Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sat, 15 Oct 2022 12:49:59 -0700
From:   Cong Wang <xiyou.wangcong@gmail.com>
To:     Bobby Eshleman <bobbyeshleman@gmail.com>
Cc:     Stefano Garzarella <sgarzare@redhat.com>,
        "Michael S. Tsirkin" <mst@redhat.com>,
        Bobby Eshleman <bobby.eshleman@gmail.com>,
        Bobby Eshleman <bobby.eshleman@bytedance.com>,
        Cong Wang <cong.wang@bytedance.com>,
        Jiang Wang <jiang.wang@bytedance.com>,
        Stefan Hajnoczi <stefanha@redhat.com>,
        Jason Wang <jasowang@redhat.com>,
        "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>, kvm@vger.kernel.org,
        virtualization@lists.linux-foundation.org, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] vsock: replace virtio_vsock_pkt with sk_buff
Message-ID: <Y0sO5yNqQkFQucjb@pop-os.localdomain>
References: <20221006011946.85130-1-bobby.eshleman@bytedance.com>
 <20221006025956-mutt-send-email-mst@kernel.org>
 <20221006073410.ahhqhlhah4lo47o7@sgarzare-redhat>
 <Yzoou4UwOv5lh0hE@bullseye>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Yzoou4UwOv5lh0hE@bullseye>
Precedence: bulk

On Mon, Oct 03, 2022 at 12:11:39AM +0000, Bobby Eshleman wrote:
> On Thu, Oct 06, 2022 at 09:34:10AM +0200, Stefano Garzarella wrote:
> > On Thu, Oct 06, 2022 at 03:08:12AM -0400, Michael S. Tsirkin wrote:
> > > On Wed, Oct 05, 2022 at 06:19:44PM -0700, Bobby Eshleman wrote:
> > > > This patch replaces the struct virtio_vsock_pkt with struct sk_buff.
> > > > 
> > > > Using sk_buff in vsock benefits it by a) allowing vsock to be extended
> > > > for socket-related features like sockmap, b) vsock may in the future
> > > > use other sk_buff-dependent kernel capabilities, and c) vsock shares
> > > > commonality with other socket types.
> > > > 
> > > > This patch is taken from the original series found here:
> > > > https://lore.kernel.org/all/cover.1660362668.git.bobby.eshleman@bytedance.com/
> > > > 
> > > > Small-sized packet throughput improved by ~5% (from 18.53 Mb/s to 19.51
> > > > Mb/s). Tested using uperf, 16B payloads, 64 threads, 100s, averaged from
> > > > 10 test runs (n=10). This improvement is likely due to packet merging.
> > > > 
> > > > Large-sized packet throughput decreases ~9% (from 27.25 Gb/s to 25.04
> > > > Gb/s). Tested using uperf, 64KB payloads, 64 threads, 100s, averaged
> > > > from 10 test runs (n=10).
> > > > 
> > > > Medium-sized packet throughput decreases ~5% (from 4.0 Gb/s to 3.81
> > > > Gb/s). Tested using uperf, 4k to 8k payload sizes picked randomly
> > > > according to normal distribution, 64 threads, 100s, averaged from 10
> > > > test runs (n=10).
> > > 
> > > It is surprizing to me that the original vsock code managed to outperform
> > > the new one, given that to my knowledge we did not focus on optimizing it.
> > 
> > Yeah mee to.
> > 
> 
> Indeed.
> 
> > From this numbers maybe the allocation cost has been reduced as it performs
> > better with small packets. But with medium to large packets we perform
> > worse, perhaps because previously we were allocating a contiguous buffer up
> > to 64k?
> > Instead alloc_skb() could allocate non-contiguous pages ? (which would solve
> > the problems we saw a few days ago)
> > 
> 
> I think this would be the case with alloc_skb_with_frags(), but
> internally alloc_skb() uses kmalloc() for the payload and sk_buff_head
> slab allocations for the sk_buff itself (all the more confusing to me,
> as the prior allocator also uses two separate allocations per packet).

I think it is related to your implementation of
virtio_transport_add_to_queue(), where you introduced much more
complicated logic than before:

-	spin_lock_bh(&vsock->send_pkt_list_lock);
-	list_add_tail(&pkt->list, &vsock->send_pkt_list);
-	spin_unlock_bh(&vsock->send_pkt_list_lock);
-
+	virtio_transport_add_to_queue(&vsock->send_pkt_queue, skb);

A simple list_add_tail() is definitely faster than your
virtio_transport_skbs_can_merge() check. So, why do you have to merge
skb while we don't merge virtio_vsock_pkt?

_If_ you are trying to mimic TCP, I think you are doing it wrong, it can
be much more efficient if you could do the merge in sendmsg() before skb
is even allocated, see tcp_sendmsg_locked().

Thanks.