[PATCH v1 0/9] Fix Allwinner D1 boot regression

[Palmer Dabbelt]

On Wed, 14 Aug 2024 10:30:48 PDT (-0700), tglx at linutronix.de wrote:
> On Wed, Aug 14 2024 at 16:56, Emil Renner Berthing wrote:
>> As described in the thread below[1] I haven't been able to boot my
>> boards based on the Allwinner D1 SoC since 6.9 where you converted the
>> SiFive PLIC driver to a platform driver.
>>
>> This is clearly a regression and there haven't really been much progress
>> on fixing the issue since then, so here is the revert that fixes it.
>>
>> If no other fix is found before 6.11 I suggest we apply this.
>
> So this mess has been ignored for two month now?
>
>>From the pastebin in the initial report:
>
> [    0.000000] irq: no irq domain found for interrupt-controller at 10000000 !
> [    0.000000] Failed to map interrupt for /soc/timer at 2050000
> [    0.000000] Failed to initialize '/soc/timer at 2050000': -22
>
> This comes back with -EINVAL. So the timer cannot find an interrupt,
> which makes it pretty obvious why the system stops to boot, unless there
> is some other timer available.
>
> This is obviously related to the SUN4I_TIMER because that message went
> away when it was disabled according to the next pastebin.
>
> Obviously that can't work because the SUN4I timer driver is using
> timer_of_init() which cannot handle deferred probing.
>
> Daniel: There was a partial fix for the sun4i driver, which you said you
> applied:
>
>   https://lore.kernel.org/all/20240312192519.1602493-1-samuel.holland@sifive.com
>
> But that thing never materialized in a pull request.
>
> And of course everyone involved ignored the problem since March 13th
> 2024, i.e. almost half a year.
>
> Seriously?
>
> Can you RISCV folks get your act together and ensure to fix things you
> broke on the way? Especially when Emil reported this nobody pointed him
> to this patch and nobody noticed that it's still not merged?
>
> It took me less than 15 minutes to find that patch and the correlation,
> but this is absolutely not my job.

Sorry, I guess I'd just sort of been ignoring the platform-specific side 
of things because it's so frustrating to deal with, but that's led to a 
bunch of breakages so it's obviously the wrong thing to do.

> I'm seriously grumpy about that. This is not how it works. If you break
> stuff, then you take care to fix it before you shove more changes into
> the tree and waste my time.
>
> I'm very much inclined to take the reverts right now, send them to Linus
> for -rc5 tagged with cc: stable and ignore/nak any irqchip related riscv
> patches until the next merge window is over.

Acked-by: Palmer Dabbelt <palmer at rivosinc.com>

if you want to take the revert.

IIUC the patch above doesn't actually fix it, that's what led to just 
sending the reverts -- at least reverts are better than breaking users.  
I'll post over there too...

And it's no big deal if we're in the doghouse for a bit.  Regressions 
should get fixed faster than this, so we deserve it.

Probably also another sign we're way too focused on getting new features 
merged, as that's coming at the expense of making existing platforms 
work.  IMO we've been way too focused on getting support for specs that 
don't even have implementations, and not enough on building real working 
systems.

> Emil, can you give that sun4i fix a test ride please?
>
> Thanks,
>
>         tglx