Currently, the spapr-vlan device is trying to flush the RX queue
after each RX buffer that has been added by the guest via the
H_ADD_LOGICAL_LAN_BUFFER hypercall. In case the receive buffer pool
was empty before, we only pass single packets to the guest this
way. This can cause very bad performance if a sender is trying
to stream fragmented UDP packets to the guest. For example when
using the UDP_STREAM test from netperf with UDP packets that are
much bigger than the MTU size, almost all UDP packets are dropped
in the guest since the chances are quite high that at least one of
the fragments got lost on the way.
When flushing the receive queue, it's much better if we'd have
a bunch of receive buffers available already, so that fragmented
packets can be passed to the guest in one go. To do this, the
spapr_vlan_receive() function should return 0 instead of -1 if there
are no more receive buffers available, so that receive_disabled = 1
gets temporarily set for the receive queue, and we have to delay
the queue flushing at the end of h_add_logical_lan_buffer() a little
bit by using a timer, so that the guest gets a chance to add multiple
RX buffers before we flush the queue again.
This improves the UDP_STREAM test with the spapr-vlan device a lot:
Running
netserver -p 44444 -L <guestip> -f -D -4
in the guest, and
netperf -p 44444 -L <hostip> -H <guestip> -t UDP_STREAM -l 60 -- -m 16384
in the host, I get the following values _without_ this patch:
Socket Message Elapsed Messages
Size Size Time Okay Errors Throughput
bytes bytes secs # # 10^6bits/sec
229376 16384 60.00 1738970 0 3798.83
229376 60.00 23 0.05
That "0.05" means that almost all UDP packets got lost/discarded
at the receiving side.
With this patch applied, the value look much better:
Socket Message Elapsed Messages
Size Size Time Okay Errors Throughput
bytes bytes secs # # 10^6bits/sec
229376 16384 60.00 1789104 0 3908.35
229376 60.00 22818 49.85
Signed-off-by: Thomas Huth <thuth@redhat.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
8836630f5d