Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 13 May 2019 19:23:22 +0200
From:   Stefano Garzarella <sgarzare@redhat.com>
To:     Jason Wang <jasowang@redhat.com>
Cc:     netdev@vger.kernel.org, "David S. Miller" <davem@davemloft.net>,
        "Michael S. Tsirkin" <mst@redhat.com>,
        virtualization@lists.linux-foundation.org,
        linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        Stefan Hajnoczi <stefanha@redhat.com>
Subject: Re: [PATCH v2 1/8] vsock/virtio: limit the memory used per-socket
Message-ID: <20190513172322.vcgenx7xk4v6r2ay@steredhat>
References: <20190510125843.95587-1-sgarzare@redhat.com>
 <20190510125843.95587-2-sgarzare@redhat.com>
 <3b275b52-63d9-d260-1652-8e8bf7dd679f@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <3b275b52-63d9-d260-1652-8e8bf7dd679f@redhat.com>
User-Agent: NeoMutt/20180716
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, May 13, 2019 at 05:58:53PM +0800, Jason Wang wrote:
> 
> On 2019/5/10 下午8:58, Stefano Garzarella wrote:
> > Since virtio-vsock was introduced, the buffers filled by the host
> > and pushed to the guest using the vring, are directly queued in
> > a per-socket list avoiding to copy it.
> > These buffers are preallocated by the guest with a fixed
> > size (4 KB).
> > 
> > The maximum amount of memory used by each socket should be
> > controlled by the credit mechanism.
> > The default credit available per-socket is 256 KB, but if we use
> > only 1 byte per packet, the guest can queue up to 262144 of 4 KB
> > buffers, using up to 1 GB of memory per-socket. In addition, the
> > guest will continue to fill the vring with new 4 KB free buffers
> > to avoid starvation of other sockets.
> > 
> > This patch solves this issue copying the payload in a new buffer.
> > Then it is queued in the per-socket list, and the 4KB buffer used
> > by the host is freed.
> > 
> > In this way, the memory used by each socket respects the credit
> > available, and we still avoid starvation, paying the cost of an
> > extra memory copy. When the buffer is completely full we do a
> > "zero-copy", moving the buffer directly in the per-socket list.
> 
> 
> I wonder in the long run we should use generic socket accouting mechanism
> provided by kernel (e.g socket, skb, sndbuf, recvbug, truesize) instead of
> vsock specific thing to avoid duplicating efforts.

I agree, the idea is to switch to sk_buff but this should require an huge
change. If we will use the virtio-net datapath, it will become simpler.

> 
> 
> > 
> > Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> > ---
> >   drivers/vhost/vsock.c                   |  2 +
> >   include/linux/virtio_vsock.h            |  8 +++
> >   net/vmw_vsock/virtio_transport.c        |  1 +
> >   net/vmw_vsock/virtio_transport_common.c | 95 ++++++++++++++++++-------
> >   4 files changed, 81 insertions(+), 25 deletions(-)
> > 
> > diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
> > index bb5fc0e9fbc2..7964e2daee09 100644
> > --- a/drivers/vhost/vsock.c
> > +++ b/drivers/vhost/vsock.c
> > @@ -320,6 +320,8 @@ vhost_vsock_alloc_pkt(struct vhost_virtqueue *vq,
> >   		return NULL;
> >   	}
> > +	pkt->buf_len = pkt->len;
> > +
> >   	nbytes = copy_from_iter(pkt->buf, pkt->len, &iov_iter);
> >   	if (nbytes != pkt->len) {
> >   		vq_err(vq, "Expected %u byte payload, got %zu bytes\n",
> > diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> > index e223e2632edd..345f04ee9193 100644
> > --- a/include/linux/virtio_vsock.h
> > +++ b/include/linux/virtio_vsock.h
> > @@ -54,9 +54,17 @@ struct virtio_vsock_pkt {
> >   	void *buf;
> >   	u32 len;
> >   	u32 off;
> > +	u32 buf_len;
> >   	bool reply;
> >   };
> > +struct virtio_vsock_buf {
> > +	struct list_head list;
> > +	void *addr;
> > +	u32 len;
> > +	u32 off;
> > +};
> > +
> >   struct virtio_vsock_pkt_info {
> >   	u32 remote_cid, remote_port;
> >   	struct vsock_sock *vsk;
> > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
> > index 15eb5d3d4750..af1d2ce12f54 100644
> > --- a/net/vmw_vsock/virtio_transport.c
> > +++ b/net/vmw_vsock/virtio_transport.c
> > @@ -280,6 +280,7 @@ static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
> >   			break;
> >   		}
> > +		pkt->buf_len = buf_len;
> >   		pkt->len = buf_len;
> >   		sg_init_one(&hdr, &pkt->hdr, sizeof(pkt->hdr));
> > diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> > index 602715fc9a75..0248d6808755 100644
> > --- a/net/vmw_vsock/virtio_transport_common.c
> > +++ b/net/vmw_vsock/virtio_transport_common.c
> > @@ -65,6 +65,9 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info *info,
> >   		pkt->buf = kmalloc(len, GFP_KERNEL);
> >   		if (!pkt->buf)
> >   			goto out_pkt;
> > +
> > +		pkt->buf_len = len;
> > +
> >   		err = memcpy_from_msg(pkt->buf, info->msg, len);
> >   		if (err)
> >   			goto out;
> > @@ -86,6 +89,46 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info *info,
> >   	return NULL;
> >   }
> > +static struct virtio_vsock_buf *
> > +virtio_transport_alloc_buf(struct virtio_vsock_pkt *pkt, bool zero_copy)
> > +{
> > +	struct virtio_vsock_buf *buf;
> > +
> > +	if (pkt->len == 0)
> > +		return NULL;
> > +
> > +	buf = kzalloc(sizeof(*buf), GFP_KERNEL);
> > +	if (!buf)
> > +		return NULL;
> > +
> > +	/* If the buffer in the virtio_vsock_pkt is full, we can move it to
> > +	 * the new virtio_vsock_buf avoiding the copy, because we are sure that
> > +	 * we are not use more memory than that counted by the credit mechanism.
> > +	 */
> > +	if (zero_copy && pkt->len == pkt->buf_len) {
> > +		buf->addr = pkt->buf;
> > +		pkt->buf = NULL;
> > +	} else {
> 
> 
> Is the copy still needed if we're just few bytes less? We meet similar issue
> for virito-net, and virtio-net solve this by always copy first 128bytes for
> big packets.
> 
> See receive_big()

I'm seeing, It is more sophisticated.
IIUC, virtio-net allocates a sk_buff with 128 bytes of buffer, then copies the
first 128 bytes, then adds the buffer used to receive the packet as a frag to
the skb.

Do you suggest to implement something similar, or for now we can use my
approach and if we will merge the datapath we can reuse the virtio-net
approach?

Thanks,
Stefano