On Mon, Apr 12, 2021 at 10:16 PM Michael S. Tsirkin [off-list ref] wrote:

quoted

On Fri, Feb 05, 2021 at 02:28:33PM -0800, Wei Wang wrote:

quoted

On Thu, Feb 4, 2021 at 12:48 PM Willem de Bruijn
[off-list ref] wrote:

quoted

On Wed, Feb 3, 2021 at 6:53 PM Wei Wang [off-list ref] wrote:

quoted

On Wed, Feb 3, 2021 at 3:10 PM Michael S. Tsirkin [off-list ref] wrote:

quoted

On Wed, Feb 03, 2021 at 01:24:08PM -0500, Willem de Bruijn wrote:

quoted

On Wed, Feb 3, 2021 at 5:42 AM Michael S. Tsirkin [off-list ref] wrote:

quoted

On Tue, Feb 02, 2021 at 07:06:53PM -0500, Willem de Bruijn wrote:

quoted

On Tue, Feb 2, 2021 at 6:53 PM Willem de Bruijn [off-list ref] wrote:

quoted

On Tue, Feb 2, 2021 at 6:47 PM Wei Wang [off-list ref] wrote:

quoted

On Tue, Feb 2, 2021 at 3:12 PM Michael S. Tsirkin [off-list ref] wrote:

quoted

On Thu, Jan 28, 2021 at 04:21:36PM -0800, Wei Wang wrote:

quoted

With the implementation of napi-tx in virtio driver, we clean tx
descriptors from rx napi handler, for the purpose of reducing tx
complete interrupts. But this could introduce a race where tx complete
interrupt has been raised, but the handler found there is no work to do
because we have done the work in the previous rx interrupt handler.
This could lead to the following warning msg:
[ 3588.010778] irq 38: nobody cared (try booting with the
"irqpoll" option)
[ 3588.017938] CPU: 4 PID: 0 Comm: swapper/4 Not tainted
5.3.0-19-generic #20~18.04.2-Ubuntu
[ 3588.017940] Call Trace:
[ 3588.017942]  <IRQ>
[ 3588.017951]  dump_stack+0x63/0x85
[ 3588.017953]  __report_bad_irq+0x35/0xc0
[ 3588.017955]  note_interrupt+0x24b/0x2a0
[ 3588.017956]  handle_irq_event_percpu+0x54/0x80
[ 3588.017957]  handle_irq_event+0x3b/0x60
[ 3588.017958]  handle_edge_irq+0x83/0x1a0
[ 3588.017961]  handle_irq+0x20/0x30
[ 3588.017964]  do_IRQ+0x50/0xe0
[ 3588.017966]  common_interrupt+0xf/0xf
[ 3588.017966]  </IRQ>
[ 3588.017989] handlers:
[ 3588.020374] [<000000001b9f1da8>] vring_interrupt
[ 3588.025099] Disabling IRQ #38

This patch adds a new param to struct vring_virtqueue, and we set it for
tx virtqueues if napi-tx is enabled, to suppress the warning in such
case.

Fixes: 7b0411ef4aa6 ("virtio-net: clean tx descriptors from rx napi")
Reported-by: Rick Jones <redacted>
Signed-off-by: Wei Wang <redacted>
Signed-off-by: Willem de Bruijn <willemb@google.com>

This description does not make sense to me.

irq X: nobody cared
only triggers after an interrupt is unhandled repeatedly.

So something causes a storm of useless tx interrupts here.

Let's find out what it was please. What you are doing is
just preventing linux from complaining.

The traffic that causes this warning is a netperf tcp_stream with at
least 128 flows between 2 hosts. And the warning gets triggered on the
receiving host, which has a lot of rx interrupts firing on all queues,
and a few tx interrupts.
And I think the scenario is: when the tx interrupt gets fired, it gets
coalesced with the rx interrupt. Basically, the rx and tx interrupts
get triggered very close to each other, and gets handled in one round
of do_IRQ(). And the rx irq handler gets called first, which calls
virtnet_poll(). However, virtnet_poll() calls virtnet_poll_cleantx()
to try to do the work on the corresponding tx queue as well. That's
why when tx interrupt handler gets called, it sees no work to do.
And the reason for the rx handler to handle the tx work is here:
https://lists.linuxfoundation.org/pipermail/virtualization/2017-April/034740.html

Indeed. It's not a storm necessarily. The warning occurs after one
hundred such events, since boot, which is a small number compared real
interrupt load.

Sorry, this is wrong. It is the other call to __report_bad_irq from
note_interrupt that applies here.

quoted

Occasionally seeing an interrupt with no work is expected after
7b0411ef4aa6 ("virtio-net: clean tx descriptors from rx napi"). As
long as this rate of events is very low compared to useful interrupts,
and total interrupt count is greatly reduced vs not having work
stealing, it is a net win.

Right, but if 99900 out of 100000 interrupts were wasted, then it is
surely an even greater win to disable interrupts while polling like
this.  Might be tricky to detect, disabling/enabling aggressively every
time even if there's nothing in the queue is sure to cause lots of cache
line bounces, and we don't want to enable callbacks if they were not
enabled e.g. by start_xmit ...  Some kind of counter?

Yes. It was known that the work stealing is more effective in some
workloads than others. But a 99% spurious rate I had not anticipated.

Most interesting is the number of interrupts suppressed as a result of
the feature. That is not captured by this statistic.

In any case, we'll take a step back to better understand behavior. And
especially why this high spurious rate exhibits in this workload with
many concurrent flows.

I've been thinking about it. Imagine work stealing working perfectly.
Each time we xmit a packet, it is stolen and freed.
Since xmit enables callbacks (just in case!) we also
get an interrupt, which is automatically spurious.

My conclusion is that we shouldn't just work around it but instead
(or additionally?)
reduce the number of interrupts by disabling callbacks e.g. when
a. we are currently stealing packets
or
b. we stole all packets

Agreed. This might prove a significant performance gain at the same time :)

quoted

quoted

Thinking along this line, that probably means, we should disable cb on
the tx virtqueue, when scheduling the napi work on the rx side, and
reenable it after the rx napi work is done?
Also, I wonder if it is too late to disable cb at the point we start
to steal pkts or have stolen all pkts.

The earlier the better. I see no benefit to delay until the rx handler
actually runs.

I've been thinking more on this. I think the fundamental issue here is
that the rx napi handler virtnet_poll() does the tx side work by
calling virtnet_poll_cleantx() without any notification to the tx
side.
I am thinking, in virtnet_poll(), instead of directly call
virtnet_poll_cleantx(), why not do virtqueue_napi_schedule() to
schedule the tx side napi, and let the tx napi handler do the cleaning
work. This way, we automatically call virtqueue_disable_cb() on the tx
vq, and after the tx work is done, virtqueue_napi_complete() is called
to re-enable the cb on the tx side. This way, the tx side knows what
has been done, and will likely reduce the # of spurious tx interrupts?
And I don't think there is much cost in doing that, since
napi_schedule() basically queues the tx napi to the back of its
napi_list, and serves it right after the rx napi handler is done.
What do you guys think? I could quickly test it up to see if it solves
the issue.

Sure pls test. I think you will want to disable event index
for now to make sure disable cb is not a nop (I am working on
fixing that).

Hi Michael and Jason,

I'd like to follow up on this issue a bit more.
I've done some more investigation into this issue:
1. With Michael's recent patch: a7766ef18b336 ("virtio_net: disable cb
aggressively"), we are still seeing this issue with a tcp_stream test
with 240 flows.
2. We've tried with the following patch to suppress cleaning tx queue
from rx napi handler for 10% of the time:

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 79bd2585ec6b..711768dbc617 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -1510,6 +1510,8 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
                return;

        if (__netif_tx_trylock(txq)) {
+               if (virtqueue_more_used(sq->vq) && !prandom_u32_max(10))
+                       goto unlock;
                do {
                        virtqueue_disable_cb(sq->vq);
                        free_old_xmit_skbs(sq, true);

@@ -1518,6 +1520,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
                if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
                        netif_tx_wake_queue(txq);

+unlock:
                __netif_tx_unlock(txq);
        }
 }

This also does not help. It turns out skipping 10% is just not enough.
We have to skip for 50% of the time in order for the warning to be
suppressed.
And this does not seem to be a viable solution since how much we skip
probably will depend on the traffic pattern.

My questions here:
1. Michael mentioned that if we use split queues with event idx, the
interrupts are not actually being disabled. Is this still the case? If
so, is that also the cause for so many spurious interrupts?
2. Michael also submitted another patch: 8d622d21d248 ("virtio: fix up
virtio_disable_cb"). I am not quite sure, would that change help
reduce the # of spurious interrupts we see if we use split queues with
event idx? From my limited understanding, that patch skips calling
virtqueue_disable_cb_split() if event_trigger is set for split queues.

BTW, I have the setup to reproduce this issue easily. So do let me
know if you have other ideas on how to fix it.

Thanks.
Wei