[PATCH] i2c: aspeed: fix master pending state handling

[Tao Ren]

On 10/10/19 4:16 PM, Jae Hyun Yoo wrote:
> On 10/10/2019 4:11 PM, Tao Ren wrote:
>> On 10/10/19 3:04 PM, Jae Hyun Yoo wrote:
>>> On 10/10/2019 2:20 PM, Tao Ren wrote:
>>>> On 10/9/19 2:20 PM, Jae Hyun Yoo wrote:
>>> [...]
>>>>>            /*
>>>>>             * If a peer master starts a xfer immediately after it queues a
>>>>> -         * master command, change its state to 'pending' then H/W will
>>>>> -         * continue the queued master xfer just after completing the
>>>>> -         * slave mode session.
>>>>> +         * master command, clear the queued master command and change
>>>>> +         * its state to 'pending'. To simplify handling of pending
>>>>> +         * cases, it uses S/W solution instead of H/W command queue
>>>>> +         * handling.
>>>>>             */
>>>>>            if (unlikely(irq_status & ASPEED_I2CD_INTR_SLAVE_MATCH)) {
>>>>> +            writel(readl(bus->base + ASPEED_I2C_CMD_REG) &
>>>>> +                ~ASPEED_I2CD_MASTER_CMDS_MASK,
>>>>> +                   bus->base + ASPEED_I2C_CMD_REG);
>>>>
>>>> Sorry for the late comments (just noticed this line while testing the patch):
>>>>
>>>> I assume this line is aimed at stopping the running master commands, but as per
>>>> AST2500 datasheet, it's NOP to write 0 to MASTER_STOP/MASTER_RX/MASTER_TX bits.
>>>> Maybe all we need is writing 1 to MASTER_STOP field?
>>>
>>> There could be two pending cases:
>>> 1. Master goes to pending before it triggers a command if a slave
>>>     operation is already initiated.
>>> 2. Master goes to pending after it triggered a command if a peer
>>>     master immediately sends something just after the master command
>>>     triggering.
>>>
>>> Above code is for the latter case. H/W handles the case priority based
>>> so the slave event will be handled first, and then the master command
>>> will be handled when the slave operation is completed. Problem is,
>>> this H/W shares the same buffer for master and slave operations so
>>> it's unreliable. Above code just removes the master command from the
>>> command register to prevent this H/W command handling of pending events.
>>> Instead, it restarts the master command using a call of aspeed_i2c_do_start when the slave operation is completed.
>>
>> Thanks for the clarify, Jae. I mean clearing these bits has no effect to
>> hardware according to aspeed datasheet; in other word, master command cannot
>> be removed from command register by this statement.
>>
>> For example, below is the description for MASTER_STOP_CMD(I2CD14, bit 5):
>>
>>    0: NOP
>>    1: Issue Master Stop Command
>>    This register will be automatically cleared by H/W when Stop Command has
>>    been issues.
> 
> It's removing before H/W fetches the the command so the pending command
> isn't cleared by H/W at the timing. If we send a stop command at here, the bus will be messed up.

I see. I didn't know we could clear the bits before hardware fetches them.


Cheers,

Tao