Re: [PATCH v2 net-next] net: snmp: add statistics for tcp small queue check

[Eric Dumazet <edumazet@google.com>]

On Tue, Nov 30, 2021 at 7:23 AM Jakub Kicinski <kuba@kernel.org> wrote:
>
> On Tue, 30 Nov 2021 22:36:59 +0800 Menglong Dong wrote:
> > On Mon, Nov 29, 2021 at 11:57 PM Jakub Kicinski <kuba@kernel.org> wrote:
> > >
> > > On Sun, 28 Nov 2021 14:01:02 +0800 menglong8.dong@gmail.com wrote:
> > > > Once tcp small queue check failed in tcp_small_queue_check(), the
> > > > throughput of tcp will be limited, and it's hard to distinguish
> > > > whether it is out of tcp congestion control.
> > > >
> > > > Add statistics of LINUX_MIB_TCPSMALLQUEUEFAILURE for this scene.
> > >
> > > Isn't this going to trigger all the time and alarm users because of the
> > > "Failure" in the TCPSmallQueueFailure name?  Isn't it perfectly fine
> > > for TCP to bake full TSQ amount of data and have it paced out onto the
> > > wire? What's your link speed?
> >
> > Well, it's a little complex. In my case, there is a guest in kvm, and virtio_net
> > is used with napi_tx enabled.
> >
> > With napi_tx enabled, skb won't be orphaned after it is passed to virtio_net,
> > until it is released. The point is that the sending interrupt of
> > virtio_net will be
> > turned off and the skb can't be released until the next net_rx interrupt comes.
> > So, wmem_alloc can't decrease on time, and the bandwidth is limited. When
> > this happens, the bandwidth can decrease from 500M to 10M.
> >
> > In fact, this issue of uapi_tx is fixed in this commit:
> > https://lore.kernel.org/lkml/20210719144949.935298466@linuxfoundation.org/
> >
> > I added this statistics to monitor the sending failure (may be called
> > sending delay)
> > caused by qdisc and net_device. When something happen, maybe users can
> > raise ‘/proc/sys/net/ipv4/tcp_pacing_ss_ratio’ to get better bandwidth.
>
> Sounds very second-order and particular to a buggy driver :/
> Let's see what Eric says but I vote revert.

I did some tests yesterday, using one high speed TCP_STREAM (~90Gbit),
and got plenty of increments when using pfifo_fast qdisc.
Yet seed was nominal (bottleneck is the copyout() cost at receiver)

I got few counter increments when qdisc is fq as expected
(because of commit c73e5807e4f6fc6d "tcp: tsq: no longer use
limit_output_bytes for paced flows")


So this new SNMP counter is not a proxy for the kind of problems that a buggy
driver would trigger.

I also suggest we revert this patch.

Thanks !