On Tue, Apr 23, 2013 at 11:41:42AM +0400, Fedorov Sergey wrote: > >Beyond that, we also want to avoid growing net queues indefinitely. If > >the hub does not implement .can_receive() then it relies on growing > >queues (keeping packets buffered in memory). > No, net_hub_receive() calls qemu_send_packet(). If the destination > queue cannot receive the packet qemu_net_queue_append() will take > care of queue->nq_maxlen.
You are right, sorry. We do discard packets at nq_maxlen. The problem with ignoring .can_receive() on the hub is that it breaks flow control. For example, net/tap.c is designed to avoid reading more packets if its peer cannot receive (see tap_can_send()). If the hub claims it can always receive we waste cycles reading packets from the tap device only to discard them. Since qemu.git already has a fix which preserves flow control, I am not going to merge your patch. Stefan
