[U-Boot] [PATCH v2 2/2] xilinx_xiic: Fix transfer initialisation

[Melin Tomas]

Hi,

+ Richard, Ben


On 6/27/19 3:05 PM, Heiko Schocher wrote:
> Hi Marek, Tomas,
>
> Am 27.06.2019 um 13:32 schrieb Marek Vasut:
>> On 6/27/19 7:53 AM, Melin Tomas wrote:
>>>
>>> On 6/27/19 4:41 AM, Marek Vasut wrote:
>>>> On 6/26/19 8:17 PM, Melin Tomas wrote:
>>>>> On 6/26/19 4:36 PM, Marek Vasut wrote:
>>>>>> On 6/26/19 3:19 PM, Melin Tomas wrote:
>>>>>>> On 6/26/19 3:48 PM, Marek Vasut wrote:
>>>>>>>> On 6/26/19 2:45 PM, Melin Tomas wrote:
>>>>>>>>> On 6/26/19 3:26 PM, Marek Vasut wrote:
>>>>>>>>>> On 6/26/19 2:19 PM, Melin Tomas wrote:
>>>>>>>>>>> On 6/26/19 2:49 PM, Marek Vasut wrote:
>>>>>>>>>>>> On 6/26/19 1:25 PM, Melin Tomas wrote:
>>>>>>>>>>>>> On 6/26/19 1:47 PM, Marek Vasut wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> On 6/26/19 12:39 PM, Melin Tomas wrote:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> As such, it's probably a good idea to keep the same 
>>>>>>>>>>>>>>> delay values here as
>>>>>>>>>>>>>>> in the original driver unless good reason to use 
>>>>>>>>>>>>>>> something else.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> As what goes for the original reasoning for 3ms, the 
>>>>>>>>>>>>>>> commit history does
>>>>>>>>>>>>>>> not mention that so I cannot comment.
>>>>>>>>>>>>>> So would you be so kind and research this ?
>>>>>>>>>>>>> Based on a short study of other i2c bus drivers it seems 
>>>>>>>>>>>>> most have bus
>>>>>>>>>>>>> busy timeout checks.
>>>>>>>>>>>>>
>>>>>>>>>>>>> The timeout values seems to differ, ranging from 
>>>>>>>>>>>>> milliseconds to seconds.
>>>>>>>>>>>> Yep
>>>>>>>>>>>>
>>>>>>>>>>>>> So probably it's just a number, after all it's just a 
>>>>>>>>>>>>> check to know if
>>>>>>>>>>>>> we are good to go.
>>>>>>>>>>>> And is the number large enough ?
>>>>>>>>>>> As mentioned, good approach is probably using value known to 
>>>>>>>>>>> work
>>>>>>>>>>> instead of
>>>>>>>>>>>
>>>>>>>>>>> guessing a new number.
>>>>>>>>>> So why did kernel pick that specific number ? Surely there 
>>>>>>>>>> was some
>>>>>>>>>> reasoning, they didn't just pull it out of /dev/random .
>>>>>>>>> Yes, history does not tell.
>>>>>>>>>
>>>>>>>>> I do see that this driver uses timeout of 1000ms for bus busy 
>>>>>>>>> when
>>>>>>>>> probing, perhaps you can enlighten how that number was 
>>>>>>>>> concluded? If
>>>>>>>>> that could give some clues about this.
>>>>>>>> I don't know.
>>>>>>> But you are author of that line?
>>>>>> +    ret = wait_for_bit_8(priv->base + XIIC_SR_REG_OFFSET,
>>>>>> +                 XIIC_SR_BUS_BUSY_MASK, false, 3, true);
>>>>>> +
>>>>>>
>>>>>> comes from 2/2 ?
>>>>> No, I was referring to the already existing wait_for_bit_8 of bus 
>>>>> busy
>>>>>
>>>>> https://gitlab.denx.de/u-boot/u-boot/blob/master/drivers/i2c/xilinx_xiic.c#L294 
>>>>>
>>>> Oh, this. When probing, you need to pick some arbitrary amount of time
>>>> after which you give up and conclude there isn't any device at that
>>>> address. 1 second should give enough time to even slow devices on slow
>>>> busses. If some device is too slow or doesn't respond, too bad. That's
>>>> also why i2c is NOT a bus which could be probed, there are simple
>>>> devices which do not respond to addressing in any way.
>>>>
>>>> So this probably didn't help you determine why you should wait some 
>>>> time
>>>> for bus busy when sending messages, since this is unrelated timeout.
>>>>
>>>>>>>> You're the patch author, it's your responsibility to know why 
>>>>>>>> you're
>>>>>>>> adding/changing the code you're adding/changing.
>>>>>>> yes, and the reasoning is:
>>>>>>>
>>>>>>> * the value has been deemed good in original driver. If it would 
>>>>>>> be bad,
>>>>>>> probably it would have been changed during course of time
>>>>>>>
>>>>>>> * the value has been tested for this driver as well with success
>>>>>> So shouldn't there be some upper bound on the bus busy time , 
>>>>>> demanded
>>>>>> either by the i2c bus spec or the xiic core spec ?
>>>>> In case some of the devices on the bus misbehaves, bus could 
>>>>> potentially
>>>>> stay
>>>>>
>>>>> busy until device reset or power-cycle.
>>>> My concern here would be e.g. EEPROM programming, which is slow, so I
>>>> wonder whether this timeout is enough. But maybe the xiic datasheet 
>>>> says
>>>> something about maximum delay , or somehow tells you how to derive the
>>>> delay from bus clock speed ...
>>>
>>> Specification only mentions to check bus busy status prior to starting,
>>>
>>> no mention about delays AFAIS.
>>>
>>> Even with standard low speed 100kHz, 3ms would equal 300 clock
>>>
>>> cycles waiting.
>>>
>>> Unless you have other suggestion, I suggest proceeding with this delay
>>> value,
>>>
>>> as explained why in previous messages.
>>
>> There was no explanation besides "Linux does it this way, so it must be
>> right", and I don't really like that. But ultimately, this is up to
>> Heiko now.
>
> May it makes sense to ask Richard or Ben which have the driver
> and the comment with "3 mS" introduced in linux?
>
> See linux commit:
> commit e1d5b6598cdc33257fe68302ae9db81d2f7bb883
> Author: Richard Röjfors <richard.rojfors at pelagicore.com>
> Date:   Thu Feb 11 10:42:00 2010 +0100
>
>     i2c: Add support for Xilinx XPS IIC Bus Interface
>
>     This patch adds support for the Xilinx XPS IIC Bus Interface.
>
>     The driver uses the dynamic mode, supporting to put several
>     I2C messages in the FIFO to reduce the number of interrupts.
>
>     It has the same feature as ocores, it can be passed a list
>     of devices that will be added when the bus is probed.
>
>     Signed-off-by: Richard Röjfors <richard.rojfors at pelagicore.com>
>     Signed-off-by: Ben Dooks <ben-linux at fluff.org>
>
>
> Beside of that, wait_for_bit_8() drops a debug message, if
> timeout, so that is good ...
>
> Remains the question, what is a good timeout value ?
>
> I tend to say, we could set the timeout higher, as the SR
> register is polled all millisecond, and if bus is not busy
> we loose no time.
ok, so increase timeout to 1000ms or so?
>
> But the problem remains, what a good timeout is here ...
>
> So may we should add here a printf, if wait_for_bit_8 returns with
> -ETIMEDOUT, so users may get in panic and report, and we can find a
> better timeout?

Sounds ok to me.

So I'll update the patch to

* have longer 1000ms timeout

* print message in case of timeout from bus busy to catch possibly 
incorrect timeout estimate

Please let me know if this approach does not sound good to any of you.


thanks,

Tomas