Re: [PATCH] spi: Fix error code checking in spi_mem_exec_op()

[Pratyush Yadav <pratyush@kernel.org>]

On Wed, Mar 13 2024, Florian Fainelli wrote:

> On 3/13/24 12:34, Florian Fainelli wrote:
>> On 3/13/24 12:29, Pratyush Yadav wrote:
>>> On Wed, Mar 13 2024, Florian Fainelli wrote:
>>>
>>>> On 3/13/24 11:28, Pratyush Yadav wrote:
>>>>> On Wed, Mar 13 2024, Michael Walle wrote:
>>>>>
>>>>>> On Wed Mar 13, 2024 at 6:10 PM CET, Florian Fainelli wrote:
>>>>>>> After commit cff49d58f57e ("spi: Unify error codes by replacing -ENOTSUPP
>>>>>>> with
>>>>>>> -EOPNOTSUPP"), our SPI NOR flashes would stop probing with the following
>>>>>>> visible in the kernel log:
>>>>>>>
>>>>>>> [    2.196300] brcmstb_qspi f0440920.qspi: using bspi-mspi mode
>>>>>>> [    2.210295] spi-nor: probe of spi1.0 failed with error -95
>>>>>>>
>>>>>>> It turns out that the check in spi_mem_exec_op() was changed to check
>>>>>>> for -ENOTSUPP (old error code) or -EOPNOTSUPP (new error code), but this
>>>>>>> means that for drivers that were converted, the second condition is now
>>>>>>> true, and we stop falling through like we used to. Fix the error to
>>>>>>> check for neither error being neither -ENOTSUPP *nor* -EOPNOTSUPP.
>>>>>>>
>>>>>>> Fixes: cff49d58f57e ("spi: Unify error codes by replacing -ENOTSUPP with
>>>>>>> -EOPNOTSUPP")
>>>>>>> Signed-off-by: Florian Fainelli <florian.fainelli@broadcom.com>
>>>>>>> Change-Id: I4159811f6c582c4de2143382473d2000b8755872
>>>>>>
>>>>>> Ha, thank you!
>>>>>>
>>>>>> Reviewed-by: Michael Walle <mwalle@kernel.org>
>>>>>>
>>>>>> FWIW in next, there is commit
>>>>>> e63aef9c9121e ("spi: spi-mem: add statistics support to ->exec_op()
>>>>>> calls")
>>>>>> that probably will conflict with this one.
>>>>>>
>>>>>> Also, - not for this patch - but with that logic, spi_mem_exec_op()
>>>>>> might return EOPNOTSUPP *or* ENOTSUPP, even for drivers which might
>>>>>> still return ENOTSUPP, because there is one condition in
>>>>>> spi_mem_exec_op() which will always return EOPNOTSUPP. That is
>>>>>> somewhat confusing, no?
>>>>> I agree. I suppose it would be better to do:
>>>>>       if (!ret)
>>>>>          return 0;
>>>>>       if (ret == -ENOTSUPP || ret == -EOPNOTSUPP)
>>>>>          return -EOPNOTSUPP;
>>>>>
>>>>
>>>> But with e63aef9c9121e ("spi: spi-mem: add statistics support to ->exec_op()
>>>> calls") applied, would not that mean duplicating the statistics gathering,
>>>> or
>>>> were the statistics gathering only intended for when ret == 0?
>>>
>>> Hmm, I didn't properly understand this. Ignore my suggestion. Your patch
>>> does the right thing.
>> What I meant is that e63aef9c9121e will increment statistics not just when we
>> return 0 from ctlr->mem_ops->exec_op, but also if we return -ENOTSUPP or
>> -EOPNOTSUPP, and I am  not sure if this is exactly what is intended. But this
>> is somewhat orthogonal.

No it won't. This is what confused me in my earlier reply as well. If
ret is either of -ENOTSUPP or -EOPNOTSUPP, the expression

    (ret != -ENOTSUPP && ret != -EOPNOTSUPP)

becomes false (along with !ret also being false). In that case, it will
_not_ go in the if statement, and not call spi_mem_add_op_stats().
Instead, it will go via the normal SPI path and that path would do the
accounting based on error or success.

>
> It looks like the handling of a non-zero return code will fall either in the
> -ETIMEDOUT category, or in the general category of an error. I suppose there is
> a question whether a operation that could not be supported should fall in the
> "error" category.

The only questionable thing I see in spi_mem_add_op_stats() is that it
increments bytes_{rx,tx} even in case of failure. It mimics what
spi_statistics_add_transfer_stats() does but perhaps that also is wrong.

-- 
Regards,
Pratyush Yadav