Re: [PATCH][RFC] workqueue: Fix kernel panic on CPU hot-unplug

[Helge Deller]

Hi Tejun,

On 2/2/24 18:29, Tejun Heo wrote:
> Hello, Helge.
>
> On Fri, Feb 02, 2024 at 09:41:38AM +0100, Helge Deller wrote:
> > In a second step I extended your patch to print the present
> > and online CPUs too. Below is the relevant dmesg part.
> >
> > Note, that on parisc the second CPU will be activated later in the
> > boot process, after the kernel has the inventory.
> > This I think differs vs x86, where all CPUs are available earlier
> > in the boot process.
> > ...
> > [    0.000000] XXX workqueue_init_early: possible_cpus=ffff  present=0001  online=0001
> ...
> > [    0.228080] XXX workqueue_init: possible_cpus=ffff  present=0001  online=0001
> ...
> > [    0.263466] XXX workqueue_init_topology: possible_cpus=ffff  present=0001  online=0001
>
> So, what's bothersome is that when the wq_dump.py script printing each cpu's
> pwq, it's only printing for CPU 0 and 1. The for_each_possible_cpu() drgn
> helper reads cpu_possible_mask from the kernel and iterates that, so that
> most likely indicates at some point the cpu_possible_mask becomes 0x3
> instead of the one used during boot - 0xffff, which is problematic.
>
> Can you please sprinkle more printks to find out whether and when the
> cpu_possible_mask changes during boot?

It seems the commit 0921244f6f4f ("parisc: Only list existing CPUs in cpu_possible_mask")
is the culprit. Reverting that patch makes cpu hot-unplug work again.
Furthermore this commit breaks the cpumask Kunit test as reported by Guenter:
https://lkml.org/lkml/2024/2/4/146

So, I've added the revert to the parisc git tree and if my further tests
go well I'll push it upstream.

Thanks for your help!!
Helge