From: Shachar Raindel <shacharr@xxxxxxxxxxxxx> The batching logic in netvsc_send is non-trivial, due to a combination of the Linux API and the underlying hypervisor interface. Add a comment explaining why the code is written this way. Signed-off-by: Shachar Raindel <shacharr@xxxxxxxxxxxxx> Signed-off-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx> --- .../ethernet/microsoft/netvsc.rst | 14 ++++++++----- drivers/net/hyperv/netvsc.c | 20 +++++++++++++++++++ 2 files changed, 29 insertions(+), 5 deletions(-) diff --git a/Documentation/networking/device_drivers/ethernet/microsoft/netvsc.rst b/Documentation/networking/device_drivers/ethernet/microsoft/netvsc.rst index c3f51c672a68..fc5acd427a5d 100644 --- a/Documentation/networking/device_drivers/ethernet/microsoft/netvsc.rst +++ b/Documentation/networking/device_drivers/ethernet/microsoft/netvsc.rst @@ -87,11 +87,15 @@ Receive Buffer contain one or more packets. The number of receive sections may be changed via ethtool Rx ring parameters. - There is a similar send buffer which is used to aggregate packets for sending. - The send area is broken into chunks of 6144 bytes, each of section may - contain one or more packets. The send buffer is an optimization, the driver - will use slower method to handle very large packets or if the send buffer - area is exhausted. + There is a similar send buffer which is used to aggregate packets + for sending. The send area is broken into chunks, typically of 6144 + bytes, each of section may contain one or more packets. Small + packets are usually transmitted via copy to the send buffer. However, + if the buffer is temporarily exhausted, or the packet to be transmitted is + an LSO packet, the driver will provide the host with pointers to the data + from the SKB. This attempts to achieve a balance between the overhead of + data copy and the impact of remapping VM memory to be accessible by the + host. XDP support ----------- diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c index dc3f73c3b33e..dc333dceb055 100644 --- a/drivers/net/hyperv/netvsc.c +++ b/drivers/net/hyperv/netvsc.c @@ -1006,6 +1006,26 @@ static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send, } /* RCU already held by caller */ +/* Batching/bouncing logic is designed to attempt to optimize + * performance. + * + * For small, non-LSO packets we copy the packet to a send buffer + * which is pre-registered with the Hyper-V side. This enables the + * hypervisor to avoid remapping the aperture to access the packet + * descriptor and data. + * + * If we already started using a buffer and the netdev is transmitting + * a burst of packets, keep on copying into the buffer until it is + * full or we are done collecting a burst. If there is an existing + * buffer with space for the RNDIS descriptor but not the packet, copy + * the RNDIS descriptor to the buffer, keeping the packet in place. + * + * If we do batching and send more than one packet using a single + * NetVSC message, free the SKBs of the packets copied, except for the + * last packet. This is done to streamline the handling of the case + * where the last packet only had the RNDIS descriptor copied to the + * send buffer, with the data pointers included in the NetVSC message. + */ int netvsc_send(struct net_device *ndev, struct hv_netvsc_packet *packet, struct rndis_message *rndis_msg, -- 2.25.1