Re: [Regression 5.6-rc1][Bisected b6231ea2b3c6] Powerpc 8xx doesn't boot anymore

On 02/13/2020 07:45 AM, Rasmus Villemoes wrote:

On 12/02/2020 15.24, Christophe Leroy wrote:

quoted

Hi Rasmus,

Kernel 5.6-rc1 silently fails on boot.

I bisected the problem to commit b6231ea2b3c6 ("soc: fsl: qe: drop
broken lazy call of cpm_muram_init()")

I get a bad_page_fault() for an access at address 8 in
cpm_muram_alloc_common(), called from cpm_uart_console_setup() via
cpm_uart_allocbuf()

Sorry about that. But I'm afraid I don't see what I could have done
differently - the patch series, including b6231ea2b3c6, has been in
-next since 20191210, both you and ppc-dev were cc'ed on the entire
series (last revision sent November 28). And I've been dogfooding the
patches on both arm- and ppc-derived boards ever since (but obviously
only for a few cpus).

Yes, this patch series should have ringed a bell in my head, looks like 
I'm the one who introduced this 4 years ago through commit 4d486e008379 
("soc/fsl/qe: fix Oops on CPM1 (and likely CPM2)")

But I had completely forgotten that patch until I did some git blame 
this morning on this lazy call.

quoted

Reverting the guilty commit on top of 5.6-rc1 is not trivial.

In your commit text you explain that cpm_muram_init() is called via
subsys_initcall. But console init is done before that, so it cannot work.

No, but neither did the code I removed seem to work - how does doing
spin_lock_init on a held spinlock, and then unlocking it, work? Is
everything-spinlock always a no-op in your configuration? And even so,
I'd think a GFP_KERNEL allocation under spin_lock_irqsave() would
trigger some splat somewhere?

Please note I'm not claiming my patch is not at fault, it clearly is, I
just want to try to understand how I could have been wrong about the
"nobody can have been relying on it" part.

It seems spin_lock_init() does just nothing.
spin_lock_irqsave() just disable IRQs and increases preempt_count.
spin_lock_irqrestore() restore IRQ state, decreace preempt_count and 
call preempt_schedule if preempt_count reaches 0.

Maybe with some debugging options like DEBUG_ATOMIC_SLEEP could detect it ?

quoted

Do you have a fix for that ?

Not right now, but I'll have a look. It's true that the patch probably
doesn't revert cleanly, but it shouldn't be hard to add back those few
lines in the appropriate spot, with a big fat comment that this does
something very fishy (at least as a temporary measure if we don't find a
proper solution soonish).

Thanks
Christophe