Re: [PATCH] init: Add lockdep annotation to kthreadd_done completer

[Helge Deller <deller@gmx.de>]

On 8/11/23 18:43, Andrew Morton wrote:
> On Fri, 11 Aug 2023 18:04:22 +0200 Helge Deller <deller@gmx.de> wrote:
>
>> Add the missing lockdep annotation to avoid this warning:
>>
>>   INFO: trying to register non-static key.
>>   The code is fine but needs lockdep annotation, or maybe
>>   you didn't initialize this object before use?
>>   turning off the locking correctness validator.
>>   CPU: 0 PID: 1 Comm: swapper/0 Not tainted 6.5.0-rc5+ #681
>>   Hardware name: 9000/785/C3700
>>   Backtrace:
>>    [<000000004030bcd0>] show_stack+0x74/0xb0
>>    [<0000000041469c7c>] dump_stack_lvl+0x104/0x180
>>    [<0000000041469d2c>] dump_stack+0x34/0x48
>>    [<000000004040e5b4>] register_lock_class+0xd24/0xd30
>>    [<000000004040c21c>] __lock_acquire.isra.0+0xb4/0xac8
>>    [<000000004040cd60>] lock_acquire+0x130/0x298
>>    [<000000004146df54>] _raw_spin_lock_irq+0x60/0xb8
>>    [<0000000041472044>] wait_for_completion+0xa0/0x2d0
>>    [<000000004146b544>] kernel_init+0x48/0x3a8
>>    [<0000000040302020>] ret_from_kernel_thread+0x20/0x28
>>
>> ...
>>
>> --- a/init/main.c
>> +++ b/init/main.c
>> @@ -682,6 +682,8 @@ noinline void __ref __noreturn rest_init(void)
>>   	struct task_struct *tsk;
>>   	int pid;
>>
>> +	init_completion(&kthreadd_done);
>> +
>>   	rcu_scheduler_starting();
>>   	/*
>>   	 * We need to spawn init first so that it obtains pid 1, however
>
> This is pretty old code, as is the page_alloc_init_late() change.  Do
> we know why this warning has just turned up lately?

I haven't debugged in depth yet, but here is what I believe is the reason
why I do see those lockdep warnings and others not.

I'm building & testing on the parisc platform.
Just recently I added lockdep support to parisc for kernel 6.4 and backported
it to v6.0+. Since then I've seen the warnings.

And I think the main reason why I see those warnings on parisc and
others on other platforms don't, is that parisc is the only architecture
where __ARCH_SPIN_LOCK_UNLOCKED_VAL is NOT zero (0x1a46 actually).
The reason is that parisc offers only one assembler instruction which
operates atomicly on memory, and which we use to lock spinlocks:
ldcw ("load and clear word").
So, a "zero" spinlock word means the lock is locked, while non-zero means
it's unlocked.
For other platforms it's the other way around.

So, for a structure in e.g. __initdata[] which is pre-initialized by the compiler,
the spinlocks are locked by default (lockword = 0) on parisc, if they haven't
been initialized correctly, and thus the kernel will complain at runtime.

Now, maybe lockdep doesn't use spinlocks per se. I have't checked in depth yet,
but I'm sure it's somehow related to the odd __ARCH_SPIN_LOCK_UNLOCKED_VAL value
of parisc.

I wonder if the same bug appears if you use a non-zero __ARCH_SPIN_LOCK_UNLOCKED_VAL
on other platforms too (if it's possible).

> I'll add cc:stable to these, but might take that away again if we can
> answer the above.

Thanks for adding the patches.
I did sent two other patches as well: for watchdog and devtmpfs..

> btw, please don't forget the "^---$" between changelog and patch.

Ok.

Helge