Re: [PATCH -next 00/13] Extensible console matching & direct earlycon

[Peter Hurley <peter@hurleysoftware.com>]

[ +cc devicetree, Mark Rutland, Grant Likely ]

On 02/26/2015 09:58 AM, Rob Herring wrote:
> On Thu, Feb 26, 2015 at 8:48 AM, Peter Hurley <peter@hurleysoftware.com> wrote:
>> Hi Rob,
>>
>> On 02/24/2015 03:20 PM, Rob Herring wrote:
>>> On Tue, Feb 24, 2015 at 1:53 PM, Peter Hurley <peter@hurleysoftware.com> wrote:
>>
>> [...]
>>
>>>>>> Direct earlycon
>>>>>>
>>>>>> This feature enables arches and proms to start an earlycon directly,
>>>>>> rather than requiring an "earlycon=" command line parameter.
>>>>>> Devicetree can already do this via the 'linux,stdout-path' property,
>>>>>> but arch and prom code requires direct coupling to the serial driver.
>>>>>>
>>>>>> This support is implemented by judicious refactoring and the same
>>>>>> construct that devicetree and early_param use: a link table containing
>>>>>> the necessary information (name and setup() function) to find and
>>>>>> bind the appropriate earlycon "driver".
>>>>>
>>>>> I've skimmed thru this and it looks like a great improvement.
>>>>>
>>>>> One problem we have currently with DT stdout-path and earlycon is a
>>>>> preferred console does not get registered, so the console will get
>>>>> switched to tty0 and you lose your console. The problem is DT does not
>>>>> know the console name to register a preferred console. It looks like
>>>>> this series may help this problem, but I'm not sure and wanted your
>>>>> thoughts.
>>>>
>>>> I thought that of_alias_scan() + of_console_check() caused DT stdout-path
>>>> to add_preferred_console() the driver console @ port registration time
>>>> via uart_add_one_port() -> of_console_check().
>>>>
>>>> Is that not how that works?
>>>
>>> Yes, I believe that is how it works with earlycon not enabled. This
>>> doesn't work when earlycon is enabled with just "earlycon" on the
>>> command line. The fix I have is here[1], but I don't like putting DT
>>> specifics into the console code.
>>
>> After much gnashing of teeth and pulling of hair yesterday, I managed
>> to mock up the situation you describe, but I need to study it in more
>> detail. Some things I did learn:
>>
>> 1. The serial console _does_ come back up when using stdout-path but the
>>    line settings don't match, because the serial core sets them to the
>>    default of 9600n8 if unspecified.
> 
> That may have been what I saw as I tested on QEMU which ignores the
> baud rate. But it does stop between the time tty0 is enabled and the
> "real" serial console which is a time period we really want the
> console.

The bug is that of_alias_scan() does not emit a preferred console, but
rather defers it until of_console_check().

In many setups, this means there is no preferred console. So when the
dummy vt console loads (which happens via console_init()), all the boot
consoles are disabled.

The trivial fix is to emit a preferred console from of_alias_scan(), even
if that console is never matched. I've sent you a patch to do just that.


>> 2. The line settings can now be set with stdout-path like,
>>      stdout-path = "serial0:115200n8"
>>    but this breaks DT earlycon (as I wrote in the other email you were
>>    cc'd on).
> 
> Right. We should fix libfdt.

Ok.

Regards,
Peter Hurley