On Mon, May 25, 2015 at 11:51:23AM +0800, Fam Zheng wrote:
> On Tue, 05/19 15:54, Stefan Hajnoczi wrote:
> > On Tue, May 19, 2015 at 10:51:01AM +0000, Fam Zheng wrote:
> > > This callback is called by main loop before polling s->fd, if it returns
> > > false, the fd will not be polled in this iteration.
> > > 
> > > This is redundant with checks inside read callback. After this patch,
> > > the data will be copied from s->fd to s->iov when it arrives. If the
> > > device can't receive, it will be queued to incoming_queue, and when the
> > > device status changes, this queue will be flushed.
> > > 
> > > Also remove the qemu_can_send_packet() check in netmap_send. If it's
> > > true, we are good; if it's false, the qemu_sendv_packet_async would
> > > return 0 and read poll will be disabled until netmap_send_completed is
> > > called.
> > 
> > This causes unbounded memory usage in QEMU because
> > qemu_net_queue_append_iov() does not drop packets when sent_cb != NULL.
> 
> I think netmap_send will use "netmap_read_poll(s, false)" to stop reading, only
> the first packet will be queued. Why is it unbounded?

I looked again and I agree with you.  It should stop after the first
packet and resume when the peer flushes the queue.

Stefan