Re: [RFC v2 6/7] virtio: in order support for virtio_ring

[Guo Zhi]

----- Original Message -----
> From: "jasowang" <jasowang@xxxxxxxxxx>
> To: "Guo Zhi" <qtxuning1999@xxxxxxxxxxx>, "eperezma" <eperezma@xxxxxxxxxx>, "sgarzare" <sgarzare@xxxxxxxxxx>, "Michael
> Tsirkin" <mst@xxxxxxxxxx>
> Cc: "netdev" <netdev@xxxxxxxxxxxxxxx>, "linux-kernel" <linux-kernel@xxxxxxxxxxxxxxx>, "kvm list" <kvm@xxxxxxxxxxxxxxx>,
> "virtualization" <virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx>
> Sent: Thursday, August 25, 2022 3:44:41 PM
> Subject: Re: [RFC v2 6/7] virtio: in order support for virtio_ring

> 在 2022/8/17 21:57, Guo Zhi 写道:
>> If in order feature negotiated, we can skip the used ring to get
>> buffer's desc id sequentially.
>>
>> Signed-off-by: Guo Zhi <qtxuning1999@xxxxxxxxxxx>
>> ---
>>   drivers/virtio/virtio_ring.c | 53 ++++++++++++++++++++++++++++++------
>>   1 file changed, 45 insertions(+), 8 deletions(-)
>>
>> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
>> index 1c1b3fa376a2..143184ebb5a1 100644
>> --- a/drivers/virtio/virtio_ring.c
>> +++ b/drivers/virtio/virtio_ring.c
>> @@ -144,6 +144,9 @@ struct vring_virtqueue {
>>   			/* DMA address and size information */
>>   			dma_addr_t queue_dma_addr;
>>   			size_t queue_size_in_bytes;
>> +
>> +			/* In order feature batch begin here */
> 
> 
> We need tweak the comment, it's not easy for me to understand the
> meaning here.
> 
How about this: if in_order feature is negotiated, is the next head to consume.
> 
>> +			u16 next_desc_begin;
>>   		} split;
>>   
>>   		/* Available for packed ring */
>> @@ -702,8 +705,13 @@ static void detach_buf_split(struct vring_virtqueue *vq,
>> unsigned int head,
>>   	}
>>   
>>   	vring_unmap_one_split(vq, i);
>> -	vq->split.desc_extra[i].next = vq->free_head;
>> -	vq->free_head = head;
>> +	/* In order feature use desc in order,
>> +	 * that means, the next desc will always be free
>> +	 */
> 
> 
> Maybe we should add something like "The descriptors are prepared in order".
> 
I will change it to this: The descriptors are made available in order if the in order feature is used. Since the free_head is already a circular list, it must consume it sequentially.

> 
>> +	if (!virtio_has_feature(vq->vq.vdev, VIRTIO_F_IN_ORDER)) {
>> +		vq->split.desc_extra[i].next = vq->free_head;
>> +		vq->free_head = head;
>> +	}
>>   
>>   	/* Plus final descriptor */
>>   	vq->vq.num_free++;
>> @@ -745,7 +753,7 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue
>> *_vq,
>>   {
>>   	struct vring_virtqueue *vq = to_vvq(_vq);
>>   	void *ret;
>> -	unsigned int i;
>> +	unsigned int i, j;
>>   	u16 last_used;
>>   
>>   	START_USE(vq);
>> @@ -764,11 +772,38 @@ static void *virtqueue_get_buf_ctx_split(struct virtqueue
>> *_vq,
>>   	/* Only get used array entries after they have been exposed by host. */
>>   	virtio_rmb(vq->weak_barriers);
>>   
>> -	last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
>> -	i = virtio32_to_cpu(_vq->vdev,
>> -			vq->split.vring.used->ring[last_used].id);
>> -	*len = virtio32_to_cpu(_vq->vdev,
>> -			vq->split.vring.used->ring[last_used].len);
>> +	if (virtio_has_feature(_vq->vdev, VIRTIO_F_IN_ORDER)) {
>> +		/* Skip used ring and get used desc in order*/
>> +		i = vq->split.next_desc_begin;
>> +		j = i;
>> +		/* Indirect only takes one descriptor in descriptor table */
>> +		while (!vq->indirect && (vq->split.desc_extra[j].flags & VRING_DESC_F_NEXT))
>> +			j = (j + 1) % vq->split.vring.num;
> 
> 
> Let's move the expensive mod outside the loop. Or it's split so we can
> use and here actually since the size is guaranteed to be power of the
> two? Another question, is it better to store the next_desc in e.g
> desc_extra?

Thanks, I will use bit operation instead of mod.

We only use one next_desc_begin at the same time, there is no need to store it in desc_extra.

> 
> And this seems very expensive if the device doesn't do the batching
> (which is not mandatory).
> 
> 
We will judge whether the device batched the buffer or not, we will only use this way for the batched buffer.
And Not in order is more expensive, because is following a linked list.

>> +		/* move to next */
>> +		j = (j + 1) % vq->split.vring.num;
>> +		/* Next buffer will use this descriptor in order */
>> +		vq->split.next_desc_begin = j;
>> +		if (!vq->indirect) {
>> +			*len = vq->split.desc_extra[i].len;
>> +		} else {
>> +			struct vring_desc *indir_desc =
>> +				vq->split.desc_state[i].indir_desc;
>> +			u32 indir_num = vq->split.desc_extra[i].len, buffer_len = 0;
>> +
>> +			if (indir_desc) {
>> +				for (j = 0; j < indir_num / sizeof(struct vring_desc); j++)
>> +					buffer_len += indir_desc[j].len;
> 
> 
> So I think we need to finalize this, then we can have much more stress
> on the cache:
> 
> https://lkml.org/lkml/2021/10/26/1300
> 
> It was reverted since it's too aggressive, we should instead:
> 
> 1) do the validation only for morden device
> 
> 2) fail only when we enable the validation via (e.g a module parameter).
> 
> Thanks
> 
> 
>> +			}
>> +
>> +			*len = buffer_len;
>> +		}
>> +	} else {
>> +		last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
>> +		i = virtio32_to_cpu(_vq->vdev,
>> +				    vq->split.vring.used->ring[last_used].id);
>> +		*len = virtio32_to_cpu(_vq->vdev,
>> +				       vq->split.vring.used->ring[last_used].len);
>> +	}
>>   
>>   	if (unlikely(i >= vq->split.vring.num)) {
>>   		BAD_RING(vq, "id %u out of range\n", i);
>> @@ -2236,6 +2271,8 @@ struct virtqueue *__vring_new_virtqueue(unsigned int
>> index,
>>   	vq->split.avail_flags_shadow = 0;
>>   	vq->split.avail_idx_shadow = 0;
>>   
>> +	vq->split.next_desc_begin = 0;
>> +
>>   	/* No callback?  Tell other side not to bother us. */
>>   	if (!callback) {
>>   		vq->split.avail_flags_shadow |= VRING_AVAIL_F_NO_INTERRUPT;