diff mbox

[v2,00/18] Multiple fixes & improvements to QIOChannel & Win32

Message ID BLUPR0301MB2034FD833FD5B3E2D7F6E4D59EB50@BLUPR0301MB2034.namprd03.prod.outlook.com (mailing list archive)
State New, archived
Headers show

Commit Message

Andrew Baumann March 11, 2016, 11:51 p.m. UTC
Hi folks,

> From: Paolo Bonzini [mailto:pbonzini@redhat.com]
> Sent: Thursday, 10 March 2016 9:37 AM
> 
> On 10/03/2016 18:26, Daniel P. Berrange wrote:
> > This series started out as an attempt to fix the Win32 problems
> > identified by Andrew Baumann
> >
> >    https://lists.gnu.org/archive/html/qemu-devel/2016-03/msg01343.html
> >
> > It turned into a significantly larger cleanup of some chardev
> > and osdep win32 portability code.
[...]

Sorry for chiming in a bit late here. I've tested these patches (the complete set, not individually), and they do appear to fix my immediate problem: socket char devices now work again. So thank you!

However, I'm now seeing a problem I don't believe we had before: very slow responses to GDB commands. From looking at a packet capture (using a localhost tcp socket between qemu and my gdb client), it seems that a couple of operations will go through just fine, and then there is a 1 second delay between my client's request and qemu's response. After fiddling with poll timeouts, it became clear that we were noticing the socket events when waking up from the poll, but the events themselves were still not waking us. It turns out that we were not calling WSAEventSelect on the accept path. At least, the following patch fixed the problem for me:


However, I'd note that both callers of tcp_chr_new_client() make the same call to set blocking to false immediately before calling tcp_chr_new_client(). Furthermore, the doc comment for qio_channel_set_blocking() appears to suggest that non-blocking mode is the default. If that's true, maybe you don't even want to rely on the caller explicitly setting blocking to false?

Cheers,
Andrew

Comments

Daniel P. Berrangé March 14, 2016, 2:10 p.m. UTC | #1
On Fri, Mar 11, 2016 at 11:51:29PM +0000, Andrew Baumann wrote:
> Hi folks,
> 
> > From: Paolo Bonzini [mailto:pbonzini@redhat.com]
> > Sent: Thursday, 10 March 2016 9:37 AM
> > 
> > On 10/03/2016 18:26, Daniel P. Berrange wrote:
> > > This series started out as an attempt to fix the Win32 problems
> > > identified by Andrew Baumann
> > >
> > >    https://lists.gnu.org/archive/html/qemu-devel/2016-03/msg01343.html
> > >
> > > It turned into a significantly larger cleanup of some chardev
> > > and osdep win32 portability code.
> [...]
> 
> Sorry for chiming in a bit late here. I've tested these patches
> (the complete set, not individually), and they do appear to fix my
> immediate problem: socket char devices now work again. So thank you!

Thanks for confirming this, these patches have now merged into
git msater.

> However, I'm now seeing a problem I don't believe we had before:
> very slow responses to GDB commands. From looking at a packet
> capture (using a localhost tcp socket between qemu and my gdb
> client), it seems that a couple of operations will go through
> just fine, and then there is a 1 second delay between my client's
> request and qemu's response. After fiddling with poll timeouts,
> it became clear that we were noticing the socket events when
> waking up from the poll, but the events themselves were still
> not waking us. It turns out that we were not calling WSAEventSelect
> on the accept path. At least, the following patch fixed the
> problem for me:
> 
> diff --git a/qemu-char.c b/qemu-char.c
> index 3bf30b5..c1be622 100644
> --- a/qemu-char.c
> +++ b/qemu-char.c
> @@ -3047,6 +3047,7 @@ static gboolean tcp_chr_accept(QIOChannel *channel,
>          return TRUE;
>      }
> 
> +    qio_channel_set_blocking(QIO_CHANNEL(sioc), false, NULL);
>      tcp_chr_new_client(chr, sioc);
> 
>      object_unref(OBJECT(sioc));
> 
> However, I'd note that both callers of tcp_chr_new_client()
> make the same call to set blocking to false immediately before
> calling tcp_chr_new_client(). Furthermore, the doc comment for
> qio_channel_set_blocking() appears to suggest that non-blocking
> mode is the default. If that's true, maybe you don't even want
> to rely on the caller explicitly setting blocking to false?

No, the docs don't intend to suggest that - the default is in
fact blocking mode, so its correct to place it into nonblocking
mode explicitly.

I think I didn't notice the problem you describe because my original
patch series had us call WSAEventSelect when creating the watch. This
indirectly puts Win32 sockets into non-blocking mode. The patches which
just merged however no longer call WSAEventSelect when creating the
watch, instead requiring the caller to explicitly set the socket into
non-blocking mode. So I think your suggested addition here is probably
the right way to address this. I'll investigate and respond with a
followup patch as needed.

Regards,
Daniel
diff mbox

Patch

diff --git a/qemu-char.c b/qemu-char.c
index 3bf30b5..c1be622 100644
--- a/qemu-char.c
+++ b/qemu-char.c
@@ -3047,6 +3047,7 @@  static gboolean tcp_chr_accept(QIOChannel *channel,
         return TRUE;
     }

+    qio_channel_set_blocking(QIO_CHANNEL(sioc), false, NULL);
     tcp_chr_new_client(chr, sioc);

     object_unref(OBJECT(sioc));