On 04/02/2015 09:32 AM, Grant Likely wrote:
> On Sat, 28 Mar 2015 13:01:24 -0400
> , Peter Hurley <peter@xxxxxxxxxxxxxxxxxx>
> wrote:
>> Hi Grant,
>>
>> On 03/27/2015 09:36 PM, Grant Likely wrote:
>>> On Sun, 01 Mar 2015 17:23:11 -0500
>>> , Peter Hurley <peter@xxxxxxxxxxxxxxxxxx>
>>> wrote:
>>>> Hi Kevin,
>>>>
>>>> On 11/24/2014 06:36 PM, Kevin Cernekee wrote:
>>>>> If an earlycon (stdout-path) node is being used, check for "big-endian"
>>>>> or "native-endian" properties and pass the appropriate iotype to the
>>>>> driver.
>>>>>
>>>>> Note that LE sets UPIO_MEM (8-bit) but BE sets UPIO_MEM32BE (32-bit). The
>>>>> big-endian property only really makes sense in the context of 32-bit
>>>>> registers, since 8-bit accesses never require data swapping.
>>>>>
>>>>> At some point, the of_earlycon code may want to pass in the reg-io-width,
>>>>> reg-offset, and reg-shift parameters too.
>>>>>
>>>>> Signed-off-by: Kevin Cernekee <cernekee@xxxxxxxxx>
>>>>> ---
>>>>> drivers/of/fdt.c | 7 ++++++-
>>>>> drivers/tty/serial/earlycon.c | 4 ++--
>>>>> include/linux/serial_core.h | 2 +-
>>>>> 3 files changed, 9 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/drivers/of/fdt.c b/drivers/of/fdt.c
>>>>> index 658656f..9d21472 100644
>>>>> --- a/drivers/of/fdt.c
>>>>> +++ b/drivers/of/fdt.c
>>>>> @@ -794,6 +794,8 @@ int __init early_init_dt_scan_chosen_serial(void)
>>>>>
>>>>> while (match->compatible[0]) {
>>>>> unsigned long addr;
>>>>> + unsigned char iotype = UPIO_MEM;
>>>>> +
>>>>> if (fdt_node_check_compatible(fdt, offset, match->compatible)) {
>>>>> match++;
>>>>> continue;
>>>>> @@ -803,7 +805,10 @@ int __init early_init_dt_scan_chosen_serial(void)
>>>>> if (!addr)
>>>>> return -ENXIO;
>>>>>
>>>>> - of_setup_earlycon(addr, match->data);
>>>>> + if (of_fdt_is_big_endian(fdt, offset))
>>>>> + iotype = UPIO_MEM32BE;
>>>>> +
>>>>> + of_setup_earlycon(addr, iotype, match->data);
>>>>
>>>> I know these got ACKs already but as you point out in the commit log,
>>>> earlycon _will_ need reg-io-width, reg-offset and reg-shift. Since the
>>>> distinction between early_init_dt_scan_chosen_serial() and
>>>> of_setup_earlycon() is arbitrary, I'd rather see of_setup_earlycon()
>>>> taught to properly decode of_serial driver bindings instead of a
>>>> stack of parameters to of_setup_earlycon().
>>>>
>>>> In fact, this patch allows a mis-defined devicetree to bring up a
>>>> functioning earlycon because the 'big-endian' property is directly
>>>> associated with UPIO_MEM32BE, which will create incompatibility problems
>>>> when DT earlycon is fixed to decode the of_serial DT bindings.
>>>
>>> That's a good point. This hasn't been merged yet, so there isn't any
>>> impact on addressing this. I would propose that for consistency, the
>>> earlycon code should always default to 8-bit access. if big-endian
>>> accesses are required, then reg-io-width + big-endian must be specified.
>>>
>>> Something like the following would do it and would be future-proof. We
>>> can add support for 16 or 64bit big or little endian access if it ever
>>> became necessary.
>>
>> I was planning on adding MEM32BE support to OF earlycon on top of my
>> patch series 'OF earlycon cleanup', which adds full support for the
>> of_serial driver DT properties (among other things).
>>
>> Unfortunately, that series is waiting on two things:
>> 1. libfdt upstream patch, which I submitted but was referred back to me
>> to add test cases. That was 3 weeks ago and I simply haven't had a free
>> day to burn to figure out how their test matrix is organized. I don't
>> think that's going to change anytime soon; I might just abandon that patch
>> and do the string manipulation on the stack.
>>
>> ATM, earlycon is still broken if stdout-path options have been set.
>
> I don't seem to have that patch. Can you send it to me please?
Will do.
> I do have a thought though. Would it be better to teach
> fdt_path_offset() to recognize the ':' delimiter? It's never a valid
> character for a path.
>
> The unittests are easy. "make check" builds and runs them. Adding a test
> is as simple as editing tests/parent_offset.c. main() calls check_path()
> several times to test calls to fdt_path_offset(). The tests can be added
> directly to that file.
Well, the patch reimplements fdt_path_offset() in terms of a new
length-limited function, fdt_path_offset_namelen(). So the existing tests
of fdt_path_offset() are already exercising the patch.
The issue is that the unit tests for fdt_path_offset_namelen() will need
additional nodes and properties defined to properly test the functionality,
and it's not clear to which fdt files these changes are required.
Not that I can't figure it out, but right now, I just have way more
pressing matters, like outstanding regressions and the 8250 split, both
of which are overdue.
Regards,
Peter Hurley
