On 10/05/17 09:56, Jason Wang wrote:
On 2017年05月10日 13:28, Anton Ivanov wrote:
On 10/05/17 03:18, Jason Wang wrote:
On 2017年05月09日 23:11, Stefan Hajnoczi wrote:
On Tue, May 09, 2017 at 08:46:46AM +0100, Anton Ivanov wrote:
I have figured it out. Two issues.
1) skb->xmit_more is hardly ever set under virtualization because
the qdisc
is usually bypassed because of TCQ_F_CAN_BYPASS. Once
TCQ_F_CAN_BYPASS is
set a virtual NIC driver is not likely see skb->xmit_more (this
answers my
"how does this work at all" question).
2) If that flag is turned off (I patched sched_generic to turn it
off in
pfifo_fast while testing), DQL keeps xmit_more from being set. If
the driver
is not DQL enabled xmit_more is never ever set. If the driver is DQL
enabled
the queue is adjusted to ensure xmit_more stops happening within
10-15 xmit
cycles.
That is plain *wrong* for virtual NICs - virtio, emulated NICs, etc.
There,
the BIG cost is telling the hypervisor that it needs to "kick" the
packets.
The cost of putting them into the vNIC buffers is negligible. You
want
xmit_more to happen - it makes between 50% and 300% (depending on
vNIC
design) difference. If there is no xmit_more the vNIC will
immediately
"kick" the hypervisor and try to signal that the packet needs to
move
straight away (as for example in virtio_net).
How do you measure the performance? TCP or just measure pps?
In this particular case - tcp from guest. I have a couple of other
benchmarks (forwarding, etc).
One more question, is the number for virtio-net or other emulated vNIC?
Other for now - you are cc-ed to keep you in the loop.
Virtio is next on my list - I am revisiting the l2tpv3.c driver in QEMU
and looking at how to preserve bulking by adding back sendmmsg (as well
as a list of other features/transports).
We had sendmmsg removed for the final inclusion in QEMU 2.1, it
presently uses only recvmmsg so for the time being it does not care.
That will most likely change once it starts using sendmmsg as well.
In addition to that, the perceived line rate is proportional to this
cost,
so I am not sure that the current dql math holds. In fact, I think
it does
not - it is trying to adjust something which influences the
perceived line
rate.
So - how do we turn BOTH bypass and DQL adjustment while under
virtualization and set them to be "always qdisc" + "always xmit_more
allowed"
Virtio-net net does not support BQL. Before commit ea7735d97ba9
("virtio-net: move free_old_xmit_skbs"), it's even impossible to
support that since we don't have tx interrupt for each packet. I
haven't measured the impact of xmit_more, maybe I was wrong but I
think it may help in some cases since it may improve the batching on
host more or less.
If you do not support BQL, you might as well look the xmit_more part
kick code path. Line 1127.
bool kick = !skb->xmit_more; effectively means kick = true;
It will never be triggered. You will be kicking each packet and per
packet.
Probably not, we have several ways to try to suppress this on the
virtio layer, host can give hints to disable the kicks through:
- explicitly set a flag
- implicitly by not publishing a new event idx
FYI, I can get 100-200 packets per vm exit when testing 64 byte
TCP_STREAM using netperf.
I am aware of that. If, however, the host is providing a hint we might
as well use it.
xmit_more is now set only out of BQL. If BQL is not enabled you
never get it. Now, will the current dql code work correctly if you do
not have a defined line rate and completion interrupts - no idea.
Probably not. IMHO instead of trying to fix it there should be a way for
a device or architecture to turn it off.
In fact BQL is not the only user for xmit_more. Pktgen with burst is
another. Test does not show obvious difference if I set burst from 0
to 64 since we already had other ways to avoid kicking host.
That, as well as this not being wired to bulk transport.
To be clear - I ran into this working on my own drivers for UML, you are
cc-ed because you are likely to be one of the most affected.
I'm still not quite sure the issue. Looks like virtio-net is ok since
BQL is not supported and the impact of xmit_more could be ignored.
Presently - yes. If you have bulk aware transports to wire into that is
likely to make a difference.
Thanks
A.
Thanks
A.
P.S. Cc-ing virtio maintainer
CCing Michael Tsirkin and Jason Wang, who are the core virtio and
virtio-net maintainers. (I maintain the vsock driver - it's unrelated
to this discussion.)
A.
On 08/05/17 08:15, Anton Ivanov wrote:
Hi all,
I was revising some of my old work for UML to prepare it for
submission
and I noticed that skb->xmit_more does not seem to be set any more.
I traced the issue as far as net/sched/sched_generic.c
try_bulk_dequeue_skb() is never invoked (the drivers I am working
on are
dql enabled so that is not the problem).
More interestingly, if I put a breakpoint and debug output into
dequeue_skb() around line 147 - right before the bulk: tag that skb
there is always NULL. ???
Similarly, debug in pfifo_fast_dequeue shows only NULLs being
dequeued.
Again - ???
First and foremost, I apologize for the silly question, but how can
this
work at all? I see the skbs showing up at the driver level, why are
NULLs being returned at qdisc dequeue and where do the skbs at the
driver level come from?
Second, where should I look to fix it?
A.
--
Anton R. Ivanov
Cambridge Greys Limited, England company No 10273661
http://www.cambridgegreys.com/
--
Anton R. Ivanov
Cambridge Greys Limited, England company No 10273661
http://www.cambridgegreys.com/
