Virtualization - Jul 2021 - [PATCH v11] i2c: virtio: add a virtio i2c frontend driver

On 2021/7/2 12:55, Viresh Kumar wrote:
> On 01-07-21, 21:24, Wolfram Sang wrote:
>>> I just noticed this now, but this function even tries to send data
>>> partially, which isn't right. If the caller (i2c device's
driver)
>>> calls this for 5 struct i2c_msg instances, then all 5 need to get
>>> through or none.. where as we try to send as many as possible here.
>>>
>>> This looks broken to me. Rather return an error value here on
success,
>>> or make it complete failure.
>>>
>>> Though to be fair I see i2c-core also returns number of messages
>>> processed from i2c_transfer().
>>>
>>> Wolfram, what's expected here ? Shouldn't all message
transfer or
>>> none?
>> Well, on a physical bus, it can simply happen that after message 3 of
5,
>> the bus is stalled, so we need to bail out.
> Right, and in that case the transfer will have any meaning left? I believe
it
> needs to be fully retried as the requests may have been dependent on each
other.
>
>> Again, I am missing details of a virtqueue, but I'd think it is
>> different. If adding to the queue fails, then it probably make sense to
>> drop the whole transfer.
> Exactly my point.
>
This is not efficient. If adding the ith request to the queue fails, we 
can still send

the requests before it. We don't need to know if it fails here (adding 
to the queue)

or there (later in the host). The "master_xfer" just need to return 
final number of

messages successfully processed.

On 02-07-21, 14:52, Jie Deng wrote:
> This is not efficient. If adding the ith request to the queue fails, we can
> still send
> 
> the requests before it.
Not really. Normally the requests which are sent together by clients, are linked
together, like a state machine. So if the first one is sent, but not the second
one, then there is not going to be any meaningful result of that.

The i2c core doesn't club requests together from different clients in a
single
i2c_transfer() call. So you must assume i2c_transfer(), irrespective of the
number of underlying messages in it, as atomic. If you fail, the client is going
to retry everything again or assume it failed completely.
> We don't need to know if it fails here (adding to
> the queue)
> 
> or there (later in the host). The "master_xfer" just need to
return final
> number of
> 
> messages successfully processed.
No, that isn't going to help and it is rather inefficient, trying to send
transfer when we already know part of it failed.

-- 
viresh