similar to: [RFC PATCH net-next 00/12] XDP batching for TUN/vhost_net

Hi all: This series tries to batch submitting packets to underlayer socket through msg_control during sendmsg(). This is done by: 1) Doing userspace copy inside vhost_net 2) Build XDP buff 3) Batch at most 64 (VHOST_NET_BATCH) XDP buffs and submit them once through msg_control during sendmsg(). 4) Underlayer sockets can use XDP buffs directly when XDP is enalbed, or build skb based on XDP

Hi all: This series tries to batch submitting packets to underlayer socket through msg_control during sendmsg(). This is done by: 1) Doing userspace copy inside vhost_net 2) Build XDP buff 3) Batch at most 64 (VHOST_NET_BATCH) XDP buffs and submit them once through msg_control during sendmsg(). 4) Underlayer sockets can use XDP buffs directly when XDP is enalbed, or build skb based on XDP

Hi all: This series tries to batch submitting packets to underlayer socket through msg_control during sendmsg(). This is done by: 1) Doing userspace copy inside vhost_net 2) Build XDP buff 3) Batch at most 64 (VHOST_NET_BATCH) XDP buffs and submit them once through msg_control during sendmsg(). 4) Underlayer sockets can use XDP buffs directly when XDP is enalbed, or build skb based on XDP

On Thu, Sep 06, 2018 at 12:05:21PM +0800, Jason Wang wrote: > This patch split out XDP logic into a single function. This make it to > be reused by XDP batching path in the following patch. > > Signed-off-by: Jason Wang <jasowang at redhat.com> > --- > drivers/net/tun.c | 84 ++++++++++++++++++++++++++++------------------- > 1 file changed, 51 insertions(+), 33

On Thu, Sep 06, 2018 at 12:05:21PM +0800, Jason Wang wrote: > This patch split out XDP logic into a single function. This make it to > be reused by XDP batching path in the following patch. > > Signed-off-by: Jason Wang <jasowang at redhat.com> > --- > drivers/net/tun.c | 84 ++++++++++++++++++++++++++++------------------- > 1 file changed, 51 insertions(+), 33

On Mon, 21 May 2018 17:04:31 +0800 Jason wrote: > This patch implement build XDP buffers in vhost_net. The idea is do > userspace copy in vhost_net and build XDP buff based on the > page. Vhost_net can then submit one or an array of XDP buffs to > underlayer socket (e.g TUN). TUN can choose to do XDP or call > build_skb() to build skb. To support build skb, vnet header were also

On Thu, Sep 06, 2018 at 12:05:26PM +0800, Jason Wang wrote: > This patch implements XDP batching for vhost_net. The idea is first to > try to do userspace copy and build XDP buff directly in vhost. Instead > of submitting the packet immediately, vhost_net will batch them in an > array and submit every 64 (VHOST_NET_BATCH) packets to the under layer > sockets through msg_control of

On Thu, Sep 06, 2018 at 12:05:26PM +0800, Jason Wang wrote: > This patch implements XDP batching for vhost_net. The idea is first to > try to do userspace copy and build XDP buff directly in vhost. Instead > of submitting the packet immediately, vhost_net will batch them in an > array and submit every 64 (VHOST_NET_BATCH) packets to the under layer > sockets through msg_control of

On Thu, Sep 06, 2018 at 12:05:19PM +0800, Jason Wang wrote: > There's no need to duplicate page get logic in each action. So this > patch tries to get page and calculate the offset before processing XDP > actions, and undo them when meet errors (we don't care the performance > on errors). This will be used for factoring out XDP logic. > > Signed-off-by: Jason Wang

On Thu, Sep 06, 2018 at 12:05:19PM +0800, Jason Wang wrote: > There's no need to duplicate page get logic in each action. So this > patch tries to get page and calculate the offset before processing XDP > actions, and undo them when meet errors (we don't care the performance > on errors). This will be used for factoring out XDP logic. > > Signed-off-by: Jason Wang

We do a get_page() which involves a atomic operation. This patch tries to mitigate a per packet atomic operation by maintaining a reference bias which is initially USHRT_MAX. Each time a page is got, instead of calling get_page() we decrease the bias and when we find it's time to use a new page we will decrease the bias at one time through __page_cache_drain_cache(). Testpmd(virtio_user +

On Mon, 21 May 2018 17:04:25 +0800 Jason wrote: > Instead of mixing zerocopy and datacopy logics, this patch tries to > split datacopy logic out. This results for a more compact code and > specific optimization could be done on top more easily. > > Signed-off-by: Jason Wang <jasowang at redhat.com> > --- > drivers/vhost/net.c | 111

On Thu, Sep 06, 2018 at 12:05:22PM +0800, Jason Wang wrote: > This will allow adding batch flushing on top. > > Signed-off-by: Jason Wang <jasowang at redhat.com> > --- > drivers/net/tun.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/drivers/net/tun.c b/drivers/net/tun.c > index 21b125020b3b..ff1cbf3ebd50 100644 > ---

On Thu, Sep 06, 2018 at 12:05:20PM +0800, Jason Wang wrote: > If we're sure not to go native XDP, there's no need for several things > like bh and rcu stuffs. > > Signed-off-by: Jason Wang <jasowang at redhat.com> True... > --- > drivers/net/tun.c | 9 ++++++--- > 1 file changed, 6 insertions(+), 3 deletions(-) > > diff --git a/drivers/net/tun.c

This patch implements XDP batching for vhost_net with tun. This is done by batching XDP buffs in vhost and submit them when: - vhost_net can not build XDP buff (mostly because of the size of packet) - #batched exceeds the limitation (VHOST_NET_RX_BATCH). - tun accept a batch of XDP buff through msg_control and process them in a batch With this tun XDP can benefit from e.g batch transmission

On Fri, Sep 07, 2018 at 03:41:52PM +0800, Jason Wang wrote: > > > @@ -556,10 +667,14 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock) > > > size_t len, total_len = 0; > > > int err; > > > int sent_pkts = 0; > > > + bool bulking = (sock->sk->sk_sndbuf == INT_MAX); > > What does bulking mean? > > The

Hi all: This series implements packed virtqueues which were described at [1]. In this version we try to address the performance regression saw by V2. The root cause is packed virtqueue need more times of userspace memory accesssing which turns out to be very expensive. Thanks to the help of 7f466032dc9e ("vhost: access vq metadata through kernel virtual address"), such overhead cold be

Hi all: This series implements packed virtqueues which were described at [1]. In this version we try to address the performance regression saw by V2. The root cause is packed virtqueue need more times of userspace memory accesssing which turns out to be very expensive. Thanks to the help of 7f466032dc9e ("vhost: access vq metadata through kernel virtual address"), such overhead cold be

Hi: This series tries to re-enable XDP_REDIRECT for mergeable buffer which was removed since commit 7324f5399b06 ("virtio_net: disable XDP_REDIRECT in receive_mergeable() case"). Main concerns are: - not enough tailroom was reserved which breaks cpumap - complex logic like EWMA and linearizing during XDP processing Fix those by: - reserve enough tailroom during refill - disable EWMA

Hi: This series tries to re-enable XDP_REDIRECT for mergeable buffer which was removed since commit 7324f5399b06 ("virtio_net: disable XDP_REDIRECT in receive_mergeable() case"). Main concerns are: - not enough tailroom was reserved which breaks cpumap - complex logic like EWMA and linearizing during XDP processing Fix those by: - reserve enough tailroom during refill - disable EWMA