[Skiboot] [RFC PATCH 2/2] Add SBE driver support

[Oliver O'Halloran]

On Wed, Feb 22, 2017 at 9:21 PM, Vasant Hegde
<hegdevasant at linux.vnet.ibm.com> wrote:
> On 02/16/2017 11:56 AM, Oliver O'Halloran wrote:
>>
>> On Wed, 2017-02-15 at 20:44 +0530, Vasant Hegde wrote:
>>>
>>> SBE (Self Boot Engine) on P9 has two different jobs:
>>>    - Boot the chip-up till core is functional
>>>    - Provide various services like timer, scom, etc., at runtime
>
>
> .../...
>
>>> +static inline struct sbe_msg *__sbe_allocmsg(void)
>>> +{
>>> +       return zalloc(sizeof(struct sbe_msg));
>>> +}
>>> +
>>> +struct sbe_msg *sbe_allocmsg(bool alloc_response)
>>> +{
>>> +       struct sbe_msg *msg;
>>> +
>>> +       msg = __sbe_allocmsg();
>>> +       if (!msg)
>>> +               return NULL;
>>> +
>>> +       if (alloc_response) {
>>> +               msg->resp = __sbe_allocmsg();
>>> +               if (!msg->resp) {
>>> +                       free(msg);
>>> +                       return NULL;
>>> +               }
>>
>> Given that we're limited to one outstanding response per SBE we could
>> have a sbe_msg inside the sbe structure itself and skip this extra
>> allocation. Up to you.
>
>
> I think I should have explained this in commit message.  Actually caller can
> queue any number of messages for given SBE. And driver will serialize
> sending message. Hence I've added two structures.

I know you can have any number of messages queued up, but like you
said sending them is serialised. As far as I can tell the receive
logic looks something like this:

msg = msg_at_top_of_queue();
remove_from_queue(msg);
if (msg->complete)
     msg->complete(msg);
sbe_free(msg);
sbe_queue_poke();

So we completely handle the message at the top of the queue before
processing the next one. The space that was allocated for the response
to the first message is free()ed immediately after calling the
completion function. What I was getting at is that because responses
are also serialised there is no need to allocate a separate response
structure for each message. Instead we could have a single response
structure (for each SBE) that is re-used when processing each message
in the queue. I think, I might have missed something ;)

Oliver