Re: [Qemu-devel] [PULL 00/14] Net patches

[Li Zhijian]

On 03/15/2016 11:07 AM, Jason Wang wrote:
> On 03/10/2016 11:51 AM, Li Zhijian wrote:
> > On 03/10/2016 10:28 AM, Jason Wang wrote:
> > > On 03/08/2016 05:54 PM, Peter Maydell wrote:
> > > > On 8 March 2016 at 16:06, Zhang Chen <address@hidden>
> > > > wrote:
> > > > > I found the reason for this problem is that
> > > > > unix_connect() have not connect to sock_path before iov_send().
> > > > > It need time to establish connection. so can we fix it with usleep()
> > > > > like this:
> > > > >
> > > > >       recv_sock = unix_connect(sock_path, NULL);
> > > > >       g_assert_cmpint(recv_sock, !=, -1);
> > > > > +    usleep(1000);
> > > > >
> > > > >       ret = iov_send(send_sock[0], iov, 2, 0, sizeof(size) +
> > > > > sizeof(send_buf));
> > > > >       g_assert_cmpint(ret, ==, sizeof(send_buf) + sizeof(size));
> > > > >       close(send_sock[0]);
> > > > >
> > > > >       ret = qemu_recv(recv_sock, &len, sizeof(len), 0);
> > > > I would prefer it if we could find a way to fix this race
> > > > reliably rather than just inserting a delay and hoping it
> > > > is sufficient. Otherwise the test is likely to be unreliable
> > > > if run on a heavily loaded or slow machine.
> > > >
> > > > thanks
> > > > -- PMM
> > >
> > > +1
> > >
> > > To make sure the connected socket to be proceeded before iov_send(), you
> > > could use some like  qmp("{ 'execute' : 'query-status'}") before
> > > iov_send(). With this we are guaranteed that connected is setting to
> > > true before iov_send().
> >
> > it seem works, but i don't know.
> > Is this because that both qemu accepting the connection and qmp
> > command are working under *iothread*,
> > so that when the qemu command returns, we can guaranteed the
> > connection is accepted ?
>
> I think the problem is the race in main loop between the two handlers.
> If the socket netdev read handler was proceed before chardev read
> handler, since connected was false the packet will be dropped silently.
> After we place whateve a qmp command in the middled, it's guaranteed
> that all handlers were proceed after qmp command was executed, so we are
> sure the connected is true when doing iov_send().

Got it, thank for your explain.

we will send V9 with '-chardev socket' and add an qmp sync point soon.

Thanks
Li Zhijian

> > Thanks
> > Li Zhijian
> > >
>
>
>
> .