What's the p2p_daemon?

[mvpel]

One of the threads in an HPMC crash dump is in that WCHAN, and I can't seem to find any information about it or the __cnx_p2p_ctl kernel routine (nm -b /stand/vmunix | grep cnx).

The fact that CNX was Convex's NYSE ticker symbol up until December 20, 1995 is about all I've been able to winkle out of Google.

Any ideas? Thanks!

[Don Morris_1 · Accepted solution]

The p2p_daemon is there to manage process-to-process pagefaults. That's to service functionality for process-to-process [across address space] bcopy operations, used for MPI performance enhancement.

[mvpel]

Okay, so the process which trapped (type 18, data memory protection fault) was trying to perform this copy, looks like, and it stepped on something. Unfortunately it's in user space so there's nothing to unwind in the crash dump.

Is the creation of a separate thread normal when this process-to-process copy is taking place? It looks like the p2p daemon was sleeping in one thread (TSDETACHED|TSSYS) and the other thread which trapped was the original (TSFIRSTTHREAD).

Is there some documentation of the normal procedure for the kernel's use of the p2p_daemon? I'm seeing some unusual numbers and I'm not sure if they're typical.

[mvpel]

Nevermind, I was able to find a live process using the p2p_daemon, and it looks like the thread is spawned at the first p2p operation and is just left there sleeping in case there's any additional p2p operations in the future.