Re: [RFC PATCH v4 03/17] vsock/virtio: support to send non-linear skb

[Arseniy Krasnov <avkrasnov@sberdevices.ru>]

On 26.06.2023 18:36, Stefano Garzarella wrote:
> On Sat, Jun 03, 2023 at 11:49:25PM +0300, Arseniy Krasnov wrote:
>> For non-linear skb use its pages from fragment array as buffers in
>> virtio tx queue. These pages are already pinned by 'get_user_pages()'
>> during such skb creation.
>>
>> Signed-off-by: Arseniy Krasnov <AVKrasnov@sberdevices.ru>
>> ---
>> net/vmw_vsock/virtio_transport.c | 37 ++++++++++++++++++++++++++------
>> 1 file changed, 31 insertions(+), 6 deletions(-)
>>
>> diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>> index e95df847176b..6053d8341091 100644
>> --- a/net/vmw_vsock/virtio_transport.c
>> +++ b/net/vmw_vsock/virtio_transport.c
>> @@ -100,7 +100,9 @@ virtio_transport_send_pkt_work(struct work_struct *work)
>>     vq = vsock->vqs[VSOCK_VQ_TX];
>>
>>     for (;;) {
>> -        struct scatterlist hdr, buf, *sgs[2];
>> +        /* +1 is for packet header. */
>> +        struct scatterlist *sgs[MAX_SKB_FRAGS + 1];
>> +        struct scatterlist bufs[MAX_SKB_FRAGS + 1];
>>         int ret, in_sg = 0, out_sg = 0;
>>         struct sk_buff *skb;
>>         bool reply;
>> @@ -111,12 +113,35 @@ virtio_transport_send_pkt_work(struct work_struct *work)
>>
>>         virtio_transport_deliver_tap_pkt(skb);
>>         reply = virtio_vsock_skb_reply(skb);
>> +        sg_init_one(&bufs[0], virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb)));
>> +        sgs[out_sg++] = &bufs[0];
> 
> Can we use out_sg also to index bufs (here and in the rest of the code)?
> 
> E.g.
> 
>         sg_init_one(&bufs[out_sg], ...)
>         sgs[out_sg] = &bufs[out_sg];
>         ++out_sg;
> 
>         ...
>             if (skb->len > 0) {
>                 sg_init_one(&bufs[out_sg], skb->data, skb->len);
>                 sgs[out_sg] = &bufs[out_sg];
>                 ++out_sg;
>             }
> 
>         etc...
> 
>> +
> 
> For readability, I would move the smaller branch above:
> 
>         if (!skb_is_nonlinear(skb)) {
>             // small block
>             ...
>         } else {
>             // big block
>             ...
>         }
> 
>> +        if (skb_is_nonlinear(skb)) {
>> +            struct skb_shared_info *si;
>> +            int i;
>> +
>> +            si = skb_shinfo(skb);
>> +
>> +            for (i = 0; i < si->nr_frags; i++) {
>> +                skb_frag_t *skb_frag = &si->frags[i];
>> +                void *va = page_to_virt(skb_frag->bv_page);
>> +
>> +                /* We will use 'page_to_virt()' for userspace page here,
>> +                 * because virtio layer will call 'virt_to_phys()' later
>> +                 * to fill buffer descriptor. We don't touch memory at
>> +                 * "virtual" address of this page.
>> +                 */
>> +                sg_init_one(&bufs[i + 1],
>> +                        va + skb_frag->bv_offset,
>> +                        skb_frag->bv_len);
>> +                sgs[out_sg++] = &bufs[i + 1];
>> +            }
>> +        } else {
>> +            if (skb->len > 0) {
> 
> Should we do the same check (skb->len > 0) for nonlinear skb as well?
> Or do the nonlinear ones necessarily have len > 0?

Yes, non-linear skb always has 'data_len' > 0, e.g. such skbs always have some
data in it.

Thanks, Arseniy

> 
>> +                sg_init_one(&bufs[1], skb->data, skb->len);
>> +                sgs[out_sg++] = &bufs[1];
>> +            }
>>
>    ^
> Blank line that we can remove.
> 
> Stefano
> 
>> -        sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb)));
>> -        sgs[out_sg++] = &hdr;
>> -        if (skb->len > 0) {
>> -            sg_init_one(&buf, skb->data, skb->len);
>> -            sgs[out_sg++] = &buf;
>>         }
>>
>>         ret = virtqueue_add_sgs(vq, sgs, out_sg, in_sg, skb, GFP_KERNEL);
>> -- 
>> 2.25.1
>>
>