PLEASE TEST: IPI deadlock avoidance patch
Brian Fundakowski Feldman
green at freebsd.org
Mon Aug 23 09:19:40 PDT 2004
On Mon, Aug 23, 2004 at 11:41:19AM -0400, Brian Fundakowski Feldman wrote:
> On Sun, Aug 22, 2004 at 12:05:39PM -0700, Doug White wrote:
> > Hey folks,
> > If you have a reasonably fast i386 or amd64 multiprocessor and/or
> > hyperthreading machine and are experiencing reproducible hangs during -j
> > buildwords and other highly parallel operations, please try this patch:
> > http://people.freebsd.org/~dwhite/smp_rv_mtx.patch
> > This patch avoids a deadlock between the smp_rendezvous() mechanism and
> > TLB shootdowns via pmap by forcing them to share a mutex. The patch has
> > been tested on several multiproc Xeon systems and a dual Opteron I have
> > here and is stable, but a couple of more eyeballs and varied loads
> > wouldn't hurt since this will be going into RELENG_5.
> > If there's anything glaringly wrong from a style point of view, let me
> > know too.
> > I plan on committing this late Monday or Tuesday, barring any serious
> > problems.
> > Thanks to alc for the nudge on the solution and rwatson for staying up
> > late and testing it on his hardware.
> I couldn't even run portupgrade without a hard hang on my Athlon!
Oops, let me clarify: "... before applying these changes." My system
is very stable under normal workload after applying them, but I also
could not make a -jwhatever buildworld hang for me previously like
Brian Fundakowski Feldman \'[ FreeBSD ]''''''''''\
<> green at FreeBSD.org \ The Power to Serve! \
Opinions expressed are my own. \,,,,,,,,,,,,,,,,,,,,,,\
More information about the freebsd-current