Re: [PREEMPT-RT] Oops in rapl_cpu_prepare()

From: Charles (Chas) Williams
Date: Fri Oct 21 2016 - 17:04:12 EST

On 10/21/2016 06:56 AM, Sebastian Andrzej Siewior wrote:
On 2016-10-20 16:27:55 [-0400], Charles (Chas) Williams wrote:
Recent 4.8 kernels have been oopsing when running under VMWare:

can you reproduce this on bare metal?

I can't get dedicated access to the specific bare metal since it is
running as a dedicated hypervisor. I haven't seen this issue anywhere
else though with the 4.8 kernel.

[ 2.270203] BUG: unable to handle kernel NULL pointer dereference at 0000000000000408
[ 2.270325] IP: [<ffffffff81012bb9>] rapl_cpu_online+0x59/0x70

can you check if pmu is NULL?

It's not. The dereference at 0x408 and pmu->cpu being fairly early in
the struct seems to indicate that pmu wasn't pointing to 0 at the time
(but fairly close). I should have noticed that earlier.

Is there a particular order guaranteed by the callbacks? Will
rapl_cpu_prepare() always happen before online/offline? Additionally,

yes, see include/linux/cpuhotplug.h. On CPU-up the array ids are invoked

Yes, I see that now. Thanks for the pointer!

If a callback (such as CPUHP_PERF_X86_RAPL_PREP) fail then we rollback
to the starting point (in case of CPU up it would be CPUHP_OFFLINE.

You'll like this, I just did a little printk debugging because it was
easier than trying to get a debugger running:

[ 3.107126] init_rapl_pmus: maxpkg 4
[ 3.107263] rapl_cpu_prepare: pmu ffff880234faa540 cpu 0 pkgid 0
[ 3.107400] rapl_cpu_prepare: pmu ffff880234faa600 cpu 1 pkgid 2
[ 3.107537] rapl_cpu_prepare: pmu ffff880234faa6c0 cpu 2 pkgid 65535
[ 3.107662] rapl_cpu_online: pmu ffff880234faa540 cpu 0 pkgid 0
[ 3.107907] rapl_cpu_online: pmu ffff880234faa600 cpu 1 pkgid 2
[ 3.108133] rapl_cpu_online: pmu ffff880234faa6c0 cpu 2 pkgid 65535
[ 3.108333] rapl_cpu_online: pmu ffff880234faa6c0 cpu 3 pkgid 65535

where pkgid is topology_logical_package_id(cpu).

I can't understand why I don't see a cpu 3 during cpu prepare, when I
see one later. The 65535 is a -1 from topology_phys_to_logical_pkg()
getting assigned to the logical_proc_id apparently.

So this is pretty puzzling. Since this is a guest running under VMWare, I
don't know that there is any particular CPU pinning or emulation of RAPL.

It looks there was a proposal to not run in guests: