Re: [BISECTED] Regression: Solidrun Clearfog Base won't boot since "PCI: mvebu: Only remap I/O space if configured"

[Baruch Siach]

Hi Russell,

Russell King - ARM Linux writes:
> On Wed, Sep 12, 2018 at 09:49:41PM +0300, Baruch Siach wrote:
>> I reproduced the same Oops on Clearfog Base without any taint:
>> 
>> [    1.476401] Internal error: Oops - undefined instruction: 0 [#1] SMP ARM
> ...
>> [    1.855954] Code: e2844004 e5972000 e3520000 0affffee (e7f001f2)
>
> That is a BUG().  Please turn on verbose bug reporting to get more
> information about the cause.
>
> There are two possibilities:
>
>         BUG_ON(addr >= end);
>
> and
>
>                 BUG_ON(!pte_none(*pte));
>
> It's probably the latter - the region is probably already mapped, that
> being the PCI IO region.

That is the one. Enabling CONFIG_DEBUG_BUGVERBOSE shows:

[    1.481927] kernel BUG at lib/ioremap.c:72!
[    1.486118] Internal error: Oops - BUG: 0 [#1] SMP ARM
[    1.491269] CPU: 0 PID: 1246 Comm: kworker/0:2 Not tainted 4.19.0-rc3 #146
...

baruch

> The original driver was setup to call pci_ioremap_io() as the very
> last thing - and as the driver is non-removable, we were guaranteed
> to never tear down this mapping (which is sensible, it's published
> to userspace.)
>
> However, the current code calls pci_ioremap_io() much earlier, in
> a path where probe failures can occur.  This breaks pci_ioremap_io()'s
> requirements - it must not be called more than once.  So:
>
> ee1604381a37 ("PCI: mvebu: Only remap I/O space if configured")
>
> is basically incorrect - pci_ioremap_io() needs to move back to a
> place where it is only called in a path which will never fail.
> However, looking at the generic host bits, I'm not sure such a place
> exists in the new effort to make stuff more generic.

-- 
     http://baruch.siach.name/blog/                  ~. .~   Tk Open Systems
=}------------------------------------------------ooO--U--Ooo------------{=
   - baruch@xxxxxxxxxx - tel: +972.52.368.4656, http://www.tkos.co.il -