On 09/21/2017 02:57 AM, Michael Ellerman wrote:

Tyrel Datwyler [off-list ref] writes:

quoted

On 09/20/2017 04:39 AM, Michael Ellerman wrote:

quoted

Rob Herring [off-list ref] writes:

<snip>

quoted

quoted

Testing a fix, will report back.

So, that patch slipped past me. Not only is the parent reference not ours to drop, but
when I went and looked at dlpar_cpu_add() I also noticed that of_node_put() was done on
the parent prior to the call to dlpar_attach_node(). With the addition of "parent" to the
dlpar_attach_node() parameter list dlpar_cpu_add() needs to be fixed up to hold the
"parent" reference until after dlpar_attach_node() returns.

Yep. I wrote the same patch :)

Rob asked me to test it, which I did, but /cpus starts out with an
elevated ref count, so you have to do ~30 (on my system) DLPAR removes
to hit the bug, which I didn't do.

Yeah, there are a lot of things that grab references to /cpus. So, I had a good idea that
I needed to loop a few times adding and removing multiple cpus to trigger the issue. Its
also obvious when using those OF trace points I wrote a while back that refcount for /cpus
is dropping off uncharacteristically in response to symmetrical adds/removes of cpus. I
saw your note about getting that patchset resubmitted. I'll try and get that queued back
up soon.

-Tyrel

I've updated my test script to do roughly $(nproc) x 10 DLPAR removes,
which is hopefully sufficient to catch these bugs in future.

cheers