Re: Eventtimers b0rking w/ math/atlas

From: Alexander Motin <mav_at_FreeBSD.org>
Date: Fri, 10 Sep 2010 18:31:03 +0300
Hi.

Thomas E. Spanjaard wrote:
> While trying to build math/atlas on a FreeBSD/amd64 9.0-CURRENT r212411,
> the kernel hangs at some point when math/atlas tries to run some tests
> (presumably the ones to profile the code and optimise). The kernel
> spouts messages about Starting event timers: LAPIC _at_ 1000Hz, HPET _at_
> 127Hz; then LAPIC _at_ 1000Hz, HPET _at_ 8128Hz (iirc), back, then back again.
> After that, the system is no longer responsive, and eventually panic()s
> because some spinlock has been held too long in the pmap TLB invalidate
> code. Couldn't get a dump, because I had no dumpdev configured.
> 
> Anyone else see this problem? I'm trying to reproduce it by building
> other ports (it has survived gcc44, gcc45, llvm and clang so far), but
> only math/atlas seems to trigger it.

It is reported deadlock between event timers and some IPI senders, like
TLB invalidation, during switching to/from profiling clock rate. In
forthcoming version of event timer patch this problem should not happen.
You can get latest version of the patch here:
http://people.freebsd.org/~mav/timers_oneshot13.patch

-- 
Alexander Motin
Received on Fri Sep 10 2010 - 13:31:23 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:07 UTC