > > > > > > Iteration 3016 > > > > 1 [main] a 4872 transport_layer_pipes::connect: lost connection > > > > to cygserver, error = 2 > > > > > > Thanks for analyzing this situation. IIUC, that means if we create the > > > pipe in listen(), and then create another pipe instance per accepting > > > thread in accept(), we will always have at least one instance of the pipe, > > > whatever the load, right? > > > > > > Are you set up to test a patch? If so, I'd propose the below patch. It > > > would be nice if you could test it in your environment. > > > > Forget it. That doesn't work. > > > > Surprisingly, the next client calling CreateFile is *not* connected to > > any server side of the pipe which calls ConnectNamedPipe, but apparently > > the pipes are used in the order of creation. Thus, the first client > > hangs waiting for the pipe instance created in the listen() method, but > > since the server never calls ConnectNamedPipe on that instance, the > > client will wait forever. > > > > Bummer. > > > > Back to the drawing board... > > Try the below one. It also connects to the listening pipe in listen(), > so there's a single instance of the pipe connected and blocked for > further use forever. This should avoid the race (*and* work...) > Please give it a try.
You can also simply try the today's snapshot version of cygserver from http://cygwin.com/snapshots/ It contains this patch, as well as the semaphore patch I mailed to the other thread. Corinna -- Corinna Vinschen Please, send mails regarding Cygwin to Cygwin Project Co-Leader cygwin AT cygwin DOT com Red Hat -- Problem reports: http://cygwin.com/problems.html FAQ: http://cygwin.com/faq/ Documentation: http://cygwin.com/docs.html Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple
