On 22/01/15 11:09, Mark Goodwin wrote:
Debian verge 32bit systems (and possibly others) either don't fake a
numa node on single CPU systems, or don't export the node-cpu mapping
info in /sysfs. This was causing some buffer overrun issues found by
valgrind when hinv.nnode is zero but it was assumed to be at least one
on all systems.
I haven't committed this because I don't have a debian/i686 system
handy to test it on. I've only tested it on a fedora/4cpu/1node
system.
Thanks Mark.
I don't think the issue is debian/fedora related ... my first failing
case I looked at was vm04 which is CentOS 5.11.
Anyway applying the patch to vm04 and remaking and installing the linux
PMDA leads to qa/873 passing (which is good).
check -g pmda.linux also passes on vm04, which suggests no glaring
regression at least for this one platform.
qa/873 was also failing for me on my Fedora20 platform (vm23).
After applying the patch there, qa/873 is passing now.
So, looks good to me.
|