On Wed, Oct 30, 2013 at 02:29:00PM +0400, Fedorov Sergey wrote:
> On 10/29/2013 06:55 PM, Stefan Hajnoczi wrote:
> >On Mon, Oct 21, 2013 at 03:44:46PM +0400, Fedorov Sergey wrote:
> >>After our discussion about this patch I decided to keep my patch in
> >>our branch until rebase onto a new release. Recently I have rebased
> >>our branch onto v1.5.3 and reverted my patch. Then I face an issue
> >>when using user-mode networking with USB network device for mounting
> >>root file system through NFS. Fragmented UDP packets from host to
> >>guest does not handled properly. Seems that some fragments is lost
> >>or somehow stalled. See guest tcpdump log below.
> >>
> >>03:16:52.259690 IP (tos 0x0, ttl 64, id 0, offset 0, flags [DF],
> >>proto UDP (17), length 164)
> >> 10.0.2.15.3369105030 > 10.0.2.2.nfs: 136 readdirplus fh
> >> Unknown/0100070004001200000000002873593C9B3C43388E23748B0BAD870C00000000
> >>512 bytes @ 0 max 4096 verf 0000000000000000
> >>03:16:52.262323 IP (tos 0x0, ttl 64, id 16, offset 0, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2.nfs > 10.0.2.15.3369105030: reply ok 1472 readdirplus
> >>POST: DIR 40777 ids 0/0 sz 4096 verf 0000000000000000
> >>03:16:52.264592 IP (tos 0x0, ttl 64, id 16, offset 1480, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2 > 10.0.2.15: udp
> >>03:16:54.462961 IP (tos 0x0, ttl 64, id 0, offset 0, flags [DF],
> >>proto UDP (17), length 164)
> >> 10.0.2.15.3369105030 > 10.0.2.2.nfs: 136 readdirplus fh
> >> Unknown/0100070004001200000000002873593C9B3C43388E23748B0BAD870C00000000
> >>512 bytes @ 0 max 4096 verf 0000000000000000
> >>03:16:54.466300 IP (tos 0x0, ttl 64, id 17, offset 0, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2.nfs > 10.0.2.15.3369105030: reply ok 1472 readdirplus
> >>POST: DIR 40777 ids 0/0 sz 4096 verf 0000000000000000
> >>03:16:54.467084 IP (tos 0x0, ttl 64, id 17, offset 1480, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2 > 10.0.2.15: udp
> >>...
> >>
> >>I didn't investigate the cause of the problem in detail. I just reverted
> >>
> >>commit 199ee608f0d08510b5c6c37f31a7fbff211d63c4
> >>Author: Luigi Rizzo <ri...@iet.unipi.it>
> >>Date: Tue Feb 5 17:53:31 2013 +0100
> >>
> >> net: fix qemu_flush_queued_packets() in presence of a hub
> >>
> >>And then applied my patch. After that everything works fine for me.
> >>See guest tcpdump log below.
> >>
> >>04:45:15.897245 IP (tos 0x0, ttl 64, id 0, offset 0, flags [DF],
> >>proto UDP (17), length 164)
> >> 10.0.2.15.3642011847 > 10.0.2.2.nfs: 136 readdirplus fh
> >> Unknown/0100070004001200000000002873593C9B3C43388E23748B0BAD870C00000000
> >>512 bytes @ 0 max 4096 verf 0000000000000000
> >>04:45:15.899686 IP (tos 0x0, ttl 64, id 15, offset 0, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2.nfs > 10.0.2.15.3642011847: reply ok 1472 readdirplus
> >>POST: DIR 40777 ids 0/0 sz 4096 verf 0000000000000000
> >>04:45:15.906253 IP (tos 0x0, ttl 64, id 15, offset 1480, flags [+],
> >>proto UDP (17), length 1500)
> >> 10.0.2.2 > 10.0.2.15: udp
> >>04:45:15.906687 IP (tos 0x0, ttl 64, id 15, offset 2960, flags
> >>[none], proto UDP (17), length 240)
> >> 10.0.2.2 > 10.0.2.15: udp
> >>
> >>So there must be something wrong with already applied patch. What
> >>could you suggest?
> >The next step is to investigate the cause.
> >
> >Perhaps hw/usb/dev-network.c:usb_net_handle_datain() is not calling
> >qemu_flush_queued_packets() every time in_buf[] is read completely.
> >This if statement looks strange to me:
> >
> >if (s->in_ptr >= s->in_len &&
> > (is_rndis(s) || (s->in_len & (64 - 1)) || !len)) {
> > /* no short packet necessary */
> > usb_net_reset_in_buf(s);
> >}
> >
> >Try placing printfs to find out whether qemu_flush_queued_packets() is
> >getting called when you see packet loss.
> >
> >Stefan
> >
>
> Seems that I have figured out the problem. net_hub_flush() does not
> flush source port. And qemu_flush_queued_packets() also returns
> after calling net_hub_flush(). So I think the problem is that
> neither qemu_flush_queued_packets() nor net_hub_flush() call
> qemu_net_queue_flush() for the source port. So I think it sould be
> fixed in qemu_flush_queued_packets() by removing the return
> statement after calling net_hub_flush(). That fix does work for me.
> So I could submit that patch after getting permission for that.
Sounds good to me.
I have CCed Luigi in case he wants to comment.
Stefan
