Re: [PATCH v5 2/8] firmware: add support for ARM System Control and Power Interface(SCPI) protocol

[Sudeep Holla <sudeep.holla@arm.com>]

On 31/07/15 14:45, Jassi Brar wrote:
> On Fri, Jul 31, 2015 at 6:38 PM, Sudeep Holla <sudeep.holla@arm.com> wrote:
>>
>>
>> On 31/07/15 11:43, Sudeep Holla wrote:
>>>
>>>
>>>
>>> On 31/07/15 11:38, Jassi Brar wrote:
>>>>
>>>> On Fri, Jul 31, 2015 at 3:10 PM, Sudeep Holla <sudeep.holla@arm.com>
>>>> wrote:
>>>>>
>>>>>
>>>>> I forgot to mention, we have a the following description in
>>>>> mbox_client_txdone which is misleading:
>>>>>
>>>>> "The client/protocol had received some 'ACK' packet and it notifies the
>>>>> API that the last packet was sent successfully. This only works if the
>>>>> controller can't sense TX-Done."
>>>>>
>>>>> which is clearly not the case in SCPI. IMO we may have to reword that.
>>>>>
>>>> Yes. And also see whether it could race with polling driven tx_tick.
>>>>
>>>
>>> Yes I am also looking at that now while I am trying to check if
>>> TXDONE_BY_ACK works on Juno, will keep you posted.
>>>
>>
>> OK, I recollect the racy condition now which I had in my mind from the
>> beginning convincing myself why we can't use that option. I was not good
>> in words to explain it so far but let me try with the ASCII art this
>> time. Note Tx ACK below means the remote setting the register flag and
>> not to be confused with the ACK packet. For simplicity Rx can be assumed
>> to be Tx ACK packet
>>
>>       Time       MHU/SCPI                        Remote SCP
>>                     |                                |
>>          1          |------------ Tx1 -------------->|
>>                     |                                |
>>          2          |<----------- Tx1 ACK -----------|
>>                     |                                |
>>          3          |------------ Tx2 -------------->|
>>                     |                                |
>>          4          |<----------- Rx1 ---------------|
>>                     |                                |
>>          5          |<----------- Tx2 ACK -----------|
>>                     |                                |
>>          6          |------------ Tx3 -------------->|
>>                     |                                |
>>          7          |<----------- Rx2 ---------------|
>>
>> Now lets consider the above scenario, suppose we have TXDONE_BY_ACK
>> and use mbox_client_txdone in Rx interrupt(i.e. response packet), we end
>> up in the race easily IIUC.
>>
>> E.g. A client would have sent Tx2(3) before Rx1 interrupt(4) and if we
>> ACK Tx1 now in (3), we will end up acknowledging Tx2(5) as the mailbox
>> core assumes only one Tx request at a time which is clearly not the case
>> in our setup. The client can then go ahead and send Tx3(6) overwriting
>> the payload while remote was processing or even result in remote missing
>> the request completely. Does that make sense or am I missing something ?
>>
> Yeah, that could happen. But the race can be fixed by checking
> last_tx_done if the controller provides that. If last_tx_done is not
> implemented, polling won't race.
>
> Please try the following...
>

Looks good to me. Sometimes it takes very long time(days) to hit race
conditions(esp. in firmware), so I need some time to think before
I cook up a patch to start stress test this on Juno so that I don't
waste time waiting for result.

Regards,
Sudeep