Re: [PATCH V5 6/7] i2c: tegra: fix PIO rx/tx residual transfer check

[Dmitry Osipenko <digetx@gmail.com>]

14.06.2019 12:50, Bitan Biswas пишет:
> 
> 
> On 6/13/19 5:28 AM, Dmitry Osipenko wrote:
>> 13.06.2019 14:30, Bitan Biswas пишет:
>>>
>>>
>>> On 6/12/19 7:30 AM, Dmitry Osipenko wrote:
>>>> 11.06.2019 13:51, Bitan Biswas пишет:
>>>>> Fix expression for residual bytes(less than word) transfer
>>>>> in I2C PIO mode RX/TX.
>>>>>
>>>>> Signed-off-by: Bitan Biswas <bbiswas@nvidia.com>
>>>>> ---
>>>>
>>>> [snip]
>>>>
>>>>>            /*
>>>>> -         * Update state before writing to FIFO.  If this casues us
>>>>> +         * Update state before writing to FIFO.  If this causes us
>>>>>             * to finish writing all bytes (AKA buf_remaining goes to
>>>>> 0) we
>>>>>             * have a potential for an interrupt (PACKET_XFER_COMPLETE is
>>>>> -         * not maskable).  We need to make sure that the isr sees
>>>>> -         * buf_remaining as 0 and doesn't call us back re-entrantly.
>>>>> +         * not maskable).
>>>>>             */
>>>>>            buf_remaining -= words_to_transfer * BYTES_PER_FIFO_WORD;
>>>>
>>>> Looks like the comment could be removed altogether because it doesn't
>>>> make sense since interrupt handler is under xfer_lock which is kept
>>>> locked during of tegra_i2c_xfer_msg().
>>> I would push a separate patch to remove this comment because of
>>> xfer_lock in ISR now.
>>>
>>>>
>>>> Moreover the comment says that "PACKET_XFER_COMPLETE is not maskable",
>>>> but then what I2C_INT_PACKET_XFER_COMPLETE masking does?
>>>>
>>> I2C_INT_PACKET_XFER_COMPLETE masking support available in Tegra chips
>>> newer than Tegra30 allows one to not see interrupt after Packet transfer
>>> complete. With the xfer_lock in ISR the scenario discussed in comment
>>> can be ignored.
>>
>> Also note that xfer_lock could be removed and replaced with a just
>> irq_enable/disable() calls in tegra_i2c_xfer_msg() because we only care
>> about IRQ not firing during of the preparation process.
> This should need sufficient testing hence let us do it in a different series.

I don't think that there is much to test here since obviously it should work.

>>
>> It also looks like tegra_i2c_[un]nmask_irq isn't really needed and all
>> IRQ's could be simply unmasked during the driver's probe, in that case
>> it may worth to add a kind of "in-progress" flag to catch erroneous
>> interrupts.
>>
> TX interrupt needs special handling if this change is done. Hence I think it should be
> taken up after sufficient testing in a separate patch.

This one is indeed a bit more trickier. Probably another alternative could be to keep GIC
interrupt disabled while no transfer is performed, then you'll have to request interrupt
in a disabled state using IRQ_NOAUTOEN flag.

And yes, that all should be a separate changes if you're going to implement them.