Date:   Wed, 7 Nov 2018 14:17:53 +0800
From:   Jason Wang <jasowang@...hat.com>
To:     jiangyiwen <jiangyiwen@...wei.com>, stefanha@...hat.com
Cc:     netdev@...r.kernel.org, kvm@...r.kernel.org,
        virtualization@...ts.linux-foundation.org
Subject: Re: [PATCH 1/5] VSOCK: support fill mergeable rx buffer in guest


On 2018/11/6 下午2:22, jiangyiwen wrote:
> On 2018/11/6 11:38, Jason Wang wrote:
>> On 2018/11/5 下午3:45, jiangyiwen wrote:
>>> In driver probing, if virtio has VIRTIO_VSOCK_F_MRG_RXBUF feature,
>>> it will fill mergeable rx buffer, support for host send mergeable
>>> rx buffer. It will fill a page everytime to compact with small
>>> packet and big packet.
>>>
>>> Signed-off-by: Yiwen Jiang<jiangyiwen@...wei.com>
>>> ---
>>>    include/linux/virtio_vsock.h     |  3 ++
>>>    net/vmw_vsock/virtio_transport.c | 72 +++++++++++++++++++++++++++++-----------
>>>    2 files changed, 56 insertions(+), 19 deletions(-)
>>>
>>> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
>>> index e223e26..bf84418 100644
>>> --- a/include/linux/virtio_vsock.h
>>> +++ b/include/linux/virtio_vsock.h
>>> @@ -14,6 +14,9 @@
>>>    #define VIRTIO_VSOCK_MAX_BUF_SIZE        0xFFFFFFFFUL
>>>    #define VIRTIO_VSOCK_MAX_PKT_BUF_SIZE        (1024 * 64)
>>>
>>> +/* Virtio-vsock feature */
>>> +#define VIRTIO_VSOCK_F_MRG_RXBUF 0 /* Host can merge receive buffers. */
>>> +
>>>    enum {
>>>        VSOCK_VQ_RX     = 0, /* for host to guest data */
>>>        VSOCK_VQ_TX     = 1, /* for guest to host data */
>>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>>> index 5d3cce9..2040a9e 100644
>>> --- a/net/vmw_vsock/virtio_transport.c
>>> +++ b/net/vmw_vsock/virtio_transport.c
>>> @@ -64,6 +64,7 @@ struct virtio_vsock {
>>>        struct virtio_vsock_event event_list[8];
>>>
>>>        u32 guest_cid;
>>> +    bool mergeable;
>>>    };
>>>
>>>    static struct virtio_vsock *virtio_vsock_get(void)
>>> @@ -256,6 +257,25 @@ static int virtio_transport_send_pkt_loopback(struct virtio_vsock *vsock,
>>>        return 0;
>>>    }
>>>
>>> +static int fill_mergeable_rx_buff(struct virtqueue *vq)
>>> +{
>>> +    void *page = NULL;
>>> +    struct scatterlist sg;
>>> +    int err;
>>> +
>>> +    page = (void *)get_zeroed_page(GFP_KERNEL);
>> Any reason to use zeroed page?
> In previous version, the entire structure of virtio_vsock_pkt is preallocated
> in guest use kzalloc, it is a contiguous zeroed physical memory, but host only
> need to fill virtio_vsock_hdr size.
>
> However, in mergeable rx buffer version, we only fill a page in vring descriptor
> in guest, and I will reserve size of virtio_vsock_pkt in host instead of write
> the total size of virtio_vsock_pkt, for the correctness of structure value,
> we should set zeroed page in advance.


I may miss something, but it looks to me only the header needs to be zeroed.


>
>>> +    if (!page)
>>> +        return -ENOMEM;
>>> +
>>> +    sg_init_one(&sg, page, PAGE_SIZE);
>> FYI, for virtio-net we have several optimizations for mergeable rx buffer:
>>
>> - skb_page_frag_refill() which can use high order page and reduce the stress of page allocator
>>
> You're right, initially I want to use a memory poll to manage the rx buffer,
> and then use this in the later optimized patch. Your advice is very great.
>
>> - we don't use fixed buffer size, instead we use EWMA to estimate the possible rx buffer size to avoid internal fragmentation
>>
> Ok, I analysis the feature and consider add it into my patches.
>
>> If we can try to reuse virtio-net driver, we will get those nice features.
>>
> Yes, after all virtio-net has a very good ecological environment, and it also
> do many performance optimization, it is actually a good idea.
>

Yes, so my suggestion is to consider to reuse them (unless we found 
something that is a real blocker) instead of duplicating codes, features 
a bugs.

Thanks

Powered by blists - more mailing lists