Re: [RFC v4 3/5] virtio_ring: add packed ring support

[Jason Wang <jasowang@redhat.com>]

On 2018年05月18日 22:33, Tiwei Bie wrote:
> On Fri, May 18, 2018 at 09:17:05PM +0800, Jason Wang wrote:
>> On 2018年05月18日 19:29, Tiwei Bie wrote:
>>> On Thu, May 17, 2018 at 08:01:52PM +0800, Jason Wang wrote:
>>>> On 2018年05月16日 22:33, Tiwei Bie wrote:
>>>>> On Wed, May 16, 2018 at 10:05:44PM +0800, Jason Wang wrote:
>>>>>> On 2018年05月16日 21:45, Tiwei Bie wrote:
>>>>>>> On Wed, May 16, 2018 at 08:51:43PM +0800, Jason Wang wrote:
>>>>>>>> On 2018年05月16日 20:39, Tiwei Bie wrote:
>>>>>>>>> On Wed, May 16, 2018 at 07:50:16PM +0800, Jason Wang wrote:
>>>>>>>>>> On 2018年05月16日 16:37, Tiwei Bie wrote:
>>>>> [...]
>>>>>>>>>>> +static void detach_buf_packed(struct vring_virtqueue *vq, unsigned int head,
>>>>>>>>>>> +			      unsigned int id, void **ctx)
>>>>>>>>>>> +{
>>>>>>>>>>> +	struct vring_packed_desc *desc;
>>>>>>>>>>> +	unsigned int i, j;
>>>>>>>>>>> +
>>>>>>>>>>> +	/* Clear data ptr. */
>>>>>>>>>>> +	vq->desc_state[id].data = NULL;
>>>>>>>>>>> +
>>>>>>>>>>> +	i = head;
>>>>>>>>>>> +
>>>>>>>>>>> +	for (j = 0; j < vq->desc_state[id].num; j++) {
>>>>>>>>>>> +		desc = &vq->vring_packed.desc[i];
>>>>>>>>>>> +		vring_unmap_one_packed(vq, desc);
>>>>>>>>>> As mentioned in previous discussion, this probably won't work for the case
>>>>>>>>>> of out of order completion since it depends on the information in the
>>>>>>>>>> descriptor ring. We probably need to extend ctx to record such information.
>>>>>>>>> Above code doesn't depend on the information in the descriptor
>>>>>>>>> ring. The vq->desc_state[] is the extended ctx.
>>>>>>>>>
>>>>>>>>> Best regards,
>>>>>>>>> Tiwei Bie
>>>>>>>> Yes, but desc is a pointer to descriptor ring I think so
>>>>>>>> vring_unmap_one_packed() still depends on the content of descriptor ring?
>>>>>>>>
>>>>>>> I got your point now. I think it makes sense to reserve
>>>>>>> the bits of the addr field. Driver shouldn't try to get
>>>>>>> addrs from the descriptors when cleanup the descriptors
>>>>>>> no matter whether we support out-of-order or not.
>>>>>> Maybe I was wrong, but I remember spec mentioned something like this.
>>>>> You're right. Spec mentioned this. I was just repeating
>>>>> the spec to emphasize that it does make sense. :)
>>>>>
>>>>>>> But combining it with the out-of-order support, it will
>>>>>>> mean that the driver still needs to maintain a desc/ctx
>>>>>>> list that is very similar to the desc ring in the split
>>>>>>> ring. I'm not quite sure whether it's something we want.
>>>>>>> If it is true, I'll do it. So do you think we also want
>>>>>>> to maintain such a desc/ctx list for packed ring?
>>>>>> To make it work for OOO backends I think we need something like this
>>>>>> (hardware NIC drivers are usually have something like this).
>>>>> Which hardware NIC drivers have this?
>>>> It's quite common I think, e.g driver track e.g dma addr and page frag
>>>> somewhere. e.g the ring->rx_info in mlx4 driver.
>>> It seems that I had a misunderstanding on your
>>> previous comments. I know it's quite common for
>>> drivers to track e.g. DMA addrs somewhere (and
>>> I think one reason behind this is that they want
>>> to reuse the bits of addr field).
>> Yes, we may want this for virtio-net as well in the future.
>>
>>>    But tracking
>>> addrs somewhere doesn't means supporting OOO.
>>> I thought you were saying it's quite common for
>>> hardware NIC drivers to support OOO (i.e. NICs
>>> will return the descriptors OOO):
>>>
>>> I'm not familiar with mlx4, maybe I'm wrong.
>>> I just had a quick glance. And I found below
>>> comments in mlx4_en_process_rx_cq():
>>>
>>> ```
>>> /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
>>>    * descriptor offset can be deduced from the CQE index instead of
>>>    * reading 'cqe->index' */
>>> index = cq->mcq.cons_index & ring->size_mask;
>>> cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
>>> ```
>>>
>>> It seems that although they have a completion
>>> queue, they are still using the ring in order.
>> I guess so (at least from the above bits). Git grep -i "out of order" in
>> drivers/net gives some hints. Looks like there're few deivces do this.
>>
>>> I guess maybe storage device may want OOO.
>> Right, some iSCSI did.
>>
>> But tracking them elsewhere is not only for OOO.
>>
>> Spec said:
>>
>> for element address
>>
>> "
>> In a used descriptor, Element Address is unused.
>> "
>>
>> for Next flag:
>>
>> "
>> For example, if descriptors are used in the same order in which they are
>> made available, this will result in
>> the used descriptor overwriting the first available descriptor in the list,
>> the used descriptor for the next list
>> overwriting the first available descriptor in the next list, etc.
>> "
>>
>> for in order completion:
>>
>> "
>> This will result in the used descriptor overwriting the first available
>> descriptor in the batch, the used descriptor
>> for the next batch overwriting the first available descriptor in the next
>> batch, etc.
>> "
>>
>> So:
>>
>> - It's an alignment to the spec
>> - device may (or should) overwrite the descriptor make also make address
>> field useless.
> You didn't get my point...

I don't hope so.

> I agreed driver should track the DMA addrs or some
> other necessary things from the very beginning. And
> I also repeated the spec to emphasize that it does
> make sense. And I'd like to do that.
>
> What I was saying is that, to support OOO, we may
> need to manage these context (which saves DMA addrs
> etc) via a list which is similar to the desc list
> maintained via `next` in split ring instead of an
> array whose elements always can be indexed directly.

My point is these context is a must (not only for OOO).

>
> The desc ring in split ring is an array, but its
> free entries are managed as list via next. I was
> just wondering, do we want to manage such a list
> because of OOO. It's just a very simple question
> that I want to hear your opinion... (It doesn't
> means anything, e.g. It doesn't mean I don't want
> to support OOO. It's just a simple question...)

So the question is yes. But I admit I don't have better idea other than 
what you propose here (something like split ring which is a little bit 
sad). Maybe Michael had.

Thanks

>
> Best regards,
> Tiwei Bie
>
>> Thanks
>>
>>> Best regards,
>>> Tiwei Bie
>>>
>>>> Thanks
>>>>
>>>>>> Not for the patch, but it looks like having a OUT_OF_ORDER feature bit is
>>>>>> much more simpler to be started with.
>>>>> +1
>>>>>
>>>>> Best regards,
>>>>> Tiwei Bie