Re: v3.9 - CPU hotplug and microcode earlier loading hits a mutex deadlock (x86_cpu_hotplug_driver_mutex)

[Toshi Kani <toshi.kani@hp.com>]

On Wed, 2013-05-08 at 14:42 -0400, Konrad Rzeszutek Wilk wrote:
> On Wed, May 08, 2013 at 12:19:45PM -0600, Toshi Kani wrote:
> > On Wed, 2013-05-08 at 12:32 -0400, Konrad Rzeszutek Wilk wrote:
> > > On Wed, May 08, 2013 at 04:29:49PM +0200, Borislav Petkov wrote:
> > > > On Wed, May 08, 2013 at 10:03:42AM -0400, Konrad Rzeszutek Wilk wrote:
> > > > 
> > > > [ … snip some funky BIOS code ]
> > > > 
> > > > > [here it shifts and continues on testing each CPU bit]
> > > > > 
> > > > > > Questions over questions...?
> > > > > 
> > > > > I probably went overboard with my answers :-)
> > > > 
> > > > Konrad, you're killing me! :-) You actually went and looked at the
> > > > BIOS disassembly voluntarily. You must be insane, I think you should
> > > > immediately go to the doctor now for a thorough checkup. :-)
> > > > 
> > > > I think I know who I can sling BIOS issues now to.
> > > 
> > > Great .. :-)
> > > > 
> > > > > > Looks like save_mc_for_early would need another, local mutex to fix that.
> > > > > 
> > > > > Let me try that. Thanks for the suggestion.
> > > > 
> > > > Ok, seriously now: yeah, this was just an idea, it should at least get
> > > > the nesting out of the way.
> > > > 
> > > > About the BIOS deal: you're probably staring at some BIOS out there
> > > > but is this the way that it is actually going to be implemented on
> > > > the physical hotplug BIOS? I mean, I've only heard rumors about IVB
> > > > supporting physical hotplug but do you even have access to such BIOS to
> > > > verify?
> > > 
> > > Unfortunatly not. I am getting an IvyTown box so hopefully that has this
> > > support. But I thought that Fenghua did since he mentioned in the patch.
> > > 
> > > Besides that I think this can also appear on VMWare if one is doing 
> > > CPU hotplug and on some HP machines - let me CC the relevant people
> > > extracted from drivers/acpi/processor_driver.c.
> > > (see  https://lkml.org/lkml/2013/5/7/588 for the thread)
> > 
> > >From the stack trace, it looks like a bug in save_mc_for_early().  This
> > function may not call cpu_hotplug_driver_lock() during CPU online.  I
> > suppose it intends to protect from CPU offline when microcode is updated
> > outside of the boot/CPU online context.  If it indeed supports updating
> > the microcode without using reboot/cpu hotplug, the lock should be held
> > when such update request is made.
> 
> 
> Hey Toshi,
> 
> The fix for this I have posted, but I am more curious whether you have
> seen on baremetal this issue? Meaning using CPU ACPI hotplug on v3.9?

Oh, I see.  No, I do not have a beremetal test machine that supports CPU
hotplug yet.  My initial target is to support CPU hotplug on VMs.  I did
not hit this issue since my test env was limited to CPU hot-delete
followed by CPU hot-add.  In such case, uci->valid is still set during
CPU hot-add and does not get into this code path.  I am not sure if
uci->valid is supposed to be set after CPU hot-delete, though.

Thanks,
-Toshi